Advances in Geoscience and Remote Sensing PDF

I
Advances in Geoscience
and Remote Sensing
Advances in Geoscience
and Remote Sensing
Edited by
Dr. Gary Jedlovec
In-Tech
intechweb.org
Published by In-Teh
In-Teh
Olajnica 19/2, 32000 Vukovar, Croatia
Abstracting and non-profit use of the material is permitted with credit to the source. Statements and
opinions expressed in the chapters are these of the individual contributors and not necessarily those of
the editors or publisher. No responsibility is accepted for the accuracy of information contained in the
published articles. Publisher assumes no responsibility liability for any damage or injury to persons or
property arising out of the use of any materials, instructions, methods or ideas contained inside. After
this work has been published by the In-Teh, authors have the right to republish it, in whole or part, in any
publication of which they are an author or editor, and the make other personal use of the work.
© 2009 In-teh
www.intechweb.org
Additional copies can be obtained from:
publication@intechweb.org
First published October 2009

Printed in India
Technical Editor: Goran Bajac
Advances in Geoscience and Remote Sensing,

Edited by Dr. Gary Jedlovec
p. cm.
ISBN 978-953-307-005-6
V
Preface
Remote sensing is the acquisition of information of an object or phenomenon, by the

use of either recording or real-time sensing device(s), that is not in physical or inti-
mate contact with the object (such as by way of aircraft, spacecraft, satellite, buoy,
or ship). In practice, remote sensing is the stand-off collection through the use of a
variety of devices for gathering information on a given object or area. Human exis-
tence is dependent on our ability to understand, utilize, manage and maintain the
environment we live in - Geoscience is the science that seeks to achieve these goals.
This book is a collection of contributions from world-class scientists, engineers and
educators engaged in the fields of geoscience and remote sensing. The content of the
book includes selected topics covering instrumentation (SAR, lidar, and radar), and
remote sensing techniques to study atmospheric air quality and clouds, oceanogra-
phy, in-land lake water quality and sea ice, soil moisture, vegetation mapping, and
earthquake damage. In also includes several chapters on web tools for GIS applica-
tions, and the use of UAVs for low cost remote sensing studies. This book provides
a particularly good way for experts in one aspect of the field to learn about advances
made by their colleagues with different research interests and for students to broad-
en their exposure to the topic area.
Dr. Gary Jedlovec

Earth Science Office, VP61
NASA/MSFC
320 Sparkman Drive
Huntsville, Alabama 35805
USA
VI
VII
Contents
Preface V
1. Enhancing the Unmixing Algorithm through the Spatial Data Modeling

for Limnological Studies 1
Enner Herenio Alcântara, José Luiz Stech, Evlyn Márcia Leão de Moraes Novo
and Cláudio Clemente Faria Barbosa
2. Application of remote sensing to the estimation of sea ice thickness distribution 21

Takenobu Toyota
3. Correction of NDVI calculated from ASTER L1B and ASTER (AST07) data based
on ground measurement 45
Hoshino Buho, Masami Kaneko and Kenta Ogawa
4. Environmental hazards in the El-Temsah Lake, Suez Canal district, Egypt 57

M.F. Kaiser, A.S. Amin and H.A. Aboulela
5. Air Pollution Determination Using Remote Sensing Technique 71

H. S. Lim, M. Z. MatJafri, K. Abdullah and C. J. Wong
6. Some Features of The Volume Component of Radar Backscatter From Thick

and Dry Snow Cover 93
Boris Yurchak
7. Theoretical Modeling for Polarimetric Scattering and Information Retrieval

of SAR Remote Sensing 141
Ya-Qiu Jin
8. Polarimetric responses and scattering mechanisms of tropical forests

in the Brazilian Amazon 183
J. R. dos Santos, I. S. Narvaes, P. M. L. A. Graça and F. G. Gonçalves
9. Discovery of user-generated geographic data using web search engines 207

Carlos Abargues, Carlos Granell, Laura Díaz, Joaquín Huerta and Arturo Beltran
10. Estimation of soil properties using observations and the crop model STICS.
Interest of global sensitivity analysis and impact on the prediction
of agro-environmental variables 229
Hubert Varella, Martine Guérif and Samuel Buis
VIII
11. Modeling of wake-vortex detection by a ground-based fiber LIDAR system 249

S. Brousmiche, L. Bricteux, G.Winckelmans, B. Macq and P. Sobieski
12. Earthquake damage mapping techniques using SAR and Optical remote sensing
satellite data 269
Marco Chini
13. Spectroscopic Microwave Dielectric Model of Moist Soils 279

Valery Mironov and Pavel Bobrov
14. Automated Detection of Clouds in Satellite Imagery 303

Gary Jedlovec
15. Remote sensing and the disaster management cycle 317

Karen E. Joyce, Kim C. Wright, Sergey V. Samsonov, and Vincent G. Ambrosia
16. Prediction of Volumetric Shrinkage in Expansive Soils (Role of Remote Sensing) 347
Fekerte Arega Yitagesu, Freek van der Meer and Harald van der Werff
17. Monitoring Soil Moisture from Spaceborne Passive Microwave Radiometers:

Algorithm Developments and Applications to AMSR-E and SSM/I 371
Hui Lu, Toshio Koike, Tetsu Ohta, David Ndegwa Kuria, Kun Yang, Hideyuki Fujii,
Hiroyuki Tsutsui and Katsunori Tamagawa
18. Multiwavelength Polarimetric Lidar for Foliage Obscured Man-Made

Target Detection 391
Songxin Tan and Jason Stoker
19. Unmixing Based Landsat ETM+ and ASTER Image Fusion for Hybrid
Multispectral Image Analysis 407
Nouha Mezned, Sâadi Abdeljaoued and Mohamed Rached Boussema
20. InSAR phase analysis: Phase unwrapping for noisy SAR interferograms 419
Riadh Abdelfattah
21. Soil moisture estimation using L-band radiometry 441

Alessandra Monerris and Thomas Schmugge
22. AggieAir: Towards Low-cost Cooperative Multispectral Remote Sensing Using

Small Unmanned Aircraft Systems 463
Haiyang Chao, Austin M. Jensen, Yiding Han, YangQuan Chen and Mac McKee
23. Possibilistic and fuzzy multi-sensor fusion for humanitarian mine action 491
Nada Milisavljević and Isabelle Bloch
24. Non specular reflection and depolarisation due to walls under oblique incidence 505
Iñigo Cuiñas, Manuel G. Sánchez and Ana V. Alejos
25. Dynamical Enhancement Technique for Geophysical Analysis of Remote

Sensing Imagery 515
Iván E. Villalón-Turrubiates
IX
26. Forest Inventory using Optical and Radar Remote Sensing 539
Yolanda Fernández-Ordóñez, Jesús Soria-Ruiz and Brigitte Leblon
27. Application of Multi-Frequency Synthetic Aperture Radar (SAR)

in Crop Classification 557
Jiali Shang, Heather McNairn, Catherine Champagne and Xianfeng Jiao
28. Multispectal Image Classification Using Rough Set Theory and Particle
Swam Optimization 569
Chih-Cheng Hung, Hendri Purnawan, Bor-Chen Kuo and Scott Letkeman
29. Derivative analysis of hyperspectral oceanographic data 597

Elena Torrecilla, Jaume Piera and Meritxell Vilaseca
30. Electromagnetic Scattering Analysis of Simple Targets Embedded in Planar

Multilayer Structures: Remote Sensing Applications 619
Sidnei J. S. Sant’Anna, J. C. da S. Lacava and David Fernandes
31. On Performance of S-band FMCW Radar for Atmospheric Measurements 645

Turker Ince
32. Use of Lidar Data in Floodplain Risk Management Planning: The Experience
of Tabasco 2007 Flood 659
Ramos J., Marrufo L. and González F.J.
33. Improving Wetland Characterization with Multi-Sensor, Multi-Temporal SAR

and Optical/Infrared Data Fusion 679
Laura L. Bourgeau-Chavez, Kevin Riordan, Richard B. Powell, Nicole Miller
and Mitch Nowels
34. Remote Sensing Rock Mechanics and Earthquake Thermal Infrared Anomalies 709
Lixin Wu and Shanjun Liu
X
Enhancing the Unmixing Algorithm through the Spatial Data Modeling for Limnological Studies 1
X1
Enhancing the Unmixing Algorithm through the

Spatial Data Modeling for Limnological Studies
Enner Herenio Alcântara 1
José Luiz Stech 1
Evlyn Márcia Leão de Moraes Novo 1
Cláudio Clemente Faria Barbosa 2
1Remote Sensing Division, Brazilian Institute for Space Research
{enner,stech,evlyn}@dsr.inpe.br
2 Imaging Processing Division, Brazilian Institute for Space Research
claudio@dpi.inpe.br
São José dos Campos, São Paulo, Brazil
1. Introduction
Conventional water quality monitoring is expensive and time consuming. This is
particularly problematic if the water bodies to be examined are large. Conventional
techniques also bring about a high probability of undersampling (Hadjimitsis et al. 2006).
Conversely, remote sensing is a powerful tool to assess aquatic systems and is particularly
useful in remote areas such as the Amazon lakes (Alcântara et al., 2008).
Data collected using this technique can provide a synoptic overview of such large aquatic
environments, which could otherwise not be observed at a glance (Dekker et al. 1995).
However, remote sensing is not easily applied to aquatic environment monitoring mainly
because the mixture of the optically active substances (OAS) in the water. Several
approaches have been proposed to cope with this issue such as derivative analysis (Goodin
et al. 1993), the continuum removal (Kruse et al. 1993), and spectral mixture analysis (Novo
and Shimabukuro, 1994; Oyama et al. 2009).
The two first approaches are more suitable for hyperspectral images, whereas the spectral
mixture analysis can be used for both hyperspectral and multispectral images. The Spectral
Mixture Model (SMM) has largely been used for spectral mixture analysis, uncoupling the
reflectance of each image pixel (Tyler et al. 2006) into the proportion of each water
component contributing to the signal. The result of a spectral mixture analysis is a set of
fraction images representing the proportion of each water component per image pixel. This
technique has been applied to TM/Landsat images to determine the concentration of
suspended particles (Mertes et al. 1993), chlorophyll-a concentration (Novo and
Shimabukuro, 1994); as well as to MODIS images, to determine the chlorophyll-a
concentration in the Amazon floodplain (Novo et al. 2006), to characterize the composition
2 Advances in Geoscience and Remote Sensing
of optically complex waters in the Amazon (Rudorff et al. 2007), and to study turbidity
distribution in the Amazon floodplain (Alcântara et al. 2008, 2009a).
Remote sensing data have been extensively used to detect and to quantify water quality
variables in lakes and reservoirs (Kloiber et al. 2002). One of the most important variables to
monitor water quality is turbidity, because it gives information on underwater light
availability (Alcântara et al. 2009b). Although turbidity is caused by organic and inorganic
particles, one unresolved issue is to distinguish between them using remote sensing (Wetzel,
2001). The Spectral Unmixing Model (SMM) can, however, be useful to analyze the turbidity
caused by inorganic particles and by phytoplankton cell scattering.
This chapter book shows how to improve the well know unmixing algorithm using the
spatial data modelling concept and also their applicability in limnological studies.
2. Study Area
The Amazon River basin drains an area of approximately 6x106 km2, which represents 5% of
the Earth surface. The Central Amazon has large floodplains covering around 300,900 km2
(Hess et al. 2003), including 110,000 km2 of the main stem ‘Varzeas’ (white water river
floodplains; Junk, 1997). At high water, the Amazon River flows into the floodplains, and
fills both temporary and permanent lakes which might merge to each other. The ‘Lago
Curuai’ floodplain (Figure 1) covers an area varying from 1340 to 2000 km2 from low to the
high water. This floodplain is located 850 km from the Atlantic Ocean, near Óbidos (Pará,
Brazilian Amazon). Formed by ‘white’ water lakes (characterized by high concentration of
suspended sediments), and ‘black’ water lakes (with a high concentration of dissolved
organic matter, and a low concentration of sediments; Moreira-Turcq et al. 2004). These
lakes are connected to each other and to the Amazon River. The floodplain also has ‘clear’
water lakes filled by both rainwater and by river- water drained from ‘Terra Firme’
(Barbosa, 2005).
The lakes are connected to each other and to the Amazon River. The Curuai floodplain is
controlled by the Amazon River flood pulse (Moreira-Turcq et al. 2004) which creates four
states (Barbosa, 2005) in the floodplain-river system (Figure 2): (1) rising water level
(January - February), (2) high water level (April – June), (3) receding water level (August –
October), and (4) low water level (November – December).
The exchange of water between the floodplain and the Amazon River is shown in Figure 2.
When the water level is high, there is very little flow, and the surface water circulation is
caused predominantly by wind. In the receding state, the exchange of water between the
river and the floodplain is inverted, i.e., the water flows from the floodplain to the river. The
water level then drops to the low water state, when the exchange of water between the river
and the floodplain is at a minimum. According to Barbosa (2005), during the rising water
state, the flow from the river to the floodplain starts at a channel located on its Eastern
border, and then migrates to small channels located on its Northwestern side.
The rate of inundation is influenced by the floodplain geomorphology, by the density of
floodplain channels, and by the ratio of local drainage basin area to lake area. A majority of
93% of the flooded area in the floodplain is between 2-6 m in depth (using a water level
reference of 936 cm). The deepest lake is the ‘Lago Grande’, and the shallowest are the ‘Açaí’
and Santa Ninha Lakes (Figure 1-c). An area of about 0.04% of the floodplain is below sea
level (with the mean altitude of the floodplain at 9 m above sea level (Barbosa, 2005).
(a) (b)
(c)
Fig. 1. Legal Amazon location into Brazil (a); (b) Location of Curuai floodplain into Amazon
Legal; (c) Landsat-5 Thematic Mapper (normal composition) imagery showing the Curuai
Floodplain and their bathymetry. The arrows indicate the main channels of connection
Amazon River-floodplain.
Fig. 2. Theoretical water level dynamic at the Curuai Floodplain (Source: Barbosa, 2005).
3. In-situ Measurements and Remote Sensing Data

The turbidity ground data acquisition was carried out from February 1st to February 14th
2004, during the rising water period. Turbidity measurements were taken at 215 sampling
stations using the HORIBA U-10 multi-sensor. This equipment provides turbidity
measurements in NTU (Nephelometric Turbidity Unit) with a resolution of 1 NTU. The
locations of the sampling stations were determined with the aid of spectral analyses of
Landsat/TM images taken at similar water level (Barbosa, 2005). These samples had
maximum, minimum and mean values of 569, 101 and 232.29 NTU, respectively.
A Terra/MODIS image, acquired as MOD09 product on February 27th 2004 was used in this
study. The spectral bands used in the analyses were band 1 (620-670 nm) and 2 (841-876)
with a spatial resolution of 250 m, and bands 3 (459-479 nm) and 4 (545-565 nm) with a
spatial resolution of 500 m. The two latter bands were re-sampled to 250 m using MODIS
Reprojection Tool software (MRT, 2002).
4. Methodological Approach
The turbidity distribution was assessed using fraction images derived from the Linear
Spectral Mixing Model, using four MODIS spectral bands (3 – blue, 4 – green, 1 – red and 2 –
near infrared) with a spatial resolution of 250 m. In order to evaluate the turbidity
distribution observed in the MODIS fraction images, in-situ measurements acquired during
in February 2004 (a few days apart of the MODIS acquisition) were used to apply the
Ordinary Least Square (OLS), spatial lag, and spatial error regression models. The kernel
estimation algorithm was used to verify the spatial correlation of the in-situ data before
performing the regression analyses. A summary of our methodological approach is
presented in Figure 3:
Fig. 3. Flow chart of the methodological approach.
4.1 The SMM

The SMM estimates the proportion of the various surface components present in each image
pixel based on their spectral characteristics (Novo and Shimabukuro, 1994). The number of
spectral endmembers used in the SMM algorithm must be less than or equal to the number
of spectral bands (Tyler et al. 2006). Given these conditions, it is possible to determine the
proportion of each component by knowing the spectral responses of pixel components
according to equation 1:
n
Ris = ∑ f j rij + ε i (1)
j =1
Where: Ris is the reflectance at each spectral band i of a pixel with one or more
components, rij is the spectral reflectance of each component j in the spectral band i; fj is
the fraction of the component within the pixel; εi is an error term for each spectral band i.
Also, j = 1, 2, ..., n ( n = number of components), and i = 1, 2, ..., m (m = number of
spectral bands), with the following constraints: ∑f j = 1, and fj ≥0 for all
components.
4.1.1 Endmember selection

As pointed out in (Rudorff et al. 2006), mixtures of dissolved or suspended materials will
always occur in natural water bodies. The desired conceptual ‘pure’ endmembers are hence
not accessible for the OAS. Thus, the SMM results will not lead to a complete separation of
the fractions. They will rather indicate a relative proportion of each endmember in which
the relationship with the actual concentration of a certain OAS will be stronger according to
its reflectance spectral dominance.
Some authors have selected the endmembers in a spectral library or laboratory
measurements and applied them in the satellite images (Mertes et al. 1993). This approach
for selecting the most ‘pure pixel’ for each component in the water sometimes does not
consider the actual characteristics of endmembers found in the local area (Theseira et al.
2003). Thus, some authors collect the endmembers directly in the image, the so called image
endmember (Novo and Shimabukuro, 1994).
The endemembers that will be selected in this work are the phytoplankton, inorganic
suspended particle and dissolved organic matter laden water. The phytoplankton and the
inorganic particle can cause the turbidity in the water and the dissolved organic matter, on
the other hand, is a representative manner of non-turbid water. The method of Alcântara et
al. (2008) was used to select the endmembers.
Fig. 4. Image endmembers used as input to run the Linear Spectral Mixing Model on the
Terra/MODIS image acquired on 27th February 2004. Chl is the phytoplankton laden water,
Dom is the dissolved organic matter laden water, and Iss is the inorganic particle laden
water.
Figure 4 shows that the spectral responses of the selected endmembers were quite different.
Chl-laden water endmember is characterized by high reflectance in the green band, typical
of phytoplankton-laden waters. Dom-laden water endmember presents low reflectance at all
wavelengths and the Iss-laden water endmember is characterized by an increasing
reflectance towards to visible wavelength, and decreasing slightly at the near infrared.
This model was used to estimate the proportion of each component in the water using the
following components: phytoplankton (Chl) laden water, dissolved organic matter (Dom)
laden water, and inorganic particle (Iss) laden water. These components (endmembers) were
selected based on the dominance of them in the surface water. Then, to estimate the
proportion of components into the water we used the following equation:
Ris = f Chl rChl + f Dom rDom + f Iss rIss + ε i (2)
Where, Ris is the reflectance at the ith band at sampling station s ; f Chl , f Dom and f Iss
are the fractions phytoplankton laden water, dissolved organic matter-laden water and
inorganic matter-laden water, respectively; rChl , rDom and rIss are the reflectance of
phytoplankton laden water, dissolved organic matter-laden water and inorganic matter-
laden water, respectively and εi is the error at the ith spectral bands.
The main idea of this study was to estimate the proportion of each of the three components
in the surface water. Ground data of Chl, Dom, and Iss concentration in water was then
used to understand the distribution of turbidity in the Curuai floodplain lake. The results
obtained with the SMM were compared with in-situ turbidity measurements through the
interpolation, using a geostatistical method, called Ordinary Kriging.
4.2 Ordinary Kriging

A turbidity map was generated with the in-situ data, for subsequent comparison with the
fraction images generated using the SMM. Thereby, it could be verified to which extent the
two data sets matched. To produce such a reference map, Ordinary Kriging was used,
interpolating in-situ turbidity measurements, as described in (Isaaks and Srivastava, 1989).
Ordinary Kriging is a technique of making optimal, unbiased of regionalized variables at
unsampled locations using the structural properties of the semivariogram and the initial set
of data values. The calculation of the Kriging weights is based upon the estimation of a
semivariogram model, as described by equation 3:
1
γ (h) = Var[ z ( s + h) − z ( s )] (3)
2
Where: γ (h) is an estimated value of the semivariance for the lag h. The estimation of a
semivariogram model relies on two important assumptions: (1) the parameter γ (h) exists
and is finite for all choices of h and s, and that it do not depend on s. The Ordinary Kriging
estimator is presented in:
n
Z ( x, y ) = ∑ wi z i (4)
i =1
Where n is the number of considered measures, zi are the corresponding attribute values,
and wi are the weights.
The experimental semivariogram was fitted with various theoretical models (spherical,
exponential, Gaussian, linear and power) using the weighted least square method. The
theoretical model that gives minimum RMSE is chosen for further analysis. In this case, the
fitted model was based on Gaussian model (Table 1).
Water Anisotropy Structure Nugget Sill >range <range Model

level direction
1º 114 16436 ε
Rising 94º 619 Gaussian
2º 7770 17924 16436
3º 1480 ∞ 17924
Table 1. The semivariogram parameters used to interpolate the in-situ turbidity.
The adjustment on Gaussian model suggests the existence of smooth spatial variance pattern
in the of study site (Burrough and Mcdonnell, 1998). The reference map was then used to
evaluate the result obtained with the SMM. The equation 5 presents the fitted model used to
interpolate the turbidity distribution during the rising water phase.
2 2 2 2
h   h   h   h 
γ (h) = 619 + 114[Gau (  94°  +  216°  )] + 7770[Gau (  94°  +  216°  )] +
 ε   16436   17924   16436 
2 2
(5)
 h  h 
1480[Gau (  94°  +  216°  )]
 17924   ∞ 
Where γ (h) is the semivariance at the lag h , hn is the semivariance due to angles of
anisotropy, ε is the range of the lower anisotropy angle, and ‘Gau’ is the fitted Gaussian
fitted model.
4.3 Spatial regression between the SMM and in situ turbidity

In general, OLS models have been used in research (Tyler et al. 2006). This approach,
however, does not consider the spatial autocorrelation of samples within an aquatic system.
When spatial autocorrelation is not considered in regression analysis, the significance of
parameters can be overestimated, and the existence of large-scale variations might lead to
spurious associations (Anselin, 1988). In our study, a method called Kernel Estimation (KE)
was applied to in-situ turbidity, thereby testing for the existence of spatial autocorrelation
(which is indicative of different spatial regimes).
The KE is a common analysis tool to determine the local density of point events and create a
field representation. The Kernel is computed, using a Gaussian function, k , which
considers m events (i.e. turbidity samples) {u i ,...u i + m − 1} contained in a bandwidth τ
around u and the distance d between the position and the ith sampling (Figure 5), through
the functions represented by equation 6:
^ 1 n
 d (u i , u ) 
λτ (u ) = Σ k , d (u i , u ) ≤ τ (6)
τ  τ
2 i =1

Fig. 5. Kernel estimation of the spatial turbidity pattern.
To improve the kernel estimation we used an adaptive bandwidth version which locally
adjusted bandwidth values at different points within the floodplain. Thereby, it ensured that
the bandwidth always contained a minimum number of samples, improving estimate
precision (Bailey and Gatrell, 1995).
The technique was also applied to the OLS, a model compiled using global in-situ turbidity
data to evaluate the spatial regression models. The OLS does not take into account spatial
dependence among samples and consequently, different spatial regimes. However, spatial
error and lag models were developed for each identified spatial regime, using the KE
algorithm, thus including spatial dependence which tends to inflate the variance of OLS
regression (Bailey and Gatrell, 1995).
To assess the relationship between the fraction images and in-situ turbidity two approaches
were adopted The first, assumes that the variance of the disturbance term is constant; we
start with the OLS model:
y = Xβ + ε , with ε ~ N (0, σ 2 ) (7)
Where y is an (Nx1) vector of observations on a dependent variable taken at each of N

locations, X is an (NxK) matrix of exogenous variables, β is an (Kx1) vector of parameters,
and ε is an (Nx1) vector of disturbances.
The second uses two alternative forms of spatial dependence models (Bivand, 1998), the
spatial lag model (presented in equation 8) and the spatial error (presented in equation 9).
y = ρWy + Xβ + ε (8)
y = Xβ + u , with u = λWu + ε (9)
Where λ is a scalar spatial error parameter, and u is a spatially autocorrelated disturbance

vector. These two models can also be related to each other in the Common Factor model.
According to (Bivand, 1998), however the spatial lag and spatial error models can only be
combined for estimation if the neighborhood specifications (here: the W matrices of the lag
and error components) differ. The spatial lag model is clearly related to a distributed lag
interpretation, in that the lagged dependent varible, Wy, can be seen as equivalent to the
sum of a power series of lagged independent variables stepping out across the map, with
the impact of spillovers declining with successively higher powers of ρ (Bivand, 1998).
Conversely, the spatial error model is a special case of a regression with a non-spherical
error term, in which the off-diagonal elements of the covariance matrix express the structure
of spatial dependence (Baltaqi, 2003).
Spatial lag and spatial error dependence tests allowed determining which of the two models
was more suitable to the data. Subsequently, spatial regression models were created for each
spatial regime. Measurements of Log likelihood fit, Akaike information criteria (AIC) and
Schwarz criteria (SC) for both models (OLS and spatial models) to verify if the inclusion of
spatial parameters improve the regression model.
The Schwarz criterion addresses the problem of selecting one of a number of models of
different dimensions (Bailey and Gatrell, 1995). However, some authors (Burrough and
Mcdonnell, 1998; Bailey and Gatrell, 1995) suggest the use of AIC to evaluate the best fit. The
Akaike information criterion is expressed in equation 10 (Pan, 2001).
AIC = −2 * LIK + 2k (10)
Where, LIK is the log of the maximized likelihood and k is the number of regression
coefficients. A small AIC value suggests a high suitability of the tested model.
4.4 Evaluation of turbidity estimations

Having selected the best model, the next step was to apply it to the Terra/MODIS image to
estimate the turbidity distribution. The resulting map was then used as a reference to
evaluate the model. Random positions were selected on the image to run correlation
analyses so as to assess the potential spatial model and the RSME. At each random
geographical position, a 3x3 pixel window was averaged, obtaining both the measured
and modeled value and computing the correlation.
5. Results and Discussion

Figure 6-a shows that the water, of the entire Curuai floodplain lake was rich in inorganic
matter (Iss), with a particularly high proportion in the Poção lake (Figure 6-b). The images
illustrating the distribution of dissolved organic matter (Dom) revealed that this was
particularly apparent in the Salé lake, and in the littoral region (i.e. the region between water
and forest). This is mainly due to the fact that some organic matter in decomposition is
transported into the floodplain by the water during the rising , phase (Figure 6-c). The
phytoplankton (Chl) fraction represented a high proportion of the ‘Grande de Curuai’ and
‘Grande do Poção’ lakes (Figure 6-d). As described in (Barbosa, 2005), the water coming
from the Amazon River was more enriched with phytoplankton than that in the Curuai
floodplain; the proportion of the water rich in Chl was found to increase from the East to the
West.
Fig. 6. Terra/MODIS imagery of the Curuai floodplain (a), and proportion images for the
tested endmembers: Iss (b), Dom (c) and Chl (d) compiled with the SMM algorithm.
The fraction images (Iss in the red channel, Dom in the blue channel, and Chl in the green
channel) unfolded a qualitative picture of the turbidity distribution within the Curuai
floodplain lake (Figure 7-a). Figure 7-b shows the SMM error image of the Region 1 (Figure
7-a) which represents the most turbid water within the floodplain. This regions has an area
of extremely high turbidity (see the circle in Figure 7-b) which actually is due to cloud cover.
Region 2 (Figure 7-a) evidence of the mixture of water masses with dominant proportions of
Chl and Iss, and low turbidity (Figure 7-b). Region 3 (Figure 7-a), finally, was dominated by
water with a high proportion of Chl, and moderate turbidity. A mixture of water masses
containing high concentrations of Chl and Iss lead to a low error of the SMM. Conversely, a
high amount of Chl entering the floodplain through the main channel from the Amazon
River leads to a moderate error in the unmixing model. Due to the occurrence of a transition
zone between the aquatic and the terrestrial environments, higher errors occurred at the
edges.
However, these results are qualitative. To obtain suitable quantitative results, an OLS
regression model was applied to the SMM results and the in situ turbidity data.
Subsequently, we checked for any signs of spatial autocorrelation between samples to
prevent the problem of spurious associations, using the result obtained by the kernel
estimator algorithm.
The results of the OLS regression model, using all 215 turbidity samples collected, and the
fractional abundance of the OAS are poorly correlated (R²= 0.10, p <0.05). This is
presumable due to the fact that the floodplain received water from different sources (rain,
black water and white water) when the water level rose. The mixture of water masses
caused a high heterogeneity in the spectral response of the surface water, causing high
standard deviation in turbidity measurements, what highlights the importance of including
the autocorrelation factor in the regression analysis.
Fig. 7. (a) The composition of the image fractions derived from the SMM (Iss in red channel,
Chl in green and Dom in blue), in comparision with the the RMS image (b).
If spatial dependence was verified the OLS method would lose validity (Anselin, 1988). In
accordance with Bailey and Gatrell (1995) the use of the non-spatial regression models
(OLS), may or may not be suitable because it assumes a stationary water conditions in the
period of ground sampling. Also, the OLS model does not consider the presence of spatial
autocorrelation among samples distributed within the floodplain. Hence, in order not to
overestimate the significance of parameters, autocorrelation must be considered. Likewise,
large-scale variations may induce spurious associations.
To check the spatial autocorrelation between turbidity samples, we applied the Kernel
estimator, and then separated the samples of fractional abundance of the OAS and in-situ
turbidity data in spatial regimes. These cluster the turbidity data from the whole floodplain
by their spatial dependence (Bailey and Gatrell, 1995).
The KE revealed that there were four spatial regimes of the Curuai floodplain lake turbidity
at the rising water level (Figure 8-a). The kernel also shows four regions of density from
high (region 1) to low density (region 4). The region 1 was characterized by the highest
spatial dependence in the study area. In this area, the spatial regime is particularly abundant
1(Figure 8-b) and includes the largest number of samples.
Region 2 included a mixture of spatial regimes 1 and 2, and represented the second largest
spatial dependence. Region 3 was characterized by regimes 3 and 4, and region 4 included a
mixture of all regimes (Figure 8-b). The number of samples grouped in each spatial regime
was 64, 54, 51 and 45 for regimes 1, 2, 3, and 4 respectively.
Fig. 8. (a) Spatial dependence of in-situ turbidity samples, as defined by Kernel algorithm,
(b) spatial regimes occurring within the Curuai floodplain lakes with corresponding
turbidity intervals values (NTU).
The different regimes are brought about by different types of water input when the Amazon
River enters into the floodplain. The water arrives in different ways, and at different times.
As a result, the water entering the different sections of the floodplain has variable physical
and chemical properties.
Table 2 indicates that the application of the spatial regression model increased the R² in
relation to the OLS model. This is presumably due to the turbidity data being characterized
by spatial auto-correlation, as observed in figure 8-a. The OLS model did not take into
account the spatial dependence among samples, thereby causing an overestimated
significance of the parameters.
OLS Spatial regression

Spatial Lag Model Spatial Error Model
Global Re1 Re2 Re3 Re4 Re1 Re2 Re3 Re4
R2 0.10 0.31 0.27 0.61 0.71 0.32 0.30 0.95 0.79
LIK -295.79 -95.07 -101.88 -35.97 -22.78 -94.93 -101.69 -35.50 -20.70
AIC 599.58 200.15 213.76 81.94 55.57 197.86 211.38 79 49.41
SC 607.54 205.61 219.44 82.93 54.53 202.22 215.92 79.79 48.57
Table 2. Results of the OLS and spatial regression models, with R2 being the coefficient of
determination, LIK being the log of the maximized likelihood, AIC being the Akaike
information criterion, and SC being the Schwarz criterion.
The approach to estimate the turbidity data in spatial regimes shows that in all regimes the
best fit was obtained using the spatial error model. According to Anselin (1988) the fit on the
spatial error model suggests that the spatial effect is a noise derived from the interplay of
several variables not included in the model.
The spatial regime 1 (Re1) has the poorest correlation compared to the others (R² = 0.10, p <
0.001), with LIK, AIC and SC being -295.79, 599.58, and 607.54 respectively. Conversely, the
spatial regime 3 (Re3) shows the best results with LIK = -35.50, AIC = 79 and SC = 79.79, and
R2 = 0.95, p < 0.001, using the spatial error model (Table 2). The equations of the four spatial
regimes (Re1, Re2, Re3, and Re4) are listed below in the equations 11, 12, 13, and 14
respectively:
Turbidity = 99.72 + 221.71Iss + 199.46Chl + 186.53Dom (11)

Turbidity = 311.61 − 254.14 Iss − 23.87Chl + 22.98 Dom (12)
Turbidity = 327.19 + 42.62 Iss − 246.85Chl − 136.06 Dom (13)
Turbidity = 428.86 − 154.66 Iss − 231.23Chl − 94.05 Dom (14)
Where, Iss is the inorganic-laden water fraction, Chl is the phytoplankton laden water
fraction and Dom is the dissolved organic matter-laden water fraction.
Equation 13 indicates the influence of the OAS proportion on this method to quantify
turbidity. According to equation 13, the turbidity in this region will be lower when the
water has high proportion of phytoplankton than inorganic particles. Conversely, in areas
where those types of water are particularly abundant, the modeled turbidity will be smaller.
Where the Iss fraction dominates, turbidity will be high. This suggests that the model will
perform better when turbidity is determined mainly by suspended inorganic particles.
Before applying the equation 13 to the MODIS image fractions, a sensitivity analysis was
carried out using Pearson’s correlation analysis for each spatial regime (Figure 9).
This sensitivity analysis revealed statistically significant correlations R² = 0.65 (p < 0.05, n =
20) and R² = 0.40 (p < 0.05, n = 20) for the spatial regimes 3 [Figure 9 (c)] and 4 [Figure 9 (d)]
respectively. In contrast, analyses for the spatial regimes 1 [Figure 9 (a)] and 2 [Figure 9 (b)]
were not significant with R² = 0.005 (p < 0.05, n = 20) and R² = 0.004 (p < 0.05, n = 20)
respectively (Figure 9). The best fit was hence found in the spatial regime 3.
(a) (b)
(c) (d)
Fig. 9. Correlation between turbidity estimated using the spatial model and observed
turbidity: (a) spatial regime 1; (b) spatial regime 2; (c) spatial regime 3; (d) spatial regime 4.
The main reason for this poor statistical performance is probably related to the
instantaneous image acquisition and the sampling turbidity design. Both cause a difference
in turbidity conditions, which makes a comparison of the image and the in-situ data
difficult. In addition, the nonlinearity presented in the relationship between the loads of
phytoplankton, dissolved organic matter, and inorganic matter in the water made the
unmixing processes difficult.
Figure 10 shows the turbidity distribution resulting from the application of this model to the
entire floodplain (Figure 10-a) as well the in-situ turbidity distribution (Figure 10-b). The
spatial distribution of the turbidity is similar, especially in regions 1, 2, and 3.
Region 1 is characterized by the highest turbidity compared to the other regions, and by
small depths when the water level rises (Barbosa, 2005), and is the main pathway from the
Amazon River into the floodplain. The water then flows through many channels, located in
the Northwest region of the floodplain, with sufficient energy to keep the inorganic particles
in suspension. Winds promote the sediment re-suspension and increment turbidity here.
Region 2 is also characterized by small depths but is protected from wind. As a result,
turbidity is related to depth rather than wind perturbation (Carper and Bachmann, 1984;
Booth et al. 2000). In Region 3, turbidity varied between 203 and 305 NTU, and the spatial
model was able to estimate turbidity values within the range measured in-situ. According to
Barbosa (2005), this is one of the regions with the highest depths in the Curuai floodplain
lake. It is not subjected to intense wind perturbation due to this characteristic. Instead,
suspended particles settle into the bottom of the lake, decreasing the possibility of re-
suspension.
Figure 10-c indicates that the spatial model overestimated turbidity in regions 4 and 6, and
underestimated it in region 5. The overestimation in 4 can be explained by cloud cover
contamination. Region 5 is in a transition zone (natural barrier) that separates the floodplain
lake in two larger zones, (i) zone Northwest and (ii) zone Southeast (Barbosa, 2005). This
transition zone causes a turbidity underestimation due to mixture of different water masses.
The overestimation in region 6 can be related to a large mixture caused by a water inlet
owing to a rising water level. The water inlet is particularly pronounced in region 6
(Barbosa, 2005).
Fig. 10. The turbidity distribution (NTU) in the Curuai floodplain, (a) using the spatial
regression model, using the OK algorithm (b) and (c) is in percent difference between (a)
and (b).
The samples to evaluate the spatial model were collected the entire lake, except for regions 4
(due to cloud cover), 5 (natural barrier), and 6 (overestimation due to intense water flowing
through this channel), as pointed out in Figure 7. The evaluation of the model resulted in a
value of R² = 0.90 (p = 0.05; n = 60) and RSM of 17 NTU for the turbidity model under the
spatial regime 3 (Figure 11). Previous studies of turbidity using the SMM in high water level
(Alcântara et al., 2008) show that the modeled turbidity had a correlation R² = 0.62 (p >
0.005; n = 20).
This difference in performance can be explained by the presence of different water types in
the floodplain. Their presence makes it difficult to apply one model, adapted to a given
regime for the entire region. Another possible error source is the in-situ turbidity-sampling
scheme. Due to the size of the lake, it was not possible to get all the samples on a single day.
Instead, the data was collected during a 13-day period, in which local turbidity may have
been affected by changes in wind intensity, light field, and other environmental factors
affecting the lake hydrodynamics. The highest errors were encountered in more turbid
waters, as opposed to clear waters, presumably due to a high mixture of the OAS present in
the floodplain.
Fig. 11. Evaluation of the turbidity estimated through the spatial model (equation 13) and in-
situ turbidity (NTU), interpolated using the OK algorithm.
Since the MODIS image was instantaneously acquired on 27 February 2007, the modeled
turbidity does not account for environmental changes, which can affect in-situ conditions. In
spite of this drawback, the unmixing model showed a good potential to assess the turbidity
in continental aquatic systems. This potential could be improved using hyperspectral remote
sensing imagery (Rudorff et al. 2007), since features in spectral responses could then be
detected in more detail (Rudorff et al., 2006; Fraser, 1998; Brando and Dekker, 2003).
6. Conclusions
This present work evaluates the suitability of the spectral unmixing model to map the
turbidity distribution in the Curuai floodplain. The main conclusions are:
1. The fraction images for the endmembers selected directly from the MODIS image based
on dominance of water components allowed assessing the turbidity in the Curuai floodplain
lake.
2. Owing to non-linearity in the Amazon floodplain waters, the unmixing model does not
work in an optimal way. This is also due to autocorrelation presented in the study area.
3. It was clear that the presence of autocorrelation in limnological studies that use spatial
distributed samples and this paper shows the possible way to solve the problem of
autocorrelation between samples spatially distributed.
4. The spatial regression between the results obtained from MODIS fraction imaging with
the map generated from in situ data using the Ordinary Kriging approach seems to be
useful to estimate water turbidity.
5. The modeling of autocorrelation helps to improve the applicability of the SMM to map the
turbidity distribution in high complexity water bodies, such as the Amazon floodplain.
7. Acknowledgements
This work was supported by the Brazilian funding agency FAPESP under Grant 02/09911-1.
The work of E. H. de Alcântara was supported by the Brazilian Council for Scientific and
Technological Development (CNPq). The authors would like to thank Carsten Reinhard
from the School of Biological Sciences, University of Bristol, for the English review.
8. References
Alcântara, E.H.; Barbosa, C.; Stech, J.; Novo, E.; Shimabukuro, Y. (2009a). Improving the
spectral unmixing algorithm to map water turbidity distributions. Environmental
Modelling & Software 24, 1051-1061.
Alcântara, E.H.; Novo, E.; Stech, J.; Lorenzzetti, J.; Barbosa, C.; Assireu, A.; Souza, A.
(2009b). The turbidity behavior in an Amazon floodplain. Hydrol. Earth Syst. Sci.
Discuss., 6, 3947–3992.
Alcântara, E.H.; Stech, J.L.; Novo, E.M.L.M.; Shimabukuro, Y.E.; Barbosa, C.C.F. (2008).
Turbidity in the Amazon floodplain assessed through a spatial regression model
applied to fraction images derived from MODIS/Terra. IEEE T. Geosci. Remote 46
(10), 2895-2905.
Anselin, L. (1988). Spatial Econometrics: Methods and Models; Kluwer Academic:
Dordrecht, Netherlands, pp. 284.
Barbosa, C.C.F. (2005). Sensoriamento remoto da dinâmica de circulação da água do sistema
planície de Curuai/ Rio Amazonas. (PhD. Thesis); INPE: São José dos Campos,
Brazil, pp. 255.
Boardman, J. W., Kruse, F. A., Green, R. O. (1995). Mapping target signatures via partial
unmixing of AVIRIS data: in Summaries, Fifth JPL Airborne Earth Science
Workshop, JPL Publication 95-1 (1), 23-26.
Baltaqi, B.H. (2003). A companion to theoretical econometrics. Blackwell: Texas, USA, pp.
709.
Bailey, T.C.; Gatrell, A.C. (1995). Interactive spatial data analysis; Prentice Hall: Upper
Saddle River, USA, pp. 432.
Bivand, R. (1998). A review of spatial statistics techniques for location studies. In
Proceedings of the CEPR symposium on New Issues in Trade and Location, Lund,
Sweden, 1-21.
Booth J.G.; Miller R.L.; Mckee B.A.; Leathers R.A. (2000). Wind-induced bottom sediment
resuspension in a microtidal coastal environment. Cont. Shelf Res. 20 (7) 785-806.
Brando, V.E.; Dekker, A.G. (2003). Satellite hyperspectral remote sensing for estimating
estuarine and coastal water quality. IEEE T. Geosci. Remote. 41 (6), 1378-1387.
Burrough, P.A.; Mcdonnell, R.A. (1998). Principles of geographical information systems;
Oxford University Press: New York, USA, pp. 333.
Carper, G.L.; Bachmann, R.W. (1984). Wind resuspension of sediments in a Prairie Lake.
Can. J. Fish. Aquat. Sci 41, 1763-1767.
Dekker, A.G.; Malthus, T.J.; Hoogenboom, H.J. (1995). The remote sensing of inland water
quality. In Advances in Environmental Remote Sensing; Danson, F.M., Plummer,
S.E., Eds., Wiley, pp. 123-142.
Fraser, R.N. (1998). Hyperspectral remote sensing of turbidity and chlorophyll a among
Nebraska Sand Hills lakes. Int. J. Remote Sens. 19 (8), 1579-1589.
Goodin, D.G.; Han, L.; Fraser, R.N.; Rundquist, D.C.; Stebbins, W.A.; Schalles, J.F. (1993).
Analysis of suspended solids in water using remotely sensed high resolution
derivative spectra. Photogramm. Eng. Rem. S. 59 (4), 505-510.
Hadjimitsis, D.G.; Hadjimitsis, M.G.; Clayton, C.; Clarke, B. (2006). Determination of
Turbidity in Kourris Dam in Cyprus Utilizing Landsat TM Remotely Sensed Data.
Wat. Res. Manag. 20 (3), 449-465.
Hess, L.L.; Melack, J.M.; Novo, E.M.L.M.; Barbosa, C.C.F.; Gastill, M. (2003). Dual-season
mapping of wetland inundation and vegetation for the central Amazon basin.
Remote Sens. Environ. 87 (4), 404-428.
Isaaks, E.H.; Srivastava, M.R. (1989). An introduction to applied geostatistics; Oxford
University Press: New York, USA, pp. 561.
Junk, W.J. (1997). The Central Amazon Floodplain: ecology of a pulsing system; Springer
Verlag: Berlin, Germany, pp. 525.
Kloiber, S.M.; Brezonik, P.L.; Olmanson, L.G.; Bauer, M.E. (2002). A procedure for regional
lake water clarity assessment using Landsat multispectral data. Remote Sens.
Environ. 82 (1), 38-47.
Kruse, F.A.; Lefkoff, A.B.; Boardman, J.W.; Heidebrecht, K.B.; Shapiro, A.T.; Barloon, P.J.;
Goetz, A.F.H. (1993). The spectral image processing system (SIPS)—interactive
visualization and analysis of imaging spectrometer data. Remote Sens. Environ. 44
(2-3), 145-163.
Mertes, L.A.K.; Smith, M.O.; Adams, J.B. (1993). Estimating suspended sediment
concentrations in surface waters of the Amazon River wetlands from Landsat
images. Remote Sens. Environ. 43 (3), 281-301.
Moreira-Turcq, P.; Jouanneau, J.M.; Turcq, B.; Seyler, P.; Weber, O.; Guyot, J.L. (2004).
Carbon sedimentation at Lago Grande de Curuai, a floodplain lake in the low
Amazon region: insights into sedimentation rates. Palaeo. Palaeoc. Palaeoe. 214 (1-
2), 27-40.
MRT. (2002). MODIS Reprojection Tool User’s Manual; Department of Mathematics and
Computer Science South Dakota School of Mines and Technology. USGS EROS
Data Center, pp. 1-57.
Novo, E.M.L.M.; Shimabukuro, Y.E. (1994). Spectral mixture of inland tropical waters. Int. J.
Remote Sens. 15 (6), 1351-1356.
Novo, E.M.L.M.; Barbosa, C.C.; Freitas, R.M.; Shimabukuro, Y.E.; Melack, J.M.; Pereira Filho,
W. (2006). Seasonal changes in chlorophyll distributions in Amazon floodplain
lakes derived from MODIS images. Limnology 7 (3), 153-161.
Oyama, Y.; Matsushita, B.; Fukushima, T.; Matsushige, K.; Imai, A. (2009). Application of
spectral decomposition algorithm for mapping water quality in a turbid lake (Lake
Kasumigaura, Japan) from Landsat TM data. ISPRS Journal of Photogrammetry
and Remote Sensing 64 (1), 73-85.
Pan, W. (2001). Akaike’s information criterion in generalized estimating equations.
Biometrics 57 (1), 120-125.
Rudorff, C.M; Novo E.M.L.M.; Galvão L.S; Pereira-Filho W. (2007). Derivative analysis of
hyperespectral data measured at field and orbital level to characterize the
composition of optically complex waters in the Amazon (in Portuguese). Acta
Amaz. 37 (2), 269-280.
Rudorff, C.M.; Novo, E.M.L.M.; Galvão, L.S. (2006). Spectral mixture analysis for water
quality assessment over the Amazon floodplain using Hyperion/EO-1 Images.
Revista Ambi-Água 1 (2), 65-79.
Theseira, M.A.; Thomas, G.; Taylor, J.C.; Gemmell, F.; Varjo, J. (2003). Sensitivity of mixture
modelling to end-member selection. Int. J. Remote Sens. 24 (7), 1559-1575.
Tyler, A.N.; Svab, E.; Preston, T.; Présing, M.; Kovács, W.A. (2006). Remote sensing of the
water quality of shallow lakes: A mixture modelling approach to quantifying
phytoplankton in water characterized by high-suspended sediment. Int. J. Remote
Sens. 27 (7-8), 1521- 1537.
Wetzel, R.G. (2001). Limnology: lake and river ecosystems, 3rd Ed.; Academic Press: San
Diego, USA, pp. 1006.
Application of remote sensing to the estimation of sea ice thickness distribution 21
X2
Application of remote sensing to the estimation

of sea ice thickness distribution
Takenobu Toyota
Hokkaido University
Japan
1. Introduction
Sea ice covers about one tenth of the world ocean surface and plays an important role in the
global climate system through its interaction with the atmosphere and ocean. Although its
thickness is only several meters at most, it forms an effective thermal insulation sheet due to
its high albedo and low thermal conductivity, leading to a significant reduction in the heat
flux from the ocean to the atmosphere, especially in winter. Moreover, when sea ice forms, it
expels brine into water to produce high density water underneath the ice, leading to the
deep vertical convection. This plays a major role in the three dimensional ocean
thermohaline circulation of the ocean on a global scale. Another notable feature is its large
variability in areal extent on seasonal to annual time scales e.g. in the Southern Hemisphere
(Figure 1). Even on a daily time scale, it varies significantly due to the combined effects of
wind and ocean swell. In turn this variability also has an impact on both climate and daily
weather conditions through the above processes. In addition, the mobility of sea ice leads to
the interaction between ice floes and ridging/rafting of the ice (dynamical pile-up), which
makes its properties more complex in comparison with other surface types in the cryosphere,
such as snow cover or frozen soil.
In these processes, thickness distribution is one of the most important parameters of sea ice.
This is firstly because the outgoing heat flux from ocean to atmosphere is highly dependent
on the ice thickness distribution, especially in winter when the areal fraction of ice with a
thickness of less than 0.4 m essentially controls the net heat input to the atmosphere
(Maykut, 1978). This implies that the areal fraction of thin ice controls the rates of
thermodynamic ice growth amount and the associated production of the high density water
beneath. Therefore, the ice thickness distribution can make a significant difference to both
the surface heat budget and deep ocean convection for a given concentration of sea ice. In
this context, it is worth mentioning that snow cover on ice also works as a more effective
insulator due to its low thermal conductivity. Secondly, and in terms of sea ice dynamics,
the thickness distribution determines ice strength, and thus controls ridging/rafting
processes and vice versa (Thorndike et al., 1975). These “pile-up” processes are essential to
the ice thickening, particularly in the seasonal sea ice zone (SSIZ) where the ice is relatively
thin and mobile. From the crystallographic analyses it has been shown that a layered
structure due to ice pile-up is prominent in the SSIZ (e.g. Gow et al., 1982; Lange and Eicken,
(a) (b)
(c) (d)
Frequency of sea ice existence
Fig. 1. Mean geographical distribution of sea ice extent (1979-2000) for the northern
hemisphere at (a) maximum extent (Feb. 28) and (b) minimum extent (Sep. 15) the southern
hemisphere at (c) maximum extent (Sep. 15) and (d) minimum extent (Feb. 28).
(Japan Meteorological Agency, 2005)
1991) and that the mean layer thickness is only about 12 cm both for Antarctic ice (Worby et
al., 1996) and that in Arctic peripheral seas e.g. the Sea of Okhotsk (Toyota et al., 2007).
Therefore, the observed thickness distribution can provide some important insights into the
role of such dynamical thickening processes. Thirdly, the long-term trend and the short-
term change in the sea ice thickness distribution is a sensitive indicator of climate change, as
observed for Arctic perennial ice (e.g. Nghiem et al., 2007). Therefore, the accurate and
consistent estimation of sea ice thickness distribution is one of the most important issues
that many researchers in this field have paid attention to for a long time.
Unfortunately, sea ice thickness is also one of the most difficult parameters to measure. It
was only two decades ago that the mean hemispheric distribution of ice thickness was first
mapped for the Arctic Ocean (Bourke and Garrette, 1987; Figure 2a), and only recently for
the Antarctic Seas (Worby et al., 2008; Figure 2b). The most direct and accurate
measurement technique is in-situ drilling with an ice auger. In the Arctic, it is said that
Nansen first conducted systematic ice thickness measurement with this method during the
drift of “Fram” (1893-96) (Wadhams, 2000). In the Antarctic, it was by drilling that the
standard measurement protocol was initiated in the late 1970’s (Ackley et al., 1979). In
addition to being labour-intensive, this technique suffers a problem of representativity.
Rothrock (1986) pointed out from a statistical perspective that as many as 560 drilling sites
would be required if we need to know the mean thickness at a point with an accuracy of 0.1
m for the Arctic ice. Thus, nowadays this method is used mainly for calibration of other
tools. To obtain thickness distribution at a large scale, remote sensing is a requisite tool.
Since 1958, when submarine sonar was applied for this purpose in the Arctic Ocean,
enormous efforts have been made to estimate it using remote sensing. Unfortunately,
submarine sonar coverage of Antarctic ice is not possible.
In this chapter several methods will be introduced to give a brief history of ice thickness
measurement, focusing on how remote sensing techniques have been applied and how they
have contributed to the improved understanding of sea ice thickness. Finally, the current
situation will be summarized and some suggestions will be made about a future work that
needs to be done.
2. Measurement techniques
In this section the remote sensing tools that have been used to measure ice thickness to date
will be introduced. There are generally two strategies: 1) direct measurement of ice draft or a
full depth of ice using an upward-looking sonar or other tools; and 2) relating the observed
surface properties with ice thickness using satellite sensors. Here, they will be described in
near-chronological order.
2.1 Submarine sonar profiling

This method provides the under-ice profile relative to sea level using an upward beamed
sonar on a submarine. By this technique, profiles of ice draft can be obtained along the track
of the submarine by recording the distance to the underside of the ice pack and subtracting
this from the depth of the sonar, using intervening regions of open water as a reference level.
The overall accuracy is about 0.3 m (0.09 m for smooth ice), and the horizontal resolution is
1.3 to 1.5 m with a surface beam diameter of 3.2 m (Wadhams and Horne, 1980). These data
are not a complete measurement of ice thickness, as ice draft corresponds to 80 to 95 % of ice
thickness, but nonetheless this data set provides the approximate thickness distribution and
detailed topography of under-ice with reasonable accuracy, assuming invariability in snow
cover thickness and density.
This device was first used during the voyage of the “Nautilus” in early August 1958 which
crossed the Arctic Basin, and much of our knowledge about the synoptic ice thickness
(a) (b)
(unit: m)
Fig. 2. Climatology of ice thickness distribution for (a) the northern hemisphere (reproduced
based on Bourke & Garrett (1987)) (b) the southern hemisphere (cited from Worby et al.
(2008)). The average periods are 1960-82 for the Arctic and 1981-2005 for the Antarctic. The
data are based on submarine sonar observations for the Arctic and on ship-based repeated
visual observations for the Antarctic.
Fig. 3. Probability density of ice thickness in the Greenland Sea. Note that the clear negative
exponential tail at the thick end can be seen for all the cases. (cited from Wadhams (1992)).
distribution in the Arctic Ocean is based on the results obtained with this device on
subsequent cruises (McLaren, 1989). With the results obtained, the regional statistics of ice
thickness distribution and ridge keel profiles were investigated in detail e.g. for the Beaufort
Sea (Wadhams and Horne, 1980), in the Fram Strait region (Wadhams, 1983, 1992), and in
Davis Strait (Wadhams et al., 1985). One of the most significant findings is that in most cases
the thickness distribution follows a negative exponential function, irrespective of regional
location (Figure 3). Although the mechanism to explain this feature has not been clarified
yet, this feature may provide considerable insight into the ridging process and is therefore a
problem worth investigating given that ridging is a key determinant of the thickness
distribution. McLaren (1989) examined the general features for the ice over the Arctic Basin
from the data sets obtained along the two identical cruises tracks in 1958 and 1970. Bourke
and Garrett (1987) then compiled the data obtained during the 17 voyages from 1960 to 1982
and mapped the ice thickness distribution for the first time (Figure 2a). According to their
results, the mean ice thickness above 65N is 2.9 m with a standard deviation of 1.8 m. This
value was taken to be the standard “baseline” value for the Arctic Ocean before the recent
thinning of the Arctic ice that has been reported (e.g. Rothrock et al., 1999).
In a further development of this method, Wadhams (1988) showed that additional
information can be gained of the three-dimensional under-ice topography of Arctic ice using
a side-scan sonar. Indeed, it was this technique that provided the first evidence for
significant thinning of the Arctic ice cover over the past few decades (Wadhams, 1990;
Rothrock et al., 1999; Tucker et al., 2001). By comparing the data obtained in 1958-1976 with
those in the 1990s, Rothrock et al. (1999) showed that the mean ice draft at the end of the
melt season decreased from 3.1 m to 1.8 m in the Arctic Ocean.
Despite all these pioneering results, the drawback to these data is that since this
measurement is operated on a military basis, the observation is not necessarily systematic in
space and time and time series of thickness data cannot be obtained within given areas
(Wadhams, 2000). Therefore, it does not provide sufficient data sets to enable monitoring of
change in ice thickness distribution on a seasonal or annual basis. Moreover, it cannot be
applied to Antarctic ice because military submarines are prohibited from operating under
Southern Ocean sea ice by the Antarctic Treaty.
2.2 Moored ice profiling sonar

This technique measures ice draft using a moored upward-looking sonar system which is
tethered to a line attached to the seafloor. The principle is similar to that of submarine sonar.
The device transmits a pulse of sound towards the surface, and the ice draft is then obtained
by converting the travel time of the pulse to vertical distance allowing for variation in water
temperature and pressure and subtracting it from the depth of the device (Figure 4). The
footprint at the underside of the sea ice is about 2 m and the overall accuracy is 0.05 to 0.10
m (Melling and Riedel, 1995). The most remarkable merit of this method is that it can
provide a time series of ice draft distribution at a fixed site with high temporal resolution (a
few seconds to minutes) for several months. Thus it enables the monitoring of daily to
seasonal variation in the ice draft distribution. Additionally, it can be deployed on the
shallow continental shelf regions, where a submarine cannot operate. This is important
because considerable deformation often occurs near the Arctic coastal region, to affect sub-
sea ice exploration.
This technique was first applied to the near-shore region of the Beaufort Sea in the late
1970’s (Hudson, 1990). From the over-winter measurements, the seasonal change of level ice
thickness and the frequency of ridge keels were first revealed. In the same region, Melling
and Riedel (1995, 1996) also deployed this system and examined the more detailed
characteristics of deformed ice and under-ice topography by combining the ice motion
obtained with a Doppler sonar. In Fram Strait, Vinje et al. (1998) examined the annual and
seasonal variation of ice draft during the period of 1990 to 1996. By empirically converting
ice draft to thickness and combining the ice drift data obtained from buoys, they estimated
the annual mean ice volume flux through the Fram Strait to be 2850 km3 yr-1. Combining this
ice thickness data set with satellite-derived ice velocity Kwok and Rothrock (1999) found a
relationship between the ice volume flux out of the central Arctic and the large-scale modes
of atmospheric circulation (i.e. the North Atlantic Oscillation). In the Sea of Okhotsk, Marko
(2003) applied this method to the east of Sakhalin Island and analyzed the ice draft data in
conjunction with intense wave energy. Fukamachi et al. (2003, 2006) deployed this system in
the coastal region of Hokkaido, Japan, and first revealed the ice thickness distribution
including ridged ice in this region from three-year data sets.
In the Antarctic, where submarine sonar data are not available, this method has been an
effective tool. However, it should be noted that there are some restrictions regarding
deployment: the mooring system has to be operated from the seafloor at a depth deeper
than 350-440 m to avoid ice bergs in the sea ice area (Massom et al., 2009), which may
damage the instrument (Worby et al., 2001). Strass and Fahrbach (1998) examined the
temporal and regional variation of ice draft across the Weddell Sea from six ice profiling
sonars deployed between the tip of the Antarctic Peninsula and Kapp Norvegia from 1990 to
1994. Based on these data, Harms et al. (2001) estimated the annual mean ice volume export
from the western Weddell Sea to be about 0.05 Sv (~ 1580 km3 yr-1). In East Antarctica,
Worby et al. (2001) analyzed the seasonal variation of ice thickness distribution with this
system.
Thus this method has significantly enhanced our knowledge of detailed ice thickness
distribution, as a second tool to submarine sonar, for about three decades. It has allowed the
estimation of various parameters related to ice thickness and in particular contributed to the
understanding the fresh water budget through Fram Strait, which relates to, and has strong
implications for climate change. Additionally, this method has enabled detection of ice ridge
keels as deep as a few tens of meters (e.g. 29 m in the Beaufort Sea [Melling and Riedel,
1995] and 17 m in the Sea of Okhotsk [Fukamachi et al., 2006]). However, the drawback is
that this device is not appropriate for obtaining the spatial distribution of ice thickness on a
large scale at a given time, and has difficulties in monitoring ice thickness for more than
decadal terms because of logistical constraints i.e. the need to recover and redeploy
instruments every few years.
2.3 Video monitoring system

This technique is basically designed as a ship-based measurement. With a downward-
looking video camera mounted on the side deck of a ship, ice conditions around the hull are
continuously recorded along-track. Ice thickness is then estimated on the video images for
ice floes which break up at the bow and turn into a side-up position alongside of the hull
(Figure 5). The scale on the video image is determined by lowering a measured stick onto
the ice surface while the ship is stationary. The reading error is less than a few centimetres. In
Fig. 4. Schematic representation of an ULS mooring set-up. (cited from Strass (1998))
(a)
(b)
Fig. 5. Schematic picture of a video monitoring method.
(a) A schematic figure (cited from Shimoda et al. (1997))

(b) A typical picture of ice which was broken and turned into a side-up position. Note the
brown color near the bottom, associated with the attachment of ice algae.
this way, spatial ice thickness distribution along the ship track can be determined in a
relatively short amount of time. However, it should be noted that deformed ice, which is
hard to turn, is beyond the measuring capability of this method, and that it is subject to the
ship route which tends to be selected to traverse regions of thinner ice. These facts can cause
some bias in the results i.e. a net underestimate in ice thickness distribution estimates.
Therefore, this method is not appropriate for the measurement of thick ice distribution. This
is in contrast with the measurement by upward-looking sonar, which is relatively free from
such artificial bias.
Nevertheless, there are several advantages unique to this method. First, it is useful for
obtaining the thickness distribution of level ice, accurate knowledge of which is important to
understand the thermodynamic growth and rafting process (stacking of two ice floes
without break-up). Secondly, it allows us to measure the overlying snow depth according to
the degree of whiteness on the video images. This is particularly effective in the marginal ice
zones, where ice is so thin that it is hard to measure snow depth directly. It also allows us to
detect brown ice containing ice algae (Figure 5b). Thirdly, this system can be operated at a
relatively low cost and does not require complicated calibration. Finally, without any risk of
the damage by icebergs, this method can safely be applied in the coastal Antarctic ice area.
The key to this method is the downsizing of the video camera to facilitate its use. This
system was first introduced to monitor the sea ice conditions in the Antarctic coastal region
during the Japanese Antarctic Research Expeditions (JARE) in late 1980’s (Shimoda et al.,
1997). In this area, theses observations were continued from 1988 to 2004, and from these
results Uto et al. (2006a) found a significant annual variation in the thickness distribution of
land-fast ice in the Lützow-Holm Bay near the Syowa Station and discussed the relationship
with the break-up events of the land-fast ice there. This system was also deployed in the Sea
of Okhotsk in 1991. Monitoring of the ice thickness distribution in the southern Sea of
Okhotsk has continued since then from the icebreaker P/V “Soya”, in collaboration with the
Japan Coast Guard. Based on the data set obtained from 1991 to 2000, Toyota et al. (2004)
showed that the thickness distribution follows a near Poisson distribution with an average
of 33±10 cm, and proposed a stochastic rafting cycle model, referring to the layered structure.
They also checked the accuracy of this method by comparing their data with the ice
profiling sonar data obtained in that region (Fukamachi et al., 2003) and confirmed that the
thickness distribution of these two methods matches well within the range of a few percent
for non-ridged ice. Their group is still continuing these observations over this region, and
the annual variations in mean ice thickness and volume are shown in Figure 6. In this figure,
an annual variation in ice thickness is apparent with a period of about 7 years, while the
trend towards a decrease in ice volume is remarkable, particularly after 2000.
In spite of several drawbacks, a video monitoring system therefore still remains an effective
tool for obtaining ice thickness distributions, and in the marginal ice zones in particular. As
this method provides direct measurements in a larger sense, it is sometimes used to validate
the results of other remote sensing tools. However, the time-consuming task of analysing
the images for measurement precludes real time monitoring and this problem should be
improved on the computer base in the future.
2.4 Electromagnetic Induction Sounding

This device measures the distance from the instrument to the ice-seawater interface (i.e. ice
bottom) using low-frequency (10 to 100 kHz) electromagnetic (EM) induction. The device is
composed of transmitter and receiver coils, which are set apart by a few meters. The
transmitter generates a primary electromagnetic field which induces eddy currents mainly
just below the ice bottom, because the conductivity of sea ice (0-50 mS m-1) is negligible
compared to that of seawater (2500 mS m-1). In turn the induced currents generate a
secondary electromagnetic field. Then the distance to the ice bottom is calculated from the
strength of the secondary electromagnetic field sensed by the receiver coil. Therefore, if this
device is used on the ice surface, the data obtained correspond directly to ice thickness.
While the accuracy is ±0.1 m for undeformed ice, it degrades to up to 30 % for ridged ice
because of the presence of seawater-filled cavities between the ice blocks and the smoothing
effect caused by a larger footprint that is approximately equal to the distance to the ice
bottom (Haas, 2003). Therefore validation is needed for ridged ice. The primary advantage
of this method is that it allows us the measurement of ice thickness from the air, which is
much easier than from below the ice. If we mount a laser profilometer on the instrument
and measure the height of the instrument above the combined snow and ice surface, ice
thickness can be obtained by subtracting it from the EM results and by incorporating
knowledge of snow cover depth and density (Figure 7a). In this way, we can use an
icebreaker or an aircraft to obtain profiles of ice thickness distribution with this device over
a wide scale of 100s of kms. However, it should be noted that this method provides the
summation of ice thickness and overlying snow depth. If we want to determine ice thickness
itself, snow depth has to be measured by another tool and its density must also be factored in.
Since this technique was first tested for the ground-based observations in the Arctic in the
1970’s, numerous measurements were done to obtain ice thickness data in the various
regions of the Arctic (Worby et al., 1999). The EM instrument used most frequently is the
Geonics EM31 which contains the coils located at each end, 3.66 m apart (Figure 7b), and
uses a frequency of 9.8 kHz. The ship-based EM system was first applied to the summer sea
ice in the Bellingshausen and Amundsen Seas, Antarctica (Haas, 1998). This study found
that the properties of summer ice, in particular seawater-filled gaps close to the ice surface,
significantly reduces the accuracy of the EM measurement, although Worby et al. (1999)
confirmed that the EM system is more robust when operating under winter and spring ice
conditions in the Antarctic. The ship-based EM system was also introduced to the Sea of
Okhotsk in 2004, where it was found that the seawater-filled cavities between the ice blocks
in this region lower the accuracy of ice thickness estimates significantly, even in winter.
Therefore, based on the calibration results from drilling on the ridged ice, Uto et al. (2006b)
developed an algorithm that converts conductivity to ice thickness, taking into account the
presence of gaps within the ice (Figure 8a). As a result, the measurement error was reduced
to about 10 % for ridged ice, and the spatial ice thickness distribution, including ridged ice,
was first revealed in this area. The comparison of ice thickness distribution with coincident
ship-based video result in the same area is shown in Figure 8b, where reasonable agreement
can be seen. However, the ground or ship-based EM measurement has a limitation in spatial
coverage and may be biased according to the selection of ice floes or ship-route. For greater
representativity, airborne measurement is desirable.
The airborne EM sounding technique was first developed with a helicopter in the Beaufort
Sea in the mid 1980’s (Kovacs et al., 1987; Kovacs and Holladay, 1989). Multala et al.(1996)
subsequently applied the technique to Baltic Sea ice using a fixed-wing aircraft. By
validating the data with drill-hole measurements, each study showed the method to be
promising. Recently, the technique was developed further and improved using two
1.0
(a)
0.8
Mean Hi (m)
0.6
0.4
0.2
0.0
1995 2000 2005 2010
150
(b)
Ice volume (km )
3
100
50
0
1995 2000 2005 2010
Fig. 6. Annual variations of (a) mean video thickness with a standard deviation (error bars)
and (b) ice volume estimated by multiplying mean thickness and sea ice area in the southern
Sea of Okhotsk, where sea ice area were offered by Japan Meteorological Agency.
(a) (b)
EM
Laser profiler
Fig. 7. Schematic picture of EM sounding system.

(a) Illustration of the principle of the measurement (cited from Uto et al. (2006b))
(b) A picture of EM set-up onboard the P/V “Soya” in the Sea of Okhotsk.
frequencies (3.7 and 112 kHz) for the systematic measurement, and since 2001 the
monitoring and observation of ice thickness in the Arctic has been operated with a
helicopter-borne EM by the German Alfred Wegener Institute for Polar and Marine
Research (AWI). The details of the technique are described by Haas et al. (2009). By
combining this data set with the ground-based EM measurements acquired since 1991, Haas
et al. (2008) showed a rapid thinning of sea ice had occurred in the Arctic Transpolar Drift in
the 2000’s and suggested a regime shift from multi-year to first-year ice.
Thus, the EM sounding technique has enabled us to remotely measure sea ice thickness
from the air, which enhances the measurement considerably in comparison with the
previous methods. With a ship- or helicopter-borne system, more detailed and spatial ice
thickness distribution was revealed in various regions. Although a ship-based measurement
may have some bias, it is effective especially in regions where relatively thin ice is dominant
or helicopter operations are prohibited by law. However, it should be noted that it still
needs validation, especially for melting ice or regions with significant ridges, and that its
total areal coverage is rather limited.
(a) (b)
Fig. 8. (a) Model ice structure used for the algorithm converting conductivity to ice thickness
developed for the Okhotsk ice. (cited from Uto et al. (2006b)). (b) Comparison of the
probability density functions of ice thickness between the results obtained by converting the
EM outputs to ice thickness with the algorithm (Figure 8a) and those obtained from video
monitoring system in the Sea of Okhotsk during 2004 – 2008.
2.5 Satellite sensors

Previous sections have concentrated on various in-situ and airborne remote sensing tools.
Although they have all contributed greatly to our knowledge about ice thickness
distribution on a regional scale, ultimately it is the operational monitoring at a global scale
via space-borne sensors that is most desirable. In this section, we evaluate the current
situation regarding several satellite remote sensing possibilities.
2.5.1 Passive microwave methods

Multi-frequency passive microwave sensors have flown in space continuously since 1979.
The merits of this sensor are: 1) the signal is not subject to the effect of cloud; 2) due to its
large swath width, daily coverage is possible at a global scale; and 3) data can be obtained
both day and night. Since the microwave emissivity of open water is much lower than that
of sea ice, the areal coverage of sea ice can be derived with relatively high accuracy and thus
several algorithms have been successfully developed to derive ice concentration using
brightness temperature data acquired at the frequencies of 19 and 37 GHz (e.g. Cavalieri et
al., 1984; Comiso, 1986). On the other hand, this has not necessarily been the case with ice
thickness. Theoretically and from tank experiments, it has been shown that sea ice becomes
opaque at all the microwave frequencies used at a thickness of less than a few centimetres
(Troy et al., 1981; Grenfell and Comiso, 1986). Thus in a practical sense, the signals received
by this sensor originate from the top few centimetres of sea ice and cannot include the direct
information of ice thickness. Even so, if there are any surface dielectric properties of sea ice
that do correlate well with ice thickness, they differentiate the emissivity and then allow the
estimation of ice thickness. Fortunately, the dielectric properties depend on brine volume
which is determined by salinity and temperature (Vant et al., 1978), and it is generally
known that ice bulk salinity decreases with the increase of thickness through desalination
over time by brine drainage and seasonal “flushing” by melt water (Kovacs, 1996).
Furthermore, it was found for the Sea of Okhotsk ice that the surface brine volume fraction
is highly correlated with ice thickness for ice thinner than 20 cm due to the multiplying
effect of salinity and temperature (Figure 9). These observational results provide the
physical grounds for the potential estimation of ice thickness with passive microwave
sensors.
In the field, measurements of microwave emission over sea ice area were initiated using
aircraft in the late 1960’s (Wilheit et al., 1972). This was aimed at a preparation for the
coming satellite observations, and they found that thin and thick ice can be discriminated at
about 30 GHz due to the contrast in their emissivities at this frequency. Since the first
scanning passive microwave sensor (Electrically Scanning Microwave Radiometer; 19.3
GHz) was flown on the Nimbus-5 in 1972, significant effort has been put into classifying sea
ice types using passive microwave signals. The ESMR sensor was succeeded by the
Scanning Multichannel Microwave Radiometer (SMMR) on Nimbus-7 (1978-1987), followed
by the Special Sensor Microwave/Imager (SSM/I) on DMSP (1987-present) and the
Advanced Microwave Scanning Radiometer-EOS (AMSR-E) since 2002. To interpret the
satellite-derived data, the fundamental microwave properties of sea ice were studied from
field experiments, and the dependence of emissivity on ice types and microwave
frequencies was investigated in detail (Troy et al., 1981; Svendsen et al., 1983). As a result, it
was revealed that while the microwave emissivity increases as sea ice grows from open
water or nilas to first-year ice, it decreases to an intermediate value for Arctic multi-year ice,
and that the difference in emissivity between ice types is large at frequencies below 40 GHz
(Figure 10). Based on theses results, Cavalieri (1994) succeeded in mapping ice type
distribution in the Arctic Ocean, categorizing sea ice into nilas (< 10 cm thick), young (10-30
cm), first-year (> 30 cm), and multi-year ice with the SSM/I 19 and 37 GHz data. More
recently, and from an aircraft campaign in the Sea of Okhotsk Naoki et al. (2008) showed
Fig. 9. Relationship between ice thickness and brine volume fraction for (a) bulk and (b)
surface (top 9cm layer) mean for the Okhotsk ice. (from Toyota et al. (2007))
Fig. 10. Emissivity response curves in the 14 to 90 GHz range for four ice types.
(cited from Troy et al. (1981))
that even the estimation of ice thickness for ice thinner 20 cm is possible with microwave
data at 19 and 37 GHz. In fact, on the basis of the ice thickness calculated from AVHRR-
derived ice surface temperature and a simple thermodynamic model, the algorithm which
estimates the thin ice thickness (< 20 cm) was successfully developed with SSM/I 37 GHz
data for the Chukchi Sea (Martin et al., 2004) and with SSM/I 37 and 85 GHz data for the
Antarctic (Tamura et al., 2007). Martin et al. (2005) further extended the analysis using
AMSR-E data for the Chukchi Sea. In addition to the four thickness categories, Tateyama et
al. (2002) and Tamura et al. (2007) proposed the use of the 85 GHz data to detect fast ice area.
In summary it is concluded that passive microwave data can be usefully used to estimate ice
thickness distribution for ice thinner than 20 cm. This is important because it is over the thin
ice area that the most intense heat exchange between ocean and atmosphere occurs in winter
(Maykut, 1978), producing high rates of brine rejection from growing ice and dense water
formation. With SSM/I-derived ice thickness information, Tamura et al. (2008) mapped the
amount of ice production around the Antarctic coast and discussed the geographical
features. However, it should be noted that the spatial resolution is rather coarse (25 km for
SSM/I 37 GHz data) and that at present this method is available only for ice at an early
stage of growth. It has not been guaranteed for deformed ice or snow-covered ice yet
because surface roughness or snow also has significant influence on the emissivity (Grenfell
and Comiso, 1986; Brown and Armstrong, 2008). For such ice conditions, further studies are
needed.
2.5.2 Active microwave methods

The key merit of this sensor is, as with passive microwave, that it is available irrespective of
either cloud or darkness. In addition, the advent of space-borne Synthetic Aperture Radar
(SAR) has made it possible to obtain the backscatter coefficient data with a horizontal
resolution of 10 to 100 m, although the swath width is limited to a few tens to hundreds of
kilometres. In general, the backscatter coefficient includes the effects of both surface
scattering and volume scattering, depending on the microwave frequency and ice conditions
such as salinity, roughness (degree of deformation) and temperature. Snow cover tends to
be largely transparent unless wet. For typical first-year ice with a salinity of 5 psu at a
temperature of -5 deg. C, the penetration depth of the microwaves is only about 10 cm at the
C-band (Hallikainen and Winebrenner, 1992). Therefore, the backscatter coefficient is
predominantly determined from surface scattering for first-year ice. Since the surface
scattering includes the effects of both the dielectric properties of the top layer and the
surface roughness of the ice, the principle of the method is to relate either of these two
parameters to ice thickness. As mentioned in the previous section, the former is well
correlated with ice thickness for ice thinner than 20 cm. Therefore if these two effects can be
separated, the estimation of ice thickness distribution becomes possible for thin ice.
Fortunately, it was found that the effect of surface roughness can be reduced by taking the
ratio of the VV (vertical) to HH (horizontal) backscattering (Winebrenner et al., 1989). In fact,
from field experiments using an airborne L-band SAR in the Sea of Okhotsk, it was revealed
that the VV-HH ratio is well correlated with ice thickness for relatively thin ice (Figure 11;
Wakabayashi et al., 2004; Nakamura et al., 2005). Thus, it was confirmed that polarimetric
data are particularly useful as a means of remotely estimating thin ice thickness distribution
in the peripheral Arctic seas.
Then what about thick ice distribution? Toyota et al. (2009) presented the usefulness of L-
band SAR backscattering to estimate thick ice thickness distribution in the SSIZ from the
airborne Pi-SAR campaign in the Sea of Okhotsk. The main results and conclusions of this
work will be summarized here.
It is generally known that in the SSIZ ridging (pile-up) activities are a key process by which
the ice thickens (e.g. Worby et al., 1996; Toyota et al., 2007). It follows therefore that the
degree of surface roughness correlates with ice thickness. As discussed by Dierking and
Busche (2006), L-band SAR data (wavelength = 24 cm) seem more suitable for this purpose
than C-band (6 cm) data. This is because the longer wavelength approximates the horizontal
Fig. 11. Relations between ice thickness and VV-to-HH backscattering ratio obtained from
airborne L-band SAR in the Sea of Okhotsk. (cited from Wakabayashi et al. (2004))
roughness scale of the ice surface more closely (Massom, 2006). Spaceborne L-band SARs
have been launched on JERS-1 and ALOS, while C-band instruments have flown on ERS-1
and -2, Radarsat-1 and -2 and Envisat. Details are given in Lubin and Massom (2006).
To verify this idea, Toyota et al. (2009) conducted ship-borne electromagnetic (EM)
induction sounding and supersonic profiling observations with an icebreaker, coordinated
with airborne L-band SAR observations in the southern Sea of Okhotsk in February 2005
(Figure 12). The general conditions of ice thickness in the study area were of undeformed ice
thickness ranging from 0.3 to 0.7 m with an average of 0.33 m, and a total average ice
thickness, including ridged ice area fraction (26%), was estimated to be 0.84 m. The surface
elevation was estimated with a supersonic profiler by representing the ship’s motion with a
low-pass filter.
Based on results obtained along a 40 km observation line, radar backscattering coefficients
correlated well with ice thickness (correlation coefficient = 0.88) and surface roughness (i.e.
0.80), defined by the standard deviation of surface elevation (Figure 13abc). Here, it should
be noted that the theoretical microwave penetration depth is up to 0.50 m at L-band for first-
year ice with a salinity of 5.1 psu (Hallikainen and Winebrenner, 1992). Therefore,
backscattering is affected by the internal structure of the sea ice, which should work to
decrease the correlation with ice thickness. Importantly and in spite of this property, the L-
band backscatter coefficient corresponds well with surface roughness. On the other hand,
the VV-HH ratio has a lower correlation with ice thickness (i.e. -0.77) than either VV or HH
(Figure 13d). This means that backscatter coefficients are useful for estimating the thickness
distribution for thick ice not through the dielectric properties but through roughness
properties. All these results shed light on the possibility for determining ice thickness
distribution in SSIZ from polarimetric radar backscatter data.
Fig. 12. Location map of the Pi-SAR experiment conducted in mid-February 2005 in the Sea
of Okhotsk. Solid lines denote the observation line selected for analysis. (cited from Toyota
et al. (2009))
Fig. 13. Scatter plots of the 1 km-averaged values between (a) surface roughness and
Backscatter coefficient (VV), (b) surface roughness and ice thickness, (c) backscatter
coefficient (VV) and ice thickness, (d) backscattering ratio (VV-HH) and ice thickness.
(cited from Toyota et al. (2009))
In summary, ice thickness retrieval for first-year ice by L-band SAR is promising for a wide
range of ice thickness by virtue of the effect on the dielectric constant of surface properties
for relatively thin undeformed ice and of surface roughness for thicker ice. On the other
hand, multi-year ice has significantly different properties in the Arctic in particular. Due to
desalination in summer, the penetration depth becomes deeper for Arctic multi-year sea ice,
and then the volume scattering becomes more significant at L-band. In this case, C-band is
more appropriate to distinguish between multi-year ice and deformed first-year ice (Rignot
and Drinkwater, 1994). Therefore, as pointed out by Rignot and Drinkwater (1994),
ultimately the combining of C- and L-band is optimal for the overall classification of ice
types. Since the L-band data are now available from the ALOS/PALSAR with polarimetric
and ScanSAR modes, combined analysis with C- and L-band is expected in the near future.
2.5.3 Altimeter methods

This sensor measures the freeboard of sea ice (i.e. its height relative to sea level), and
freeboard is then converted to ice thickness by assuming that a floating ice floe is in isostatic
balance. At present there are two kinds of altimeters available for this type of ice thickness
estimation: radar and Laser profiler. Radar altimeters operating at a frequency of 13.8 GHz
on the ERS-1 and ERS-2 satellites were launched in 1991 and 1995, respectively. The
coverage area for these data extends to 81.5°N and S with a footprint of ~1 km. Since it has
been confirmed from the laboratory experiments that reflected echoes come from the
interface between sea ice and snow (Beaven et al., 1995), this sensor basically measures the
height of snow/ice interface relative to sea level with an accuracy of 10 cm. The relatively
large footprint limits the applicable area to closely packed ice area, however the key
advantage remains that the data can be obtained irrespective of weather conditions, albeit
along a profile. Conversely, the laser altimeter (Geoscience Laser Altimeter System or GLAS)
onboard ICESat satellite, launched in January 2003, operates at a wavelength of 1064 nm and
covers the area up to 86°N and S. In this case, the sensor measures the elevation of the ice or
snow surface within a footprint of 70 m spaced at 170 m and an overall accuracy of 14 cm
(Kwok and Cunningham, 2008). Although such a small footprint promises detailed analysis,
it should be noted that this sensor is subject to cloud contamination, which may be critical
for the polar research, as pointed by Haas (2003). A schematic figure is shown in Figure 14
and the calculation for each sensor is as below:
Radar: h = ρw ρs (1)
i hf + hs
ρ w − ρi ρ w − ρi
and
Laser: hi =
ρw ρw − ρs (2)
hts − hs
ρ w − ρi ρ w − ρi
, where ρ i , ρ s , and ρ w are the densities of sea ice, snow and seawater, respectively, and
the notation of h is given in Figure 14. For both altimeters, the key need is to estimate the
reference sea level, and to incorporate accurate estimates of the density of seaweater, snow
and sea ice, and the snow depth.
This method was first applied in the Arctic Ocean by Laxon et al. (2003), using the ERS radar
altimeter. In this case, the reference sea level was analyzed using individual echoes with
some corrections for orbits, tides, and atmospheric pressure (Peacock and Laxon, 2004),
Fig. 14. Schematic picture showing the variables of snow depth (hs), ice thickness (hi), ice
freeboard (hf), and total freeboard (htf).
and the density of sea ice and seawater was assumed to be constant (i.e. 915 and 1024 kg m-
3).The density and depth of snow were taken from a monthly climatology. Submarine sonar
data were used for validation of the ice thickness product. As a result, the authors succeeded
in extracting the geographical distribution of the Arctic ice thickness and showed the
availability of this method for ice thicker than 1 m. Furthermore and on the basis of an 8-
year data set (1993-2001), they found out that the interannual variation is much larger than
predicted by numerical models, and that this large variability could be better explained by
the length of melting season than expected changes of atmospheric and oceanic circulation.
Similarly the results from the algorithm used by the ICESat GLAS laser altimeter has
recently been analyzed by Kwok et al. (2004), Kwok & Cunningham (2008) to the Arctic
Ocean, and by Zwally et al. (2008) in the Weddell Sea. Kwok & Cunningham (2008) used the
ECMWF outputs for snow data, and tested two cases of sea ice density as a constant value
(925 kg m-3) and a function of ice thickness, while Zwally et al. (2008) used AMSR-derived
snow depth and constant densities for snow (300 kg m-3) and sea ice (915 kg m-3). The
method for estimating the reference sea level was developed individually. As a result, these
studies all succeeded in showing the potential usefulness of this method. However, as
mentioned by both of the authors, the improved knowledge about snow is crucial because of
the sensitivity of laser altimetry to the temporal variation of snow depth and density.
In summary, it has been shown that satellite altimeters are useful as a means of estimating
ice thickness distributions over a wide scale, and for thick ice in particular. The primary
merit is the ability to obtain hemispheric ice thickness distribution on a seasonal time scale,
which allows us to monitor trends in ice thickness change and validate the results from
numerical models. In turn, this will feed back to an improvement in the way dynamical
processes are prescribed in the sea ice models. However, it should also be noted that
improved knowledge about densities of snow and sea ice is still needed for this method,
since observational data show that they range widely and also evolve over time according to
the meteorological conditions. In addition, we need to consider the effect of the seawater-
filled cavities between ice blocks on the altimeter estimates. Since such cavities increase the
bulk ice density, the further investigation will be needed. An additional concern relates to
the footprint size. Since this method is based on the assumption of isostatic balance, research
needs to be carried out to determine whether the important underlying assumption of
isostatic balance holds at the scale of the footprints.
3. Concluding Remarks
In this chapter, a brief history of ice thickness measurement has been presented, stressing
how remote sensing tools have contributed to an improved understanding of the sea ice
thickness distribution and its importance in sea ice research. Since Weyprecht first
conducted the observation of the ice thickness growth at the Franz-Joseph Island (80°N)
during the 1873/74 winter (Zubov, 1945), ice thickness has been one of the most challenging
and interesting topics in sea ice research. As seen above, in general the tools have been
developed to cover a wider region. Most recently satellite altimeters have allowed us to
monitor the ice thickness distribution at a hemispheric scale and on a seasonal time scale,
albeit limited to thick ice. Since the response of sea ice associated with recent global
warming receives much attention now, it is a notable achievement to be able to discuss the
ice thickness trends not only from the products of numerical models but also on the basis of
observational results. This will also be helpful for the improvement of physical processes in
the models, and validation of their output.
In spite of these advances, it is unlikely that all of the various methods developed to date
will be replaced by satellite sensors. The individual tools are appropriate for measurement
at their optimum scale. In general, ice thickness varies in a hierarchical fashion, as suggested
by Rothrock and Thorndike (1980), and the properties at each scale have an influence on those
at a different scale. For example, the ice thickness distribution at a scale of a few to tens of
meters is directly involved in the formation of ridges, while it is the thickness properties at a
scale of a few kilometres that determine the external forcing for ridging. Thus, a more
comprehensive understanding of sea ice behaviours depends upon acquisition of accurate
information on the ice thickness distribution at both local and large scales. Considering that ice
formation and melting are basically localized processes, the interaction between different
scales needs to be further clarified when using the different though complementary tools at
different scale; this factor will become more important to the sea ice research in the future.
It should be mentioned that there are other methods widely used for ice thickness
measurement. The representative one is visual observation on the basis of ASPeCT protocol,
which was designed for the ship-based observations in the Antarctic seas by Australian
researchers based on a previous method used by other groups in the 1980’s (Worby &
Allison, 1999). The observation is conducted every hour from the ship’s bridge along-track,
including the thickness of sea ice and snow, floe size, and surface properties. Although there
may be some problems with accuracy and bias from the ship-track, this is a convenient
method and the statistics obtained represent useful general characteristics of the ice cover. A
climatology of sea ice thickness obtained in this fashion are presented in Figure 1.2b. The
continuation of this type of observing program is to be encouraged.
Finally, I would like to close this chapter by concluding that the history of ice thickness
measurement is closely related to the development of sea ice research on both macro- and
micro- scale properties, and that many researchers from a variety of backgrounds have been
involved in it. Therefore, this is truly an interdisciplinary issue, and further collaboration
with many researches is highly desirable in future.
4. Acknowledgments
The author expresses his sincere gratitude to Dr. Robert Massom (Antarctic Climate and
Ecosystems Cooperative Research Centre and Australian Antarctic Division) for the critical
reading of this manuscript and valuable comments and to Dr. Guy Williams (Hokkaido
Univ.) for the proof-reading. They are very helpful for the improvement of the manuscript.
5. References
Ackley, S.F. (1979). Mass-balance aspects of Weddell Sea pack ice, J. Gloaciol., Vol. 24, No.90,
pp. 391-405.
Beaven, S.G.; Lockhart, G.L.; Gogineni, S.P.; Hossetnmostafa, A.R.; Jezek, K.; Gow, A.J.;
Perovich, D.K.; Fung, A.K. & Tjuatja, S. (1995). Laboratory measurements of radar
backscatter from bare and snow-covered saline ice sheets, Int. J. Remote Sens., Vol.
16, No. 5, pp. 851-876.
Bourke, R.H. & Garrett, R.P. (1987). Sea ice thickness distribution in the Arctic Ocean, Cold
Regions Science and Technology, Vol.13, pp.259-280.
Brown, R. & Armstrong, R.L. (2008). Snow-cover data: measurement, products, and sources,
In Snow and Climate, Armstrong, R.L. & Brun, E. (ed.), pp. 181-216, Cambridge
University Press, ISBN 978-0-521-85454-2, Cambridge, UK.
Cavalieri, D.J.; Gloersen, P. & Campbell, W.J. (1984). Determination of sea ice parameters
with the NIMBUS 7 SMMR, J. Geophys. Res., Vol. 89, No.D4, pp. 5355-5369.
Cavalieri, D.J. (1994). A microwave technique for mapping thin sea ice, J. Geophys. Res.,
Vol.99, No.C6, pp. 12561-12572.
Comiso, J.C. (1986). Characteristics of Arctic winter sea ice from satellite multispectral
microwave observation, J. Geophys. Res., Vol.91, No.C1, pp.975-994.
Dierking, W. & Busche, T. (2006). Sea ice monitoring by L-band SAR: An assessment based
on literature and comparisons of JERS-1 and ERS-1 imagery, IEEE Trans. Geosci.
Remote Sensing, Vol. 44, No.2, pp. 957-970.
Fukamachi, Y.; Mizuta, G.; Ohshima, K.I.; Melling, H.; Fissel, D. & Wakatsuchi, M. (2003).
Variability of sea-ice draft off Hokkaido in the Sea of Okhotsk revealed by a
moored ice-profiling sonar in winter of 1999, Geophys. Res. Lett., Vol.30, No.7, 1376,
doi:10.1029/2002GL016197.
Fukamachi, Y.; Mizuta, G.; Ohshima, K.I.; Toyota, T.; Kimura, N. & Wakatsuchi, M. (2006).
Sea ice thickness in the southwestern Sea of Okhotsk revealed by a moored ice-
profiling sonar, J. Geophys. Res., Vol.111, C09018, doi:10.1029/2005JC003327.
Gow, A.J.; Ackley, S.F.; Weeks, W.F. & Govoni, J.W. (1982). Physical and structural
characteristics of Antarctic sea ice, Ann. Glaciol., Vol.3, pp.113-117.
Grenfell, T.C. & Comiso, J.C. (1986). Multifrequency passive microwave observations of
first-year sea ice grown in a tank, IEEE Trans. Geosci. Remote Sensing, Vol.GE-24,
No.6, pp. 826-831.
Haas, C. (1998). Evaluation of ship-based electromagnetic-inductive thickness measure-

ments of summer sea-ice in the Bellingshausen and Amundsen Seas, Antarctica,
Cold Reg. Sci. Technol., Vol. 27, pp. 1-16.
Haas, C. (2003). Dynamics versus thermodynamics: The sea ice thickness distribution, In Sea
Ice, Thomas, D.N. & Dieckmann, G.S. (Ed.), pp. 82-111, Blackwell Publishing, ISBN
0-632-05808-0, Oxford, U.K.
Haas, C.; Pfaffling, A.; Hendricks, S.; Rabenstein, L.; Etienne, J.-L. & Rigor, I. (2008).
Reduced ice thickness in Arctic Transpolar Drift favors rapid ice retreat, Geophys.
Res. Lett., Vol.35, L17501, doi:10.1029/2008GL034457.
Haas, C.; Lobach, J.; Hendricks, S; Rabenstein, L. & Pfaffling, A. (2009). Helicopter-borne
measurements of sea ice thickness, using a small and lightweight, digital EM
system, J. Appl. Geophys., Vol.67, pp.234-241.
Hallikainen, M. & Winebrenner, D.P. (1992). The physical basis for sea ice remote sensing, In
Microwave Remote Sensing of Sea Ice, Carsey (ed.), pp. 29-46, American Geophysical
Union, Washington D.C.
Harms, S.; Fahrbach, E. & Strass, V.H. (2001). Sea ice transports in the Weddell Sea, J.
Geophys. Res., Vol.106, No.C5, pp. 9057-9074.
Hudson, R. (1990). Annual measurement of sea-ice thickness using an upward-looking
sonar, Nature, Vol.344, pp.135-137.
Japan Meteorological Agency (2005). The Results of Sea Ice Observations, Vol. 23, (2004-2005).
(CD-ROM)
Kovacs, A.; Valleau, N.C. & Holladay, J.S. (1987). Airborne electromagnetic soundings of sea
ice thickness and sub-ice bathymetry, Cold Reg. Sci. Technol., Vol.14, pp. 289-311.
Kovacs, A. & Holladay, J.S. (1989). Development of an airborne sea ice thickness measure-
ment system and field test results, CRREL Report, Vol.89-19, pp. 1-47.
Kovacs, A. (1996). Sea Ice. Part I. Bulk salinity versus ice floe thickness. USA Cold Regions
Research and Engineering Laboratory, CRREL Rep., Vol.97-7, pp. 1-16.
Kwok, R. & Rothrock, D.A. (1999). Variability of Fram Strait ice flux and North Atlantic
Oscillation, J. Geophys. Res., Vol.104, No.C3, pp.5177-5189.
Kwok, R.; Zwally, H.J. & Yi, D. (2004). ICESat observations of Arctic sea ice: First look,
Geophys. Res. Lett., Vol. 31, L16401, doi:10.1029/2004GL020309.
Kwok, R. & Cunningham, G.F. (2008). ICESat over Arctic sea ice: Estimation of snow depth
and ice thickness, J. Geophys. Res., Vol. 113, C08010, doi:10.1029/2008JC004753.
Lange, M.A. & Eicken, H. (1991). Textural characteristics of sea ice and the major mechanism
of ice growth in the Weddell Sea, Ann. Glaciol., Vol. 15, pp. 210-215.
Laxon, S.; Peacock, N. & Smith, D. (2003). High interannual variability of sea ice in the Arctic
region, Nature, Vol. 425, pp. 947-950.
Lubin, D. & Massom, R. (2006). Polar Remote Sensing, Volume 1: Atmosphere and Oceans, Praxis
Publishing Ltd and Springer, ISBN 3-540-43097-0, Chichester, UK and Berlin,
Germany.
Marko, J.R. (2003). Observation and analyses of an intense wave-in-ice event in the Sea of
Okhotsk, J. Geophys. Res., Vol.108, No.C9, 3296, doi:10.1029/2001JC001214.
Martin, S.; Drucker, R.; Kwok, R. & Holt, B. (2004). Estimation of the thin ice thickness and
heat flux for the Chukchi Sea Alaskan coast polynya from Special Sensor
Microwave/Imager data, 1990-2001, J. Geophys. Res., Vol. 109, No.C10012, doi:
1010.29/ 2004JC002428.
Martin, S.; Drucker, R; Kwok, R. & Holt, B. (2005). Improvements in the estimates of ice
thickness and production in the Chukchi Sea polynyas derived from AMSR-E,
Geophys. Res. Lett., Vol. 32, L05505, doi:10.1029/2004GL022013.
Massom, R. (2006). Basic remote-sensing principles relating to the measurement of sea ice
and its snow cover, In Polar Remote Sensing Vol. I: Atmosphere and Oceans, Lubin, D.
& Massom, R. (ed.), pp. 356-380, Praxis Publishing Ltd, ISBN 3-540-43097-0,
Chichester, UK and Berlin, Germany, 756 pp.
Massom, R.; Hill, K.; Barbraud, C.; Adams, N.; Ancel, A.; Emmerson, L. & Pook, M.J. (2009).
Fast ice distribution in Adelie Land, East Antarctica: Interannual variability and
implications for emperor penguins (Aptenodytes forsteri), Marine Ecol. Progr. Ser.,
Vol. 374, pp. 243-257.
Maykut, G.A. (1978). Energy exchange over young sea ice in the central Arctic, J. Geophys.
Res., Vol. 83, No.C7, pp. 3646- 3658.
McLaren, A.S. (1989). The under-ice thickness distribution of the Arctic Basin as recorded in
1958 and 1970, J. Geophys. Res., Vol.94, No.C4, pp. 4971-4983.
Melling, H. & Riedel, D.A. (1995). The underside topography of sea ice over the continental
shelf of the Bering Sea in the winter of 1990, J. Geophys. Res., Vol.100, No.C7,
pp.13641-13653.
Melling, H. & Riedel, D.A. (1996). Development of seasonal pack ice in the Beaufort Sea
during the winter of 1991-1992: a view from below, J. Geophys. Res., Vol.101, No.C5,
pp.11975-11991.
Multala, J.; Hautaniemi, H.; Oksama, M.; Lepparanra, M.; Haapala, J.; Herlevi, A.; Riska, K.
& Lensu, M. (1996). An airborne electromagnetic system on a fixed wing aircraft for
sea ice thickness mapping, Cold Reg. Sci. Technol., Vol.24, pp.355-373.
Nakamura, K.; Wakabayashi, H.; Naoki, K.; Nishio, F.; Moriyama, T. & Uratsuka, S. (2005).
Observation of sea ice thickness in the Sea of Okhotsk by using dual-frequency and
fully polarimetric airborne SAR (Pi-SAR) data, IEEE Trans. Geosci. Remote Sensing,
Vol. 43, 2460-2469.
Naoki, K.; Ukita, J.; Nishio, F.; Nakayama, M.; Comiso, J.C. & Gasiewski, Al (2008). Thin sea
ice thickness as inferred from passive microwave and in situ observations, J.
Geophys. Res., Vol. 113, No.C02S16, doi:10.1029/2007JC004270.
Nghiem, S.V.; Rigor, I.G., Perovich, D.K.; Clemente-Colon, P.; Weatherly, J.W. & Neumann,
G. (2007). Rapid reduction of Arctic perennial sea ice, Geophys. Res. Lett., Vol. 34,
L19504, doi:10.1029/2007GL031138.
Peacock, N.R. & Laxon, S.W. (2004). Sea surface height determination in the Arctic Ocean
from ERS altimetry, J. Geophys. Res., Vol. 109, C07001, doi:10.1029/2001JC001026.
Rignot, E. & Drinkwater, M.R. (1994). Winter sea-ice mapping from multi-parameter
synthetic-aperture radar data, J. Glaciol., Vol. 40, No. 134, pp. 31-45.
Rothrock, D.A. & Thorndike, A.S. (1980). Geometric properties of the underside of sea ice, J.
Geophys. Res., Vol. 85, No. C7, pp. 3955-3963.
Rothrock, D.A. (1986). Ice thickness distribution – measurement and theory. In: The
Geophysics of Sea Ice (ed. N. Untersteiner), Plenum, New York, pp. 551-575.
Rothrock, D.A.; Yu, Y. & Maykut, G.A. (1999). Thinning of the Arctic sea-ice cover. Geophys.
Res. Lett., Vol.26, No.23, pp. 3469-3472.
Shimoda, H.; Endoh, T.; Muramoto, K.; Ono, N.: Takizawa, T.; Ushio, S.; Kawamura, T. &
Ohshima, K.I. (1997). Observations of sea-ice conditions in the Antarctic coastal
region using ship-board video cameras, Antarctic Record, Vol.41, No.1, pp.355-365.
(in Japanese with English summary)
Strass, V.H. (1998). Measuring sea ice draft and coverage with moored upward looking
sonars, Deep-Sea Res. I, VOl.45, 795-818.
Strass, V.H. & Fahrbach, E. (1998). Temporal and regional variation of sea ice draft and
coverage in the Weddell Sea obtained from upward looking sonars, In: Antarctic Sea
Ice, Physical Processes, Interactions and Variability, Jeffries, M.O. (Ed.), pp.123-139,
American Geophysical Union, ISBN 0-87590-902-7, Washington, D.C.
Svendsen, E.; Kloster, K.; Farrelly, B.; Johanessen, O.M.; Johanessen, J.A.; Campbell, W.J.;
Gloersen, P.; Cavalieri, D.J. & Matzler, C. (1983). Norwegian remote sensing
experiment: evaluation of the Nimbus 7 scanning multichannel microwave
radiometer for sea ice research, J. Geophys. Res., Vol.88, No.C5, pp. 2781-2792.
Tamura, T.; Ohshima, K.I.; Markus, T.; Cavalieri, D.J.; Nihashi, S. & Hirasawa, N. (2007).
Estimation of thin ice thickness and detection of fast ice from SSM/I data in the
Antarctic Ocean, J. Atmos. Ocean. Tech., Vol. 24, pp.1757-1772.
Tamura, T.; Ohshima, K.I. & Nihashi, S. (2008). Mapping of sea ice production for Antarctic
coastal polynyas, Geophys. Res. Lett., Vol.35, No.L07606, doi:10.1029/
2007GL032903.
Tateyama, K.; Enomoto, H.; Toyota, T. & Uto, S. (2002). Sea ice thickness estimated from
passive microwave radiometers, Polar Meteorology and Glaciology, Vol. 16, pp.15-31.
Thorndike, A.S.; Rothrock, D.A.; Maykut, G.A. & Colony, R. (1975). The thickness
distribution of sea ice, J. Geophys. Res., Vol. 80, No. 33, pp. 4501-4513.
Toyota, T.; Kawamura, T.; Ohshima, K.I.; Shimoda, H. & Wakatsuchi, M. (2004). Thickness
distribution, texture and stratigraphy, and a simple probabilistic model for
dynamical thickening of sea ice in the southern Sea of Okhotsk, J. Geophys. Res.,
Vol.109, C06001, doi:10.1029/2003JC002090.
Toyota, T.; Takatsuji, S.; Tateyama, K.; Naoki, K. & Ohshima, K.I. (2007). Properties of sea ice
and overlying snow in the southern Sea of Okhotsk, J. Oceanogr., Vol.63, No.3, pp.
393-411.
Toyota, T.; Nakamura, K.; Uto, S.; Ohshima, K.I. & Ebuchi, N. (2009). Retrieval of sea ice
thickness distribution in the seasonal ice zone from airborne L-band SAR, Int. J.
Remote Sens., Vol. 30, No. 12, pp. 3171-3189.
Troy, B.E.; Hollinger, J.P.; Lerner, R.M. & Wisler, M.M. (1981). Measurement of the
microwave properties of sea ice at 90 GHz and lower frequencies, J. Geophys. Res.,
Vol. 86, No.C5, pp. 4283-4289.
Tucker, W.B. ; Weatherly, J.W. ; Eppler, D.T. ; Farmer, L.D. & Bentley, D.L. (2001). Evidence
for rapid thinning of sea ice in the weatern Arctic Ocean at the end of the 1980s,
Geophys. Res. Lett., Vol.28, No.14, pp. 2851-2854.
Uto, S.; Shimoda, H. & Ushio, S. (2006a). Characteristics of sea-ice thickness and snow-depth
distributions of the summer landfast ice in Lützow-Holm Bay, East Antarctica, Ann.
Glaciol., Vol.44, pp. 281-287.
Uto, S.; Toyota, T.; Shimoda, H.; Tateyama, K. & Shirasawa, K. (2006b). Ship-borne
electromagnetic induction sounding of sea-ice thickness in the southern Sea of
Okhotsk, Ann. Glaciol., Vol. 44, pp. 253-260.
Vant, M.R.; Ramseier, R.O. & Makios, V. (1978). The complex-dielectric constant of sea ice at
frequencies in the range 0.1-40 GHz, J. Appl. Phys., Vol. 49, No.3, pp.1264-1280.
Vinje, T.; Nordlund, N. & Kvambekk, A. (1998). Monitoring ice thickness in Fram Strait, J.
Geophys. Res., Vol.103, No.C5, pp.10437-10449.
Wadhams, P. & Horne, R.J. (1980). An analysis of ice profiles obtained by submarine sonar
in the Meaufort Sea, J. Glaciol., Vol.25, No.93, pp. 401-424.
Wadhams, P. (1983). Sea ice thickness distribution in Fram Strait, Nature, Vol.305, pp. 108-
111.
Wadhams, P.; McLaren, A.S. & Weintraub, R. (1985). Ice thickness distribution in Davis
Strait in February from submarine sonar, J. Geophys. Res., Vol.90, No.C1, pp. 1069-
1077.
Wadhams, p. (1988). The underside of Arctic sea ice imaged by sidescan sonar, Nature,
Vol.333, pp.161-164.
Wadhams, P. (1990). Evidence for thinning of the Arctic ice cover north of Greenland,
Nature, Vol.345, pp. 795-797.
Wadhams, P. (1992). Sea ice thickness distribution in the Greenland Sea and Eurasian Basin,
May 1987, J. Geophys. Res., Vol.97, No.C4, pp. 5331-5348.
Wadhams, P. (2000). Ice in the Ocean, Gordon and Breach Science Publishers, ISBN 90-5699-
296-1, Amsterdam, The Netherlands.
Wakabayashi, H.; Matsuoka, T.; Nakamura, K. & Nishio, F. (2004). Polarimetric
characteristics of sea ice in the Sea of Okhotsk observed by airborne L-band SAR,
IEEE Trans. Geosci. Remote Sensing, Vol. 42, No. 11, 2412-2425.
Wilheit, T.; Nordberg, W.; Blinn, J.; Campbell, W. & Edgerton, A. (1972). Aircraft
measurements of microwave emission from Arctic sea ice, Remote Sensing of
Environment, Vol. 2, pp.129-139.
Winebrenner, D.P; Tsang, L.; Wen, B & West, R. (1989). Sea ice characterization
measurements needed for testing of microwave remote sensing models, IEEE J.
Oceanic Eng., Vol. 14, No.2, pp. 149-158.
Worby, A.P.; Jeffries, M.O.; Weeks, W.F.; Morris, K. & Jana, R. (1996). The thickness
distribution of sea ice and snow cover during late winter in the Bellingshausen and
Amundsen Seas, Antarctica, J. Geophys., Res., Vol.101, pp. 28441-28455.
Worby, A.P.; Griffin, P.W.; Lytle, V.I. & Massom, R.A. (1999). On the use of electromagnetic
induction sounding to determine winter and spring sea ice thickness in the
Antarctic, Cold Reg. Sci. Technol., Vol. 29, pp.49-58.
Worby, A.P. & Allison, I. (1999). A technique for making ship-based observations of
Antarctic sea ice thickness and characteristics, ANTARCTIC CRC Research Report,
No.14, pp. 1-63.
Worby, A.P.; Bush, G.M. & Allison, I. (2001). Seasonal development of the sea-ice thickness
distribution in East Antarctica: measurements from upward-looking sonars, Ann.
Glaciol., Vol.33, pp.177-180.
Worby, A.P.; Geiger, C.A.; Paget, M.J.; Van Woert, M.L.; Ackley, S.F. & DeLiberty, T.L.
(2008). Thickness distribution of Antarctic sea ice, J. Geophys. Res., Vol.113, C05S92,
doi:10.1029/2007JC004254.
Zubov, N.N. (1945). Arctic ice, Izdatel’stvo Glavsermorputi, Moscow, 360pp. (Translated to
Japanese by Institute of Low Temperature Science, Hokkaido Univ. in 1958).
Zwally, H.J.; Yi, D.; Kwok, R. & Zhao, Y. (2008). ICESat measurements of sea ice freeboard
and estimates of sea ice thickness in the Weddell Sea, J. Geophys. Res., Vol. 113,
C02S15, doi:10.1029/2007JC004284.
Correction of NDVI calculated from ASTER L1B
and ASTER (AST07) data based on ground measurement 45
X3

and ASTER (AST07) data based on ground
measurement*
Hoshino Buho,
Masami Kaneko,
Kenta Ogawa
Rakuno Gakuen University, Japan
Abstract
atmospheric correction of the satellite data, for example correction of visible and
near-infrared spectral bands, requires removing the adjacency effect that is due to the
reflection from contiguous pixels. To evaluate the accuracy of the ATCOR software
atmospheric correction method of Terra/ASTER data (Jun 30, 2002) we compared the NDVI
parameters taken from different sources, such as ground measurement using ground
radiometric measurement data (ASD's FieldSpec® Pro), atmospheric corrected ASTER L1B
data and ASTER surface reflectance product (AST07) data. The study area named Sarobetsu
Marsh is located in the coastal area of Hokkaido Island, Japan. It has been found that 5% of
the scattering radiation is contained within the ASTER Green band, 47% of radiation was
absorbed in the ASTER NIR band and 17% of radiation was absorbed in the ASTER SWIR6
band, respectively. There was no significant difference between ASD's measurement values
and the ATCOR software output values in the ASTER reflection band and absorption bands
of chlorophyll (i.e. NIR-band and Red-band). However, differences were seen in the ASTER
scattering bands (i.e. visible Green band) and soil reflection bands (i.e. ASTER SWIR bands).
Comparison of the original ASD's ground measurement data with the AST07
(©NASA/EOSDIS ASTER surface reflectance product data (L2B) values has revealed very
low reflectance of a NIR band.
Keywords: NDVI, ASTER L1B & AST07, ASD’s FieldSpec®, surface reflectance, ground
measurment.
1. Introduction
The electromagnetic (EM) energy travels through the atmosphere from the sun to target
remote sensing sensors. Radiation from the Earth's surface undergoes significant interaction
with the atmosphere before it reaches the satellite sensors. Regardless of the type of analysis
that is performed on the remotely sensed data, it is important to understand the effect the
atmosphere has made to the radiance responses [1]. In order to acquire an exact radiation of
target, we must correct the atmosphere effect of satellite imagery. Correction of image data
for the effects of atmospheric propagation can be carried out essentially in three ways [2]. It
is respectively, based on atmospheric scattering and absorption characteristics of the
physical model; based on pre-calibration, on-board calibration against targets of known
reflectance method and based on dark-pixel subtraction method. The physically based
methods of attempt to model (for example, Look-up table (LUT) approach and top-of
atmosphere (TOA) radiance) is the most rigorous approach, and also the most difficult to
apply [3] [4]. The atmospheric scattering and absorption characteristics area calculated by a
computer model such as the LOWTRAN-7 [4], MODTRAN [5] [6] and 6S [7] require to input
meteorological, seasonal and geographical data. In practice, these variables may not be
available with sufficient spatial or temporal resolution, and, in particular, estimation of the
contribution of atmospheric aerosols is difficult [3] [8]. In the calibration based atmospheric
correction of VNIR (visible and near infrared), SWIR (short-wave infrared) imagery method,
these targets can be artificially constructed or naturally occurring, but they need to satisfy a
number of criteria[3]: (1) their reflectances must be known sufficiently accurate, in the same
spectral bands as are used by the imager; (2) the range of reflectances represented by the
calibrators must span the range of interest in the sensor; (3) each calibrator should cover an
area of at least several resolution elements; (4) the calibrators should be well distributed
over the entire scene, so that possible variation of atmospheric conditions from place to
place can be assessed and if necessary, allowed for [3]. Dark pixel subtraction is a technique
that determines the pixel in the image with the lowest brightness value. This method is quite
crude: it is that the minimum reflectance in each band is zero, that the atmospheric
correction can be modeled adequately as an additive effect, and that the correction does not
vary from place to place within the scene. To some extent, visual inspection of an image can
determine whether these assumptions are likely to be valid. Zero-reflectance resolution
element can be provided by shadows, and in the near-infrared region, by water bodies [3]
[5] [7]
In this study, we evaluated the accuracy of the atmosphere correction with ATCOR
atmospheric correction algorithm based on ground radiometric measurement data, and
compared also with the radiative transfer code (RTC) based atmospheric corrected ASTER
L2B standard products surface reflectance (AST07) data simultaneously.
1.1 ATCOR algrithm atmospheric correction method

Rayleigh scattering of sunlight in clear atmosphere is the main reason why the sky is blue.
Since blue light is at the short wavelength end of the visible spectrum, it is more strongly
scattered in the atmosphere than long wavelength red light. The result is that the human eye
perceives blue when looking toward parts of the sky other than the sun [9]. So the
atmospheric effects are much larger in these blue band images. As is known to all, healthy
live green plants absorb solar radiation in the photosynthesis active radiation spectral (i.e.
visible Red) region, which they use as a source of energy in the process of photosynthesis.
Leaf cells have also evolved to scatter solar radiation in the near-infrared (NIR) spectral
region, because the energy level per photon in that domain is not sufficient to be useful to
synthesize organic molecules: a strong absorption here would only result in over-heating the
plant and possibly damaging the tissues. Hence, live green plants appear relatively dark in
the PAR (Photosynthesis Active Radiation) and relatively bright in the NIR [10]. By contrast,
clouds and snow tend to be rather bright in the red (as well as other visible wavelengths)
and quite dark in the NIR.
The ATCOR program can correct the path radiance, adjacency radiation and terrain
radiation reflected to the pixel in order to calculate the reflected radiation from the viewed
pixel. ATCOR2 algorithm atmospheric correction algorithm is for a flat terrain working with
an atmospheric database, and ATCOR3 algorithm can correct terrain radiation reflected to
the pixel (from opposite hills, according to the terrain view factor). The database that
contains the atmospheric correction functions is stored in LUT. ATCOR does the
atmospheric correction by inverting the results obtained from MODTRAN, are stored in a
Look up Table. If anything, the ATCOR algorithm method is kind of applied to the
above-mentioned method of the physically based methods attempt to model [11] [12] [13].
1.2 The ASTER surface reflectance product (AST07) algorithm

The validated version of the VNIR/SWIR surface leaving radiance and reflectance products
(product name: ((c) NASA/EOSDIS) AST07) provide an estimate of the total radiance
leaving the surface including both the reflected solar and sky components for ASTER bands
1-9. The atmospheric correction for the VNIR and SWIR is based upon LUT approach using
results from a Gauss-Seidel iteration radiative transfer code [14]. The method has its basis in
the reflectance-based, vicarious-calibration approach of the Remote Sensing Group at the
University of Arizona [15]. We are applying the knowledge learned from our calibration
methods to the atmospheric correction of the VNIR and SWIR bands for ASTER. Specifically,
the RTC we have used for the past 10 years is used as a basis for LUT approach to
atmospheric correction. The method currently assumes atmospheric scattering optical
depths and aerosol parameters are known from outside sources. Using these parameters, a
set of piecewise-linear fits are determined from the LUT that relates the measured satellite
radiances to surface radiance and surface reflectance [16] [17].
2. The dara analyzed in this study
2.1 Resampling the original ASTER data

Wetland monitoring, particularly wetland vegetation classification, is crucial for preserving
valuable wetland ecosystems. The development of remote sensing techniques for wetland
monitoring is urgent. To improve the accuracy of vegetation classification, we have
investigated wetland vegetation classification with multi-temporal ASTER images. However,
for many quantitative applications of ASTER VIR imagery (e. g. calculated the ASTER
NDVI), it is desirable to correct the data for the effects of atmospheric propagation. If the
data is accurately calibrated, the variable that is measured is the radiance (W.sr-1.m-2)
reaching the sensor. But the variable that is wanted is the reflectance (%) of the surface. In
this study we approach to make an atmospheric correction of ASTER VIR imagery based on
calibration against targets of known reflectance. The ground targets object is wetland area,
and the land cover types including water surface, bare soil and green vegetation.
The Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) onboard
NASA’s satellite Terra is a high resolution multispectral radiometer with 14 bands that
covers the visible and near-infrared (VNIR), short wave infrared (SWIR) and thermal
infrared (TIR), and is effective in studying the Earth's surface land cover, vegetation and
mineral resources, etc. We used data from Terra/ASTER original Level 1B VNIR /
SWIR/TIR Data (Time of day (UTC): 1:30, June 30, 2003, Path-108/Row-835, and 1:30, July
12, 2004, Path-109/Row-837, the subset coordinate of the UL Geo N45° 08’, E141° 36’)
supplied by the Earth Remote Sensing Data Analysis Center, Tokyo, Japan (©ERSDAC). In
the ATCOR software, if a 14-bands ASTER image is loaded the default Layer-Band
assignment will be set that input layer 13 (thermal band 13) is set to layer 10 and the output
image will be restricted to 10 bands. The reason for this is that from the 5 ASTER thermal
bands only band 13 is used in ATCOR. In order to carry out calculation between bands, we
re-sampled (layer stacking) this 3 layers with different spatial resolution ASTER VNIR (15
m), SWIR (30 m) and TIR (90 m) data to one layer that has the same spatial resolution (15 m)
dataset, and used this dataset input to ATCOR software.
2.2 ATCOR input parameters

With the ASTER data (Path-108/Row-835), ATCOR input parameters include: Solar zenith
(degrees): 24.8; Solar azimuth (degrees): 147.7; Scene Visibility (km) = 30m; Model for solar
region; fall/spring/rural; various aerosol types: rural; Model for thermal region: fall. Input
satellite data: subset ASTER VNIR-SWIR-TIR, 10-bands one layer data. In the calibration of
ASTER data the Level 1B data is in terms of scaled radiance. The unit conversion coefficients
(defined as radiance per 1 DN) are shown in Table 1. Radiance (spectral radiance) is
expressed in units of W/(m2*sr*um). The true radiance at sensor can be obtained from the
DN values as follows:
L = c0 + c1 × DN (1)
Where, L is radiance, c0 (offset) and c1 (gain) is conversion coefficient; DN is digital

number.
Band No. c0 c1
1 -0.1 0.0676
2 -0.1 0.0708
3 -0.1 0.0862
4 -0.1 0.02174
5 -0.1 0.00696
6 -0.1 0.00625
7 -0.1 0.00597
8 -0.1 0.00417
9 -0.1 0.00318
Table 1. The calibration of ASTER original L1B data (Unit: mW/cm2 sr micron)
3. The study area

The study area Sarobetsu Marsh, is the largest registered wetland of 7000 ha located in
coastal area of northwestern Hokkaido, Japan (Figure 1) and nominated by the Ramsar
Convention on Wetlands in 2005. Test areas are mostly swamp with Sphagnum sp.,
Moliniopsis japonica while the western coastal zone area is dominated by isolated small hills
covered with broad-leaf trees. The terrain of the study area is very flat with the elevation
distributed between 5 m to 15 m. The area was selected because of its natural environment
and perseverance by national and municipal governments. On the other hand recently the
expansion of a non-moor plant (bamboo grass) has been detected in a swamp. So far, for the
monitoring, the classification of the wetland territory into a high moor, low moor and non
moor types, to clarify the invasion front of bamboo grass is very important.
(d) Test sites

er
cent
s4 to r
Visi
Wooden road
(a) s4 s1
s3 s2 s2
n
pa
s3
Ja
s1 s2
s1 s2
P1
(b) Hokkaido
N
s1 - Bamboo grass
s2 - Sphagnum
s3 - Grass land (c) Sarobetsu marsh P2
s4 – Pasture land (after moving)
Test sites
Fig. 1. Location of study area and test sites
4. Result
The atmospheric correction processing evaluated within this chapter is to a major part based
on the technology of the ATCOR-3 atmospheric correction procedure [11][12]. In this
algorithm, the total signal at the sensor consists of four components: (1)-path radiance;
(2)-reflected radiation from the viewed pixel; (3)-scattered radiation from the neighborhood;
and (4)-terrain radiation reflected to the pixel. Figure 2. shows the result of ATCOR
atmospheric corrected ASTER data in the test area. The satellite scene has to be
ortho-rectified to a DEM (digital elevation model) before the ATCOR3 processing starts. The
influence of the neighborhood is neglected. A start value for the ground reflectance of the
surrounding topography is employed whilst the path radiance component is subtracted
from the signal. Shadow cast from surrounding topography is included. However, as our
study area is very flat, the topography effect is rare. Haze removal is the important steps
prior to the application of imagery. This result shows the ATCOR algorithm is a more
effective haze removal and atmospheric correction modeling which combined several
improved methods (see Fig. 2(c)).
The global flux on the ground depends on the large-scale (1 km) average reflectance. The
global flux in the atmospheric LUT's is calculated for a fixed reflectance=0.15 . This iteration
performs the update for the spatially varying average reflectance map of the current scene, if
the adjacency range R > 0 [10]. The empirical BRDF correction is areas of low illumination
(see Fig. 2(b)).
(a) (b) (c) (d)
Fig.2. Result of the ATCOR3 correction. (a) original ASTER/L1B data; (b) the illumination
azimuth angle ; (c) hazeoverlay; (d) atmospheric corrected ASTER data.
Comparison of the ASTER data before and after ATCOR software atmospheric
correction has revealed following results:
(a) The mean values of ASTER band 1 and band 2 decrease after atmospheric correction
(Table 2). This means that the visible green and red band has included not only the
radiance from a target, but also radiance other than an atmospheric scattering is also
included.
(b) Comparison of mean values of NIR and SWIR bands before/after atmospheric correction
shows that the radiance values became larger after atmospheric correction. It means that
the radiation from the target is absorbed by atmosphere before it reaches the satellite
sensor. Atmospheric scattering primarily affects the direction of visible Green and Red
band, and atmospheric absorption primarily affects the direction of NIR and SWIR
bands.
The most significant interaction that undergoes by the thermal infrared radiation when it
passes through the atmosphere is its absorption, primarily due to ozone and water vapor
particles in the atmosphere. At the visible shorter wavelengths (i.e. Green or Red band),
attenuation occurs by scattering due to clouds and other atmospheric constituents, as well as
reflection. The type of scattering in which the energy undergoes is depends upon the size of
the particle. Rayleigh scattering occurs when radiation interacts with air molecules smaller
than the radiation's wavelength, such as oxygen and nitrogen. The degree of scattering is
inversely proportional to the fourth power of the wavelength. When particles are
comparable in size to the radiation wavelength, such as aerosols, it results in Mie scattering
type [18]. The effect of scattering on the visible wavelengths is significant and must be
compensated for when developing empirical relationships through time [19] [20].
Atmospheric scattering primarily affects the direction of short wave radiation. There are
four types of atmospheric scattering: Rayleigh, Mie, Raman and non selective. The most
significant of these types of scattering is Rayleigh scatter, which effects the short visible
wavelengths and results in haze. For ASTER data the scattering is four times as great in
Green band of the electromagnetic spectrum as in the NIR band [21],[22].
DN of original ASTER L1B data (before correction)

Band No. Min Max Mean St dev
1 (Green) 0 255 50.93 34.97
2 (red) 0 210 30.83 23.15
3 (NIR) 0 179 69.82 49.02
4 (SWIR) 0 111 45.21 32.59
5 (SWIR) 0 105 27.82 20.52
6 (SWIR) 0 144 29.90 22.66
7 (SWIR) 0 155 27.49 20.41
8 (SWIR) 0 191 23.83 18.10
9 (SWIR) 0 133 21.01 15.20
DN after ATCOR correction of ASTER data
1 (Green) 0 229 26.47 20.16
2 (Red) 0 255 30.04 26.17
3 (NIR) 0 255 93.53 66.98
4 (SWIR) 0 224 90.89 65.64
5 (SWIR) 0 175 46.17 34.13
6 (SWIR) 0 251 51.75 39.30
7 (SWIR) 0 241 42.42 31.55
8 (SWIR) 0 255 35.86 27.33
9 (SWIR) 0 157 24.34 17.69
Table 2. Comparison of DN before and after ATCOR atmospheric correction of ASTER data
((Path-108/Row-835, UTC: 1:30, June 30, 2002)
Figure 3(1) and 3(2) shows that the ASD's measurement values and the ATCOR output
values have no big difference in the ASTER reflection bands and absorption bands of
chlorophyll (i.e. NIR-band and Red-band); the difference has come out in scattering band (i.e.
ASTER Green band) and soil reflection bands (i.e. ASTER SWIR bands). However, in the
©EOSDIS AST07 (ASTER surface reflectance products), the values are considerably different
in ASTER NIR band. The problem is in low values of NIR after atmospheric correction. In
this research, the results of ATCOR software correction were better than those of AST07
products. Figure 3(1) and 3(2) shows the comparison of ATCOR software atmospheric
correction result and ASTER surface reflectance products (AST07) data, in non moor and
high moor plant samples. In a swamp (high moor plant), the background soil and the open
water area will be incorrectly recognized as moving haze of ATCOR software method. After
ATCOR correction the ASTER SWIR bands values are becomes larger following the changes
of vegetation due to dryness.

50
Bamboo grass
40 (100%) (ASD's)
REFLECTANCE (%)
30 Bamboo grass
dense (ATCOR)
20
Bamboo grass
10 dense (AST07)
0
1 2 3 4 5 6 7 8 9
ASTER SPECTRAL BANDS
Fig. 3 (1) Comparison of the spectral reflectance of non-moor plant Bamboo grass calculated
form ASD's measurement method, ATCOR method and EOSDIS AST07 method.
ATCOR have rectified more correctly scattering with short wavelength visible ((i.e. Band 1
(Green)) and absorption with NIR band (Band 3). The Fig. 4(1) and 4(2) clearly shows that
5% of scattering radiation is contained with the green band and 47% of radiation was
absorbed in the NIR band and 17% of radiation was absorbed in the SWIR6 band.
40
REFLECTANCE (%)
Sphagnum dense
30 (80%) (ASD's)
Sphagnum dense
20 (ATCOR)
10 Sphagnum dense
(AST07)
0
1 2 3 4 5 6 7 8 9
ASTER SPECTRAL BANDS
Fig. 3 (2) Comparison of the spectral reflectance of high moor plant Sphagnum marsh
calculated form ASD's measurement method, ATCOR method and EOSDIS AST07 method.
The input (x: DN of original ASTER L1B data) and output (y: DN of after ATCOR software
atmospheric corrected ASTER L1B data) expression of the ASTER data using ATCOR are as
follows:
Green band (band 1): y = 0.95x - 28.56 (2)
NIR band (band 3): y = 1.473x - 21.12 (3)
SWIR band (band 6): y = 1.171x - 1.23 (4)
Comparing NDVI from ground ASD's measurement, corrected ASTER data and not
corrected original ASTER L1B data (see Figure 4(3)), we found that values from the ground
NDVI and atmospheric corrected NDVI did not greatly differed. However, the value of
NDVI of ASTER L1B is smaller than the value of grand NDVI. The formula of the
correlation of a NDVI-Corrected value and an original ASTER L1B NDVI value is as follows:
NDVI corrected = 1.27 × NDVI L1B + 0.04 (5)
This formula showed that the NDVI value after atmospheric correction became larger than
that before atmospheric correction.
70
ATCOR OUTPUT BAND 1 y = 0.9528x - 28.564
60
R^2=0.755
50
40
30
20
10
0
40 60 80 100
ASTER ORIGINAL BAND 1
Fig. 4(1) The correlation coefficient of the ASTER original band 1 (Green) and ATCOR
output band 1 (Green).
300
ATCOR OUTPUT BAND 3
250 y = 1.4729x - 21.121

R^2=0.994
200
150
100
50
0
0 50 100 150 200 250
ASTER ORIGINAL BAND 3
Fig. 4(2) The correlation coefficient of the ASTER original band 3 (NIR) and ATCOR output
band 3 (NIR)
0.9
NDVI-Corrected
ASTER NIR Band
0.8 NDVI-Ground
NDVI-L1B
0.7
0.6
0.5
0.4
0.3
0.5 0.6 0.7 0.8 0.9
ASTER Red band
Fig. 4(3) Comparison of NDVI of ground measurement, atmospheric corrected ASTER L1B
data and not atmospheric corrected original ASTER L1B data
5. Conclusion
Many techniques have been developed to determine the contribution of atmospheric
scattering has on the radiation detected by the satellite sensor. The radiance received from a
target against a background surface by the satellite sensor comes from a combination of
three sources; first, the intrinsic radiance reflected by the target and then directly
transmitted by the atmosphere; secondly, the radiant energy scattered diffusely by the
atmosphere which then further interacts with the target background; thirdly, the radiant
energy scattered diffusely by the atmosphere. The radiant energy reflected by the target
carries the direct energy from the target. The other two sources produce a combined effect.
Atmospheric measurements and modeling involve the theoretical determination of the path
radiance contribution of the atmosphere for the particular time of the overpass. To calculate
the contribution of the scattering on the reflected radiance requires that many atmospheric
variables at the time of the satellite overpass be recorded and input into theoretically
derived equations to determine the effect of the atmosphere on each spectral band.
Correction of image data for the effects of atmospheric propagation can be carried out
essentially in three ways. It is respectively, 1) based on atmospheric scattering and
absorption characteristics of model; 2) based on pre-calibration, on-board calibration against
targets of known reflectance method and 3) based on dark-pixel subtraction method. The
physic based methods are the most rigorous approach, and also the most difficult to apply.
The atmospheric scattering and absorption characteristics area calculated by a computer
model requires meteorological, seasonal and geographical variables. In practice, these
variables may not all be available with sufficient spatial or temporal resolution, and in
particular, estimation of the contribution of atmospheric aerosols is difficult. Other methods
mentioned also have a number of requirements that are not easy to satisfy.
In this study, the accuracy of the atmospheric correction with ATCOR software algorithm
was based on the usage of ground radiometric measurement data, which were compared
also with the radiative transfer code (RTC) that is based on atmospheric corrected ASTER
L2B standard products surface reflectance (AST07) data simultaneously.
The NDVI data calculated from ground measurement, atmospheric corrected ASTER L1B
data and ASTER surface reflectance product (AST07) data was used to evaluate the accuracy
of the ATCOR software atmospheric correction of Terra/ASTER data (Jun 30, 2002). Ground
measurements were done using ground radiometric measurement data (ASD's FieldSpec®
Pro) at the study area named Sarobetsu Marsh located in coastal area of Hokkaido, Japan.
The study showed that the background soil and leaf area affected the accuracy of ATCOR. It
has been found that 5% of scattering radiation is contained with the ASTER Green band and
47% of radiation was absorbed in the ASTER NIR band and 17% of radiation was absorbed
in the ASTER SWIR6 band. The ground measurement values and the ATCOR software
output values were of no big difference in the ASTER reflection band and absorption bands
of chlorophyll (i.e. NIR-band and Red-band). However, the difference was seen in the
ASTER scattering bands (i.e. visible Green band) and soil reflection bands (i.e. ASTER SWIR
bands). Compared with the data of ASD's measurement, the AST07 (©NASA/EOSDIS
ASTER surface reflectance product data (L2B)) values are too low in a NIR band.
For the ASTER original L1B data (Jun 30, 2001), the statistics mean value of Green band and
NIR band is 50.9 and 69.8 after atmospheric correction, the statistics mean value of Green
band and NIR band is 26.5 and 90.9. (For the ASTER/July 12, 2004 case, the values was 53.8
and 83.5 to 27.5 and 131.4 respectively). The value of NDVI after atmospheric correction is
larger than atmospheric correction before, and this rate of change is (NDVI-Corrected) =
1.27 (NDVI-L1B) + 0.04.
Comparison of accuracy of the ATCOR software atmosphere correction of non-moor plant and
high moor plant area ASTER imagery showed that the background soil and leaf area affected
the accuracy of ATCOR. In the case of a moor plant, the error in ASTER Green band is large.
6. Acknowledgment
This work was supported by Grant-in-Aid for Scientific Research (A) 21370005 and the
Global Environmental Research Fund (F-092) by the Ministry of the Environment, Japan.
7. References
[1] Schott J.R. and Henderson-Sellers, A., “Radiation, the Atmosphere and Satellite
Sensors, Contribution in Satellite Sensing of a Cloudy Atmosphere: Observing the
Third Planet (Ed: Henderson-Sellers), ” 1984, pp45-89.
[2] Campbell, J. B., “Introdution to Remote Sensing: Second Edition,” The Guilford Press,
New York. 1996, pp. 444-476.
[3] W.G.Rees, “ Physical Principles of Remote Sensing, Second Edition, ” Cambridge
University Press. 2001, pp. 76-85.
[4] Kneizys, F.X., “Atmospheric Transmittance Radiance: computer code ” LOWTRANG
AFGL-TR-83-0187 (1983).
[5] Berk, A., L. S. Bernstein, and D. C. Robertson, “MODTRAN. A Moderate Resolution
Model for LOWTRAN 7,”. GL-TR-89-0122, Phillips Laboratory, Geophysics
Directorate, Hanscom Air Force Base, Massachusetts, 1989.
[6] Berk, A., L.S. Bernstein, G.P. Anderson, P.K. Acharya, D.C. Robertson, J.H., “Chetwynd
and S.M. Adler-Golden, MODTRAN Cloud and Multiple Scattering Upgrades with
Application to AVIRIS, ,” Remote Sens. Environ, 1998, pp.65:367-375.
[7] E. Vermote, D. Tanr, J. Deuz, M. Herman, and J. Morcette, “Second simulation of the
satellite signal in the solar spectrum 6S: An overview, ” IEEE Trans. Geosci. Remote
Sensing, vol. 35, no. 3, 1997, pp. 675—686.
[8] Fraser, R. S, R. A. Ferrare, Y. J. Kaufman, B. L. Markham, S. Mattoo, “Algorithm for
atmospheric correction of aircraft and satellite imagery, ” Int'l. J. Rem. Sens., 1992,
pp.13:541-557.
[9] Nave, Carl Rod, 2005. Hyperphysics-Quantum Physics, Department of Physics and
Astronomy, Georgia State University, CD.
[10] Gates, David M. (1980) Biophysical Ecology, Springer-Verlag, New York, pp. 611.
[11] R. Richter, “A fast atmospheric correction algorithm applied to Landsat TM, ” Int. J.
Remote Sensing 1990, pp.11:159-166.
[12] R. Richter, "Correction of atmospheric and topographic effects for high spatial
resolution satellite imagery", Int. J. Remote Sensing 18:1099-1111 (1997).
[13] Richter, R. and Schläpfer, D., “Geo-atmospheric processing of airborne imaging
spectrometry data. Part 2:Atmospheric/Topographic Correction”. International
Journal of Remote Sensing, 23(13): 2631-2649 (2002).
[14] Herman, B. M., & Browning, S. R., “A numerical solution to theequation of radiative
transfer, ” Journal of the Atmospheric Sciences, 22, 1965, pp.559-566.
[15] Slater, P. N., Biggar, S. F., Holm, R. G., Jackson, R. D., Mao, Y., Moran, M. M., Palmer, J.
M., & Yuan, B., “Reflectance- and radiance-based methods for the in-flight absolute
calibration of multispectral sensors, ” Remote Sensing of Environment, 1987, pp.22,
11 – 37.
[16] .Aosier, B. Kaneko, M. Takada, M., Comparison of NDVI of ground measurement,
atmospheric corrected ASTER L1B data and ASTER surface reflectance product
(AST07) data, Geoscience and Remote Sensing Symposium, 2007 IEEE
International; Barcelona,Spain, Publication Date: 23-28 July 2007, pp. 1806-1811.
[17] Buhe Aosier, K. Tsuchiya, M. kaneko, S. J. Saiha, “Comparison of Image Data
Acquired with AVHRR, MODIS, ETM+, and ASTER Over Hokkaido
Japan,” Advances in Space Research, Vol. 32, No. 11, 2003, pp.2211-2216.
[18] Drury, S.A., “Image Interpretation in Geology. Allen and Unwin (Publishers) Ltd,
London. Forster, B.C. (1980). Urban residential ground cover using Landsat digital
data. Photogram, ” Eng. Remote. Sen. 46, 547, 1987, pp.58.
[19] Coops, N.C., Culvenor, “Utilizing local variance of simulated high-spatial resolution
imagery to predict spatial pattern of forest stands, ” Remote Sensing of
Environment. 71(3): 2000, pp.248-260.
[20] Jensen, J.R., 1986. ‘Introductory Digital Image Processing : A Remote Sensing
Perspective’, New Jersey : Prentice-Hall.
[21] Crippen, R.E., “The Regression Intersection Method of Adjusting Image Data for Band
Ratioing, ” International Journal of Remote Sensing. Vol. 8. No. 2. 1987, pp. 137 155.
[22] Chavez, P. S., Jr., “An improved dark-object subtraction technique for atmospheric
scattering correction of multispectral data ,” Remote Sens. Environ., 24, 1988, pp.
459-479.
[23] Herman, B. M., & Browning, S. R., “A numerical solution to theequation of radiative
transfer, ” Journal of the Atmospheric Sciences, 22, 1965, pp.559-566.
[24] Slater, P. N., Biggar, S. F., Holm, R. G., Jackson, R. D., Mao, Y., Moran, M. M., Palmer, J.
M., & Yuan, B., “Reflectance- and radiance-based methods for the in-flight absolute
calibration of multispectral sensors, ” Remote Sensing of Environment, 1987, pp.22,
11 – 37.
* corresponding author Hoshino Buho: e-mail: aosier@rakuno.ac.jp

Environmental hazards in the El-Temsah Lake, Suez Canal district, Egypt 57
4
X
Environmental hazards in the El-Temsah Lake,

Suez Canal district, Egypt
1Kaiser, M. F.
Geology Department, Faculty of Science, Suez Canal University
Egypt
2Amin, A. S.
Botany Department, Faculty of Science, Suez Canal University
Egypt
3Aboulela, H. A.
Marine Science Department, Faculty of Science, Suez Canal University
Egypt
1. Abstract
Chemical and biological analyses were integrated using remote sensing GIS techniques to
evaluate the environmental pollution of El-Temsah Lake, in the Suez Canal, in order to
provide critical data to enhance development planning and economic projects within the
study area. Fifty-six samples were collected from seven sites in the lake from July 2005 to
May 2006. Samples were collected in each of four seasons, and included 28 surface sediment
samples and 28 water samples. Sediment samples were analyzed for Fe, Pb, Ni, Co, Cu, Mg,
K and Na. The results showed an increase in beach sediment pollution from summer to
winter. Taxonomic analysis of phytoplankton samples revealed 102 taxa, including 56
Bacillariophyceae, 8 Chlorophyceae, 18 Dinophyceae and 20 Cyanophyceae. Chlorophyll a
concentrations ranged from 0.3 to 26 µg l-1, with the highest values during the winter and
lowest values during the summer. These results suggest that beach sediment pollution is
highest in the winter and, at the same time, the water quality conditions in El-Temsah lake
favor oxidation conditions which maximize phytoplankton productivity. In contrast,
sediment pollution and phytoplankton productivity are lowest during summer, which also
corresponds to more alkaline water conditions. The images were rectified and analyzed by
ERDAS IMAGINE 8.9. A 1968 topographic map and enhanced 2005 Landsat Thematic
Mapper images (30 m resolution) were utilized to determine the coastline positions using
ERDAS Imagine 8.9. Image processing techniques were applied using ENVI 4.2 to analyze
the ETM+ image data. Image enhancement was applied. Image data was enhanced
spectrally to verify surface water pollution detected from chemical and biological analyses
and to detect the sources of untreated domestic, industrial and agricultural waste water. In
general, the lake has been subjected to successive shrinking due to human activities,
primarily through extensive building along the shoreline. The uncontrolled growth of cities
is associated with seismic hazards, affecting on buildings and infrastructures, mostly due to
insufficient knowledge of earthquakes activity. Seismic epicentres were recorded along the
Suez Canal from 1904 to 2006. Widespread moderate to micro earthquakes were identified
around the western side of the lake, with scattered events along the eastern side. In general,
water pollution in El-Temsah Lake has been mitigated over the last decade due to successive
dredging and improved water treatment. Most of untreated water was discharged along the
western side of the lake. The eastern part of the lake is less polluted and is, therefore, more
suited for fishing, tourism, urban planning and navigation activities, although higher use of
eastern portion of the lake could accelerate water and sediment quality deterioration in that
region.
2. Introduction
El-Temsah Lake formed in a depression situated in a fault trough (Holmes, 1965 and El
Shazley et al., 1974) covered with Nile Delta sediments. The trough originated tectonically as
part of the "Clysmic Gulf" that represented the first stage of separation along the Red Sea-
Suez Rift during the Late Oligocene-Early Miocene period (El-Ibiary, 1981). The lake covers
about 15 km2 between 300 32` and 300 36` north latitude and 320 16` and 320 21` east
longitude, and is located near the middle of the Suez Canal, at a point 80 km south of Port
Said. The depth of the lake ranges between 6 and 13 m. Following the construction of High
Dam (completed in 1970), lake water quality changed from saline to fresh water due, to a
large degree, to the precipitation of gypsum and mud lamina. The lake is the backbone of a
tourism industry that attracts a large number of holiday visitors. In addition to the visitors,
the tourism and fishing industries employ local citizens and provide a significant portion of
the district revenues. Unfortunately, the increasing number of temporary and permanent
residents has also created higher volumes of waste, including raw liquid and solid
municipal sewage, agricultural runoff and industrial wastewater, all of which end up in El-
Temsah Lake. The lake is also a sink for aliphatic and aromatic hydrocarbons originating
from shipping activities including accidental and incidental oil pollution, ballast water
release and general vessel and facility maintenance. In addition to hydrocarbons, several
other potential chemical contaminants originate from various pollutant sources. The
concentrations of Pb, Zn, Cu and Cd metals have been shown, for example, to be
significantly higher in this area (Abd El Shafy and Abd El Sabour, 1995). Elevated nitrate
and nitrite levels are due to the discharge of sewage, fertilizers and pesticides.
Through 1996, water quality along the northern and western boundaries of the lake has
experienced substantial deterioration due to the rapid development of tourism projects and
continuous waste discharge at Ismailia City (ETPS, 1995). In order to rejuvenate the lake,
improve water quality and help re-establish the fishing and tourism industries, local
authorities have embarked on a national program, using dredging to remove significant
portions of contaminated lake sediment. Moreover, the re-evaluated assessment and
mitigation of seismcity risk around any area, especially lake area is important due to the fast
growing and developing of great and important constructions. The earliest attempts to
review the seismicity and tectonic activity of the overall Suez Canal region were given by
Aboulela, 1994. The main objective of the research described here was to assess the
environmental status of El-Temsah Lake, focusing on heavy metal analysis of beach
sediments and the distribution of phytoplankton. Furthermore, try to make main zoom on
the seismicity pattern around concerned area.
The diversity and density of biological communities are impacted by the physicochemical
characteristics of the ecosystem, including temperature, pH and dissolved oxygen (DO). In
addition, any single species is inevitably impacted by the other species present within the
system (Moss, 1998). Phytoplankton species, for example, are strongly influenced by water
quality, act as the primary source of autochthonous organic material in many lakes and are
the main food for many filter-feeding, primary consumers (Boney, 1989). Monitoring
phytoplankton populations, therefore, is often used as a means of evaluating the health of
fresh, brackish and saltwater ecosystems. Phytoplankton diversity generally shows
substantial temporal variation. In the Suez Canal, 126 diatom species were recorded in the
winter and summer from 1969-1971 (Dorgham, 1974); 102 species were identified in 1983
(Dorgham, 1985); 94 species were observed in 1991 (El-Sherief and Ibrahim, 1993). Because
of the existing historical data on phytoplankton populations in the Suez Canal, the
importance of phytoplankton to overall community structure and the sensitivity of
phytoplankton to environmental perturbations, they were used as the biological monitor in
this study. Earthquakes epicentres distribution occurred overall the study area and its
surrounding were using by recently information recorded data. Geographical information
system (GIS) techniques were used to integrate chemical, biological and seismicity data in
order to evaluate current conditions and make recommendations for sustainable coastal
zone management.
3. Methodology
3.1 Field sampling locations
Six study sites in El-Temsah lake were selected to assess environmental pollution (Fig. 1).
Samples were collected during four seasons between June 2005 and April 2006 from the
following locations: El-Taween, El- Fayrooz, the Bridge, El-Osra, El-Forsan, Beach clubs.
Sediment appearance and composition was dark-gray to blackish, soft, silty sand at the
western sites (El-Taween, El- Fayrooz , the Bridge, El-Osra and El-Forsan), but a lighter-
colored sand at the eastern El-Temsah sites. During the field visits, it was noticed that the El-
Taween area was subject to substantial dumping or leakage of industrial wastes from ships
passing through the lake, as well as disposal of domestic wastewater from businesses,
particularly clubs, along the shore. The El-Fayrooz and Bridge sites are located near the
junction of a western lagoon and El- Temsah lake, and receive large quantities of domestic,
agricultural and industrial wastewater. The El-Osra and El-Forsan sites, in the northern part
of the lake, receive domestic wastewater. In contrast, the eastern part of the lake (Beach club-
1 and Beach club-2) is characterized by much lower pollution exposure. From each of the
seven sampling locations, sediment, water and phytoplankton samples were collected from
the lake. Sediment was collected at plastic bags and was kept for a week in a fridge before
analyses. Subsurface water samples were taken at all sites for phytoplankton identification
and counting. Water samples for physico-chemical studies were collected in polyethylene
bottles of one liter capacity, and laboratory analysis started within few hours from the time
of collection.
3.2 Environmental analyses
3.2.1 Water and sediment analyses

In the field, dissolved oxygen (DO), pH and temperature were determined. Water samples
were analyzed for salinity, conductivity, nitrate and phosphorus, according to Standard
Methods (APHA 1992). For analysis of trace metal concentrations, the 28 sediment samples
were dried at 600C for one day, weighed, and then digested with hot concentrated acid
under reflux conditions. Metal concentrations in all samples were determined by following
the methods described by More and Chapman (1986).
3.2.2 Biological investigations

Phytoplankton samples were preserved in 500 ml glass bottles with lugol's solution
(Utermöhl's, 1958). The standing crop was calculated as the number of cells per liter.
Qualitative analysis was carried out using the preserved as well as fresh samples. These
were examined microscopically for the identification of the present genera and species. The
algal taxa were identified according to the following references: Van-Heurck (1986), Prescott
(1978), Humm and Wicks (1980) and Tomas (1997). Species richness (alpha diversity) was
calculated for each phytoplankton algal group as the average number of species per season.
Relative concentration and relative evenness were expressed according to the equations of
Whittaker (1972), Pielou (1975) and Magurran (1988).
3.2.3 Seismicity pattern investigation
Seismicity activity occurred around the El-Temsah lake and its surrounding was assessed in
this study using information on revealed recorded earthquakes. Seismicity data from 1904 to
2006 were supplied from the National Earthquake Information Center (NEIC), International
Seismological Centre (ISC) and the Egyptian National Seismic Network (ENSN).
3.3 Remote Sensing and GIS analyses

Satellite remote sensing is a potentially powerful tool for detecting shoreline change.
Enhanced Landsat Thematic Mapper images ETM-7 (30 m resolution) for 2004 (Figure 2)
were geometrically corrected and digitized to extract vectors representing shoreline
positions in 2004. ERDAS IMAGINE 8.9 (Leica Geosystems, Norcross, GA, USA) was
utilized for digitizing and editing data input. The outline of El-Temsah lake in 2004 was
compared to the topographic map in 1968. Arc GIS 9.2 (ESRI, Redlands, CA, USA) platforms
were used to organize, analyze and manipulate available data and generate new data for
heavy metal, seismic and biological analyses.
Image segmentation techniques were used to produce vector maps for the El Temsah
coastline to detect the shoreline position and identify changes between 1968 and 2004. In this
study, image registration was carried out using ERDAS IMAGINE 8.9. A second-order
polynomial was used to provide an adequate transformation for registering a full Thematic
Mapper scene to geographical coordinates.
Thematic Mapper band 7 (short-wave infrared) was used to produce a vector map of the
shoreline. Image segmentation was used to delineate the land/sea boundary along the
coastline. A segmentation approach was also necessary in order to compare the results with
previous shoreline maps produced.
4. Results and Discussion

4.1 Metal analyses
Data collected during this study indicate that, in the summer, concentrations of Na, K, Fe
and Mg are elevated along the western side of the lake at the Bridge and El-Taween sites
(Fig. 2). Low levels were identified along the northeastern side, at the El-Forsan and Beach
clubs (Table 1). Concentrations ranged from 9.03 to 28.9 ppm (Na), 2.28 to 10.47 ppm (K),
3.85 to 12.71 ppm (Fe) and 0.42 to 9.69 ppm (Mg). In autumn, concentrations of some metals,
including Na, K, Mg, Fe and Co were higher along the western side of the lake, at the Bridge
area. Concentrations ranged from 65.74 to 127.63 ppm (Na), 5.27 to 34.19 ppm (K), 6.01 to
32.71 ppm (Mg), 7.97 to 102.44 ppm (Fe) and 0.12 to 0.42 ppm (Co). Mg and Na levels were
higher in beach sediments in the winter and spring than during either the summer or
autumn, ranging from 841.5 to 2145.12 ppm (Na) and from 29.82 to 676.89 ppm (Mg). Levels
of both K (10.4 to 27.12 ppm) and Fe (27.59 to 133.25 ppm) in winter-collected sediment were
higher than in the summer. Spring sediment concentrations tended to be very similar to
winter levels. The lowest metal concentrations were identified along the northeastern side of
the lake; samples collected along the western perimeter had the highest levels in the spring.
Concentrations ranged from 8.65 to 18.8 ppm (K), 17.54 to 72.98 ppm (Fe), 29.82 to 340.51
ppm (Mg) and 0 (below analytical detection) to 1.51 ppm (Co). The concentrations of Cu, Pb
and Ni were negligible at all study sites.
In summary, the concentrations of metals, including common ions such as Na and Mg, were
noticeably elevated along the western corner of the lake at El-Taween, El Fayroos and the
Bridge, but declined in the northeastern corner at El-Forsan and the eastern beach clubs.
High concentrations of Fe at the El-Taween and Bridge areas are likely due to the disposal of
raw sewage from nearby housing, and the impact of industrial waste residues in the Suez
Canal. In addition, high concentrations of Mn are most likely associated with agricultural
sewage, as well as wastewater from substantial industrial activities associated with ship
construction and maintenance.
4.2 Water quality

While average ocean salinity ranges from approximately 32 to 35 ppt, local salinity –
particularly near shore - can vary according to freshwater input, evaporation and
temperature. Low salinity (31.4 ppt) was recorded when the temperature was low (14.5°C)
in the winter. Salinity reached a maximum value (44.9 ppt) during the summer when
temperature was highest (31°C), (Table.2). Conductivity varied between 42 µmhos cm-1 in
the winter and 107 µmhos cm-1 in the summer. Freshwater discharged from sewage
treatment systems, untreated sources and nonpoint source runoff, enter the lake on the
western side, causing a salinity decline (31.4 – 32.2 ppt); the highest salinity (44.9 – 44.8 ppt)
was recorded on the eastern side. The pH values recorded in the lake tended to be
somewhat alkaline (7.9 – 8.9) in the summer. Dissolved oxygen in the surface water
fluctuated between 5.6 mg l-1 at El-Osra in the summer, to 10.2 mg l-1 at El Fayroos in the
winter. Nitrate (NO3-N) concentrations ranged from15.6 µg l-1 at El-Taween to 130.4 µg l-1 at
El-Osra. SiO3 – Si values fluctuated between 2.7 mg.l-1 at El Forsan site and 6.8 mg.l-1 at the
beach clubs. Its highest values occurred in spring and the lowest values in autumn. The
importance of silicate lies in its significance for the construction of the cell wall of diatoms.
This element is not a limiting factor for diatoms in the lake. Willem (1991) reported
dissolved silica content between 0.03 mg.l-1 and 0.20 mg.l-1as limiting level for diatom
growth. High level of PO4-P is attributed to eutrophication caused by waste and input of
drain terminating in the lake along the western side. It reached its maximum value 29.6 µg.l
at El Fayroos site in winter and its minimum value of 6.9 µg.l-1 at the beach clubs during
spring. The N:P ratio is an important water quality indicator which can significantly affect
phytoplankton taxonomic structure in a lake. In El-Temsah lake it was quite variable,
ranging from 1.07 at El Taween to 9.45 the beach clubs (Table 2). A low ratio of N:P
contributes to eutrophication and excessive blue-green algae which is often caused by
wastewater, particularly domestic, input (Knuuttila et al., 1994). In summary, the western
and northern parts of El-Temsah lake subjected higher levels of water pollution from
various domestic, industrial and agricultural sources, which result in less favorable water
and sediment quality conditions than are found in the eastern portions of the lake.
4.3 Phytoplankton analyses

Phytoplankton algal community and hydrographic parameters were investigated at the
same sites and during the same seasons. A total of One-hundred and two taxa have been
identified. Most of them belong to Chlorophyceae (8); Bacillariophyceae (56); Dinophyceae
(18); Cyanophyceae (20) were recorded. The result suggested the presence of phytoplankton
continuum in the area with dispersion of some species in certain microhabitat conditions.
The maximum occurrence of the total phytoplankton density (1788 & 1718 cells.l-1) was
recorded in winter at the El-Taween and El Fayrooz sites. The productivity was high for
Bacillariophyta (60.2 & 59.5 %) and Cyanophyceae (19.3 & 19.7 %) which may be attributed
to concentrations of NH4 – N, PO4 – P and NO3 – N, NO2 -N and relatively higher values of
salinity (33.8-32.2 ppt). Where the species richness (Alpha Diversity) was ranged between
6.586 during summer and 14.287 in winter while relative evenness (H') was fluctuated
between 1.684 in summer and 1.902 in winter while ( C ) value was measured as 0.015 in
winter and 0.033 in summer. Chlorophyll- a varied from 0.3 µg.l-1 at the El-Taween to 26
µg.l-1 at the El-Fayrooz sites. It exhibited high values during winter, while the lowest one
during summer. The N:P ratio in the El Temsah Lake ranged between 1.07 and 9.45. The
lowest value occurred due to eutrophication caused by wastewater and domestic input into
the lake. Abundance of phytoplankton species may be attributed to High PO4 lower water
temperature; NH4–N and NO3–N were relatively high.
Phytoplankton biomass in the El-Temsah water samples was higher during the winter; the
lowest values were observed during the summer. A total of 102 taxa were recorded,
including Chlorophyceae (8), Bacillariophyceae (56), Dinophyceae (18) and Cyanophyceae
(20). Species richness (Alpha Diversity) varied from 4 to 7% during the summer and 19 to
20% in the winter (Fig. 2). Maximum phytoplankton density (1788 cells l-1) was recorded in
the winter in the western side of the lake.
5. Risk Assessment using GIS

Image data extracted through remote sensing techniques were integrated with data collected
from topographic and morphologic maps in order to monitor urbanization along the El-
Temsah lake perimeter from 1968 to 2004. Shoreline changes were monitored using vector
data, analyzed by image segmentation and growing techniques. Vector data indicated
shrinking at the boundary of El-Temsah Lake due to human activities, including
construction of roads, buildings and recreation centers (Fig. 3). Regulations mandate that
development and urban projects constructed along the coastal zone of El-Temsah include
assessment studies to enhance the economic viability of the projects and to protect the
environment. The quantitative assessment of heavy metals in beach sediments is considered
an important tool to trace anthropogenic pollution that affects water quality (Guerrin et al.,
1990). A thick layer of sludge has accumulated in the near-shore lake sediments due to the
disposal of multi-source wastes. Historical chemical and biological data clearly indicate that
El-Temsah lake is suffering from significant human-related pressures, resulting in
substantial water and sediment contamination from heavy metals, petroleum hydrocarbons,
pesticides and general wastewater contaminants, thus making it a primary candidate for
periodic dredging. Seismicity data was plotted on a map overall study area and its
surrounding representing the magnitude variations (Fig. 4 A). Recent Epicentres
distribution pattern show widespread micro to moderate earthquakes (3 ≤ Mb ≥ 5.4) along
the western side, with scattered events on the eastern side, all of which could affect the
various settlement projects. Hazards assessment maps indicate a positive relationship
between the spatial distribution of heavy metals in beach sediments and biological
assessment parameters, suggesting the health of phytoplankton populations (and other
organisms through food-web linkages) is intimately connected to the degree of water
pollution in the El-Temsah coastal zone.
6. Conclusion
Heavy metal analyses were conducted to determine the concentrations of Fe, Pb, Ni, Co, Cu,
Mg, K and Na in beach sediments. The results show a temporal increase in levels from
summer to winter: 4-22 ppm to 27-71 ppm (Fe), 0.5-10 ppm to 91-677 ppm (Mg), 2-10 ppm to
15-27 ppm (K) and 9-29 ppm to 1650-1850 ppm (Na). The need for mitigation of higher
contaminant levels is indicated along the western border of the lake, including the Bridge,
El-Fayrooz and El-Taween, due to the disposal of irrigation drainage water. The El-Fayrooz
club area is apparently less polluted than the El-Taween and Bridge sites, probably because
it is further away from areas where agricultural sewage enters the lake. Chemicals and
biological analyses during the 2005-2006 study period indicate that the most appropriate
zones for activities such as swimming, fishing and general water-related recreation are
along the eastern and north eastern parts of the lake during the summer and spring. The El-
Taween club in the southwestern corner of the lake was also found to be one of the least
polluted areas in the autumn. Increased recreational and industrial activities in the areas
that currently have better water and sediment quality, however, could lead to accelerated
environmental degradation.
Concentrations of Mg and Na were highest in the winter and spring. These data indicate
that lake conditions in the winter are conducive to increased phytoplankton productivity, as
well as higher beach sediment pollution under acidic condition. In contrast, sediment
pollution and phytoplankton productivity are lower during the summer and the lake water
tends to be more alkaline. Abundance of Chlorophyceae, Bacillariophyceae and
Cyanophyceae increase from the summer to the winter, as do dissolved oxygen, nitrate and
(not surprisingly) chlorophyll a concentrations. Epicentres distribution pattern show
widespread micro to moderate (3 ≤ Mb ≥ 5.4) earthquakes around the western sides, and
scattered events along the eastern side of the lake as shows in Fig.4 A and B. Integrated GIS
was used to construct land use and hazards maps of the study area. The eastern portion of
the lake, because of lower pollution levels, is the most appropriate site for fishing, tourism,
urban planning and navigation. Fortunately, environmental pollution has shown a decline
over the last decade due to successive sediment dredging and improvements in water
purification systems.
Figure 1 (A) Location of El-Temsah Lake in the Suez Canal. (B) sample locations within El-
Temsah Lake .
Fig 1. (C) Landsat image for year 2004 of the Suez Canal District.
Fig 2. Enhanced Landsat Thematic Mapper images ETM-7 for year 2004 of the El-Temsah Lake
Fig 3. Concentrations of metals in beach sediments during the summer and winter
Fig 4. Phytoplankton species richness (Alpha Diversity) in El-Temsah lake during the
summer and winter
Fig 5. Locations of activities affecting the boundaries of El-Temsah lake

Fig 6. (A): Map showing distribution of recently local and regional recorded earthquakes
with magnitude (3.0 ≤ Mb ≥ 4.5) of the study area and its surroundings. (B): Earthquake
hazards assessment map of El-Temsah lake.
Magnitudes (Mb): d≤ 3 ( b) 3.1 -4-4 ; 4.5-5.4

Table 1. Heavy metal analyses result of water samples collected at El-Temsah Lake during
summer 2005-spring 2006
Table 2. Physicochemical parameters of water samples collected at El Temsah Lake during

summer 2005-Spring 2006
7. References
Abdel-Shafey, H.I. and Abdel-Sabour, A. (1995). Levels of heavy metals in the environment
of the river Nile and some Egyptian surface water German Egyptian conference,
Ismailia, 6-7, February, pp 36-53
APHA. (1989). Standard methods for the examination of water and wastewater, 16th edition.
American Public Health Association, Washington, DC.
Aboulela, H. (1994). A study on the tectonic activity and seismicity in the Suez Canal region.
M.Sc. Thesis, Suez Canal University, Faculty of Science, Ismailia, 133 pp.
Boney, A.D. (1989). Studies in Biology, No. 52, Phytoplankton, Edward Amold, London, 116 pp.
Dorgham, M.M. (1985). Distribution of phytoplankton in spring along the Suez Canal. Rapp.
Comm. Int. Mer. Medit., 29, pp 123-124.
Dorgham, M.M. (1990). The winter phytoplankton of the north Suez Canal, January 1990,
Rapp. Comm. Int. Mer Medit., 32, pp 121-125
El-Shazly, E.M., Abdel Hady, M.A., El-Ghawaby, M.A., El-Kassas, I.A. and El-Shazly, M.M.
(1974). Geology of Sinai Peninsula from ERTS-1 satellite images. Remote sensing
research project, Academy of Scientific Research and Technology, Cairo, Egypt.
El-Sherif, Z.M. and Ibrahim, A.M. (1993). Phytoplankton production, diversity and chlorophyll
a in the Suez Canal, Egypt. Bull. Nat. Oceanogr., & Fish., A.R.E., 19, pp 191-212.
El-Ibiary, M.G. (1981). Shallow subsurface geological and geophysical studies in Lake
Timsah, Suez Canal area, Egypt. M. Sc. Thesis submitted to Geology Department,
Faculty of Science, Tanta University, Egypt, 200 pp.
ENSN, Egyptian National Seismic Network. Periodical Bulletines Press of National Research
Institute of Astronomy and Geophysics, Helwan, Egypt.
ETPS. (1995). Environmental testing of pollution status in Lake Temsah, Abu-Attwa Water
Reuse Center Research and Training, Ismailia, Egypt, 356 pp.
Guerrin, F., Burgat-Sacaze, V. and Saqui-Sanes, P. (1990). Levels of heavy metals and
organochlorine pesticides of eyprinid fish reared four years in wastewater treated
pond. Bull. Environ. Contam. Toxicol., 44, pp 461-467.
Holmes, A. (1965). Principals of physical geology, 2nd Edition. ELBS and Nelson, London, 190 pp.
Humm, H. and Wicks, S. (1980). Introduction and guide to the marine blue-green algae.
Wiley- Inter-Science publication, London, 193 pp.
ISC, online at, http://www.isc.ac.uk/
Magurran, A. (1988). Ecological Diversity and its Measurement. Princeton University Press,
Princeton, 179 pp.
More, P. and Chapman, S. (1986). Methods in plant ecology, second edition, Blackwell
Scientific Publications.
Moss, B. (1998). Ecology of fresh waters, Man and Medium, Past to Future, 3rd edition.
Blackwell Science Ltd., London, 557 pp.
NEIS, online at, http://earthquake.usgs.gov/regional/neic/
Pielou, E. (1975). Ecological Diversity. A Willy-Inter-Science Publication, 165 pp.
Prescott, A. (1978). How to know the freshwater algae, 3rd edition, Wm. C. Brown Company
Publishers, Dubuque. 293 pp
Tomas C.R. (1997). Identifying Marine Phytoplankton. Academic Press, Har Court Brac &
Company.
Utermohl, H. (1958). Zur Vervolkmmimg derquantitativenphytoplankton methodic. Mitt.
Int. Verein. Limnol., 9: 1-39.
Van-Heurck H. (1986). Attreatise on Diatomaceae. William Wesley and Son, London.
Whittaker, G. (1972). Evolution and measurement of species diversity. Taxonomy, 21, pp 213-
251.
Willem, E. (1991). Planktonic diatoms-ecological review. Algological Studies, 62, pp 69-106.
Air Pollution Determination Using Remote Sensing Technique 71
X5
Air Pollution Determination

Using Remote Sensing Technique
H. S. Lim, M. Z. MatJafri, K. Abdullah and C. J. Wong
School of Physics, Universiti Sains Malaysia
11800 USM, Penang, Malaysia.
1. Introduction
Environmental pollution become more and more serious due to today’s development all
around the world. Environmental pollution is our concern nowadays because all the daily
activities are related to the environment. In order to ensure this situation is under control,
Malaysia’s government has already established a network throughout Malaysia to monitor
this situation. Air pollution is one of the most important environmental problems, which
concentrates mostly in cities. Air pollution in Asian cities has grown with the progressing
industrialization and urbanization. Air quality standards often refer to respirable suspended
particulate matter (PM), being aerosols with a diameter smaller than 10 µm (PM10) (UNEP).
Fine particles are of the greatest concern since they are capable of being easily transported
over long distances on currents of air. Also, fine particles may be drawn into the respiratory
airways where they may adversely affect health. Recently, the attention of scientists has
been drawn towards studying the PM2.5 fraction and even smaller particles, which can
penetrate the very deepest parts of the lung (Environmental Protection Service Tameside
MBC Council Offices, 2008). The effects of environmental pollution in the study area are
shown in Figure 1. The objective of air quality monitoring is to obtain an estimate of
pollutant (total suspended particulates TSP) concentrations (Ung, et al., 2001a, 2001b). First,
compared to atmospheric gases, aerosol is highly inhomogeneous and variable; that is,
aerosol observations have to be global and continuous. Second, the available accuracy of
aerosol characterization is often not sufficient. For instance, in situ measurements
traditionally considered as the most reliable observations are inappropriate for global
monitoring of aerosol radiative forcing parameters and usually do not characterize the
aerosol in the total atmospheric column (Dubovik, et al., 2002).
Remote sensing has been widely used for environmental application such as for air quality
and water quality studies. But the atmosphere affects satellite images of the Earth’s surface
in the solar spectrum. So, the signal observe by the satellite sensor was the sum of the effects
from the ground and atmosphere. Tropospheric aerosols act to significantly alter the Earth’s
radiation budget, but quantification of the change in radiation is difficult because
atmospheric aerosol distributions vary greatly in type, size, space and time (Penner, et al.
2002). Surface reflectance is a key to the retrieval of atmospheric components from remotely
sensed data. Optical atmospheric effects may influence the signal measured by a remote
sensor in two ways: radiometrically and geometrically. This means that they can modify the
signal’s intensity through scattering or absorption processes and its direction by refraction
(Sifakis and Deschamps, 1992).
The problem of particulate pollution in the atmosphere has attracted a new interest with the
recent scientific evidence of the ill-health effects of small particles. Aerosol optical thickness
in the visible (or atmospheric turbidity), which is defined as the linear integral of the
extinction coefficient due to small airborne particles, can be considered as an overall air
pollution indicator in urban areas (Sifakis, et al., 1998). Air pollution has long been a
problem in the industrial nations of the West. It has now become an increasing source of
environmental degradation in the developing nations of East Asia. The lack of detailed
knowledge of the optical properties of aerosols results in aerosol being one of the largest
uncertainties in climate forcing assessments. Monitoring of atmospheric aerosol is a
fundamentally difficult problem. The problem of particulate pollution in the atmosphere has
attracted a new interest with the recent scientific evidence of the ill-health effects of small
particles. Air pollution is one of the most important environmental problems, which
concentrates mostly in cities. Aerosols are liquid and solid particles suspended in the air
from natural or man-made sources (Kaufman, et al., 1997).
The main objective of the present study is to test the performance of our proposed algorithm
for mapping PM10 using Landsat satellite images. In situ measurements were needed for
algorithm calibration. We used a DustTrak Aerosol Monitor 8520 to collect the in situ data.
We collected the PM10 data simultaneously during the satellite Landsat overpass the study
area. An algorithm was developed to determine the PM10 concentration on the earth
surface. The efficiency of the proposed algorithm was determined based on the correlation
coefficient (R) and root-mean-squares deviation, RMS. Finally, the PM10 map was generated
using the proposed algorithm. In addition, the PM10 map was also geometrically corrected
and colour-coded for visual interpretation.
Before
After
(a)
Source: vincentchow, 2009
Before
Source: The Digital Awakening-Haze in Penang

After
(b)
Source: The Digital Awakening-Haze in Penang Update - 10.48am
Fig. 1. Air pollution was found at (a) Kuching and (b) Penang, Malaysia
2. Remote Sensing
Remote sensing is a technique for collecting information about the earth without taking a
physical sample of the earth’s surface or touching the surface using sensors placed on a
platform at a distance from it. A sensor is used to measure the energy reflected from the
earth. This information can be displayed as a digital image or as a photograph. Sensors can
be mounted on a satellite orbiting the earth, or on a plane or other airborne structure.
Because of this energy requirement, passive solar sensors can only capture data during
daylight hours. The major applications of remote sensing include environmental pollution,
land cover/use mapping, urban planning, and earth management. We have to understand
the basic concept of electromagnetic waves well enough for applying to the remote sensing
techniques in our studies. We classify electromagnetic energy by its wavelength. This
electromagnetic radiation gives an energy source to illuminate the target except the sensed
energy that is being emitted by the target (Figure 2).
Fig. 2. Remote sensing instrument (Jensen, 2006)
There are two basic types of sensors: passive and active sensors (Figure 3). Passive remote
sensors detect reflected energy from the sun back to the sensor; they do not emit energy
itself. But active sensors can emit energy or provide its own source of energy and detect the
reflected energy back from the target.
(a)
(b)
Fig. 3. (a) Passive sensor and (b) Active sensor (Fundamentals of Remote Sensing -
http://www.ccrs.nrcan.gc.ca/resource/tutor/fundam/pdf/fundamentals_e.pdf.)
3. Study Area
The study area is the Penang Island, Malaysia, located within latitudes 5o 9’ N to 5o 33’ N
and longitudes 100o 09’ E to 100o 30’ E. The map of the study area is shown in Figure 4.
Penang Island is located in equatorial region and enjoys a warm equatorial weather the
whole year. Therefore, it is impossible to get the 100 % cloud free satellite image over
Penang Island. But, the satellite image chosen is less than 10 % of cloud coverage over the
study area. Penang Island located on the northwest coast of Peninsular Malaysia.
Penang is one of the 13 states of the Malaysia and the second smallest state in Malaysia after
Perlis. The state is geographically divided into two different entities - Penang Island (or
“Pulau Pinang” in Malay Language) and a portion of mainland called “Seberang Perai” in
Malay Language. Penang Island is an island of 293 square kilometres located in the Straits of
Malacca and “Seberang Perai” is a narrow hinterland of 753 square kilometres (Penang-
Wikipedia, 2009). The island and the mainland are linked by the 13.5 km long Penang
Bridge and ferry.
Penang Island is predominantly hilly terrain, the highest point being Western Hill (part of
Penang Hill) at 830 metres above sea level. The terrain consists of coastal plains, hills and
mountains. The coastal plains are narrow, the most extensive of which is in the northeast
which forms a triangular promontory where George Town, the state capital, is situated. The
topography of “Seberang Perai” is mostly flat. Butterworth, the main town in “Seberang
Perai”, lies along the “Perai” River estuary and faces George Town at a distance of 3 km (2
miles) across the channel to the east (Penang-Wikipedia, 2009).
The Penang Island climate is tropical, and it is hot and humid throughout the year. with the
average mean daily temperature of about 27oC and mean daily maximum and minimum
temperature ranging between 31.4oC and 23.5oC respectively. However, the individual
extremes are 35.7oC and 23.5oC respectively. The mean daily humidity varies between 60.9%
and 96.8%. The average annual rainfall is about 267 cm and can be as high as 624 cm
(Fauziah, et al, 2006).
Fig. 4. The study area

4. Satellite Remote Sensing Data

For our research in USM, we use satellite images from passive sensors in our analysed.
Images acquired by Landsat TM 5. On March 1, 1984, NASA launched Landsat 5 were used
in this study, the agency’s last originally mandated Landsat satellite. Landsat 5 was
designed and built at the same time as Landsat 4 and carried the same payload: the
Multispectral Scanner System (MSS) and the Thematic Mapper (TM) instruments.
Fig. 5. Landsat Satellite TM
5. Algorithm Model
The atmospheric reflectance due to molecule, Rr, is given by (Liu, et al., 1996)
τ r Pr (Θ)
Rr =
4µ s µ v (1)
Where
τr = Rayleigh optical thickness

Pr( Θ ) = Rayleigh scattering phase function
µv = Cosine of viewing angle
µs = Cosine of solar zenith angle
W e assume that the atmospheric reflectance due to particle, Ra, is also linear with the τa
[King, et al., (1999) and Fukushima, et al., (2000)]. This assumption is valid because Liu, et
al., (1996) also found the linear relationship between both aerosol and molecule scattering.
τ a Pa (Θ)
Ra =
4µ s µ v (2)
where
τa = Aerosol optical thickness

Pa( Θ ) = Aerosol scattering phase function
Atmospheric reflectance is the sum of the particle reflectance and molecule reflectance, Ratm,
(Vermote, et al., (1997).
Ratm=Ra+Rr (3)
Where
Ratm=atmospheric reflectance
Ra=particle reflectance
Rr=molecule reflectance
τ P (Θ) τ r Pr (Θ) 
R atm =  a a + 
 4 µ s µ v 4µ s µ v 
1
R atm = [τ a Pa (Θ) + τ r Pr (Θ)]
4µ s µ v (4)
The optical depth is given by Camagni and Sandroni, (1983), as in equation (5). From the
equation, we rewrite the optical depth for particle and molecule as equation (6)
τ = σρs (5)
Where
τ = optical depth
σ = absorption
s = finite path
τ = τ a +τ r (Camagni and Sandroni, 1983)

τ r = σ r ρr s (6a)
τ p =σ pρps
(6b)
Equations (6) are substituted into equation (4). The result was extended to a three bands
algorithm as equation (7) Form the equation; we found that PM10 was linearly related to the
reflectance for band 1 and band 2. This algorithm was generated based on the linear
relationship between τ and reflectance. Retalis et al., (2003), also found that PM10 was
linearly related to τ and the correlation coefficient for linear was better that exponential in
their study (overall). This means that reflectance was linear with PM10. In order to simplify
the data processing, the air quality concentration was used in our analysis instead of using
density, ρ, values.
1
Ratm = [σ a ρ a sPa (Θ) + σ r ρ r sPr (Θ)]
4µ s µ v
s
R atm = [σ a ρ a Pa (Θ) + σ r ρ r Pr (Θ)]
4µ s µ v
s
R atm (λ1 ) = [σ a (λ1 ) PPa (Θ, λ1 ) + σ r (λ1 )GPr (Θ, λ1 )]
4µ s µ v
s
Ratm (λ 2 ) = [σ a (λ 2 ) PPa (Θ, λ 2 ) + σ r (λ 2 )GPr (Θ, λ 2 )]
4µ s µ v
P = a 0 R atm (λ1 ) + a1 R atm (λ 2 ) (7)
Where
P = Particle concentration (PM10)

G = Molecule concentration
Ratmi = Atmospheric reflectance, i = 1 and 2 are the band number
aj = algorithm coefficients, j = 0, 1, 2, … are then empirically determined.
6. Data Analysis and Results

Remote sensing satellite detectors exhibit linear response to incoming radiance, whether
from the Earth’s surface radiance or internal calibration sources. This response is quantized
into 8-bit values that represent brightness values commonly called Digital Numbers (DN).
To convert the calibrated digital numbers to at-aperture radiance, rescaling gains and biases
are created from the known dynamic range limits of the instrument.
Radiance, L (λ) = Bias (λ) + [Gain (λ) x DN (λ)] (8)
where
λ = band number.
L is the radiance expressed in Wm-2 sr-1µm-1.
The spectral radiance, as calculated above, can be converted to at sensor reflectance values.
πL(λ )d 2
ρ =
*
E0 (λ ) cos θ (9)
Where
ρ* = Sensor Reflectance values

L (λ) = Apparent At-Sensor Radiance (Wm-2 sr-1µm-1)
d = Earth-Sun distance in astronomical units
= {1.0-0.016729cos [0.9856(D-4)]} where (D = day of the year)
E0 (λ) = mean solar exoatmospheric irradiance (Wm-2 µm-1)
θ = solar Zenith angle (degrees)
The rescaling gain and offset values for Landsat TM 5 used in this paper are listed
in Table 1 (Chander, et al. 2007) and Solar Exoatmospheric spectral irradiances are given in
Table 2 (Chander and Markham, 2003).
Landsat TM 5
Rescaling Gain and Bias
Processing Mar 1, 1984 - May May 5, 2003 - Apr Apr 2, 2007 – Present
Date 4, 2003 1, 2007
Acquisition Mar 1, 1984 - May May 5, 2003 - Apr Mar 1, 1984 - Dec Jan 1, 1992 -
Date 4, 2003 1, 2007 31, 1991 Present
Band Gain Bias Gain Bias Gain Bias Gain Bias
1 0.602431 -1.52 0.762824 -1.52 0.668706 -1.52 0.762824 -1.52
2 1.175100 -2.84 1.442510 -2.84 1.317020 -2.84 1.442510 -2.84
3 0.805765 -1.17 1.039880 -1.17 1.039880 -1.17 1.039880 -1.17
4 0.814549 -1.51 0.872588 -1.51 0.872588 -1.51 0.872588 -1.51
5 0.108078 -0.37 0.119882 -0.37 0.119882 -0.37 0.119882 -0.37
6 0.055158 1.2378 0.055158 1.2378 0.055158 1.2378 0.055158 1.2378
7 0.056980 -0.15 0.065294 -0.15 0.065294 -0.15 0.065294 -0.15
Table 1. Rescaling gains and biases used for the conversion of calibrated digital numbers to
spectral radiance for Landsat TM 5
Unit: ESUN = Wm-2µm-1

Band Landsat TM 5
1 1957
2 1826
3 1554
4 1036
5 215.0
7 80.67
Table 2. Solar Exoatmospheric spectral irradiances in Wm-2 µm-1 for Landsat TM 5
Landsat TM satellite data set was selected corresponding to the ground truth measurements
of the pollution levels. The PCI Geomatica version 10.1 image processing software was used
in all the analyses. The Landsat TM 5 satellite images were acquired on 15th February 2001
(Figure 6), 17th January 2002 (Figure 7), 6th March 2002 (Figure 8) and 5th February 2003
(Figure 9).
Raw digital satellite images usually contain geometric distortion and cannot be used directly
as a map. Some sources of distortion are variation in the altitude, attitude and velocity of the
sensor. Other sources are panoramic distortion, earth curvature, atmospheric refraction and
relief displacement. So, to correct the images, we have to do geometric correction. After
applying the correction, the digital data can then be used for other processing steps
(Anderson, et al. 1976). Image rectification was performed by using a second order
polynomial transformation equation. The images were geometrically corrected by using a
nearest neighbour resampling technique. Sample locations were then identified on these
geocoded images. Regression technique was employed to calibrate the algorithm using the
satellite multispectral signals.
Fig. 6. Raw Landsat TM satellite image of 15th February 2001

Fig. 7. Raw Landsat TM satellite image of 17th January 2002
Fig. 8. Raw Landsat TM satellite image of 6th March 2002

Fig. 9. Raw Landsat TM satellite image of 5th February 2003
It should be noted that the relfectance values at the top of atmospherr is the sum of the
surface reflectance and atmospheric relfectance. The signals measured in each of these
visible bands represent a combination of surface and atmospheric effects, usually in
different proportions depending on the condition of the atmosphere. Therefore, it is
required to determine the surface contribution from the total reflectance received at the
sensor. In this study, we extracted the surface reflectance from mid-infrared band because
the surface reflectance at various bands across the solar spectrum are correlated to each
other to some extended. The surface reflectances of dark targets in the blue and red bands
were estimated using the measurements in the mid-infrared band (Quaidrari and Vermote,
1999). Over a simple black target, the observed atmospheric reflectance is the sum of
reflectance of aerosols and Rayleigh contributions (Equation 10). This simplification,
however, is not valid at short wavelengths (less than 0.45 pm) or large sun and view zenith
angles (Vermote and Roger, 1996). In this study, a simple form of the equation was used in
this study (Equation 11). This equation also used by other research in their study (Popp,
2004).
Rs – TRr = Ratm (10)

Rs – Rr = Ratm (11)
where:
Rs = reflectance recorded by satellite sensor

Rr = reflectance from surface references
Ratm = reflectance from atmospheric components (aerosols and molecules)
T = transmittance
It should be noted that the relfectance values at the top of atmosphere was the sum of the
surface reflectance and atmospheric relfectance. In this study, we used ATCOR2 image
correction software in the PCI Geomatica 9.1 image processing software for creating a
surface reflectance image. And then the reflectance measured from the satellite [reflectance
at the top of atmospheric, ρ(TOA)] was subtracted by the amount given by the surface
reflectance to obtain the atmospheric reflectance. And then the atmospheric reflectance was
related to the PM10 using the regression algorithm analysis (Equation 7). In this study,
Landsat TM signals were used as independent variables in our calibration regression
analyses. The atmospheric reflectances for each band corresponding to the ground-truth
locations were determined. The atmospheric reflectance were determined for each band
using different window sizes, such as, 1 by 1, 3 by 3, 5 by 5, 7 by 7, 9 by 9 and 11 by 11. In
this study, the atmospheric reflectance values extracted using the window size of 3 by 3 was
used due to the higher correlation coefficient (R) with the ground-truth data. The extracted
atmospheric reflectance values were regressed with their respective ground -truth data and
the proposed algorithm to obtain the regression coefficients. PM10 maps for all the images
were then generated using the proposed calibrated algorithm and filtered by using a 3 x 3
pixel smoothing filter to remove random noise.
The data points were then regressed to obtain all the coefficients of equation (7). Then the
calibrated algorithm was used to estimate the PM10 concentrated values for each image. The
proposed model produced the correlation coefficient of 0.8 and root-mean-square error 16
µg/m3. The PM10 maps were generated using the proposed calibrated algorithm. The
generated PM10 map was colour-coded for visual interpretation [Landsat TM 5 - 15th
February 2001 (Figure 10), 17th January 2002 (Figure 11), 6th March 2002 (Figure 12) and 5th
February 2003 (Figure 13)]. Generally, the concentrations above industrial and urban areas
were higher compared to other areas.
Fig. 10. Map of PM10 around Penang Island, Malaysia-15/2/2001 (Blue < 40 μg/m3, Green =
(40-80) μg/m3, Yellow = (80-120) μg/m3, Orange = (120-160) μg/m3, Red = (>160) μg/m3
and Black = Water and cloud area)
7. Conclusion
This study indicates that Landsat TM satellite data can provide very useful information for
estimating and mapping air pollution. The proposed algorithm is considered superior based
on the values of the correlation coefficient, R=0.8 and root-mean-square error, RMS=16
µg/m3. This technique has been proved to be reliable and cost effective for such
environmental study. Further study will be carried out to verify the results.
8. Acknowledgements
This project was supported by the Ministry of Science, Technology and Innovation of
Malaysia under Grant 06-01-05-SF0298 “ Environmental Mapping Using Digital Camera
Imagery Taken From Autopilot Aircraft.“, and also supported by the Universiti Sains
Malaysia under short term grant “ Digital Elevation Models (DEMs) studies for air quality
retrieval from remote sensing data“. We would like to thank the technical staff who
participated in this project. Thanks are also extended to USM for support and
encouragement.
9. References
Camagni. P. & Sandroni, S. (1983). Optical Remote sensing of air pollution, Joint Research
Centre, Ispra, Italy, Elsevier Science Publishing Company Inc
Dubovik.; O., Holben, B.; Eck, T. F.; Smirnov, A.; Kaufman, Y. J.; King, M. D.; Tanre, D. &
Slutsker, I. (2002). Variability of absorption and optical properties of key aerosol
types observed in worldwide locations, American Meteorological Society, 590 – 608
Environmental Protection Service Tameside MBC Council Offices, (2008). [Online] available:
http://www.tameside.gov.uk/airquality
Fauziah, Ahmad; Ahmad Shukri Yahaya & Mohd Ahmadullah Farooqi. (2006),
Characterization and Geotechnical Properties of Penang Residual Soils with
Emphasis on Landslides, American Journal of Environmental Sciences 2 (4): 121-
128
Fukushima, H.; Toratani, M.; Yamamiya, S. & Mitomi, Y. (2000). Atmospheric correction
algorithm for ADEOS/OCTS acean color data: performance comparison based on
ship and buoy measurements. Adv. Space Res, Vol. 25, No. 5, 1015-1024
Fundamentals of Remote Sensing, [Online] available:
http://www.ccrs.nrcan.gc.ca/resource/tutor/fundam/pdf/fundamentals_e.pdf
Jensen, J. R. (2006), Remote Sensing of the Environment: An Earth Resource Perspective, 2nd
Ed., Upper Saddle River, NJ: Prentice Hall. [Online] available:,
http://www.cas.sc.edu/geog/Rsbook/Lectures/Rse/index.html
Kaufman,; Y. J., Tanr, D. C.; Gordon, H. R.; Nakajima, T.; Lenoble, J.; Frouins, R.; Grassl, H.;
Herman, B. M..; King, M. D. & Teillet, P. M. (1997). Passive remote sensing of
tropospheric aerosol and atmospheric correction for the aerosol effect. Journal of
Geophysical Research, 102 (D14):16,815-16,830
Liu, C. H.; Chen, A. J. ^ Liu, G. R. (1996). An image-based retrieval algorithm of aerosol
characteristics and surface reflectance for satellite images, International Journal Of
Remote Sensing, 17 (17), 3477-3500
King, M. D.; Kaufman, Y. J.; Tanre, D. & Nakajima, T. (1999). Remote sensing of tropospheric
aerosold form space: past, present and future, Bulletin of the American
Meteorological society, 2229-2259
Penner, J. E.; Zhang, S. Y.; Chin, M.; Chuang, C. C.; Feichter, J.; Feng, Y.; Geogdzhayev, I. V.;
Ginoux, P.; Herzog, M.; Higurashi, A.; Koch, D.; Land, C.; Lohmann, U.;
Mishchenko, M.; Nakajima, T.; Pitari, G.; Soden, B.; Tegen, I. & Stowe, L. (2002). A
Comparison of Model And Satellite-Derived Optical Depth And Reflectivity.
[Online} available: http://data.engin.umich.edu/Penner/paper3.pdf
Popp, C.; Schläpfer, D.; Bojinski, S.; Schaepman, M. & Itten, K. I. (2004). Evaluation of
Aerosol Mapping Methods using AVIRIS Imagery. R. Green (Editor), 13th Annual
JPL Airborne Earth Science Workshop. JPL Publications, March 2004, Pasadena,
CA, 10
Quaidrari, H. dan Vermote, E. F. (1999). Operational atmospheric correction of Landsat TM
data, Remote Sensing Environment, 70: 4-15.
Retalis, A.; Sifakis, N.; Grosso, N.; Paronis, D. & Sarigiannis, D. (2003). Aerosol optical
thickness retrieval from AVHRR images over the Athens urban area, [Online]
available:
http://sat2.space.noa.gr/rsensing/documents/IGARSS2003_AVHRR_Retalisetal_
web.pdf
UNEP Assessment Report, Part I: The South Asian Haze: Air Pollution, Ozone and Aerosols,
[Online] available:
http://www.rrcap.unep.org/issues/air/impactstudy/Part%20I.pdf
Ung, A.; Weber, C.; Perron, G.; Hirsch, J.; Kleinpeter, J.; Wald, L. & Ranchin, T. (2001a). Air
Pollution Mapping Over A City – Virtual Stations And Morphological Indicators.
Proceedings of 10th International Symposium “Transport and Air Pollution”
September 17 - 19, 2001 – Boulder, Colorado
Ung, A.; Wald, L.; Ranchin, T.; Weber, C.; Hirsch, J.; Perron, G. & Kleinpeter, J. (2001b).
Satellite data for Air Pollution Mapping Over A City- Virtual Stations, Proceeding
of the 21th EARSeL Symposium, Observing Our Environment From Space: New
Solutions For A New Millenium, Paris, France, 14 – 16 May 2001, Gerard Begni
Editor, A., A., Balkema, Lisse, Abingdon, Exton (PA), Tokyo, pp. 147 – 151, [Online]
available:
http://www-cenerg.cma.fr/Public/themes_de_recherche/teledetection/title_tele_
air/title_tele_air_pub/satellite_data_for_t
Sifakis, N. & Deschamps, P.Y. (1992). Mapping of air pollution using SPOT satellite data,
Photogrammetric Engineering & Remote Sensing, 58(10), 1433 – 1437
Sifakis, N. I.; Soulakellis, N. A. & Paronis, D. K. (1998). Quantitative mapping of air
pollution density using Earth observations: a new processing method and
application to an urban area, International Journal Remote Sensing, 19(17): 3289 –
3300
The Digital Awakening-Haze in Penang, (2009). [Online] available:
http://www.petertan.com/blog/2005/08/12/haze-in-penang/
The Digital Awakening--Haze in Penang Update - 10.48am, [Online] available:
http://www.petertan.com/blog/2005/08/13/haze-in-penang-update-1048am/
Vincentchow, (2009). Available Online:
http://www.vincentchow.net/289/haze-in-malaysia
Vermote, E. & Roger, J. C. (1996). Advances in the use of NOAA AVHRR data for land
application: Radiative transfer modeling for calibration and atmospheric correction,
Kluwer Academic Publishers, Dordrecht/Boston/London, 49-72
Vermote, E.; Tanre, D.; Deuze, J. L.; Herman, M. & Morcrette, J. J. (1997). 6S user guide
Version 2, Second Simulation of the satellite signal in the solar spectrum (6S),
[Online] available:
http://www.geog.tamu.edu/klein/geog661/handouts/6s/6smanv2.0_P1.pdf
Some Features of The Volume Component
of Radar Backscatter From Thick and Dry Snow Cover 93
6
X
Some Features of The Volume

Component of Radar Backscatter
From Thick and Dry Snow Cover
Boris Yurchak
University of Maryland, Baltimore County, GEST
USA
1. Introduction
Radar monitoring of thick snow cover in polar regions with optical thickness (a product of
the depth of the snow and the extinction coefficient) of order 1 or more from elevated above-
ground and space platforms is of great importance for registration and for understanding
glaciology processes caused by climatic change. Applicable to this issue, the volume
component of the backscatter coefficient has a notable contribution to the total backscatter
(Noveltis, 2005). Although the idea of the radar-cross section (RCS) term and its derivatives,
like the backscatter coefficient, aims to separate as far as possible the sensor (radar) and
target parameters, this distinction can rarely be fulfilled in the case of spatially extended
geophysical targets (SEGT), such as atmospheric clouds and rain, as well as the thick snow
cover that is the focus of this paper. Due to this feature, the analysis of backscatter from
SEGT strongly depends on the relationship between the technical and physical-geometrical
properties of the radar and target, respectively. The main parameters that govern the radar-
target configuration for snow sounding are wavelength, antenna characteristics, pulse
duration, sounding direction, extinction coefficient and the geometrical depth of the snow.
A correct assessment of the volume component of the backscatter coefficient and an
understanding of the realm of applicability of any backscattering model is possible only
when the size of the scattering volume within a snow slab is known. This parameter
depends on the factors mentioned above and should be distinguished for different
situations in the practice of radar sounding of snow-covered terrain by Synthetic Aperture
Radar (SAR), a scatterometer or an altimeter. This work attempts, at first, to determine the
radar-target configurations inherent to volume scattering estimations of the thick snow
cover under the different radar applications mentioned above. Next, we analyze the range of
applications of the incoherent approach for backscatter magnitude estimation, currently one
of the main techniques for snow characteristics assessment. The simplest incoherent
approach, based on the so-called “particle” or “discrete” approximation, leads to the
dependence of backscatter on the sixth moment of the particle size distribution function
(PSDF) and the mean amount of particles in the scattering volume (Siegert & Goldstein,
1951; Battan, 1959; Ulaby et al., 1982). For a medium with losses (such as a thick snow slab)
the modification of this approach is referred to as the semi-empirical model (Attema &
Ulaby, 1978; Ulaby et al., 1982; Ulaby et al., 1996). There is some evidence of less backscatter
occurring than expected by virtue of the conventional reflectivity factor and the backscatter
coefficient for incoherent scatter from homogeneous thick snow slab (e.g., Rott et al., 1993).
For dense media (mean distance between particles is less than the wavelength) there are
many references in the literature that the backscatter is determined by media
inhomogeneities (Naito & Atlas, 1967; Gossard & Srauch, 1983; Fung, 1994). To attempt to
evaluate the contributions of these inhomogeneities, the radiative transfer (RT) and the
dense media radiative transfer (DMRT) models were developed (e.g., Ulaby et al., 1982;
Fung, 1994; Tsang et al., 2007). A review of these models and their modifications provided
by Noveltis (2005) stated that in certain aspects, these approaches showed some successes.
However, these models have not the close-form solutions that make difficult to use them
and analyze results obtained. Kendra et al. (1998), based on experiments with artificial snow
of varying depths, concluded that both conventional and dense-medium radiative transfer
models fail to adequately explain the observed results.
Finding an appropriate and relatively simple approach to calculate the volume component
of backscatter to explain the observable deviations from the classical (incoherent) model is
therefore a relevant task. The majority of previous studies, focusing on measurements of a
dry snow at temperate latitudes, where the snow depth seldom exceeds 1 m and the
corresponding optical thickness is much less than 1, have led to the conclusion that such an
approach is not feasible because of the much weaker interaction of the electromagnetic
radiation of radar wavebands with ice particles within the snow compared with the
backscatter from the soil beneath. The current study focuses on the case of the Greenland ice
sheet, however, where the depth of snow significantly exceeds that found in temperate
latitudes. This favorable condition, in conjunction with the recent finding that the incoherent
approach is only a specific case of a more comprehensive description of electromagnetic
wave interactions with spatially-extended individual scatterers (Yurchak, 2009), provides a
reason for more detailed investigations of the possibilities of the semi-empirical model for
interpreting the observable features of radar backscatter from thick snow cover.
2. Condition of complete burial of the probing pulse into snow medium

To estimate the volume component of backscatter from a snow slab, it is first necessary to
understand whether the probing pulse is completely buried within the snow slab or
whether the illuminated volume is only part of the probing pulse volume. In weather radar
meteorology, applicable to rain and clouds, this problem is known as “partial (or
incomplete) beam filling,” with a corresponding factor included in the weather radar
equation (e.g., Clift, 1985). For sounding thick snow cover, this issue is practically not
discussed and is different compared with the sounding of meteorological targets. Complete
burial depends on the sounding configuration and snow slab depth. The main criterion for
complete burial is the location of the pulse scattering volume with angular size equal to the
angular antenna beam width and with spatial length equal to one half of the actual spatial
duration of the transmitted pulse into a snow layer. Thus, for complete burial, the angular
and radial sizes of the backscatter volume should be matched with the snow slab depth,
which is usually known only roughly for a particular geographical region. This condition
poses problems for practical applications. Nevertheless, to better understand possible
situations where complete burial is feasible, the requirements for the angular antenna beam
size and the probing pulse duration will be analyzed separately for different major
sounding configurations.
2.1. Condition of complete burial of the angular (transverse) size of the pulse volume
2.1.1. Oblique sounding, flat surface, plane wavefront
The scheme of this configuration is shown in Figure 2.1. Here, and everywhere below, the
snow slab is assumed to have limited depth and an unbounded horizontal extent.
φ 0.5
θ
R0
a b c
db
h θ
p g
Fig. 2.1. Scheme of sounding for flat surface and plane wavefront
Mathematically, the condition for complete burial in this case is:
db ≤ h (2.1)
where db is the burial depth of the advance point (p) of pulse edge (leading or trailing)
when the lagging point (c) reaches the surface and h is the snow depth. As follows from the
geometry of Figure 2.1:
R ϕ
d b = cg = pc × sin ∠cpg ; ∠cpg = θ ; pc ≈ 2 0 tg 0.5 , and hence:
cosθ 2
R0 ϕ (2.2)
db = 2 tg 0.5 sin θ
cosθ 2
For φ0.5<<1 one can state tg ϕ 0.5 ≈ ϕ 0.5 , and condition (2.1) has the form:
2 2
h 1 (2.3)
ϕ 0.5 ≤
R0 tgθ
1.25 −5
If h~10 m and R0=800 km, then ϕ 0.5 << 10 and for θ~400, for example, the burial
tgθ
condition (2.1) is satisfied at roughly φ0.5<<0.0010. Obviously, this condition can not be
fulfilled for any space-based radar system with a real aperture, but it is possible for the SAR
if one assumes that the effective synthetic aperture radar beam illuminating an element of
spatial resolution Δx on a flat surface can be equal to an extremely narrow, pencil-like beam
∆x
with ϕ 0.5 ~ . For example, RadarSAT-1 has Δx ~12.5 m with an orbit height of
R
approximately 800 km and thus, φ0.5~0.9*10-3 degrees. If one assumes that the typical size
of the main lobe of the conventional antenna pattern is of order ~10, the complete burial of
the transverse size of the pulse volume is possible for h ≥ ϕ 0.5tgθ ~ 2 ⋅ 10− 2 tgθ . This condition
R0
can definitely be fulfilled for airborne and above-surface elevated radars. The estimate
obtained above should be considered only a rough approximation, because at nadir
sounding (θ=0) and unbounded horizontal extent of snow slab, the wavefront can not be
considered planar for the assessment provided. A more precise estimate is given below.
2.1.2. Oblique sounding, flat surface, spherical wavefront

A real wavefront within a snow slab has a spherical shape. The impact of this shape on the
estimation of db as the look angle θ (Figure 2.1) decreases begins at the moment when the
line tangent to the spherical front at point p coincides with the horizontal line pg, Figure 2.2.
This situation takes place when θ becomes equal to the angle between pc and the tangent:
θ =γ (2.4)
Spherical Tangent
wavefront to point p
b c
π/2 β
θ γ=φ0.5/2
p g
Fig. 2.2. To an assessment of sphericity impact
As follows from geometrical sketch, the auxiliary angle β is equal to: β = π − ϕ 0.5 and, thus,
2
π ϕ
γ= − β = 0.5 . Therefore, the impact of sphericity should be taken into account only for
2 2
small look angles, when
ϕ 0.5 (2.5)
θ≤
2
ϕ 0.5
Thus, the estimate (2.3), provided in the previous subsection, is valid for θ > . An
2
estimate for configurations close to nadir sounding will be carried out in the next subsection.
2.1.3. Nadir sounding, flat surface spherical wavefront
φ0.5
R0
db
h
Fig. 2.3. Schematic of sounding in nadir direction
This configuration is shown schematically in Figure 2.3 and the corresponding condition for
complete burial is:
d b( flt ) ≤ h (2.6)
From the geometrical configuration and φ0.5<<1, it follows that:
  ϕ
  2 sin 2 0.5
( flt ) 1 4 ≈ 1 R ϕ2 (2.7)
db = R0  − 1 = R0
 ϕ 0.5  ϕ 0.5 8
0 0.5
 cos  cos
 2  2
This estimate can be accepted for elevated platforms (for example, airborne radars and those
mounted above the surface) but needs to be analyzed further for space platforms due to the
sphericity of the surface, as the corresponding area illuminated on the Earth’s surface by
conventional radar is large.
2.1.4. Nadir sounding, spherical surface, spherical wavefront

In the spherical surface approach, the depth of complete burial in altimeter mode is more
than that for a flat surface described above by an increment Δz in the center of the beam.
This effect is illustrated in Figure 2.4, and can be written as:
d b( sp ) = d b + ∆z ≤ h (2.8)
φ0.5
n0
q
R0
db
h
Δz y d b( sp )
RE
Fig. 2.4. Estimation of complete burial of an increment of pulse due to the sphericity of the
Earth
Based on the geometrical relations of the sketch, the following system of equations may be
composed
 y ϕ 0.5
 R + q = tg 2 (2.9)
 0
 R 2 = (R − q )2 + y 2
 E E
ϕ 0.5
where q = ∆z ⋅ cos , and y is the distance between the vertical axis and the beam
2
periphery touching the spherical surface. Solving this system, the increment due to the
sphericity is equal to:
 2 
ϕ  2ϕ  ϕ 
RE − R0tg 0.5 
2
1 + tg 0.5  R0tg 0.5  
(2.10)
2 1 ± 1 −  2  2  
q=
 ϕ   2 
2  2 ϕ 0.5 
1 + tg 0.5   R
 E − R tg  
 2   
0
2 
 
ϕ 0.5 RE
where RE is the Earth’s radius. Taking into account that << 1, > 1 , and that the
2 R0
minus sign should be taken before the square root for physical reasons, the above formulae
can be simplified significantly, yielding an increment equal to:
2 −1
1 R0  ϕ 0.5   ϕ  (2.11)
∆z ≈ R0    cos 0.5 
2 RE  2   2 
A numerical assessment under the parameters values listed above gives Δz~5 m. The
estimation of the parameter d b( sp ) can be represented in the form:
 
  2
( sp )  1  1 R0  ϕ 0.5  1 1  R  (2.12)
db = R0 −1 + R0   ≈ R0ϕ 02.5 1 + 0 
 ϕ  2 RE  2  ϕ 8  RE 
 cos cos 0.5
0.5

 2  2
This relationship coincides with the estimate obtained by Barrick (1972) for a spherical wave
increment over the spherical mean sea surface. For R0=800 km, RE=6400 km and ϕ 0.5 = 1 ,
0
( sp )
one calculates d b ≈45 m. This result means that the Earth’s sphericity increases the
complete burial parameter compared with the flat surface case (see 2.7) by a factor of
R
~( 1+ 0 ). For R0 ≈ 1 the complete burial parameter equals to 1.125, i.e., a 12.5% increase.
RE RE 8
Based on known assessments of penetration depth of 2-5 m for Ku-band (e.g., Davis, 1996),
the wave sphericity in practice does not allow for consideration of the complete burial of the
transversal size of the probing pulse for a space-based altimeter and scatterometer. This
sphericity causes the area illuminated on the surface to change as the probing pulse is
buried into a snow slab. It can be easily shown that the radius of the illuminated area
changes with the burial increment ΔR as rill ≈ 2 R0 ⋅ ∆R .
2.1.5. Wavefront flattening

The speed of an electromagnetic wave within a snow slab is lower than that in an air
by ε ′ , where ε’ is the real part of the dielectric permittivity of snow. For dry snow within
a density range of 0.2-0.5 g/cm3, parameter ε’ changes roughly from 1.35 to 1.95 (Tiuri et al.,
1984). Recently, a decrease in wave speed was also confirmed by direct measurements in
snow (Scott et al., 2006). Due to the sphericities of the Earth and the wavefront, the paths of
the wave for different rays into a snow slab for equal time are not the same that results from
the distortion of the initial spherical wave front, as demonstrated in Figure 2.5.
φ0.5
Snow
surface
a
b
u
g
Spherical Flattened real
wavefront wavefront
Fig. 2.5. Illustration of the flattening of a spherical wavefront within a snow slab
The peripheral ray (ray Ob, for example) has a different path into a snow slab as compared
with the central ray (Og), which that has a portion ag extending into the snow. This feature
causes the flattening of the spherical wavefront. Thus, the snow slab works like dielectric
lens antenna (e.g., Lo & Lee, 1993). Let us make a rough estimate of the impact of this
phenomenon on the increment Δz. For a spherical wavefront (dashed curve in Figure 2.5),
the following equation is valid: Ob=Oa+au+ug. Dividing both sides of this equation by
speed of the wave propagation, c, yields the time of the wavefront arrival at point g:
Ob Oa + au + ug
t g( sp ) = = .
c c
Due to delays in snow in part of ag=au+ug the equation for the wavefront arrival at point g
Oa au + ug
is: t g( flt ) = + , where v is the wave speed in snow. Because v<c, t g( flt ) > t g( sp ) ,
c v
providing proof of this flattening. To rate this phenomenon, let us suppose that the
flattening front crosses the vertical at point u. This condition implies that the wave arrival
Ob Oa au
times at points u and b are the same. This yields the equations: = + and
c c v
v
au = (Ob − Oa ) = v ag . Hence, the relative shortage (flattening) of the wavefront path in
c c
the center of the beam compared with propagation in the free space (air) is equal to:
au v 1
= = , and the corrected increment Δz (2.11) should be decreased by ε ′ ~1.2…1.4
ag c ε′
times.
Moreover, due to the increase in snow density with depth, the dielectric permittivity also
increases, additionally impacting the rays’ path configuration as takes place in a Luneburg
lens. Thus, the actual meaning of the burial parameter lies between the two estimates carried
out above:
d b( flt ) ≤ d b ≤ d b( sp ) (2.13)
Based on the above estimates, one can say that the surface sphericity practically contributes
a relatively small amount to the estimate of the complete burial condition of the transverse
dimension of the probing pulse, and consequently, the surface can be roughly considered as
flat for any radar configurations regarding volume component assessment.
2.1.6. Modes of volume scattering for spherical wave

By analogy to spherical wave scattering from a surface (e.g., Moore & Williams, 1957) it is
also reasonable to distinguish two modes of the volume component in the case of spherical
wave scattering, as shown in Figure 2.6. For simplicity, the bounds of the snow slab are
considered flat. Let us call the mode depth-limited (Figure 2.6a) if at the moment when the
central point of the spherical wavefront reaches the bottom of a snow slab or a penetration
depth (Dp), the peripheral point of the wavefront, crossing the surface, is still within the
beamwidth footprint on the surface. This condition takes place when:
2
h 1  ϕ 0.5  if h ≤ D , or (2.14a)
≤   p
R0 2  2 
D p 1  ϕ 0.5  2 otherwise (2.14b)
≤  
R0 2  2 
a) b)
φ0.5
R0 R0
φ0.5
h, (Dp) h, (Dp)
Fig. 2.6. Modes of the volume scattering; a) depth-limited, b) beamwidth-limited
Let us call the mode beamwidth-limited (Figure 2.6b) if the peripheral points of the spherical
wavefronts, crossing the surface, are beyond the beamwidth footprint on the surface for a
significant part of the scattering volume. This situation takes place when:
h if h ≤ D
ϕ 0.5 << 2 2 p
R0
or (2.15)
Dp
ϕ 0.5 << 2 2 otherwise
R0
Obviously, that the depth-limited mode is inherent to space based altimeter and
scatterometr and the beamwidth-limited mode is realized for SAR. If the “significant part”
in the definition of this mode is replaced with a more exact term, the “determined part,”
the condition (2.15) can be modified and written in the form:
h  ∆h  if h ≤ D
ϕ 0.5 ≤ 2 2 1 −  p
R0  h 
or (2.15a)
Dp  
ϕ 0.5 ≤ 2 2 1 − ∆h  otherwise
R0  D p 
∆h ∆h
where or is the relative part of the scattering volume in which the beamwidth-
h Dp
∆h
limited mode is fulfilled. For example, if 90% of the scattering volume ( =0.9) is under
h
h
the beamwidth-limited mode, the condition (2.15a) is ϕ 0.5 ≤ 2 0.2 for h ≤ D p .
R0
2.2. Condition of complete burial of radial size of the pulse volume

Suppose that some area of snow cover is illuminated by a radar located in orbit as shown in
Figure 2.7. For simplicity, this scenario is depicted with a plane wavefront and flat surface.
Radar
α φ0.5
R0 h0=cτo
n
R θ Δh
( p)
hsct
h hθ
Fig. 2.7. Assessment of the complete burial of radial size of the probing pulse
Let us denote the height of the orbit as R0, the distance from the radar to the center of the
area as R, the anglular width of the main lobe of the antenna pattern as φ0.5 and, the look
angle as α. In addition, we assume that the cross-section of the main lobe is the circular. The
illuminated area is assumed to be flat horizontally, and the bottom boundary surface of the
snow slab is also flat and parallel to the top surface. Thus, the incidence angle, θ is equal to
the look angle, α (θ=α).
The main scaling parameters determining radial propagation are:
1) The radial length of the pulse scattering volume, equal to half of its spatial extent in the
snow medium:
( p) vτ 0 (2.16)
hsct =
2
c
where v = is the wave propagation speed in snow, c is the wave propagation speed in
ε′
the air, ε ′ is the real part of the snow permittivity and τ0 is the duration of the probing
pulse;
2) A one-way path in the snow, where the incidence power is decreased by “e” times,
usually called the “penetration depth” in the literature
D p = k e−1 (2.17)
where ke is the extinction coefficient, which characterizes the attenuation properties of the
medium due to scattering and absorption. One important remark is necessary. Since radar
sounding of snow is often performed in off-nadir mode, and the main lobe of the antenna
pattern has a finite angular size, it should be
surface
 
L p = k e−1
θ
D ∗p = L p cosθ
D p = ke−,1↓
h
bottom
Fig. 2.8. Relationships between the penetration depth (Dp), penetration path (Lp) and
propagation depth ( D ∗p ) in a snow slab
underlined that the extinction coefficient and the “penetration depth” are measured along
the direction of the wave (ray) propagation and not only to the vertical. Therefore, calling
the term ∝ k e−1 as the “…depth” should be considered a little bit confusing. A more
appropriate name for this term, from a physical point of view, is the “penetration path” (Lp),
keeping in mind that the penetration depth (Dp) is its value in the vertical direction. The
depth reached by an electromagnetic wave propagating at an angle θ to the vertical and
attenuating by “e” times represents the vertical component of the penetration path, and can
∗
( )
be called the propagation depth D p , as illustrated in Figure 2.8. Thus, from this point
onward in this paper, the following definitions and relationships are used:
 
Penetration path: L p = ke−1 (2.18)
Propagation depth: D = L p cosθ
∗
p
(2.19)
Penetration depth: D p = k −1 (2.20)
e,↓
where k e ,↓ is the extinction coefficient in the vertical direction. In general, in

inhomogeneous medium, the penetration path is a function of distance and
direction L p = f (R, θ ,ψ ) , where ψ is the azimuthal angle. For homogenous medium
( p)
D p = L p . We consider the probing pulse to be short if hsct ≤ L p ; otherwise it is considered
long. Obviously, the mode of sounding with current wave (CW) always belongs to the long
probing pulse configuration. Also one should distinguish a case with fully scattering snow
slabs, when the geometrical snow depth is equal to or less than the propagation depth
( h ≤ D ∗p ). Otherwise, one has a case of partially (because the wave does not penetrate to the
bottom) scattering slow slab. The condition for complete burial, as it follows from Figure
2.7, is
( p)
hsct ≤ hθ − ∆h (2.21)
h ϕ tgθ tgθ
where hθ = is the slant snow depth and ∆h = 2 R0 tg 0.5 ≈ R0ϕ 0.5 .
cosθ 2 cosθ cosθ
All of the above conditions are summarized in Table 2.1.
# Condition Description Comments

1 ( p) Short/long probing pulse (P.P.) Complete burial is satisfied
hsct ≤ L p
for:
2 h ≤ D ∗p a) short P.P. and fully
Fully/partially scattering snow slab
or scattering snow slab;
hθ ≤ L p b) long P.P. and
3 ( p) Complete burial of the radial size of partially scattering
hsct ≤ hθ − ∆h
the probing pulse snow slab
Table 2.1. Conditions for complete burial of the radial size of the probing pulse
Concluding remarks:
If the conditions for complete burial of the radial size of the probing pulse are completed, it
can at least be said that one sample of the return signal is formed by the scattering volume
( p) (2.22)
Vsct = hsct Aill
where Aill is the illuminated base of the scattering volume. If, in addition, the conditions for
complete burial of the transverse size of the probing pulse (see previous subsection) are also
satisfied, that the scattering volume is determined by the entire pulse volume:
2
( p)  ϕ  (2.23)
Vsct = V p = hsct π  R0 0.5 
 2 
Otherwise, under ∆R ≤
1  ϕ 0,5  ,
R0  
2  2 
( p) (2.24)
Vsct (∆R) ≈ hsct 2πR0 ∆R
2.3. An assessment of the scattering volume under incomplete burial condition

If the parameters of the probing pulse and of the snow slab do not satisfy the conditions
presented in the table above, the radial size of the scattering volume is determined by the
geometry of the snow slab. Several practically important cases for practical application are
discussed below.
2.3.1. Flat surface, plane wavefront, long probing pulse and fully scattering snow slab
In this scenario, the illuminated area, Aill on the snow cover (Figure 2.1) is an ellipse, with
 ϕ 0.5 
the minor semi-axis equal to the radar beam cross-section radius rmin =  R  and major
 2 
semi-axis the same divided by the cosine of the incidence angle: rmjr =
rmin . Thus:
cos θ
2
 Rϕ 0.5  1
Aill = π   (2.25)
 2  cos θ
If one uses the height of the satellite orbit R0 = R cos θ , (2.23) transforms to:
2
R ϕ  1
Aill = π  0 0.5  3
(2.26)
 2  cos θ
The pattern of the scattering volume for this case is depicted in Figure 2.9, where the
scattering volume is bounded by the surface and bottom planes.
2rmin
θ
Ail
О b d0
V
V h
h a d
V
Fig. 2.9. An assessment of the scattering volume for a fully scattering snow slab
In this case, the size of the scattering volume is the sum of the volume of a circular cylinder
V0 with cylindrical element d0 and two similar volumes of the truncated cylinders V1 that
mutually add up to a completed circular cylinder with cylindrical element d1 (the bases of
all volumes are the same and equal to Aill cos θ ):
Vsct = V0 + 2V1 = Aill cosθ (d 0 + d1 ) (2.27)
Because d 0 + d1 = hθ , we have, taking expression (2.25) into account:
2
 Rϕ 0.5  (2.28)
Vsct = Aill h = πhθ  
 2 
Thus, the equivalent scattering volume is an elliptic cylinder with a base equal to the
illuminated area on a surface and with a height element equal to the slab depth (under the
assumption that the slab depth is less than the half of the spatial duration of the probe radar
pulse). This result was obtained by Matzler (1987).
The duration of the SAR probing pulse duration is equal to several tens of microseconds
(e.g., 42 μs for RadarSAT-1 and 37.1 μs for ERS-1). Due to frequency chirp, the compressed
probing pulse duration decreases by many times, resulting in a volume radial size equal to
only several meters (5-13 m for RadarSAT-1 and 9.7 m for ERS-1, for example). Numerical
data are provided based on Alaska Satellite Facility documents (“RadarSAT-1 Standard
Beam SAR Images”, 1999; and “ERS-1 and ERS-2 SAR Images”, 1996).
Snow surface φ0.5 aa’=Dp
a Scattering
volume
b bottom
a’
Snow slab
b’ bb’=Lp bottom
Fig.2.10. Cross-section of the scattering volume for a spherical surface and wavefront. The
position of the flattened wavefront at the moment when its central point (a’) reaches the
penetration depth is depicted with a convex dashed line
Because SAR images represent the backscatter pulse train from the entire path of a wave
into a snow slab for any look angle as a point on a surface of some mean brightness (return
power), the radial size of the scattering volume can be considered equal to either the slant
size of the snow layer or the penetration path (whichever is smaller). That is, although the
SAR pulse has a finite compressed spatial duration; it works like a long pulse due to the
absence of radial discrimination in the sense of conventional radar terminology.
2.3.2. Long probing pulse, spherical wave, spherical surface, partially scattering slab
For this case, the bottom of the scattering volume is the geometrical placement of points
located at a distance Lp from the snow surface along the family of rays within the solid angle
of the main antenna lobe, as shown in Figure 2.10. The defining feature of this configuration
is that the central and peripheral points of the spherical wavefront pass the length Lp at
different times, resulting in different radii of curvature for the top and bottom bounded
surfaces.
3. Volume component of the backscatter coefficient

3.1. Semi-empirical model for flat surface, plane wave and long probing pulse
Based on the considerations provided above, one can formulate an estimation of the volume
component of the backscatter within the framework of the so-called semi-empirical model
(Attema & Ulaby, 1978), as for the case of incomplete burial of the probing pulse with a flat
surface, plane wave and a long probing pulse, which better fits the conditions for sounding
of a thick snow slab with SAR.
The total radar backscatter from the illuminated area is composed of four components
(Fung, 1994):
σ t = σ as + σ s + σ g + σ gv (3.1)
where σas is the radar cross-section (RCS) due to backscatter from the air/snow (top)
interface, σs is the RCS due to backscatter from the snow volume, σg is the RCS due to
backscatter from the snow/ground (bottom) interface and σgv is the RCS due to rescattering
between ground and snow volume irregularities. As was summarized by Koskinen (2001),
only the volume component is essential for dry snow. Therefore the essential portion of the
total backscatter consists of the volume component and the contribution of the ground
beneath:
σt = σ s +σ g (3.2)
Obviously, the radar becomes sensitive to the properties of the snow only when σ s >> σ g .
In accordance with Ulaby et al. (1982), the general form for the backscatter coefficient from a
surface is:
σt (3.3)
σ t0 =
Aill
This notation presupposes that the illuminated areas on the slab top and on the ground
(bottom of the slab) are the same. This assumption applies for small angle divergences of the
radar beam, as is valid for strong directed antennae, low refraction on the air-snow interface
and a ground surface that is flat and parallel to the surface of the snow slab. Accordingly,
the volume component of the backscatter coefficient is:
σs (3.4)
σ s0 =
Aill
N
In the incoherent approach, σ s = ∑ σ i ; i.e., σs represents the total radar cross-section
i =1
(RCS) of N scatterers contributing backscatter from a volume Vsct = Aill h of a snow slab (see
subsection 2.3.1). Due to attenuation of the electromagnetic wave upon propagation within a
snow mass, the total RCS of the snow slab is equal to:
N
σ s = ∑σ iα i (3.5)
i =1
where the attenuation coefficient, inherent to the i-th particle with a distance of ξi from the
coordinate origin O, is equal to:
α i = exp(− 2k eξ i ) (3.6)
where ke is the extinction coefficient. The factor α takes into account the two-way distance of
forth and back wave propagation. The assessment of the scattering volume described above
relates only to the absolute value of the volume. In the case of wave directed propagation
and, consequently, directed attenuation, the summation should be performed along the
propagation axis, ξ. Because N>>1, the summation in (3.5) can be replaced by integration. As
was mentioned above, this integration should be carried out along the direction of wave
propagation, i.e., along the ξ axes. Small changes in the propagation direction at the
air/snow interface are ignored due to minor differences in the corresponding refraction
coefficients. Due to the random spatial distribution of scatterers within the scattering
volume, this discussion considers only the mean backscatter characteristics averaged over
several illuminated areas, with particles having independent spatial locations and RCSs.
N hθ hθ
σs = ∑ σ iα i =
i =1
∫ σ (ξ )α (ξ )dξ = ∫ σ (ξ ) α (ξ )dξ (3.7)
0 0
 2
where σ(ξ) is the running RCS (by the unit of a distance along axis ξ, dimension is  L  ) of
L
the scattering volume. Assuming that the attenuation properties of snow remain the same
along the propagation path (k e ≠ f (ξ )) , one can write:
α (ξ ) = exp(−2k eξ ) (3.8)
If the scatterers’ RCSs are independent of their locations within the scattering volume
(condition of incoherent approach), we get:
hθ
(3.9)
σs = σ ∫ α (ξ )dξ
0
where <σ> is the mean running RCS of the scattering volume.

Based on (3.4) and taking into account (3.9), the mean backscatter coefficient can be written
as:
σs σ hθ
(3.10)
σ s0 =
Aill
=
Aill ∫ α (ξ )dξ
0
Taking into account that hθ is the radial size of the scattering volume (along axis ξ ),
σ s∗ = σ ⋅ hθ is the mean total RCS of snow (ice) particles within the scattering volume (an
σ s*
oblique cylinder) while ignoring the attenuation and σv = is the corresponding
V
mean volume specific backscatter coefficient, we can transform the ratio σ to the form:
Aill
σ σ hθ σ hθ σs ∗
= ⋅ = cosθ = cosθ = σ v cosθ (3.11)

Aill Aill hθ Aill h V
Therefore, substituting (3.11) into (3.10), we can state:
hθ
σ s0 = σ v cosθ ∫ α (ξ )dξ (3.12)

0
Taking into account (3.8) and assuming the homogeneity of snow slab (k e ≠ f (ξ , θ ,ψ )) , we
can conduct the integration of (3.12) and finally arrive at:
1
σ s0 = σ v [1 − exp(− 2k e hθ )]⋅ cosθ (3.13)
2k e
This expression is the mathematical formulation of the semi-empirical model and has been
obtained by Attema & Ulaby (1978) and Ulaby et al. (1982). In this paper, this model will be
cited as the “A-U model.” For this case, the backscatter coefficient (3.13) depends only on
the specific volume backscatter of the snow medium and not on the pulse volume sizes. This
remarkable feature is due to (1) “overcomplete” burial of the pulse length into the snow
medium, and (2) the backscattering normalization factor Aill (see 3.4) is the basis of the
probing pulse.
The obvious imperfection of the model, as applied to sounding of thick snow, is the
assumption of a constant extinction coefficient within the snow slab. The problem can be
solved by designing an appropriate stratification model for the selected study area and
modeling the spatial distribution of the extinction coefficient (e.g., Drinkwater et al., 2001).
The next limitation of the model is due to wave sphericity. Under depth-limited mode
conditions (see section 2), the illuminated area changes, and the running RCS in (3.7) can not
be assumed to be statistically homogeneous within the scattering volume. The same is true
regarding the extinction coefficient in (3.8) as well. Thus, for this case, the A-U model in
form (3.13) should be used with care. For the beamwidth-limited mode, the form in (3.13)
can be used, taking in mind the “determined part” of the scattering volume (see comments
to 2.15a).
The ground component of the backscatter coefficient is localized by the bottom location and
can thus be expressed by its backscatter coefficient, taking into account the attenuation:
σ g0 = σ g0, max exp(− 2k e hθ ) (3.14)
where σ g0, max is the backscatter coefficient of the ground bottom surface governed only by
the surface properties and its orientation in regards to the incidence of radar illumination..
3.2. Effective snow depth

Given the expression
1   h 
L p cosθ 1 − exp − 2  = H eff (3.15)
2  
 L p cosθ 
(3.13) can be presented in the simple form:
σ s0 = σ v H eff (3.16)
where Heff is the effective depth of snow sounding (EDS); i.e., the depth such that the
backscatter from which occurs as if without attenuation, with backscatter equal to that
which would occur from a slab of larger real depth with attenuation due to absorption and
scattering.
hθ
Parameter k e hθ = = τ θ is the optical thickness of snow (SOT) along the wave
Lp
propagation direction. It is useful to express the backscatter coefficient through the
0 H eff
dimensionless EDS using the geometrical snow depth, h: H eff =
h . Given that
expression, equation (3.16) can be written in the form:
σ s0 = σ v hH eff0 = σ s0 H eff0 (3.17)

max
where
σ s0 = σv ⋅h (3.18)
max
is the maximal value of the backscatter coefficient for a given snow depth, h, ignoring both
energy losses due to attenuation and scattering and the angular dependence of the
0
scattering volume. The normalized effective depth of snow sounding (nEDS≡ H eff ) plays the
role of a correction factor, and is equal to:
1
0
H eff = [1 − exp(− 2τ θ )] (3.19)
2τ θ
And the main formula of the model (3.2) can be written as:
σ s0 = σ s0 max
⋅ H eff0 + σ g0 (3.20)
It is useful to find the dependence of the correction factor on the SOT magnitude and the
incidence angle. Let us consider two extreme cases:
a) “shallow” snow: τ θ << 1
1
0
H eff ≈ [1 − (1 − 2τ θ )] = 1 (3.21)
2τ θ
Given this relation,
σ t0 = σ v h + σ g0 (3.22)
meaning that the EDS is approximately equal to the geometrical snow depth. Nevertheless,
the small magnitude of the volume component compared with the backscatter from the
bottom layer ( σ v h < σ g ) makes this dependence difficult to use in practice. For example,
0
a snow layer with h<~1 m is practically transparent to electromagnetic irradiance of the C-

band.
b) “thick” snow: τ θ >> 1
1 L p cosθ 1 D ∗p (3.23)
H eff0 ≈ = =
2τ θ 2h 2 h
Given this relation, and in accordance with (3.18) and (3.20), one gets:
∗
σ t0 = σ s0 max
+ σ g0 (3.24)
1
where σ s
0 ∗
= σ v D ∗p (θ ) . Equation (3.24) shows that the backscatter depends only on
max 2
the penetration path (depth) under a constant incidence angle. In the “thick” snow regime,
no additional snow accumulation contributes to the total backscatter due to the saturation
effect. On the other hand, in this regime, the backscatter coefficient demonstrates an angular
dependence, as the angle of incidence affects the propagation depth ( D ∗p ). The sensitivity of
the backscatter coefficient to the changes in snow thickness takes place in the so-called
“intermediate” regime, when τ θ ~ 1. Field data gives 20-30 meters of the penetration depth
for the C-band (e.g., Hoen & Zebker, 2000), and yield values that can be considered the
“working” range for probable snow measurements in this wave band. Values of the nEDS
(3.19) can be assessed from the plot of the nEDS as a function of SOT, as shown in Figure 3.1.
1
0.9
normalized effective snow depth
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0.1 1 10
<- "shallow" "thick" ->
optical thickness of snow
Fig. 3.1. Plot of the normalized effective depth of snow sounding versus the snow optical
thickness
To evaluate the angular dependence of the nEDS, we consider the normalized Snow Depth
(nSD), which is the depth of snow normalized to the penetration path:
h (3.25)
nSD =
Lp
Comparing nSD with the definition of the SOT yields:
nSD = τ θ ⋅ cosθ (3.26)
Substituting (3.26) into (3.19), we have
1 cosθ   nSD  (3.27)

H eff0 = 1 − exp − 2 cosθ 
2 nSD   
A plot of (3.27) reduced to zero dB at θ=0 for different nSD is shown in Figure 3.2. As
follows from this plot , the angular dependence of the volume component of the backscatter
coefficient in a practically appropriate range of angles, 0-500 is rather weak and equal to
~2dB for the thick snow regime. For the intermediate regime, with nSD~1, its range of
variation is about 1 dB. Because this range is of the same order as the errors, it is difficult to
expect a notable angular dependence in practical measurements inherent to the intermediate
regime.
-2
nEDS*2*nSD, dB
-4
-6
-8
-10
0 20 40 60 80
incidence angle, degree
nSD>>1 nSD=1 nSD=0.5 nSD=0.2
Fig. 3.2. The angular dependence of the normalized effective depth of snow sounding for
the “thick” (nSD>>1) and “intermediate” regimes
On the other hand, a notable angular dependence of the backscatter coefficient indicates that
snow depth is greater than the penetration path. The intermediate regime represents
practical interest for snow measurements. Suppose, for instance, the linear dependence of
the geometrical snow depth on a horizontal distance S over the flat surface:
h( s ) = k ⋅ S (3.28)
where k is the proportionality factor. Given that relation and taking into account (3.18) -
(3.20), the normalized backscatter coefficient as a function of a distance is:
σ s0 (S ) 1   k ⋅ S  (3.29)
= cosθ 1 − exp − 2 ∗ 
σ s max 2
0
  D p ( S ,θ ) 

To specify the proportionality factor, let us assume that the geometrical snow depth reaches
D ∗p
the propagation depth value at a distance, for example, equal to S0, i.e., k = . Thus,
S0
σ s0 (S ) 1
= cosθ [1 − exp(− 2 S ′)] (3.30)
σ s0 2
max
0 2
normalized snow
coefficient, dB
-2 1.5
backscatter
normalized
depth
-4 1
-6 0.5
-8 0
0.1 0.4 0.7 1 1.3
normalized distance
n_BC_dB n_H
Fig. 3.3. Illustration of the expected relative spatial behavior of the backscatter coefficient
under a linear dependence of the snow depth on a distance
S
where S ′ = is the normalized distance. For the intermediate regime, we can ignore the
S0
angular dependence of the backscatter coefficient. Let us also assume that the specific
volume scattering is approximately the same for all points along that distance. The
2
corresponding plot of expression (3.27), additionally normalized by a factor of , is
cosθ
depicted in Figure 3.3:
2 σ s0 (S ) (3.31)
= 1 − exp(− 2 S ′)
cosθ σ s0
max
Thus, in accordance with the semi-empirical model, the expected profile of the backscatter
coefficient should have an exponential pattern for a transect taken over terrain with a linear
increase in snow depth within the intermediate regime condition.
4. Comparison the A-U model estimates and the experimental radar data on
the Greenland ice sheet
For accurate comparison of the model with any experimental data, the study area must
satisfy the “thick” snow condition to exclude the impact of the ground. Second, the snow
pack must contain primarily Rayleigh ice particles with sizes less than approximately one-
tenth of the wavelength. The most appropriate area matching this condition is the dry
snow zone of the Greenland ice sheet. This region is not impacted by melt-freezing
processes due to permanent sub-zero temperatures year-round. Therefore, the auspicious
conditions for creation the ice lenses, tubes and large size aggregates, which inherent, for
instance, to the percolation zone, are absent here. Additionally, the snow (firn) is up to
several tens of meters deep (e.g., Zwally & Brenner, 2001), corresponding to the
intermediate or even the thick snow regime. Due to known difficulties regarding in situ
measurements in Greenland, there are no examples of the comparison of backscatter with
snow depth in the literature for the intermediate or thick snow regimes.
Fig. 4.1. Transect across hill (straight line crossing the dark spot on the SAR signature) in the
north-eastern region of Greenland (left), and the corresponding backscatter (solid line,
convex downwards curve) and elevation (dotted line, convex upwards curve) profiles (right)
Nevertheless, a certain qualitative example of the sensitivity of backscatter to the snow

depth relative to the intermediate regime can be provided based on a topographic feature (a
hill) in the northeast region of Greenland between 810 and 820 N, as shown in Figure 4.1.
The counter-phase behavior of the backscatter coefficient and the elevation profiles plotted
above can be explained supposing that the snow depth is minimal on the hilltop and
increases proportional to the distance from the top of the hill down to its base. This
assumption closely matches the field and modeled data provided by Jaedicke et al. (2000).
For quantitative assessment of the semi-empirical model more regular terrain should be
chosen. A typical dry snow area is presented on the RadarSAT-1 SAR image of eastern
Greenland in Figure 4.2.
As is clearly seen, the magnitude of the backscatter coefficient is about -20 dB in the dry
snow zone. This assessment coincides with data in the known literature (e.g., Drinkwater et
al., 2001; Partington, 1998; Baumgartner et al., 1999; Forster et al., 1999) which also gives
values of -12…-20 dB. We can estimate the corresponding value using the A-U model. We
assume that the thick snow regime is valid, and that the snow slab consists primary of
particles of roughly the same sizes; i.e., it is a monodisperse medium.
For these conditions (see 3.24), ignoring backscatter from the bottom surface, we get:
Fig.4.2. Illustration of the backscatter coefficient values (left) range for an arbitrary transect
(straight line on RadarSAT-1 SAR signature, right) from the percolation to the dry snow
areas within the east margin of the Greenland Ice Sheet (directions are from left to right for
the plot of the backscatter profile and contra versa for the SAR signature; date of data
acquisition: November 2, 2007; data granule ID: R1_62609_SWB_271)
L p cos θ 1 cos θ
σ s0 = σ v = σv (4.1)
2 2 ke
Further calculations below aim to discover the specific volume backscatter, σv and the
extinction coefficient, ke. The estimates listed below will be conducted with the Rayleigh
approach. In accordance with Ulaby et al. (1986), this approach bounds the upper electric
2π ru
size of particle, ζ u = , with the inequality ζ u m  ≤ 0.5 , where m is the complex
λ
refractive index of the particle matter and ru is the corresponding upper size of a particle.
For ice particles with m ≈ 1.7776 observed with C-band radar (λ=5.6 cm), the maximal size
satisfying the Rayleigh approach is 2.6 mm. Following known relationships (e.g., Ulaby et
al, 1986), the specific volume scattering is:
1
σ v = Zs (4.2)
λ4
where
26 5 N
∑r
2
Zs = π K 6 (4.3)
i
V i =1
In this formulae
2
2 m 2 − 1 (4.4)
K = 2
m + 2
where m = n + jη is the refractive index of the particle matter and n and η are the real and
imaginary parts of the refractive index, respectively. The value of these components can be
derived from the value of the complex dielectric permittivity of a material ε = ε ′ + jε ′′ ,
where ε’ and ε” are the real and imaginary parts of the dielectric permittivity.
0.5
n = 0.5 ⋅  (ε ′)2 + (ε ′′)2 + ε ′  (4.5)
  
0.5
η = 0.5 ⋅  (ε ′)2 + (ε ′′)2 − ε ′  (4.6)
  
For monodisperse media ( ri=r ) equation (3.3) has the form:
2
Z s = 26 π 5 K n0 r 6 (4.7)
where n0 is the particle concentration.

The extinction coefficient can be calculated as a sum of the absorption (ka) and scattering (ks)
coefficients (Ulaby et al., 1986):
k e = k a + k s = n 0 [Qa + Qs ] (4.8)
where Qa and Qs are the absorption and scattering cross-sections respectively. These
parameters, under the Rayleigh approach, can be calculated by the following formulas:
λ2 3
Qa = ζ Im{− K } (4.9)
π
and
λ2 6 2
(4.10)
Qs = 2 ζ K
3π
In accordance with (Ulaby et al., 1982)
3ε ′′
Im{− K } ≈ (4.11)
(ε ′ + 2)2
The real part of the dielectric permittivity for ice, contained in the formulas above, is equal
to 3.15 and in practice does not depend on temperature or wavelength. The imaginary part
of the dielectric permittivity for ice is not constant with changes in temperature or the
illumination frequency. In accordance with Matzler (1987) it can be expressed through the
following empirical formula:
A
ε ′′ = + BF C (4.12)
F
where F is frequency in GHz and A, B and C are the empirical coefficients. For a
temperature of -150C, which is more appropriate for the dry snow case, the coefficients are
equal to: A=3.5*10-4, B=3.6*10-5, C=1.2. Now, the expressions obtained for the backscatter
coefficient calculation may be combined. Substituting (4.2) in (4.1), we get:
1 Z  (4.13)
σ s0 =  4 s  cosθ
2λ k
 e 
Taking into account (4.7) and (4.8)-(4.10), one can get:
2
1 26 π 5 K n0 r 6 (4.14)
σ s0 = cosθ
2λ4 λ 2
2λ 6 2  2
n 0  ζ 3 Im{− K } + ζ K 
 π 3π 
Simplifying this expression, we have:
1 cosθ (4.15)
σ s0 =
2 1 2
2
Im{− K } +
ζ3K 3
An important feature of this expression is its independence with regards to the particle
concentration, a reflection of the property inherent to the saturation regime mentioned
above in section III. Since the scattering volume is less than the spatial duration of the
probing pulse an increasing particle number is equivalent to an increasing particle
concentration. Thus, no additional amount of the snow over the snow pack with a thickness
greater than the penetration depth can cause a notable increase in the backscatter coefficient.
This circumstance is also very useful from a simulation point of view, since there is no
microstructure parameter that needs to be assumed. For the numerical calculation listed
below we assumed that the snow within the dry snow zone consists of snow grains that are
actually grains of ice that look like grains of rice, with a typical size of ~1 mm. Taking into
2
account the fact for ice: ε’=3.15 and ε’>> ε” (Tiuri et al., 1984; Matzler, 1987), K = 0.16 .
Given that information, equation (4.15) for an intermediate incidence angle of ~400
(cosθ=0.77) is equal to:
0.385 (4.16)
σ s0 =
6.25 2
Im{− K } +
ζ3 3
Taking into account (4.11) and (4.12), we get:
0.5775 (4.17)
σ s0 =
1.06  A C 
 + BF  + 1
ζ3 F 
The plot of (4.17) is depicted in Figure 4.3 for C-band radar with F=5.3 GHz (λ≈ 5.6 cm).
0
backscatter coefficient, dB
-5
-10
-15
-20
-25
-30
-35
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
particle radius, mm
Fig. 4.3. The dependence of the backscatter coefficient on particle size in accordance with
the A-U model for C-band radar and snow depth greater than the penetration depth
(“thick” snow regime)
It follows from this plot that the observable backscatter coefficient values of -12…-20 dB can
be caused by particles with radii of less than 0.35-0.17 mm. The current estimates are slightly
lower compared with the same made by Partington (1998), who found that the backscatter
coefficient would range from -20 dB for a mono-distribution of grain sizes with a mean
radius of 0.25 mm to -2 dB for a mean grain radius of 1 mm. Our results showed a 0.17 mm
mean radius for -20 dB and a 1 mm mean radius for -3 dB. The small discrepancy is
probably due to the temperature dependence of the imaginary part of the ice’s dielectric
permittivity. The grain sizes values assessed do not coincide with the field data. In
accordance with Table 4.1, the mean particle radius within a ~1 m depth surface layer of dry
snow is 0.5…2.0 mm. As seen in Figure 4.3., these values should produce a backscatter
coefficient of approximately from -8 dB to 0 dB. Thus, there is a discrepancy of ~12 dB
between the observable data and the values forecast by the semi-empirical model. Therefore,
the A-U model overestimates the backscatter coefficient significantly. It should be noted, in
addition, that the particle radius can only increases with the snow depth due to the depth-
dependent grain radius model (Alley et al., 1982):
C ⋅h (4.18)
r 2 (h) = r02 +
Ha
where r0 is the mean radius at the surface, C is the crystal growth rate and Ha is the mean
annual layer thickness. Given this model, various simulation experiments (e.g., Forster et al.,
1999; Drinkwater et al., 2001) and field measurements (e.g., Jezek et al., 1994; Woods, 1994;
Lytle & Jezek, 1994), the mean particle radius for the entire scattering volume is even greater
than the surface values and the aforementioned discrepancy becomes even more firmly
grounded.
# Mean ice particle radius, Source Location

mm
1 0.5-2.0 Benson, 1996 Greenland
2 1.0-1.5 Schytt, 1964 Spitsbergen
3 Less than 0.75 mm Higham and Craven, 1997 Antarctic
4 0.2-0.6 Woods, 1994 Greenland
5 0.1-0.7 Lytle and Jezek, 1994 Greenland
Table 4.1. Literature data on the mean ice particle size near the surface of dry snow cover
Thus, although the A-U model gives a qualitative assessment of the main features of
backscatter behavior from a snow slab, the result of quantitative comparisons with field data
does not match the theoretical predictions.
5. Enhanced semi-empirical model of the volume component of the

backscatter coefficient
The problem highlighted in the previous section can be resolved by considering the
statistical properties of the small-scale fluctuations in the particle concentration and its
scattering properties within the scattering volume. The appropriate method for doing so
that is the so-called “slice” approach, which was primarily suggested in weather radar
meteorology (Marshal & Hitchfeld, 1953; Smith, 1964). This approach exploits the known
radar feature in accordance with which the particles of a “cloud” located close to the front of
the incident radar wave are considered to be approximately at the same distance from the
radar and reflect incident electromagnetic wave almost coherently. One can consider that
these particles are embedded in a fictitious thin cylindrical volume (“slice”), whose base
coincides with the surface of spherical wave front and side-bounded by the main lobe of the
antenna pattern. Thus, the scattering volume into a snow slab can be represented as an
adjoining series of these slices, as illustrated in Figure 5.1. Each slice is much narrower than
the radar wavelength in the wave propagation direction (Δs<<λ).
Scattering
volume
ΔS
slice
Fig.5.1. Simplified slice approach scheme. The arbitrary position of a single slice is shown
on the cross section of the scattering volume
A slice’s radial size, Δs can actually be considered as a minimal spatial scale of the
backscatter property fluctuation, if this scale is much smaller than the wavelength. Using
this approach, it was shown by the author (Yurchak, 2009), that the specific volume
component of the backscatter from any spatial extended geophysical target (SEGT),
included the snow as well, can be presented as the incoherent sum of the radar cross-
sections of individual scatterers only if the number of particles in the slices (n) are
distributed in accordance with the Poisson law; i.e., the variance of the number of particles
(Var(n) ) is equal to the mean number ( n ). Otherwise, the “classical” specific volume
component should be corrected by a so-called deviation factor Y(ξa, χ):
σv = σv class
Y (ξ a , χ ) (5.1)
where
ξ a2 + χ
Y (ξ a , χ ) = (5.2)
ξ a2 + 1
and Var (n) is the Poisson index, Stdev(a ) is the variation coefficient of the
χ= ξa =
n a
≡ a p ≡ σ p , σp is the radar cross-section of an
particle radar equivalent length (PREL
individual particle). Formulas (5.1) and (5.2) reflect the fundamental physical principal that
the fluctuations of the medium parameters (inhomogeneities) are the cause of the
electromagnetic wave scattering (e.g., Atlas, 1964; Fabelinskii, 1968). The parameter ξa might
be expressed through the measurable parameters of the snow particle size distribution
function (PSDF), such as the variation coefficient of particle radius, ξ r =
Stdev(r ) , and the
r
skewness coefficient, Sk:
−3 ξ r−1 + 2 ⋅ Sk (5.3)
ξ a = 3ξ r 2
ξ r−3 + 3ξ r−1 + Sk
Following Shi et al. (1993), if the PSDF for snow can be presented by the lognormal
distribution, the parameters of relationship (5.3) are: ξ r = ( )
exp S r2 − 1 , and
( ) [ ( )]
Sk = exp S r2 − 1 ⋅ 2 + exp S r2 . Given these relations, Sk = ξ r3 + 3ξ r . Thus, the PREL variation
coefficient (5.3) depends, via ξr only on the geometric standard deviation, Sr. The
corresponding plot of the deviation factor (5.2) is depicted in Figure 5.2 for different values
of the Poisson index χ.
8
6
4
deviation factor, dB
χ=5
2
χ=2
0
χ=1
-2
χ=0.5
-4
χ=0.1
-6
χ=0
-8
-10
-12
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
particle size variation coefficient
Fig.5.2. Deviation factor versus particle radius variation coefficient for different deviations of
Poisson index χ from the Poisson law
As can be concluded from analysis of the plot, the deviation of ~12 dB emphasized in the
previous section can be explained by the homogeneous microstructure of the dry snow
zone, with approximately identical particle sizes (ξr<0.1) and negligible variations of the
particle concentration (χ<<1) within the scattering volume. As an example of data
supporting the current considerations, the results of comparative measurements of C-band
backscatter from firn and in situ measurements of statistical parameters of firn dielectric
permittivity for a study area in Antarctic (Zahnen et al., 2002) can be taken into account. In
this study, dielectric permittivity was considered to be a measure of the heterogeneity of the
snow/firn pack. A notable positive correlation was observed between backscattering signal
strength and the standard deviation of the dielectric constant collected along 4 and 15
meters depth interval of several drills within a dry firn area. At the same time, a negative
correlation takes place for a mean dielectric constant. Because the dielectric constant is
approximately linearly proportional to the density of dry snow (e.g., Tiuri et al., 1984), and
therefore linearly proportional to the number of ice particles per unit volume as well, these
correlations match the behavior of the deviation factor (5.2) on the Poisson index. The index
and, hence, scattering increase with increasing variations in particle number and decrease
with increasing n . Of course, more comprehensive field data are needed to confirm this
hypothesis.
The considerations provided above are related to the simplest case of complete burial of the
probing pulse into a snow slab in the beamwidth-limited mode. Obviously, that the result is
still the same for incomplete burial of the radial size of the scattering volume. The transverse
size change of a slice within the scattering volume in the depth-limited mode results in an
additional gradient in particle number among slices, and the contribution of this additional
gradient to the backscatter should be analyzed in the future.
6. Normalized snow depth

6.1. Derivation of the normalized snow depth from the backscatter coefficient based
on the enhanced semi-empirical model
After enhancing the semi-empirical model by the “slice” approach, one resume assessment
of the snow depth, the key parameter in all mass-balance studies. To derive this parameter,
let us to write down the enhanced semi-empirical model (ESEM) in a form more fitting for
the Greenland ice sheet:
  h   
σ t0 = σ s0 + σ i0 = σ v
1
L p cos θ 1 − exp − 2 θ Y + σ i0, max exp − 2 hθ  (6.1)
2  Lp   Lp 
    
where σ i0 is the backscatter coefficient from the ice layer, Lp is the penetration path (2.18), Y
h
is the deviation factor (5.2), and hθ = is the slant snow depth. Let us also denote:
cosθ
Lp
σv ⋅ cos θ = σ ∞0 as the saturation backscatter coefficient. Given these expressions, the
2
h
normalized Snow Depth ( nSD = , see (3.25)) as a function of distance S along snow cover
Lp
with a variable snow depth, h, can be derived from (6.1) to take the form:
 σ (S ) 
1 − 
h(S ) 1 Yσ ∞  (6.2)
nSD( S ) = = − cos θ ( S ) ⋅ ln 
Lp 2  σi 
1 − 
 Yσ ∞ 
Here, σ (S ) ≡ σ t0 (S ) , σ ∞0 ≡ σ ∞ and σ i0 ≡ σ i to simplify the notation. If Y ⋅ σ ∞ >> σ i then

1  σ (S ) 
nSD( S ) = − cosθ ( S ) ⋅ ln 1 −  (6.3)
2  Y ⋅σ ∞ 
The equation above indicates that the estimated normalized snow depth depends on
number and kind of scatterers (σ∞) and their statistical characteristics (Y) at small (less then
wavelength) scales. Thus, to extract the nSD from measured radar data, σ (S ) , one needs to
know the product of the deviation factor and the saturation coefficient. To find the physical
snow depth, the penetration path must also be known. It should
Fig.6.1. Typical backscatter coefficient profile across a transect within the margin of the
Greenland Ice Sheet (in the vicinity of the Swiss camp)
be underlined, additionally, that the nSD  h  is not quite the same as the snow optical
=
 L p 

thickness  = 1h h 
= θ . Obviously, τ θ ≥ nSD . nSD is a combination of the
 L cosθ L 
 p p 
electromagnetic wave direction (Lp) and vertical geometrical (h) scaling factors of the snow
depth. These two terms coincide only for sounding in the nadir direction.
From the point of view of practical measurements, detailed analysis recommends stopping
the nSD estimate at a distance when  σ (S )  ≤ −2dB . It should also be noted also that the
 Y ⋅σ 
 ∞  dB
deviation factor (Y) might be less or more than unit (Figure 5.2). Because the product Y ⋅ σ ∞
is generally unknown, it is reasonable to establish its value as ~2 dB higher than the value of
the backscatter profile plateau shown, for instance, in Figure 6.1. The backscatter profile has
the main features inherent to the theoretical profile in the linear approach, as shown in
Figure 3.3.
Fig. 6.2. The normalized snow depth profile derived from the backscatter coefficient profile
in Figure 6.1
The profile has an exponential-like increase at the beginning of transect that transitions into
a plateau-like, approximately constant level. Most probably, this level represents the
saturation level caused by thick snow of depth greater than the propagation depth. The
corresponding nSD is shown in Figure 6.2.
The nSD profile plotted in Figure 6.2 is obtained under the assumption that the deviation
factor and the extinction coefficient are constant across the entire profile distance. Although
the approach demonstrated here does not give the absolute value of the snow depth without
certain assumptions, its advantage is that the assumptions are clearly delineated and put
forth. In turn, if one has a calibration point (spot), the technique allows extension of the
remote sensing methodology on vast remote areas with similar electromagnetic properties.
In addition, the nSD is a relative indicator of the snow environment stability and
redistribution.
The nSD in form (6.3) is derived from the ESEM, which in turn is based on consideration of a
flat surface, plane wavefront and long probing pulse. Hence, the result obtained is most
appropriate for SAR sounding of a thick snow slab.
6.2. An example of time transformation of nSD profiles within a selected Greenland

marginal area
The normalized snow depth is sensitive to the geometrical snow depth and to the extinction
coefficient. Assuming the snow microstructure to be homogeneous within a limited area,
one can suppose that nSD depends only on the snow depth and thus apply nSD for snow
mass-balance assessment. An illustrative example of an nSD profile change after a 10-year

period is shown in Figure 6.3.
Fig.6.3. Study area #3 (box with center: 75.070N and 54.300W) within the Greenland ice sheet
margin with several transects normal to the coastline (left). Normalized snow depth profiles
(right) for transect 3a (lower line in the study area) derived from RadarSAT-1 data. Distance
scale: 1 sample=50 m. Solid line is nSD profile on 11/16/1997, dotted line is nSD profile on
11/12/2007
The nSD profiles were derived from the RadarSAT-1 data archive from the Alaska Satellite
Facility in Fairbanks (granules R1_10617_SWB_261 from 1997 and R1_62753_SWB_261 from
2007 with coordinates of the center scene 75.230N/54.290W and 75.220N/54.370W
respectively, descending modes for both). As follows from the center locations of the data,
the SAR shots were performed from practically the same point in the orbit and thus can be
analyzed in parallel without additional corrections.
A comparative assessment of the nSD profiles indicates notable differences between the
profiles. Generally, these changes can be caused snow mass redistribution within the 10-
year period, variation of the extinction coefficient due to possible snow metamorphism, or
both factors simultaneously. The detailed analysis of these factors and their “weights” in the
deviations of these profiles is beyond the scope of this work but is planning for further
consideration.
7. Model spectral dependence of the backscatter

7.1. Spectral dependence of the backscatter in accordance with the A-U model
The spectral (frequency) dependence of the backscatter coefficient within the framework of
the semi-empirical model (A-U and ESEM) is the topic of the current section. The aim of this
discussion is to find out how spectral features depend on the thick snow properties.
Particularly, Davis & Moore (1993), namely assumed that a two-frequency radar system can
provide quantitative estimates of snow physical properties.
The model consists of two components that are definitely frequency-dependent: the specific
volume component (4.2) and the normalized effective depth of sounding (3.19).
c
Transforming the wavelength (λ) into the frequency  F =  , the combination of these two
 λ
components gives, for the “intermediate” snow regime:
 Z s 4   L p ( F ) cosθ   2h  
σ s0 = Y F h  1 − exp −   (7.1)
 c4   2h  
 L p (F ) cosθ 
 
The frequency-dependent core of (7.1) is
1 σ s0  F  4     2hθ  

A( F , hθ ) = 2 =     L p ( F ) 1 − exp −  (7.2)
Z s Y ⋅ cosθ  c      L p (F )  
   
h
is the slant snow depth; L p ( F ) = k e (F ) . The dimension of function (7.2)
−1
where hθ =
cosθ
is [L-3].
7.2. Estimation of spectral differences based on field penetration depth data

For further analysis, it is convenient to deal with the dimensionless spectral function, which
can be obtained by multiplying the spectral core (7.2) by the cube of the wavelength:
 F     2hθ  
B( F , hθ ) = λ3 A( F , hθ ) =    L p ( F ) 1 − exp −  (7.3)
 c     L p ( F )  
  
To evaluate (7.3) the spectral dependence of the penetration path (depth) should be
estimated. One can state that the high frequency irradiance interacts primarily with the
surface layer of snow with the relative small particles. And the low frequency irradiance
penetrates deeper and, thus, is affected by particles which sizes are larger in accordance
with the grain growth model (e.g., 4.18). For example, the grain radius changes from 0.25 to
0.6 mm within the depth of 10 m is in model of Drinkwater et al., (2001). In model of Forster
et al., (1999) the range of the particles’ changes is even more (up to 1 mm) and additionally
depends on the accumulation rate. Thus, to get more realistic spectral dependence, we
should take into account the depth of penetration of the electromagnetic wave into the snow
and the corresponding grain size change within this depth. Therefore, there is an obvious
paradox: to calculate frequency dependence of the penetration depth we should know the
latter a priory. Moreover, due to chain: “frequency-penetration depth-particle size” the
mean particle radius impacted the radar irradiance turns out to be dependent on the
irradiance frequency. In additional, the A-U model was derived assuming the constant
extinction coefficient, as it was mentioned in section 3. The solution of these questions
should be evaluated in further research. Taking in mind the restrictions listed above, the
penetration depth is assumed to be dependent only on frequency in the following
estimation. Since the significance of the particles’ size change compared with the fixed
mean size in the extinction coefficient calculation is currently unclear, it is reasonable to take
the field data available in the literature (Table 7.1) for rough assessment the spectral
dependence of the penetration depth.
# Frequency, GHz Penetration depth, m Literature source

5 18
10 9
1
20 4.3 Hofer & Matzler, 1980
35 1.5
5 20
10 10
2
20 4.8 Matzler, 1987
35 1.4
5.2 21.7
6.6 17.8
10.3 8.1
3 Rott et al., 1993
10.7 10.4
18.0 3.2
37 0.85
4 10 3.4±1.3 Devis & Poznyak, 1993
5 13.5 3.5±1.5 Davis, 1996
6 5.3 27±4 Hoen & Zebker, 2000
Table 7.1. Literature data on the experimental penetration path (depth)

These field data can be approximated by a simple empirical power function
L p ≈ 223.44 F −1.44 (7.4)
with the coefficient of determination equals to 0.9. In (7.4) the penetration path is in meters
and frequency is in GHz. It should be noted that field data, provided in the first three rows
of Table 7.1, are obtained based on measurements within the surface layer of the snow (up
to 3 m). It means that these data also do not take into account adequately the actual
stratification of the thick dry snow cover of the Greenland ice sheet and, therefore, the
further estimation should be considered only as a first approximation. A plot of (7.3) is
shown in Figure 7.1. This plot demonstrates a notable discrimination in behavior of the
spectral function B(F,hθ) for different snow depths. Particularly, the magnitudes of the
curves at a frequency of 5 GHz are notably different. Differences of the backscatter
magnitude for frequencies of 5 and 14 GHz as a function of the slant snow depth are
depicted in Figure 7.2.
28 60
26 50
24
y = 12.137e0.6032x
h=1m R2 = 0.9817 40
Snow depth, m
22 h=5m
B, dB
h=10m 30
20 h=20m
h=50m 20
18
10
16
14 0
4 9 14 19 24 29 -5 -4 -3 -2 -1 0 1 2 3
frequency, GHz Spectral difference (5-14 GHz), dB
Fig. 7.1. Dimensionless spectral characteristics Fig. 7.2. Slant snow depth as a function of
of the backscatter for several values of the the magnitude difference between the
slant snow depth based on the empirical dimensionless spectral components of the
spectral dependence of the penetration depth backscatter coefficients at 5 and 14 GHz
As follows from the approximation results,
hθ = 12.14 ⋅ exp(0.6 ⋅ ∆B5−14 ) (7.5)
10
5
y = 1.7323x - 25.835
0 R2 = 0.9528 GITS
sigma_0, dB
-5 NASA_U
-10 Linear (NASA_U)
-15 Linear (GITS)

y = 1.6822x - 31.012
-20 R2 = 0.9967
-25
0 5 10 15 20
frequency, GHz
Fig. 7.3. Spectral pattern of the experimental data of (Baumgartner et al., 1999) for an
incidence angle of 400 for two test sites in Greenland
where the slant snow depth is in meters and the spectral difference ∆B5−14 is in dB. An
illustrative example of the spectral difference observed in Greenland can be provided from the
known literature. Baumgartner at al. (1999) conducted field measurements of the backscatter
coefficient from the dry snow site GITS (77.10N, 610W) and from the site NASA-U (730N,
50.50W) located between the dry snow and percolation zone for four wavelengths, 5.3, 10, 13.5
and 17 GHz, at temperature from -210C to -150C, with the ground-based radar. The primary
goal of this experiment was to study the angular dependence of the backscatter. For our
purposes, the spectral dependence related to the incidence angle of 400 (where the backscatter
is expected to be due mainly to the volume scattering) can be extracted for the four points and
σ s0
is shown in Figure 7.3. Taking into account that B ( F , hθ ) = µ , where μ is an unknown
F3
coefficient that we assume to be the same for all wavelengths, one can obtain the spectral ratio
or the spectral difference in the logarithmic units for a pair of frequencies F1 and F2:
[ ]
∆BdB , F1 − F2 = B(F1 )dB − B(F2 )dB = σ s0 ( F1 ) dB − σ s0 ( F2 ) dB − 30[lg F1 − lg F2 ] (7.6)
If, for example, F1=5 GHz and F2=14 GHz, we get:
[ ]
∆BdB , 5−14 = σ s0 (5) dB − σ s0 (14) dB + 13.41 (7.7)
Differences in the backscatter coefficients of these frequencies for both test sites are
approximately the same and equal roughly to -13 dB, as shown in Figure 7.3. Assuming the
backscatter error of ~±0.5 dB, the corresponding spectral difference (from 7.7) is
approximately equal to 0.4±0.5 dB. In accordance with (7.5), the slant snow depth is ~6.7-
20.8 m. Since these data are related to an incidence angle of 400, the estimate of actual snow
depth is 5.1-15.9 m. This value is in accordance with a note in the paper that the depth of
snow was ~7-8 meters for the NASA-U site, at least.
Thus, a rough assessment of the snow depth within several selected dry snow areas, based
on the spectral dependence of the backscatter in the frames of the SEM (ESEM), gives results
comparable with in-situ measurements.
7.3. Spectral characteristics of the penetration depth for model monodisperse cold
snow
It is interesting to compare the empirical (7.4) and theoretical frequency dependence of the
penetration path (depth). Calculation of the extinction coefficient has been completed in the
same manner as in section 3, with the additional estimation of particle concentration n0 (see
(4.8)):
−1
4  (7.8)
n0 = fv ⋅  π r 3 
 3 
where f v = ρ s is the fractional volume of ice in snow (0.2-0.4), ρs and ρi are density of snow
ρi
and ice, respectively. We also should demarcate the frequency and particle size inherent to
the Rayleigh approach ( 2π r F m ≤ 0.5 ). The penetration depth (path) in snow is calculated for
c
several combinations of snow density and particle size for a temperature of -150C. The result
of this calculation is shown in Figure 7.4 together with the empirical dependence (7.4). The
gently sloping of the empirical curve compared with the theoretical ones indirectly support
the idea, mentioned above, about the contribution of the particle size in spectral behavior of
the penetration depth (see the corresponded comments in the previous subsection).
In Figure 7.5, the results obtained from these calculations for T= -150C and the snow density
of 0.3 g/m3 are compared with the known values (Ulaby et al., 1986, Fig. 19.64, p.1608) for
T= -10C and density of 0.24 g/m3. The plot indicates the increasing of the penetration depth
at low temperatures for the low frequency region, a trend most obvious for small particles of
0.5 mm radius. This behavior is explainable because the losses in snow caused by absorption
decrease as the temperature decreases (Ulaby et al., 1986, Fig.Е.3, p. 2027), and the
absorption mechanism prevails over scattering in the low frequency region. Pointed
changes become more apparent for small particles due to more significant role absorption
plays for them as compared with large particles.
10000 10000
1000 1000
Penetration depth, m
Penetration depth, m
100 100
10
10
1
1
0.1
0.1
0.01
0 5 10 15 20 25 0.01
frequency, GHz 0 5 10 15 20 25
ρ=0.3g/m3; r=0.5 mm ρ=0.3g/m3; r=1.0 mm ρ=0.3g/m3; r=2.0 mm frequency, GHz

ρ=0.6g/m3; r=0.5 mm ρ=0.6g/m3; r=1.0 mm ρ=0.6g/m3; r=2.0 mm
r=0.5 mm, T=-1C r=0.5mm, T=-15C
Rayleigh_limit Empirical r=2 mm, T=-1C r=2.0 mm, T=-15C
Fig.7.4. Theoretical spectral dependence under the Fig.7.5. Comparison of the spectral
Rayleigh approach of the penetration depth for a dependence of the penetration depth
temperature of -150C (our calculations) with calculated with Mie formulas for a
comparison of the empirical relationship (7.4) temperature of -10С (Ulaby et al., 1986)
with our calculations under the
Rayleigh approach for a temperature of
-150С
8. Backscatter depth profile model for snow sounding with broadband pulses
of short effective (compressed) length
8.1. Phenomenological approach in the case of very short probing pulse
The above considerations apply when the scattering volume was less then the probing pulse
volume. In the case of altimeter sounding of thick snow cover with short broadband chirp
pulses this condition is not satisfied, the backscatter coefficient is represented not by the
point of a radar signature but in the form of a backscatter profile along the wave
( p) ( p)
propagation path. When hsct << hθ and hsct << L p , the volume component of the total
backscatter from a snow area with the center located at a slant distance of R beneath the
surface should simply mirror the backscatter coefficient depth distribution taking
attenuation into account:
σ s (R )
σ s0 (R ) = exp(− 2k e R ) (8.1)
Aill
N
where σ s (R ) = ∑ σ i is the total radar cross-section (RCS) of N scatterers contributing
i =1
backscatter from a scattering volume Vsct from a distance R. Since the scattering volume is
( p)
Vsct = hsct Aill cosθ , and taking the correction on the deviation factor into account (see section
5), the equation (8.1) can be presented in the form:
σ v (R )  R 
σ s0 (R ) = Y ( R) Vsct exp(− 2k e R ) = Y ( R)σ v (R ) hsct
( p)
cosθ exp − 2 (8.2)
Aill 
 L p (R ) 
σs
where σ v = is the specific volume backscatter, and the return signal is proportional to
Vp
the cosine of the incidence angle and has the envelope of the exponential form distorted by
the variability of the statistical, scattering and absorption properties of the snow medium
along the wave propagation path. The illuminated area Aill (transverse size of the scattering
volume) changes during the burial of the probing pulse into the snow slab due (mainly) to
the wave sphericity in the depth-limited mode (see section 2). However, this effect does not
impact the accuracy of the specific volume backscatter coefficient presented here, due to its
the mutual cancellation in the nominator (via Vsct) and the denominator of equation (8.2). As
a result of this cancellation, for very short probing pulses, the backscatter envelope depends
only on changes in the specific volume backscatter with distance.
8.2. Analytical derivation for commensurate length of the scattering volume and snow
depth
In general case, when h ( p ) ~ h and hsct
( p)
~ L p , the interactions between a probing pulse
sct
cosθ
with a normalized form U(R) and an extended target such as snow cover is described by the
convolution integral (Moore & Williams, 1957):
Pr (t ) = PT (t ) ⊗ Ps (t ) (8.3)
where Pr(t) is the return signal, PT(t) is the transmitted pulse profile and Ps(t) is a term that
includes the distribution range of the scattering facets, their scattering properties and the
effects of the antenna gain. As applied our consideration of volume backscatter, one can
write
∞
 x 
σ (R ) = cosθ ∫ dx ⋅ U (x − R ) ⋅ Y ( x)σ v (x ) exp − 2
0 (8.4)
L p ( x ) 
s
0 
If, for example, the probing pulse has a Gaussian form, which is usually an adequate
approximation for short pulse systems (e.g., Brown, 1977), the weighting function of the
 2 
scattering volume is U ( R) = exp − β R  , where Δq is the spatial duration of the scattering
 ∆2q 

 β
volume weight function at level q = exp −  . For level q=0.5 (-3 dB), β=-4ln0.5=2.8. If, in
 4
addition, the microstructural, statistical and attenuation properties of the snow slab do not
change along the propagation path within the snow pack, one can write:
∞
 x 
σ s0 (R ) = cos θ ⋅ Yσ v ∫ dx ⋅ U (x − R ) ⋅ exp − 2 
 (8.5)
 Lp 
0
For short pulses, if their envelope can be represented by the delta-function, i.e.,
U ( x − R) ~ δ (x − R ) , equation (8.5) yields the corresponding result obtained earlier (8.3) for
the very short pulses configuration. For the long pulses, where ( p)
hsct > L p and
L
 x 
σ s0 ~ ∫ dx exp − 2  , which yields to the standard form of the A-U model (3.13). For the

0  Lp 
general case, a closed-form solution of the equation (8.5) can be obtained.
Taking into account the tabulated integral (Abramowitz & Stegun, 1972):
∞ b2 ∞ b2
∫ dx exp(− ax )
2 π  b + 2ax  1 π  b 
− bx = e a
4 Φ   = e 4 a 1 − Φ   (8.6)
2 a  2 a 0 2 a  2 a 
0
x
exp(− t )dt is the error function, one can get the integral of equation
2
where Φ (x ) =
π ∫
2
0
(8.5) for a Gaussian probing pulse:
∞
 x  1 π  1  ∆ 2    ∆  
( )    q  − 2 R  ⋅ 1 − Φ 1  q − β R   (8.7)
∫0 dx ⋅ U x − R ⋅ exp

− 2 =
L p  2
∆ q
β
exp
 β  L p  Dp    β  L p ∆ q  
       
∆q
If one denotes = ∆ , then (8.5) has the form:
β
 2  
π   
σ s0 (R ) = cosθ ⋅ Yσ v ∆ exp ∆  − 2 R  ⋅ 1 − Φ  ∆ − 1 R 
  L p ∆  (8.8)
2  L p  Lp    
  
1  ∆  R
Finally, if one denotes: σ max
∗
= σ v L p cos θ (see 3.24),   = γ and S = , the closed-
2  Lp  Lp
form solution of equation (8.6) using these dimensionless parameters can be written as:
  S 
σ s0 (S ) = Yσ max
∗
π γ exp(γ 2 − 2S ) ⋅ 1 − Φ γ −  (8.9)
  γ 
Distance normalization in (8.9) is chosen through the parameter S for the purpose of
comparing the backscatter profile calculated here with the modeled “clear” backscatter
profile derived for a very short pulse (8.2) and governed only by attenuation. The equation
obtained is close to the volume component of the SV model of Davis & Moore (1993). The
distinctions of the current approach are that it uses a slightly different description of the
pulse-snow interaction, the end result is written in terms of the backscatter coefficient depth
profile, and the state of homogeneity of the snow mass is taken into account through the
deviation factor, Y. It is interesting, that equation (8.9) coincides with the backscatter model
from the sea surface, obtained by Barrick (1972), exactly by the form with, of course, other
meaning of model parameters.
8.3. Comparison of the results obtained with the phenomenological approach
8.3.1. Very short probing pulse

For the limit case when γ << 1 , and taking into account that Φ ( − x ) = −Φ ( x ) and
lim Φ (x ) = 1 (Abramowitz & Stegun, 1972), the corresponding backscatter coefficient is:
x →∞
  1    1 
σ s0 (S )γ <<1 = Yσ max
∗
π γ exp[− 2S ]⋅ 1 − Φ − S  = Yσ max
∗
π γ exp[− 2S ]⋅ 1 + Φ S  ≈ (8.10)
  γ    γ 
2Yσ max∗
π γ exp(− 2S ) = cos θ Yσ v ∆ π exp(− 2S )
This result confirms relationship (8.2) obtained above based on phenomenological

considerations.
8.3.2. Long probing pulse

Taking into account the asymptotic value of the error function
1  1 
Φ (x ) ≈ 1 − e − x
2
1 − 2 − ...  (Abramowitz & Stegun, 1972), and taking only the first
x >>1 πx 2 x 
term of the extension, we get the value of (8.9) for γ>>1:
 2 S  1 
2 −1
S
σ (S )γ >>1 ≈ Yσ
0
s
∗
max ( 2
)
π γ exp γ − 2S exp − γ + 2S −   
γ
 γ − 
γ
(8.11)
     π  
For long pulses, the available distance variation is much less than the pulse length,
i.e.,  S  R
= << 1 . Hence,
γ 
  γ >>1 ∆
  S 2   S 
−1
1
σ (S )γ >>1 ≈ Yσ
0
s
∗
max exp −    1 − 2  ≈ Yσ max
∗
= ⋅ Yσ v D ∗p (8.12)
  γ    γ γ >>1 2
This expression also confirms the result of the A-U model (enhanced) under thick snow
condition (3.24). Thus, an important positive feature of the model (8.9) is that its results
coincide with those obtained with phenomenological approaches for extreme values of the
scattering volume (much less and much more than the penetration path respectively).
Nevertheless, it should be noted that ignoring the change in the illumination area during the
burial of the probing pulse into snow media is, strictly speaking, valid only for a short
probing pulse or for the beamwidth-limited mode. For long and intermediate-length
probing pulses with wavefront sphericity in the commonly-existing depth-limited
configuration, this effect should be taken into account because there are slices (see sections 2
( p) .
and 5) of different transverse sizes within hsct
8.4. Numerical calculation of the return signal for intermediate length of the probing
pulse
A plot of the normalized backscatter coefficient σ ∗ (S , γ ) = σ s (S , γ ) (Smax is the normalized
0
σ s0 ( S max , γ )
distance where function (8.7) has the maximum value at a given γ) is shown in Figure 8.1.
Fig.8.1. Illustration of the reproducibility of the snow backscatter uniform profile taking into
account only attenuation (bold dash line) and finite probing pulse length normalized to the
penetration depth (parameter γ = ∆ L ).
p
Legend: solid line, γ=0.01; dot line, γ=0.035; dashed line, γ=0.06; dashed-dot line, γ=0.085;
dashed-dot-dot line, γ=0.11; long dashed line, γ=0.135
This figure illustrates the contribution of the pulse length to the return signal envelope
pattern. The shorter pulse clearly gives a better approximation of the real backscatter profile.
Contemporary altimeters use a probing pulse with a physical duration of 20-50
microseconds and a corresponding effective duration (due to intrapulse frequency chirping)
equal to 3.125 nanoseconds (e.g., Rosmorduc et al., 2006). In spatial units, these values are
cτ 0
equivalent to ∆ ≈ ≈0.5 m. For Lp~10-50 m, the expected values of the parameter γ range
2
from 0.05 to 0.01. An important feature of this plot is that the value of the maximum of the
return signal for any maximum position (Smax) and values of parameter γ are the same:
σ s∗ (S max , γ ) = π σ max
∗
Y (8.13)
9. Summary
The semi-empirical model of the volumetric component of the backscatter coefficient,
proposed by Attema & Ulaby (1978) (SEM or A-U model) is considered in more detail. The
basic advantage of the SEM is its mathematical simplicity and clear physical sense but the
SEM, derived using the incoherent approach, does not explain some remote sensing
experimental data collected over the dry snow zone of the Greenland ice sheet. We discuss
the main configurations of radar sounding of snow, including the relationships between the
technical and positional parameters of radar on one hand and the geometrical and
electromagnetic properties of snow cover on the other. Based on this consideration, the field
of applicability of the SEM is figured out for the further assessments. Some qualitative
assessments of the configuration of the scattering volume were carried out using plane and
spherical surface approaches. More detailed study of the diffraction of space-bounded
spherical electromagnetic wave on a large and strongly absorbing dielectric sphere with a
rough surface is needed to evaluate this problem further. Detailed SEM estimates and
experimental radar data over the dry snow zone of the Greenland ice sheet are provided for
the negative temperatures inherent to the winter seasons. Data analysis shows a significant
discrepancy between the expected and actually measured values of the backscatter
coefficient. Possible reasons for the observed deviations of the SEM results are discussed
based on the enhanced semi-empirical model (ESEM) and analysis of the dependence of the
backscatter coefficient on the snow optical thickness and its spectral function. The ESEM is
obtained on the basis of the so-called "slice" approach, taking into account partially-coherent
backscattering due to finite wave front size and considering the contribution of the statistical
characteristics of small-scale fluctuations of concentration and of sizes of the ice particles in
a snow pack. The known semi-empirical model is only a special case of the ESEM. Using the
ESEM, some observable radar signatures of the Greenland ice sheet, including the dry snow
zone, can be interpreted in terms of the statistical characteristics of the snow microstructure.
The model obtained retains the mathematical simplicity and physical clarity of the known
A-U model and significantly expands its application for the interpretation of experimental
data. Available field data, which confirm the consistency of the ESEM with observed values
of the backscatter coefficient, are also provided. As a consequence of this analysis, the
concept of the normalized snow depth (nSD) is formulated, and its practical application for
monitoring snow mass displacement is described. The frequency-dependent core of the
backscatter has been analyzed as well. Empirical and theoretical spectral characteristics of
the penetration depth are also presented. The result allow for improvements in our
understanding of opportunities for a two-frequency method of sounding thick snow covers
and promote more accurate interpretations of the known experimental data. The extension
of the ESEM for sounding thick snow cover with broadband pulses of short effective length
is also discussed, resulting in an exact description of the return signal envelope as a function
of the electromagnetic and microstructural characteristics of the snow cover. The validity of
the model is supported by the coincidence of the model estimates with assessments obtained
for several specific cases using phenomenological approach. In conclusion, this work
demonstrates that the enhanced version of the semi-empirical model can be applied
successfully in the analysis of radar backscatter from thick snow cover.
10. Acknowledgements
This work was supported by NASA’s Cryospheric Sciences Program. The author would like
to thank Dr. W. Abdalati for his permanent attention to this work and useful discussions.
11. References
Abramowitz, M., & I.A. Stegun, Eds. (1972). Hanbook of mathematical functions with
formulas, graphs, and mathematical tables. New York: Wiley.
Alley, R.B., Bolzan, J.F., & Whillans, I.M. (1982). Polar firn densification and grain growth.
Ann Glaciol., vol.3, pp.7-11.
Atlas, D. (1964). Advances in radar meteorology. Advances in Geophysics, vol.10, pp. 317-
479.
Attema, E., & Ulaby, F. (1978). Vegetation modeled as water cloud. Radio Science, vol.13,
pp. 357-364.
Barrick, D. E. (1972). Remote sensing of sea by radar. In Derr, V. ed. Remote sensing of the
troposphere. Washington, DC, U.S. Government Printing Office, pp. 12.1-12.46.
Battan, L.J. (1959). Radar Meteorology, Univ. of Chicago Press, Chicago, Ill.
Baumgarnter F., Jezek K., Forster R.R., Gogineni S.P., & Zabel I.H.H. (1999). Spectral and
angular ground-based radar backscatter measurements of Greenland Snow facies.
IGARSS’99 Proceedings, 28 June-2 July 1999, Humburg, vol.2, pp. 1053-1055.
Benson, C. S. (1996). Stratigraphic studies in the snow and firn of the Greenland ice sheet.
U.S. Army Cold Reg. Res. and Eng. Lab (CRREL), Rep.70, 120 pp. Hanover, NH.
(Revised edition of 1962 report).
Brown, G.S. (1977). The average impulse response of a rough surface and its applications.
IEEE Transactions on Antennas and Propagation, vol. AP-25, No. 1, January 1977,
pp. 67-74.
Clift, G.A. (1985). Use of radar in meteorology. Secretariat of the World Meteorological
Organization, Geneva.
Davis, C.H. (1996). Temporal change in the extinction coefficient of snow on the Greenland
Ice sheet from an analysis of Seasat and Geosat altimeter data. IEEE Trans. Geosci.
Remote Sens., vol.34, No.5, September 1996, pp. 1066-1073.
Davis C.H., & Pozdnyak V.I. (1993). The depth of penetration in Antarctic firn at 10 GHz.
IEEE Trans. Geosci. Remote Sens., vol. 31, No.5, September 1977, pp. 1107-1111.
Davis, C.H., & Moore, R.K. (1993). A combined surface- and volume scattering model for
ice-sheet radar altimeter. Journal of Glaciology, vol. 39, No. 133, pp. 675-686.
Drinkwater M.D., Long D.G., & Bingham A.W. (2001). Greenland snow accumulation
estimates from satellite radar scatterometer data. Journal of Geophysical Research,
vol. 106, No. D24, December 2001, pp. 33,935-33,950.
ERS-1 and ERS-2 SAR Images. ASF document. September 18, 1996, 21 p. Available online:
http://www.asf/alaska.edu/content/reference/pdf/ERSproducts.pdf
Etheridge D.M., Steele L.P., Francey R.J., & Langenfelds R.L. (2001). Historic CH4 records
from Antarctic and Greenland ice cores, Antarctic firn data, and archived air
samples from Cape Grim, Tasmania. Available online:
http://ediac.ornl.gov/trends/atm_meth/lawdome_meth.html
Fabelinskii I. L. (1968). Molecular scattering of light. Translated from Russian, New York,
Plenum Press.
Forster R.R., Jezek, K.C., Bolzan, J., Baumgartner, F., & Gogineni, S.P. (1999). Relationship
between radar backscatter and accumulation rates on the Greenland ice sheet. Int. J.
Remote Sensing, vol. 20, No. 15&16, pp. 3131-3147.
Fung A.K. (1994). Microwave Scattering and Emission Models and Their Applications.
Boston, Artech House, Inc.
Gossard E.E., & Srauch, R.G. (1983). Radar observation of clear air and clouds. Elsevier, New
York.
Higham, M., & Craven, M. (1997). Surface mass balance and snow surface properties from
the Lambert Glacier Basin traverses 1990-94. Tech. Rep. Res Rep. 9, Antarctic CRC,
Hobard, Australia.
Hoen E.W., & Zebker H.A. (2000). Penetration depths inferred from interferometric volume
decorrelation observed over the Greenland Ice Sheet. IEEE Trans. Geosci. Remote
Sens., vol.38, No.6, November 2000, pp. 2571-2583.
Hofer R., & Matzler C. (1980). Investigations of snow parameters by radiometry in the 3 to
60 mm wavelength region. Journal of Geophysical Research, vol. 85, pp. 453-460.
Jaedicke, C., T. K. Thiis & Bang, B. (2000). The snowdrift around a small hill in the High
Arctic. Proceedings of the 4th Int. conference on Snow Engineering. Trondheim,
Norway, 19-21 June 2000. Hjoth-Hansen, Holand, Loset & Norem (eds). Balkema,
Rotterdam, pp. 75-80.
Jezek, K. C., P. Gogineni & Shanableh, M. (1994). Radar measurements of melt zones on the
Greenland Ice Sheet. Geophysical Research Letters, vol. 21, No. 1, pp. 33-36.
Kendra J.R., K. Sarabandi, & Ulaby, F.T. (1998). Radar measurements of snow: experiment
and analysis. IEEE Trans. Geosci. Remote Sens., vol. 36, No. 3, May 1998, pp. 864-
879.
Koskinen J. (2001). Snow monitoring using microwave radars. Thesis for the degree of
Doctor of Technology. Helsinki University of Technology, Espoo, January 2001,
Report 44, 31p.
Liu, T.W. (2002). Progress in scatterometer application. Journal of Oceanography, vol. 58,
pp. 121-136.
Lo, Y.T., & S.W. Lee, Eds. (1993). The antenna handbook. Volume II, Antenna theory. New
York, Van Nostrand Reinhold.
Lytle, V.I., & Jezek, K. (1994). Dielectric permittivity and scattering measurements of
Greenland firn at 26.5-40 GHz. IEEE Trans. Geosci. Remote Sens., vol. 32, March
1994, pp. 290-295.
Matzler, C. (1987). Application of Interaction of Microwaves with the natural snow cover.
Remote Sensing Reviews, vol. 2, pp. 259-387.
Moore, R.K., & Williams, C.S. (1957). Radar terrain return at near vertical incidence.
Proceedings Institute of Radio Engineers, vol. 45, pp. 228-238.
Naito, K., & Atlas, D. (1966). On microwave scatter by partially coherent clouds, In Proc.
12th Weather Radar Conf., AMS, Norman, OK, October 17-20, 1966, pp. 7-12.
Newkirk, M.H., & Brown, G.S. (1996). A waveform model for surface and volume scattering
from ice and snow. IEEE Trans. Geosci. Remote Sens., vol. 34, No. 2, March 1996,
pp. 444-456.
Newkirk, M.H., & Brown, G.S. (1992). Issues related to waveform computations for radar
altimeter applications. IEEE Trans. Geosci. Remote Sens., vol. 40, No. 12, December
1992, pp. 1478-1488.
Noveltis, C.R. (2005). Feasibility study of imaging the Antarctic ice using a spaceborne P-
band radar. WP 200: Electromagnetic Model Review Document, ESA Contract
Number 18195/04/NL/CB, p. 91, 2005.
RadarSAT -1 Standard Beam SAR Images. ASF document. June 21, 1999, 20 p. Available
online:
http://www.asf/alaska.edu/content/reference/pdf/R1stanprod.pdf
Rosmorduc, V., J.Benveniste, O. Lauret, M. Milagro, & Picot, N. (2006). Radar Altimetry
Tutorial, J. Benveniste and N. Picot (Eds)., 301p. Available online:
http//www.altimetry.info.
Rott H., Sturm K., & Miller H. (1993). Active and passive microwave signatures of Antarctic
firn by means of field measurements and satellite data. Annals of Glaciology, vol.
17, pp. 337-343.
Schytt, V. (1964). Scientific results of the Swedish glaciological expedition to
Nordaustlandet, Sptisbergen, 1957 and 1958. Geografiska Annaler, vol. 44, pp. 243-
281.
Scott, J.B.T., Mair, D., Nienow, P., Parry, V., & Morris, E. (2006). A groud-based radar
backscatter onvestigation in the percolation zone of the Greenland ice sheet.
Remote Sensing of Environment, vol. 104, pp. 361-373.
Shi, J., R.E. Davis, & Dozier, J. (1993). Stereological determination of dry-snow parameters
for discrete-scatterer microwave modeling. Ann. Glaciol., vol. 17, pp. 295-299.
Siegert A.J.F., & Goldstein, H. (1951). Coherent and incoherent scattering from assemblies of
scatterers, /In: Propagation of short radio waves. D.E. Kerr, (Ed), McGraw-Hill,
New York.
Tiuri, M.E., A.H. Sihvola, E.G.Nyfors, & Hallikainen, M.T. (1984). The complex dielectric
constant of snow at microwave frequencies. IEEE Journal of Oceanic Engineering,
vol. OE-9, No. 5, pp. 377-382, December 1984.
Tsang, L., J. Pan, D. Liang, Z. Li, D. W. Cline, & Tan, Y. (2007). Modeling active microwave
remote sensing of snow using dense media radiative transfer (DRMT) theory with
multiple-scattering effects. IEEE Trans. Geosci. Remote Sens.. vol. 45, No. 4, April
2007, pp. 990-1004.
Ulaby F.T., Moore R.K., & Fung A.D. (1982). Microwave Remote Sensing, Active and
Passive. Vol. II. Radar Remote Sensing and Surface Scattering and Emission
Theory. Artech House, Inc., Norwood.
Ulaby F.T., Moore R.K., & Fung A.D. (1986). Microwave Remote Sensing, Active and
Passive. Vol. III. From Theory to Applications. Artech House, Inc., Norwood.
Ulaby, F.T. , P. Siqueira, A. Nashashibi, & Sarabandi, K. (1996). Semi-empirical model for
radar backscatter from snow at 35 and 95 GHz. IEEE Trans. Geosci. Remote Sens.,
vol. 34, No. 5, pp. 1059-1065, September 1996.
Woods, G.A. (1994). Grain growth behavior of the GISP2 ice core from central Greenland.
Technical Report Series 94-002, Earth System Science Center, Penn State University,
USA.
Yurchak, B.S. (2009). Radar volume backscatter from spatially extended geophysical targets
in a “slice” approach. IEEE Trans. Geosci. Remote Sens., vol. 47, issue 10, (in press).
Zahnen, N., F. Jung-Rothenhausler, H. Oerter, F. Wilhelms & Miller, H. (2002). Correlation
between Antarctic dry snow properties and backscattering characteristics in
Radarsat SAR imagery. Proceedings of EARSeL-LISSIG_Workshop Observing our
Cryosphere from Space, Bern, March 11-13, 2002, pp. 140-148.
Zwally, H.J., & Brenner, A.C. (2001). Ice sheet dynamics and mass balance. In book: Satellite
altimetry and earth sciences. L.L. Fu and A. Cazenave, (Eds). International
Geophysics series, vol. 69, pp. 351-369.
Theoretical Modeling for Polarimetric
Scattering and Information Retrieval of SAR Remote Sensing 141
X7

Scattering and Information Retrieval
of SAR Remote Sensing
Ya-Qiu Jin
Fudan University
China
1. Introduction
Synthetic aperture radar (SAR) imagery technology is one of most important advances in
space-borne microwave remote sensing during recent decades. Completely polarimetric
scattering from complex terrain surfaces can be measured. Fully understanding and
retrieving information from polarimetric scattering signatures of natural media and SAR
images have become a key issue for the SAR remote sensing and its broad applications.
This paper presents some research progress in Fudan University on theoretical modeling of
the terrain surface for polarimetric scattering simulation and Mueller matrix solution, mono-
static and bistatic SAR image simulation, new parameters for unsupervised surface
classification, DEM inversion, change detection from multi-temporal SAR images, and
reconstructions of buildings from multi-aspect SAR images, etc. Some applications are
discussed.
2. Model of Vegetation Canopy and Mueller Matrix solution

As a polarized wave Einc (  , ) is incident upon the natural media, the scattering field is
written as
 Evs  eikr  Svv Svh   Evi  eikr
      S  Einc (  , ) (1)
 Ehs  r  S hv Shh   Ehi  r
where 22-D (dimensional) complex scattering amplitude function S can be measured by

the polarimetry. The incident polarization is indicated by the elliptic and orientation angles
(  , ) . Using the Mueller matrix solution of vector radiative transfer equation (Jin, 1994)
and Eq. (1), the scattered Stokes vector (four Stokes parameters) can be obtained or
measured as (Jin, 2005)
I s ( , )  M ( , ;  0 ,0 )  I i (  , ) (2)
The Mueller matrix solution of a layer of scatterers model, as shown in Figure 1, is written as
M ( , ;   0 ,0 )  M 0  M1  M 2  M 3  M 4 (3)
M 0  exp   d sec κ e ( , )   R ( , ;    0 ,0 )  exp   d sec 0κ e (   0 ,0 ) (4a)
0
M1  sec  d dz exp  z sec κ e ( , )   P ( , ;    0 ,0 )  exp  z sec 0 κ e (   0 ,0 )  (4b)
0 2  /2
M 2  sec  d dz exp  z sec κ e ( , )   0 d 0 d  sin  P ( , ; , )
(4c)
 exp (d  z)sec κ e ( , )   R ( , ;    0 ,0 )  exp  d sec 0 κ e (   0 , 0 ) 
2  /2
M 3  sec exp   d sec κ e ( , )   0 d 0 sin  d R ( , ;   , )
0
(4d)
 d dz exp ( d  z)sec κ e (   , )  P(   , ;    0 ,0 )  exp  z sec 0 κ e (   0 ,0 ) 
2  /2
M 4  sec exp   d sec κ e ( , )   0 d 0 d  sin  R ( , ;   , )
2  /2
 exp  ( d  z)sec κ e (   , )  0 d 0 d  sin  P (   , ; , ) (4c)
 exp  ( d  z)sec κ e ( , )  R ( , ;    0 ,0 )exp   d sec 0 κ e (   0 ,0 ) 
where κ e , P, R are the extinction matrix, phase matrix of non-spherical scatterers, and
reflectivity matrix of underlying rough surface, respectively. Five contributions from
scatterers and underlying surface are illustrated in Figure 1. κ e , P can be expressed by
 S pq S st*  of S in Eq. (1). The co-polarized and cross-polarized backscattering coefficients,
 c and  x , polarization degree ms for scattered Stokes echo with partial polarization and
other functions can be numerically calculated (Jin and Chen, 2002).
Eqs. (4a-c) will be applied to numerical simulation of polarimetric scattering from the terrain
surface with physical parameters of scatterers (Jin, 2005), as described in next sections.
M2 M 3 M4
M0
M1
Fig. 1. A model of non-spherical scatterers for vegetation canopy
The Mueller matrix is a 4×4-D real matrix with complex eigenvalues and eigenvectors. To be
physically realizable, this matrix must satisfy the Stokes criterion together with several other
restrictive conditions. Unfortunately, however, these restrictions do not have any direct
physical interpretation in terms of the eigenstructure of the Mueller matrix.
The coherency matrix C is applied to the study of polarimetric scattering of SAR image (Jin
and Cloude, 1994). Define the scattering vector as
k P  12 [ Svv  S hh , Svv  S hh , Svh  S hv , i ( Svh  S hv )]T  12 [ A, B, C , iD]T (5)
where A, B, C, D are sequentially defined in Eq. (5), and the subscript P denotes Pauli
vectiorization. The coheremcy matrix is defined as
  AA*   AB*   AC*   AD*  

  BA*   BB*   BC*   BD*  
1
C  k Pk P   (6)
4   CA*   CB*   CC*   CD*  
 
 DA*   DB*   DC*   DD*  
The eigenvalue and eigenvector of the coherency matrix are defined as
4
C  k Pk P   i k Pik Pi  (7)
i 1
All eigenvalues are real and non-negative, and 1  2  3  4 .

The coherency matrix is also related with the Mueller matrix as
1
M i 1, s 1    Trace(σ i σ t σ s σ u )[C]tu (8)
t u 2
where σ i , i  0,1, 2,3 are the Pauli matrices.

Usually, the case of Svh  S hv is simply assumed. The entropy H is defined by the
eigenvalues of the coherency matrix as
3
H    Pi log 3 Pi , Pi  i /(1  2  3 ) (9)
i 1
The entropy is an important feature since it relates the randomness of scatter media with
other physical parameters, such as canopy configuration, biomass, etc.
The coherency matrix is 4×4-D positive semidefinite Hermitian and as such has real non-
negative eigenvalues and complex orthogonal eigenvectors. Amplitude and difference of
four eigen-values are functionally related with polarimetric scattering process of the terrain
canopy. Eigen-analysis of the coherency matrix yields better physical insight into the
polarimetric scattering mechanisms than does the Mueller matrix and further, it can be
employed to physically identify a Mueller matrix by virtue of the semi-definite nature of the
corresponding coherency matrix.
For the first-order solution in small albedo, cross polarization is small and correlations such as
the terms  CA* ,  CB* ,  DA* ,  DB*  etc. in Eq. (6) are always very small, and might
be neglected in our following derivations. Actually, for weak assumption of azimuthal
symmetric orientation, these correlations have been proved to be zero. Applying this
approximation to Eq. (6), the eigenvalues of C have been derived as (Jin and Cloude, 1994)
1 1
1,2  [| A |2   | B |2   (| A |2   | B |2 ) 2  4  A* B  AB*  ] , 3  | C |2  (10)
8 4
It is interesting to see that if  AB*  A * B  AA*  BB*  at an ordered case, it yields
2  0 .
Substituting A,B,C and D of Eq. (6) into Eq. (10), the eigen-values are derived using the
power terms of | Svv |2 , | S hh |2 , | Shv |2  as follows,
1 1 | Svv |2 | Shh |2 

1  (| Svv |2   | S hh |2 )  2 , 2  (1  ) , 3 | S hv |2  (11)
2 2 (| Svv |2   | Shh |2 )
where the configuration parameter
 Svv S hh*  Svv*Shh 

 (12)
| Svv |2 | Shh |2 
varies from 0 of totally random media to 1 of ordered media, i.e. no-random case.
It can be seen that the first eigen-value 1 now indicates the total vv and hh co-polarized
power subtracting their coherence defined by the parameter  of Eq. (12), which varies from
1 at ordered case to 0 in totally random medium. The second eigen-value 2 now takes into
account for the vv and hh coherent power. As  approaches to one, 2  0 for ordered
media, and as  approaches to zero, 2 would increase to indicate the loss of the vv and hh
power coherency for totally random media. The third eigen-value 3 is now due to
depolarization caused by the media randomness.
The (p,q=v,h)-polarized backscattering coefficient is
 pq  4 cos | S pq |2  (13)
Substituting Eq. (13) into Eqs. (11,12), the functions Pi , i  1, 2, 3 , of Eq. (9) are then
derived as
P1  1  P2  P3 , P2  X (1  )(1   ), P3   (1   )  1 (14)
where
 hh /  vv 2 hv
X ,  (15)
(1   hh /  vv ) 2  vv   hh
Note that all functions Pi , i  1, 2, 3 , can be now calculated by polarized backscattering

coefficient  pq , p, q  v, h , not by Eq. (7) from full Mueller matrix solution. Then, the
entropy H of Eq. (9) is calculated using Pi of Eq. (14). It can be seen that the function P1 is
related with the total vv and hh co-polarized power  hh +  vv ; P2 is due to the difference
between  hh and  vv ; and P3 is due to depolarization  hv . They are modulated by the

media configuration and randomness via the parameters  and  . As the medium
becomes more random or disordered,  increases and  approaches zero. Vice versa, as
the medium becomes more ordered,  decreases and  increases. Thus, the relationship
between the entropy H and backscattering measurements  pq ( pq  vv, hh, hv) is established.
Define the co-polarized and cross polarized indexes, respectively, as
CPI  10log10 ( hh /  vv ) ( dB ) , XPI   (16)
The entropy H is directly related with backscattering indexes CPI and XPI via Eqs. (14~16).
Figure 2(a) presents an AirSAR image of total power of  vv ,  hh and  vh at the L band. The
image data from some typical areas such as the lake, an island surface, sparse trees, and
thick forest (see the frames in the figure) are chosen, as shown in the discrete points of
Figure 2(b), and are compared with theoretical results of H, CPI and XPI. The H is first
rigorously calculated by polarimetric data of the Mueller matrix, i.e. the eigenvalues of
coherency matrix C . Using the AirSAR data  vv ,  hh ,  vh and the averaged  in the
respective region, the lines are calculated with appropriate  to match the center of discrete
data.
0.85
Lake
0.80
Island
0.75 Sparse Trees
0.70 Thick Forest
0.65
Entropy H
0.60 1
0.55 2
0.50 3
0.45
0.40
0.35
0.30
4
0.25
0.20 5
0.15
0 1 2 3 4 5 6 7 8 9 10
CPI (dB)
(a) An AirSAR image (b) Entropy H vs CPI
Fig. 2. An AirSAR image of total power  vv ,  hh ,  vh at L band and Entropy H vs CPI.

Line 1:   0.257 ,   0.176 , Line 2:   0.318 ,   0.156 , Line 3:   0.365 ,   0.118
Line 4:   0. ,   0. , Line 5:   0.731 ,   0.0265
3. SAR Imaging Simulation (MPA)

Given the sophisticated mechanisms of electromagnetic (EM) wave scattering and SAR
signal collecting, it is hard to tackle SAR imagery without the aid of forward simulation of
scattering based on physical modeling. Simulation of SAR imaging presents a different

mode to study inhomogeneous terrain scene under SAR observation. It may evaluate or
predict the SAR observation, interpret the physical scattering behavior, and take account of
the parameterized characteristics of terrain surface objects. Some approaches to SAR image
simulation in previous literature might be classified into two catalogues: one focuses on the
statistical characteristics of SAR images, and the other on physical scattering process.
Natural scene is more complicated including randomly, inhomogeneously distributed,
penetrable or impenetrable objects, such as vegetation canopies, manmade structures and
perturbed surface topography. Particularly, volumetric scattering through penetrable
scatterers, e.g. timberland forests, crops, green plants etc., plays an essential or dominant
role in SAR imagery. It is meaningful to develop computerized simulation of SAR imaging
over comprehensive terrain scene with heterogeneous terrain objects.
we present an novel approach of the Mapping and Projection Algorithm (MPA) to
polarimetric SAR imaging simulation for comprehensive scenarios, which takes account of
scattering, attenuation, shadowing and multiple scattering of spatially distributed
volumetric and surface scatterers (Xu and Jin, 2006). In this approach, scattering
contributions from scatterers are cumulatively summed on the slant range plane (the
mapping plane) following geometric principles, while their extinction and shadowing
effects are cumulatively multiplied on the ground range plane (the projection plane). It
finally constructs a general imaging picture of the whole terrain scene via mapping and
projection operations. A MPA is then devised to speed up the simulation of the whole
process of scattering, extinction, mapping and projection in association with a grid partition
of the comprehensive terrain scene. Our SAR simulation scheme incorporates polarimetric
scattering, attenuation or shadowing of several typical terrain surfaces, as well as the
coherent speckle generation.
3.1 The mapping and projection algorithm (MPA)

Since the slant range is much larger than the synthetic aperture, the incidence angle to the
same target is nearly invariant during the interval of radar flying over a length of synthetic
aperture. Thus in the imaging simulation, the whole scene is divided into cross-track lines
along the azimuth dimension, and then the scattering contribution from the terrain objects
lying in each line (included in the incidence plane) are calculated sequentially. It finally
constructs a scattering picture or a scattering map of the whole scene.
Following from VRT, we derived a general expression of the scattering power at range r of
x -position in azimuth dimension as
1
 r   r0 
S ( x, r )   exp    dr  e ( x, r , )  P ( x, r , )exp    dr  e ( x, r , )  rdxdrd (17)
0  r0   r 
where the phase function P stands for scattering,  e ( x, r , ) ,  e ( x, r , ) for the backward
and forward extinction of an differential element dv at the position  x, r ,  of the imaging
space.  is incident angle. Integrating S ( x, r ) over the n, i -th pixel i.e. x   xn , xn 1  ,
s   si , si 1  , the corresponding discrete scattering map S n ,i can be obtained. Considering
polarimetric scattering, the computational discrete form of Eq. (17) for both volume and
surface scatterer is given as

mn 1 1 pi 1 1 Q 1 p  0 
S n ,i      Eo (m, p, q )S o (m, p, q )  Eo (m, p, q ) (18)
m  mn p  pi q  0 p   0 p p
vef S vol ( m, p, q ) for volume scatterer

S o ( m, p , q )   (19)
aef S surf (m, p, q ) for surface scatterer
exp   d κ e ( m, p, q)  for volume scatterer
   ef 
E (m, p, q )  
o (20)
 0 for surface scatterer
where vef   v0 is the effective volume of volume scatterer, calculated from particle number
 and volume v0 ; aef  xs is the effective area of surface scatterer; d ef is the effective
penetration depth. All of them take into account the proportion coefficients of each scatterer
in the same discrete unit. The subscript o denotes scattering or extinction of a single
scatterer. Eq. (18) presents a general description of SAR imaging which deals with single
scattering, attenuation or shadowing of all possible objects in the imaging space.
We further devise a mapping and projection algorithm (MPA) to fast compute Eq. (17) The
first step is to partition the scene and terrain objects lying in the horizontal plane xoy into
multi-grids. Each grid unit contains its corresponding parts of the terrain objects overhead,
e.g. ground, vegetation, building, etc.
As shown in Figure 3, two arrays E  and S are deployed for temporary storage of
scattering and extinction (the under-bar denotes an array). The mapping plane and
projection plane are equally divided into discrete cells corresponding to elements of S , E .
Attenuation or shadowing effect of each grid unit is counted in the increasing sequence of
y axis and cumulatively multiplied into the array E .
SI1
SIk
y SIK
M apping Plane ( Array S )
volume
scatterer
Multiple
L1 Scattering
Lk r Object SIef
LK
r Path
r Grou nd
r ef Rr
surface
O scatterer
EIK EIk EI1EIGround EIObject y

subsidied
Proje ction Plane ( Array E )
R r sin
scatterer
Fig. 3. The MPA algorithm for natural media scene.

The MPA calculating sequence of the terrain objects in one grid unit is: first to calculate the
scattering power of the current terrain object and accumulate onto the array S , where the
attenuation or shadowing from other terrain objects are obtained from array E ; then, to
calculate the attenuation or shadowing caused by the current terrain object itself, and
cumulatively multiply onto the array E ; ultimately when all grid units of the current line
are counted, the scattering map is exactly produced in the array S .
As shown in Figure 3, a vertical segment (length L ) of volume scatterers above the current
grid unit is firstly divided into K sub-segments Lk in order to make each sub-segment
mapping into one cell SI k in mapping plane, while its mid-point is projected to cell EI k in
projection plane. After the mapping and projecting, the elements of array S are refreshed as
k
S[SI k ] : S[SI k ]  E [ EI k ]  S o  E [ EI k ], k  1,.., K (21)
k
S o is the scattering of the k -th sub-segment. Assuming that volume scatterers always
uniformly fill the whole sub-segment, then the effective volume in Eq. (19) is vef  xyLk f s .
Therefore, we have
k
S o  xyLk f s S vol  vef S vol (22)
where f s is the fractional volume of scatterers. Otherwise, it is necessary to estimate the

total number of scatterers  within the sub-segment, and calculate scattering from Eq. (19).
The next step is to refresh the elements of the array E  as
E  [ EI ] : Eo, k  E [EI ] k  1,.., K

 k k
 
(23)
E [ EI k ] : E [EI k ]  Eo,k k  1,.., K
 

where Eo, k  exp(d ef κ e ) is the attenuation of the k -th sub-segment. Due to discrete
calculation for attenuation, the array E  should store the averaged attenuation. Thus, the
ratio of shadowed area of the sub-segment to the whole projection cell should be taken into
account, when counting the contribution of the sub-segment to the mean attenuation. The

effective depth in Eo, k is calculated as
vef xyLk xyLk

d ef    tan  (24)
Rx Rr tan  Rx Rr tan  Rx Rr
It can be seen from the effective depth d ef that if we redistribute all volume scatterers of the
sub-segment and make them cover the whole projection cell under the invariance of density,
then the layer depth of redistributed scatterers along the projection line is seen as the
effective depth.
In a similar way, larger surface scatterers, like vertical wall surfaces, cliff slopes etc., are
segmented according to mapping cells. But, for nearly horizontal surfaces, e.g. ground or
roof surface, the facet cut by one grid unit is small enough and is projected into one
projection cell, i.e. K  1 . Hence, given the normal vector of facet n̂ , the effective area and
its scattering in Eq. (21) are calculated in two ways:
k xySsurf cos 2 cos 1 K 1

S o  aef Ssurf   (25)
 xLk Ssurf cos 2 sin 1 K 1
where 1 is the angle between n̂ and z axis; 2 is the angle between n̂ and iˆ incidence.
Lk is the length of sub-segment. Moreover, the array E  should be set to zero over all
projection interval from the first projection cell to the last one, i.e.
E [ j ] : 0 EI1  j  EI K (26)
The elements of the array E  are initialized to unit matrices at the start point of calculation
for each line, while the array S should be saved before cleared to start the next line. Note
that each line is strictly computed in the sequence of increasing y in order to precisely
count attenuation caused by the terrain objects ahead.
Double and triple scatterings between terrain objects and the underlying ground are
regarded as shown in Figure 3. First, the corresponding ground patch of multiple scattering
can be located by the ray tracing technique. Then the length of propagation path, i.e. the
effective range ref can be given as
 rObject  rGround  rPath  2 for double scattering

ref   (27)
rObject  rPath for triple scattering
In the same way, multiple scattering of a segment of scatterers is mapped into the array S
by ref . The same step of dividing into sub-segments is employed if the mapping interval
exceeds one cell.
For double scattering, the attenuation or shadowing suffered through the propagation paths
of rObject , rGround are the same with single scattering of the terrain object and ground patch,
respectively. It means that the array E  can be used as well. The attenuation during the way
from the terrain object to ground patch rPath , is omitted for simplicity. Therefore, the double
scattering contribution can be written as
2+ 2 2 2
S[SI ef ] : S[SI ef ]  E [ EI Ground ]  S Ground   S Object  E [ EI Object ]  E  [ EI Object ]   S Object  S Ground  E [EI Ground ] (28)
where EIGround , EIObject are the projection indices of the terrain object and the ground patch
2 2
respectively. SGround , S Object are the Mueller matrices of the ground patch and the terrain
object along the forward and backward directions of double-scattering, respectively. The
coefficient   vef or aef is calculated similarly to Eqs. (19,22,25).
For triple scattering (object-ground-object), the expression can be directly written as:
2 3 2
S[SI ef ] : S[SI ef ]  E  [ EI Object ]  S Object  S Ground   S Object  E [EI Object ] (29)
3
where SGround is the Mueller matrix of ground patch along the triple scattering path;   vef
or aef .
Correct sequence for calculation of each grid unit is first to count single scattering, then
multiple scattering, at last its attenuation or shadowing. Notice that the ground patch of
multiple scattering could be located at any place around the terrain object. For simplicity,
we take the projection cell in the current incidence plane as an approximate substitute. Some
other coding techniques are adopted to guarantee the correct computation and make it more
efficient.
3.2 Scattering models for terrain objects

For vegetation canopies, VRT model of a layer of random non-spherical particles (Jin, 1994)
is employed. Leaves, small twigs or thin stems are modeled as non-spherical dielectric
particles under the Rayleigh or Rayleigh-Gans approximation, while branches, trunks or
thick stems are modeled as dielectric cylinders. A tree model is composed of crown and
trunk. The crown is a cloud with a simple geometrical shape containing randomly oriented
non-spherical particles. The trunk is an upright cylinder with the top covered by the crown.
Oblate or disk-like particles and elliptic crown are adopted for broad-leaf forest, while
prolate or needle-like particles and cone-like crown are for needle-leaf forest. In addition,
the crown shapes take a small random perturbation. Similarly, a farm filed is modeled as a
layer of randomly oriented non-spherical particles with perturbed layer depth. If the grid
units are small enough, the segment of crown or crops cut by one grid unit can be regarded
as a segment fully filled with particles. Differences between the trunk and crown are: (a) the
trunk is solid within a grid unit i.e. fractional volume is set to 1; (b) double scattering
between the trunk and ground is taken into account.
Scattering of the building is seen as the surface scattering from its wall and roof surfaces and
multiple interactions with the ground surface. The integral equation method (IEM) is
employed to calculate surface scattering. The building is modeled as a parallelepiped with
an isoscales triangular cylinder layered upon it. Due to the orientation of wall surface and
roof surface, the incident angle must be transferred into the local coordinates of the surface,
as well as the polar basis of wave propagation. Geometrical relationships among the wall
surfaces, the roof surfaces, as well as the double and triple scatterings between the wall and
ground can be found in Franceschetti et al. (2002). Figure 2 illustrates the building model, its
image and shadowing in SAR imagery.
incide nce
H
r
A
o
B C O
D
E
Hctg F
G
Htan y
x
images on the line
A-B : Roof Scattering + Wall Single Sca ttering
B-C : Wall Single Scattering
D : Wall / Ground Double Scattering
C-F : Wall / Ground Triple Scattering (no ground scattering)
F-G : Shadow (no ground scattering)
Fig. 4. Image and shadow of a building in SAR imagery
For ground facet, the tangent vectors along x , y axes on the current grid unit are used to
construct the facet plane.
The MPA approach is based on the VRT theory for incoherent scattering power, not
coherent summation of scattering fields. However, speckle due to coherent interference is
one of the most critical characteristics of SAR imagery, especially for studies of SAR image
filtering and optimization. Assuming Gaussian probability distribution of the scattering
vector k L   S hh , Shv , Svh , Svv  , random scattering vector is then generated as follows:
T
1
k L  k 0  C2 , k 0i  I i  jQi , i  1,.., d (30)
where I i , Qi are independent Gaussian random numbers with the zero mean and unit
variance. Given the positive semidefinite property of C , its square root can be obtained by
Cholesky decomposition.
3.3 Simulation results

The configurations and parameter settings of the radar and platform in simulation are
selected follow the AIRSAR sensor of NASA/JPL.
A virtual comprehensive terrain scene is designed, as shown in Figure 5(a), as according to a
true DEM of Guangdong province, south China. It contains different types of forests
covering on the hill, crops farmland, ordered or random buildings in urban and suburban
regions, roads and rivers.
A simulated scattering map (decibel of normalized scattering power, from -50dB to 0dB,
pseudo color: R-HH, G-HH, B-HV.) at L band with 12m resolution (pixel spacing is 5m) is
displayed in Figure 5(b). The simulated SAR image at L band is shown in Figure 5(c).
Figures 6(a,b) give the simulated scattering map and SAR image at C band, respectively. In
this higher band, vegetation scatterings are significantly increased, as well as attenuation is
increased and wave penetrable depth is decreased. It causes that trees become thinner while
the shadowing becomes darker.
2000
river
1800 build.
tree
1600
soil
1400
1200
a z im u t h (m )
1000
800
600
400
200
200 400 600 800 1000 1200 1400 1600 1800 2000
range(m)
(a) Designed Scene (b) Scattering coefficient map (c) simulated image of total power
Fig. 5. Simulated SAR image at L band
(a) Scattering coefficient map (b) Simulated image of total power

Fig. 6. Scattering and Simulated SAR image at C bands
At the upper-right corner, the blocks images appear like parallel lines, which reveal the
dominance of double scattering in the urban area. However, on the other side below the
river, the buildings are oriented nearly 45 degrees to the radar flying direction. As a result,
the cross-pol scattering (blue) is stronger, while double scattering is reduced. Timberlands
can be classified into two classes: broad leaves in yellow color due to balance between HH
and VV scattering, and narrow leaves in green color due to weaker scattering of HH. The
narrow leaves trees make weaker attenuation of HH and therefore stronger HH scattering of
the shadowed ground, which is the reason why most areas of narrow leaves forest become
red. Overlay and shadowing effects caused by mountainous topography are particularly
perceptible. Croplands are generally uniform and dense, and appear like patches in the
image. Its VV scattering (green) is always stronger than HH. They have distinguishable
brightness due to different density, sizes, shapes etc. More detail discussion can be seen in
(Xu and Jin, 2006)
3.4 Bistatic SAR imaging

Bistatic SAR (BISAR) with separated transmitter and receiver flying on different platforms has
become of great interests. Most of BISAR research is focused on the engineering realization
and signal processing algorithm with few on land-based bistatic experiments or theoretical
modeling of bistatic scattering. The MPA provides a fast and efficient tool for monostatic
imaging simulation. It involves the physical scattering process of multiple terrain objects,
such as vegetation canopy, buildings and rough ground surfaces (Xu and Jin, 2008a).
Tx
z
Rx p v-
p v+
dv
The mapping track - m v

Upward Projection
m
pu
Downward Projection
pd
Array S y
Array E 
x
Fig. 7. Bistatic MPA
Similar to mono-static MAP, the bistatic-MAP steps are described as follows,

(1) The arrays E , E are initialized as unit matrices, S is as zero matrices. Then, visit
the grid units, sequentially, along  y direction.
(2) Perform 3D projections of the current grid unit along incidence and scattering
directions to the cells pd , pu of E , E , respectively.
(3) Determine the mapping position of the current grid unit based on the information of its
synthetic aperture and Doppler history, which corresponds to the cell m of the array S .
(4) Obtain or calculate the scattering matrix S 0 and the upward/downward attenuation

matrix E0 of the current grid unit.
(5) Refresh the elements of S as
S[m] : S[m]  E  [ pu ]  S 0  E [ pd ] (31)

Refresh the elements of E , E as
 
E [ pd ] : E0  E [ pd ], E  [ pu ] : E  [ pu ]  E0 (32)
(6) Return to Step (2), and continue to visit the next grid at the same coordinate of y or,
if all grids at this coordinate have been visited, step forward to a larger coordinate of y
till the whole scene is exhausted.
Figures 7,8 explain the bistatic imaging process of a tree and a building. One of the
differences between bistatic and monostatic cases is the shadowing area projected in both
incidence and scattering directions. Additionally, due to the split incidence and scattering
directions, double scattering terms of object-ground and ground-object are different. One
reflects at the ground and diffusely scatters from the object, the other scatters at the object
and then reflects from the ground. These two scattering terms have different paths which
give rise to separated double-scattering images.
Fig. 8. Composition of a tree and a house in BISAR image
A comprehensive virtual scene as shown in Figure 9(a) is designed for simulation and
analysis. Simulated BISAR images of the configurations, AT (across track) and TI
(translational invariant) are shown in Figures 9(b,c), respectively.
(a) Ground truth (b) AT span/dB (c) TI span/dB

Fig. 9. Simulated BISAR images of AT and TI
3.5 Unified bistatic polar bases and polarimetric analysis

It is found in the BISAR simulation that in AT BISAR image, bistatic scattering of terrain
objects preserves polarimetric characteristics analogous to monostatic case. while the major
disparity is on total scattering power.
However, the polarimetric parameters behavior is different in the image of general TI mode.
First, H is generally higher, probably due to the fact that large bistatic angle is more likely
to reveal the randomness and complexity of the object, and makes the scattering energy
more uniformly distributed on different polarizations. Second, almost all terrain surfaces
show   45 . It is found that scattering energy is concentrated on the 4-th component of
Pauli scattering vector k P all over the image.
It seems that  might lose its capability to represent the polarimetric characteristics under
certain bistatic configuration. Generally speaking, all parameters  ,  ,  in monostatic SAR
image cannot reflect polarimetric characteristics in the general TI case.
Conventional polarimetric parameters such as  ,  ,  may largely depend on angular
settings of the sensors rather than instinct properties of the target. It becomes inconvenient
to interpret the physical meaning of polarimetric parameters when they are severely
involved with the bistatic configuration.
As proposed in many studies, inverse problems of bistatic scattering are usually discussed
in a coordinates system determined by the bisectrix. We believe it is important to first
transform the conventional bistatic polar bases to a new one defined by the bisectrix. The
unified bistatic polar bases are defined as
bˆ  kî kˆ  kî
hî  , vî  hî  kî , bˆ  s (33)
ˆ
b  kiˆ kˆs  kî
where b̂ denotes the bisectrix, which is defined as the bisector of the incident and scattered
wave vectors in the plane, as shown in Figure 10(a).
hi z
b hi
vi
ki
ks hs
hs
vs
y
x
(a) Unified bistatic polar bases (b) Bistatic (TI) image
Fig. 10. Unified bistatic polar bases and bistatic (TI) pseudo color images coded by NPC after
unified bistatic polar bases transform
The relationship between the unified bistatic polar bases and the conventional polar bases
can be described by polar basis rotation, i.e. E, S defined in (vî , hî , kî ) are re-defined as
E, S in (vî, hî, kî ) . It is written as
T  vî  vî hî  vî 

Ei  U i  Ei , S  U s  S  U i , Ui    (34)
ˆ ˆ ˆ
vî  hi hi  hi
where U i is the rotation matrix. Figure 10(b) gives the bistatic TI image after transfom of Eq.
(34). The normalized Pauli components (NPC) are defined as Pi | k P ,i | / | k P |, i  1,.., 4 . The
appearance of the NPC color-coded image after transform agrees with the convention of
human vision.
We modify the definition of Cloude’s parameters as
 cos  cos  exp( j1 ) 

sin  cos  exp( j ) 
kP  kP  2 
(35)
 sin  sin  exp( j3 ) 
 
 cos  sin  exp( j4 ) 
The redefined  ,  ,  are calculated and plotted in Figure 11 for bistatic image of Figure
10(b). It can be seen that the redefined  ,  ,  are more capable to represent or classify the
polarimetric characteristics of different scattering types.
Redefinition of Cloude’s parameters is only a preliminary attempt of bistatic polarimetric
interpretation. It remains open for further study in a systematic way and for verification
using both simulation and experiments.
(a) redefined  (b) redefined  (c) redefined 

Fig. 11. Redefined  ,  ,  after bistatic polar bases transform
The MPA is also applied to simulation of SAR image of undulated lunar surface (Fa and Jin,
2009). Based on the statistics of the lunar cratered terrain, e.g. population, dimension and
shape of craters, the terrain feature of cratered lunar surface is numerically generated.
According to inhomogeneous distribution of the lunar surface slope, the triangulated
irregular network is employed to make the digital elevation of lunar surface model. The
Kirchhoff approximation of surface scattering is then applied to simulation of lunar surface
scattering. The SAR image for cratered lunar surface is numerically generated. Making use
of the digital elevation and Clementine UVVIS data at Apollo 15 landing site as the ground
truth, an SAR image at Apollo 15 landing site is simulated.
4. Terrain Surface Classification using De-Orientation Theory

Classification of complex terrain surfaces using polarimetric SAR imagery is one of most
important SAR applications. An unsupervised method based on the entropy H and target
decomposition parameters has been well developed by Cloude et al., which extracts the
target decomposition parameter  and entropy H from eigen-analysis of coherence
matrix and construct an unsupervised classification spectrum on the   H plane.
Scattering of the terrain targets functionally depends on the scatter orientation, shape,
dielectric property, and scattering mechanism etc. Scatter targets of complex terrain surfaces
are often randomly oriented and cause randomly fluctuating echoes. It is difficult to make
classification of randomly oriented and randomly distributed scatter targets. Different
scatters with different orientations can make the similar scattering, and vice versa, the same
scatters with random orientation can make different scattering to make confused
classification.
In this section, a transformation of the target scattering vector to rotate the target along the
sight line is derived. Deorientation is introduced to transform the target orientation into
such fixed state with minimization of cross-polarization (min-x-pol). And meanwhile, the
angle is extracted to indicate the angle deviation of the target orientation from this min-x-
pol state.
A set of new parameters u , v, w from the target scattering vector is defined to indicate the
ratio and phase difference of the two co-polarized (co-pol) backscattering terms, and the
significance of cross polarized (x-pol) term. All these parameters as well as the entropy H
are applied to classification of random surface targets.
Numerical simulations of polarimetric scattering of a single small non-spherical particle,
and a layer of random non-spherical particles above a rough surface are studied to show the
effectiveness of the parameters u , v, , H and the capability of u , v, H for classification of
complex terrain surfaces (Xu and Jin, 2005).
As examples, the terrain surface classifications for a SIR-C and an AirSAR images are
presented.
4.1 De-orientation and parameterization

From Eq. (1), the scattering vector is usually defined as
k P  [ S hh  Svv , S hh  Svv , 2 S x ]T / 2 , or k L  [ S hh , 2S x , Svv ]T (36)
where S x  S hv  Svh . Cloude et al. defined the parameterization of k P in terms of the

parameters  ,  etc. as
k P | k P | [cos  e j1 ,sin  cos  e j2 ,sin  sin  e j3 ]T (37a)
similar parameterization of k L is defined as
k L | k L | [sin c cos ae j0 ,cos c e jx ,sin c sin ae j (0  2b ) ]T (37b)

It yields
1
a  tan 1 (| Svv | / | S hh |), b  arg( Svv / S hh ), c  cos 1 ( 2 | S x | / | k L |) (38)
2
Rotating the angle  of the polarization base along with the sight line, the electrical field
vector E becomes E as
cos  sin 
E  [ R ]  E, [ R]   (39)
 sin cos 
In backscattering direction, this rotation makes the scattering vector k P to be k P , which can
be expressed as
1 0 0 
k P  [U ]  k P , [U ]   0 cos 2 sin 2  (40)
 0  sin 2 cos 2 
Applying minimization of cross polarization (min-x-pol) to k P , i.e. let
 | k P ,3 |2  2 | k P ,3 |2
 0, 0
  2
it yields the deorientation angle  m is obtained as
 
 m  sgn cos(2  3 )   (41)
 2 
2
It means that such rotation of the angle  m along the sight line makes k P to the min-x-pol
status as deoriented k dP .
New parameters u , v, w are then defined from the deoriented scattering vector of Eqs. (38,40)
as:
u  sin c cos 2a, v  sin c sin 2a cos 2b, w  cos c . (42)
In the case of non-deterministic targets, we obtain the uncorrelated scattering vectors, i.e.
eigenvectors, through eigen-analysis of coherency matrix. Here the most significant
eigenvector i.e. principal eigenvector is considered as the representative scattering vector of
non-deterministic targets. Thus, the deorientation of non-deterministic targets is merely
conducted on the principal eigenvector.
4.2 Numerical simulation

Based on the model of Figure 1, polarimetric scattering is calculated, and the scattering
vector and Mueller matrix are obtained. A model of a layer of random non-spherical
particles above a rough surface for scattering from terrain surfaces, as shown in Figure 1, is
applied to numerical simulation of u , v, H at L band in various cases.
Figures 12(a,b) show numerical relationship between the parameters u , v and the particle
parameters: Eular angle orientation and particle shape (oblate or prolate spheroids). It can
be seen that | u | indicates the non-symmetry of particle’s projection along with the sight line,
and the dielectric property of the particle, |  s | , affects | u | , and  s /  s affects v .
Figure 12(a) presents the distributions of some typical terrain surfaces on the | u |  H plane,
which are obtained from the scattering model of Figure 1. It can be seen that H indicates
the complexity of layer structures of terrain surfaces. Different dielectric property of the soil
land and oceanic surface also makes u identifiable on the | u | axis.
Figure 12(b) presents the distribution of scattering terms on the plane | u | v , where M i
denotes the i-th order scattering. Concluded from these figures, H indicates the canopy
randomness, u is useful for distinguishing different terrain targets and v is helpful to take
account of scattering mechanisms of different-orders.
Soil Ground
1.0 M
Sea Surface 1.0   
   M0
Vegetation Type1
 M1
Vegetation Type2
0.8  M2
Forest
0.5  M3
More Rough More Thrive

0.6
More Random
More Dense 
H
0.0
v
0.4
-0.5
0.2 
-1.0
0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
|u| |u|
(a) Typical surfaces on |u|-H plane, (b) Scattering mechanism on |u|-v plane
Fig. 12. Distributions of typical surfaces and different scattering mechanism
4.3 Surface classification

The flow-chart of the classification method of PolSAR Data based on deorientation and new
parameters u , v, is shown in Figure 13. The classification decision tree of parameters
u , v, H is displayed on Figure 14.
Pol-SAR Eigenanalysis of
Data Eigenvalues i H
Coherency Matrix
Target principal
scattering vector u
Orientation
Classification Analysis
Deorientation v
Fig. 13. Flow-chart of the surface classification.

H < 0 .8 A N D
(v < -0 .2 O R v >0 .2 )
Y
N
T im b er N
v<0
Y
H < 0 .5 |u|> 0 .7
Y
N
Y
N
|u |< 0 .7 |u |< 0 .7 Ur ban 2 U rb a n 1
Y
N
N
Y
C a nop y2 C an op y1 S u rf ace 1 |u|< 0 .3

Y
N
S u rf ace 2 S u rf ace3
Fig. 14. Classification decision tree of parameters u , v, H .
An AirSAR data at L band of the Boreal area in Canada with rich resources of vegetation is
chosen for classification and orientation-analysis, as shown in Figure 15(a). Figure 15(b) is
the deorientation classification over this area. Terrain surfaces are divided into 8 classes
following the decision-tree: Timber, Urban 2, Urban 1, Canopy 2, Canopy 1, Surface 3,
Surface 2 and Surface 1.
(a) A SIR-C SAR polarimetric image, (b) Classification using the parameters u, v, H
Fig. 15. A SIR-C SAR image and surface classification
Figure 16 shows the data distributions on the planes | u |  H for parameter v  0.2 and
v  0.2 , and corresponding classification.
v>0.2 v<-0.2
1.0 1.0
sea
0.9 0.9 soil
timberland
0.8 0.8 urban area
 suburban
0.7 0.7
 shore
0.6 0.6
H
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0.0 0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
|u| |u|
Fig. 16. Classification by u , v, H for 19 classes
The orientation distribution of several classes are selected to be displayed in Figures 17(a,b).
It can be seen that
(a) While in the Urban 1 (sparse forest of vertical trunks, instead of artificial
constructions), there are uniformly vertical orientations indicating orderly trunks.
(b) Random orientation in the Canopy 1 means that the vegetation canopy in this area
might be the disordered bush. Note that the region with the roads inside the forest might
show randomness confused by the bush vegetation on the roadsides.
Following the above orientation analysis, the terrain surfaces are further classified into
the subclasses and are identified by their types.
A
B
Timber Urban1
(a) Classificaction of AirSAR data, Angle  distribution: (b) timber (c) urban
Fig. 17. Classification of AirSAR data, Boreal area.
4.4 Faraday rotation and Surface classification

To obtain the moisture profiles in subcanopy and subsurface, it needs UHF/VHF bands
(435MHz, 135MHz) to penetrate through the dense subcanopy (~20kg/m2 and more) with
little scattering and reach subsurface. However, when the wave passes through the
anisotropic ionosphere and action of geomagnetic field, the polarization vectors of
electromagnetic wave are rotated, which is called the Faraday rotation (FR). The FR depends
on the wavelength, electronic density, geomagnetic field, the angle between the direction of
wave propagation and geomagnetic field, and the wave incidence angle.
Assuming that the propagation direction is not changed passing through homogeneous
ionosphere, ds  seci dz (where ẑ is the normal to the surface), and geomagnetic field keeps
constant as one at 400 km altitude, FR is simply written as
 F  2620 e B(400) 2 cos  B seci (radians) (43)
where  B is the angle between the directions of electromagnetic wave propagation and
geomagnetic field,  e is the total electron content per unit area ( 1016 electron/m2), and
B (400) is the intensity of geomagnetic field (T) at 400 km altitude.
The FR may decrease the difference between co-pol backscattering, enhance cross-polarized
echoes, and mix different polarized terms. Thus, the satellite-borne SAR data at low
frequency becomes distorted due to FR effect. Since the FR is proportional to the square
power of the wavelength, it yields especially serious impact on the SAR observation
operating at the frequency lower than L band. The FR angle at P band can reach dozens of
degrees.
As a polarized electromagnetic wave passes through the ionosphere, the scattering matrix
F
S with FR (indicated by superscript F ) is written by the scattering matrix S without FR as
follows
 ShhF S hvF   cos  sin    S hh S hv   cos  sin  
 F F
    (44a)
 Svh Svv    sin  cos    S hv Svv    sin  cos  
S hhF  Shh cos 2   Svv sin 2 , SvvF   S hh sin 2   Svv cos 2 
(44b)
S hvF  S hv  ( S hh  Svv )sin  cos , SvhF  S hv  ( S hh  Svv )sin  cos 
F F
The measured polarimetric data with FR, S or M , are distorted and cannot be directly
applied to terrain surface classification. However, fully polarimetric 4×4-D M without FR
F
can be inverted from M . And it shows that the remained  / 2 ambiguity error does not
affect the classification parameters: u , v, H , , A . Based on intuitive assumption of gradual
change of FR degree with geographical position, a method of 2D phase unwrapping method
with random benchmark can be employed to eliminate the  / 2 ambiguity error is
designed. Some example shows the fully polarimetric SAR without FR at low frequency,
such as P band, can be fully inverted (Qi and Jin, 2007).
5. Change Detection of Terrain Surface from Multi-Temporal SAR Images

Multi-temporal observations of SAR remote sensing imagery provide fast and practical
technical means for surveying and assessing such vast changes. One direct application is to
detect and classify the information on changes in the terrain surfaces. It would be laborious
to make an intuitive assessment for a huge amount of multi-temporal image data over a vast
area. Such assessment based on qualitative gray-level analysis is not accurate and might lose
some important information. How to detect and automatically analyze information on
change in the terrain surfaces is a key issue in remote sensing.
In this section, two-thresholds EM and MRF algorithm (2EM-MRF) is developed to detect

the change direction of backscattering enhanced, reduced and unchanged regimes from the
SAR difference image (Jin and Luo, 2004, Jin and Wang, 2009).
On May 12, 2008, a major earthquake, measuring 8.0 on the Richter scale, jolted
southwestern China’s Sichuan Province, Wenchuan area. To evaluate the damages and
terrain surface changes caused by the earthquake, multi-temporal ALOS PALSAR image
data before and after the earthquake are applied to detection and classification of the terrain
surface changes. Using the tools of Google Earth for surface mapping, the surface change
situation after the earthquake overlapped the DEM topography can be demonstrated in
multi-azimuth views as animated cartoon. The detection and classification are also
compared with the optical photographs. It is proposed that multi-thresholds EM and MRF
analysis may become traceable when multi-polarization, multi-channels, multi-sensors
multi-temporal image data become available.
5.1 Two thresholds expectation maximum algorithm

Consider two co-registered images X 1 and X 2 with size I  J at two different times (t1 , t2 ) .
Their difference image is X  ( X 2  X 1 ) , and denote X D  { X (i, j ), 1  i  I ,1  j  J } (in
some cases, X  X 2 / X 1 also can be used).
As usually, one-threshold expectation maximum (EM) algorithm has been employed to
classify two opposite changes: the unchanged class n and the changed class c . The
probability density function P ( X ), X  X D was modeled as a mixture distribution of two
density components associated with n and c , i.e.
p( X )  p ( X | n ) P(n )  p ( X | c ) P(c ) (45)
Assume that both the conditional probabilities p ( X | n ) and p ( X | c ) are modeled by

Gaussian distributions. The iterative equations for estimating the statistical parameters and
the a priori probability for the class n are the following
P t (n ) p t ( X (i, j ) | n )

X ( i , j ) X D p t ( X (i, j ))
P t 1 (n )  (46a)
IJ
P t (n ) p t ( X (i, j ) | n )
 X (i, j )
X ( i , j ) X D p t ( X (i, j ))
mean  nt 1  (46b)
P t (n ) p t ( X (i, j ) | n )

X ( i , j ) X D p t ( X (i, j ))
P t (n ) p t ( X (i, j ) | n )
 [ X (i, j )   nt ]2
X ( i , j ) X D p t ( X (i , j )
variance ( n ) 
2 t 1
(46c)
P t (n ) p t ( X (i, j ) | n )

X ( i , j ) X D p t ( X (i, j ))
where the superscripts t and t +1 denote the current and next iterations.
Analogously, these equations can also be used to estimate p (c ) with c and  c2 .
The estimates are computed starting from the initial values by iterating the above equations
until convergence. The initial value of the estimated parameters can be obtained by the
analysis of the histogram of the difference image. A pixel subset S n likely to belong to n
and a pixel subset Sc likely to belong to c can be obtained by preliminarily selecting two-
threshold Tn and Tc on the histogram. This is equivalent to solving the ML boundary To for
the two classes n and c on the difference image. The optional threshold To requires
P(c ) p( X | n )
 (47)
P (n ) p ( X | c )
Under the assumption of Gaussian distribution, Eq. (3) yields
  P (n ) 
( n2   c2 )To2  2(  n c2  c n2 )To  c2 n2   n2 c2  2 n2 c2 ln  c 0 (48)
  n P (c ) 
to obtain optimal To .
Figures 18 (a,b) are the ALOS PALSAR images (L band, HH polarization) in February 17
and May 19, 2008 before and after earthquake in Beichuan County, Wenchuan area,
respectively. Spatial resolution is 4.7m4.5m. Figure 18(c) is the difference image of Figures
18(a,b), which seems very difficult to evaluate the terrain surface changes only using man’s
vision, especially to accurately classify the change classes in large scale.
(a) 17 Feb 2008 before earthquake, (b) 19 May after earthquake, (c) Difference image
Fig. 18. ALOS PALSAR image in in Beichuan
The pixels of the difference image (i.e. X D   D0   20   10 ) are classified into three classes:
c1 of  D0 -enhanced, n of  D0 -unchanged and c 2 of  D0 -reduced. Thus, the probability
density function P ( X ) is modeled as a mixture density distribution consisting of three
components:
p( X )  p ( X | c1 ) P(c1 )  p( X | n ) P(n )  p( X | c 2 ) P (c 2 ) (49)
The parameter To1 is first obtained by application of the EM algorithm to the enhanced class
c1 and no-enhanced class n1 ( n and c 2 ). Then, we obtain To 2 by applying the EM to
the reduced class c 2 and no-reduced class n 2 ( n and c1 ), where n1  n  c 2 ,
n 2  n  c1 . Finally, the two results are superposed to get the final classification.
0.030
P(n)p(X|n)
0.025
Grayscale probability density
0.020
0.015 P(c2)p(X|c2)
P(c1)p(X|c1)
0.010
0.005
0.000
-20 -15 -10 -5 0 5 10 15
Grayscales after the quantification of the difference image (dB)
(a) The histogram of the difference image from Fig. 18(c) (b) Change detection by 2EM
Fig. 19. The histogram of the difference image and change detection using 2EM method
As an example, Figure 19(a) shows the grayscale histogram of a small part chosen from the
difference image, Figure 18(c). Three classes of the probability distributions are outlined in
this figure. The histogram is normalized in accordance with the probability density
distributions.
The initial statistical parameters of the subsets of
S n1  { X (i, j ) | 0  X (i, j )  Tn1} , Sc1  { X (i, j ) | X (i, j )  Tc1 (50a)
are first derived, and then derive the parameters of
S n 2  { X (i, j ) | Tn 2  X (i, j )  0} , Sc 2  { X (i, j ) | X (i, j )  Tc 2 } (50b)
The initial statistical parameters are related to the classes n1 , c1 and the classes n 2 , c 2 ,
respectively. Then, Eqs. (46a-c) are sequentially used to perform the iterations on the four
classes, i.e. the a priori probability and other statistical parameters [ P (n1 ) ,  n1 ,  n21 ],
[ P (c1 ) , c1 ,  c21 ], [ P (n 2 ) ,  n 2 ,  n22 ] and [ P (c 2 ) , c 2 ,  c22 ] . Solving Eq. (48), the thresholds
To1  3.1643 dB and To 2  2.8381 dB are obtained.
Figure 19(b) is the change detection using EM algorithm, where red color indicates the area
with scattering enhanced, blue color indicates the area with scattering reduced, and green
color denotes no-changed. It makes three classes change of the difference image.
5.2 The change detection using 2EM-MRF

Actually, the pixels of the image are spatially correlated, i.e. a pixel belonging to the class
k is likely to be surrounded by pixels belonging to the same class. To take account of
spatial texture may yield more reliable and accurate change detection.
Let the set C  C ,1    L represent the possible sets of the labels in the difference image:
C  C (i, j ),1  i  I ,1  j  J  ， C (i, j )  c1 , c 2 , n  (51)
where L  3IJ . According to Bayes rule for minimum error, the classification result should
maximize the posterior conditional probability,
Ck  arg max P(C | X D )  arg max  P(C ) p ( X D | C ) (52)

C C C C
where P (C ) is the prior model for the class labels, and p ( X D | C ) is the joint density
function of the pixel values in the difference image X D given the set of labels C .
The MRF algorithm is employed with a spatial neighborhood 55 pixels system to take
account of spatial texuture. The generation of the final change-detection map involves the
labeling of all the pixels in the difference image so that the posterior probability of Eq. (52) is
maximized.
The MRF algorithm is equivalent to the minimization of the Gibbs energy function. The
MRF algorithm is iteratively carried out.
Figure 20(a) presents the final result of 2EM-MRF for detection and classification of the
terrain surface changes from the difference image, Figure 18(c). The numbers 1~8 indicate
some areas with typical changes.
It can be seen that, for example, in the area 1 the river was largely blocked up with
landslide; in the area 2, there were landslides causing large scale blocks; the area 3 is
Beichuan town, where the terrain surface was significantly undulated or roughed due to
landslide and building collapse; in the areas 4 and 5 the river was blocked up due to
landslides along river lines; in the area 6 the highway was significantly blocked up with
landslide; in the area 7 there was collective landslides almost towards one azimuth
direction, and the flat area with reduced scattering along the river might be due to seasonal
risen water in May than February; in the area 8 collapse of mountain blocks caused the
terrain surface undulating to enhance stronger scattering.
Figure 20(b) gives the topography and contour lines of the same area from the google
website. It is useful to locate where and which kind of the changes are happening. It might
be seen that the landslides, e.g. in the areas 1 and 7, are correlated with direction and
magnitude of slopes. A quantitative analysis to assess the landslides and surface damage
can be further developed.
1
N
B
7 1 5
5
A
4 4
8
3 3
2 2
B
A
6
6
(a) Change detection of 2EM-MRF,(b) The topography, (c) Comparison with photograph
Fig. 20. Automatic detection and classification of terrain surface changes using 2EM-MRF
As an example, Figure 20(c) gives a comparison of the change at the region 3 with a
photograph distributed to public from the website of the Ministry of the State Resources of
China. It can be clarified in both figures that A shows lanslides and makes the surface
smooth, and B shows the highway blocked. It can be seen that optical shadowing actually
does not confuse these classifications.
Using the tool of Google Earth mapping with the 2EM-MRF results, the terrain surface
change situation classified by three types overlapped the DEM topography can be showed
in multi-azimuth views as a animated cartoons. Since all process of 2EM-MRF are automatic
and carried out on real time, it should be helpful, especially for commanding the rescue
works in disaster scene. Figure 21 shows four views.
It can be seen that information retrieval from a single amplitude image (scattered power)
with one-frequency, mono-polarization at one time is very limited. In fully polarimetric
technology, for example, using target decomposition theory, deorientation transform etc.,
the physical status of the terrain surfaces such as vegetation canopy, road, building, river etc.
can be well classified and would be of greatly helpful to change detection. Four Stokes
parameters of polarimetric measurement can be feasible to show the surface slope variation
and anisotropy, and INSAR has been well applied to retrieval of surface digital elevation.
All of these progress show superiority over a single amplitude image analysis and manual
vision estimation. It is also helpful to fuse CFAR (constant false alarm rate) for detection of
the object, e.g. edge and block, from a SAR image.
As multi-polarized, multi-channels and multi-temporal image data become available, 2EM-
MRF can be further extended to utilization of principal characteristic parameters of the
difference image, such as VV+HH, VV-HH, entropy, angular  ,  ,  , for change detection
and classification. More information about terrain surface changes can be retrieved.
Fig. 21. Multi-azimuth views in animated cartoon showing terrain surface changes on
Google Earth mapping
6. DEM Inversion from a Single POL-SAR Image

Co-polarised or cross-polarised backscattering signature is the function of the incidence
wave with the ellipticity angle  and orientation angle  . Recently, polarimetric INSAR
image data has been utilized to generate digital surface elevation and to invert terrain
topography. When the terrain surface is flat, polarimetric scattering signature has the
maximum largely at the orientation angle   0 . However, it has been shown that as the
surface is tilted, the orienattion angle  at the maximum of co-polarized (co-pol) or cross-
polarized (cross-pol) signature can shift from   0 . This shift can be applied to convert the
surface slopes. Making use of the assumption of real co-pol and zero cross-pol scattering
amplitude functions, the  shift is expressed by the real scattering amplitude functions.
Since both the range and azimuth angles are coupled, two- or multi-pass SAR image data
are required for solving two unknowns of the surface slopes. This approach has been well
demonstrated for inversion of digital elevation mapping (DEM) and terrain topography by
using airborne SAR data.
However, scattering signature is an ensemble average of echo power from random scatter
media. Measurable Stokes parameters as the polarized scattering intensity should be
directly related to the  shift. In this section, using the Mueller matrix solution, the  shift
is newly derived as a function of three Stokes parameters, I vs , I hs ,U s , which are measurable
by the polarimetric SAR imagery.
Using the Euler angles transformation between the principal and local coordinates, the
orientation angle  is related with both the range and azimuth angles,  and  , of the
tilted surface pixel and radar viewing geometry. These results are consistent with Lee et al.
(1998) and Schuler (2000), but are more general.
It is proposed that the linear texture of tilted surface alignment is used to specify the
azimuth angle  . The adaptive threshold method and image morphological thinning
algorithm are applied to determine the azimuth angle  from image linear textures. Thus,
the range angle  is then solved, and both  and  are utilized to obtain the azimuth
slope and range slope. Then, the full multi-grid algorithm is employed to solve the Poisson
equation of DEM and produce the terrain topography from a single pass Polarimetric SAR
image (Jin, 2005; Jin and Luo, 2003).
6.1 The  shift as a function of the Stokes parameters

Consider a polarized electromagnetic wave incident upon the terrain surface at (  i ,i ) .
Incidence polarization is defined by the ellipticity angle  and orientation angle  (Jin
1994). The 22-D complex scattering amplitude functions are obtained from polarimetric
measurement as
T
1 1 
I i (  , )  [ I vi , I hi ,U i ,Vi ]T   (1  cos 2  cos 2 ), (1  cos 2  cos 2 ),  cos 2  sin 2 ,sin 2   (53)
2 2 
 c  4 cosi Pn (54)
where
Pn  0.5 I vs (1  cos 2  cos 2 )  I hs (1  cos 2  cos 2 )  U s cos 2  sin 2  Vs sin 2   (55)
When the terrain surface is flat, co-pol backscattering  c versus the incidence polarization
(  , ) has the maximum at   0 . However, it has been shown that as the surface is tilted,
the orientation angle  for the  c maximum is shifted from   0 .
Let Pn /   0 at the maximum  c and   0 of symmetric case, it yields
0  ( I hs  I vs )sin 2  U s cos 2  0.5( I hs  I vs ) ' 0.5U s sin 2  0.5( I hs  I vs ) 'cos 2 (56)
where the superscript of prime denote  /  . It can be shown that
0.5( I hs  I vs )' ~ ( M 13  M 23 )cos 2 ~ (Re  Svv Svh

* *
  Re  S hv S hh )cos 2 (57a)
0.5U s ~ M 33 cos 2 ~ Re  S S  S S  cos 2
*
vv hh
*
vh hv (57b)
2 2
( I hs  I vs ) ~ 0.5( M 11  M 22 )cos 2 ~ 0.5( Svv    S hh )cos 2 (57c)
0.5( I hs  I vs )' ~ 0.5( M 13  M 23 )cos 2 ~ 0.5(Re  S S   Re  S hv S )cos 2
*
vv vh
*
hh (58a)
* *
U s ~ 0.5( M 31  M 32 )~Re  S S   Re  S S 
vv hv vh hh (58b)
It can be seen that (57a) and (57b) are much less than (57c), and (58a) is much less than (58b),
so the last three terms on RHS of Eq. (56) are now neglected. Thus, it yields the  shift at
the  c maximum expressed by the Stokes parameters as follows
Us
tan 2  (59)
I hs  I vs
It can be seen that the third Stokes parameter U s  0 does cause the  shift.
By the way, if both U s and I hs  I vs approach zero, e.g. scattering from uniformly oriented
scatterers or isotropic scatter media such as thick vegetation canopy,  cannot be well
defined by Eq. (59).
6.2 The range and azimuth slopes and DEM inversion

The polarization vectors hî , vî of the incident wave at (  i ,i ) in the principal coordinate
( xˆ , yˆ , zˆ ) are defined as (Jin, 1994)
zˆ  kî
hî    sin i xˆ  cos i yˆ and vî  hî  kî (60a)
| zˆ  kî |
where the incident wave vector
kî  sin i cos i xˆ  sin i sin i yˆ  cosi zˆ (60b)
As the pixel surface is tilted with local slope, the polarization vectors should be re-defined
following the local normal vector zˆb as follows
zˆ  kˆ
hˆb  b ib and vˆb  hˆb  kîb (61)
| zˆb  kîb |
By using the transformation of the Euler angles (  ,  ) between two coordinates ( xˆ , yˆ , zˆ ) and
( xˆb , yˆb , zˆb ) (Jin, 1994), it has
xˆ  cos  cos  xˆb  sin  yˆb  cos  sin  zˆb (62a)

yˆ   sin  cos  xˆb  cos  yˆb  sin  sin  zˆb (62b)
zˆ  sin  xˆb  cos  zˆb (63c)
Substituting above relations into Eqs. (60,61), it yields
hî   cos  sin(  i ) xˆb  cos(  i ) yˆb  sin  sin(   ) zˆb (63)
kîb  (cos  sin i cos(  i )  sin  cosi ) xˆb  sin i sin(  i ) yˆb
(64)
 [sin  sin i cos(  i )  cos  cosi ]zˆb
Using Eq. (64) to Eq. (61), the polarization vector hˆb for local surface pixel is written as
axˆ  byˆ b
hˆb  b (65)
a2  b2
where a   sin i sin(  i ) , b  sin i cos  cos(  i )  sin  cosi . Thus, Eqs. (63,65) yield the
orientation angle as
cos  sin i  cos(  i )sin  cosi

cos  hˆb  hî  (66a)
a 2  b2
tan  sin(  i )
tan  (66b)
sin i  cos i tan  cos(  i )
Thus, the range and azimuth slopes of the pixel surface can be obtained as
S R  tan  cos  , S A  tan  sin  (67)
Since a single  shift cannot simultaneously determine two unknowns of  and  in Eq.
(64), two- or multi-pass SAR image data are usually needed.
If only the single-pass SAR image data are available, one of two unknown angles,  or  ,
should be first determined. The azimuth alignment of tilted surface pixels can be visualized
as a good indicator of the azimuth direction. We apply the adaptive threshold method and
image morphological thinning algorithm to specify the azimuth angle in all SAR image
pixels. The algorithm contains the following steps:
(1) Make speckle filtering over the entire image.
(2) Apply the adaptive threshold method to produce a binary image. The global
threshold value is not adopted because of the heterogeneity of the image pixels.
(3) Apply the image morphological processing for the binary image, remove those
isolated pixels and fill small holes. Referring to the part of binary’s “1” as the foreground
and the part of binary’s “0” as the background, the edges from the foreground are
extracted.
(4) Each pixel on the edge is set as the center of a square 2121 window, and a curve
segment through the centered pixel is then obtained. Then, applying the polynomial
algorithm for fitting curve segment in the least-squares sense, the tangential slope of the
centered pixel is obtained. It yields the azimuth angle of the centered pixel. Make a mark
on that pixel so that it won’t be calculated in the next turn.
(5) Removing the edge in Step (4) from the foreground, a new foreground is formed.
Repeat Step (4) until the azimuth angle of every pixel in the initial foreground is
determined.
(6) Make the complementary binary image, i.e. the initial background now becomes the
foreground. Then, the Steps (4) and (5) are repeated to this complementary image until
the azimuth angle of every pixel in the initial background is determined.
This approach provides a supplementary information to firstly determine the angle 
over whole image area if there is no other information available.
(a) (b) (c) (d)

Fig. 22. Morphological thinning algorithm to determine the azimuth angle
As an example for DEM inversion, the L-band polarimetric SAR data is shown in Figure
23(a). As the azimuth angle  of each pixel is obtained by the adaptive threshold method
and thinning algorithm as described in the above steps, the  angle of each pixel can be
determined by Eq. (66b). Taking i  0 , it yields
tan sin i
tan   (68)
sin   tan cos  i cos 
where the orientation angle  is calculated using Eq. (59), while the incident angle i is
determined by the SAR view geometry.
Substituting  and  into Eq. (67), the azimuth slope S A and range slope S R are obtained.
We utilize the two slopes to invert the terrain topography and DEM. The slopes S A and S R
of all pixels in the SAR image are calculated as shown in Figures 23(b,c).
(a) SAR image of total power (b) Azimuth slope (c) Range slope
Fig. 23. SAR image and azimuth/range slopes
The DEM can be generated by solving the Poisson equation for a M  N rectangular grid
area. The Poisson equation can be written as
 2 ( x, y )   ( x, y ) (69)
where  2 is the Laplace operator. The source function  ( x, y ) consists of the surface
curvature calculated by the slopes S ( x, y ) , where the xˆ , yˆ -directions are used as the range
and azimuth directions, respectively. In this section, the full multi-grid (FMG) algorithm is
employed to solve the Poisson equation. The benefits of FMG are due to its rapid-
convergence, robustness and low computation load. The inverted DEM of SAR image in
Figure 23(a) is given in Figures 24(a,b), and its contour map, Figure 24(b) has been validated
in comparison to the map of China Integrative Atlas (1990).
(a) Inverted DEM (b) Inverted contour map

Fig. 24. A DEM inversion
7. Reconstruction of Buildings Objects from Multi-Aspect SAR Images

Reconstruction of 3D-objects from SAR images has become a key issue for information
retrieval for SAR monitoring. 3D reconstruction of man-made object is usually based on
interferometeric technique or fusion with other data resources, e.g. optical and GIS data. It
has been noted that scattering from man-made objects produces bright spots in sub-metric
resolution, but present strips/blocks image in metric resolution. This difference is largely
attributed to the different imaging ways employed for different resolutions, for example,
spotlight for sub-metric resolution, stripmap for metric resolution.
An automatic method for detection and reconstruction of 3D objects from multi-aspect
metric-resolution SAR images. The steps are as follows:
The linear profile of the building objects is regarded as the most prominent characteristics.
The POL-CFAR detector, Hough transform, and corresponding image processing techniques
are applied to detection of parallelogram-like façade-images of building objects. Results of
simulated SAR images as well as real 4-aspect Pi-SAR images are given subsequently after
each step.
A probabilistic description of the detected façade- images is presented. Maximum-likelihood
estimation (ML) of building objects from multi-aspect observed façade-images is given.
Eventually, in association with a hybrid priority rule of inversion reliability, an automatic
algorithm is designed to match multi-aspect façade- images and reconstruct the building
objects. Moreover, taking advantage of the multi-aspect coherence of building-images, a
new iterative co-registration method is presented.
Finally, reconstruction results are presented, and good performance is evaluated comparing
with ground truth data. It is also concluded that the reconstruction accuracy closely
depends on the number of available aspects. In the end, a practicable scheme of the 3D-
object reconstruction from satellite-borne SAR image is proposed (Xu and Jin, 2007).
7.1 Object image detection

As shown in Figure 25, the scattering image of a simple building is produced by direct
scattering from the wall/façade, roof and edges, double scattering from wall-ground and
shadows projected etc. In metric resolution, the scattering produces bright strips/blocks
from respective parts of the object. A key step of reconstruction is to identify and extract
these strips/blocks.
For the case of a smooth wall/façade, the only double scattering term to be considered must
follow a specific propagation path, i.e. wall (reflect) to ground (diffuse). Simple building
object is taken into account and is modeled as a cuboid, and the spatial distribution of the
building objects is assumed not crowded, i.e. without serious shadowing and superposition.
The cuboid object is described by seven geometric parameters: 3D position, 3D size and
orientation. Besides, the flat roof is assumed with much less scattering compared with the
edges and façades.
Figure 26 shows (a) a model of the cuboid object, (b) its simulated SAR image, (b) using the
mapping and projection algorithm, (c) a photo of real rectangular building, and (d) its image
of Pi-SAR observation, respectively.
It can be seen from Figure 26 that the longer wall of the cuboid-like building, called as major
wall henceforth, plays the dominant role in a SAR image. In this section, main attention is
focused on the major wall image. At the first step, the edge detectors of constant false alarm
rate (CFAR) such as ratio gradient are used for edge detection.
incidence
H
r
A
o
B C O
D
E
Hctg F
G
Htan y
x
ima ge s on the line
A-B : Roof Sca tte ring + Wa ll Single Scatte ring
B-C : Wa ll Single Scatte ring
C-D : Wa ll / Ground Double Scattering
C-F : Wa ll / Ground Triple Sca tte ring (no ground sca tterin
F-G : Sha dow (no ground sca ttering)
Fig. 25. SAR imaging of a simple building model and its image composition
To our experience, the POL-CFAR detector derived from complex Wishart distribution can
fulfill the segmentation requirements of medium- and small-scale building-images. Besides,
the ridge filter applied after this step can accommodate segmentation error to some extent.
(a) model (b) simulated SAR image (c) a photo (d) PI-SAR image
Fig. 26. A cuboid object in m-resolution SAR images
In order to improve the segmentation precision, the detected edge by window-operator

needs to be further thinned. Using the edge intensity produced in the POL-CFAR detection,
an edge thinning method of ‘ridge filtering’ is presented. Taking an 8-neighbor 3×3 window
as an example, the center pixel is regarded as on the ridge if its intensity is higher than the
pixels along two sides.
4-aspect Pi-SAR images acquired over Sendai, Japan by a square-loop flying path (Flight No.
8708, 8709, 8710, 8711, X-band, pixel resolution 1.25m) are taken as an example of real SAR
image study. The region of interest (ROI) is the Aobayama campus of Tohoku University.
Figure 27(a) shows an aspect Pi-SAR images as an example.
Figures 27(b-d) show the results processed by the edge detection of Pi-SAR images.
The most distinct feature of a building object in SAR image is parallel lines. The Hough
transform is employed to detect straight segments from the thinned edges, and
parallelogram outlines of the façade-images can then be extracted. It is carried out in tiling
manner in this paper, i.e. the original picture is partitioned into blocks, each of which is
detected independently via Hough transform.
(a) An aspect PI-SAR image (b) edge detection (c) Level 4 ridge (d) Level 2 ridge
Fig. 27. A PI-SAR image and edge detection
The detection steps of parallel line segments are: i) find bright spots in transform domain
with a minimum distance between every two of them so as to avoid re-detection; ii) search
the segments consisted of successive points along the corresponding parallel lines in spatial
domain, which is longer than a minimum length, and the distance between two successive
points is shorter than a maximum gap; iii) only the pairs of points lying on two lines and
facing directly are taken into account for segment searching.
(a) Hough transform (b) Post-processing (c) End adjusting (d) Offset-removed
Fig. 28. Extraction of strip-like building-images
Figures 28(a-d) show the detection from 4 aspects Pi-SAR images. Few building-images are
not detected and some false images are wrongly produced. In fact, there is a compromise
between over-detection and incomplete-detection. We prefer to preserve more detected
building-images, so as to control the non-detected rate, while the false images are expected
to be eliminated through the subsequent auto-selection of effective images for
reconstruction. However, there always remain some undetectable images, attributable to
shadowing of tree canopy, overlapping of nearby buildings or with too complicated wall
structures.
The detected parallelogram of a homogenous scattering area could be direct scattering from
façade, double scattering of wall-ground, combination of direct and double scatterings,
projected shadow of building, or even strong scattering of strip-like objects (e.g. metal fence
or metal awning), which is not considered in classification.
Shadowing is identified if scattering power of that area is much lower than the vicinity.
Specifically, first set up two parallel equal-length strip windows on its two sides and then
calculate the median scattering powers of the three regions. If the middle one is weaker than
two sides, it is classified as shadow instead of building image.
The wall-ground double scattering can be differentiated from direct scattering based on
polarimetric information. The de-orientation parameter v indicating the scattering
mechanism is used to identify double scattering.
To implement target reconstruction from multi-aspect SAR data, calibrating multi-aspect
data requires that at least one aspect is calibrated beforehand, and other aspects are then
calibrated with respect to this calibrated aspect. A natural object, such as flat bare ground, is
usually chosen as a reference target, which is expected to preserve identical scattering for all
aspects. Thereafter, the channel imbalance factors are estimated from distribution of the
phase difference and amplitude ratio of co-polarized, hh and vv, echoes of the reference
target, and are then used to compensate the whole SAR images (Xu and Jin, 2008b).
7.2 Building reconstruction

Complexity of objects-terrain surfaces and image speckles incorporate certain uncertainty
for detection of object images. To well describe multi-aspect object image, the parameterized
probability is introduced for further proceeding of automatic reconstruction. For
convenience, the detected building-images are parameterized. Generally, the edge pixels
detected by CFAR are randomly situated around the real, or say, theoretical boundary of the
object. It is reasonable to presume that the deviation distances of the edge points follow a
normal distribution.
The original edge can be equivalently treated as a set of independent edge points. The line
detection approach is considered as an equivalent linear regression, i.e. line fitting from
random sample points. According to linear regression from independent normal samples,
the slope of the fitted line follows the normal distribution.
All parameters have normal distributions and their variances are determined by the
variance of the edge points deviation. After counting the deviation of the edge points in the
vicinity of each lateral of all detected building-images and making its histogram, the
variance can be determined through a minimum square error (MSE) fitting of the histogram
using normal probability density function (PDF). The 4-aspect Pi-SAR images are counted.
z Sensor
i
l
Wall
h
Ground ψ
e
Double Scattering
(x,y) i li y
Drect
bi (x i ,y i ) Scattering
di
ψi
Imaging Plane
x
Fig. 29. Geometry of a wall/façade imaging
Given a group of building-images detected from multi-aspect SAR images, the

corresponding maximum- likelihood probability can be further used as an assessment of the
coherence among this group of multi-aspect images. A large maximum-likelihood
probability indicates a strong coherence among the group of multi-aspect building-images,
and vice versa.
Multi-aspect co-registration, as a critical pre- processing step, is necessary when dealing
with real SAR data.
Given the specification of a region of interest (ROI), e.g. the longitudes and latitudes, the
corresponding area in SAR image of each aspect can be coarsely delimited according to its
orbit parameters. It can be regarded as a coarse co-registration step. Manual intervention is
necessary if the orbit parameters are not accurate enough.
Only parameters of the building-images are needed to be co-registered to the global
coordinates, but rather than the original SAR images. It is regarded as a fine co-registration
step. In this study, only linear co-registrations are considered. A simple and effective
method should be manually choosing ground control points. In general, a featured terrain
object or building is first selected as the reference of zero-elevation, and then its locations of
different aspects are pinpointed. Coherence among multi-aspect building-images is the basis
of automatic reconstruction.
Finally, an automatic multi-aspect reconstruction algorithm is designed. The building

objects are reconstructed from the 4-aspect simulated SAR images. The inversion accuracy is
very good. shows the 3D reconstructed objects on the true DEM. It seems the inverted
elevations also match well with the true DEM.
Reconstructed
Ground Truth
Fig. 30. 3D reconstructed objects and comparisons with optical picture
Due to the difficulty for authors to collect ground truth data, a high-resolution satellite
optical picture (0.6m QuickBird data) is used as a reference. Geometric parameters of each
building manually measured from the picture are taken as ground truth data to evaluate the
accuracy of reconstruction.
There is a trade-off between the false and correct reconstruction rates. If we increase the
false alarm rate of edge detector, relax the requirements of building-image detection and/or
increase the false alarm rate of judging correctly matched group, it will raise the
reconstruction rate, but also boost the false reconstruction rate. Efficiency will be
deteriorated if the false reconstruction rate goes too high. On the contrary, the false
reconstruction rate can be reduced and the accuracy of reconstructed buildings can be
improved, but the reconstruction rate will also decline.
A critical factor confining the reconstruction precision should be the number of effective
aspects, hereafter referred as effective aspect number (EAN).T he reconstruction result will
become better if more SAR images of different aspects are available.
Main error of reconstruction is caused by the boundary-deviation of detected building-images,
which is originated from complicated scattering and interactions of spatially distributed
objects and backgrounds. In probabilistic description of detected building-image, the
presented boundary in SAR image is seen as the same as reality. However, the real detected
building-image might be biased or even partly lost due to the obstacle and overlapping.
In addition, it is noticed that large-scale buildings might not be well reconstructed. The
reason is partly attributed to the premise of Wishart distribution of POL-CFAR detector.
Since the images of large-scale buildings might reveal more detail information about the
texture feature and heterogeneity, it deteriorates the performance of edge detector. To
develop a new edge detector based on certain specific speckle model for high-resolution
images can improve the results. Another more feasible way is to employ a multi-scale
analysis, through which building-images of different scales can present Gaussian properties
in their own scale levels. Of course, the expense for multi-scale analysis might be a loss in
precision.
Another issue to be addressed is the exploitation of multi-aspect polarimetric scattering
information. In heterogeneous urban area, the terrain objects appear distinctively in
different aspects, which gives rise to a very low coherence between their multi-aspect
scatterings. Hence, polarimetric scattering information may not be a good option for the
fusion of multi-aspect SAR images over urban area
After the ROI is coarsely chosen in each aspect image, edge detection and object-image
extraction are carried out, subsequently. Then the object-images are parameterized and
finely co-registered. As long as multi-aspect object-images are automatically matched, 3D
objects are reconstructed at the same time.
The merits of this approach are the process automation with few manual interventions, the
fully utilization of all-aspect information, and the high efficiency for computer processing.
Making a reconstruction on a 4-aspect Pi-SAR dataset (500×500) takes less than 10 min CPU
time (CPU frequency 3GHz). Complexity of this approach is about O( KN 2 ) , where K is the
aspect number and N is the size of SAR images (K=4, N=500 in this case).
It is tractable to extend this approach to reconstruction of other kinds of objects. For other
types of primary objects, given the a priori knowledge, new image detection methods have to
be developed. It is possible to treat more complicated buildings as combinations of different
primary objects, which can be reconstructed separately and then combine together.
Considering a space-borne SAR with the functions of left/right side looking (e.g. Radarsat2
SAR) and forward/backward oblique side looking (e.g. PRISM in ALOS), six different
aspect settings are available for ascending orbit. There is a 20° angle between ascending and
descending flights for sun-synchronous orbit. Therefore, it can observe from 12 different
aspects. Suppose the sensor use different aspect in each visit and the repeat period is 15
days, a set of SAR images acquired from 12 aspects can be obtained through a 3-month
observation. Then, application of SBT reconstruction with acceptable precision.
8. Acknowledgment
This work was supported by the National Science Foundation of China with the grant No.
NSFC 406370338.
9. Reference
Jin, Y.Q. (2005). Theory and Approach of Information Retrievals from Electromagnetic Scattering
and Remote Sensing, Springer, ISBN-10 1-4020-4029-6, The Netherlands.
Jin, Y.Q. (2000). Information of Electromagnetic Scattering and Radiative Transfer in Natural
Media. Science Press, ISBN 7-03-008259-1, Beijing.
Jin, Y.Q. (1994). Theory and Approach of Information Retrievals from Electromagnetic Scattering
and Remote Sensing, Sworld Scietific, ISBN-981-02-1648-3, Singapore.
Jin, Y.Q. (1992). A Mueller matrix approach to complete polarimetric scattering from a layer
of non-uniformly oriented, non-spherical scatterers. Journal of Quantitative
Spectroscopy and Radiative Transfer, 48(3): 295-306, ISSN 0022-4073.
Jin, Y.Q. (1994). Polarimetric scattering from a layer of random clusters of small spheroids.
IEEE Transactions on Antenna and Propagation, 42(8): 1138-1144, ISSN 0018-926X.
Jin, Y.Q. & Cloude, S. (1994). Numerical eigenanalysis of the coherency matrix for a layer of
random nonspherical scatterers. IEEE Transactions on Geoscience and Remote Sensing,
32(6): 1179-1185, ISSN 0196-2892.
Jin, Y.Q. (1995). Polarimetric scattering and emission from a layer of random clustters of
small spheroids and dense spheres. Microwave Radiometry and Remote Sensing of the
Environment, ed. by Solimini, D., VSP, 419-427, The Netherlands, ISBN-90-6764-189.
Jin, Y.Q. (1995). Polarimetric scattering from heterogeneous particulate media. Journal of
Applied Physics, 78(8): 4835-4840, ISSN 1087-3848.
Jin, Y.Q., Zhang, N., Chang, M. (1999). The Mueller and coherency matrices solution for
polarimetric scattering simulation for tree canopy in SAR imaging at C band.
Journal of Quantitative Spectroscopy and Radiative Transfer, 63(2): 699-612, ISSN 0022-
4073.
Jin, Y.Q. & Zhang, N. (2002). Statistics of four Stokes parameters in multi-look polarimetric
SAR imagery. Canadian Journal of Remote Sensing, 28(4): 610-619, ISSN 0703-8992.
Jin, Y.Q. & Chen, Y. (2002). An improved method of the minimum entropy for refocusing
moving target image in the SAR observation. The Imaging Science Journal, 50(6): 147-
152, ISSN 1368-2199.
Jin, Y.Q. & Chen, F. (2002). Polarimetric scattering indexes and information entropy of the
SAR imagery for surface monitoring. IEEE Transactions on Geoscience and Remote
Sensing, 40(11): 2502-2506, ISSN 0196-2892.
Jin, Y.Q. & Liang, Z. (2003). Iterative solution of multiple scattering and emission from
inhomogeneous scatter media. Journal of Applied Physics, 93(4): 2251-2256.
Jin, Y.Q. & Chen, F. (2003). Scattering simulation for inhomogeneous layered canopy and
random targets beneath canopies by using the Mueller matrix solution of the pulse
radiative transfer. Radio Science, 38(6): 1107-1116, ISSN 0048-6604.
Jin, Y.Q. & Liang, Z. (2004). Iterative inversion from the multi-order Mueller matrix solution
of vector radiative transfer equation for a layer of random spheroids. Journal of
Quantitative Spectroscopy and Radiate Transfer, 83(4): 303-311, ISSN 0022-4073.
Jin, Y.Q. & Liang, Z. (2004). An approach of the three-dimensional vector radiative transfer
equation for inhomogeneous scatter media. IEEE Transactions on Geoscience and
Remote Sensing, 42(2): 355-360, ISSN 0196-2892.
Jin, Y.Q., Chen, F. & Chang, M. (2004). Retrievals of underlying surface roughness and
moisture for stratified vegetation canopy using polarized pulse echoes in the
specular direction. IEEE Transactions on Geoscience and Remote Sensing, 42(2): 426-
433, ISSN 0196-2892.
Jin, Y.Q. & Liang, Z. (2004). Iterative retrieval of canopy biomass and surface moisture by
using multi-order Mueller matrix solutions of polarimetric radiative transfer.
Canadian Journal of Remote Sensing, 30(2): 169-175, ISSN 0703-8992.
Jin, Y.Q. & Luo, L. (2004). Terrain topography inversion using single-pass polarimetric SAR
image data. Science in China (F), 47(4): 490-500, ISSN 1009-2757.
Jin, Y.Q., Chen, F. & Luo, L. (2004). Automatic analysis of change detection of multi-
temporal ERS-2 SAR images by using two-thresholds EM and MRF algorithms. The
Imaging Science Journal, 52: 234-241, ISSN 1368-2199.
Jin, Y.Q. & Liang, Z. (2004). Scattering and emission from inhomogeneous vegetation
canopy and alien target by using three-dimensional vector radiative transfer (3D-
VRT) equation. Journal of Quantitative Spectroscopy and Radiative Transfer, 92(3): 261-274,
ISSN 0022-4073.
Jin, Y.Q. (2007). Theory and application for retrieval and fusion of spatial and temporal
quantitative information from complex natural environment. Frontiers of Earth
Science in China, 1(3): 284-298, ISSN 1673-7385.
Jin, Y.Q., Xu, F. & Fa, W. (2007). Numerical simulation of polarimetric radar pulse echoes
from lunar regolith layer with scatter inhomogeneity and rough interfaces. Radio
Science, 42, RS3007, doi: 10.1029/ RS2006003523: 1-10, ISSN 0048-6604.
Jin, Y.Q. & Wang, D. (2009). Automatic Detection of the Terrain Surface Changes after
Wenchuan Earthquake, May 2008 from ALOS SAR Images Using 2EM-MRF
Method, IEEE Geoscience and Remote Sensing Letters, 6(2): 344-348, ISSN 0196-2892.
Cao, G.Z. & Jin, Y.Q. (2007). A hybrid algorithm of BP-ANN/GA for data fusion of landsat
ETM+ and ERS-2 SAR in urban terrain classification. International Journal of Remote
Sensing, 28(2): 293-305, ISSN 0143-1161.
Cao, G. & Jin, Y.Q. (2007). Automatic detection of main road network in dense urban area
using microwave SAR image. The Image Science Journal, 55: 215-222, ISSN 1368-2199.
Chang, M. & Jin, Y.Q. (2002). The temporal Mueller matrix solution for polarimetric scattering
from a layer of random non-spherical particles with a pulse incidence. Journal of
Quantitative Spectroscopy and Radiative Transfer, 74: 339-353, ISSN 0022-4073.
Chang, M. & Jin, Y.Q. (2003). Temporal Mueller matrix solution for polarimetric scattering
from inhomogeneous random media of non-spherical scatterers under a pulse
incidence. IEEE Transactions on Antenna and Propagation, 51(4): 820-832, ISSN 0018-
926X.
Dai, E., Jin, Y.Q., Hamasaki, T. & Sato, M. (2008). 3-D stereo reconstruction of buildings
using polarimetric images acquired in opposite directions. IEEE Geoscience and
Remote Sensing Letters, 5(2): 236-240, ISSN 0196-2892.
Fa, W., & Jin, Y.Q. (2009). Image Simulation of SAR Remote Sensing over Inhomo-geneously
Undulated Lunar Surface. Science in China (F), 52(4): 559-574, ISSN 1009-2757.
Liang, Z. & Jin Y.Q. (2003). Iterative approach of high-order scattering solution for vector
radiative transfer of inhomogeneous random media. Journal of Quantitative
Spectroscopy and Radiative Transfer, 77: 1-12, ISSN 0022-4073.
Qi, R. & Jin, Y.Q. (2007). Analysis of Faraday rotation effect in satellite-borne SAR
observation at P band. IEEE Transactions on Geoscience and Remote Sensing, 45(5):
1115-1122, ISSN 0196-2892.
Wei, Z., Xu, F. & Jin, Y.Q. (2008). Phase unwrapping for SAR interferometry based on ant
colony optimization algorithm. International Journal of Remote Sensing, 29(3): 711-725,
ISSN 0143-1161.
Xu, F. & Jin, Y.Q. (2005). Deorientation theory of polarimetric scattering targets and
application to terrain surface classification. IEEE Transactions on Geoscience and
Remote Sensing, 43(10): 2351-2364, ISSN 0196-2892.
Xu, F. & Jin, Y.Q. (2006). Multi-parameters inversion of a layer of vegetation canopy over
rough surface from the system response function based on the Mueller matrix
solution of pulse echoes. IEEE Transactions on Geoscience and Remote Sensing, 44(7):
2003-2015, ISSN 0196-2892.
Xu, F. & Jin, Y.Q. (2006). Imaging simulation of polarimetric synthetic aperture radar remote
sensing for comprehensive terrain scene using the mapping and projection
algorithm. IEEE Transactions on Geoscience and Remote Sensing, 44(11): 3219-3234,
ISSN 0196-2892.
Xu, F. & Jin, Y.Q. (2007). Automatic reconstruction of 3-D stereo buildings from multi-aspect
high-resolution SAR images, IEEE Transactions on Geoscience and Remote Sensing.
45(7): 2336-2353, ISSN 0196-2892.
Xu, F. & Jin, Y.Q. (2008a). Imaging simulation of bistatic SAR and polarimetric analysis.
IEEE Transactions on Geoscience and Remote Sensing, 46(8): 2233-2248, ISSN 0196-
2892.
Xu, F. & Jin, Y.Q. (2008b). Calibration and validation of a set of multi-aspect airborne
polarimetric Pi-SAR data. International Journal of Remote Sensing, 29(23): 6801-6810,
ISSN 0143-1161.
Xu, F. & Jin, Y.Q. (2009). Raw signal processing of stripmap bistatic SAR. IET-Radar Sonar
and Navigation, 3(3): 233-244, ISSN 1751-8784.
Polarimetric responses and scattering mechanisms of tropical forests in the Brazilian Amazon 183
X8
Polarimetric responses and scattering

mechanisms of tropical forests
in the Brazilian Amazon
J. R. dos Santos1, I. S. Narvaes1, P. M. L. A. Graça2 and F. G. Gonçalves3
(1) INPE
- National Institute for Space Research
Av. dos Astronautas, 1758 CEP.:12.227-010 São José dos Campos – SP., Brazil
Email: {jroberto, igor}@dsr.inpe.br
(2) INPA - National Institute of Amazonian Research
Av. André Araujo, 2936 CEP.: 69060-001 Manaus – AM., Brazil,

Email: pmlag@inpa.gov.br
(3 ) Oregon State University
Department of Forest Ecosystems and Society, Corvallis, OR 97331 EUA,

Email: fabio.goncalves@oregonstate.edu
Abstract
This chapter discusses the polarimetric responses of PALSAR (L-band) data and scattering
mechanisms of some tropical forest typologies based on target decomposition. The
fundamentals of polarimetric theory related to both SAR topics under development are
summarized. For representation of polarimetric signatures, the cross-section of the forest
target (σ) was plotted on a bi-dimensional graphic as a function of the orientation angle,
ellipticity angle and the intensity of co-polar components of the radar signal. The analysis of
scattering mechanisms was done by the association of entropy and mean alpha angle values
of each sample, plotted in a bi-dimensional classification space. This study improves the
understanding of the interaction mechanisms between L-band PALSAR signals and
structural parameters, supporting the forest inventory in the Brazilian Amazon region.
KEYWORDS: tropical forest, polarimetric signatures, target decomposition, PALSAR,
monitoring.
1. Introduction
The ongoing particular and governmental struggle over the Brazilian Amazon’s future
mirror broader current discourses on the “environment” (Andersen et al., 2002), experts and
layperson alike vary widely along continuum perspectives, based on two scenarios: (a) the
line of defenders of global ecological services (“conservationists”) and (2) the focus of
development interests in the countries hosting these forests (“developmentalists”). They
both present an important view points and significant arguments. While these discussions
occur in the frame of a possible territorial planning, the degradation process with habitat
fragmentation is still going on, with an estimative of rate gross deforestation in the Amazon
region of around 11,968 km2 (timeframe of 2007-2008), with an error of approximately 4%
(<http://www.obt.inpe.br/prodes/index.html>). Due to the advancement of economic
activities in the domain of the Brazilian Amazon tropical rainforest, remote sensing have
being used as a fundamental tool to characterize the causes of degradation (conversion of
natural vegetation to agriculture and cattle raising, selective logging, charcoal production,
etc.), and also to monitor the impact of human activities over these large ecosystems.
Taking into account that in tropical regions there is, all over the year, a high percentage of
cloud cover which impedes the inventory and updating of the forest cover by optical data, it
is important to use sensor systems which operate at microwave wavelengths. A multi-
temporal analysis of TM-Landsat data (timeframe of 12 years) used to estimate the annual
gross deforestation rate in the Brazilian Amazon region showed that in average 21.12% of
the imaged areas was cloud covered (Santos et al., 2008). Since there is a significant dynamic
in environments of the Amazon region due to seasonal variations or from the different
conditions resulting from human action, it is important to use radar images to understand
the transformation process in the forest landscape.
To briefly introduce a reader on this subject, including concepts and applications of radar
data in forestry studies, the following papers are recommended: Ulaby and Elachi (1990),
Beaudoin et al. (1994), Henderson and Lewis (1998) and Coops (2002). These authors studied
multitemporal SAR (Synthetic Aperture Radar) backscatter for discriminating forest types,
discussing those conditions of scattering and attenuation of radar signals at different
frequencies interacting with structural vegetation features. Pope et al. (1994) developed
indices based on ratios and normalized different of multipolarimetric radar data, which can
be related to certain characteristics of vegetation cover, such as e.g. biomass, canopy
structure, volume scattering, applied to tropical forest in Central America. Hoekmann and
Quiñones (2000), Hawkins et al. (2000) and Santos et al. (2002; 2003) correlated information
obtained from radar with forest biomass levels, discussing the saturation of backscatter
saturation caused by the high density of tropical forests. Power spectrum analysis was made
using autoregressive moving-average (ARMA) models, for evaluation of forest canopy
structure from airborne SAR (X–band) data in the Brazilian Amazon (Neeff et al., 2005a).
This approach is novel where the radar spectra match those obtained from ground data,
establishing the basis for the use of this relationship in spatial structure analysis of
ecosystems from remote sensing sources. In more recent studies, and exploring the radar
potentialities beyond its traditional polarimetric characteristics, Kugler et al. (2006) and
Treuhaft et al. (2006) used interferometric radar data to estimate biophysical parameters in
tropical forest areas. Using another approach, Neeff et al. (2005b) explored, through the local
maximum filter technique (Wulder et al., 2000) and statistical Markov processes (Cressie,
1993) the interferometric height derived from airborne SAR (X- and P- bands) for the
analysis of spatial forest patterns (three dimensional canopy structure and distances of very
large trees). The above mentioned studies exemplify the importance of radar information in
the analysis of tropical environments.
Within this frame, the objective of this chapter is to explorer the polarimetric mode, by the
analysis of the graphic representation of signatures from PALSAR/ALOS data (L-band) in
primary forest, secondary succession (initial, intermediate and advanced levels of natural
recovery) and also, forests with timber exploitation. Additionally, we also discuss an
exploratory analysis of scattering mechanisms of tropical forest typology, in accordance

with the alternative procedure of SAR image classification based on target decomposition.
1.1 Theoretical background

In order to understand the theme of this chapter, it is necessary to introduce some concepts
about SAR polarimetry and of the classification technique based on target decomposition, as
described below.
1.1.1 Concepts of polarized waves

Through different polarizations of radar imaging, the polarimetric signature of a certain
target can be obtained, as well as the characterization of the dominant mechanism which
controls the interaction of the microwave signal with the structure of the target under
consideration.
The principle of polarimetry is related to the state of polarization of the wave, represented
by definition as the electric field from the transmitting and receiving antenna. The state of
polarization is described as an arbitrary set of axis used to describe horizontal polarized
waves in the x axis (H polarization) and at the vertical polarized waves in the y axis (V
polarization) along the wave propagation plan (z), which corresponds to the pointing
direction of the antenna to the target. In this way, four types of polarization are obtained:
HH for transmission and reception at the horizontal plan; VV for transmission and reception
at the vertical plan; HV for transmission at the horizontal plan and reception at the vertical
plan; and finally, VH for vertical transmission and horizontal reception. Ulaby and Elachi
(1990) and Henderson and Lewis (1998) discuss how the electric field of an irradiated
electromagnetic wave could be mathematically represented by a polarization ellipse. The
polarization ellipse (Figure 1) is described by the three geometric parameters: the orientation
angle (-90º< ψ < +90º), the ellipticity angle (+45º> χ >-45º) and amplitude (A). The flattening
factor from the ellipsis is given by angle χ, and there is a linear polarization when χ = 0; the
circular polarization to the left and right when χ = - 45º and χ = + 45º, respectively. At the
condition of χ = 0º and ψ = 0º there is horizontal polarization and at χ = 0º and ψ ± 90º a
vertical polarization. The analyst, who will discuss the polarization ellipse based on some
literature references, must pay attention at the representation scale from the orientation axis,
because graphical plots are presented in two forms: one of them with the polarization power
varying between 0º and 180º, and the other varying from -90º to +90º, which can cause an
interpretation error.
Fig. 1. Polarization pattern in the x-y plane, rotation angle ψ, ellipticity angle χ and auxiliary
angle α. The size of the ellipse is governed by the maxima of the horizontal and vertical
electrical field components. Source: Woodhouse (2006).
It is important to comment that the description of the wave polarization can be different
according to the convention being considered. In the Forward Scattering Alignment (FSA)
system, the wave propagation axis changes its direction after interaction with the target (this
case represents the waves scattered by forest targets). In the Back Scattering Alignment
(BSA) system the x and y axis are aligned in the same direction, and the wave propagation
axis (z) is considered parallel, with the advantage of some mathematical simplifications on
how the data are described (Woodhouse, 2006). In this way the coordinate system does not
influence the nature of polarization, but just the way how the data are described.
There are other ways to represent the polarization state: (1) the graphic representation with
the Poincaré sphere (Ulaby and Elachi, 1990), where the circular polarizations to the right
and left are represented at the north and south poles respectively, while the intermediate
polarization states are located at the regions between these poles. In this representation the
meridians correspond to different inclination angles of the largest axis of the ellipse (ψ = 0)
and the parallels to the ellipticity of the electric field (χ = arctan (−B/A)); (2) the
mathematical representation by Jones vector, where the complex data received from each
pulse emitted after the wave interacted with a target is demodulated at in-phase (In) and
quadrature (Q) and converted to the digital format. Therefore, this mathematic
representation is important because there is a direct relation with the polarization ellipse
(van der Sanden, 1997; Woodhouse, 2006); (3) the Stokes vector where the polarized waves
can be characterized by four real and measurable parameters, with the same physical
dimension, but only three are independent and describe the polarization state as:
I02 = Q² + U ² + V² (1)
Where: Q = is the tendency of more vertical (Q > 0) or horizontal polarization (Q < 0);
U = expresses the tendency to be polarized at +45º (U > 0) and - 45º (U < 0);
V = polarization referring to the rule of right (V < 0) and left (V > 0) hand rules.
The parameters U and V are sometimes referred as being co-variances in-phase (real) and
quadrature (imaginary), respectively between the components of the vertical and horizontal
field of the wave. This expression is applied to completely polarized waves (amplitude and
phase are invariant in time). Additionally the Stokes vector has a direct relation with the
Poincaré sphere, because its ray is the Stokes parameter Io (Woodhouse, 2006).
It is important to mention that the first imaging radars known as “conventional” recorded
only information about the amplitude of the electric field at components x and y, i.e. the
maximum magnitude of the electric field from a wave as a measurable indicator of size
(van der Sanden, 1997). The amplitude takes into account the texture and brightness of the
image, influenced by the surface roughness (macro-, meso- and micro-scales) and by the
dielectric constant (moisture content of the target). With the advancement of technology,
imaging radars started to record, phase information (the angular position δ of an oscillating
movement at a defined instant in time t, as a function of the distance radar to target) in
addition to amplitude, being called polarimetric radars. Such phase information provides
information about the scattering mechanisms occurring in the imaged scene.
To improve the knowledge of the microwave interaction process with a certain target, it is
necessary to understand the types of scatterers involved. The deterministic scatterers, also
called coherent or punctual targets, are those where the interaction with electromagnetic
waves reflects completely polarized waves, preserving the polarization of the incident wave
(ESA, 2009a). Examples of this are urban targets. The non-deterministic targets, also known
as incoherent targets or partially polarized or random distributed scatterers, are those that
have more than one center of scattering, where the measured signal is the overlapping of a
large amount of waves with variable polarization (ESA, 2009a). This is the case of forest
targets. The change of polarization state is known as depolarization, caused mainly by
multiple scattering due to surface roughness and to volumetric interaction with the target.
Besides that, there is an almost specular reflection for smooth undulating surfaces, whose
physical concept and representations will be presented below, when we consider the target
decomposition approach.
1.1.2 Targets decomposition

In the polarimetric imaging the four complex components derived from the backscattering
in each resolution cell (SHH, SHV, SVH, SVV) can be expressed by a scattering matrix [S] which,
according to Ulaby and Elachi (1990) is given by:
 SVV SVH 
[S ] =   (2)
 SHV SHH 
According to Woodhouse (2006), the scattering matrix [S] is applied to deterministic targets
(coherent), and models how the scatterers transform the components of the incident electric
field (Ei) in the scattered electric field (Es).
 Evs  e − ik0 r  SVV SVH   Evi 

 s =     (3)
 Eh  R  SHV SHH   Ehi 
Where R is the distance from the antenna to the target and k = 2π / λ
Non-deterministic targets (incoherent) are represented by second order matrices. When the
polarization of the electromagnetic wave is represented by the modified Stokes vector, the
relation between the polarization of the incident electric field Fi and the scattered field Fs is
given by the Müller matrix [M] (or Stokes matrix) in the FSA system, or also by the
Kennaugh matrix [K] in the BSA system (Woodhouse, 2006), according to the equations
below:
1
F
s
= 2
[M ] F i (4)
r
1
F
s
= 2
[K ] F i (5)
r
Besides these matrices, the scattering of non-deterministic targets can be represented,
according to Cloude and Pottier (1996), by power matrices of covariance [C] and coherence
[T]. Since the radar image pixels contain many scatterers (especially in the forest targets),
with different scattering properties, the covariance and coherence can be used to represent
the matrix [S]. The covariance matrix [C] is generated by multiplying the vector-target
(Kennaugh) and its joint complex.
 
[C ] = K c .K c*T (6)
where:

kc = ( S hh , 2 S hv , S vv )T ;

K p = 1 / 2 [(S hh + S vv )(S hh − S vv )(S hv + S vh )i (S hv − S vh )]
T
<…> represents the spatial average.
When one is working with the average value of scattering mechanisms from a pre-defined
window in the imaged scene, the characterization of the targets by polarization power is a
consequence of a spatial average (normally of pixels), and this operation is called multi-look
processing.
On the other hand, the coherence matrix [T], which is related to the physical and geometric
properties of the scattering processes, is given by:
 
[T ] = K p .K *pT (7)

Where: Kp is the scattering vector written in the Pauli basis (ESA, 2009b)
The matrices [C] and [T] contain the same information about amplitude, phase and
correlations, besides its’ eigenvalues are real and similar. As a complement, the matrices
[M], [K], [C] and [T] are linearly related among them and there are formal methods for
transformation among them (Woodhouse, 2006).
The known decompositions to model the nature of coherent targets (represented by the
matrix [S]) are those from Pauli, Krogager (SHD), Cameron and Hyunen (Touzi et al., 2004;
Zhang et al., 2008), but they are beyond the scope of discussion at this chapter. For non-
deterministic targets (represented by second order matrices) there are two decomposition
mechanisms implemented at most software packages, which are used for the discrimination
of scattering type occurring in forest targets: the decompositions of Freeman-Durden and
Cloude-Pottier.
The Freeman-Durden decomposition models the matrix [C] as a contribution of the surface
scattering (fs), the double-bounce (fd), and volumetric type (fv) where each type of
contribution is represented by the theoretical target, related to the respective type of
scattering which occurs at the forest (Freeman and Durden, 1998).
The polarimetric decomposition of Cloude and Pottier (1997) is based on three canonic
elements (entropy, anisotropy and alpha angle) defined as a function of the decomposition
of eigenvalues and eigenvectors from the matrix [T]:
3
T3 = ∑ [T3i ] = λ1 (e1e1*T ) + λ2 (e2 e2*T ) + λ3 (e3e3*T ) (8)
i =1
Where: λi and ei are eigenvalues and eigenvectors of the matrix [T] respectively.
The entropy (H) is a parameter that indicates the type of target involved in the imaged scene
and the number of dominant scattering mechanisms (H~0 = mechanism of unique
scattering; H~1 = multiple scattering mechanisms). In a dense tropical forest, the H values
are normally elevated, indicating a larger variation of structural arrangements due to the
diversity of species composition, the different levels of competition and growth, as well as
the inter-specific association among species. In short, the entropy represents the randomness
of the scattering.
The alpha angle (α) indicates the type of scattering dominating the scene, where α = 0°
(surface scattering), α = 45° (volumetric scattering) and α = 90° (double-bounce scattering).
The anisotropy (A) is a complementary parameter that measures the relative importance of
λ2 and λ3 (Equation 8). When H > 0.7, the anisotropy contributes with additional information
to characterize the target. Below this value, the anisotropy is noisy and has limited
importance. The intensity of these canonic elements identifies the presence of primary or
secondary scatterers, and the dominant type of scattering.
The association of entropy (H) and alpha angle (α) planes is one of the most efficient and
usual means of improving the understanding of forest targets. In this classification
procedure based on target decomposition, the H - α components are plotted in a plane with
finite zones (Figure 2), which represent scattering mechanisms defined by Cloude and
Pottier (1997). These attribute spaces help defining the dominant type of scattering which
can be multiple, volumetric or superficial. The theory behind this target decomposition
method assumes that the formation of an image is a stochastic process, where an image with
a high value of entropy (0.5 < H < 0.9) presents much more details.
Fig. 2. Schematic diagram of the bi-dimensional classification based on entropy (H) and
average orientation alpha angle (α).
Source: Modified from Cloude and Pottier (1997).
In this procedure of classification by target decomposition, the scattering mechanisms are

defined in 9 distinct zones (Figure 2), namely: Z1- High entropy multiple scattering; Z2 -
High entropy vegetation scattering; Z3 - High entropy surface scatter; Z4 - Medium entropy
multiple scattering; Z5 – Medium entropy vegetation scattering; Z6 – Medium entropy
surface scatter; Z7 – Low entropy multiple scattering events; Z-8 – Low entropy dipole
scattering; Z9 – Low entropy surface scatter. Within the partition of the plan entropy and
angle α in several regions which determine and label the intensity of each scattering
mechanism, it is important to mention that Z3 is outside de representation areas from curves

I and II (Figure 2) and so it is not a non-feasible region (Hajnsek et al., 2003), because the
polarimetric SAR systems are not able to discriminate high entropy (H > 0.9). The curves I
and II represent the borders of values with maximum and minimum observable α angle, as
a function of entropy, whose limits are determined by the H- α variation for a coherence
matrix with the degenerate minor eigenvalues with amplitude m(0 ≤ m ≤ 1), according to the
Cloude and Pottier (1997).
1.1.3 Polarization synthesis and signatures

An important aspect for the characterization of forest targets is the polarization synthesis,
because when the matrix [S] of a target is known, it is possible to synthesize its backscatter
(σ) to any combination of transmitted and received polarization (Ulaby and Elachi, 1990;
Woodhouse, 2006). This is a good advantage of polarimetric data, synthesized according to
equations 9 and 10:
κ r 
Prec (ψ r , χ r ,ψ t , χ t ) = 2
. F .[K ].F t (9)
r
 Prt rec 
σ rt (ψ r , χ r ,ψ t , χ t ) = lim 4πr 2  trans 
 (10)
P
r →∞

Where: ψ = orientation angle; χ = ellipticity angle; K = vector of target; r = radius.
The polarization states of the electric fields Ev, and Eh which are associated with the radar
backscatter (σ) and the dependence of amplitude on polarization, can be represented
graphically as a function of ellipticity ( χ ) and orientation (ψ ) angles of the transmitted
wave, defining a three-dimensional surface plot called polarization signature or polarization
response (Figure 3). The polarization signatures can be used to characterize targets,
considering the more detailed knowledge of the mechanisms responsible for the scattering
of the radar signal, and the intrinsic target characteristics (van Zyl et al., 1987). It can also be
used for the polarimetric calibration of radar data. It is very important to point out that the
polarized signature of a certain target is not unique, since different combinations of
scattering mechanisms can give a similar configuration, as will be shown in this chapter.
This can be due to the structural composition of the forest target, such as the density of trees,
regular spatial distribution of trees, trunk diameter and density of twigs and branches, the
moisture content in the leaves and in the soil, the dielectric constant of the target under
study, etc. In addition, a target can also present distinct polarimetric behavior shown in its
signatures, as a function of variations from radar frequency and/or local incidence angle.
Fig. 3. Polarimetric response of a secondary succession area in the Brazilian Amazon.
It is important to observe in Figure 3 the representation of values of transmitted and

received power, where there is a minimum value (pedestal height) of polarization intensity.
This minimum value represents the amount of the non-polarized power in the received
signal, and thus is related to the polarization level of the scattered wave (CCRS, 2001). If a
single target is scattering and the backscattered wave is fully polarized, or if the signature is
calculated from a single unaveraged measurement, the pedestal height is zero. But if the
signature is calculated from an average of several samples, and there are multiple, dissimilar
scatters present or there is noise in the received signal, the pedestal height will be non-zero.
Thus the pedestal height is also a measure of the number of different types of scattering
mechanism found in the averaged samples. The pedestal of the polarimetric signature may
also be a worthy source of information for target characterization (McNairn et al., 2002). In
fact, the pedestal corresponds to the ratio of the extreme power received when the antennas
are co-polarized or cross-polarized. Therefore, the pedestal provides information equivalent
to that obtained with the coefficient of variation by van Zyl et al. (1987). The latter should be
more effective because it is not limited to co-polarizations and cross-polarizations (Touzi et
al., 2004).
2. ALOS PALSAR data over the Brazilian tropical forest

2.1. Area under study
In order to understand the mechanisms controlling the interaction radar-forest target, a test-
site was selected in the Tapajós region (NW Pará State, Brazil), delimited by the
geographical coordinates 3º 01’ 60” - 3º 10’ 39” S and 54º 52’ 45” - 54º 59’ 53” WGr (Figure
4). This region is characterized by a low rolling relief, typical of the lower Amazon plateau
and the upper Xingu-Tapajós Plateau. It is dominated by a continuous cover of primary
tropical rainforest characterized by the presence on emergent trees and an uniform
vegetation cover (Dense Ombrophilous Forests), as well as sections of low to dissected
plateaus with few emerging individuals and a high density of palm trees (Open
Ombrophilous Forests). The land use is primarily subsistence agriculture, few cash crops,
cattle raise and selective logging activities (Figure 5). Areas of secondary succession are also
found in different growth stages, resulting from land tracts abandoned after a short period
of use, with intensive subsistence agricultural practice or cattle raising.
Fig. 4. Study area (dotted rectangle) as shown on LANDSAT/ETM+ (RGB = B3, B4, B5) and
ALOS/PALSAR (RGB = HH, HV, VV) images.
Forest with selective

Primary Forest
logging
Agriculture Pasture
Fig. 5. Land use and land cover classes in the Tapajós region.
2.2 PALSAR image and field data analysis

The Advanced Land Observing Satellite” (ALOS) launched on January 24th 2006 has a “Phased-
array L-band Synthetic Aperture Radar” (PALSAR) onboard, which was developed by the
Japanese Aerospace Exploration Agency (JAXA), with a strategy for multi-temporal
imaging, at both regional and global levels. All imaging modes of PALSAR (L- band =
23cm), named as FBS (Fine beam), FBD (Fine beam dual), SCANSAR, DSN (direct
transmission) and POL (polarimetric) were calibrated and validated using a total of 500
well-distributed points around the world (Shimada et al., 2007). In this study we used fully
polarimetric data (HH, HV, VH and VV polarizations ascending mode), acquired on April
23rd 2007, with a spatial resolution of 4.50m in range and 9.50m in azimuth and an incidence
angle of ~ 24º. Using these data studies were done to characterize polarimetric signatures
and to gain a detailed understanding of the scattering mechanisms that control the radiation
response to the SAR antenna.
2.2.1 Exploratory data analysis

A methodological approach was conducted to explore the sensitivity of two polarimetric
techniques in the assessment of structural variations of some land cover classes: primary
forest (PF), forest with old timber exploitation (SL) and also, initial (ISS), intermediate
(IntSS), and advanced secondary succession (ASS).
The first polarimetric technique involved the generation of graphic representations of the
co-polarized SAR response of each forest types, as described by van Zyl et al. (1987). In
order to generate the polarization response plots, we used the polarimetric module
SPW /PCI (SAR Polarimetry W orkstation) was used (PCI Geomatics, 2007). In this
representation, a cross section of a certain type of forest cover (σ) was plotted on a bi-
dimensional graphic, as a function of all combinations of orientation angles (ψ) and
ellipticity (χ), related to the polarization ellipsis (Ouarzeddine et al., 2007). In order to derive
the polarization response for a given region of interest (ROI) - sample area that includes a
sufficient number of representative pixels of the theme, reducing the statistical uncertainties
and the influence of the speckle noise - at the SAR image, we used the average complex
value of all pixels within that ROI. The resulting surfaces were compared by visual
inspection, using information on structural characteristics derived from field inventory. It is
important to mention that all ROIs represent georeferenced forest inventory plots, where
field-based measurements were performed.
In the second exploratory analysis the objective was to performed target decomposition,
which expresses the average scattering mechanism of each forest types as the sum of
independent elements associated with physical mechanisms. The entropy values and the
average alpha angle values of each ROI, (resulting from the decomposition of eigenvalues
and eigenvectors of the coherence matrix – see section 1.2.2) were plotted in the bi-
dimensional classification space (H , α ) , following the procedure introduced by Cloude
and Pottier (1997).
After including the SAR response of each forest type in the (H , α ) bi-dimensional space,
the resulting classifications were pair-wise compared using regression analysis technique
(Freese, 1964; Neter, 1996). First a linear model was fitted for each classification pair by
regressing the percentage number of pixels classified in each zone of the (H , α ) plane. The
estimated regression coefficients where then tested to check whether ther fitted line was a 45
degree line through the origin (F test of b1 = 0 versus b1 ≠ 0; t test of b0 = 0 versus b0 ≠ 0; and t
test of b1 = 1 versus b1 ≠ 1; where b0 and b1 are the regression coefficients). This result would
indicate no significant difference between the two tested classifications at the 5%
significance level, and therefore no success of the technique in discriminating between the
two forest types.
2.2.2 Field survey

In August/September 2007 (same year as the PALSAR data acquisition) a forest inventory
was made for the physiognomic-structural characterization of these typologies, to obtain
information about the five forest types described above. Tropical forest sections were
discriminated by the occurrence (sustainable timber exploitation - SL) or non-occurrence
(PF) of human disturbances. In addition, areas of regeneration were studied in order to
represent three secondary succession stages, namely: initial (ISS) with age intervals of below
6 years; intermediate (IntSS) from 6 to 15 years; and advanced (ASS) above 15 years. This
stratification is considered the age of the natural regrowth as well as structural and floristic
characteristics of each seral stage.
Several transects were inventoried to represent each forest typologies. Within these
transects, several biophysical parameters were measured, including: diameter at breast
height (DBH), commercial and total heights, canopy gap fraction, and location of each
arboreal individual in the plot. Additionally, a botanical identification was also performed,
according to the field survey procedure described by Santos et al. (2003). As an example to
illustrate the importance of accounting for the biophysical parameters collected during the
forest inventory to analyze radar data, Figure 6 shows the structural differences between
two thematic types, referring to a primary forest and a young natural regrowth.
(a)
(b)
(c)
(d)
Fig. 6. Structural profile and dendrometric parameters of primary forest (a, c) and initial
secondary succession (b, d), respectively.
2.2.3 Results and discussion

In order to familiarize the reader with the physiognomic-structural features of the forest
typology of this section from the Amazon region under study, some results of the field
survey are briefly presented.
The initial stage ‘‘capoeiras’’ regrowth (ISS) had an average height of 8m and an average
DBH of 9cm. Floristic composition was dominated by species such as Vismia guianensis
(Aubl.) Choisy (Guttiferae), Cecropia leucoma Miquel. (Moraceae), Guatteria poeppigiana Mart.
(Annonaceae), Aegiphila sp. (Verbenaceae) and Cordia bicolor A. D.C. (Boraginaceae). The
intermediate secondary succession (IntSS) had an average height of 11m and an average
DBH of 10cm, with a higher abundance of Cecropia sciadophylla Mart (Cecropiaceae),
Schizolobium amazonicum Huber (Fabaceae), Cassia pentandra Raddi (Fabaceae), Vismia
japurensis Reichardt and Vismia guianensis (Aubl.) Choisy. (Guttiferae). The advanced
secondary succession (ASS) had an average height of 18m and average DBH of 25cm with
species Bagassa guianensis Aubl. (Moraceae), Guatteria poeppigiana Mart. (Annonaceae),
Casearia decandra Jacq. (Flacourtiaceae), Cassia pentandra Raddi (Fabaceae), Couratari
oblongifolia Ducke & R. Knuth (Lecythidaceae), and Apeiba albiflora Ducke (Malvaceae).
The primary forest stands (PF) present 3 to 4 strata with about 360 trees per hectare, whose
upper stratum shows individuals with an average height of 21 m and DBH of 27 cm. The
botanical diversity of the primary forest was significantly higher compared to areas of
secondary succession, being dominated by species such as: Protium apiculatum Swart
(Burseraceae), Picrolemma sprucei Hook (Simaroubaceae), Rinorea guianensis (Aubl.) Kuntze
(Violaceae), Cordia bicolor A.DC. in DC. (Boraginaceae), Jacaranda copaia D.Don
(Bignoniaceae), Pouteria sp. (Sapotaceae) and Inga sp. (Fabaceae). The primary forest stands
that were affected by timber exploitation (SL) present a decrease of timber volume of around
27.3 m3/hectare, which corresponds to an average density of exploration of 3.3 trees/ha, in
accordance with the forest management plan formerly established by IBAMA/PNUD
(1997). At the gaps resulting from the extraction of trees with commercial value and also at
the trails where tree trunks were dragged, quickly pioneer species install themselves, such
as Cecropia leucoma Miquel, Cecropia sciadopylla Mart., and Guatteria poeppigiana Mart.,
modifying the environment and providing resources for other species.
The description above suggest that there is a distinct frequency of trees by diameter and
height class, as well as the occurrence of characteristic species in each forest types, which
model the physical space under study. There is a positive gradient in these two ecological
parameters, increasing from initial secondary succession until a mature forest. One observes
that at the initial regrowth there is a more uniform canopy formed by 1 or 2 strata of low
height, with pioneer species with thinner trunks found repeatedly in the area, from a few
botanical families (Figure 6). The older secondary succession stages, the higher is the
structural complexity, characterized by a wide variation in tree size, with a stepwise
replacement of pioneer species and the formation of multiple canopy layers. A detailed
description of the floristic composition and structure of forest typology from the Tapajós
region can be found at Gonçalves (2007) Gonçalves and Santos (2008). The considerations
above are important to facilitate the understanding of the polarized response.
At Figure 7 the shape features of co-polarization signatures derived from the PALSAR data
of the 11 ROIs are presented, which were selected to represent the diversity of the tropical
forest in Tapajós region. In all the polarization signature plots presented, the maximum
intensity has been normalized to 1.0, according to Ouarzeddine et al. (2007). Before
discussing this subject it is important to remind that in theory, in the co-polarization
response, taking into account the representation in a plane (-90º< ψ < +90º), the highest
values of intensity for an ellipticity angle at 0º and an orientation angle at 0º, the polarization
is horizontally oriented (Evans et al., 1988; Mc Nairn et al. 2002).
P1 P2
P3 P4
P5 P6
P7 P8
P9 P10
P11
Fig. 7. Polarization responses derived from at PALSAR data acquired in the Tapajós region,
Brazil. Targets include primary forest (P1, P2), forest with old selective logging (P3, P4, P5),
advanced secondary succession (P6, P7), intermediate secondary succession (P8, P9) and
initial secondary succession (P10, P11).
The polarization responses of samples of ASS (P6, P7), IntSS (P8) and ISS (P10, P11) show a
strong similarity to the theoretical responses of short and thin conducting cylinders (i.e.
radius and lengths much shorter than the wavelength). This suggests that the scattering
produced by twigs and small branches has an important contribution for the total
backscattering of these regrowth types. The co-polarized response of samples P6 and P7
(both old regrowth) with ψ = - 70º and P8 (intermediate regrowth) with ψ = ± 90°, show
stronger returns at VV than HH, consistent with increased double reflections and present
diffusers with preferentially vertical orientation. The responses of the other analyzed areas
(i.e. P10, P11) indicate the predominance of branches oriented at ±45º in relation to the
horizontal plane, possibly due to more homogeneous canopies of these younger recovery
sections with low species diversity and a quite uniform growth rate (height intervals of 6 to
9 m corresponds to 90% of the trees) (Figure 6). One must emphasize that, taking into
account the distinction of the stands investigated, the polarization responses of sections P7
and P10 are similar to those presented by Ulaby and Elachi (1990), but with maximum co-
polarized power at ψ = - 60º and χ = 0º for P7 sample and maximum co-polarized response
at ψ = 30º and χ = -5º for P10 site and with a thin cylinder oriented vertically.
The polarization signatures of plots P1 and P2 (primary forest), P4 and P5 (forest under
exploitation) and also P9 (intermediate secondary succession), are similar those of trihedral
corner reflectors, which have a double bounce scattering geometry, when one of the reflector
sides is parallel to the line connecting the radar antenna and the corner reflector, as
mentioned by Zebker and Norikane (1987). For sections of more complex forest structures,
the highest σ values occur in linear polarizations, with a certain independence of the
orientation angle (ψ). This can be observed in plots P1, P2, P3 and P4, which represent the
mature forest (with or without legal selective logging), whose concentrated section of
maximum co-polarized power occurs in a range of ψ = 0° to ±20º when χ = 0º. As for forest
areas affected by timber exploitation, one expects predominance of double-bounce
scattering, due of the larger number of gaps resulting from the removal some trees with
commercial value and also because this type of mature forest has a well-defined vertical
structure in various strata with a much thinner undergrowth. Specifically, for the plot P3
presents most diffusers oriented horizontally (ψ = 0º) and it has a stronger return at HH
than VV, where the maximum for this signature seems slightly shifted from the horizontal
position towards the vertical. This effect might be explained by the presence of the
topographic slopes in the azimuth direction. In additional, medium or high impact logging
practices can increase the structural heterogeneity, causing changes in the polarization
response of the forest target investigated.
At the analysis of polarimetric signatures, the height of the pedestals must also be
considered because it provides an additional source of information for target
characterization, and to provide information similar to that obtained with the variation
coefficient (van Zyl et al, 1987; Mc Nairn et al, 2002). Contrary to expected, in the co-
polarized response derived from PALSAR, the highest pedestals occur for areas of
secondary succession if compared to those of forest typologies, suggesting the existence of a
considerable variation on the scattering properties of the adjacent resolution elements. The
highest values of 0.313 and 0.333 were found respectively for areas of initial (P10) and
intermediate (P8) recovery, respectively. Old-growth forests, with or without timber
exploitation (P1, P2, P3 and P4) showed pedestal height ranging from 0.068 to 0.195. Within
the characteristics of natural forest stands, the lower branch density can explain the lower
values of pedestal height, which results at a lower non-polarized power component in the
backscatter, indicating a smaller amount of volume scattering (CCRS, 2001). Note that, in
contrast to our findings, previous studies conduced in tropical zones using SAR data have
shown that a higher age and diversity of species and consequently, a stronger variability of
forest structure, results at an increase of the pedestal size on the polarization responses.
In tropical areas with complex vegetation types, the polarimetric signature of each type is
really not unique. One can simply observe that areas of early secondary succession could,
for instance, could have its’ growth influenced by aspects, such as: previous land use, soil
fertility level, bio-ecological level of fragmentation, etc. From measurements performed
during field survey we also verified that the canopy was more open in sites of initial
regrowth, localized over areas with low soil fertility and a history of more intensive use,
compared to those with higher soil fertility and a less intensive land use, both areas with the
same age of natural succession. During the intermediate phase of secondary succession, the
canopy remains more open, and the mortality of some pioneer species which are not
tolerant to shadow is higher, but presenting a larger vertical stratification. With the increase
of age from the natural regrowth, the horizontal and vertical structures become more
complex, presenting biophysical characteristics close to those of primary forest. Such

biophysical arrangements, which include the structure/density of the canopy, the
occurrence of a larger or smaller number of arboreal/bush individuals (scatterers) in these
types of vegetation formations, can lead to distinct polarimetric intra-class and inter-class
configurations, with pedestal height values which are also different, as observed on Figure 7.
Figure 8 shows the distribution of pixels extracted from the eleven ROIs in the bi-
dimensional classification space (H , α ) , whose numerical results (i.e., percent of pixels by
scattering zone) are presented in Table 1. At Figure 8 one observes that all 7 ROIs
representative of primary forest, forest with timber exploitation and advanced secondary
succession, show a concentrated pixel distribution (~70%) at zones Z4, Z5 and Z9. This
indicates that the polarimetric response of such classes from already defined forest structure
(containing 3 or 4 strata) is based mainly on multiple and volumetric scattering (the
backscattering signal is mainly a function of geometry of scatters elements from dense
cover) at zones of medium entropy. Additionally there is also some influence of a surface
scattering but derived from a zone of low entropy. The two representative plots of classes
from intermediate recovery have a higher distribution of pixels (>60%) at the zone of
medium entropy (Z5), due to dipole scattering objects and in zones of low entropy (Z9, Z7),
with scattering objects regularly distributed in space, and also as a consequence of the
double-bounce effect caused by scattering objects such as twigs-trunk-soil-trunk. The initial
secondary succession is positioned preferentially in the space of attributes of low entropy
(Z9) and also configured by double-bounce interaction mechanisms (medium entropy
multiple scattering – Z4), affected by the propagation of the canopy, as a result of the
growth uniformity of pioneer species where there is a predominance of only an upper
homogeneous stratum, and a lower dispersed one, formed by small plants which germinate
and develop in a late succession process, in the shadow of the pioneers.
Fig. 8. L-band PALSAR α-H distribution for scattering mechanism: green colour are pixels of
primary forest, forest with old selective logging and advanced secondary succession; red
colour are initial and intermediate secondary succession. Source: Santos et al. (2009).
ROIs (Vegetation Types)

Zones PF SL ASS IntSS ISS
P1 P2 P3 P4 P5 P6 P7 P8 P9 P10 P11
Z1 0 0 0 0 0 0 0 0 0 0 0
Z2 0 0 0 0 0 0 0 0 0 0 3.7
Z3 0 0 0 0 0 0 0 0 0 0 0
Z4 8.16 20.4 14.89 35.29 24.49 24.24 42.86 18.52 0 10 33.33
Z5 46.94 26.5 29.79 21.57 14.29 30.3 17.86 14.81 50 26.67 3.7
Z6 8.16 10.2 10.64 11.76 10.2 18.18 3.57 11.11 11.54 13.33 14.81
Z7 10.2 18.4 10.64 17.65 18.37 15.15 7.14 7.41 11.54 13.33 18.52
Z8 12.24 4.1 0 3.92 2.04 0 0 7.41 7.69 6.67 0
Z9 14.29 20.4 34.04 9.8 30.61 12.12 28.57 40.74 19.23 30 25.93
Table 1. Percentual of pixels distribution of each ROI at zones of two-dimensional
classification space (H , α ) .
The statistical tests used to detect variations among classifications of pairs of ROIs, at 5%
significance values, indicate that there present no significant differences. Low p values were
observed when comparing SL/ASS, PF/ASS and SL/IntSS, with values of 0.1725, 0.2661 and
0.2754 respectively, which correspond to the largest differences found in polarimetric
responses among pairs of ROIs, but they are not significant. Since there are physiognomic-
structural differences in forested targets (with or without timber exploitation) and those at
younger secondary successional stages, one can infer that, at this level of significance (5%),
this classification procedure (H , α ) by target decomposition of the PALSAR data was not
enough robust to detect such a variability. As observed during the previous discussion, we
made the identification and definition of intensity for each scattering mechanism of the
pixels contained in the PALSAR image, corresponding to each type forest typology, within
these zones of the alpha-entropy space. This procedure should facilitate the attribution of
each zone to a certain thematic class. But, due to the structural complexity and floristic
diversity of the tropical landscape investigated and of its intrinsic scatterers, in this case
there was a limitation to the procedure described by Cloude and Pottier (1997), because
there is a fixed position of the decision borders of the classifier used. Nevertheless, from this
partition of the alpha-entropy plan, new approaches which consider statistical modelling to
treat POLSAR data (Pottier and Lee, 2000; Ferro-Famil et al., 2008) are being used with a
relative success, as shown by Freitas et al. (2008), for the analysis of land use/land cover of a
section from the Amazon region.
3. Conclusions
This chapter gave an overview of the utility from SAR polarimetric data to discriminate
tropical forest types, whose analysis procedures were based on the knowledge of the shape
from the co-polarized signature and also on the type and intensity of the dominant
scattering mechanisms. All this allows the analyst to distinguish the space of attributes
occupied by features of a tropical landscape and to establish which parameters are more
efficient for a thematic discrimination. It allows also to define the level of agglutination
among typologies, of subtle physiognomic-structural gradients at a classification from SAR
images.
Taking into account the experience gathered with the use of ALOS/PALSAR data (L-band)
in the Brazilian Amazon, based on the classification by target decomposition supported by
information derived from forest inventories, the following conclusions are made: (a) at
several types of landscapes studied there is a predominance of scattering processes with
medium entropy, with volumetric and multiple scattering, influencing mainly those
structures with more forests; (b) areas under natural recovery (initial or intermediate)
present mechanisms of surface type (low entropy) and double-bounce affected by canopy
propagation (medium entropy); (c) the classification method by target decomposition based
on entropy values and of the mean alpha angle was not robust enough to detect the floristic-
structural variability existing among certain land cover classes; (d) the parallel polarization
responses obtained from several forest cover classes show different configurations,
indicating that the backscatter of certain plots was dominated by distinct physical
mechanisms. A more detailed analysis of species composition and its arrangements and
proportions as dominant and co-dominant in the vertical structure of forest stands must be
done, in order to improve understanding of different co-polarized signatures and the
interaction mechanisms that command the radar signal and the target. The analysis of
PALSAR cross-polarized mode data is another approach under development. Presently, we
are analyzing cross-polarized PALSAR data aiming at a higher accuracy at the thematic
classification as a consequence we are expecting an improvement of biomass modelling in
tropical forest.
4. Acknowledgements
We acknowledge to CNPq for the research grants; the IBAMA/MMA and LBA for the
logistic support in the study. We also thank our colleagues J. C. Mura, F. F. Gama, A. S.
Queiroz, S. J. S. Sant’Anna, H. J. H. Kux (INPE) and Jakob van Zyl (JPL/NASA) for
discussions during the development of this chapter.
5. References
Andersen, L.; Granger, C.; Reis, E.; Weinhold, D.; Wunder, S. (2002). The dynamics of
deforestation and economic growth in the Brazilian Amazon. Cambridge:
Cambridge University Press. 259p. [ISBN 0 521 81197 X].
Beaudoin, A.; Le Toan, T.; Goze, S.; Nezri, E.; Lopez, A.; Mougin, E.; Hsu, C.C.; Han, H.C.;
Kong, J.; Shin, R.T. (1994). Retrieval of forest biomass from SAR data. International
Journal of Remote Sensing, 15(14): 2777-2796.
Canada Centre of Remote Sensing - CCRS, (2001). Radar polarimetry: advanced radar
polarimetry tutorial. In: Fundamentals of Remote Sensing. 97p.
<http://ccrs.nrcan.gc.ca/resource/tutor/polarim/pdf/polarim_e.pdf>.
Cloude, S. R.; Pottier, E. (1997). An entropy based classification scheme for land application of
polarimetric SAR. IEEE Transactions on Geoscience and Remote Sensing, 35(1): 68-78.
Cloude, S.R.; Pottier, E. (1996). A review of target decomposition theorems in radar
polarimetry. IEEE Transactions on Geoscience and Remote Sensing, 34 (2): 498-518.
Coops, N.C. (2002). Eucalyptus forest structure and synthetic aperture radar backscatter: a
theoretical analysis. Trees, 16: 28-46.
Cressie, N. A. (1993). Statistics for spatial data, New York: John Wiley & Sons. 900p.
European Space Agency – ESA. (2009a). Single vs multi polarization SAR data. 28p.
<http://earth.esa.int/polsarpro/tutorial.html>.
European Space Agency – ESA. (2009b). Polarimetric decompositions. 28p.
<http://earth.esa.int/polsarpro/tutorial.html>.
Evans, D. L.; Farr, T. G.; van Zyl, J. J; Zebker, H. A. (1988). Radar polarimetry: analysis tools and
applications. IEEE Transactions on Geoscience and Remote Sensing, 26(6): 774 – 789.
Ferro-Famil, L.; Pottier, E.; Lee, J. S. (2001). Unsupervised classification of multifrequency
and fully polarimetric SAR images based on the H/A/Alpha-Wishart classifier.
IEEE Transactions on Geoscience and Remote Sensing, 39(11): 2332-2342.
Freeman, A.; Durden, S. L. A. (1998). Three-component scattering model for polarimetric
SAR data. IEEE Transactions on Geoscience and Remote Sensing, 36(3): 963-973.
Freese, F. (1964). Linear regression methods for forest research. U.S. Forest Services
Research, Madison, Wisconsin.138p.
Freitas, C. C.; Soler, L. S.; Sant’Anna, S., J. S.; Dutra, L. V.; Santos, J. R.; Mura, J. C.; Correia,
A. H. (2008). Land use and land cover mapping in the Brazilian Amazon using
polarimetric airborne P-band SAR data. IEEE Transactions on Geoscience and
Remote Sensing, 46(10): 2956 – 2970.
Gonçalves, F. G. (2007). Avaliação de dados SAR polarimétricos para estimativa volumétrica de
florestas tropicais. São José dos Campos, Brazil, Master in Degree in Remote Sensing
at National Institute for Space Research. INPE:14777-TDI/1230. 107p.
Gonçalves, F. G.; Santos, J. R. (2008). Composição florística e estrutura de uma unidade de
manejo florestal sustentável na Floresta Nacional do Tapajós, Pará. Acta
Amazonica, 38(2): 229-244.
Hajnsek, I.; Pottier, E.; Cloude, S. R. (2003). Inversion of surface parameters from polarimetric
SAR. IEEE Transactions on Geoscience and Remote Sensing, 41(4): 727-739.
Hawkins, R. K.; Attema, E.; Crapolicchio, R.; Lecomte, P.; Closa, J.; Meadows, P. J.;
Srivastava, S. K. (2000). Stability of Amazon backscatter at C-band: spaceborne
results from ERS-1/2 and Radarsat-1. In: SAR Workshop – CEOS Committee on
Earth Observation Satellites. Toulouse, France, 26-29 Oct., 1999. Edited by Robert A.
Harris and L. Ouwenhand. Publisher: Paris, European Space Agency, ESA-SP,
vol.450, p.99-108. (ISBN:9290926414).
Hendersen, F. M.; Lewis, A. J. (1998). Radar fundamentals: The geoscience perspective. In:
Ryerson, R.A. ed., Principles & Applications of Imaging Radar. Manual of Remote
Sensing, 3. ed. New York, John Wiley & Sons, Inc., vol.2, cap.3, p.131-181.
Hoekman, D. H.; Quinones, M. J. (2000). Land cover type and biomass classification using
AirSAR data for evaluation of monitoring scenarios in the Colombian Amazon,
IEEE Transactions Geoscience and Remote Sensing, 38 (2):.685–696.
Instituto Brasileiro do Meio Ambiente e dos Recursos Naturais Renováveis/Programa das
Nações Unidas para o Desenvolvimento - IBAMA/PNUD. (1997). Plano de manejo
florestal da Floresta Nacional do Tapajós, Pará, Brasil. IBAMA, Brasília, Distrito
Federal. 109pp.
Kugler, F.; Papathanassiou, K. P.; Hajnsek, I. (2006). Forest height estimation over tropical
forest by means of polarimetric SAR interferometry. In: Seminário de Atualização
em Sensoriamento Remoto e Sistemas de Informações Geográficas Aplicados à
Engenharia Florestal, 7., Curitiba, Paraná, 17-19 out., 2006. Anais. pp.504-512.
[CDROM].
McNairn, H.; Duguay, C.; Brisco, B.; Pultz, T.Z. (2002). The effect of soil and crop residue
characteristics on polarimetric radar response. Remote Sensing of Environment,
80(2): 308-320.
Neeff, T.; Biging, G. S.; Dutra, L. V.; Freitas, C. C.; Santos, J. R. (2005b). Markov point
processes for modeling of spatial forest patterns in Amazonia derived from
interferometric height. Remote Sensing of Environment, 97: 484 – 494.
Neeff, T. ; Dutra, L. V. ; Santos, J. R.; Freitas, C. C.; Araujo, L. S. (2005a). Power spectrum
analysis of SAR data for spatial forest characterization in Amazonia. International
Journal of Remote Sensing, 26(13): 2851-2865.
Neter, J.; Kutner, M. H.; Nachtsheim, C. J.; Wasserman, W. (1996). Applied linear statistical
models. 4.ed. Boston: McGraw-Hill. 1408 p.
Ouarzeddine, M.; Souissi B.; Belhadj_Aissa, A. (2007). Classification of polarimetric SAR
images based on scattering mechanisms. In: Spatial Data Quality, 12-15 June, 2007.
ITC, Enschede. http://www.itc.nl/ISSDQ2007/proceedings/poster.html
PCI Geomatics. (2007). Geomatica SPW 10.1 User Guide (SAR Polarimetry Workstation),
PCI Geomatics Enterprises Inc., Richmond Hill Ontario, Canada, 179 pp.
Pope, K. O.; Rey-Benayas, J. M.; Paris, J. F. (1994). Radar remote sensing of forest and
wetland ecosystems in the Central American tropics. Remote Sensing of
Environment, 48(2): 205-219.
Pottier, E.; Lee, J. S. (2000). Unsupervised classification scheme of POLSAR images based on
the complex Wishart distribution asnd the H/A/Alpha – Polarimetric
decomposition theorem. In: European Conference on Synthetic Aperture Radar, 3,
2000, Munich, Germany. Proceedings. [CDROM].
Santos, J. R.; Freitas, C. C.; Araujo, L. S.; Dutra, L. V.; Mura, J. C.; Gama, F. F.; Soler, L. S.;
Sant’Anna, S. J .S. (2003). Airborne P-band SAR applied to the aboveground
biomass studies in the Brazilian tropical rainforest. Remote Sensing of
Environment, 87(4): 482-493.
Santos, J. R., Mura, J. C., Paradella, W. R., Dutra, L. V. ; Gonçalves, F. G. (2008). Mapping
recent deforestation in the Brazilian Amazon using simulated L-band MAPSAR
images. International Journal of Remote Sensing, 29(16): 4879 - 4884.
Santos, J. R.; Narvaes, I. S.; Graça, P. M. L. A.; Gonçalves, F.G. (2009). ALOS PI 2008
Symposium. Proceedings, Nov., 3-7th, 2008., Island of Rhodes, Greece. (ESA-664).
Santos, J. R.; Pardi Lacruz, M. S.; Araujo, L. S.; Keil, M. (2002). Savanna and tropical
rainforest biomass estimation and spatialization using JERS-1 data. International
Journal of Remote Sensing, 23(7):1217-1229.
Shimada, M.; Isoguchi, O; Tadono, T., Higuchi, R.; Isono, K. (2007). PALSAR CALVAL –
Summary and update 2007. IEEE Transactions on Geoscience and Remote Sensing,
p. 3593-3596.
Touzi, R.; Boerner, W. M.; Lee, J. S.; Lueneburg, E. (2004). A review of polarimetric in the
context of synthetic aperture radar: concepts and information extraction. Canadian
Journal of Remote Sensing. 30 (3): 380-407.
Treuhaft, R. N.; Chapman, B.; Dutra, L.V.; Gonçalves, F. G.; Santos, J. R.; Mura, J. C.; Graça,
P. M. L. A.; Drake, J. (2006). Estimating 3-dimensional structure of tropical forest
from radar interferometry. Ambiência, 2(1): 111-119.
Ulaby, F. T.; Elachi, C. (1990). Radar polarimetry for geoscience applications. Norwood:
Artech House, Inc., 388 p.
van der Sanden, J. J. (1997). Radar remote sensing to support tropical forest management.
Tropenbos-Guyana Series 5: Georgetown, Guyana, PhD Thesis Wageningen
Agricultural University. ISBN: 90-5485-778-1. 330p.
van Zyl, J. J., Zebker H. A., Elachi C. (1987). Imaging radar polarimetric signatures: theory
and observation. Radio Science, 22(4): 529-543.
Woodhouse, I. H. (2006). Introduction to microwave remote sensing. Boca Raton: CRC Press
Taylor & Francis Group, 370p.
Wulder, M.; Niemann, K.; Goodenough, D. (2000). Local maximum filtering for the
extraction of tree locations and basal area from high spatial resolution imagery.
Remote Sensing of the Environment, 73(1): 103-114.
Zebker, H. A.; Norikane, L. (1987). Radar polarimeter measures orientation of calibration
corner reflectors. In: Proceedings of the IEEE, vol. 75, n° 12, pp. 1686-1688.
Zhang, L.; Zhang, J.; Zou, B.; Zhang, Y. (2008). Comparison of methods for target detection
and applications using polarimetric SAR image. PIERS Online, 4(1): 140-145.
Discovery of user-generated geographic data using web search engines 207
X9
Discovery of user-generated geographic

data using web search engines
Carlos Abargues, Carlos Granell, Laura Díaz, Joaquín Huerta
and Arturo Beltran
Universidad Jaume I de Castellón
Spain
1. Introduction
Traditionally the discovery of geographic data has been performed locally, from a
centralized, desktop-based perspective. In the last decade, though, the notion of Spatial Data
Infrastructure (SDI) (Masser, 2005) appeared as a net of interconnected SDI nodes to build
spatial information infrastructures at different scales and levels (local, regional, national,
European and even global) (INSPIRE, 2007). These SDI nodes represent an effort to link,
share, and exchange geographic data from different and disperse sources. Discovery
geographic data through this immense network of SDI nodes has been usually delegated to
the use of catalogues services (http://www.opengeospatial.org/standards/cat). These
specialized services base their functionality on the use of metadata that describes the
geographic content they register to offer users and client applications effective interfaces for
searching and publishing that content.
With the advent of the popular Web 2.0 the traditional Web moved towards a more social
platform where the roles of content producer and content consumer have been diluted and
often played by the same actor. The appearance of new technologies and applications
(AJAX, Web APIs, social networks, etc.) offered new possibilities that facilitated and
accelerated the content creation and sharing using the Web as the basic environment. Some
of them deal specifically with geographic content and provide a new way to create and
publish geographic content by people independently of their technical skills or knowledge
on Geographic Information System (GIS). This new trend is currently known as
Neogeography (Turner 2006). Millions of people are currently potential data providers and
are able to create and share fresh geographic content for a given area acting as truly sensors
(Goodchild 2007). All these quantities of geographic data on the Web have made the concept
of the Geospatial Web (Scharl & Tochtermann, 2007) to gain relevancy. This term defines the
use of the Web in such a way that the content has a geographic component defining its
geospatial location in the world that is used to access to it and also linked it with the
resources. In deep the Geospatial Web could be seen as the product of organizing the Web
based in the location of its resources.
The openness, simplicity and easy of share of the new Web 2.0 tools and services also
facilitate the content creation what results in an increase of geographic data in a proportion
never seen till the date. The rate and speed at which the quantity of content increases and
the nature of most of the content creators, most of them without deep technical skills, make
the use of catalogues with its strict metadata requirements and its user-directed process of
content submission an ineffective solution for its management. For these reasons new
methods for discovering and managing this user-generated content should be used. During
the evolution of the Web its size started to reach a dimension that made the task of
searching for information along directories with linked documents a problem. This problem
could happen again with the almost unmanageable quantity of content. In the first case the
web search engines (Google, Yahoo, AltaVista) appeared with successful results becoming
an essential tool for almost any Web user and for innovative and distinct uses (Al-Masri &
Mahmoud, 2008). Maybe these applications can represent again a solution for the discovery
of geographic data in the actual Web.
2. Background
2.1 Web 2.0 and the Geospatial Web.
Despite the fuzziness that use to follow its definition the term Web 2.0 does not refer to any
new and different version of the traditional Web but a change in the technology and design
to improve its functionality, communications, information sharing, creativity and
collaboration along it. This new vision of the Web could be considered as the evolution and
merge of different streams such as the technological improvement, the appearance of new
applications and also the socialization of the Web, which comprises the participation and
contribution of the users in it (Vossen & Hagemann 2007).
Still a few years ago the model of content consumption on the Web was represented by a
commonly used client server model. In this model the content provider acted as server,
producing content that was directly used by the client or content consumer defining clearly
each role. With the spreading of the Web 2.0 philosophy this model completely changed
dissolving the limits of each role. Nowadays the Web is full of services and tools with one
major objective, sharing content (photos, videos, bookmarks, knowledge, etc.) among users.
The Geospatial Web represents the merging of different types of information that are
already present on the Web (i.e. HTML pages, images, etc.) with geographic content. This
facilitates discovering and searching any type of content based on its geographical
component. In other words, the Geospatial Web structures each piece of Web content
(photos, videos, web pages, 3D models, etc) according to its geospatial location, what is
called georeferencing. This linkage between data and its geographic location enables their
discovery and use by location approaching, as defined by the Geospatial Web.
2.2 Geographic services and user-generated data

Since the appearance of the first web-based GIS systems a lot of tools and services have
emerged. It is especially in these last years when more and more services appeared probably
thanks to the entering of big companies in the business of the geographic information (i.e.
Google, Yahoo or Microsoft), the increase and improvement of the Internet communications
and the spread use of positioning devices such as GPS receivers. Probably the most
common of these tools are the web mapping services (Mitchell 2005). There exist many web
mapping service implementations available, both from private companies such as Google
Maps, Yahoo! Maps, Bing Maps (formerly Microsoft Live Maps), MapQuest and much
others, and integrated in Geoportals (Bernard et al., 2005) present usually in SDI nodes.
Not only the web mapping services have become popular but also a new type of desktop
applications have recently emerged to visualize geographic content. Geobrowsers (virtual
globes or digital earths) offer a three-dimensional view of the earth where the geographic
content is displayed over that virtual representation of the planet’s surface. These tools offer
totally new ways for visualizing data and they have started a process of change from the
two-dimensional to a more realistic method of visualizing data. Users can find geobrowsers
such as Google Earth, NASA World Wind and ESRI’s ArcGIS Explorer. These services are
not limited to show to end users simple cartography and are also enriched with different
sources of information. For instance users can find addresses and directions, business, traffic
conditions or even content created by other users. These applications do not offer just the
visualization of content but also provide a searching mechanism over the geographic data
they manage.
The use of the web-based and desktop applications is extended with the release of
Application Programming Interfaces (API) and Software Development Kits (SDK). In the
first case APIs enable third parties developments to use geographic services and data
through calls to the API in their own projects. For instance, anybody can create geographic
data and visualize it using either Google Maps or Google Earth in their own website. The
use of these APIs popularized the term mashups that denotes the creation of new web
applications for specific purposes by combining and integrating remote data from other
distributed sources. The SDK offers the opportunity of freely reuse or extend desktop
applications such as NASA World Wind or ESRI’s ArcGIS Explorer enabling the users to
create their own applications with this capabilities.
The emerge of all these new brand services and tools along with the vision of content
creation and sharing in Web 2.0, have led to new ways to create and distribute geographic
content. Probably one of the pioneering and most important movements nowadays is the
OpenStreetMap project whose objective is the collaborative creation of freely available
cartography all around the world (Haklay & Weber 2008). This cartographic data is
collected, edited and uploaded by its own users. Another example is the concept of Public
Participation GIS (PPGIS) (Nyerges et al., 1997) that enables participation of a given
community in the future urbanism development of their own area. Not only projects such as
the OpenStreetMap promote this type of geographic content creation but also private
companies offer their own tools to do so. This is the case of Google My Maps and Google
Map Maker, this last one intended also to facilitate the creation of cartography for those
countries that experience a lack of it.
2.3 Keyhole Markup Language: The Geospatial Web’s HTML.

The Keyhole Markup Language (KML) is a XML-based language designed to express
geographic annotation and visualization. The geographic visualization includes not only the
representation of the graphical data but also establishes order and control over the data
navigation. KML is used in a broad range of applications such as web mapping services,
two-dimensional maps including those used in mobile devices and geobrowsers. Most of
these applications or services also make use of KMZ files. These ones are basically
compressed files containing a KML file and other resources such as images or icons that
could be referenced on it. KMZ files facilitate the sharing and distribution of content
encapsulating in one single file all the possible resources needed to visualize and work with
the geographic content.
KML was originally created by Keyhole Inc. a company founded in 2001 and specialized in
software development for geospatial data visualization. Its main application suite was called
Earth Viewer that was transformed into Google Earth after the acquisition of the company
by Google in 2004. Currently a broad range of applications dedicated to visualize and
operate with geographic data are using KML as one of their supported file formats. Some
examples of these applications that make use of KML files are ESRI’s ArcGIS Explorer
(http://www.esri.com/software/arcgis/explorer/index.html), NASA’s World Wind
(http://worldwind.arc.nasa.gov), Google Earth (http://earth.google.com) and Google
Maps (http://maps.google.com). Not only applications that could be considered clients are
employing this format but also what could be considered as geographic data servers export
information in KML format. This is the case of GeoServer (http://geoserver.org), one of the
most spread used software to serve geographic content that supports KML and KMZ output
for Web Mapping Service (WMS) (http://www.opengeospatial.org/standards/wms)
requests. The success of KML is a reality partly thanks to its use and promotion by Google but
also because its adoption as OGC standard
(http://www.opengeospatial.org/standards/kml). KML is since 2008 one of the OGC
standards what ensures its continuity, improvement and interconnection with other OGC
standards. This last point opens KML to work with other OGC standards such as WMS and
WFS. This interconnection and collaboration between standards could be seen in most of the
actual geobrowsers since KML allows the inclusion of links that represent WMS requests
and which response is directly visualized over the virtual globe (Foerster et al., 2008)
2.3.1 Structure
The KML language structures information by means of specific elements or tags. It is based
on an object-oriented model where KML defines a set of objects employed to build the
corresponding files. Some of these objects are abstract which from an object-oriented
perspective means that these are not implemented or used in KML files however they play
an important role for structuring the other objects within the languages definition and to
keep a hierarchical structure. For instance some of these abstract objects are used as parents
of others letting in this way inheriting common properties. Figure 1 represents the object
hierarchy for the KML language, in which a given object inherits their parents’ attributes
and elements and these are inherited in turn by their child elements.
Fig. 1. KML object-oriented structure.
In KML everything inherits from the abstract element Object that only specifies the use of an
id (identifier). Whatever is represented in KML derives from the abstract element Feature.
The different features defined by the standard are:
• NetworkLink represents a link to remote resources, including images, photos or even
WMS requests;
• Placemark represents a Feature with an associated geometry;
• PhotoOverlay, ScreenOverlay and GroundOverlay used to superpose images over the
virtual globe represented in the geobrowsers; and
• Folder and Document elements used to order and structure KML files as it could be
done with files and folders in any Operating System.
All these features have some common elements inherited from Feature and other self-
elements and attributes. Probably the most important elements in any feature are those that
textually describe them. These elements are for instance the Title, Description and Snippet.
The first two elements are self-explanatory and the last one represents a short description
(usually composed by no more than a couple of lines) that represents a quick description of
the feature and that uses to be found on the lists or tree views that represents the different
elements visualized in the geobrowser.
One interesting characteristic regarding the element Description is its acceptance of a
restricted subset of HTML tags to be used as the feature’s description. This fact allows the
use of HTML not just to enhance the visual appearance of the feature’s description but it
also opens the door to use images, hyperlinks to related resources or even embed more
complex elements such as Adobe Flash Video videos from services like YouTube.
2.3.2. Geographic annotation

KML can represent a set of different types of geometries that can be used depending on the
user’s requirements. These geometries include Point, LineString, LinearRing represented by a
totally closed line, Polygon, MultiGeometry that represents a compound of other geometries
and finally Model, usually employed to represent 3D models.
The standard defines other elements such as Region to specify concrete areas on the virtual
globe. These regions are defined by a set of parameters including coordinates and height to
define an area of interest. All the elements associated with regions become active as users
navigate over such regions. This KML element is especially useful when working with
complex KML files that contain a big quantity of features. Enabling only the features just on
the region where they are placed reduce the possible overload caused by their visualization.
2.3.3 Visualization aspects

KML offers a long list of elements focused on the visualization of geographic content. For
instance the user can find elements to define the style of any KML element that could be
visualized such as lines, polygons, icons and others. The user has also control to define how
the end-user would visualize the content on a geobrowser. To understand this it is helpful to
imagine a virtual camera controlled with the keyboard or the mouse. This camera has some
parameters such as its position in the three dimensions or orientation. These parameters can
be specified within a KML file defining what could be called views for driving the end-users
attention to a given feature or position over the globe.
2.3.4 Advanced aspects

Because events take place in a concrete moment in time and they use to exhibit an evolution
on it, KML defines elements to associate features with punctual moments or periods of time.
These elements become really useful for representing the evolution of a given fact (e.g.
meteorological data) on geobrowsers, enabling the representation of features not just in
space but also in time and approaching the use of four dimensions (three spatial plus one
temporal). Another important element present in the standard is the ExtendedData, which
allows the insertion of custom XML code into the KML file. Depending on the usage of the
ExtendedData child elements the users can specify their own XML Schema
(http://www.w3.org/XML/Schema) or import already existing ones to be used within
features defined in the file. This element represents a great chance in order to enrich the
KML file with extra information and metadata. As we will see more in deep in the next
section, these elements represent an effective solution for adding structured metadata or
reuse existing one that follows a given XML Schema.
Finally KML represents a really flexible language allowing its extension with new and
specific features (aggregating new object derived elements). Good examples of new
functionalities implemented by extending its object-oriented model are the Tour object by
which users experience virtual tours over the globe using Google Earth. In this case Google
extended the KML version 2.2 specification with new elements to enable this type of feature
not supported by the default standard.
Thanks to the growth in its use, its capabilities to visualize and annotate geographic content,
its relative simplicity and its easy to extend model, some people consider that KML it is in
the process of becoming the Geospatial Web’s HTML.
3. Traditional discovery mechanisms

When somebody goes to a bookstore looking for a good thriller he or she would probably
start looking for in the thriller’s specific section. Then this potential buyer would start
looking for a specific title that probably better match what is looking for based for instance
in the period of history or place where the story would take place. It is also probable that at
first sight the information provided by the title is not enough for the buyer so reading the
summary in the back cover would become necessary since it offers more information about
the book and sometimes the author and the critique. If this is not enough maybe the user
could open the book looking for the editor details or year of publication and in last instance,
start reading the book to try to see if the book is the one that better matches the buyer’s desire.
This example can be taken to illustrate the steps followed when somebody tries to discover
some kind of information. The first step is decide where to start looking for. In the example
the buyer is looking for a thriller so the obvious action is to start looking in the thriller
section in the bookstore. Its equivalent in the Web and Web 2.0 would be probably open in
the web browser a web search engine website such as Google, Yahoo or Microsoft’s Bing or
a more specialized site depending on the user’s preferences and knowledge. The second step
in the process would be to check the book information based on a given criteria. This implies
the use of information about the content. In the book’s case it is quite easy since it is
composed by text however this is not always the case when searching for resources in a
heterogeneous environment, as it is the Web. This is full of none textual content such as
images, audio or video files and in last instance geographic content. Since most of the
searching interfaces present on the Web are based on filling in some searching criteria, it is
necessary to describe properly these resources in order to find them. Resource descriptions
are called metadata and represent a key element in the discovery of any type of content and
mostly in those distributed and heterogeneous networks like the Web.
3.1 Metadata
Metadata could be defined as data about data (Craglia et al., 2007) and in some cases like in
GIS area also data about services. Metadata aims to explain the meaning of this data and
services and to facilitate its understanding and use by different users or even automatic
agents. The use of metadata it is essential for a broad range of activities and applications
including the discovery of information, cataloguing of resources and for other ways of data
processing. In the Geographic Information area and more precisely in the study area of the
Spatial Data Infrastructures (SDI) metadata is considered an essential component that acts as
the glue that keeps all the pieces together (Masser, 2005; Rajabifard et al., 2002).
The study on metadata is not a new one and a lot of effort has been invested on it. Since the
appearance of the first studies on metadata in computing science some standards have
appeared. Probably the most used and most well known is the Dublin Core Metadata
Initiative (DCMI) or simply Dublin Core (DC) (http://dublincore.org). This format defines a
multipurpose metadata standard composed by a set of 15 basic elements designed to
facilitate the discovery of electronic resources. DC is used in a broad range of applications
since its design allows the addition of basic information to any type of content.
Among others reasons DC is probably one of the most extended metadata standard because
its simplicity and general purpose. This is in general a good characteristic for a standard
however in given areas or applications more domain specific information could be required
for a given resource. This is the case for instance of the Geographic Information area. For
this reason, several initiatives and standards have born and are currently in use in the
geospatial domain. This is the case of the ISO 19115 (ISO 19115:2003) and the ISO 19139 (ISO
19139:2007). The ISO 19115 is an international standard for geographic content metadata and
was created in 2003 by the International Standardization Organization (ISO) providing an
abstract model for the organization of geospatial metadata. The ISO 19139 standardises the
expression of 19115 metadata using XML and its derived logical model.
The complexity and requirements of most of the actual metadata standards for geographic
information reserves them for the more specialized users. These metadata formats do not
seem a solution for the description of content created by non-specialized users that maybe
are not as interested in exhaustively describing content as more experienced user and
professionals could be. It is then necessary to find ways to add the minimal description to all
this new content in such a way that could be easily performed by any user and at the same
time offer enough information for data discovery and cataloguing. These new solutions
should be as simple as possible, allowing the use of existing resources and standards and
probably observed from a format-dependent perspective in a first stage.
3.2 Catalogues
Catalogues are currently the most used method for geographic content indexing in the GI
field. These systems perform all the operations that defines the term indexing when talking
about searching systems. Therefore the catalogues collect, process, and store the data for its
posterior retrieval. The collection process commonly requires of the user’s interaction with
the system since in most of the cases the content must be uploaded with its metadata
manually. This metadata is the key element in the indexing process since it is the
information used for the content retrieval and not the content itself.
GeoNetwork (http://geonetwork-opensource.org) is one of these catalogue applications
that also implements the OGC Catalogue Service standard specification. Data providers
simply upload their geographic data and metadata directly to the catalogue service
application in order to make them available to others. Geographic data and metadata are
then processed and stored in the application for being retrieved through discovery interfaces
exposed by Geonetwork. This process involves the immediate publication, storing and
accessibility of the content but also the human interaction and supervision during the process.
3.3 Web search engines

Besides the catalogues, web search engines are probably the most used discovery
mechanism for general-purpose searching. Web search engines carry out three main tasks:
crawling, indexing and searching. Web crawlers also known as spiders or bots perform the
crawling automatically. This task consists of visiting as much resources as possible
collecting the information for its indexing and also adding all the links present in the
resources that are registered for its posterior analysis. The task of indexing in the web and
more specifically in the search engine area refers to the process by which data is collected,
processed and stored in convenient manner to facilitate its retrieval as the catalogues do
specifically for geographic content. Basically all the relevant information is extracted from
the resources and stored conveniently in a database or index for its rapid retrieval. Finally,
the task of retrieval or searching is performed through specific interfaces. Usually these
interfaces allow the insertion of text or keywords that will be searched in the index.
4. Mechanisms for discovering user-generated data

4.1 KML as a metadata container
Although KML is probably popular for its simplicity and its visualization capabilities, this
standard defines a set of elements that allow adding more metadata about the content to
visualize in formats that could be reused for other means. Rather than representing a format
focused on the visualization of data, KML can act as a metadata container, i.e., a format
suitable to transport metadata together with the data itself.
4.1.1 Identifying elements

When adding metadata within a KML file the first question to answer to is where to add
descriptive information. As it has been already explained in the Background section, the
KML standard specification defines a set of elements that can contain not just the
description of a given feature but also more detailed information. Usually metadata is
expressed using text. This fact limits the number of KML elements that can contain metadata
since some of these elements are not designed to offer textual information. This is the case of
those elements that just offer information about visualization, for instance all those
employed to specify the virtual camera position or geometry of a feature.
However KML offers some other elements that could transport textual metadata effectively.
Looking at the KML schema in Figure 1 the best options seem all those elements derived
from the abstract element Feature. This makes sense as far as somebody could be interested
in adding metadata about a placemark, about other resources linked within the document,
imagery or documents and folders of KML resources. Thanks to these last two KML
elements the user can also decide at which level needs to add metadata. The elements
Document and Folder are designed to organize and contain other KML features such as
Placemark, NetworkLink or even other Folder and Document elements. Considering this, the
user is able to add metadata to a given Placemark, however it is also possible to add
metadata to the Document or Folder that contains it. Imagine the scenario where a set of
Placemark elements is organized inside a Document element and this is part of a Folder that
contains other Document elements with different Placemarks. In this scenario the user is able
to add placemark-specific metadata for each of the Placemark elements, add metadata at
Document level with information that is common to all the Placemark in the Document and
finally add more information shared by all the documents at Folder level. In brief, KML
allows choosing in more or less detail the granularity when adding metadata to a project.
4.1.2 Identifying child attributes

Once the suitable elements that could act as metadata containers are identified, the next step
is to choose which of their child elements can contain metadata information.
Since all the elements previously described inherit from the Feature element, all of them
share some common child elements. Some of them allow the inclusion of textual information
or metadata such as Title, Snippet and Description. The objective of the element Title seems
obvious, to describe the title or name of a Feature within the KML file. In almost all the
existing metadata standards the title or name is present. The element Snippet allows the
users to add a short description about the feature that is associated with it. This description
although it is intended to be short (around 2 lines of text), could contain important
information about a feature that could help in its understanding and use. The purpose of the
element Description seems clear, to offer a comprehensive description about a given feature.
Although its objective is the same as Snippet, the element Description offers some advantages
since it supports unlimited text extension and it can store fragments of HTML code on it.
The user can add metadata about a given feature using this element however some
considerations should be observed. First, the user should consider that the main objective of
the Description element is to offer a description about the related feature, however it is
possible to add not just the description but also other information that could be useful as
well. In this way it is possible to add a text description including all the information that
could be found in a DC file expressing metadata about a given geographic feature. Although
this could be a valid solution since all the metadata is accessible and carried within the file
and with the described feature, its unstructured format could become problematic for
automatic processing tasks.
To solve the problem of adding unstructured information and to allow the inclusion of
custom XML data within a KML file, the standard offers the ExtendedData and its child
elements. This data can involve any kind of information including metadata about a given
feature. Depending on these last elements the user has three methods to add fragments of
XML code in a KML file:
• The first method specifies how the users can add simple key-value pairs in XML
format to any feature in the KML file using the ExtendedData element and its child
element Data. The element Data contains an attribute called name intended to store
the key of the pair. Data has two child elements as well, the first one the
displayName element to specify alternative names for visualization and the element
value that stores the pair’s value. This structure does not follow any previously
defined schema allowing the user to insert data in an ad-hoc fashion however it
does not allow the creation of complex or nested XML structures. The users can use
this method to add structured metadata however restricted with the previous
limitations.
• The second method involves the specification of a schema at Document level within
a KML file. This schema is created using the KML element Schema and its child
element SimpleField. The Schema element contains an attribute storing the schema’s
name and another one doing the same with the id that in this case is mandatory
since the schema could be referenced by this value. The different elements in the
schema are specified using the element SimpleField that offers attributes to specify
the name and type of the custom element and an alternative name for visualization.
The Schema element neither allows the creation of complex structures with nested
elements. Once the schema is defined, this can be used via the element SchemaData.
This element references a given schema that can be placed in the same KML file or
in another indicating its identifier. The child element SimpleData makes reference to
any SimpleField declared in the schema and allows its filling with values of the type
declared by its referenced SimpleField element.
• Finally the last and probably most powerful method allows the import and use of
any externally defined XML Schema without any limitation about the elements or
structure defined by it. In this case it is possible to import a given schema
associating it with a namespace in the KML file. Using this namespace any element
defined in the original schema can be used within the KML file. This method also
allows the creation of nested or complex XML structures. For instance, thanks to
this method the users can import a metadata schema such as the ISO19115 into
their KML files and create a metadata structure as defined by the ISO standard for
describing the content.
Thanks to the above elements the users can add all that metadata required within a KML
file. By default KML also offers fields to add information about the authoring of the content.
This information is provided by the use of elements imported from the Atom schema
(http://www.ietf.org/rfc/rfc4287.txt) and includes the author name and link to related
website or email. This information although short it is still useful and could act as contact
information, common information in most of the metadata standards.
4.2 Reusing existing metadata

Since the appearance of metadata standards such as Dublin Core or ISO19115 a lot of
geographic content has been already catalogued and documented using these or other
metadata formats. Since KML is a really recent format and OGC standard the quantity of
content expressed in this format is relatively low compared with the quantity of information
in older formats such as OGC’s GML (http://www.opengeospatial.org/standards/gml) or
ESRI’s Shapefiles (http://www.esri.com/library/whitepapers/pdfs/shapefile.pdf). Thanks
to its spread use in a broad range of services, some of them implying the user participation
through web platforms, the quantity of content generated in KML is rapidly increasing.
KML is increasingly considered an output format for some geographic content servers such
as GeoServer. In the same way the number of tools to transform from and to KML is also
growing. This transformation allows the use of already existing geographic content in KML
format and then its reuse and merging with other evolving and fresh geographic content
build up thanks to some of the geographic services running in the Web 2.0.
Since the content can be transformed and reused so the metadata can. Independently if the
metadata it is placed internally or in a separate file, when transforming from a format such
as Shapefile to KML the metadata could be also transported within the file in its new format.
As previously seen there exists various alternatives for adding metadata in KML however
the vast majority of the already existing metadata for geographic content is expressed in
XML-based formats. This means that ExtendedData is the KML element of choice for
transporting existing metadata. By using this element it is possible to add custom XML data
to any feature what makes also possible to include the existing metadata for that feature.
The first two methods described early do not allow the creation of complex data structures
but could be used to emulate metadata schemas such as Dublin Core easily. Although this
could represent a solution it is clearly not the best option since it represents the metadata
schema but it does not really makes use of it. For this reason, the last method that combines
the ExtendedData element, the import and use of external XML schemas seems the best one.
In this case the user only needs to import the schema assigning a namespace to it that will be
used in any element along the KML file but declared in the imported schema. Now there is
no limitation about the structure, types or any other factor. With this option the inclusion of
the converted files’ metadata can be also trivially implemented.
4.3 General search engines for indexing geographic content

Nowadays if somebody checks any web traffic analytics services such as Alexa
(http://www.alexa.com) and examines the top sites on visits per day that person would be
able to see some really representative data about the greatest trend in the use of the Web.
The first 10 positions in the list would be usually distributed among search engines and sites
or services of the Web 2.0 such as video sharing services, social networks, wikis or blogs.
These results represent the huge impact and relevance of these types of services. It is also
known that some of the Web 2.0 services are based on the user participation for creating
fresh content and nowadays these are becoming more and more popular counting its users
in hundreds of millions.
The search engines however are services that appeared years ago to solve the problem of
finding resources and information in a Web in expansion. These search engines are still
extremely used and most important still useful. In fact it is maybe now when they are more
necessary than ever since the quantity of content increases in such a proportion that makes
the task of finding specific information almost impossible without the use of these tools as
far as the corresponding address is unknown. Over the years with the appearance of new
data types and formats that have become popular these tools have also been uploaded to
look for and index this type of information. For example, currently it is possible to search
information on specific data files expressed in formats such as Adobe Acrobat PDF or
Shockwave Flash format, Microsoft Word or Autodesk DWF among others.
Web search engines have extended their capabilities to search over popular file formats and
more recently also support discovery and indexing of geographic content expressed in KML
and KMZ. The interest of the major search engines in the geographic content is not limited
to search the content but also the different ways in which this content can be used. In this
way these companies provide visualization services such as web mapping or geobrowsers to
represent all the geographic content they gather, edition tools to create more content that
will increase in quantity improving its utility, APIs and mashups to create new applications
and services based on the geographic content they manage. Currently Google, Yahoo and
Microsoft are as the owners of the three major search engines on the web scenario and
provide services and tools for geographic content.
The way in which content is published for its discovery by web search engines differs
notably from the way the content is used for publication in catalogues. In short, search
engines use the crawling processes to browse automatically the web navigating between
different resources following the hyperlinks present in the web sites and collect desired
information for indexing. The first advantage of crawling technique is that it does not
require user supervision further than placing the content in a publicly accessible site (i.e.
public web server). In contrast indexed content is not immediately discoverable as it

happens with the catalogues since the crawling process, due to the quantity of information,
takes some time to process and index the information. Also, since the users have no control
over the crawling process the uncertainty about the success and performance of the process
is another price to pay for the ease of an automatic and batch processing. We can conclude
that instead of the user’s detailed publication method followed by the catalogues the search
engines offer an automatic batch collecting of geographic content. This new method could
represent an effective solution for sorting all the content constantly created that do not fit in
the traditional cataloguing process.
5. Use case: Indexing user-generated data in Google

Google is one of the most representative companies in the geospatial web advent that has
found its place also in the GI business. This company offers a long range of geographic
services, applications and data that is being continuously improved and increased. These
services and applications include the web mapping service of Google Maps, the geobrowser
or virtual globe Google Earth and the Google Maps and Google Earth APIs that allow the
programming of third-party applications on the web interacting and using the previous
services. Although these are the main services, Google also supports the development and
extension of KML and 3D models that can be added for visualization in Google Earth. The
company also offers editing tools for creating new geographic content like My Maps or even
new cartography for those countries where this is limited using tools such as Google Map
Maker. These last two tools represent a clear example of the integration between the GI and
the Web 2.0 that results in a collaborative tool for geographic and cartographic content
creation and sharing. Finally Google actually supports the use of KML, KMZ and GeoRSS to
express geographic content on their services.
5.1 Geo-Index
Google maintains an index with the processed results of crawling processes. This index
primary contains textual information extracted from HTML documents and alike what
seems evident considering the quantity of this type of resources in the Web. This index is the
one queried when the users perform searches through the main Google’s website.
Regarding the geographic content it seems that Google uses a specific index for searches
done through its geographic services and tools. We will refer to this index as Google’s Geo-
Index in the rest of the document. The Geo-Index is feed with a variety of sources that could
be divided into two main groups depending on if the content is stored in Google’s servers or
not. The first group involves all of geographic content generated through the different
services and tools for creating and editing tasks such as My Maps or Google Map Maker.
This information is directly stored into Google’s servers and therefore its insertion in the
Geo-Index is very fast if not immediate. The other group of content is spread along the Web
and includes a broad range of distributed sources.
One important source worth to mention is the main index Google maintains for the general
searches on the web. Apparently when building the index some formats are automatically
identified permitting to perform custom searches specifying the desired file format for the
results. Among other formats the users can choose also to get only KML and KMZ files as
result. Another example of locally stored data and remote resources can be found in the list
of business that Google maintain and the services associated to it also known as Google
Local Business Center. These services advertise companies and their products in yellow
pages alike way. The interesting point is that the users can search for companies or services
in a given area obtaining those business closed to the desired area through Google Maps or
Google Earth. To conform the business directory Google offers to the business owners the
possibility to register their business storing in its service the information but Google also
extracts information automatically from distributed sources such as Yellow Pages and others.
Not all the content available is about commercial products or business and Google also
obtains and displays content from other sources including for example geotagged videos
and photos published in online galleries, articles in web based encyclopaedias referencing a
given location or even KML or KMZ files created by anonymous users or entities and stored
in remote servers accessible by crawlers. This content is also known as user-generated
content and the process that follows since its creation by the author to the visualization by
other users represents the best example of discovery of geographic content in the web
performed by search engines and compliant with the principles of the Web 2.0 where users
freely create and consume the content.
5.2 Experiment
As previously explained allocating KML or KMZ in a public server would be enough to get
those files discovered, indexed and then accessible through any of the geographic services that
use Google’s Geo-Index. However the crawling process is not totally effective and in some
situations content available on the web is not found by the crawlers. Despite this inconvenient,
the crawling process still offers a lot of advantages and represents a new, easy and practical
method for discovering user-generated geographic content.
To prove the above-explained method an experiment was conducted in two rounds. The main
objective of the study was to obtain realistic measures about the performance and effectiveness
of the discovery process considering the crawling and indexing as its key elements. The
retrieval of the indexed geographic content is a fundamental operation performed by the
search engine system and called for execution by the end user using Google Maps, Google
Earth and their corresponding APIs. This process involves really sophisticate techniques and a
long list of parameters, however most of them remain unknown and most of times kept secret
(i.e. ranking algorithm for ordering results). Because of this, the retrieval process was not
carefully analyzed in the study focusing in in measures for the crawling and indexing
processes evaluation.
5.2.1 Assessment indicators

The experiment consisted in reproducing a real case scenario for user-generated content
publishing a set of different KML files with specific information and concrete characteristics in
a web server configured for allowing the Google’s crawler access. The proposed indicators to
assess the study were:
1. The elapsed time in crawling the content. Although this method of sharing the user-
generated content could offer some advantages, an excessive time in discovering the
content by the search engine system could make this an unviable solution for some
use cases.
2. The elements within a KML file that become indexed. As seen before the KML files
offers different elements where storing information about the data visualized.
Although Google recommends some good practices about the creation of content
formatted in KML, there exist not too much information about all the fields that
could store information or metadata neither the ones that Google finally uses.
Establishing where the descriptive information or metadata can be placed within a
KML file for its use by Google would improve this type of file’s design and the
success in their indexing.
3. The effectiveness of the process in number of files indexed. As it happens with the
normal HTML pages not all the content becomes indexed and accessible to the end
user due to some aspects related most of times with the access to the content or the
content itself. In the case of a low number of indexed files this could mean that the
method results ineffective for the majority of the use cases.
5.2.2. Data preparedness

Google provides to content creators some advices to improve the chance of getting their
content indexed. These advices are focused mainly on improving the visibility of the content
helping Google to find it by following the next advices:
1. Create the KML or GeoRSS content. Be sure to add attribution tags, which will
appear in the Google Search results for your content.
2. Post your files on a public web server.
3. Create the Sitemap file. Copy this file to the directory of your website.
4. Submit your Sitemap to Google.
Google also recommends providing the content with meaningful and descriptive
information in specific parts of the file:
• Give your document a meaningful name.
• Provide a relevant description for each placemark so that the user can see the
context of the search results.
• If you have a big quantity of data, divide it into topic-specific layers.
• Give each feature an “id” so that the search result can link directly to it.
These parts include the file’s name and the KML elements Name and Description however
nothing mentions any other elements that could be used. Another important point to
consider is where to place the metadata within a KML file (Placemark or Container derived
elements such as Document or Folder). Again there is not too much information and since this
represents an important advantage of using KML it should be also tested.
Following the Google’s advices the first step is the content creation in either KML or
GeoRSS. The experiment was composed by different test datasets conformed each one by a
set of different KML files. In each dataset the files differ between them in the KML elements
used to store the information, in the level inside the file at which the information is placed
and also in the use of descriptive or non-descriptive file names. The KML elements used in
the study were the same elements considered previously as suitable to contain metadata:
Name, Description, Snippet, NetworkLink and the ExtendedData (including its specific child
elements designed for storing custom XML within KML code). The idea was to place
information that could represent metadata for the content described within the file in a
given element. This would allow discovering which of these elements and at which level is
Google-friendly concerning the indexing of the information. However since one of the
Google’s recommendations implies the use of both Name and Description in a file, both
elements standing alone but also combined were used in the study.
The NetworkLink element was used to link to PHP scripts that returned as output a KML file
and discover which of the files get indexed. The purpose of using NetworkLink elements in
this way was to discover if the KML file containing this element also appeared as result
when the information in the PHP script’s output was indexed since the first is linking to it.
This chaining would be useful when creating linked structures of information as for instance
when grouping in one unique KML file different sources of information via NetworkLink
elements. It is also worth to mention the use of the ExtendedData element and child elements
in this test. Using the three different techniques the ISO19115 metadata standard has been
translated recreating a similar structure with the same elements using the elements Data,
Schema and SchemaData and it also has been directly used importing its schema. The KML
elements used that are not restricted to appear in a specific position in the KML structure
(i.e. Schema at Document level, SchemaData at Placemark level, etc.) were placed in both logical
levels considered in the study and in each of them separately. These two logical levels are
feature and document despite the vocabulary used in KML. The feature level refers to the
level where the simple or atomic elements of geographic content reside. Placemark and
NetworkLink elements can be considered as this type of atomic elements. The document level
refers to that one where a container for atomic elements is defined. In terms of KML
specification this could coincide with the Container child elements Document and Folder,
which can contain other features grouping them in a logical structure. The creation of these
logical structures could be extended and new container elements be nested in other
containers and so on. However for the purpose of the study a simple structure with two
logical levels is enough to discover the effectiveness on the indexing of files with
information placed in them. Finally the combinations of KML elements used and levels at
which these were placed gave the relation showed in Table 1:
KML element name Logic level

Name Document, Placemark, Document &
Placemark
Description Document, Placemark, Document &
Placemark
Name & Description Document, Placemark, Document &
Placemark
Snipped Document, Placemark, Document &
Placemark
NetworkLink (linking Document, Placemark, Document &
a file with elements Placemark
Name & Description)
Data Document, Placemark, Document &
Placemark
Schema Document
SchemaData Placemark
Custom XML (data in Document, Placemark, Document &
ISO19115 format) Placemark
Table 1. KML elements and logic levels used.
The second step consisted on posting the content in a public server meaning by public that
the crawlers and the end users can access the files. The following step requires the
creation and convenient placement of a Sitemap file (http://www.sitemaps.org). The
Sitemap file specification defines a XML derived language designed to give specific
information about the web server’s public resources (i.e. HTML pages, images, etc.). The
specification defines a set of elements to define lists and groups of resources and also
individual properties such as resource’s URL, type, creation date and frequency of update
among others. This kind of files help the crawlers in the collection of the public elements
within the server and also add valuable information for some tasks as for example the
update periods. With the discovery of user-created geographic content Google has created
an extension over the Sitemap specification specially designed for referencing geographic
content. The extension allows the users to indicate it by using the tag geo by which the
user can also specify the content’s file type (KML, KMZ and GeoRSS). Finally it is also
important to note that users can indicate as geographic content any file that can be
interpreted as one of these files without necessarily being one of them. One of the most
interesting aspects is that users can for instance mark as geographic content programmed
scripts (i.e. scripts in python, PHP, etc.) or applications which execution output is
effectively data in KML, KMZ or GeoRSS format allowing also their crawling and
indexing.
The last step involved the submission of the sitemap file to Google. This action gives
information to Google about the users’ sites that require to be crawled and also offers
feedback to the users’ about the process. This can be achieved thanks to the use of the
different web-based tools (through the Google Webmaster Tools service) to improve the
visibility and facilitate sites and resources discovery. Among other utilities the users can
register a website URL and then submit a sitemap file with the resources to be indexed.
5.2.3 First-round results and discussion

During the experiment all of these steps were followed in order to obtain a measure for
the best-case scenario. The results obtained in the different test sets gave an average time
of three weeks since the publication of the KML files in the server to the appearance of the
first results querying the Google Web Search website and also Google Maps and Google
Earth.
Apparently three weeks time could be an acceptable interval if rapid publication and
availability is not a requirement. Otherwise, other alternatives should be considered
including the use of traditional catalogues or the direct creation of content using Google’s
tools that offer an easy way for getting the content rapidly indexed since this is stored
directly in Google’s servers. The latter has the limitation that the content is subjected to
the edition tools offered. For instance, the currently available tools just offer a set of
features of KML and are running over a web-based platform and therefore the creation of
3D content is not possible or fairly complicated. In addition, it is also necessary to
consider those cases where the index is frequently updated or at least this is updated more
frequently than the search engine system updates its content. In this case a possible
approximation to solve the problem without representing a complete solution could be
the separation of metadata and data. This separation could be implemented adding the
metadata in the suitable KML elements but loading the data using the NetworkLink
element. Therefore the metadata keeps static and successfully indexed without modifying
the Geo-Index and data is updated on a periodical basis. In this case the NetworkLink is
used as a pipe to the data becoming the file with the metadata the starting point for
visualizing the updated content. Finally in those cases where both change on time other
solutions should be found.
Once the crawling time was obtained only the part regarding the effectiveness of the
process and the elements that intervene in the process were remaining. The results
obtained after three weeks were considerable negative considering the number of files
published for indexing. From a total number of 129 KML files including the KML files and
the PHP scripts used only 56 files were found using the main Google website and filtering
results by KML type and seven were found when using either Google Maps or Google
Earth. These elements represent respectively the 43% and the 5% of the total amount of
the test files.
These figures represent a really low number of files present in the general index and a
much lower quantity in the Geo-Index. It results evident that between the publication and
its indexing in the general index some prune is performed but most evident is that
between the two indexes there is another phase where a high quantity of files are
discarded.
It is well known that there are a lot of factors that affect the correct indexing of websites
by Google. Some of these factors are meant to avoid some malicious behaviour to improve
the visibility of a website by the use of no Google-friendly actions that could be punished
(http://www.google.com/support/webmasters/bin/answer.py?hl=en&answer=35769).
One the most common of these factors is the content duplication usually performed to
improve the number of results in which a given site appears and that is commonly detected
by Google crawlers. When this behaviour is detected only one of the sites containing that
information is kept in the index dismissing the rest and sometimes applying punishing
actions against the site or the content creator. Since a lot of information, although placed in
different KML elements, was completely the same in a great number of files the possibility
that this could affect the process was considered. In fact this behaviour affects the indexing
of content for the general index and could explain the low rate in the first case.
Considering also that not only the information but also the position (coordinates) was
shared among the files the avoidance of all this duplicated information should be
considered in further tests. The duplicated content could affect the indexing of the files in
the general index however there is another prune between the general and the Geo-Index.
The reason could be the same as in the first case however this implies that the Geo-Index
applies other methods to avoid the duplicated content much more restrictive than the
ones corresponding to the general one. Although this is possible there could be other
factors that could affect the indexing including the use of suitable KML elements for
indexing in the files.
The seven files successfully indexed shared one common characteristic; all of them
contained information in the element Name at feature level. Although some of them also
contained information in the element Description seemed that the feature’s name was the
key element. Also the results presented random presence of descriptive and none-
descriptive names what probably meant that this was not a representative factor for their
indexing. It is worth to mention that some of the indexed files corresponded to PHP
scripts confirming that the indexing process works with static and dynamically generated
KML files however none of the files with NetworkLink elements were indexed. Therefore,
it seems clear that both duplicated content and the correct chose of KML elements
containing the metadata affected the indexing process.
5.2.4. Second-round results and discussion

Given the low number of files indexed and the need to know which KML elements are
suitable to place metadata, a second test was conducted. This time the files were designed
to avoid any of the problems caused by duplicated content and then being able to analyze
the impact of the KML elements in the indexing.
The new test was composed this time by a unique test data set with 23 KML files. These
files were completely different among them in terms of coordinates, information, elements
and level where this information was placed.
This second-round test resulted in four files indexed, representing the 16% of the total.
Again the results gave a really low rate of success. One meaningful result was that in this
second attempt the results did not appear in the Google’s general index but only in the
Geo-Index. The indexing system not only serves to work with the data the crawlers collect
but also update periodically this data adding the new version and deleting all content. It
seems that in a first stage, the Geo-Index is feed or takes information from the general
index. However once this information is detected as geographic content the main index
removes that content which indexing process is updated and managed by the Geo-Index
exclusively.
Again all the files present in the search results presented the same trend: the metadata
was placed in the element Name at feature level. Once the file is indexed all the content
allocated on it gets indexed as well. For instance, in those indexed test files where both
Name and Description elements were used, the information contained in the element
description was also accessible and used in the searches demonstrating that this
information is also present in the Geo-Index. These results demonstrated that the
existence of information in that element at the given level is a requisite for the indexing of
a KML file. However this is not a guarantee for getting the user-generated content
indexed. During the test all the elements that finished in the index had this characteristic
however not all the elements with the information allocated in that position got indexed.
This means that there is still a prune of files despite the use of the right KML elements. A
comparison like the one in Figure 2 between the number of files with information in the
element Name at feature level and the files that got finally indexed during the different
tests clearly shows that the content duplication could also affect in the case of the
geographic content.
Fig. 2. Percentage of suitable files indexed per test
This assumption seems true since in the first test no more than the 33% of the files with the
same content was successfully indexed. Instead the rate of indexing it was increased to the
80% of the total number of files when avoiding the use of duplicated content.
Finally we can conclude that despite the existence of other factors the use of the right KML
elements and the content duplication has a great impact in the indexing process. The use of
element Name does not necessary mean any problem since there is no other requirement
about it but its existence within a file. However the rejection of all that content that Google
considers duplicated could become a real problem in specific situations. Consider the case of
a study in a small area about measures on a specific thematic. In those cases where the
measures share a certain quantity of content the chance of being indexed decreases. This is a
real problem since lot of content in the GI area share a high quantity of information still
being different.
6. Conclusions
The catalogues are extremely useful and powerful tools for discovery geographic content.
Their use could be however more focused on professional users since catalogues require
to some extent some degree of knowledge or experience for metadata creation and
publishing.
Currently there exist a lot of web-based tools and services designed to create, modify,
share and visualize geographic content. These tools in conjunction with the spread use
and better availability of positioning devices such as GPS receivers create an ideal
scenario for new users. Non-experts users can now create and share geographic content, a
task reserved previously to the professionals. Despite some issues still remaining like data
quality, user-generated content is being created in huge quantities and rapidly. These
factors limit the use of catalogues as effective solutions to manage the continuous
proliferation of fresh content.
In this case methods for batch discovery could mean effective solutions to collect and
order all this data. The web search engines are performing effectively this task for usual
web content such as HTML pages for years. In the recent years and with the appearance
and popularization of the KML format, web search engines are also working with the
geographic content encoded in KML. This work has analyzed how KML files can contain
and carry metadata represented either by a simple textual description or complicated
XML structures such as the one defined by the standard ISO19115 allowing the reuse of
existing content. Our analysis indicates that KML can be a metadata container because it
offers encapsulation and flexibility. The former allows the transport of metadata and data
within one single file, while the latter permits to specify the level of granularity at which
the content is described.
Google represents one of the companies that most invests in geographic services and tools
freely available and addressed for a broad public. This company through its search engine
also indexes geographic content expressed in KML or KMZ. The use of Google’s search
engines has been demonstrated to be effective for most of non-professional uses. This
approach represents an extremely useful way to publish geographic content and an
effective way for discovering it. However the use of search engines also have associated
some restrictions as it could be the time spend by the system in discovering and indexing
the content or the process’ efficiency. This measure is subject to factors such as the use of
concrete elements within a KML file and at specific levels and also the quantity of content
duplicated among files. These last two factors makes the use of this solution complicated
in those cases where there exist considerable quantities of content duplicated published
by the same data provider. Despite these minor problems, the use of web search engines
complements the use of catalogues because they can manage huge quantities of content
spread along the Web. Finally another important aspect is that these systems also perform
their tasks in an automatic manner without requiring too much interaction with the user
neither complex requirements about the data or metadata since the proliferation of
geographic content among amateur public.
7. Acknowledgements
This work was partially funded by the project “CENIT España Virtual”, funded by the CDTI
in the program “Ingenio 2010” through Centro Nacional de Información Geográfica (CNIG).
8. References
Al-Masri, E. & Mahmoud, Q.H. (2008). Discovering Web Services in search engines. IEEE
Internet Computing. 12, 3, May-Jun 2008, 74-77
Bernard, L.; Kanellopoulos, I.; Annoni, A. & Smits, P. (2005). The European Geoportal -- one
step towards the establishment of a European Spatial Data Infrastructure.
Computers, Environment and Urban Systems, 29, 1, 15-31
Craglia, M.; Kanellopoulos, I. & Smits, P. (2007). Metadata: where we are now, and where
we should be going, Proceedings of 10th AGILE International Conference on
Geographic Information Science 2007, Aalborg University, Denmark.
Foerster, T.; Schaeffer, B.; Brauner, J. & Jirka, S. (2009). Integrating OGC Web Processing
Services into Geospatial Mass-Market Applications, Proceeding of the Intl.
Conference on Advanced Geographic Information Systems & Web Services, 98-103,
Cancun, Mexico
Goodchild, M (2007). Citizens as Voluntary Sensors: Spatial Data Infrastructure in the World
of Web 2.0. International Journal of Spatial Data Infrastructures Research, 2, 24-32
Haklay, M. & Weber, P. (2008). OpenStreetMap: User-generated Street Maps. IEEE Pervasive
Computing, 7, 4, 12-18
INSPIRE EU Directive (2007). Directive 2007/2/EC of the European Parliament for
establishing an Infrastructure for Spatial Information in the European Community
(INSPIRE). Official Journal of the European Union, L 108/1, Vol 50, 25 April 2007.
ISO 19115 (2003). Geographic information - Metadata. ISO/TS 19115:2003, International
Organization for Standardization (ISO).
ISO 19139 (2007). Geographic information – Metadata – XML schema Implementation.
ISO/TS 19139:2007, International Organization for Standardization (ISO).
Masser, I. (2005). GIS Worlds: Creating Spatial Data Infrastructures, ESRI Press, ISBN: 1-
58948-122-4, Redlands, California
Mitchell, T. (2005). Web Mapping Illustrated: Using Open Source GIS Toolkits, O’Reilly
Media, ISBN 9780596008659
Nyerges, T; Barndt, M. & Brooks, K. (1997). Public participation geographic information
systems, Proceedings of Auto-Carto 13 - American congress on surveying and
mapping, pp. 224–233, Seattle, WA
Rajabifard, A.; Feeney, M-E. & Williamson, I. P. (2002). Future directions for SDI
development. Intl. Journal of Applied Earth Observation and Geoinformation, 4,
11–22
Scharl, A. & Tochtermann, K. (2007). The geospatial web: how geobrowsers, social software
and the Web 2.0 are zapping the network society. Springer Verlag, ISBN: 978-1-
84628-826-5
Turner, A. (2006). Introduction to Neogeography, O’Reilly Media, ISBN: 9780596529956
Vossen, G. & Hagemann, S. (2007). Unleashing Web 2.0: From Concepts to Creativity.
Morgan Kaufmann, Burlington, MA.
Estimation of soil properties using observations and the crop model STICS. Interest
of global sensitivity analysis and impact on the prediction of agro-environmental variables 229
10
X
Estimation of soil properties using observations

and the crop model STICS. Interest of global
sensitivity analysis and impact on the
prediction of agro-environmental variables
Hubert Varella, Martine Guérif and Samuel Buis
INRA-UAPV EMMAH
France
1. Introduction
Dynamic crop models are very useful to predict the behaviour of crops in their
environment and are widely used in a lot of agro-environmental work such as crop
monitoring, yield prediction or decision making for cultural practices (Batchelor et al.,
2002; Gabrielle et al., 2002; Houlès et al., 2004). These models usually have many
parameters and their estimation is a major problem for agro-environmental prediction
(Tremblay and Wallach, 2004; Makowski et al., 2006). For spatial application, the
knowledge of soil parameters is crucial since they are responsible for a major part of the
variability of the crop model output variables of interest (Irmak et al., 2001; Launay and
Guérif, 2003; Ferreyra et al., 2006). These parameters may be estimated from different
techniques: either with soil analysis at the different points of the study area, from a soil
map and the application of pedotransfer functions (Reynolds et al., 2000; Murphy et al.,
2003), from remote sensing images (Lagacherie et al., 2008) or by using electrical
resistivity measurements (Golovko and Pozdnyakov, 2007). The choice of the first method
is difficult because of practical limitations, as well as time and financial constraints.
Detailed soil maps adapted to the scale of precision agriculture and even to the scale of
catchment are scarcely available (King et al., 1994), while the use of remote sensing
images or electrical resistivity is still hampered by a lack of robust interpretation of the
signal (Lagacherie et al., 2008). Moreover, these techniques do not permit to access the
values of all the soil parameters required to apply a complex crop model. Fortunately,
techniques derived from remote sensing images (Weiss and Baret, 1999; Houborg and
Boegh, 2008) or yield monitoring (Blackmore and Moore, 1999; Pierce et al., 1999) allow
soil parameters being estimated through the inversion of crop models.
Estimating parameters of complex models such as crop models may be not so easy
(Tremblay and Wallach, 2004; Launay and Guérif, 2005). One of the reasons for the
difficulties encountered may be a lack of sensitivity of the observed variables to the
parameters, making the estimation process inefficient. Another reason may be that the
influence of the parameters on the observed variables takes place mainly through
interactions, making it difficult to identify the relevant factors (Saltelli et al., 2000). For
complex non-linear models such as crop models, global sensitivity analysis (GSA)
methods are able to give relevant information on the sensitivity of model outputs to the
whole range of variation of model inputs. Many studies have focused on this subject,
namely, how to choose the main parameters to be estimated for the model calibration
(Campolongo and Saltelli, 1997; Ruget et al., 2002; Gomez-Delgado and Tarantola, 2006;
Makowski et al., 2006) and ranked the importance of the parameters by calculating global
sensitivity indices: first-order indices (the main effect of the parameter on the output) and
total indices (sum of all effects involving the parameter, including the interactions with
other parameters). The common practice is consistent with the principles expressed by
Ratto et al. (2007). Small total sensitivity indices indicate a negligible effect of the
parameter on the model output concerned. These parameters can be fixed at a nominal
value (“Factor Fixing setting”). High first-order indices reveal a clearly identifiable
influence of the parameter on the model output concerned, and therefore the parameters
need to be determined accurately (“Factor Prioritization setting”). Small first-order
indices combined with large interaction indices result in a lack of identification. In
practice, the two first rules are commonly used to select the set of parameters to be
estimated in a calibration problem. GSA can also be used to evaluate the quantity of
information contained in a given set of observations for estimating parameters and thus to
determine which is the best observation set for estimating the parameters (Kontoravdi et
al., 2005). Although the results of GSA are often used to design the estimation process, the
link between GSA indices and the quality of parameter estimation has never been
quantified.
Our objectives in this study are twofold. Firstly, we propose to use GSA results in order to
measure the quantity of information contained in different sets of observations and to
illustrate the link between this measurement and the quality of parameter estimates.
Secondly, we propose to study the impact of the quality of parameter estimates on the
prediction of variables of interest for agro-environmental work. As the performance of the
estimation process is supposed to depend on several conditions such as soil type,
cropping conditions (preceding crop and climate) or available observations, we chose to
conduct the study on synthetic observations in order to be able to generate variability in
parameter retrieval performance as well as in sensitivity structure of the observed model
outputs to soil parameters and in the prediction performance. This choice also allows
eliminating the impact of model errors, which may complicate the interpretation of the
results. Finally, we considered in this study the STICS-wheat crop model and various
synthetic observations on wheat crops: derived from remote sensing images (LAI and
absorbed nitrogen) as well as grain yield.
2. Material and methods

2.1 The crop model, output variables and soil parameters
2.1.1 The STICS model

The STICS model (Brisson et al., 2002) is a nonlinear dynamic crop model simulating the
growth of various crops. For a given crop, STICS takes into account the climate, type of
soil and cropping techniques to simulate the carbon, water and nitrogen balances of the
crop-soil system on a daily time scale. In this study, a wheat crop is simulated. The crop is
essentially characterized by its above-ground biomass carbon and nitrogen, and leaf area
index. The main outputs are agronomic variables (yield, grain protein content) as well as
environmental variables (water and nitrate leaching). Yield, grain protein content and
nitrogen balance in the soil at harvest are of particular interest for decision making,
especially for monitoring nitrogen fertilization (Houlès et al., 2004). Nitrogen absorbed by
the plant and leaf area index are also important to analyze the health and growth of the
plant during the crop’s growing season.
The STICS model includes more than 200 parameters arranged in three main groups:
those related to the soil, those related to the characteristics of the plant or to the genotype,
and those describing the cropping techniques. The values of the last group of parameters
are usually known as they correspond to the farmer’s decisions. The parameters related to
the plant are generally determined either from literature, from experiments conducted on
specific processes included in the model (e.g. mineralization rate, critical nitrogen dilution
curve etc.) or from calibrations based on large experimental database, as is the case for the
STICS model (Hadria et al., 2007). The soil parameters are difficult to determine at each
point of interest and are responsible for a large part of the spatial variability of the output
variable. That is why the sensitivity analysis and parameter estimation processes
described in this study only concern soil parameters.
2.1.2 Output variables considered

In this study, we focus on two types of STICS output variables. First, those corresponding
with observations that may be done on wheat canopy by automated measurements. They
consist in:
- the leaf area index (LAIt) and the nitrogen absorbed by the plant (QNt) at various dates t
during the crop season - as potentially derived from remote sensing image inversion
(Weiss and Baret, 1999; Houborg and Boegh, 2008),
- the yield at harvest (Yld) as potentially provided by yield monitoring systems.
These output variables, hereafter referred to as “observable variables” can be observed at
different dates during the growing season.
Second, a main objective of this study, beyond the estimation of soil parameter values, lies
in the prediction of some output variables of interest, and its improvement as compared to
the prediction obtained with a lack of precise values on soil parameters. They consist in:
- yield at harvest (Yld),
- protein in the grain at harvest (Prot),
- nitrogen contain in the soil at harvest (Nit).
Yield, grain protein content and nitrogen balance in the soil at harvest are of particular
interest for decision making, especially for monitoring nitrogen fertilization (Houlès et al.,
2004). Nitrogen absorbed by the plant are also important to analyze the health and growth
of the plant during the crop’s growing season (Baret et al., 2006).
2.1.3 The soil parameters estimated

The STICS model contains about 60 soil parameters. In our case, in order to limit the
problems of identifiability, the number of soil parameters to be estimated has been
reduced. First, among the available options for simulating the soil system, the simplest
was chosen, by ignoring capillary rise and nitrification. These assumptions define the
domain of validity of the model considered and hence, of the results that are found. We
then considered the soil as a succession of two horizontal layers, each characterized by a
specific thickness parameter. From the observation of the tillage practices in the region
around our experimental site of Chambry (49.35°N, 3.37°E) (Guérif et al., 2001), the
thickness of the first layer was set at 0.30 m. We performed a first sensitivity analysis on
the 13 resulting soil parameters. This allowed us to fix those whose effects on the
observed variables were negligible: for each parameter we computed the values of its
effects on all the observed variables considered for a lot of soil, climate and agronomic
conditions, and dropped the parameters for which all these values were less than 10% of
the total effects generated by the 13 parameters. We thus restricted the study to 7
parameters.
The 7 soil parameters considered (Table 1) characterize both water and nitrogen processes.
They refer to permanent characteristics and initial conditions. Among the permanent
characteristics, clay and organic nitrogen content of the top layer are involved mainly in
organic matter decomposition processes and nitrogen cycle in the soil. Water content at
field capacity of both layers affects the water (and nitrogen) movements and storage in the
soil reservoir. Finally, the thickness of the second layer defines the volume of the
reservoir. The initial conditions correspond to the water and nitrogen content, Hinit and
NO3init, at the beginning of the simulation, in this case the sowing date.
Parameter Definition Range Unit

argi Clay content of the 1rst layer 14-37 %
Norg Organic nitrogen content of the 1rst layer 0.049-0.131 %
epc(2) Thickness of the 2nd layer 0-70 or 50-130* cm
HCC(1) Water content at field capacity (1rst layer) 14-30 g g-1
HCC(2) Water content at field capacity (2nd layer) 14-30 g g-1
Hinit Initial water content (both layers) 4-29 % of weight
NO3init Initial mineral nitrogen content (1rst layer) 4-21.5 or 25-86** kg N ha-1
* the first range is for a shallow soil and the second for a deep soil; ** the first range is for a wheat
cultivated after sugar beet and the second for a wheat cultivated after pea
Table 1. The 7 soil parameters and ranges of variation.
2.2 Global sensitivity analysis
In this study, we chose a variance-based method of global sensitivity analysis which

allows calculating the sensitivity indices for a non-linear model such as STICS. More
precisely, the method we chose is Extended FAST.
2.2.1 Variance decomposition and sensitivity indices

We denote a given output variable of the STICS model as Y. The total variance of Y,
V (Y ) , caused by variation in the 7 selected soil parameters θ , can be partitioned as
follows (Chan et al., 2000):
7
V (Y ) = ∑ Vi + ∑V ij +  +V1, 2,, 7 (1)
i =1 1≤i < j ≤ 7
where Vi = V [E (Y θ i )] measures the main effect of the parameter θ i , i = 1,  ,7 , and the

other terms measure the interaction effects. Decomposition (2) is used to derive two types
of sensitivity indices defined by:
Vi
Si = (2)
V (Y )
V (Y ) − V−i
STi = (3)
V (Y )
where V −i is the sum of all the variance terms that do not include the index i.
S i is the first-order (or main) sensitivity index for the ith parameter. It computes the
fraction of Y variance explained by the uncertainty of parameter θ i and represents the
main effect of this parameter on the output variable Y.
STi is the total sensitivity index for the ith parameter and is the sum of all effects (first and
higher order) involving the parameter θi .
S i and STi are both in the range (0, 1), low values indicating negligible effects, and
values close to 1 huge effects. STi takes into account both Si and the interactions between
the ith parameter and the 6 other parameters, interactions which can therefore be assessed
by the difference between STi and S i . The interaction terms of a set of parameters
represent the fraction of Y variance induced by the variance of these parameters but that
cannot be explained by the sum of their main effects. The two sensitivity indices S i and
STi are equal if the effect of the ith parameter on the model output is independent of the
values of the other parameters: in this case, there is no interaction between this parameter
and the others and the model is said to be additive with respect to θ i .
2.2.2 Extended FAST

Sobol’s method and Fourier Amplitude Sensitivity Test (FAST) are two of the most widely
used methods to compute S i and STi (Chan et al., 2000). We have chosen here to use the
extended FAST (EFAST) method, which has been proved, in several studies (Saltelli and
Bolado, 1998; Saltelli et al., 1999; Makowski et al., 2006), to be more efficient in terms of
number of model evaluations than Sobol’s method. The main difficulty in evaluating the
first-order and total sensitivity indices is that they require the computation of high
dimensional integrals. The EFAST algorithm performs a judicious deterministic sampling
to explore the parameter space which makes it possible to reduce these integrals to one-
dimensional ones using Fourier decompositions. The reader interested in a detailed
description of EFAST can refer to (Saltelli et al., 1999).
We have implemented the EFAST method in the Matlab® software, as well as a specific
tool for computing and easily handling numerous STICS simulations. The uncertainties
considered for the soil parameters are assumed independent and follow uniform
distributions. These uncertainties are based to the measurements made in Chambry and
correspond to the ranges of variation presented in Tab. 1. A preliminary study of the
convergence of the sensitivity indices allowed us to set the number of simulations per
parameter to 2000, leading to a total number of model runs of 7×2 000=14 000 to compute
the main and total effects for all output variables and parameters considered here. One
run of the STICS model taking about 1s with a Pentium 4, 2.9 GHz processor, the overall
simulation process takes about 4h.
2.2.3 Criteria based on GSA indices

GSA provides main and total indices per parameter for each output variable considered.
In order to summarize this information, we propose to create different criteria.
(i) The first one is a global measure of the information contained in a set of observations to
estimate each parameter:
The Global Mean Sensitivity (GMSi) computes the mean of the main effect of parameter
θ i minus its interactions with the other parameters for all observed variables, each
component being weighted by the degree of dependence of the corresponding output
variable with the other variables:
1
∑ (1 − α )(S
K
GMS i =
K
k i
k
− Rik ) (4)
k =1
where k is a given observed output variable in a subset composed of K variables among

{LAIt, QNt, t=1,…,T and Yld}, Rik = STi k − S ik is the sum of all interaction terms including
parameter θi for the observed variable k, 0 ≤ α k ≤ 1 is the mean of the absolute values of
the correlation coefficients rkk ' between the variable k and the other variables k’
1
(calculated on the model simulations required for GSA): αk = ∑ rkk ' , K>1.
K − 1 k '≠k
The GMSi criterion is based on the following rules:
- if STi is low (and thus S i ), observation k is assumed not to contain enough information
to estimate parameter θ i : in this case the corresponding part of the criteria should be low,
- ifS i is high (and thus STi ), observation k is assumed to contain sufficient information
to estimate parameter θ i : in this case the corresponding part of the criteria should be
high,
- if S i is low and STi is high, then the model is over-parameterized and difficulties in
identifying parameter θi are expected (Ratto et al., 2007): in this case the corresponding
part of the criteria should be low,
- high correlation between output variables indicates that the information contents of
these variables are redundant: in this case the weights of the corresponding sensitivity
indices should be reduced.
GMSi varies within the range [-1, 1]. It tends to 1 when S i is close to 1 for all observed
variables and when all the observed variables are perfectly uncorrelated: in this case the
model has an additive structure for the parameter θ i and this parameter has a clearly
identifiable influence on the K observed variables. GMSi tends to -1 when S i and Rik are
close to 0 and 1 respectively for all observed variables and when all the observed variables
are perfectly uncorrelated: in this case problems of identification of the parameter θ i are
expected.
(ii) The second criterion is calculated at the whole parameter set level.
The Total Global Mean Sensitivity (TGMS), is the sum of the GMSi for all parameters:
1
∑ (1 − α )(S
K
)
7 7
TGMS = ∑ GMS i = ∑ k i
k
− Rik (5)
i =1 i =1 K k =1
The TGMS criterion varies within the range [-7, 1]. It tends to 1 when Rik is close to 0 for
all parameters and all observed variables and when all the observed variables are
perfectly uncorrelated: in this case the model is additive. TGMS tends to -7 when Rik is
close to 1 for all parameters and all observed variables and when all the observed
variables are perfectly uncorrelated: in this case the model is expected to be
unidentifiable.
2.3 Parameter estimation

We chose a bayesian method which allow to take into account existing information on the
parameters to be estimated (this improves the quality of the estimation process) and to
compute an estimate of the posterior probability distribution of parameter values
(Makowski et al., 2002; Gaucherel et al., 2008). More precisely, the method we chose is
Importance Sampling.
2.3.1 The Importance Sampling method

The posterior parameter distribution is given by Bayes’ theorem:
π (Y / θ )π (θ )
π (θ / Y ) = (6)
π (Z )
where Y is the vector of total observations of size K, π (θ / Y ) is the posterior parameter

distribution, π (θ ) is the prior parameter distribution, π (Y ) is a constant of
proportionality determined by the requirement that the integral of π (θ / Y ) over the
parameter space equals 1, and π (Y / θ ) is the likelihood function. The likelihood is the
probability of the data Y given the parameters θ . Its value is determined from the
probability distribution of the errors of modelled and observed data. It is readily seen that
both the prior distribution and the new data affect the posterior parameter distribution.
The principle of the Importance Sampling method (Beven and Binley, 1992; Beven and
Freer, 2001) is to approximate the posterior parameter distribution π (θ / Y ) given in (7)
N
by a discrete probability distribution ( θ n , pn ), n=1,...,N, ∑p
n =1
n = 1 , where p n is the
probability associated with the parameter vector θ n . In our case, the method proceeds as
follows:
(1) Randomly generate N vectors θn , n=1,...,N, from the prior parameter distribution
π (θ ) ,
(2) Calculate the likelihood values π (Y / θ n ) for n=1,...,N, associated with the different
generated parameter vectors,
(3) Calculate p = π (Y / θ n )
n N
∑ π (Y / θ )
m =1
m
The pairs ( θ n , pn ), n=1,...,N, can be used to determine various characteristics of the

posterior distribution, including the mean of the posterior joint distribution of θ,
N
θ post
= ∑ p nθ n .
n =1
In this study, we assume that the errors of simulated and observed data are independent
between dates and variables and follow normal distributions of zero mean and standard
deviation σ k . Thus, we use the following likelihood function:
K
1  1 
π (Y / θ ) = ∏ exp− [ y k − f k (θ , x )]2  (7)
 2σ k
2
k =1 2πσ 2
k 
The parameters are assumed to be independent in our case. The prior distribution π (θ ) is
thus the product of the different marginal prior distributions. Accordingly to the
distributions based on the experimental fields of Chambry we assumed them to be
uniform and correspond to the uncertainties given in Tab. 1.
We have implemented the Importance Sampling method in the Matlab® software. A
preliminary study of the convergence of the estimates allowed us to set the total number
of generated parameter vectors N at 100 000.
2.3.2 Criterion expressing the quality of parameter estimation

(i) For each parameter, we created a criterion noted REi (for Relative Error of the parameter
i), to quantify the quality of the parameter estimation. It computes the ratio between the
error of the estimated parameter θ i post and the error of the prior information θ i prior :
RMSE (θ i post )
REi = (8)
RMSE (θ i prior )
1 P true
where RMSE (θ i post ) =
P p =1
(
∑ θ i, p − θ i,post
p )2
, θ i,true
p
is the true value of soil parameter
θi for a given vector p and θ i,post

p
is the corresponding estimation given by the Bayesian
method. REi quantify how much the estimation given by the Bayesian method improves
(REi<1) or not (REi≥1) the prior knowledge about the parameter value.
(ii) For all parameters, the criterion called Total Relative Error (TRE), is defined by the
mean of the 7 values of REi:
1 7 1 7 RMSE (θ i post )
TRE = ∑ i 7∑
7 i =1
RE =
i =1 RMSE (θ i
prior
)
(9)
2.3.3 Criterion expressing the quality of prediction

We created a criterion to quantify the quality of the prediction of the 3 agro-
environmental variables defined above. This criterion computes the ratio between the
error of prediction obtained from the mean of the posterior distributions of the
parameters, θ post
and the one obtained from the mean of the prior distributions, θ prior
.
It is called Relative Error of Prediction and is defined as follows:
RMSEPj (θ post
)
REPj = (10)
RMSEPj (θ prior
)
Qj
1
∑∑ ( f
P
where RMSEPj (θ post
)=
P × Qj
j
q 2
)
(θ ptrue ) − f jq (θ ppost ) , θ p
true
is the true
p =1 q =1
values of soil parameters θ for a given vector p, θ ppost is the corresponding estimation
given by the Importance Sampling method, and f (θ

j
q true
p ) is assumed to be one of the Qj
observations of the predicted variable j, for the pth vector of true values of soil parameters.
2.4 Numerical experiments
2.4.1 Generation of observations for parameter estimation

The STICS model output variables depend on the soil, climate and agronomic conditions
for which the wheat crop is simulated. In view of this, we use different configurations in
our study, as presented in Table 2: 4 contrasting climates, 2 different soil depths (shallow
and deep), and 2 preceding crops (sugar beet and peas). The climatic data used were
obtained from the meteorological station of Roupy (49.48°N, 3.11°E). Four different sets of
data were chosen to characterize a dry climate (1975-1976), a wet climate (1990-1991), a
medium-dry climate (1979-1980) and a medium-wet climate (1972-1973). The distributions
of soil parameters used in our study (GSA, creation of observations and prior information
for GLUE) are independent and uniform and deduced from the experimental data
acquired in Chambry (see Tab. 1). In this application, we assume that the type of soil
depth (shallow or deep) and the preceding crop (sugar beet or pea) are known. As a
consequence, two different ranges were considered for the depth of soil epc(2) and for the
mineral nitrogen content at the beginning of the wheat crop simulation NO3init.
Climatic Soil Preceding Configuration

conditions depth crop label
Dry Shallow Sugar beet dry-beet
Medium-dry Shallow Sugar beet mdry-beet
Medium-wet Shallow Sugar beet mwet-beet
Wet Shallow Sugar beet wet-beet
Dry Deep Sugar beet dry+beet
Medium-dry Deep Sugar beet mdry+beet
Medium-wet Deep Sugar beet mwet+beet
Wet Deep Sugar beet wet+beet
Dry Shallow Pea dry-pea
Medium-dry Shallow Pea mdry-pea
Medium-wet Shallow Pea mwet-pea
Wet Shallow Pea wet-pea
Dry Deep Pea dry+pea
Medium-dry Deep Pea mdry+pea
Medium-wet Deep Pea mwet+pea
Wet Deep Pea wet+pea
Table 2. Description of the 16 configurations based on soil, climatic and agronomic
conditions.
We consider observations on wheat crops obtained for the different configurations

described before. These observations consist of LAIt and QNt available at 10 dates t,
distributed through the wheat growing season: November 15, December 12, January 15,
February 16, March 15, April 05, April 19, May 03, May 17 and June 07; and Yld. Three
possible sets of observations (see Table 3) were considered for the parameter estimation
experiments and the computation of the criteria based on the GSA and GLUE results. In
order to compute observations, 50 vectors of true values θ true were randomly generated
from the distributions defined above. The number P is thus equal to 50. Corresponding
values of STICS-wheat model output variables were simulated for each configuration
leading to 50x16 simulations. Observations y q ,t were then computed by adding a random
error term to the simulated values of the variables and dates defined above:
( )
y q ,t = f q ,t θ true , x + ε q ,t (11)
where f q ,t is the STICS model output q (Yld, LAIt or QNt) calculated on date t (harvest for
Yld or t=1,…,T for LAI and QN), x is the vector of explanatory variables and ε q,t is the
observation error term. Following the assumptions made in Section 2.3.1 to compute the
likelihood function of the GLUE method, the vector of observation error is given by:
ε q ,t ~ N (0, σ q2,t ) where σ q ,t = σ q0 f q ,t (θ true , x ) , σ Yld
0
=9%, σ LAI
0
=17% and σ QN
0
=30%
according to measurements realized in agricultural plots (Machet et al., 2007; Moulin et
al., 2007).
Set number Variables used Size K

1 LAIt on dates t=1,...,10 K=10
2 LAIt and QNt on dates t=1,...,10 K=20
3 LAIt and QNt on dates t=1,...,10, and Yld K=21
Table 3. Description of the 3 sets of observations.
2.4.2 Generation of observations for prediction

The prediction of variables of interest is performed on independent wheat crop seasons
from those used in the estimation process. From each of the 50 vectors defined above, 120
configurations of prediction were studied and are composed by the corresponding type of
soil depth, 10 climatic data, 3 different sowing dates and 4 different cropping techniques (2
amount of fertiliser and 2 types of preceding crops). The number Qj is thus equal to 120 The
climatic data were obtained from the meteorological station of Roupy (49.48°N, 3.11°E) and
are different from those used in the parameter estimation process.
From each vector θ true , type of soil depth and configuration, the values of synthetic
observations of the predicted variables of interest (Yld, Prot and Nit at harvest) are simulated
with STICS-wheat. The values of the permanent properties (argi, Norg, epc(2), HCC(1) and
HCC(2)) of θ true are the same as those used to create the synthetic observations in the
estimation step and the initial conditions (Hinit and NO3init) are randomly generated from
the distributions defined in Tab. 1: we assume that the values of the initial conditions are not
known for the predicted season. For each parameter estimation experiment defined above,
each estimated values of the permanent properties are used to predict the output variables
of wheat crop through the STICS model. Assuming that the initial conditions are unknown
for the predicted season, they are fixed at the mean of their distribution.
3. Results
3.1 Relationship between criteria based on GSA results and the quality of estimates
3.1.1 At a single parameter level

For each soil parameter, we present here the results about the relationship between the
criterion based on GSA results and the criterion related to the quality of the parameter
estimate, for the three sets of observations and the 16 soil, climatic and agronomic
conditions.
Figure 1 shows that a good link exists between GMSi and REi: the relationship seems to be
linear, the higher the GMSi criterion, the lower the REi and the better the quality of
estimation of the ith parameter. The results show clusters of parameters: h (Hinit) at high
GMSi and low REi values, e (epc(2)) at intermediate GMSi and REi values, and the other
parameters all grouped in the same cluster at low GMSi and high REi values. Within the
scattering around the relationship, the position of the parameter depends on the
configuration and especially the soil depth.
Fig. 1. Scatter diagrams of the criteria REi and GMSi of the 7 soil parameters for the 3 sets
of observations and the 16 configurations. a) Label a corresponds to argi, N to Norg, e to
epc(2), H1 to HCC(1), H2 to HCC(2), h to Hinit and n to NO3init. b) Symbol ○ corresponds
to shallow soils and □ corresponds to deep soils.
The case of parameters for which the observations contain enough information to estimate
them precisely can be illustrated by the parameter Hinit. For example, for the dry-beet
configuration and the observation set #1, Hinit has a big main effect (equal to 0.582 on
average), leading to a high value of GMSi=0.155. In this case, the parameter Hinit has a
low value of REi (REi=0.499) meaning a good improvement of its uncertainty through the
parameter estimation process (see Figure 1a). In general, the results show that for high
values of GMSi, the reduction of the estimation error is large: a high GMSi indicates good
parameter estimation.
Small first-order indices combined with a large interaction are thought to induce
problems of identification. In our case, this applies to most of the parameters. For
example, HCC(1) has a small main effect (0.089 on average) and a large interaction (equal
to 0.251 on average) for the wet-beet configuration and the first observation set, leading to
a negative value of GMSi=-0.099. In this case, the parameter HCC(1) has a high value of
REi (REi=0.921), meaning a poor improvement in parameter estimation (see Figure 1a). In
general, the results show that for negative values of GMSi, the reduction of the estimation
error is small: a negative value of GMSi reveals a bad quality of the parameter estimation.
The values of both criteria for parameter epc(2) vary a lot between the different
configurations and especially between the types of soil depth. As it is shown is Figure 1b,
only configurations with a shallow soil (with the symbol○) allow retrieving the parameter
epc(2). For shallow soil the observed variables are quite sensitive to epc(2), leading to a
quite good estimate, while for deep soil the observed variables are no longer sensitive to
epc(2), leading to a poor estimate. For example, the configuration dry-beet and the set #3
leads to intermediate GMSi and REi values (GMSi=0.073 and REi=0.724), while dry+beet
and set #3 leads to lower (resp. higher) GMSi (resp. REi) values (GMSi=-0.008 and
REi=0.871).
In order to quantify the quality of the relationship illustrated in Figure 1, we propose to
compute the Spearman rank correlations coefficient (Spearman, 1904) between GMSi and
REi, for each configuration and observation set. This coefficient allows comparing the
relationship between two ranking lists. The analysis is made after discarding the
parameters having a negative GMSi which have always a poor quality of estimation and
whose rank would still be high. The results displayed in Table 4 show that the averaged
Spearman's correlation between GMSi and REi is satisfactory (about 75.4 %). The GMSi
criterion is thus considered to be effective for ranking the accessible parameters (for
which the criterion is positive) with respect to their quality of estimates.
Parameters* Climates Observation sets

(REi, GMSi) 75 % (TRE, TGMS) 72 % 91 %
* calculated for parameters having a positive value of GMSi.
Table 4. Averaged Spearman’s rank correlation coefficient of criteria (REi, GMSi) and
(TRE, TGMS). The first pair of criteria is involved in the parameter ranking and the
second pair is involved in the climates and observation sets ranking.
3.1.2 At the whole parameter set level

In Figure 2 the values of TRE and TGMS have been averaged for each observation set and
climate over the two soil depths and the two preceding crops. The relationship between
TRE and TGMS appears satisfactory. The TRE criterion never reaches low values (the
minimum value is about 0.8) even for high TGMS values (about 0.21), due to the relatively
large number of parameters which are not easily retrievable. The effect of climate is
striking. Configurations with a dry climate have the higher values of TGMS (between 0.16
and 0.2) and they correspond to the best quality of estimation of the parameter set (TRE
between 0.81 and 0.86), unlike configurations with a wet climate (TGMS below 0.03 and
TRE above 0.89).
The greater the number of observations considered in the estimation process (from set #1
to set #3), the lower is the TRE. As expected and seen in Figure 2, TGMS often decreases
when the number of observations increases. Although some of the observed variables are
mutually correlated (the average correlation coefficient between set #1 and set #2 is about
61 % while it is about 37 % between set #2 and set # 3), they each improve the quality of
the parameter set estimation.
Finally, the Spearman's correlation coefficients between TGMS and TRE were computed
for each type of soil depth, preceding crop and observation set, in order to quantify the
ranking of the 4 climates given by both TGMS and TRE. The averaged Spearman's
correlation presented in Table 4 between TGMS and TRE is satisfactory (about 72 %).
Secondly, the Spearman correlations between TGMS and TRE were computed for each soil
depth, preceding crop and climate, in order to quantify the ranking of the three
observation sets given by both TGMS and TRE. The averaged Spearman's correlation
between TGMS and TRE is very satisfactory (about 91 %).
Fig. 2. Scatter diagrams of the criteria TRE and TGMS at the whole parameter set level, for
the 3 sets of observations and the 4 types of climate. Each of the 12 points is an average of
the 2 soil depth and the 2 preceding crop configurations. The symbol ○ corresponds to the
set #1, ◊ to set #2 and ■ to set #3.
This study shows that the quality of parameter estimation can be explained by the results
of GSA. Suitable empirical criteria have been proposed to summarize the results of GSA
which allow ranking (i) the parameters with respect to their quality of estimate and (ii) the
configurations (particularly the climate and the observation set) with respect to the
quality of estimation of the whole parameter set. These criteria are thus shown in our case
to be useful tools for estimating the potential of given configurations of observations for
retrieving soil parameter values. They may be used also for optimizing the type of
observations to be acquired and the dates of acquisition.
3.2 Impact of the quality of estimates on the quality of prediction

The quality of the prediction in now analysed. Figure 3 and 4 shows the results in term of
REPj for the prediction of the variables of interest concerning wheat crop, by using the
estimated values of the permanent soil parameters (argi, Norg, epc(2), HCC(1) and HCC(2))
in place of their prior values. The initial conditions (Hinit and NO3init) are assumed to be
unknown for the prediction and are fixed at the mean of their distributions. The values of
REPj are calculated for the 2 types of soil depth and the 3 sets of observations.
Fig. 3. Results of the prediction of the variables of interest. The results are presented for
the 2 types of soil depth and the 3 sets of observations and averaged on the 4 climates.
In Figure 3, REPj values are averaged on the 4 climates. It can be seen that 2 of the 3
variables of interest (Yld and Prot) have a significant lower REPj and therefore a greatly
improved quality of prediction when using the estimated values of the permanent
parameters, as compared to when using prior information on the parameters. In that case,
Yld and Prot seem to be quite sensitive to the permanent soil parameters. The output
variable Nit is not or slightly affected by the estimation of the soil parameters because it is
sensitive to the initial conditions, which are fixed at a nominal value for the prediction,
and not to the permanent parameters. Through the estimation of the permanent soil
properties, the size of the observation set slightly improves the quality of prediction: the
bigger the observation set size the better the permanent parameter estimates and the
better the prediction. The most important improvement between two sets of observations
concerns the output variable Yld in a shallow soil: REPj is about 0.66 for set #2 and about
0.52 for set #3. In that case, a lot of information is provided by the observation of Yld.
The type of soil depth affects a lot the quality of the prediction and especially for the
output variables Yld and Prot, which have a lower REPj when the type of soil is shallow. It
is not surprising, accordingly to the results of parameter estimation, because the
parameter epc(2) has a better quality of estimates in the case of shallow soil and because
Yld and Prot are also quite sensitive to this parameter, as well as the observed variables
(see Section 3.1.1). The output variable Nit is not affected by the soil depth because of its
lack of sensitivity to epc(2).
Fig. 4. Results of the prediction of the variables of interest. Effect of the climates on
prediction of the variables of interest. The results are averaged on the 3 observation sets
for a shallow soil.
The effect of climate on the quality of prediction of the variables of interest is illustrated in
Figure 4 on a shallow soil and averaged over the 3 sets of observations. The dryer the
climate the better the quality of the prediction. As the observed variables are more
sensitive to the soil parameters in dry climatic conditions (see Section 3.1.2), that allows a
better quality of the parameter estimates in that case. The results of the prediction are thus
affected by the quality of the estimates and are better in dry climatic conditions, because
of their sensitivity on the parameters. The REPj value of the output variables Yld and Prot
significantly decreases from 0.71 to 0.5 and 0.65 to 0.47. As previously, the output variable
Nit is not accurately predicted and slightly affected by the climate: REPj decreases from
0.92 to 0.86.
It was shown in this study that some soil parameters can be retrieved by considering
observations on crops and that the estimated values of the permanent soil parameters
allow reducing the uncertainty on the prediction, because of the sensitivity of the
predicted variables on these parameters. The quality of the prediction is manly affected by
two factors: the soil depth and climate. These results are closely linked to the quality of
the parameter estimates.
4. Conclusion
In our results the link between the quality of parameter estimation and GSA results was
illustrated through three types of behavior: high first-order indices are associated with
good quality of estimation, low total indices are associated with bad quality of estimation,
and high total indices combined with low first-order indices are associated with poor
estimates because of interactions between parameters. Many other studies show that the
parameter estimation performance can be explained by the results of GSA (Tremblay and
Wallach, 2004; Gaucherel et al., 2008; Manache and Melching, 2008). Given the large
number of output variables and dates considered in this application, the GSA indices had
to be summarized to study the link between GSA and parameter estimation results. We
proposed the GMSi criterion and show its relation with the criterion REi which measures
the quality of estimation of parameter i. The criterion GMSi proves to be effective for
ranking the accessible parameters with respect to their quality of estimation. For a given
configuration, GSA is able to provide information on which parameters can be estimated
and which can be fixed as they do not deserve an accurate determination (Ratto et al.,
2007). We show in this work that the parameters having GMSi close to zero are not
accessible from the observations and the STICS model.
The total criterion TGMS can be used to predict the ranking of the configurations with
respect to their ability to retrieve the whole set of parameters, and in particular the
ranking of the climates and the observation sets: it is possible to predict which type of
climate and observation set will lead to the better estimation of the whole parameter set.
These results are particularly interesting for screening the possibility of estimating
parameters from a given set of available observations in a given agro-environmental
context, and, following Kontoravdi et al. (2005), promote GSA as an excellent precursor to
optimal experimental design.
From observations on crop status, it is possible to retrieve the soil parameters and the
estimated values allow improving the quality of the prediction of agro-environmental
variables. Among them, some variables are strongly affected by the quality of the
parameter estimates, such as grain yield and protein of the grain, because of their large
sensitivity on the permanent soil parameters. This result is particularly interesting for
agro-environmental work because the criteria based on GSA also allow screening the
possibility of a given set of available observations to predict, through soil parameter
estimates, the variables of interest for crop management.
Finally, it would be helpful to conduct such a study on real data to assess the impact of
model errors on both soil parameter retrieval and link between the proposed criteria.
5. Acknowledgements
The financial support provided by CNES and Arvalis-Institut du Végétal is gratefully
acknowledged.
6. References
Baret, F., Houles, V.,Guerif, M., 2006. Quantification of plant stress using remote sensing
observations and crop models: the case of nitrogen management. Symposium on
Imaging Techniques for Understanding Plant Reponses to Stress held at the
Society-for-Experimentaal-Biology Meeting. Canerbury, ENGLAND.
Batchelor, W.D., Basso, B.,Paz, J.O., 2002. Examples of strategies to analyze spatial and
temporal yield variability using crop models. European Journal of Agronomy 18
141-158.
Beven, K.,Binley, A., 1992. The Future of Distributed Models - Model Calibration and
Uncertainty Prediction. Hydrological Processes 6 279-298.
Beven, K.,Freer, J., 2001. Equifinality, data assimilation, and uncertainty estimation in
mechanistic modelling of complex environmental systems using the GLUE
methodology. Journal of Hydrology 249 11-29.
Blackmore, S.,Moore, M., 1999. Remedial correction of yield map data. Precision
Agriculture 1 53-56.
Brisson, N., Ruget, F., Gate, P., Lorgeou, J., Nicoullaud, B., Tayot, X., Plenet, D., Jeuffroy,
M.H., Bouthier, A., Ripoche, D., Mary, B.,Juste, E., 2002. STICS: a generic model
for simulating crops and their water and nitrogen balances. II. Model validation
for wheat and maize. Agronomie 22 69-92.
Campolongo, F.,Saltelli, A., 1997. Sensitivity analysis of an environmental model an
application of different analysis methods. Reliability Engineering & System
Safety 57 49-69.
Chan, K., Tarantola, S., Saltelli, A.,Sobol, I.M., 2000. Variance-based methods. In: Saltelli,
A., Chan, K.,Scott, E.M., (Eds.), Sensitivity analysis. Wiley: New York.
Ferreyra, R.A., Jones, J.W.,Graham, W.D., 2006. Parameterizing Spatial Crop Models with
Inverse Modeling: Sources of Error and Unexpected Results. Transactions of the
ASABE 49 1547-1561.
Gabrielle, B., Roche, R., Angas, P., Cantero-Martinez, C., Cosentino, L., Mantineo, M.,
Langensiepen, M., Henault, C., Laville, P., Nicoullaud, B.,Gosse, G., 2002. A
priori parameterisation of the CERES soil-crop models and tests against several
European data sets. Agronomie 22 119-132.
Gaucherel, C., Campillo, F., Misson, L., Guiot, J.,Boreux, J.-J., 2008. Parameterization of a
process-based tree-growth model: Comparison of optimization, MCMC and
Particle Filering algorithms. Environmental Modelling & Software 23 1280-1288.
Golovko, L.,Pozdnyakov, A.I., 2007. Electrical geophysical methods in agriculture.
Progress of Information Technology in Agriculture 457-471.
Gomez-Delgado, M.,Tarantola, S., 2006. GLOBAL sensitivity analysis, GIS and multi-
criteria evaluation for a sustainable planning of a hazardous waste disposal site
in Spain. International Journal of Geographical Information Science 20 449-466.
Guérif, M., Beaudoin, N., Durr, C., Machet, J.M., Mary, B., Michot, D., Moulin, D.,
Nicoullaud, B.,Richard, G., 2001. Designing a field experiment for assessing soil
and crop spatial variability and defining site specific management strategies.
Proceedings 3rd European Conference on Precision Agriculture. Montpellier,
France.
Hadria, R., Khabba, S., Lahrouni, A., Duchemin, B., Chehbouni, A., Carriou, J.,Ouzine, L.,
2007. Calibration and validation of the STICS crop model for managing wheat
irrigation in the semi-arid Marrakech/Al Haouzi plain. Arabian Journal for
Science and Engineering 32 87-101.
Houborg, R.,Boegh, E., 2008. Mapping leaf chlorophyll and leaf area index using inverse
and forward canopy reflectance modeling and SPOT reflectance data. Remote
Sensing of Environment 112 186-202.
Houlès, V., Mary, B., Guérif, M., Makowski, D.,Juste, E., 2004. Evaluation of the crop
model STICS to recommend nitrogen fertilization rates according to agro-
environmental criteria. Agronomie 24 1-9.
Irmak, A., Jones, J.W., Batchelor, W.D.,Paz, J.O., 2001. Estimating Spatially Variable Soil
Properties for Application of Crop Models in Precision Farming. Transactions of
the ASAE 44 1343-1353.
King, D., Daroussin, J.,Tavernier, R., 1994. Development of a Soil Geographic Database
from the Soil Map of the European Communities. Catena 21 37-56.
Kontoravdi, C., Asprey, S.P., Pistikopoulos, E.N.,Mantalaris, A., 2005. Application of

global sensitivity analysis to determine goals for design of experiments: An
example study on antibody-producing cell cultures. Biotechnology Progress 21
1128-1135.
Lagacherie, P., Baret, F., Feret, J.B., Netto, J.M.,Robbez-Masson, J.M., 2008. Estimation of
soil clay and calcium carbonate using laboratory, field and airborne
hyperspectral measurements. Remote Sensing of Environment 112 825-835.
Launay, M.,Guérif, M., 2003. Ability for a model to predict crop production variability at
the regional scale: an evaluation for sugar beet. Agronomie 23 135-146.
Launay, M.,Guérif, M., 2005. Assimilating remote sensing data into a crop model to
improve predictive performance for spatial applications. Agriculture Ecosystems
& Environment 111 321-339.
Machet, J.M., Couturier, A.,Beaudoin, N., 2007. Cartographie du rendement du blé et des
caractéristiques qualitatives des grains. In: Guérif, M.,King, D., (Eds.),
Agriculture de précision. Quae: Versailles.
Makowski, D., Naud, C., Jeuffroy, M.H., Barbottin, A.,Monod, H., 2006. Global sensitivity
analysis for calculating the contribution of genetic parameters to the variance of
crop model prediction. Reliability Engineering & System Safety 91 1142-1147.
Makowski, D., Wallach, D.,Tremblay, M., 2002. Using a Bayesian approach to parameter
estimation; comparison of the GLUE and MCMC methods. Agronomie 22 191-
203.
Manache, G.,Melching, C.S., 2008. Identification of reliable regression- and correlation-
based sensitivity measures for importance ranking of water-quality model
parameters. Environmental Modelling & Software 23 549-562.
Moulin, S., Zurita, R.M.,Guérif, M., 2007. Estimation de variables biophysiques du couvert
par ajustement de modèles de transfert radiatif sur des réflectances. In: Guérif,
M.,King, D., (Eds.), Agriculture de précision. Quae: Versailles.
Murphy, B., Geeves, G., Miller, M., Summerell, G., Southwell, P.,Rankin, M., 2003. The
application of pedotransfer functions with existing soil maps to predict soil
hydraulic properties for catchment-scale hydrologic and salinity modelling.
International Congress on Modelling and Simulation. Townsville, AUSTRALIA.
Pierce, F.J., Nowak, P.,Roberts, P.C., 1999. Aspects of Precision Agriculture. Advances in
Agronomy 67 1-85.
Ratto, M., Young, P.C., Romanowicz, R., Pappenberger, F., Saltelli, A.,Pagano, A., 2007.
Uncertainty, sensitivity analysis and the role of data based mechanistic modeling
in hydrology. Hydrology and Earth System Sciences 11 1249-1266.
Reynolds, C.A., Jackson, T.J.,Rawls, W.J., 2000. Estimating soil water-holding capacities by
linking the Food and Agriculture Organization soil map of the world with global
pedon databases and continuous pedotransfer functions. Water Resources
Research 36 3653-3662.
Ruget, F., Brisson, N., Delecolle, R.,Faivre, R., 2002. Sensitivity analysis of a crop
simulation model, STICS, in order to choose the main parameters to be estimated.
Agronomie 22 133-158.
Saltelli, A.,Bolado, R., 1998. An alternative way to compute Fourier amplitude sensitivity
test (FAST). Computational Statistics & Data Analysis 26 445-460.
Saltelli, A., Tarantola, S.,Campolongo, F., 2000. Sensitivity analysis as an ingredient of

modeling. Statistical Science 15 377-395.
Saltelli, A., Tarantola, S.,Chan, K.P.S., 1999. A quantitative model-independent method for
global sensitivity analysis of model output. Technometrics 41 39-56.
Spearman, C., 1904. The proof and measurement of association between two things. Amer.
J. Psychol. 15 72-101.
Tremblay, M.,Wallach, D., 2004. Comparison of parameter estimation methods for crop
models. Agronomie 24 351-365.
Weiss, M.,Baret, F., 1999. Evaluation of canopy biophysical variable retrieval
performances from the accumulation of large swath satellite data. Remote
Sensing of Environment 70 293-306.
Modeling of wake-vortex detection by a ground-based fiber LIDAR system 249
0
11
Modeling of wake-vortex detection by a

ground-based fiber LIDAR system
S. Brousmiche, L. Bricteux,
G. Winckelmans, B. Macq and P. Sobieski
Ecole Polytechnique de Louvain, Université catholique de Louvain
Belgium
1. Introduction
Due to its lift force, an aircraft releases large scale swirling flows in its wake. As these vortices
can impact significantly the trajectory of a following aircraft, their study is of great importance
for practical applications concerning safety in aircraft traffic management. A well-adapted
system to detect them is the heterodyne Doppler LIDAR (Light Detection And Ranging).
Moreover, fiber-based LIDAR are known for their higher flexibiliy and compacity despite that
they are limited in pulse energy. However, to reach optimal performances, it is often needed
to precisely model the whole measurement process including atmospheric effects such as the
refractive turbulence. Some parameters have to be tuned depending on the application as,
for example, the pulse energy and duration, the pulse repetition frequency or the telescope
dimensions. Wake vortices detection and intensity estimation also require a good compro-
mise between spatial resolution and velocity precision which also depends on the scanning
configuration.
The aim of this chapter is twofold. On the one hand, it describes a complete simulation method
to evaluate the performance of a Doppler LIDAR for aircraft wake vortices detection near the
ground. The principal interest of the simulation presented here is to combine a LASER beam
propagation method with fluid dynamics simulations. On the other hand, it proposes effective
wind velocity map reconstruction algorithms. The performance of spectral and correlogram
accumulation algorithms in the vicinity of wake vortices is analyzed. The effect of velocity
gradients on the estimation error is thus addressed.
After introducing the wake vortex phenomenon and its detection using a heterodyne Doppler
LIDAR , the section 2 will discuss the measurement simulation. The numerical simulation of
the wake in ground effect is then detailled in section 3. The processing algorithms are finally
explained in section 4.
1.1 Wake-vortex fundamentals

The lift force acting on an aircraft is due to the pressure difference between the lower side
(pressure side) and the upper side (suction side) of its wing. As a result of this pressure
difference, there is a spanwise flow at the edges of the wing from the pressure side to the
suction side. This pressure difference forces the suction side streamlines to converge toward
the center of the wing and the pressure side streamlines to diverge from it. This spanwise
flow combined with the free-stream velocity produces a swirling motion of the air trailing
downstream of the wing. Just behind the trailing edge, a vortex sheet is shed which rolls-up
rapidly within a few span lengths to form a pair of counter rotating vortices of equal strength.
The circulation, Γ [m2 /s], of the velocity field [m/s], u, for each vortex is a quantity of primary
importance as it gives an image of the vortex strength. For a closed contour 𝒞 = ∂Ω enclosing
a patch of vorticity defined by the field ω = ∇ × u,
∫ ∮
Γ= ω ⋅ n dS = u ⋅ dl . (1)
Ω 𝒞
In the particular case of a vortex alone, it is natural to choose a circular integration contour 𝒞
and since the velocity field has the form u = u θ (r ) êθ it follows that:
Γ (r ) = 2π r u θ (r ) (2)
The velocity profile is thus closely related to the circulation profile. The circulation and the
velocity tangential distribution are given by the following relations:
r2
Γ (r ) = Γ0 ( ), (3)
r2 + r2c
Γ (r )
u θ (r ) = . (4)
2πr
1.5
0.5
u z /Vθmax 0
−0.5
−1
−1.5
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
y/b0
Fig. 1. Vertical velocity profile in the far wake of an aircraft. The dashed-dotted lines represent
Γ0 max = Γ0 in the case of
the potential velocity profile Vθ = 2π r . The curve is normalized by Vθ 4πr c
a low order algebraic model.
The parameters are Γ0 [m−2 /s], the total circulation of the vortex and rc [m], the radius of
maximum induced tangential velocity. It is an indicator of the core size of a vortex; typically it
amounts, after the initial roll-up phase, to 3 − 5% of the wing span b. A typical velocity profile
for a pair of vortices is illustrated in Fig. 1. A downwash velocity will alter the trajectory of an
airplane passing between the two vortices. If it passes on one of the sides of the vortex pair, it
will feel an upwash velocity. If the airplane enters in a vortex zone, a strong rolling moment
Leader aircraft Follower aircraft Separation Time delay [s]

max take-off weight (metric tons) Nautical miles Approach speed: 70m/s
Heavy Heavy (≥ 136 T) 4 106
Heavy Medium(< 136 T) 5 132
Heavy Light (≤ 7 T) 6 159
Medium Light 5 132
Table 1. International Civil Aviation Organization (ICAO) aircraft separation distances to
avoid wake vortex encounter, in nautical miles (1 NM = 1.852 km). For all other combina-
tions, the separation is 3 NM. Table reproduced from Gerz et al. (2001).
will be applied to it. The momentum conservation allows to bind the initial circulation Γ0 of
an aircraft vortex pair to the lift force itself, equal to the weight of the aircraft:
L = M g = ρ U b0 Γ 0 , (5)
where U [m/s] is the flight velocity and ρ is the fluid density. The circulation is thus propor-
tional to the mass M of the aircraft and inversely proportional to its flight speed U. Typical
values for the circulation range in Γ0 ≈ 400 − 600 m2 /s for heavy aircrafts. The length b0 is
the distance between the vortex cores; its value is approximately b0 = π4 b (assuming elliptical
loading) with b the wingspan. The vortex spacing b0 can be deduced following the lifting line
theory (Prandtl, 1957). Wake vortices generated at take-off are very strong since the airplane
evolves at low speed and is fully loaded. Aircraft wake vortices are also the source of the drag
induced by the lift force acting on the aircraft: the so-called induced drag.
Fig. 2. Classical heterodyne Doppler LIDAR architecture based on a fiber LASER source.
1.2 Detection using a heterodyne Doppler LIDAR

A general architecture for a heterodyne Doppler LIDAR is given at Fig. 2. The heart of the
system is a continuous LASER source producing a quasi-monochromatic optical wave. The
1.55 µm fiber-based LASER sources are used as an alternative to classical sources due to their
higher compacity and flexibility. Moreover, although the pulse energy is rather low (i.e. a few
hundreds of microjoules), they are able to produce a higher pulse repetition frequency. When
passing through the acousto-optic modulator, the wave is shifted in frequency and modulated
to produce a LASER pulse with a global Gaussian shape. The intermediate frequency is typ-
ically chosen around 70 MHz. The pulse duration has a duration varying from 100 ns to 1 µs
depending on the spatial resolution needed. This pulse is then amplified and emitted through
the atmosphere using a telescope. The telescope configuration is usually monostatic which
means that the energy is sent and received through the same aperture. This requires however
the use of a circulator. The monostatic configuration has less sensitivity to higher spatial scales
of the atmospheric turbulence. As it propagates, the emitted beam is scattered by the aerosols
particles. A portion of the backscattered energy is received at the telescope and is mixed with
the local oscillator beam coming from the LASER source. A balanced photodetector produces
the heterodyne current which oscillates around the intermediate frequency depending on the
Doppler effect induced by the movement of the aerosol particles along the line-of-sight. The
signal is then processed to estimate wind velocity maps, detect the wake vortices and estimate
their circulation.
Most of the LIDAR systems used for wake-vortex detection are ground-based with a 2D scan-
ning perpendicular to the runways as illustrated in Fig. 3. This allows to better discriminate
the velocity signal due to the vortices from that due to the background turbulence, which is
weaker. The angular resolution of the system is defined by the angular scanning velocity and
the pulse repetition frequency.
Fig. 3. Lidar scanning pattern on a runway for aircraft wake vortices released close to the
ground.
Axial scanning configuration can also be potentially used, especially for onboard LIDAR de-
tection. Nevertheless, the system performance needed are more difficult to attain since the
measured radial velocity inside the vortices has a much lower dynamic.
2. Simulation of a Coherent Doppler LIDAR

In heterodyne detection systems, the power of the heterodyne signal at the output of the pho-
todetector is determined by the correlation between the backscattered and the LO fields. This
means that any atmospheric phenomena causing the propagating beam to loose its spatial co-
herence will alter the system efficiency. The main processus involved at low altitude are due
to the atmospheric turbulence. The Speckle effect also induces a loss in the signal temporal
coherence but it has a limited impact because of its lower time constant. It can therefore be
compensated by increasing the level of accumulation during the velocity estimation. The low
energy combined with the sensitivity to turbulence increases the need of a precise LIDAR
simulation in order to achieve optimal performances. On the one hand, it allows to determine
the optimum LASER source and telescope parameters for a given application. Depending on
the constraints of the technology used, it gives the global limits of the system. In the case of
fiber LASER sources, the most important parameters are the pulse energy, the pulse duration,
the repetition frequency as well as the telescope parameters. On the other hand, it provides
simulated signals for the design and validation of velocity estimators and velocity map re-
construction algorithms. The system efficiency in terms of velocity estimation variance and
spatial resolution are some of the results obtained from this analysis.
The study of LIDAR performance has largely been investigated using either analytical devel-
opments (Frehlich & Kavaya, 1991; Frehlich, 1993) or numerical simulation techniques (Bel-
monte, 2000; 2003; Frehlich, 2000). The first is somewhat limited to weak fluctutation due to
refractive turbulence or the propagation of a untruncated Gaussian beam whereas numerical
simulation methods, such as the phase-screen technique has a larger panel of application. The
equations describing the LIDAR signal in the temporal domain have first been introduced in
Salamitou et al. (1995).
The main originality of the proposed simulation method is the integration of both optical
and fluid dynamics numerical methods to take into account the signal coherence loss due to
the refractive turbulence, the speckle effect as well as the fine structures of the wake vortex
velocity field. It includes also the generation of realistic LIDAR signals in order to develop ef-
ficient algorithms for velocity profile estimation and wind velocity map reconstruction. More
precisely, the simulation combines the classical signal generation method in the time domain
(Salamitou et al., 1995; Van Trees, 2001) with the phase-screen technique used to consider the
double-path problem and to compute the heterodyne efficiency and the SNR profiles for a
given configuration. To generate the Doppler shifts and induce the signal coherence loss due
to wake vortices, a fluid flow simulation is interfaced with the main module. It consists in a
large eddy simulation (LES) of incompressible turbulent flow using a high order parallel finite
difference code.
In this section, we will describe the time-domain simulation method as well as the phase-
screen techniques.
2.1 The heterodyne LIDAR equations

Let assume that the LIDAR system transmits a Gaussian sounding pulse through the atmo-
sphere along a given elevation angle θ. Thus,
√ {√ }
s T ( t ) = 2ℜ Et f˜(t)e j2π f c t (6)
1
where Et [J] is the pulse energy, f c [Hz] is the intermediate frequency and f˜(t) [s− 2 ] is the
complex envelope of the transmitted pulse defined by
1 ( )
˜
f (t) = exp − t2 /2σ2 (7)
π 1/4 σ1/2
where σ is the 1/e intensity radius. f˜(t) is normalized such that
∫ ∞
˜ 2
f (t) dt = 1 (8)
−∞
Suppose that the atmosphere is divided into layers of thickness dR along the line-of-sight.
We transmit a signal sT and examine the signal at time t backscattered from the range interval
[R,R+dR]. The signal is the superposition of many reflexions by the aerosol particles contained
in this layer. As the number of aerosol particles illuminated is large, the measured signal is
well described by a zero-mean Gaussian random process b̃(t, R) and a variance equal to
{ } AR
E ∣ b̃(t, R)∣2 = K 2 ( R) β( R) 2 ηS ( R, t) (9)
R
where E {} represents the statistical average, K ( R) is the one-way irradiance extinction taking
into account the molecular absorption at the wavelength λ [m], β( R) [m−1 sr−1 ] is the aerosol
backscatter coefficient, AR [m2 ] is the receiver area and ηS is the system-antenna efficiency de-
scribing the portion of the backscattered field effectively converted to heterodyne signal. The
system-antenna efficiency ηS ( R) term can be computed by using first order analytical equa-
tions (Frehlich, 1993) although this approach is limited to weak turbulences and untruncated
beams. The contribution of an atmospheric slice at range R to the heterodyne signal at time t
is given by
√
s̃R (t, R) = S Et PLO f˜(t − 2R/c)b̃ (t, R)exp(−4πvR ( R)t/λ)dR (10)
where PLO [W] is the local oscillator power, S [A/W] is the detector quantum efficiency and
vR ( R) [m/s] is the mean radial velocity of the aerosol particles. The return from the entire
atmosphere is the superposition of the returns from the incremental intervals. The complex
envelope is
√ ∫ +∞
s̃R (t) = S Et PLO f˜(t − 2R/c)b̃ (t − R/c, R)exp(−4πvR ( R)t/λ)dR (11)
0
The measured heterodyne current signal [A] at the output of the photodetector is given by
√ { }
i (t) = 2ℜ s̃R (t)e jωc t + iN (t) (12)
where iN (t) is noise signal mostly due to the photodetector shot-noise current. It can thus be
modeled by an additive Gaussian noise that has a bandpass spectrum and a variance
{ }
E ∣iN ∣2 = 2SeBw PLO (13)
where e [C/electron] is the electronic charge and Bw [Hz] is the detector bandwidth. The
signal-to-noise ratio is then
{ }
E ∣ s̃R (t)∣2
SNR(t) =
E {∣iN ∣2 }
∫
η Q Et ∞ 2 AR
= K ( R) β( R) 2 ∣ f (t − 2R/c)∣2 ηS ( R)dR (14)
h̄νBw −∞ R
The discrete form of the heterodyne signal complexe envelope is given by
√ Lc ( )
2
s̃R (kTs ) = S Et PLO ∆R ∑ f˜ kTs − l∆R b̃(l∆R)exp (−4πjvR (l∆R)kTs /λ) (15)
l =− L
c
c
with Ts [s] is the sampling interval and L c is the summation length chosen to include most of
the illuminated aerosol targets. In (Frehlich, 1997), it is proposed to use L c = 3.72σ(2∆R/c). In
order to completely simulate the LIDAR signal, we need to compute the Doppler shift term
as well as the variance of b̃. The first will be obtained by sampling a simulated wake vortex in
ground effect velocity field presented in the section 3. The later will require a complete beam
propagation simulation described in the next paragraph.
2.2 Simulation of LASER beam propagation

The phase-screens technique is used here to compute realizations of the system-antenna ef-
ficiency ηS intervening in the signal variance. The analysis of its mean value ⟨ηS ( R)⟩ and
standard deviation ⟨ηS2 ( R)⟩ are used to optimize the LASER as well as the telescope param-
eters. This numerical method is particularly suited to simulate the propagation of LASER
beams through random media as well as the crossing through the telescope optics.
The scales of turbulence have different effects on the beam depending on their size. Large
scales, in the order of the outer scale L0 of turbulence, cause beam wandering and tilt whereas
small ones, with a size close to the inner scale of turbulence l0 affect the beam coherence.
In monostatic configuration, only the latter have a real impact on the system efficiency. A
measure of the strength of the turbulence is given by the refractive index structure constant
Cn2 . Typical values range from 10−18 m−2/3 for weak turbulence to 10−12 m−2/3 for strong
turbulence. It is a constant for horizontal paths and decreases with altitude.
The sounding path is divided into Nr layers of thickness ∆R′ ≫ ∆R. A random phase-screen is
simulated at each layer. Hence, the LASER beam undergoes phase distortion passing through
the screen and is then diffracted within each layer. The propagation of the laser beam ET ( R, r)
through layer [R,R + ∆R′ ] can be written (Banakh et al., 2000) :
{ ( ) }
E ( R + ∆R′ , s) = ℱ −1 exp − jπλ∆z′ κ2 ℱ [ E ( R, r) exp ( jψ (r))] (16)
where r and s are the transverse spatial coordinates, Ψ(r) is the random phase screen, κ the
spectral coordinate, and ℱ the Fourier transform operator. The phase-screens are generated
using a Von Karman spectrum with a outer scale L0 of 1 m and a inner scale l0 of 1cm. Its
equation is given by
exp(−κ2 /κ2m )
Φ n (κ ) = 0.033Cn2 ,0 ≤ κ < ∞ (17)
(κ2 + κ02 )11/6
where Cn2 [m−2/3 ] is the refractive-index structure constant, κm = 5.92/l0 and κ0 = 1/L0 . The
generation of a phase-screen Ψ(u ) from the refractive-index spectrum Φ n (κ ) is given by:
{ [√ ]}
Ψ(r) = 2πk2 ℜ ℱ Φ n (κ ) ξ (κ ) (18)
where k = 2π/λ [m−1 ] is the optival wavenumber and ξ (κ ) is a zero-mean, unit variance,
complex Gaussian random variable. For convenience, the heterodyne mixing is usually com-
puted at the target plane using the backpropagated local oscillator formulation (BPLO). For a
general configuration, the emitted as well as the BPLO field must be simulated. The system-
antenna efficiency at layer [R,R + ∆R′ ] can therefore be calculated as follows
∫
λ2 R 2 ∞
ηS ( R ) = < jT ( R, w) jBPLO ( R, w) > dw. (19)
A R −∞
where jT ( R, w) and jBPLO ( R, w) [m−2 ] are respectively the irradiance of the normalized trans-
mitted and BPLO fields. The emitted beam is collimated and perfectly Gaussian before pass-
ing through the telescope. The LIDAR parameters used for the following simulation are given
in Table 2. Due to computational considerations, the PRF has been fixed to 400 Hz which is
lower than the PRF provided by a fiber LIDAR (a few kHz).
Simulation parameters
Laser wavelength 1.55 µm
Pulse energy 150 µJ
Detector bandwidth 50 MHz
Pulse repetition frequency 400 Hz
Intermediate frequency 80 MHz
Sampling frequency 250 MHz
Angular resolution 1.25 mrad
1/e radius of the emitted beam 30mm
Radius of the transmitter aperture 50mm
Aerosol backscattering coefficient 10−7 m−1 sr−1
Linear extinction coefficient 7.10−5 m−1
Table 2. LIDAR parameters used for the simulations
We use a numerical grid for the screen of 512 × 512 with 2 mm resolution. We simulate a
continuous random medium by using 10 phase-screens and the statistics have been obtained
by running 400 samples. Fig. 4. gives the SNR as well as the system efficiency at low and
medium refractive turbulence level for both monostatic and bistatic configurations. We see
on the one hand that an increase of the turbulence level degrades significantly the system
performances with range in the bistatic configuration. On the other hand, performance en-
hancement is observed in the monostatic case which is due to the fact that the emitted and
backpropagated beams propagate through the same turbulence. This effect is limited in range
and the SNR eventually becomes lower than the one obtained at low Cn2 .
−4
0.1
−5
0.08
−6
⟨SNR(R)⟩
ηS ( R)
0.06
−7
−8 0.04
−9 0.02
−10 0
0 200 400 600 800 1000 200 400 600 800 1000
Range R [m] Range R [m]

Fig. 4. Signal-to-noise ratio (left) and system efficiency (right) with range obtained for dif-
ferent measurement conditions : free-space (solid); monostatic configuration, Cn2 = 10−15
m−2/3 (dash); bistatic configuration, Cn2 = 10−13 m−2/3 (dot) and monostatic configuration,
Cn2 = 10−13 m−2/3 (dash-dot).
3. Simulation of wake vortices in ground effect

In this section, we describe the fluid simulation that have been performed to analyze the detec-
tion capability of the heterodyne LIDAR in the presence of wake vortices. Since we consider
here a ground-based system, simulating the interaction of wake vortices with the ground is of
great importance since it produces secondary vortices.
3.1 Description of the code

The governing equations of the flow considered are the Navier-Stokes equations for incom-
pressible flow, and supplemented by a term which models the effect of the small dissipatives
scales of turbulence (not captured by the grid) on the big scales:
∇⋅u = 0 (20)
∂u
+ (u ⋅ ∇) u = 2 τM
−∇ P + ν∇ u + ∇ ⋅ τ , (21)
∂t
where P is the reduced pressure, ν is the kinematic viscosity, and τ M is the subgrid scale
stress tensor model. These equations are solved using a fractional-step method with the delta
form for the pressure (Lee et al., 2001). This form allows simple boundary conditions for
the pressure and the intermediate velocity field. The convective term is integrated using an
Adams-Bashforth 2 scheme and the diffusion term using a Crank-Nicolson scheme. The time-
stepping scheme reads
u∗ − un 1( n )
= − 3H − Hn−1 − ∇φn
∆t 2
1
+ ν ∇2 ( u ∗ + u n )
2(
1 ( )n ( ) n −1 )
M M
+ 3 ∇⋅ τ − ∇⋅ τ (22)
2
1
∇2 ϕ = ∇ ⋅ u∗ (23)
∆t
u n +1 − u∗
= −∇ ϕ (24)
∆t
φ n +1 = φn + ϕ , (25)
where Hn is the convective term, u ∗ is the intermediate velocity field and φn is the modified
pressure. The model term is integrated explicitly, also using the Adams-Bashforth 2 scheme.
The subgrid scale model used here is the WALE regularized variational multiscale model as
defined in (Bricteux, 2008). The equations are discretized in space using fourth order finite
differences and the scheme of Vasilyev (Vasilyev, 2000) for the discretization of the convective
term, which conserves energy on Cartesian stretched meshes and is therefore particularly
suited for direct or large-eddy simulations of turbulent flows. The Poisson equation for the
pressure is solved using a multigrid solver with a line Gauss-Seidel smoother. The code is
implemented to run efficiently in parallel.
3.2 Radial velocity interpolation technique

In order to compute the Doppler term appearing in the equation giving the heterodyne current
(Equation 15), the radial velocity in each layer l and for each LOS is mandatory. As the fluid
flow velocity field (v(y, z), w(y, z)) issued from the Navier-Stokes simulation is computed on
cartesian grid, it is necessary to interpolate the data onto the LIDAR line-of-sight and perform
a projection of the velocity vector on the LOS in order to have the radial velocity.
The interpolation scheme selected in this work to interpolate information from a fluid simula-
tion grid grid to the LIDAR measurement points is the M4′ scheme (Winckelmans, 2004). This
technique is widely used in vortex particles methods to redistribute vorticity from a distorded
set of particles onto a regular cartesian grid. In this case we perform the inverse operation
as we transfer information from a cartesian rectangular grid onto a polar one. This scheme is
defined as:
M −1 N −1
φ I (⃗x ) = ∑ ∑ W (⃗x − ⃗xijS )φijS , (26)
i =0 j =0
where ⃗x is the lidar sensing location while ⃗x S is the location at which the information is com-
puted. The kernel for the M4′ scheme is defined as:
⎧ � �2 � �3 � �
⎨ 1 − 25 � xh � + 32 � xh �
 if � hx � ≤ 1
� � � � �
1 2 − �x� 2 1 − �x�
� � � � �
W (x) =
 2 h h if 1� <� � xh � ≤ 2 (27)
⎩ � x�
0 if h > 2
In Fig. 5, one observes the number of points on the cartesian grid which influence the target
point on a LOS. The radial velocity profiles are then obtained by computing the projection
of the interpolated velocity profile along the different line-of-sight constituting the scanning
pattern.
Fig. 5. Interpolation Stencil: the value at a target point (square) located on a given LOS at
a given range is obtained by interpolation using the 16 neighbouring points (circles) of the
cartesian fluid dynamics simulation grid with a spacing h.
In order to evaluate the estimation error using the fluid simulation velocity field and compare
the results with different pulse durations, we have to take into account the radial velocity
profile filtering during the measurement process. Hence, the estimated velocity is a filtered
version of the exact radial velocity. The weighting function is the square norm of the transmit-
ted pulse defined in Equation 7. A good approximation of the mean radial velocity estimate
is given by (Frehlich, 1997):
∫ ∞ 2

v pulse ( R) = vR (ρ) ˜
f (( R − ρ)2/c) dρ (28)
0
3.3 Results
The results are given with a dimensionless time τ defined as τ = tV0 /b0 , where V0 =
Γ0 /(2πb0 ) is the intial descent velocity of the vortex pair. The flow is evolving in the following
way: due to the mutually induced velocity, the vortex pair moves down. The impermeabil-
ity condition at the wall can be modeled by a mirror vortex pair, resulting in an hyperbolic
trajectory. However, as the induced tangential velocity at the wall is nonzero, the no-slip con-
dition is at the origin of the development of boundary layers beneath the vortices. Those are
clearly visible in Fig. 6. Starting from the stagnation point on the symmetry plane between the
vortices, the pressure gradient forcing the boundary layers is first favorable and then adverse.
This leads to their separation, when the primary vortices are close enough to the wall. The
detached vortex-sheet then rolls-up and forms two coherent secondary vortices. They induce
an upward velocity on the primary vortices and force the latter to rise; this is the rebound
phenomenon. At this time, the flow is still essentially two-dimensional. But the secondary
vortices are unstable with respect to short-wavelength modes. These perturbations are grow-
ing while the secondary vortex is orbiting around the primary vortex, as can be seen in Figs. 6
at time τ = 2. The instability mechanism can be related to the elliptic instability of Widnall.
The flow eventually evolves to a fully turbulent vortex system.
The Γ5−15 is an important hazard criterion and is defined as the circulation profile average
between 5 and 15 meters. Its evolution is reported in Fig. 7. The initial 2-D decay phase is
governed by the molecular viscosity and is due to the slow spreading of the core size, it is
however negligible in the case of a high Re number flow. The fast decay phase starts at τ ≈ 2.8
and the final circulation level attained is reduced by 40% compared to the initial one. During
the first part of the descent, the trajectoriy follows the inviscid theory. As soon as the boundary
layer starts to separate, the rebound occurs and the trajectory is significantly different. One
observes the loop trajectory induced by the rebound.
The radial velocity maps in the early case (τ=2.5) and the late case (τ=5) are given Fig. 8.
4. Velocity map reconstruction

Assuming that the LIDAR parameters (pulse energy, duration and PRF) have been optimized
for the detection problem, the only constraint subsisting on the signal processing algorithms is
to guarantee the best spatial resolution while keeping the radial velocity estimation variance
as low as possible. This constraint is related to the use of accumulation methods to compensate
the low SNR. As circulation estimation method are usually based on a model-fitting method,
a poor spatial resolution inside vortices will induce inherently a higher estimation variance
on the vortex circulation. The goal of this section is therefore to compare the efficiency of
periodogram-based and correlogram-based accumulation methods in terms of spatial resolu-
tion in the vicinity of the wake vortex.
4.1 Radial velocity estimation

Each time a pulse is sent, along a particular line-of-sight (LOS) of given elevation angle de-
fined by the current position of the telescope, a signal s(nTs ) is measured where Ts is the
sampling period. In order to evaluate the velocity profile along this LOS, we estimate the
energy spectral density of s(nTs ) at each position of a sliding window of N time samples. The
observation time is then Tobs = NTs which corresponds to a range gate size ∆p = Tobs c/2. The
time offset of the pth window is t p and the spatial offset is then R p = t p c/2. A direct spectral
estimate is the periodogram defined by
2
1 N −1

P̂s (k∆ f ) = ∑ h(nTs )s(t p + nTs )exp (−2πikn/N ) (29)
N n =0
If no spectral interpolation is performed, the frequency resolution is ∆ f = Fs /N. An estimate

of mean velocity, v̂R , is given by evaluating
( )
λ Fs { }
v̂R = f c − argmaxk P̂s [ k] (30)
2 N
where f c [Hz] is the intermediate frequency and h is the signal weighting function (usually
Gaussian). A parametric spectral estimation method is also possible usually leading to better
velocity estimates at moderate noise level. The signal s(nTs ) at a given position of the window
can be modeled by a p-order autoregressive model, noted AR(p). The difference equation
describing this model is
p
x (n ) = − ∑ ak x (n − k) + w(n ) (31)
k =1
where { ak } are the parameters of the model and w(n ) is the input sequence of the model, i.e.
a white noise with zero mean and variance σ2 . The corresponding power spectrum estimate
can then be expressed in term of the AR coefficients
σ2 σ2
P̂sAR (k∆ f ) ≡ = 2 (32)
∣ A(k)∣2 p
1 + ∑k=1 ak exp(− j2πk f )
The { ak } parameters are calculated in this work by the Yule-Walker method and the Levinson-
Durbin algorithm for which an estimation of the signal correlogram is needed. The spectrum
of an AR(p) process has at most p poles, i.e. frequencies minimizing A(k). Those models are
good candidates to represent sharp spectral peaks but not spectral valleys since they have no
zeros. However, if p is chosen large enough, AR techniques can approximate adequately any
kind of spectral shapes. The model parameters estimation technique is described in (Proakis,
1965). The velocity estimate is then computed using Equation 30. We use a fixed model order
of p=16.
τ = 0.0
τ = 2.0
τ = 5.0
⃗ × u∥b2 /Γ0 = 1 and 10.

Fig. 6. Visualization of the flow field using isosurfaces of ∥∇ 0
0.8
0.6
Γ5−15
Γ0
0.4
0.2
0
0 1 2 3 4 5 6
τ
Fig. 7. Temporal evolution of the Γ5−15 .
z[ m]
z[ m]
y[ m]
Fig. 8. Radial velocity maps vR [m/s] of the wake vortices in ground effect at dimensionless
time τ=2.5 (top, early case) and τ=5 (bottom, late case).
4.2 Accumulation strategies

Due to the low pulse energy provided by fiber-based LIDAR , the velocity estimates obtained
from a single spectral density estimate are rather poor. In order to reduce the estimation vari-
ance, two different kinds of accumulation strategies are used: a periodogram accumulation
and a correlogram accumulation (Rye & Hardesty, 1993). We apply these concepts here in the
case of 2D velocity map reconstruction in order to compare their performances when wake
vortices are present in the sensing volume.
Each signal window is defined by a vector k with polar coordinates R = 2c (t p + Nw Ts /2)
[m] and θ [rad] being the elevation angle defining the line-of-sight. The processing vector
set {k } is denoted 𝒦. We define Ik as either the periodogram or the correlogram computed
from the signal window sθ (t p + nTs) with n = 0, . . . , Nw − 1. The main idea is to compute
an averaged estimate I p at a set of spatial positions p using the estimations Ik included in

a circular region-of-interest 𝒞 around this point p. The accumulation algorithm is given in
Table 3. Once the radial velocity has been estimated at each point p, the estimation error is
computed by the mean square error between the estimated radial velocity v̂R and the exact
filtered radial velocity v pulse (p ) :
MSE(p ) = ∣ vˆR (p ) − v pulse (p )∣2 (33)
4.3 Results
The accumulation level, defined as the number of estimates Ik used for one radial velocity
estimation v̂R , is given at Fig. 9. In order to have the same spatial resolution at each point p a
constant radius ρ with distance R have been chosen. As the measurement is performed on a
polar grid, the accumulation level decreases with R. The vortex pair is situated around 400 m
from the LIDAR. At this distance, a radius ρ of 1, 2 and 4 m corresponds to an accumulation
level per velocity estimate of respectively 20, 60 and 120 k-points. For a given radius, the
accumulation level depends also on the pulse repetition frequency.
Using a fixed radius implies the variance of the spectral density estimates to increase with
range R due to variable accumulation level and SNR fluctuation. However, it is more interest-
ing to choose a range-dependent radius by determining an optimal trade-off between spatial
resolution (accumulation level) and velocity resolution (spectral density estimate variance).
Fig. 10 gives the spatial variation of the MSE compared with the exact radial velocity field.
We observe that the estimation error increases inside the vortex where the velocity gradients
along the propagation direction are the most important. High velocity gradients decrease the
temporal coherence of the signal and consequently induce a broadening of the signal spec-
trum with a decrease of its maximum value. Moreover, the accumulation process combined
with the argmax estimation tends to promote narrow spectra computed outside the vortex.
This effect must be take into account since it distorts the retrieved vortex shape and increases
the error made on subsequent model-fitting algorithms dedicated to circulation estimation.
Accumulation algorithm :
∀k ∈ 𝒦,
Estimates Ik from sθ (t p + nTs), n = 0, . . . , Nw − 1
∀ p ∈ ℛ,
Define 𝒞 with radius ρ = ρ( R).
n ← 0;
∀k ∈ 𝒦 ∣ dist(p, k) ≤ ρ( R)
I p ← (nI p + Ik ) 1+1 n ;
n ← n + 1;
Estimate spectral density from I p
Estimate v̂ (p )
Table 3. Accumulation algorithm. Ik is either the periodogram or the correlogram of the signal.
250
200
Accumulation level
150
100
50
200 300 400 500 600 700 800 900 1000
Range [m]
Fig. 9. Accumulation level for different radius ρ of the circular region-of-interest: ρ = 1 m
(dot), ρ = 2 m (dash-dot) and ρ = 4 m (solid).
100
4
2
z [m]
50 0
−2
−4
0
100
2.5
2
z [m]
50
1.5
0 1
200 250 300 350 400 450 500 550 600
y [m]
Fig. 10. Influence of the accumulation process on the estimation error. Radial velocity map vR
(top) and standard deviation σv [m/s] (bottom).
The processing results obtained for wake vortices in the late case with a pulse duration of
400 ns and a refractive index Cn2 = 10−15 m−2/3 is given Fig. 11. The window length has been
fixed to N = 64 samples which gives a range gate of 38.4 m. Comparing this results with Fig. 8
(bottom) allows to illustrate the radial velocity filtering inherent to wind measurement with a
pulsed LASER .
Fig. 12. shows the standard deviation of the radial velocity estimates averaged over all the
line-of sight ⟨σv ⟩ for different values of the pulse energy Et . It has been obtained in the early
case (young vortices) with a pulse duration of 200 ns and strong turbulence Cn2 = 10−13 m−2/3 .
The algorithm used is the correlogram-accumulation method. We see that reducing the pulse
z[ m]
z[ m]
y[ m]
Fig. 11. Signal processing results: estimation of the radial velocity map (top) and MSE
(bottom) obtained in the late case. The pulse duration is 400 ns and the refractive index is
Cn2 = 10−15 m−2/3 .
0.4
0.35
0.3
0.25
0.2
⟨σv ⟩ [m/s]
0.15
0.1
0.05
0
200 300 400 500 600 700 800 900 1000
Range [m]
Fig. 12. Influence of the pulse energy on σv . Et = 50µJ (dash); Et = 150µJ (dot-dot); Et = 300µJ
(dash-dot) and Et = 600µJ (solid) obtained with the correlogram accumulation method (ρ = 2
m). The pulse duration is 200 ns. The refractive index is Cn2 = 10−13 m−2/3 .
0.4
0.35
0.3
0.25
0.2
0.15
⟨σv ⟩ [m/s]
0.1
0.05
−0.05
−0.1
200 250 300 350 400 450 500 550 600
Range [m]
Fig. 13. Comparison of the correlogram (dash-dot) and periodogram (solid) accumulation
methods at 200 ns for young vortices. The accumulation radius is 2 m.
1.5
1
⟨σv ⟩ [m/s]
0.5
0
0 50 100
Range [m]
Fig. 14. Comparison of the correlogram (CA) and periodogram (PA) accumulation methods:
PA, 200ns (solid); CA, 200ns (dash-dot); PA, 400ns (dash); CA, 400ns (dot). The refractive
index is Cn2 = 10−15 m−2/3 . The range is given such that R = 0 m at the vortex center.
energy has no influence inside vortices whereas it monotically increases with range. It must
be noted that, as pointed by Equation 14, for the same telescope configuration and beam sizes,
the SNR is identical provided that we keep the ratio βEt /Bw constant. Hence, decreasing the
pulse energy by a factor of 2 for example, gives the same estimation variance than decreasing
by the same factor the aerosol particles density. The comparison between the two accumu-
lation strategies are given at Fig. 13 and Fig. 14. The standard deviation of the velocity
estimate are given after having removed the averaged evolution due to SNR variation. The
correlogram accumulation gives better results inside the vortices at 200 ns and 400 ns. The
error profiles are both narrower (spatial resolution) and smaller (velocity resolution) which
induces better results for subsequent circulation estimation algorithm based on model-fitting
as described in (Brousmiche et al., 2009).
5. Conclusions
The main originality of the simulation framework proposed here is to link numerical simula-
tion techniques belonging to optical propation and fluid dynamics theory in order to assess the
feasibility of wake vortex detection using a fiber-based heterodyne Doppler LIDAR. More-
over, the generation of realistic LIDAR signals with temporal decorrelation due to Speckle
and beam propagation through wake vortices allows to develop efficient velocity estimation
algorithms. Two accumulation-based algorithms have been tested and their performances
in the vicinity of the vortices have been compared. It results that the correlogram accumu-
lation method gives the best compromise between velocity and spatial resolution than the
periodogram accumulation method generally used in such application. This feature is impor-
tant since the main objective is the estimation of wake vortex circulation in order to evaluate
their strength (usually by means of model-based algorithms).
Our analysis has principally focused on the detection of wake vortices interacting with the
ground. An interesting study would be to consider cross wind as it appears in real atmo-
spheric conditions. The transmission of radial velocity estimation errors on the estimation of
vortex strength would also be considered.
6. Acknowledgements
The fluid and LIDAR simulations were done using the facilities of the Calcul Intensif et Stock-
age de Masse (CISM) on the UCL Lemaı̂tre cluster funded by Fond National de Recherche
Scientifique (FNRS) under a FRFC Project and by UCL (Fonds Spéciaux de Recherche). Re-
sults were obtained in the framework of the LASEF project funded by the Région Wallonne of
Belgium.
7. References
Banakh, V. A., Smalikho, I. N. & Werner, C. (2000). Numerical simulation of the effect of refrac-
tive turbulence on coherent lidar return statistics in the atmosphere, Applied Optics,
Vol. 39(No. 30): 5403-5414.
Belmonte, A. & Rye, B. J. (2000). Heterodyne lidar returns in the turbulent atmosphere: perfor-
mance evaluation of simulated systems, Applied Optics, Vol. 39(No. 15): 2401-2411.
Belmonte, A. (2003). Analyzing the efficiency of a practical heterodyne lidar in the turbulent
atmosphere: telescope parameters, Optics express, Vol. 11(No. 17): 2041-2046.
Bricteux, L. (2008). A new multiscale model with proper behaviour in both vortex flows and
wall bounded flows, Direct and Large-Eddy Simulation VII book, ERCOFTAC Series,
Springer.
Brousmiche, S., Bricteux, L., Sobieski, P., Winckelmans, G., Macq, B. & Craeye, C. (2009). Pa-
rameters Estimation of Wake Vortices in Ground Effect, Proceedings of the 15th Coherent
Laser Radar Conference, Toulouse, pp. 149-152.
Frehlich, R. G. and Kavaya, M. G. (1991). Coherent laser radar performance for general atmo-
spheric refractive turbulence, Applied Optics, Vol. 30(No. 36): 5325-5351.
Frehlich, R. G. (1993). Effect of refractive turbulence on coherent laser radar, Applied Optics,
Vol. 32(No. 12): 2122-2139.
Frehlich, R. G. (1997). Effect of Wind Turbulence on Coherent Doppler Lidar Performance,
Journal of Atmospheric and Oceanic Technology, Vol. 14.
Frehlich, R. G. (2000). Simulation of laser propagation in a turbulent atmosphere, Applied Op-
tics, Vol. 39(No. 3).
Gerz, T., Holzapfel, F. & Darracq D. (2001). Aircraft wake vortices: a position paper, Wakenet,
the european thematic network on wake vortex position paper.
Levin, M. J. (1965). Power Spectrum Parameter Estimation, IEEE Trans. on Information theory,
Vol. IT-11: pp 100-107
Lee, M. J., Oh, B. D. & Kim, Y. B. (2001). Canonical Fractional-Step Methods and Consistent
Boundary Conditions for the Incompressible Navier Stokes Equations, J. Comp. Phys.,
Vol. 168: 73-100.
Martin, J. M. & Flaté, S. M. (1990). Simulation of point-source scintillation through three-
dimensional random media, J. Opt. Soc. Am. A, Vol. 7(No. 5): 838-847.
Prandtl, L. & Tietjens, O.G (1957). Fundamentals of hydro- and aero-mechanics, Dover publ.
Proakis, J. G. (1965). Digital Signal Processing : Principles, Algorithms and Applications, Prentice
Hall.
Rye, B. J. & Hardesty, R. M. (1993). Discrete Spectral Peak Estimation in Incoherent Backscatter
Heterodyne Lidar. I: Spectral Accumulation and the Cramer-Rao Lower Bound, IEEE
Trans. on Geoscience and Remote Sensing, Vol. 31(No. 1): 16-27.
Rye, B. J. & Hardesty, R. M. (1993). Discrete Spectral Peak Estimation in Incoherent Backscat-
ter Heterodyne Lidar. II: Correlogram Accumulation, IEEE Trans. on Geoscience and
Remote Sensing, Vol. 31(No. 1): 28-35 .
Salamitou, P., Dabbas, A. & Flamant, P. (1995). Simulation in the time domain for heterodyne
coherent laser radar, Applied Optics, Vol. 34(No. 3): 499-506.
Van Trees, H. (2001). Detection, Estimation and Modulation Theory. Radar-sonar Processing and
Gaussian Signals in Noise, Wiley, 2001
Vasilyev, O. V. (2000). High order finite difference schemes on non-uniform meshes with good
conservation properties, J. Comp. Phys., Vol. 157(No. 2): 746-761.
Winckelmans, G. S. (2004). Chapter 5: Vortex Methods, Encyclopedia of Computational Mechan-
ics, Erwin Stein, René de Borst, Thomas J.R. Hughes (Eds), John Wiley & Sons.
Earthquake damage mapping techniques using SAR and Optical remote sensing satellite data 269
12
X
Earthquake damage mapping techniques using

SAR and Optical remote sensing satellite data.
Marco Chini
Istituto Nazionale di Geofisica e Vulcanologia
Italy
1. Introduction
The objective of this chapter is to show how it is possible to determine damages caused by
seismic events in urban areas using optical and radar data, and automatic or semi-automatic
remote sensing techniques. These techniques have revealed themselves a suitable
monitoring tool for disaster management since they provide a quick detection of land
changes in wide areas, especially in remote areas or where the infrastructures are not well
developed to ensure the necessary communication exchanges. Indeed, in the aftermath of
these severe disastrous events one of the most urgent needs is to estimate with sufficient
reliability and rapidity the amount of population and infrastructures affected for different
degrees of damage (Voigt et al., 2007).
The use of these sensors has been addressed to scenario classification procedures, which
allow also to recognize objects and changes occurred in the area. Automatic classification
procedure concerning changes in urban areas could be used for mapping damage caused by
earthquakes as well, being seismic events not so frequents.
Actually, timely and accurate change detection of Earth’s surface features is extremely
important for understanding relationships and interactions between human and natural
phenomena, especially in supporting better decision making. In general, change detection
involves the application of multi-temporal datasets to quantitatively analyse the temporal
effects of the phenomenon. Because of the advantages such as repetitive data acquisition,
synoptic view, and digital format suitable for computer processing, remote sensing data are
primary sources which have been using extensively for change detection since the last
decades (Lu et al., 2004).
The contribution of space technologies has been demonstrated to be effective for
regional/continental damage assessment using low- or medium-resolution remotely sensed
data (ranging from 30m to 1 km), and both automatic and manual interpretation approaches
have been successfully used for extraction of information at a nominal scale ranging from 1 :
100 000 to 1 : 1 000 000 (Belward et al., 2007).
The space technologies have also been demonstrated their effectiveness for damage
assessment at local scale, ranging from 1 : 10 000 to 1 : 25 000 nominal scales. The
information extracted at this level was crucial for calibration and estimation of the reliability
of low- and medium-resolution assessment, for planning logistics for relief action on the
field immediately after the event, and for planning the resources needed for recovery and
reconstruction.
Today, local or detailed damage assessment can also be addressed using Very High
Resolution (VHR) satellite data with a spatial resolution ranging from 0.6 to 1m. At the
beginning, , the operational methodology for extracting the information at this scale of
details was based on manual photo-interpretation of the satellite images, which were
processed on the screen by the photo-interpreter as for any other aerial imagery. However,
the traditional photo-interpretation methodology has some drawbacks firstly linked to the
time (and cost) needed for manual processing of the data and, secondly, to the difficulty in
maintaining coherent interpretation criteria in case a large number of photo-interpreters,
working in parallel in wide areas in a short time, is available. The long required processing
time is in conflict with the need for rapid damage estimation, and the solution to involve
parallel photo-interpreter teams often leads to an increase of time-consuming organizational
problems and additional coherency lack in the information produced (Pesaresi et al., 2007).
Accordingly, some automatic procedures for exploiting these data have been developed in
order to give information at this scale of detail. The most innovative automatic approaches
will be described in this chapter.
For an operational use of this kind of techniques the major limitation is the availability of the
images within a short time to manage the crisis. This is a key point for Civil Protections who
needs a fast and draft overview of the epicentral area, quick information relative to the
extension and distribution of damages, and the evaluation of infrastructure (roads, bridges)
conditions. A single satellite can provide access time to a specific site in the order of some
days, as a result the necessity to use any kind of satellites data available and an integration
of those data is mandatory to increase the chance to collect information on near real time.
The following paragraphs are addressed to the analysis of the different aspects leading to
obtain maps representative of damage caused by earthquakes.
2. SAR and Optical satellite sensors and missions

Earth observation satellites have a major mission to observe the entire Earth, so they circle
the most suitable Sun-synchronous sub-recurrent orbit. A Sun-synchronous orbit refers to
the orbit where the positioning between the satellite and the Sun is always identical .
Namely, the incidence angle of sunlight to the land surface is always the same. This orbit
characteristic allows to accurately observing radiation and reflection from the land surface.
While the sub-recurrent orbit means that after a certain number of days, the satellite repeats
its original orbit. Thus, Earth observation satellites utilize the combination of these two
orbits, which allow observing the same area at regular intervals under the condition of the
same sunlight angle to the land surface.
Furthermore, sensors can be divided into two types: Optical and Microwave sensors.
Optical sensors observe visible lights and infrared rays (near infrared, intermediate infrared,
thermal infrared). There are two kinds of observation methods using optical sensors:
visible/near infrared remote sensing and thermal infrared remote sensing. The first one
consists in acquiring visible light and near infrared rays of sunlight reflected by objects on
the ground. By examining the strength of reflection, we can understand conditions of land
surface, e.g., distribution of plants, forests and farm fields, rivers, lakes, urban areas. During
period of darkness this method cannot be observe. Instead, the latter method acquires
thermal infrared rays radiated from land surface heated by sunlight. It can also observe the
high temperature areas, such as volcanic activities and fires. By examining the strength of
radiation, we can understand surface temperatures of land and sea, the status of volcanic
activities and forest fires.
Microwave sensors transmit and/or receive microwaves, longer wavelength than visible
light and infrared rays, and the observation is time and weather independent. There are two
types of observation methods using microwave sensor: active and passive. In the first one
the sensor aboard earth observation satellite emits microwaves and observes microwaves
reflected by land surface, as the case of Synthetic Aperture Radar (SAR). The second one
observes microwaves naturally radiated from land surface. It is suitable to observe sea
surface temperature, snow accumulation and thickness of ice.
While optical sensors are affected by cloud cover limitations, SAR is widely used in
environmental studies due to its fairly synoptic view in almost completely weather and time
independent conditions. Moreover, like the optical sensors which have reached resolution
less than one meter (e.g. QuickBird, Ikonos, WorldView – 1, GoeEye-1 ) since 2001, SAR has
overtaken its own limitations in terms of ground resolution with the launch of the new
generation satellite missions TerraSAR-X and COSMO-SkyMed .
Destructive earthquakes challenge satellite remote sensing damage mapping techniques to
demonstrate their usefulness in supporting intervention and relief actions. The use of
remote sensing data in a disaster context has been widely investigated from a theoretical
point of view, but only recently the developed methods seem to be useful for the operational
use (Chini et al., 2008d).
Nowadays the implementation of satellite constellations is reducing the access time with the
same sensor to 12 hours, as in the case of the Italian COSMO-SkyMed system. One the most
important aspects of this new satellite mission is that the COSMO-SkyMed system is a
constellation of four satellites (three of which are already in orbit, the first of which was
launched in June 2007), developed to provide fast, meter level-resolution, all-weather
imagery for disaster management (Stramondo et al., 2008). COSMO-SkyMed is a project in
cooperation between the Italian Space Agency (ASI) and the Italian Defence Ministry (MD)
and it is the first satellite constellation devoted to the Earth’s observation for both civil and
military purposes. The system hosts a flexible, multimode X-band synthetic aperture radar
(SAR), with right and left looking imaging capabilities, an incidence angle range of 20°–60°,
and a minimum revisit time of 12 hours. The fixed antenna has electronic scanning
capabilities in both the azimuth and the elevation planes. It has been designed to implement
three different operation modes in order to acquire the images at different resolutions: i)
Spotlight mode, for metric (1m) resolutions over small images; Stripmap mode for metric
resolutions (5 m) over tenth of km images; iii) ScanSAR mode, for medium to coarse (100 m)
resolutions over large swaths (Boni et al, 2008).
3. Damage mapping by SAR data

Multi-temporal observations from SAR can be used to detect urban changes either looking
at the image intensity changes, as in the case of optical images, but also taking advantage
from the information on the phase of the returned signal. This is specific of the radar
technique and in particular of the interferometric SAR (InSAR). The intensity correlation
and the InSAR complex coherence are two important features for recognizing changes on
the surface caused by an earthquake. These two features hold different information
concerning changes in the scene (Stramondo et al., 2006). The complex coherence is
prevalently influenced by the phase difference between radar returns, a distinctive
parameter measured by a coherent sensor. It is particularly related to the spatial
arrangement of the scatterers within the pixel and thus to their possible displacements.
Conversely, the intensity correlation is more related to change in the magnitude of the radar
return (Bignami et al., 2004).
Changes in SAR intensity and phase backscattering have been largely studied by many
scientists for earthquake damage mapping purposes. An index to estimate damage level
from SAR data by combining image intensity changes and the related correlation coefficient
has been achieved and applied to some case studies: the Hyogoken-Nanbu earthquake
(Aoki et al., 1998; Matsuoka & Yamazaki, 2004) and the Izmit and Gujarat seismic events
(Matsuoka & Yamazaki, 2002; Stramondo el al., 2006). Yonezawa & Takeuchi (2001) have
compared changes in the SAR intensity and phase backscatter with damage observed in
Kobe. Ito et al. (2000) have assessed different SAR change indicators, derived from L- and C-
band sensors, and evaluated the frequency-dependent effects of spatial and temporal
decorrelations. Chini et al. (2008a) have detected wide uplift and subsiding areas, as well as
large modifications of the coastline associated to the Indonesian earthquake of 2004, using
only pre- and post-earthquake SAR backscattering.
It is commonly acknowledged that due to speckle effects (Li & Goldstein, 1990), single-pixel
classification of SAR images leads to unsatisfactory results, and this seems to hold true also
when damage assessment is concerned. Satisfying results may be achieved if the damage is
assessed at a block level, mitigating the speckle noise (Bazi ert al., 2005) by means of
averaging in somehow the unreliable results of pixel-wise comparing pre-and post-event
images (Gamba et al., 2007). The partition of the image into blocks can be made by visual
inspection or using a GIS layer provided by local authorities during crisis phase; otherwise
procedures for segment images starting from SAR or optical images are reported in
literature. Indeed, city blocks are generally marked by their respective bounding urban
roads, which allow to segmenting the urban area.
Nowadays the new very high resolutions SAR sensors, few meters resolution, allow to
classifying urban areas, producing reliable land use maps exploiting contextual information
from backscattering data. Chini et al., 2009b have investigated the use of contextual
information with TerraSAR-X backscattering images for classifying urban land-use. The
anisotropic multi-scale morphological filters, coupled with the pruning neural network as a
features selection tool, was able to provide urban land-use maps with accuracies of about
0.90 in terms of K-coefficient. This kind of techniques will allow producing damage map at
building scale using very high resolution SAR data as it is already done using Very High
Resolution (VHR) optical images.
4. Damage mapping by Very High resolution Optical data

The presence of shadows, variations in solar illumination, and geometric distortions may
prevent the use of automatic damage detection procedures in optical images. Because of
these problems, visual image inspection is still the most widely used method to produce a
realistic and reliable inventory of damage (Saito et al., 2004; Yamakaki et al., 2004).
Sakamoto et al. (2004) compared with an automatic technique combined with visual
interpretation results using QuickBird data. Matsuoka et al. (2004) proposed to detect
damage by analyzing edges in high-resolution images. More in general, automatic change-
detection algorithms using either QuickBird (Pacifici et al., 2007; Chini et al., 2008) or SAR
images (Inglada & Mercier, 2007) are also present in the literature, even if using SAR data
have been only exploited the medium resolution.
One of the most important issue when we set a change detection classifier for mapping
damages is the rapidity for producing damage maps, especially if we are supporting rescue
teams during crisis phase. This characteristic is related to the number of inputs, which are
the bands used for classifying, the number of samples for training the supervised classifier
and the time spent for selecting training samples representative of the classes we want to
identify.
Usually, the unsupervised classifiers are used for overcoming the time consuming step for
selecting training pixels, since the post processing task for labelling the classes of interest is
faster. The use of an unsupervised classifier, with QuickBird panchromatic images, has
provided a damage map at a pixel scale of 0.6 m which allow to detect the complete or
partial collapse of individual buildings in the Bam city (Iran) caused by the earthquake
occurred on 2003 (Chini et al., 2009a). Moreover, the mathematical morphology
(Benediktsson et al., 2003; Benediktsson et al., 2005; Soille, 2003) has proved to be a powerful
tool for automatically analyze the panchromatic images. Furthermore, the unsupervised
classifier, the mathematical morphology and very high resolution images have permitted to
give damage level at building scale which is a more realistic damage degree and more
useful for civil protection activity. The damage level at building scale is closer to damage
index provided by ground survey team (Chini et al. 2008c).
When the number of training pixels is fixed or limited by whatever factor (including the cost
of gathering them by in situ campaigns), experience has shown an initial increase in the
classification accuracy by adding features, followed by a subsequent decay in classification
skill as more features are considered (Hughes, 1968).
The increase in dimensionality (i.e., number of inputs) of the data worsens the parameter
estimates, overcoming the expected increase in class separability associated with the
inclusion of additional features. Therefore, one needs to use a robust method to estimate
parameters or to reduce the number of inputs. For instance, principal component analysis is
used to diminish the number of features (Landgrebe, 2003). Regularization methods have
also been proposed in the literature, as discussed in (Berge & Solberg, 2004), attempting to
stabilize the estimated class covariance matrix by replacing it with a weighted sum of the
class sample covariance matrix or the common (pooled) covariance matrix. Additional
methods include regularizing discriminant analysis (Friedman, 1989), leave-one-out
covariance estimation (Hoffbeck & Landgrebe, 1996), and pruning methods when we deal
with neural networks (Pacifici et al., 2009).
A robust method for estimating parameters has been proposed by Chini et al., (2008b),
which is different respect to method presented previously. The number of inputs is not
reduced, but they are composed differently in the new architecture to overcome the
difficulty of having poor training samples for some classes. This method speeds up the
overall analysis time (computation plus error correction) and increases the change detection
accuracies. It consists on using a parallel approach to exploit the multi-spectral and multi-
temporal characteristics of two scenes, thus combining two approaches which are typically
used for supervised change detection, the post classification comparison (Del Frate et al.,
2004; Serra et al., 2004; Sunar Erbek et al., 2004) and the direct multi data classification
(Sunar Erbek et al., 2004; Castelli et al., 1998).
The post classification comparison detects changes by comparing the classification maps
obtained by independently classifying the two sequential remote sensing images of the same
scene. This method does not require normalization to compensate for atmospheric
conditions and sensor differences between the acquisition dates. At the same time, this
method does not exploit the correlation between multi-temporal data sets, and its accuracy
depends on the product of the accuracy of single classification maps (Yuan & Elvidge, 1998).
The change detection in direct multi data classification is directly performed by considering
each transition as a class in a unique vector obtained by stacking the features related to the
individual images; this method is more sensitive to solar and atmospheric conditions.
An efficient method for reducing the number of inputs and selecting the most important
features for classifying a certain typology of urban landscape has been implemented by
Pacifici et al., (2009), using QuickBird and Worldview-1 panchromatic images. This work
has investigated on the potential of these very high resolution panchromatic imageries (the
most resolute optic satellite sensors) to classify the land-use of urban environments. Usually
spectral-based classification methods may fail with the increased geometrical resolution of
the data available. Indeed, improved spatial resolution data increases within-class variances,
which results in high interclass spectral confusion. In many cases, several pixels are
representative of objects, which are not part of land-use classes defined. This problem is
intrinsically related to the sensor resolution and it cannot be solved by increasing the
number of spectral channels. To overcome the spectral information deficit of panchromatic
imagery, it is necessary to extract additional information to recognize objects within the
scene. The multi-scale textural (Haralick, 1979; Haralick et al., 1973) approach exploited the
contextual information of panchromatic images. Moreover, neural network pruning
(Kavzoglu & Mather 1999; Del Frate et al., 2005) and saliency (Yacoub & Bennani, 1997; Tarr,
1991) measurements have allowed to give an indication of the most important textural
features for sub-metric spatial resolution imagery and different urban scenarios.
This has to be a preparatory work, it should be done before the emergency phase, because it
gives very useful information concerning the minimum number of inputs necessary for
obtain reliable classification maps. Another important aspect of this study is that it uses a
very small dataset, only one panchromatic image, all others inputs are contextual
information, second order spatial statistic, extracted from the original panchromatic image.
The latter aspect is crucial during crisis, where it is necessary processing data as soon as
they are acquired from satellites.
5. Conclusions
In this chapter the sensitivity to the urban damage level of different parameters extracted
from different types of remotely sensed images has been analyze. It has been highlighted
that the parameters to be extracted are strictly related to the type of exploited data (i.e., their
resolution and spectral channels). Additionally, this chapter aimed to demonstrate how very
high resolution (VHR) images can detect damage at the pixel scale (≤ 1 m) with automatic
techniques. Associated problems have been pointed out and ways to overcome them have
been proposed.
The chapter has provided a quite realistic view of what remote sensing can offer for this
application and which methods and sensors worth to be further developed in the future.
From our experience in this field it is possible to affirm that damage products can be
obtained at different ground scales and with different satellite data. Those products can be
valuable for managing different phases of the crises, and especially mitigation of the effect
in the course of the event and precise inventory of the damage in the post-event phase.
It has been pointed out that an operational use of remote sensing for earthquake damage
assessment strongly depends on the number of available images, their type (SAR, optical, or
both), and quality and timeliness of the data sets (i.e., time delay of the post-seismic images
with respect to the destructive event). In particular, the availability of images within a short
time to manage the crisis is a key point for Civil Protections who needs a fast and draft
overview of the epicentral area, quick information relative to the extension and distribution
of damages, and the evaluation of infrastructure (roads, bridges) conditions. Since, a single
satellite can provide access time to a specific site in the order of some days, the use of any
type of satellites data available, an integration of those data and the exploitation of data
provided by constellation missions is mandatory to increase the chance to collect
information on time.
The conclusive indication on the effectiveness of remote sensing for earthquake damage
detection requires the accumulation of many analyses by different scientists and the work
done here will hopefully contribute to this end. For all of these reasons, research of change
detection techniques is still an active and very challenging topic and new techniques are
needed to effectively use the increasingly diverse and complex remotely sensed data
available.
6. References
Aoki, H.; Matsuoka, M. & Yamazaki, F. (1998). Characteristics of satellite SAR images in the
damaged areas due to the Hyogoken-Nanbu earthquake, Proceedings of 19th Asian
Conference on Remote Sensing 1998, Vol. C7, pp. 1–6.
Bazi, Y.; Bruzzone, L. & Melgani, F. (2005). An unsupervised approach based on the
generalized Gaussian model to automatic change detection in multitemporal SAR
images, IEEE Transaction of Geoscience and Remote Sensing, Vol.43, No.4, pp.
874-887.
Belward, A. S.; Stibig H. J.; Eva, H.; Rembold, F.; Bucha, T.; Hartley, A.; Beuchle, R.;
Khudhairy, D.; Michielon, M. & Mollicone, D. (2007). Mapping severe damage to
land cover following the 2004 Indian Ocean tsunami using moderate spatial
resolution satellite imagery, International Journal of Remote Sensing, Vol. 28, No.
13–14, pp. 2977–2994.
Benediktsson, J. A.; Pesaresi, M. & Arnason, K. (2003). Classification and feature extraction
for remote sensing images from urban areas based on morphological
transformations, IEEE Transactions on Geoscience and Remote Sensing, Vol. 41,
No. 9, pp. 1940–1949.
Benediktsson, J. A.; Palmason, J. A. & Sveinsson, J. R. (2005). Classification of hyperspectral
data from urban areas based on extended morphological profiles, IEEE
Transactions on Geoscience and Remote Sensing, Vol. 43, No. 5, pp. 480–491.
Berge, A. & Solberg, A. S. (). A comparison of methods for improving classification of

hyperspectral data, Proceedings of IEEE IGARSS 2004, Anchorage (Alaska, USA),
Vol. 2, pp. 945–948, September 20-24, 2004.
Bignami, C.; Chini, M.; Pierdicca, N. & Stramondo, S. (2004). Comparing and combining the
capability of the detecting earthquake damages in urban areas using SAR and
optical data, Proceedings of IEEE IGARSS 2004, Vol. 1, pp.55-58, Anchorage
(Alaska, USA), September 20-24, 2004.
Boni, G.; Castelli, F.; Ferraris, L.; Pierdicca, N.; Serpico, S. & Siccardi, F. (2007). High
resolution COSMO/SkyMed SAR data analysis for civil protection from flooding
events, Proceeding of IEEE IGARSS 2007, pp. 6 – 9, Denver (CO - USA), July 23-28,
2007.
Castelli, V.; Elvidge, C. D.; Li, C. S. & Turek, J. J. (1998). Classificationbased change
detection: Theory and application to the NALC data set, In: Remote Sensing
Change Detection: Environmental monitoring methods and applications, pp. 53-73,
R. S. Lunetta and C. D. Elvidge, Eds. Chelsea, MI: Ann Arbor Press, 1998.
Chini, M.; Bignami, C.; Stramondo S. & Pierdicca, N. (2008a). Uplift and subsidence due to
the 26 December 2004 Indonesian earthquake detected by SAR data, International
Journal of Remote Sensing, Vol. 29, No. 13, pp. 3891–3910.
Chini, M.; Pacifici, F.; Emery, W. J.; Pierdicca, N. & Del Frate, F. (2008b). Comparing
statistical and neural network methods applied to very high resolution satellite
images showing changes in man-made structures at Rocky Flats, IEEE Transactions
on Geoscience and Remote Sensing, Vol. 46, No. 6, pp. 1812–1821.
Chini, M.; Bignami, C.; Stramondo, S.; Emery, W. J. & Pierdicca, N. (2008c). Quickbird
Panchromatic Images for Mapping Damage at Building Scale Caused by the 2003
Bam Earthquake, Proceedings of IEEE IGARSS 2008, Vol. 2, pp. 1029-1031, Boston
(USA), July 7 – 11, 2008.
Chini, M.; Bignami, C.; Brunori, C.; Atzori, S.; Tolomei, C.; Trasatti, E.; Kyriakopoulos, C.;
Salvi, S.; Stramondo, S. & Zoffoli, S. (2008d). The SIGRIS project: a remote sensing
system for seismic risk management, Proceedings of IEEE IGARSS 2008, Vol. 3, pp.
624-627, Boston (Massachusetts, USA), July 7 – 11, 2008.
Chini, M.; Pierdicca, N. & Emery, W. J. (2009a). Exploiting SAR and VHR optical images to
quantify damage caused by the 2003 Bam earthquake. IEEE Transaction on
Geosciences and Remote Sensing, Vol. 47, No. 1, pp. 145 - 152.
Chini, M.; Pacifici, F. & Emery, W. J. (2009b). Morphological operators applied to X - band
SAR for urban land use classification, Proceedings of IEEE IGARSS 2009, Cape
Town (South Africa), July 13 – 18, 2009.
Del Frate, F.; Schiavon, G. & Solimini, C. (2004). Application of neural networks algorithms
to QuickBird imagery for classification and change detection of urban areas,
Proceedings of IEEE IGARSS 2004, Anchorage (Alaska, USA), pp. 1091–1094,
September 20-24, 2004.
Del Frate, F.; Iapaolo, M.; Casadio, S.; Godin-Beekmann, S. & Petitdidier, M. (2005). Neural
networks for the dimensionality reduction of GOME measurement vector in the
estimation of ozone profiles, Journal of Quantitative Spectroscopy & Radiative
Transfer, Vol. 92, No. 3, pp. 275−291.
Friedman, J. (1989). Regularized discriminant analysis, Journal of the American Statistical
Association, Vol. 84, No. 405, pp. 165–175.
Gamba, P.; Dell'Acqua, F. & Trianni, G. (2007). Rapid Damage Detection in the Bam Area
Using Multitemporal SAR and Exploiting Ancillary Data, IEEE Transaction of
Geoscience and Remote Sensing, Vol.45, No.6, pp.1582-1589.
Haralick, R. M.; Shanmuga, K. & Dinstein, I. H. (1973). Textural features for image
classification, IEEE Transactions on Systems, Man, and Cybernetics, Vol. SMC-3,
No. 6, pp. 610−621.
Haralick, R. M. (1979). Statistical and structural approaches to texture. Proceedings of the
IEEE, Vol. 67, No.5, pp. 786−804.
Hoffbeck, J. & Landgrebe, D. (1996). Covariance matrix estimation and classification with
limited training data, IEEE Transactions on Pattern Analysis and Machine
Intelligence, Vol. 18, no. 7, pp. 763–767.
Hughes, G. F. (1968). On the mean accuracy of statistical pattern recognizers, IEEE
Transactions on Information Theory, Vol. IT-14, No. 1, pp. 55–63.
Inglada, J. & Mercier, G. (2007). A new statistical similarity measure for change detection in
multitemporal SAR images and its extension to multiscale change analysis, IEEE
Transactions on Geoscience and Remote Sensing, Vol. 45, No. 5, pp. 1432–1445.
Ito, Y.; Hosokawa, M.; Lee, H. & Liu, J. G. (2000). Extraction of damaged regions using SAR
data and neural networks, Proceedings of 19th ISPRS Congress 2000, vol. 33, pp.
156–163, Amsterdam, (NETHERLANDS), July 16–22, 2000.
Kavzoglu, T. & Mather, P. M. (1999). Pruning artificial neural networks: An example using
land cover classification of multi-sensor images. International Journal of Remote
Sensing, Vol. 20, N. 14, 2787−2803.
Landgrebe, D. A. (2003). Signal Theory Methods in Multispectral Remote Sensing, Hoboken,
NJ: Wiley, 2003.
Li, F. K. & Goldstein, R. M. (1990). Studies of multibaseline spaceborne interferometric
synthetic aperture radar,” IEEE Transaction of Geoscience and Remote Sensing,
Vol. 28, No. 1, pp. 88–97.
Lu D.; Mausel, P.; Brondizio, E. & Moran, E. (2004). Change detection techniques.
International Journal of Remote Sensing, Vol. 25, No. 12, pp. 2365–240720.
Matsuoka, M. & Yamazaki, F. (2002). Application of the damage detection method using
SAR intensity images to recent earthquakes, Proceedings of IEEE IGARSS 2002,
Vol. 4, pp. 2042–2044, Toronto (CANADA), June 24-28, 2002.
Matsuoka, M. & Yamazaki, F. (2004). Use of satellite SAR intensity imagery for detecting
building areas damaged due to earthquakes, Earthquake Spectra, Vol. 20, No. 3, pp.
975–994.
Matsuoka, M.; Vu, T. T. & Yamazaki, F. (2004). Automated damage detection and
visualization of the 2003 Bam, Iran, earthquake using high resolution satellite
images, Proceedings of 25th Asian Conference on Remote Sensing 2004, pp. 841–
845, Chiang Mai (THAILAND), November 22-26, 2004.
Pesaresi, M.; Gerhardinger, A. & Haag, F. (2007). Rapid damage assessment of built-up
structures using VHR satellite data in tsunami-affected areas, International Journal
of Remote Sensing, Vol. 28, No. 13–14, pp. 3013–3036.
Pacifici, F.; Del Frate, F.; Solimini, C. & Emery, W. J. (2007). An innovative neural-net
method to detect temporal changes in high-resolution optical satellite imagery,
IEEE Transactions on Geoscience and Remote Sensing, Vol. 45, No. 9, pp. 2940–
2952.
Pacifici, F.; Chini, M. & Emery, W. J. (2009). A neural network approach using multi-scale
textural metrics from very high resolution panchromatic imagery for urban land-
use classification, Remote Sensing of Environment, Vol. 113, No. 6, pp. 1276 – 1292.
Saito, K.; Spence, R. J. S.; Going, C. & Markus, M. (2004). Using high resolution satellite
images for post-earthquake building damage assessment: A study following the 26
January 2001 Gujarat earthquake, Earthquake Spectra, vol. 20, No. 1, pp. 145–169.
Sakamoto, M.; Takasago, Y.; Uto, K.; Kakumoto, S. & Kosugi, Y. (2004). Automatic detection
of damaged area of Iran earthquake by high-resolution satellite imagery,”
Proceedings of IEEE IGARSS 2004, vol. 2, pp. 1418–1421, Anchorage (Alaska, USA),
September 20-24, 2004.
Serra, P.; Pons, X. & Saurì, D. (2004). Post-classification change detection with data from
different sensors: Some accuracy considerations, International Journal of Remote
Sensing, Vol. 24, No. 16, pp. 3311–3340.
Soille, P. (2003). Morphological Image Analysis—Principles and Applications, 2nd ed.
Berlin, Germany: Springer-Verlag, 2003.
Stramondo, S.; Bignami, C.; Chini, M.; Pierdicca, N. & Tertulliani, A. (2006). The radar and
optical remote sensing for damage detection: results from different case studies,
International Journal of Remote Sensing, Vol. 27, No. 20, pp. 4433 – 4447.
Stramondo, S.; Chini, M.; Salvi, S.; Bignami, C.; Zoffoli, S. & Boschi, E. (2008). Ground
Deformation Imagery of the 12 May 2008 Sichuan Earthquake, EOS Transaction
American Geophysical Union, Vol. 89, No. 37, pp. 341-342.
Sunar Erbek, F.; Ozkan, C. & Taberner, M. (2004). Comparison of maximum likelihood
classification method with supervised artificial neural network algorithms for land
use activities, International Journal of Remote Sensing, Vol. 25, No. 9, pp. 1733–
1748.
Tarr, G. (1991). Multi-layered feedforward neural networks for image segmentation, Ph. D.
dissertation, School of Engineering, Air Force Institute of Technology. Wright-
Patterson AFB OH.
Voigt, S.; Kemper, T.; Riedlinger, T.; Kiefl, R.; Scholte, K. & Mehl, H. (2007). Satellite image
analysis for disaster and crisis-management support, IEEE Transaction on
Geoscience and Remote Sensing, Vol. 45, No. 6, pp. 1520–1528.
Yacoub, M. & Bennani, Y. (1997). HVS: A heuristic for variable selection in multilayer
artificial neural network classifier. In: Intelligent engineering systems through
artificial neural networks, pp. 527-532, Dagli, C.; Akay, M.; Ersoy, O.; Fernandez, B.
& Smith, A. (Eds.), Missouri, USA: St Louis.
Yamakaki, F.; Matsuoka, M.; Kouchi, K.; Kohiyama, M. & Muraoka, N. (2004). Earthquake
damage detection using high-resolution satellite images, Proceedings of IEEE
IGARSS 2004, Vol. 4, pp. 2280–2283, Anchorage (Alaska, USA), September 20-24,
2004.
Yonezawa, C. & Takeuchi, S. (2001). Decorrelation of SAR data by urban damage caused by
the 1995 Hoyogoken-Nanbu earthquake, International Journal of Remote Sensing,
Vol. 22, No. 8, pp. 1585–1600.
Yuan, d. & Elvidge, C. (1998). NALC land cover change detection pilot study: Washington
D.C. area experiments, Remote Sensing of Environment, Vol. 66, No. 2, pp. 168–178.
Spectroscopic Microwave Dielectric Model of Moist Soils 279
13
X
Spectroscopic Microwave Dielectric

Model of Moist Soils
Valery Mironov and Pavel Bobrov
Kirensky Institute of Physics Siberian Branch of the Russian Academy of Sciences
Russian Federation
1. Introduction
Microwave dielectric models (MDM) of moist soils are an essential part of the algorithms
used for data processing in the radar and radio thermal remote sensing (Ulaby et al., 1986).
So far, the MDMs for moist soils developed in (Wang & Schmugge, 1980; Dobson et al., 1985;
Peplinski et al., 1995, a, b) have become a routine mean for creating processing algorithms,
which concern radar and radio thermal remote sensing of the land. In our recent
publications (Mironov et al., 2002; Mironov, 2004; Mironov et al., 2004; Mironov et al.
2009 a, b) the error of dielectric predictions for moist soils has been noticeably decreased
due to taking into account, on a physical bases, the impact of bound soil water (BSW). This
is an essential distinction of the dielectric models developed in (Mironov, et al. 2002;
Mironov, 2004; Mironov et al., 2004; Mironov et al., 2009 a, b) from the previous dielectric
models, which are mainly of regression origin when it concerns impact of BSW. In these
works, the MDM for moist soil was worked out using an extensive continuum of
experimental dielectric data, both measured by the authors and borrowed from the
literature. As a result, the newly developed MDM for moist soils attained features of a
physical law and ensured more accurate dielectric predictions, with the soil clay content
being an intrinsic texture parameter of the soil. Further research in this area has to be aimed
at broadening the ensemble of soil types, temperatures, and frequency ranges over which
the new MDM can be effectively used for the radar and radio thermal remote sensing of the
land, including the ground penetrating radars and TDR applications.
In this chapter, a methodology to develop physically based MDM for moist soils thawed is
presented. The moist soil consists of a matrix of mineral and organic particles, air, and
aqueous solutions, the latter being further referred to as soil water. It is evident that, when
developing a physically based MDM, the law of wave propagation through such a mixture
is the first problem to be solved.
Let us consider the propagation of plane harmonic wave in a lossy, nonmagnetic, and
unbounded mixture. In this case, the wave amplitude as a function of time, t, and
propagation distance, x, may be written in the form:
exp[i(kx-ωt)] (1)
where i   1 is imaginary unit, k= k0 n* − wave propagation constant in a mixture, k0 =

ω/c − wave propagation constant in a vacuum, ω =2πf− circular wave frequency, f− wave
frequency, c − velocity of light in a vacuum, n*=n+iκ − complex index of refraction (CIR),
n=Rek/ k0 − index of refraction (IR), κ=Imk/k0 − normalized attenuation coefficient (NAC),
with n*, n, and κ being related to a mixture. The CIR is a square root of complex dielectric
constant (CDC), ε= ε′+ iε″:
n*   (2)
For the real, ε′, and imaginary, ε″, parts of the CDC, there will be used terms dielectric
constant (DC) and loss factor (LF), respectively. According to (2), the DC and LF can be
expressed through the IR and NAC and vise versa:
ε′=n2 - κ2, ε″ = 2n κ (3)
n  2   2    2,    2   2    2, (4)
In general, a frequency and temperature, T, dependent CDC for a mixture containing N

components can be written in the form
εs=εsm(m1,…,mN, ε1,…, εN ,f,T)= εsv(v1,…,vN, , ε1,…, εN, f,T) (5)
where εp , mp and vp (p=1,…,N) are respectively CDC, gravimetric and volumetric

percentages of individual components. We consider the following individual components
mixed in moist soil (Ulaby et al., 1986):
1. Mineral matrix consisting of particles of various minerals, which is characterized by
gravimetric percentages of clay, C, sand, S, and silt, L=100%-C-S;
2. Air contained in soil pores;
3. Bound soil water (BSW) consisting of conglomerates of water molecules bound on the
surface of the soil mineral matrix;
4. Free soil water (FSW), which is located in soil pores and exists in a form of capillary
aggregations;
The CDC of air is considered to be equal to one, εa=1. It has been shown (Ulaby et al., 1986)
that the CDC of mineral matrix, εm(C, S), depends on gravimetric percentages of clay, C,
sand, S, and silt, 100%-C-S, being practically independent on the temperature and wave
frequency. On the contrary, due to water molecules relaxation (Hasted J., 1973) the CDCs of
BSW, εb(f, T, C, S), and FSW, εu(f, T, C, S), are dependent on frequency and temperature. In
addition, these values are considered to be dependent on the soil texture parameters. The
latter is justified by bound water molecules interaction with the soil mineral matrix surface
and the mineral matrix impact on geometric form and size of capillary water aggregations.
At the same time, we assume the CDCs of BSW and FSW to be mutually independent on
their percentages, mb and mu. With these suggestions, the CDC of moist soil can be expressed
in the form
εs(vm, va, vb, vu, f,T) =εs(vm, va, vb, vu, εm(C, S), εb(f, T, C, S), εu(f, T, C, S)). (6)
As seen from (6), the assumptions made allowed to decompose the variables in (5)
representing the percentages of moist soil components and consider the CDCs of individual
components as parameters of a mixture dielectric model, which are in their turn dependent
on frequency, temperature, and soil texture characteristics. The equation (6) suggests the
CDCs of individual components to be independent on their percentages. Such mixtures can
be identified as conservative, once their individual components do not affect each other.
This assumption may not be true, for instance, in the case of saline soils, with the values of
εb(vb, f, T, C, S), εu(vu, f, T, C, S) varying due to possible change of soil solute concentration,
with the percentages vb and vu being changed. There have been proposed plenty of mixture
models, which are discussed in (Wang & Schmugge, 1980). From all of those, the refractive
mixing dielectric model (RMDM) of (Birchak, 1973)
(εs)=∑ vp(εp), p=m, a, b, u (7)
was used as a basis for the moist soil dielectric models suggested in
(Wang & Schmugge,1980) and (Dobson et al., 1985). In formula (7), α is a constant shape
factor, and the subscripts m, a, b, and u designate mineral matrix, air, bound water, and free
water, respectively. As seen from (7), the RMDM meets a conservative mixture condition. In
the following section, there are discussed some specific equations for the RMDMs
contemporarily used.
2. Refractive mixing dielectric models

Let us consider the propagation of plane wave through a mixture using the geometrical
optics approximation. Given the plane wave ray is straight line, a statistically averaged
complex phase of the wave that propagated the distance L in a mixture can be written in the
form
Nm Na Nb Nu
(8)
ns* k0 L  * * * *
 nm k0 Lmi   na k0 Lai   nb k0 Lbi   nu k0 Lui
i 1 i 1 i 1 i 1
In equation (8), the symbol <> designates statistical averaging over an ensemble of rays, Nm,
Na, Nb, Nu and Lmi, Lai, Lbi, Lui are respectively the numbers of wave paths and the distances
propagated through the media of soil matrix, air, BSW, and FSW, as related to an individual
cell of the soil. After statistical averaging and dividing both parts of (8) by k0L, the moist soil
CIR can be represented as follows:
ns*  nm* Lav  na* Lav  nb* Lbav  nu* Lav (9)
m a u
where the symbols Lav av av av

m , La , Lb , and Lu stand for the normalized by L mean wave paths
in the media of soil mineral matrix, air, BSW and FSW, respectively. Assuming the wave
beam to be non-divergent when propagating in the mixture, one can link the normalized
mean wave paths to the volumetric contents of respective substances contained in a mixture:
Lav
p  Vp / V (10)
where V and Vp are respectively the volumes of the mixture as a whole and the ones related
to specific components contained in the mixture (p=m, a, b, and u). Finally, substituting (10)
in (9) yields a refractive mixing dielectric model in the form of (7), with the shape factor, α,
being equal to:
α =1/2. (11)
Following the paper (Mironov et al., 1995), we re-arrange equation (7) as follows;
nd  nb  1 W , W  Wt , (12)

ns  
nd  nb  1 Wt  nu  1 W  Wt , W  Wt
   b W , W  Wt (13)
s   d
 d   b Wt   u W  Wt , W  Wt
where the volumetric soil moisture, W= Vw/V, is a ratio of the total soil water content, Vw, to
the given volume, V, of dry soil sample, Wt is the maximum bound water fraction (MBWF)
in a specific type of the soil. The MBWF is such an amount of water in soil that any
additional water added to the soil in access of this amount behaves as free soil water. As
seen from (1), the CIR is a piecewise linear function of soil moisture, with the MBWF being a
transition point in terms of slope angle between the two linear legs relating to the bound,
W≤Wt, and free, W>Wt, moisture ranges. In formulas (12) and (13), the variables nd and d
are respectively the IR and NAC for a dry soil. With the use of formula (7), these values can
be derived in the form
nm  1
nd  1    d  1  (nm  1)  (1  P ) (14)
m
m
d    d   m  (1  P ) (15)
m
where m and d are the specific and bulk soil densities, respectively, P=Va/Vd is the soil
porosity, which is equal to the ratio, of the air volume in a dry soil sample, Va, to a total
volume of dry soil sample, Vd. The RMDM was initially proposed in (Birchak, 1973), with no
distinction being made between the bound and free soil water. Later (Mironov et al., 1995), it
was generalized in the form given by the equation (12), (13), which made possible to
distinguish between contributions in the moist soil CIR, generated by the bound and free
types of soil water.
The term (nm*-1)/ρm can be considered collectively as a single mineral parameter of the soil
CIR. In the formulas (12)-(15), the values n*p=np+iκp, (p= s, m, d, b, u), Wt are understood as
functions of the temperature, soil texture, and frequency: n*d=n*d(T, C ,S), n*m=n*m(T, C ,S),
n*p=n*p(T, C, S,f); p=s, b, u; Wt=Wt(T, C, S).
3. Parameters of the RMDM

As seen from (12)-(13), the major advantage of the RMDM is that this model decomposed
the soil moisture and soil density variables, thus reducing the problem of developing the
moist soil dielectric model, as a function of temperature, soil texture, and frequency, to the
one pertaining to the bound. free soil water. and to the dry soil . To obtain building blocks
for developing frequency dependent dielectric models for the bound and free soil water, one
has to measure, as a function of frequency, the CIRs for these types of water, as parameters
in the formulas (12), (13), using the data for the moist soil CIRs measured at varying
moistures and frequencies (Mironov et al., 2004).
When fitting the IRs and NACs measured as a function of soil moisture, W, with the
formulas (12), (13), the values of nd and d can be obtained as the interceptions of the
respective fits with the axis of ordinates at W=0. While the values of MBWF, Wt, can be
identified and measured as a point at which the legs of the piecewise dependences (12), (13)
intercept each other. Also, the bound and free soil water RIs and NACs can be derived from
the slopes of the linear segments in the fits made with (12) and (13), respectively. As a result,
the RMDM parameters n*d, Wt, n*b, and n*u can be obtained for a specific soil, with the
texture parameters, C and S, temperature, T, and wave frequency, f, being fixed. Fitting the
CIR data for moist samples measured at varying frequencies, one can obtain the spectra
n*b(C, S, T, f), and n*u(C, S, T, f). The latter are further applied to derive the Debye relaxation
parameters for both the bound and free soil water (Mironov et al., 2004). In our publications,
the CIRs measured for moist soils were shown to be well fitted with the equations (12) and
(13).
To quantitatively demonstrate a degree of correlation of the RMDM model with the CIRs
measured for different types of soil at varying frequencies and temperatures, let us
transform the equations (12) and (13) into a universal piecewise linear moisture
dependences: which are able to accommodate all the data measured:
nb  1 Wb , Wb  Wt
ns   (16)
nb  1 Wt  nu  1 Wu , Wu  Wt
 W , Wb  Wt
s   b b (17)
 b Wt   u Wu , Wu  Wt
where nb , nu and  b ,  u are the bound and free soil water IR and NAC reduced values,
respectively, Wt is the reduced value of the MBWF assigned for the universal piecewise
linear RMDM. In the further analysis, we use the following numerical values for these
parameters: Wt  0.3 ; nb  1  10 / 3 ; nu  1  9 , kb  2 / 3 ; ku  8 / 3 .
This universal RMDM gives the values of reduced RI, ns , and NAC,  s :
ns (W )  (ns (W )  nd )(nb  1)W /( nb  1), W  Wt ;

(18)
ns (W )  [ns (W )  nd  (nb  1)Wt ]W (nu  1) /( nu  1), W  Wt ;

 s (W )  [ s (W )   d ]W  b /  b , W  Wt ;
(19)
 s (W )  [ s (W )   d   bWt ]W  u /  u , W  Wt
as a function of the reduced soil moistures,
W  W (Wt / Wt ) , W  Wt
(20)
W  Wt  W  Wt , W  Wt
To transform the measured IR, nsm, and NAC, sm, data as a function of the measured
moistures, Wm, to a respective universal piecewise linear RMDM, the following steps have to
be performed. First, fitting these data with the equations (12), (13) yields the measured
values of the IR, NAC, and MBWFs related to the dry soil, bound and free soil water, that is,
ndm, nbm, num, , dm, bm, um, , and Wtm. Second, using these values, together with the values
nb , nu , k b , k u , and Wt , and applying the formulas (18), (19), and (20), the experimental
values of the reduced RI, nsm , and NAC,  sm , can be found, as well as the reduced soil
moistures, Wm . As a result, any IR and NAC data measured as a function of moisture for a
given specific type of soil, frequency, and temperature can be presented with the formulas
(18), (19), and (20) in a form of the universal piecewise linear RMDM.
With this approach, we analyzed the IR and NAC data measured for moist soils, which are
available in the literature, to clear up whether the RMDM can be treated as a physical law
and to estimate the error of this model with regard to the total experimental data set
analyzed.
The IR and NAC experimental data attained for some soils at the temperatures of 20°C to
25°C, are shown in the form of the universal RMDM in Figures. 1 and 2. The soil texture
classes for these soils are presented in Table 1.
Data source.
Order Geographical
Texture class Clay % Sand % MBWF Frequencies
number location
measured
1. F2 17.3 56.0 0.12
(Wang
2. H7 34.7 19.3 0.21
& Scmugge,
3. Harlingen clay 61.0 2.0 0.42
1980)
4. Vernon clay loam Mississippi 28.0 16.0 0.198
1.4; 5.0
State USA
GHz
5. Miller clay Texas, TX 62.0 3.0 0.24
6. Loam 8.53 41.96 0.077 (Dobson et
7. Sandy Loam 13.43 51.51 0.112 al., 1985)
8. Silt Loam Kansas State 13.48 30.63 0.1 Hallikainen
9. Silty Clay USA et al., (1985)
47.38 5.02 0.175 1.4 to 18
GHz
10. Toyoura sand (S) 0.3 99.3 0.0

11. Red-yellow soil 50.0 20.9 0.25
(HC)
12. Brown forest soil 30.4 35.3 0.12
(Topsoil) (LiC) Kyushu (Miyamoto
13. Brown forest soil province, 61.8 7.1 0.26 & Chikushi,
(Subsoil) (HC) Japan 2006)
14. Andisol 18.9 61.5 0.24 5.0 GHz
(Miyazaki) (SCL)
15. Heavily loamy 48.2 61.8 0.158
chernozem
(Komarov
16. Loamy 31.28 68.72 0.113
& Mironov,
chernozem
2000;
17. Middle loamy 35.0 65.0 0.07
Rychkova,
chernozem Altai Krai,
1997)
14. Meadow Russia 17.12 82.88 0.05
0.6 to 1.43
chernozem
GHz
18. Chestnut loam 13.0 87.0 0.036
soil
19. Silt Loam 0.08 (Lukin et
Krasnoyarsk
al., 2008) 0.5
Krai
to 10 GHz
20. Shrub tundra soil 0,16 (Savin et al.,
Alaska State
2007) 1.0 to
USA
16 GHz
Table 1. Soil texture classes measured.
Fig. 1. Measured RI and NAC presented as values reduced with equations (18), (19) versus
soil moistures reduced with formula (20). The IRs and NACs assigned for the universal
piecewise linear RMDM are as follows: Wt  0.3 ; nb  1  10 / 3 ; nu  1  9 , kb  2 / 3 ;
ku  8 / 3 . Figures from (a) to (e) pertain to the texture classes shown in Table 1: a) Soils 6
to 9; b) Soils 1 to 5; c) Soils 15 to 19; d) Soil 20; e) Soils 10 to 14.
Fig. 2. The measured RI and NAC as a function of moisture presented in the form of the
reduced RMDM equations (16)-(20). The data represent all the texture classes shown in
2
Table 1. The correlation coefficients for IR and NAC, Rn and R, are equal to: Rn  0.996 ,
R2  0.966 .
As seen from Figures. 1 and 2, the IR and NAC for the soils listed in Table 1, which
represents a wide ensemble of textures and geographical locations, proved to be well
correlated with those given by the RMDM. Taking into account a broad variety of measured
soil types and frequencies, the RMDM can qualify to be treated as a general physical law,
which describes, according to (1), wave propagation velocity, v=c/n, and amplitude
attenuation coefficient, k0, dependence on soil moisture. At this moment, it is worth
noticing that, in regression dielectric models of (Wang&Schmugge, 1980) and (Dobson et al.,
1985), the index of power, , in the mixing formula (7) was defined as  =1 and =0.65,
respectively, which are other then the value of  =0.5 substantiated by the analysis
conducted in Figure 1 and 2.
Given the RMDM parameters are obtained in some frequency range through fitting (12) and
(13) to the soil IR and NAC data measured, the IR and NAC frequency spectra bor both the
bound and free soil water become available, as well as those for the DC, ′, and LF, ″, (see
formula (3)), for further consideration as a function of soil texture parameters. Taking into
account the formulas in (3). The same is true with regard to the soil water DCs, ′b and ′u,
and LFs, ″b and ″u, pertaining to the bound and free water types, respectively. From the
analysis conducted, follows that the RMDM can be applied to predict the DC and LF of
moist soisl as a function of moisture, using the equations (12), (13), provided the DCs and
LFs for the dry soil, bound and free soil water have been derived from the moisture
dependences measured for a specific soil, at given temperatures and frequencies.
In spite of the fact that the RMDM proved applicable to the soil types on a broad scale, the
soil dielectric data base developed on the basis of the RMDM would take an excessive
amount of laboratory dielectric measurements, as the domain of wave frequencies, soil
textures, and temperatures is too large for such a project to be ever completed. Hence,
further parameterizations are needed to bring the RMDM into practical usage. In the next
section, based on the Debye dipole relaxation laws, the parameterization of soil water
spectra is considered.
4. Frequency dependent RMDM

To parameterize the soil water dielectric spectra, one has to be sure that these related to both
the bound and free soil water types follow the Debye dipole relaxation law, as is the case for
the water solutions out of the soil (Hasted, 1973). The Debye relaxation formula for the CDC
has (Ulaby, 1986) the following form:
 
 ' p i " p  n*p
2
  p 
 p 0   p

i p
1  i 2f p 2f r
(21)
where ε'p ε"p are DC and LF for water, p is any one of b, u,; εp0 and εp∞ are the low- and
high- frequency limit dielectric constants, τp is the relaxation time, and σp is the ohmic
conductivity, each specific to the different forms of soil moisture. Finally, εr = 8.854 pF/m is
the DC of vacuum.
The DC and LF in (21) as a function of circular frequency :
 p ( )   p  [ 0 p   p ] /[1  ( p ) 2 ] , (22)
 
 p ( )  [ 0 p   p ] /[1  ( p ) 2 ]   p  ( p p /  r ) /  p (23)
have some characteristic features. The DC monotonically decreases with the frequency,
having an inflection at the frequency of fp=1/p, which can be derived from the equation
d2′p()/d2=0. At this frequency, the DC undergoes maximum frequency dispersion. While
the LF given by (23) has maximum at  = r, which is known as the frequency of relaxation
and minimum at  = n. This minimum originates due to decrease and increase with
frequency decreasing of the terms ([0p - ∞p]/[1+(p)2])p and (pp/r)/p, respectively.
Taking into account that these terms represent the relaxation and ohmic losses, respectively,
the frequency n can be treated as the frequency of transition between the ranges in which
the soil water dielectric mainly has losses of relaxation or ohmic origin, respectively. Hence
the frequency intervals 0<<pn and pn<<∞ can be identified as ohmic and relaxation loss
ranges.
The relaxation and transition frequencies can be obtained as solutions of the equation
d(″p())d=0:
 
( rp p ) 2   1  2s p  (1  2s p ) 2  4 s p (1  s p )  2  (1  s p ) ,
 
(24)
2

2


( np p )   1  2s p  (1  2s p )  4 s p (1  s p )  2  (1  s p ) 
here s p   p ( 0 p   p ) ,  p   p p  r , is a combined parameter to characterize the
relationship between the ohmic conductivity, p, relaxation time, p, and the range of DC
dispersion, 0p-∞p, for both the bound, p=b, and free, p=u, soil water types. The equations
(24) have real values provided the combined parameter, s, lies in the range 0≤s≤(1/8). Hence,
in the case of DC and LF spectra for soil water, which dielectric parameter s exceeds the value
of 1/8, neither the relaxation nor the transition frequencies are observed, with the LF spectra
being a function monotonically increasing with decreasing frequency.
From the equations of (24), the ranges for the relaxation and transition frequency variations as
a function of the combined parameter s, can be determined as follows:
1 3 p   rp  1 /  ; ; 0   np  1 3 p . (25)
The least and largest limits of these ranges are calculated from (24), with the values of
combined characteristic being equal to s=1/8 (at this value, the least limit for relaxation loss
range occur and largest limit for ohmic loss range ) and to s=0 (at this value, least limit for the
ohmic loss range occur and largest limit for the relaxation range). In addition, according to
(24), the relaxation and transition frequencies are related to each other by the equations given
below in two alternative forms:
np  (1/  p ) 1  (rp p ) 2  1  3(rp p ) 2  and
rp  (1/  p ) 1  (np p ) 2  1  3(np p ) 2  or

(26)
  
np p
2

1  (rp p )   rp p 
2 2

1  (np p )   1 / 2
2
The Debye equations (22), (23) can be recast as a pair of equations which are linear in the
Debye parameters. These equations can be then solved via linear regression to yield least
squares estimates of the Debye parameters. The first of the two linear equations is
 ' p  f   z p   p z p  f  (27)
where
z p   p 0   p p /  0 (28)
is a constant and
z p  f   2f " p  f  (29)
is a reduced variable, which is supposed to be calculated from dielectric measurements. A

linear regression with regard to the measured pairs (ε'p, zpτ) of (27) yield the estimates of the
parameters τp and zpσ from the measured pairs (ε'p, ε''p).
The second equation is
 ' p  f    p 0   p 0   p z p  f  (30)
where the variable
1  " p  f  / 2f p
 1 (31)
z p  f  z p   ' p  f 
is thought to be known from measurements, with the values of τp and zpσ having been obtained
from the previous regression with the use of (27). At this point, all four Debye parameters can
be found using the values of τp, zpσ, εp0, and (εp0- εp∞) attained from the regressions. The linear
fitting of the data measured for ′(f) and ″(f) with the use of (27) and (30) also provide
correlation coefficients and standard deviations to estimate compliance between the measured
data and the fitting ones calculated with the Debye equations (22), (23).
To universally represent various dielectric spectra of soil water measured for bound or free soil
water, different types of soils and temperatures, The linear functions (27) and (30) representing
the Debye equation (22), (23) can be can be modified in a universal linear function:
 ( f )  1  z pj ( f )
 pj (31)
where  pj ( f ) and z pj ( f ) are the reduced DCs and frequencies which can be obtained after
some algebraic manipulations performed with the pairs (27), (29) and (30), (31):
 
z p1 ( )   p p ( )   0 p   p   0 p   p 
(32)
 
 p1 ( )   p z p1 ( )    p  0 p   p 
 
z p 2 ( )   0 p   p   p ( )  0 p   p  ( p ( )  p )   p ( ) 
(33)
   
 p 2 ( )   p z p 2 ( )   p  0 p   p 
The values j =1 and j=2 in (32), (33) correspond to the pairs (26), (29) and (30), (31), respectively.
As a result, the equation (30) in conjunction with the expressions (32), (33) can universally
represent any DC, ′p(f), and LF,″p(f), spectra for both the bound, p=b, and free, p=u, soil water
measured for a specific type of soil, at a given temperature, provided these spectra follow the
Debye equation (21). From equations (22), (23), (32), (33) follow that the reduced functions
 p1 ( f ) ,  p2 ( f ) and their arguments z p1 ( f ) , z p2 ( f ) are defined across the same segment
from zero to unit, [0,1].
Substituting the values of p() and p() given by the formulas (22) and (23) in the expressions
for the reduced frequencies z p1 ( f ) and z p2 ( f ) of (31), (32), we come up with the following
simple formula:

z p1 ( f )  z p 2 ( f )  z p ( f )  ( p ) 2 1  ( p ) 2  (34)
which can be applied only in the case of formal transforming the equations (27), (28) and
(30), (31) to the equations (31), (32), and (33), but not for fitting the equations (31), (32), and
(33) to the DC and LF data measured. In addition, using the expressions (25) and (34) and
the eqution f=1, we find for the reduced maximal dispersion frequency, z p ( f ) , and the
ranges for relaxation, z p ( r ) , and ohmic, z p ( n ) , frequencies to be as follows:
z p ( f )  1 / 2 ; 0  z p ( n )  1 / 4 ; 1 / 4  z p ( r )  1 / 2 ; (35)
When presenting the data measured in the form of reduced frequencies, one can easily identify
using the formula (35) specific frequency ranges which these data belong to. These ranges appear
to be dimensionless due to the respective transformation conducted with (31)-(33).
An example for the DC and LF spectra following the Debye formulas (22), (23) are shown in
Figure 3. They were calculated for a sample of saline water, having solution concentration and
temperature of 1.88 g/liter and 20ºC, respectively, with the use of the empirical formulas
obtained in (Stogryn, 1971). In Figure 3b, the same spectra for the DC and LF reduced according
to (32), (33) are plotted as a function of reduced frequency, z p ( f ) , calculated with (34). The
following parameters of the Debye spectra (22), (23) were used (τu=9.27 ps ; ε0u=79.44; σu=0.3;
ε∞u=4,9), regarding the liquid saline water, to perform transition to reduced spectrum and
frequency, as given by (32), (33). The spectroscopic parameters for saline water listed above were
determined with empirical formulas of (Stogryn, 1971). Here and henceforth, we understand the
low and high frequency limits as the values related exclusively to a specific Debye relaxation
frequency range, within which the measured data are located and the low and high frequency
limits have been determined using these data. These limits can not be understood as the values of
real DCs which may occur beyond the frequencies measured. Such a situation may take place if
some other relaxation processes considerably contribute in the values of real DCs, making, for
instance due to the Maxwell-Wagner relaxation (Hasted, 1973), real values of ε′ a lot greater then
the low frequency limit, ε0, determined via data measured for its deriving. In the range of
frequencies exceeding those measured, the same understanding with regard to the high
frequency limit, ε∞ , has to be used, as many other relaxation processes may also occur when
moving towards the optical range of frequencies.
The vertical lines in Figures. 3a and 3b mark the maximal dispersion frequency, ff, maximal
relaxation loss frequency, fr, and the frequency of transition from relaxation to ohmic losses, fn , as
these are given with the formulas (24) and (32), (33), to display the intermediate range of
frequencies in which both the relaxation and ohmic losses are observed, as well as to show the
position of maximal dispersion frequency.
As any Debye spectra can be universally presented like in Figure 3, for the spectra measured a
possibility arises to align those along one and the same straight line, which makes possible not
only to test whether the measured spectra belong to the Debye class, but also to monitor the
position of measured spectra with regard to the Debye spectra characteristic frequencies, that is,
ff, fr, and fn, pictured for all the data measured in one and the same graph.
Fig. 3. Relaxation frequency spectra for the DC and LF in the case of saline liquid water,
having the following values of Debye parameters: τ=9.27 ps ε0=79.44 σ=0.3 ε∞=4,9. a) Initial
values of DC, LF and frequency; b) Reduced values of DC, LF, and frequency. With vertical
lines, are marked both the real (in Fig. 3a) and reduced (in Fig. 3b) positions of the maximal
dispersion frequency (ff, and z f , solid lines), maximal relaxation loss frequency, (fr and z r ,
dotted lines), and the frequency of transition to ohmic losses (fn and z n , dotted lines).
While their numerical values are placed next to the respective lines.
Fitting formulas (12) and (13) to the data for DCs and LFs as a function of moisture for the soils
listed in Table 1, first, the values of DCs and LF were obtained with the technique outlined in
Section I. Second, the DC and LF spectra for both the bound and free soil water obtained were
fitted with the formulas (27)-(31) in order to derive the parameters present in the Debye equation
(22), (23), that is the low and high frequency limits,  and , relaxation time, , and conductivity,
σ, regarding both types of soil water. The values thus obtained are given in Table 2.
Soil number in Table 1 7 9 19 20

Wave frequencies measured 4 to 18 GHz 4 to 18 GHz 0.5 to 10 GHz 3.01 to 15 GHz
Soil water type p=b p=u p=b p=u p=b p=u p=b p=u
Low frequency limit of DC, 0p 58.5 103 38.60 97.90 65.08 73.15 43.45 79.52
High frequency limit of DC, ∞p 4.9 4.9 4.9 4.9 4.9 4.9 14.2±1.2 14.1±1.2
Relaxation time, p(ps) 1.14 8.3 11.30 8.00 21.41 9.23 65.71 11.57
Conductivity, pSp/m) 0.7 1.03 1.24 2.15 3.48 1.29 0,061 3,31
Table 2. Debye relaxation parameters and conductivities for some soils listed in Table 1.
We did not have the dielectric data measured for the frequencies larger then the maximal
dispersion frequency, ff, which, according to Figure 3, could make possible to derive the high
frequency limit with reliable accuracy, except for the soil 20. Therefore, the value of 4.9 was
preset before fitting with the formulas (27)-(31), in accordance with previously estimated data for
this parameter (Mironov et al., 2004).
As follows from Table 2, the Debye relaxation parameters and conductivities relating to the
bound and free water can be clearly distinguished via their magnitudes. The most important
feature, seen in Table 2, is that not only the low frequency limits but also the relaxation times, as
well as conductivities are found to noticeably differ from each other. It implies that, in accordance
with the analysis conducted above the maximum relaxation frequencies and transition to ohmic
losses frequencies pertaining to the different types of soil water must be also different.
At the same time, in the dielectric model developed in (Dobson et al., 1985), both the bound and
free water spectra were assumed to be similar to that of free liquid water existing out of soil. The
mixing formula for the moist soil CDC used in (Dobson et al., 1985) can be presented as follows:
 s ( f , va , vm , vw , C , S )  va ( a )  vm m (C , S )  vw ( w ( f )  vw (  1 (C , S )  1)  1)
 s( f , v w , C , S )  v w  2 (C ,S )
  w ( f ,  (  d , C , S ))
were subscrip w stands for the water out of soil content and w(f) is a preset function
characterizing the water out of soil. At that, only the power index a and the values 1(C, S),
2(C,S) and (d,C,S) were adjusted to the dielectric data measured via regression analysis. While
the mixing formula applied in (Wang&Schmugge, 1980) suggested the following expression for
the CDC of moist soil:
 s ( f , va , vm , vw , C , S )  va a  vm m (C , S )  vw  Y   w ( f ) if vw  vt ;
 s ( f , v a , v m , v w , C , S )  v a  a  v m  m (C , S )  v t  Y   w ( f )
if v w  v t ;
 (v w  v t ) w ( f )
with the values of vt and Y being adjusted to the dielectric data measured only at 1.4 and 5 GHz.
From the expressions above pertaining to the dielectric models of (Dobson et al.,1985) and
(Wang&Schmugge, 1980) follows that the intrinsic properties of soil bound water spectra in
terms of all the Debye relaxation parameters and conductivity could neither been measured nor
taken into account on a physical ground.
The detailed consideration of both the bound and free soil water spectra measured for the soils
listed in Table 2 is conducted in Figure 4 with regard to reduced spectra.
Fig. 4. Reduced spectra (slanting dotted line) for the bound and free water pertaining to
some soils listed in Table 1. Position of the maximal dispersion frequency z fu  z fu  0.5
is marked with vertical soilid lines. While the positions of maximal relaxation loss
frequency, z rp , and the frequency of transition to ohmic losses range, z np , for the BSW,
p=b, and FSW, p=u, are shown with vertical dotted and solid lines, respectively. a) Soil 7,
z rb  0.498 , z nb  0.002 , z ru  0.49 , znu  0.01 ; b) Soil 9, z rb  0.448 , z nb  0.05 ,
z ru  0.478 , z nu  0.02 ; c) Soil 19, z ru  0.477 , z nu  0.02 ; d) Soil 20, z ru  0.421 ,
z nu  0.08 .
In contrast to the ideal reduced spectra shown in Figure 3b, some of measured data given in
Figure 4 , for reduced DC,  ( f ) and frequency, z ( f ) , fall beyond the range 0≤  ( f ) , z ( f )
≤ 1. This occurred not only due to error in the spectroscopic parameters derived, which are
taken from Table 2 to be used for calculating the reduced value with the formulas (32), but also
because the frequencies z1 ( f ) and z2 ( f ) were calculated using in formulas (27)-(31) the data
for  ( f ) measured with some error. As seen from Figure 4, for both the bound and free soil
water, the measured spectra for the soils 7, 9,and 19 appeared to be mainly located in the
frequency range from the transition to ohmic losses frequency, fn, to the relaxation loss
frequency, fr, with only a few frequencies exceeding the maximal dispersion frequency, ff . For
this group of soils, there is only a little difference between the characteristic frequencies related
to the bound or free soil water types. Though, as to the soil 19, the intermediate frequency
range, from fn to fr, did not appeared to be observed at all, because , in accordance with the
formula (24) and data given in Table 2, the condition 0≤s≤(1/8) was not met. Finally, we
found the spectral properties of the bound and free soil water contained in the arctic shrub
tundra soil to be quite different in terms of their frequency ranges As seen from Figure 4 d, the
frequencies measured appeared to be located below and higher the respective maximal
dispersion frequencies of the bound and free soil water types, respectively. In addition, there
was no interim frequency interval for the bound water spectrum observed. Once in the case of
this soil, the frequencies measured happened to be higher then the maximal relaxation one, the
high frequency limit was measured with an acceptable error. The numerical value of this
parameter is given in Table 2. As seen from Table 2, for the organic rich (95%) soil 20, the high
frequency limits appeared to be about three times as large compared to the value of 4.9, which
is a conventional estimate for the mineral soils (Mironov et al., 2004).
Similar to Figure 4, all the spectra measured for the soils 7, 9, 19, and 20, are presented in the
reduced form in Figure 5.
Fig. 5. Reduced spectra for both the bound (filled symbols) and free (empty symbols) soil
water regarding the soils 7, 9, 19, and 20. The linear fits related to the bound and free water
are drawn with dotted and solid lines, respectively. The correlation coefficient and standard
deviation regarding the linear fit are equal to R= 0.988 and SD= 0.0471, respectively.
As seen from Figure 5, the data measured follow quite well the Debye relaxation formula (21)
for both the bound and free soil water spectra regarding rather a broad variety of soils and
frequencies measured.
The results of the analysis conducted in this Section not only substantiates the use of Debye
spectra for the soil water dielectric spectra, but also propose the methodology for deriving soil
water Debye parameters and conductivity. Previously, the Debye spectrum model was used in
the soil dielectric models (Wang&Schmugge, 1980; Dobson et al., 1985; Mironov et al., 2004),
though the applicability of this spectrum class with regard to the soil water types, especially in
the case of bound water, has never been substantiated. Summing up the result of Sections 2
and 3, let us summarize the FD RMDM parameters that have to be derived from dielectric
measurements to ensure the DC and LF predictions for a specific type of moist soil. As can be
seen from the equations (3), (4), (12), (13), (22), (23), a certain type of moist soil, in terms of its
dielectric spectra, can be completely determined via a set of the following FD RMDM
parameters:
1 DC, ε′d(T,C,S), for dry soil;
2 LF, ε″d(T,C,S), for dry soil;
3 Value of MBWF, Wt(T,C,S);
4 Low frequency limit dielectric constants, ε0b(T,C,S) and ε0u(T,C,S), for bound and free
soil water;
5 High frequency limit dielectric constants, ε∞b(T,C,S) and ε∞u(T,C,S), for bound and
free soil water;
6 Relaxation times, τb(T,C,S) and τu(T,C,S), for bound and free soil water;
7 Conductivities, σb(T,C,S) and σu(T,C,S), for bound and free soil water.
As shown in Sections 2 and 3, for a specific type of soil, all of these parameters can be
derived with the use of conventional dielectric measurements of moist soils. Therefore, to be
employed in the microwave remote sensing processing algorithms, the FD RMDM requires
prior dielectric measurements to be carried out for a set of the individual soils involved in
remote sensing data processing, and the error of dielectric predictions for each individual
soil be tested. Some tests to prove the predictive power of the FD RMDM have been already
conducted in (Mironov et al., 2004). Further, we use the dielectric data of (Hallikainen et al.,
1985; Dobson et al., 1985, Curtis et al. 1995) to analyze prediction errors of the FD RMDM
over a wide ensemble of soil types in terms of texture, as well as extended ranges of
frequencies measured. Based on this analysis, the FD RMDM can be generalized to predict
DCs and LFs of moist soils, with their textures being an input parameter. That frequency
and texture dependent refractive mixing dielectric model (FTD RMDM) is outlined in the
next section, following the results of (Mironov et al., 2009a).
5. Frequency and texture dependent RMDM

The FD RMDM was tested in (Mironov et al., 2004 a,b; Mironov et al., 2009a) with the
dielectric data measured in (Hallikainen et al., 1985; Dobson et al, 1985; Curtis et al., 1995).
Given the numerical values for the FD RMDM parameters (see Table 2), the ensemble of the
FD RMDM formulas (3), (4), (12), (13), (22), (23) can be applied to calculate the DCs and LFs
predicted with this model. Texture characteristics of the combined set of soils measured in
(Hallikainen et al., 1985; Dobson et al, 1985; Curtis et al., 1995) are given in Table 3.
The ensemble of dielectric data in (Curtis et al., 1995) appeared to be incomplete, in terms of
a number of moistures measured, to be applied for deriving the FD RMDM spectroscopic
parameters as it was done in Sections 2 and 3. Based on the results of these sections, we
suggested the FD RMDM to be applicable to the dielectric data of (Curtis et al., 1995)
measured in the domain of moistures (from nearly dry soils to field capacity values), and
frequencies (from 0.45 to 26.5 GHz).
Order Sample Clay Sand

Texture class MBWF
number code % %
1. B Sand (SP), Light Gray 0 98 0.005
2. L Sand (SP), Brown 0 99 0.001
3. I Sand (SP), White 0 100 0.001
4. C Silty Sand (SM), Brown 4 88 0.02
5. E Clayey Silt (ML), Brown 7 0 0.1
6. K Clayey Silt (ML), Brown; Trace of Sand 7 4 0.11
7. F2 Loam 8.53 41.96 0.075
8. G Clayey Sand (SC), Dark Brown 13 55 0.07
9. F1 Sandy Loam 13.43 51.51 0.07
10. F3 Silt Loam 13.48 30.63 0.073
11. D Silty Sand, (SM), Reddish Brown 14 77 0.05
12. H Clay (CH), Gray 34 2 0.15
13. F5 Silty Clay 47.4 5.02 0.17
14. J Silt (ML), White 54 0 0.15
15. A Clay (CH), Light Gray 76 2 0.28
Table 3. Texture characteristics for the soils measured in (Hallikainen et al., 1985; Dobson et
al, 1985; Curtis et al., 1995).
Therefore, there was applied a procedure of fitting simultaneously the FD RMDM formulas
to all the DC and LF spectra corresponding to the whole set of moistures available for each
particular type of soil presented in Table 3, as was done in (Mironov et al., 2006; Mironov et
al. 2009a). As an example of fitting effectiveness, in the case of soil with code D, the fits and
the experimental spectra are shown in Figure 6. The results presented in this figure proves
the FD RMDM ability to accurately fit the measured CDs and LFs.
Fig. 6. DC, ε΄, and LF, ε, spectra for soil D measured (dots) and fitted (solid line) with the
use of the GRMDM model.. Shown data correspond to the following soil volumetric
moistures, W(%): 1) 3.2; 2) 8; 3) 8.8; 4) 13.2; 5) 18.4; 6) 29.1; 7) 29.7; 8) 38.2; 9) 39.4
With this technique of fitting, the FD RMDM spectroscopic parameters were attained for the
whole set of soils listed in Table 3. Thus obtained, the FD RMDM parameters are shown
with symbols in Figure 7 as functions of clay content.
Fig. 7. The GRMDM spectroscopic parameters for the soils presented in Table 1 as a
function of gravimetric clay content. The squares represent values derived through fitting
the dielectric spectra. The numbers shown in squares correspond to those given in Table 3.
To estimate the error of the FD RMDM predictions for the DCs and LFs over this ensemble
of soils, there were applied equations (3), (4), (12), (13), (22), (23) in conjunction with the FD
RMDM spectroscopic parameters given in Figure 7. The measured DCs and LFs versus the
predicted ones are demonstrated in Figure 8, with the error of prediction being given in the
figure caption. According to Figure 8, the error in terms of standard deviation, correlation
coefficient and squint of the linear regression with regard to 1 to 1 line appeared to be quit
acceptable, given a wide domain of frequencies, soil textures, and moistures.
Fig.8. Correlation of the FD RMDM predictions, ε΄p, εp, for DCs (a) and LFs (b) with the
measured ones, ε΄m, εm, in the case of soils measured in (Curtis et al., 1995). Solid and
dotted lines represent bisectors and linear fits, respectively. Correlation coefficients, RDC
and RLF, and standard deviations, SDDC and SDLF, are equal to: RDC=0. 995, RLF=0.991,
SDDC=1.023, SDLF=0.4899.The linear fits are expressed as follows: ε΄m=-0.2753 + 1.013 ε΄p,
εm=-0.0943 + 1.058 εp.
Generalizing the above analysis, it is possible to assert that the FD RMDM, having been
adjusted to each specific soil, as in Figure 6, can ensure acceptable accuracy of dielectric
predictions. This result makes possible to use the FD RMDM as a building block to create a
frequency and texture dependable dielectric model (FTD RMDM) for moist soils. For this
purpose, the whole set of FD RMDM parameters in Figure 7 was fitted with the functions of
clay percentage (Mironov et al., 2009a), yielding the equations given in Table 4.
nd=1.634-0.539·10-2C+0.2748·10-4C2 b =79.8-85.4C+32.7·10-4C2 u =100

kd=0.03952-0.04038·10-2C b =1.062·10-11+3.450·10-12·10-2C u =8.5·10-12
Wt=0.02863+0.30673·10-2C b =0.3112+0.467·10-2C u =0.3631+1.217·10-2C
Table 4. Fits for the FD RMDM as a function of clay percentage
The clay content, C, and relaxation times, τb and τu, in the formulas of Table 4 are expressed
in percentages and seconds, respectively. Further on, the equations in Table 4 are referred to
as the FTD RMDM fits, while the values calculated with these equations will be identified as
the FTD RMDM spectroscopic parameters. Finally, the DCs and LFs calculated with
formulas (3), (4), (12), (13), (22), (23) in conjunction with the spectroscopic parameters given
in Figure 7 are identified as the FTD RMDM dielectric predictions. Clay percentage, C, is the
only input parameter of the FTD RMDM in terms of soil texture to account for a specific
type of soil. It should be especially noticed that the equations in Table 4 generalize all the
information presented in Figure 7 on the FD RMDM spectroscopic parameters regarding the
15 soils of Table 3.
The same correlation analysis, as shown in Figure 8 in the case of the FD RMDM, was
conducted with respect to the FTD RMDM. It proved the FTD RMDM prediction errors to
be on the same order as those given in Figure 8. This result signifies that, over the whole
variety of soil types, moistures, and frequencies measured in (Curtis et al., 1995) and
(Hallikainen et al., 1985; Dobson et al. 1985) the FTD RMDM can predict the DCs and LFs
with the same accuracy, in terms of correlation coefficient and standard deviation, as the FD
RMDM does.
Nevertheless, it is worth mentioning that the FTD RMSD prediction error has been tested
only over the dielectric data, which were used for its developing. In case the test is carried
out regarding the dielectric data with which a respective regression model is created, it can
reveal only an error of regression analysis itself, as was done in (Wang&Schmugge, 1980;
Dobson et al. 1985). To verify an intrinsic predictive power of the FTD RMDM, it has to be
tested over the dielectric data set other than the one used for deriving formulas like in Table
4. To perform such a validation of the FTD RMDM, the formulas in Table 4 were obtained,
first, with the use of spectroscopic parameters relating to all the soils excluding the ones
coded as F1, F2, F3, and F5 and, second, only with the use of soils F1, F2, F3, and F5. As a
result, these two new versions of the FTD RMDM could be treated as independent with
regard to the soil sets which were not used in their developing. Then the tests like that in
Figure 8, were conducted for the both FTD RMDM versions with respect to the DC and LF
data independently measured. With these tests, the errors for both FTD RMDM versions was
found (Mironov et al., 2009a) to be on the same order as the ones seen in Figure 8. In addition,
we have tested the FTD RMDM against the independently measured dielectric data, which are
available in (Wang&Schmugge, 1980; Sabburg et al. 1997; Mironov et al., 2004; 2005; 2006),
finding the error of the FTD RMDM predictions to be on the same order as that in Figure 8 is.
As seen in Figure 9 borrowed from (Mironov et al., 2009a], in contrast to the physically
based FTD RMDM, the regression model of (Dobson et al., 1985) provided for DC and LF
predictions with considerably greater error than that observed in Figure 8, pertaining to the
FTD RMDM.
Fig.9. Correlation of the semiempirical model (Dobson et al. 1985) predictions, ε΄p, εp, for
DCs (a) and LFs (b) with the measured ones, ε΄m, εm, in the case of soils measured in
(Curtis et al.,1995). Solid and dotted lines represent bisectors and linear fits, respectively.
Correlation coefficients, RDC and RLF, and standard deviations, SDDC and SDLF, are equal to:
RDC=0.942, RLF=0.882, SDDC=3.391, SDLF=1.695. The linear fits are expressed as follows:
ε΄m=-0.753 + 0.902ε΄p, εm=1.483 + 0.881εp.
It is worth noting here that the methodology employed for developing the FTD RMDM is
based on two key elements. The first one is the FD RMDM, which ensures accurate dielectric
predictions using a cluster of the FD RMDM, parameters pertaining to the respective cluster
of soil types with specific textures. The second key element is the completeness of such soil
clusters in terms of texture diversity. An example of such a cluster is shown in Figure 7.

Within the FTD RMDM methodology, the cluster of the FTD RMDM parameters pertaining
to a specific ensemble of soils is turned into a specific collection of the regression equations
given in Table 4, which represent the example of a specific version of the FTD RMDM
related to the ensemble of soils listed in Table 4. The more complete ensemble of soils is
available the less error of prediction with the respective version of the FTD RMDM can be
attained, especially in regard with the soils other than those used to develop the particular
version of the FTD RMDM. Therefore, the FTD RMDM prediction error, regarding the
whole variety of natural soils, must depend on both the predictive capability of the FD
RMDM, which is a physically based building block of the FTD RMDM, and the
completeness of the ensemble of soils employed, in terms of diversity of their texture, which
is a regression analysis building block of the respective version of the FTD RMDM. The
more complete such a data set is the less regression analysis error in the equations similar to
those given in Table 4 can be obtained, thus ensuring less error of predictions related to the
respective FTD RMDM version.
One more problem regarding the FTD RMDM prediction error needs to be discussed. It
concerns the question whether the FTD RMDM trained on a specific variety of soils is able to
provide for dielectric spectra (22), (23) to the data measured predictions in the case of an
individual soil, which belong to that variety. Let us consider such a situation. In Figure 10,
experimental DCs and LFs are plotted versus the FTD RMDM predictions in the case of
specific soil D measured in (Curtis et al., 1995). At that, in Figure 10, there was used the
version of FTD RMDM based on the formulas from Table 4, which correspond to the whole
variety of soils listed in Table 1. The data predicted with the FTD RMDM instead of the
initial FD RMDM parameters as shown in Figure 6, appeared to follow the measured ones
quite well for DCs and good enough in the most cases for LFs, excluding the data for greater
moistures (graphs 7 and 9) in the range of frequencies below 2 GHz. As seen in Figure 7, this
deviation arises from a noticeable difference between the free water conductivities derived
with the FD RMDM fitting technique and calculated with the FTD RMDM formulas of
Table 4.
Fig. 10. DC, ε΄, and LF, ε, spectra for soil D measured (dots) and calculated with the use of
FTD RMDM (solid lines). Shown data correspond to the following soil volumetric
moistures, W(%): 1) 3.2; 2) 8; 3) 8.8; 4) 13.2; 5) 18.4; 6) 29.1; 7) 29.7; 8) 38.2; 9) 39.4
As seen from Figures. 6 and 10, the error of prediction in the case of an individual soil D,
with the FTD RMDM trained on a composite variety of soils listed in Table 3 is on the same
order as the one obtained with the use of the FD RMDM (Figure 6) is, the latter being trained
specifically on dielectric data from (Curtis et al., 1995) pertaining to soil D. As a result, it can
be stated that the physically based FTD RMDM has demonstrated noticeably less prediction
error, compared to the regression dielectric model of (Dobson et al., 1985), in the case of both
a group of soils or an individual soil. falling out of the varieties of soils used to develop
respective models.
6. Conclusions
Summing up the results, the following has to be stated as primary findings of this research.
First, the physical law describing both the index of refraction and attenuation coefficient as a
function of moisture content has been substantiated. This law has a form of the well known
refractive mixing dielectric model, with the index of power being equal to 0.5, regarding the
complex dielectric constants of the components mixed.
Second, the bound and free soil water dielectric spectra were shown to follow the Debye
relaxation law. The technique to derive the high and low frequency limits of dielectric
constant, relaxation time, and conductivity for both type of soil water has been proposed
and tested.
Third, there was confirmed ability for the frequency dependent refractive mixing dielectric
model recently developed in (Mironov et al., 2004) to generate good dielectric predictions
over the ensemble consisting of 15 natural soils having the textures covering almost all
natural soils. These predictions proved to be accurate, concerning the frequency range from
0.3 to 26.5 GHz, volumetric moistures spanning from nearly dry condition to field capacity
moistures, with the temperatures being of from 20 to 22○C. This model has error about three
times as small with regard to model which is currently employed in practice of the
geoscience and remote sensing.
Forth, the frequency dependent refractive mixing dielectric model was generalized. As a
result the frequency and texture dependent refractive mixing dielectric model for moist
soils, which proved to provide for complex dielectric constant spectra predictions with error
on the same order as the frequency dependent refractive mixing dielectric model does. At
that, the soil clay percentage is the only input parameter in terms of soil texture for the
frequency and texture dependent refractive mixing dielectric model.
7. Future work
The results discussed in this chapter are applicable only in the narrow range of
temperatures from 20 to 22º, in which the dielectric data were attained (Dobson et al. 1995;
Curtis et al., 1995). To extend these results over a larger range of temperatures, a
temperature and frequency dependent refractive mixing dielectric model must be
developed, as in (Mironov et al., 2009a), to become a building block for a temperature and
texture dependent refractive mixing dielectric mode, the first approach to which has already
been made in (Mironov&Fomin, 2009). There is also a lot of work to be done regarding the
tests of the dielectric models developed over the existing dielectric data and radar and radio
thermal remote sensing data available in the literature. This must help define the domains of
their applicability in the practice of geoscience and remotes sensing.
8. Acknowledgements
This work is supported by the Siberian Branch Russian Academy of Sciences under project
No. 2.5.1.1: “Dielectric spectroscopy of natural media in the radio frequencies band.” (2007-
2011). The authors would also like to express gratitude to their colleagues Lyudmila
Kosolapova, Sergej Fomin, Yurij Lukin, and Igor Savin for processing a great many of
dielectric data and preparing some figures and tables, when developing this paper for
publication. As well as the Russian Foundation for Basic Research under project No. 09-05-
91061: “Development of a soil dielectric constant model.” (2009-2011).
9. References
Birchak, J. R.; Gardner, C. G.; Hipp, J. E. & Victor, J. M. (1974). High dielectric constant
microwave probes for sensing soil moisture, Proc. IEEE, Vol. 62, No. 1, January
1974 pp. 92-98.
Curtis, J.O.; Weiss, C.A. Jr. & Everett, J.B. (1995). Effect of Soil Composition on Dielectric
Properties, Technical Report EL-95-34, U.S. Army Corps of Engineers, Waterways
Experimental Staton, Vicksburg, MS, December 1995.
Dobson, M. C.; Ulaby, F. T.; Hallikainen, M. T. & El-Rayes, M. A. (1985). Microwave
dielectric behavior of wet soil: Part II – Dielectric mixing models, IEEE Trans.
Geosci. Remote Sensing, Vol. 23, No. 1, January 1985 pp. 35-44.
Hasted, J. B. (1973). Aqueous Dielectrics, Chapman and Hall, London.
Hallikainen, M. T.; Ulaby, F. T.; Dobson, M. C.; El-Rayes, M. A. & Lin-Kun-Wu. (1985).
Microwave dielectric behavior of wet soil: Part I – Empirical model and
experimental observations, IEEE Trans. Geosci. Remote Sensing, Vol. 23, No. 1,
January 1985 pp. 25-34.
Komarov, S. A. & Mironov, V. L. (2000). Microwave Remote Sensing of Soils, Publishing
House of the Siberian Branch of the Russian Academy of Sciences, Novosibirsk,
Russia.
Lukin, Yu. I.; Mironov, V.L. & Komarov, S.A. (2008). Investigation of dielectric spectra from
moist soils during freezing-thawing processes. Russian Physics Journal, Vol. 51,
No. 9, September 2008 pp. 907-911, ISSN 1064-8887.
Mironov, V. L.; Komarov, S .A.; Rychkova, N. V. & Kleshchenko, V. N. (1995). Study of the
Dielectric Properties of Wet Grounds at Microwave Frequencies, Earth Obs. Rem.
Sens., Vol. 12, No. 4, April 1995 pp. 495-504.
Mironov V. L.; Dobson, M. C.; Kaupp, V. H.; Komarov, S. A. & Kleshchenko, V. N. (2002).
Generalized refractive mixing dielectric model for moist soils, Proceedings of
IGARSS’02, Vol.VI, pp. 3556-3558, Toronto, Canada.
Mironov, V.L. & Bobrov, P.P. (2003). Soil Dielectric Spectroscopic Parameters Dependence
on Humus Content, Proceedings of IGARSS’03, Vol. II, pp. 1106-1108, Toulouse,
France.
Mironov, V.L. (2004). Spectral Dielectric Properties of Moist Soils in The Microwave Band,
Proceedings of IGARSS’04, Vol. V, pp. 3474 – 3477, Anchorage, USA.
Mironov, V.L.; Dobson, M. C.; Kaupp, V. H.; Komarov, S. A. & Kleshchenko, V. N. (2004).
Generalized refractive mixing dielectric model for moist soils, IEEE Trans. Geosci.
Remote Sensing, Vol.42, No.4, April 2004 pp. 773–785.
Mironov V.L.; Bobrov, P.P.; Bobrov, A.P.; Mandrygina V.N. & Stasuk V.D. (2005).
Microwave Dielectric Spectroscopy of Moist Soils for a Forest-tundra Region,
Proceedings of IGARSS’05, Vol. V, pp. 4485-4488, Seoul, Korea.
Mironov V.L.; Bobrov, P.P.; Kosolapova, L.G.; Mandrygina, V.N. & Fomin S.V. (2006). Data
Processing Technique for Deriving Soil Water Spectroscopic Parameters in
Microwave Range, Proceedings of IGARSS’06, Vol. VI, pp. 2957-2961, Denver, USA.
Mironov, V.L.; Savin, S.V. & Roo R.D. (2007). Dielectric Spectroscopic Model for Tussock
and Shrub Tundra Soils, Proceedings of IGARSS’07, July 2007 pp. 726–731,
Barcelona, Spain.
Mironov, V.L.; Kosolapova, L.G. & Fomin, S.V. (2009a). Physically and mineralogically
based spectroscopic dielectric model for moist soils, IEEE Trans. Geosci. Remote
Sensing, Vol.47, No.7, Part 1, July 2009, pp.2059 - 2070.
Mironov V.L. & Fomin S.V. (2009b). Temperature Dependable Microwave Dielectric Model
for Moist Soils, Proceedings of PIERS 2009, March 23-27 2009 pp. 831-835, Beijing,
CHINA.
Miyamoto, T. & Chikushi, J. (2006). Time domain reflectometry calibration for typical
upland soils in Kyushu (Japan), JARQ, Vol. 40, No. 3, 2006 pp.225-231,
http://www.jircas.affrc.go.jp.
Peplinski, N. R.; Ulaby, F. T. & Dobson, M. C. (1995a). Dielectric properties of soils in the 0.3-
1.3-GHz range, IEEE Trans. Geosci. Remote Sensing, Vol. 33, No. 3, March 1995 pp.
803-807.
Peplinski, N.A.; Ulaby, F.T. & Dobson, M.C. (1995b). Correction to Dielectric properties of
soils in the 0.3-1.3 GHz range, IEEE Trans. Geosci. Remote Sensing, Vol. 33, No. 6,
June 1995 pp.1340.
Rychkova, N. (1997). Investigation of Radiophysical Properties of Soil with regard to Remote
Sensing Problems, Ph.D. Thesis, Alatai State University, Russia.
Sabburg, J.; Ball, J. A. R. & Hancock, N. H. (1997). Dielectric behavior of moist swelling clay
soils at microwave frequencies. IEEE Trans. Geosci. Remote Sensing, Vol. 35, No. 1,
January 1997 pp. 784-787.
Stogryn, A. (1971) Equations for Calculating the Dielectric Constant of Saline Water IEEE
Transactions on Microwave Theory and Techniques, Volume 19, Issue 8, Aug 1971
Page(s):733 - 736
Ulaby, F. T.; Moor, R. K. & Fung A. K. (1986). Microwave Remote Sensing, Active and
Passive, vol. III, Artech House, Dedham, MA.
Wang, J. & Shmugge, T. (1980). An empirical model for the complex dielectric permittivity
of soils as a function of water content. IEEE Trans. Geosci. Remote Sensing, Vol. 18,
No. 4, April 1980 pp. 288-295.
Automated Detection of Clouds in Satellite Imagery 303
14
X
Automated Detection of Clouds

in Satellite Imagery
Gary Jedlovec
NASA Marshall Space Flight Center
USA
1. Introduction
The detection of clouds in satellite imagery has a number of important applications in
weather and climate studies. The presence of clouds can alter the energy budget of the
Earth-atmosphere system through scattering and absorption of shortwave radiation and the
absorption and re-emission of infrared radiation. The scattering and absorption
characteristics of clouds vary with the microphysical properties of clouds, hence the cloud
type. Thus, detecting the presence of clouds over a region in satellite imagery is important in
order to derive atmospheric (e.g., optical depth, phase, temperature, etc.) that give insight
into weather and climate processes. For many applications however, clouds are a
contaminant whose presence interferes with retrieving atmosphere or surface information.
In these cases, the detection of cloud contaminated pixels in satellite imagery is important to
isolate cloud-free pixels used to retrieve atmospheric thermodynamic information (e.g.,
temperature and moisture information, ozone content, and even trace gas concentrations) or
surface geophysical parameters (e.g., land and sea surface temperature, vegetation
information, etc.) from cloudy ones.
The ability to derive an accurate cloud mask from geostationary and polar orbiting satellite
data under a variety of conditions has been a research topic since the launch of the first
Earth observing satellite TIROS-1 in 1960. The limited success of some early studies
(Coakley and Bretherton 1982; Rossow and Garder 1993; and those discussed by Goodman
and Sellers 1988) suggests that the accurate detection of clouds in satellite imagery both
during the day and at night is a challenging problem. In more recent work, the probability
of detecting clouds has been reported to exceed 90% (Saunders and Kriebel 1988; Merchant
et al. 2005; Jedlovec et al. 2008; Reuter et al. 2009) but the performance varies seasonally,
regionally, with time of day and retrieval technique. While traditionally both spatial and
spectral techniques have been employed to identify cloud contaminated pixels in polar
orbiting and geostationary satellite data, sensor spatial resolution, the lack of surface –
atmospheric boundary layer temperature contrast, and surface emissivity variations all
present performance challenges to a given cloud detection approach. Thus, any one
technique may not be best suited for all applications, but may perform quite well in a
particular environment (usually the environment in which the algorithm was developed
and tested). The key to the success of most of these algorithms lies in the selection of the
thresholds for various spectral tests. In more robust algorithms, spatially and temporally
varying thresholds, which better capture local atmospheric and surface effects, are used to
improve their performance and broaden their application over algorithms with fixed
thresholds for cloud tests.
In this chapter, a review of several multispectral cloud detection techniques is presented.
Emphasis is placed on techniques which use multispectral approaches applicable to a wide
variety of current and future satellite sensors. The detailed methodology used in several
recent and widely used algorithms is highlighted. The performance of two cloud detection
approaches is compared for the same observational conditions.
2. Theoretical approaches
Clouds have a high solar reflectivity at visible wavelengths compared to that of most surface
features as shown in Figure 1. Spatial differentiation and thresholding techniques can be
used with Earth viewing satellites measuring reflected solar energy to distinguish clouds
from less reflective land and ocean surfaces. These methods are applicable during the day
when solar illumination angles are sufficiently large and the reflected sunlight provides
contrast in image features.
Fig. 1. Typical reflectance values for snow (blue), bare soil (black), forest canopy (pink), and
cirrus (red) and stratus (green) clouds as a function of wavelength (micrometers).
Automated detection methods which use spatial analysis techniques to detect the contrast
between reflected energy from clouds and surrounding (less reflective) scenes can be used
to determine the extent of cloud cover over a region. This approach has difficulties under
low solar illumination conditions and when other highly reflective non-cloud surfaces are
present (e.g., snow, ice, sand) or other complicating optical conditions exist (e.g., scattering
due to aerosols). Predetermined energy threshold values based on solar illumination angles
can be used to delineate cloudy from cloud-free pixels in satellite imagery as well. This
approach provides higher resolution cloud information (down to the pixel level) but is
heavily dependent on the determination of the threshold values. Appropriate threshold

values usually change temporally, seasonally, and spatially, making the use of a fixed set of
values limiting. The use of multiple channels in the visible (0.4-0.7 micrometers) and near or
reflective infrared portion of the energy spectrum (0.7 – 2.5 micrometers), can help in the
detection of clouds. Figure 2 presents an example of how reflectance information from
satellite imagery can be used in a compositing technique to differientiate clouds from snow
on the ground. The image on the left is a natural color composite which uses information
from three channels in the visible region of the solar spectrum to replicate what the human
eye would see from space. From the reflectance signatures in Figure 1, one would expect
that snow and clouds would look similar at the visible wavelengths (and in fact this is the
case for the natural color composite). The image on the right is a false color composite which
uses one visible channel and two reflective infrared channels at 1.6 and 2.2 micrometers to
differentiate clouds (white) from snow on the ground (red). Snow has a much lower
reflectance value at the longer wavelengths as indicated in Figure 1. The use of multispectral
measurements from satellite sensors can add significant skill in the detection of clouds from
space.
Fig 2. Multispectral natural (left) and false (right) color composite images derived from
MODIS data on February 11, 2004 at 2035 UTC over the northwest portion of the United
States. The false color composite uses near infrared channels to delineate clouds (white)
from snow on the ground (red).
Clouds also emit thermal energy propotional to the product of their emissivity and
temperature. Clouds generally take on their ambient environment temperature which,
under normal lapse rate conditions, decreases with height in the atmosphere. Therefore
when emission is converted to equivalent blackbody temperature it can be used to
distinguish the presence opaque clouds from a warm surface with various thresholding
techniques. Thermal imagery (as measured from satellites in the 8.0-12.0 micrometer
window region) usually provides only limited contrast between cloudy and clear regions
however, allowing for the detection of only high (cold) clouds with single channel
techniques. Most cloud detection approaches use thermal emission variations at short (3.5-
4.0 micrometers) and long thermal wavelengths (10.0-12.0 micrometers) to differentiate
clouds from cloud-free scenes based on spectral difference thresholds. As with the visible
and near infrared spectral tests, emission and reflectance characteristics of the surface and
clouds vary considerable with time, season, and region, limiting the use of fixed thresholds
for successful cloud detection. The successful use of infrared detection methods relies on the
determination of appropriate test thresholds.
An underlying principle applied with spectral difference tests using longwave and
shortwave infrared window channels is that the difference between the emissivity of clouds
at these wavelengths (e.g., 11.0 and 3.9 micrometers) varies from that for the surface (land or
ocean). This difference can be inferred from channel brightness temperature differences and
used to detect various types of clouds. The spectral emissivity varies at both of these
wavelengths and with surface and cloud type, with the emissivity at the shortwave infrared
wavelengths being smaller than at the longwave infrared wavelengths, resulting in lower
emission temperatures at the shorter wavelengths. However, during the day reflected solar
radiation (in the 3.0-4.0 micrometer region) makes the effective brightness temperatures
(sum of emission and reflective components) at the shorter wavelengths larger than the
brightness temperatures at the longer wavelengths even though the emissivity is less. For
cloudy pixels, the longwave minus shortwave brightness temperature difference has a large
negative value during the day, and at night a positive value for opaque clouds (thick water
clouds and fog) because of the absence of solar radiation, and a negative value for thin ice
clouds. Even though the emissivity of thin ice clouds is greater at long wavelengths than at
short wavelengths, much of the energy sensed by the satellite comes from the Earth's surface
and atmosphere below the cloud and the 3.9 micrometer channel's response to warm sub-
pixel temperatures is greater than it is at 11.0 micrometer, resulting in negative difference
values both during the day and during the night. Because the difference between shortwave
and longwave emissivities is on average smaller for land and water than for clouds, cloud-
free pixels will have a small negative temperature difference value during the day and a
small negative or positive value at night.
The spectral properties of clouds and the Earth’s surface allow for the delineation of clear
and cloudy regions in thermal image as a discontinuity in the longwave minus shortwave
brightness temperature difference image. Emissivity variation with cloud type affects the
reflected component of the shortwave channel (due to varying solar input at the surface or
cloud top) and makes the use of these channel differences for cloud detection a useful but
challenging option in cloud detection. Figure 3 shows an example of a nighttime 11.0 – 3.9
micrometer difference image (top) and the corresponding infrared image (bottom) from the
GOES satellite operated by NOAA. Notice that the warmer water clouds such as those over
Texas and over the Atlantic Ocean, indicated by mid-to-light gray in the infrared image,
have positive difference image values (indicated by yellow in the difference image). The
colder thinner clouds shown as very light gray to white in the infrared image such as seen
over New Mexico, Colorado and Iowa, have negative values (aqua colored clouds in the
difference image). The clear land and water pixels have small negative or positive values,
and are shown in gray shades in the difference image.
Figure 3 also highlights some of the complexities of using the difference imagery. Notice
how in the infrared image the clouds over Kentucky and Tennessee appear to be of similar
temperature as those over Arkansas and Louisiana, but the majority of the clouds in the
former group have positive difference image values compared to the latter group having
negative values. However, there are clouds that are easily seen in the infrared image, such
as the large cloud feature over Louisiana, that are not indicated as clouds in the difference
image, but have difference values similar to the surface. The key to the successful detection
of clouds with these properties lies in the selection of an appropriate threshold value for the
11 - 3.9 micrometer difference image test which separates cloud-free pixels (where the
sensor observes the land) from cloudy pixels. The use of a fixed threshold will not produce
good results in this situation.
Fig. 3. Typical GOES nighttime 11.0-3.9 micrometer difference image (top) and 11.0
micrometer image (bottom) over the eastern two-thirds of the U.S.
2.1 Traditional approaches

Numerous approaches for the automatic detection of clouds have been applied to visible
and infrared data from geostationary and polar-orbiting satellites. Some of these approaches
consider the spatial variance of infrared radiances within a region to separate the points into
clear or cloudy scenes (e.g., Coakley and Bretherton, 1982). The International Satellite Cloud
Climatology Project (ISCCP) uses visible and infrared data that includes a series of both
spatial and temporal tests (Rossow and Garder, 1993). Other researchers such as Simpson
and Gobat (1995a) present a clustering algorithm for GOES data over land and an adaptive
thresholding algorithm over ocean (Simpson and Gobat, 1995b), both for daytime scenes.
2.2 Advanced and operational techniques

An approach used operationally by the NOAA National Environmental Satellite, Data and
Information Service (NESDIS) for the detection of clouds in GOES Imager and Sounder data
for atmospheric applications is discussed by Schreiner et al. (2001) and Schriner (2001). Their
approach derives cloud information using a limited number of visible and infrared channels
both during the day and at night. The scheme delineates between cloud-free pixels used for
atmospheric product retrievals and cloudy pixels which are used to derive cloud parameters
such as cloud top pressure, temperature, effective cloud amount, and other atmospheric
products. The NESDIS approach uses four infrared window channels (at 12.7, 12.1, 11.0 and
3.9µm) and the 13.4 micrometer carbon dioxide channel on GOES in the determination of
sky conditions. The approach is broken up into a series of channel tests; the first two are
applied to a 3 x 3 array of pixels, and the last one applied to individual pixels. The first set of
tests uses a simple 11.0 micrometer threshold test (with a cloud threshold derived from an
estimated surface temperature minus 20 K), a visible channel reflectance test (with a
background cloud-free albedo threshold value of 25% over land and 7% over water), and a
nighttime inversion test (based on the shortwave and longwave infrared channel difference).
A second set of tests includes a cold sea test, skin temperature retrieval test, and a reflected
sunlight test (based on the infrared channel difference compared to a 25 K threshold value).
The final set of tests is applied to each individual field of view and includes tests for snow
(visible reflectance), stratus (11.0 -12.0 micrometer against a fixed threshold value based on
channel noise), and thin cirrus (based on 13.4 micrometer band comparison to an assumed
"cloud-free" value). This method is both complex in its application and requires the use of a
number of different threshold values in the channel tests.
Another popular cloud detection technique was developed to detect clouds in AVHRR data
on the NOAA polar orbiting satellites (Saunders and Kriebel 1988). The method uses data
from visible and infrared channels on the AVHRR instrument in several cloud threshold
tests to determine if pixels are cloudy or cloud-free. The first test is an infrared threshold test
which compares the 11.0 micrometer channel brightness temperature to an estimate of SST
or land surface temperature. The test is applied day and night and thresholds are
determined by computing weekly averages of SSTs, and subjectly determined thresholds
over land based on observed data. The second test is a spatial coherence test which
evaluates the standard deviation of brightness temperatures in a 3 x 3 array of 11.0
micrometer pixels (as in Coakley and Bretherton 1982). The test is appled over the ocean
during the day and land and ocean at night. The presence of clouds in a 3 x 3 pixel region
will cause the standard deviation to exceed background scene variations. The third test is a
dynamic reflectance test which is applied during the day. Histograms of visible reflectances
(over land) and near infrared reflectences (over water) for a 50 x 50 pixel region are used to
determine thresholds which isolate pixel values corresponding to clouds. An additional test
involved evaluating the ratio of the near infrared to the visible channel reflectences during
the day. Values close to unity indicate cloudy pixels. Assigned thresholds above and below
1.0 are used to determine cloud-free pixels. A fifth test evaluates a group of split window
channel differences between the 11.0-3.7, 3.7–12.0, and 11.0–12.0 micrometers for each pixel.
The oceans group at NESDIS have also applied a probability-based cloud detection
approach to mask clouds in the generation of a global sea surface temperature product from
GOES and AVHRR data (Maturi et al. 2008). This approach is an adaptation of the Bayesian
method of Merchant et al. (2005) in which an estimate of the probability of clear sky is made
for each pixel in the night-time imagery. This approach uses observational texture elements
consisting of the standard deviation of the 3.7 and 11.0 micrometer brightness temperatures
over a 3 x 3 pixel box as the observation vectors. The background state vector which
describes the prior knowledge of the atmosphere and surface thermodynamic state is taken
from forward radiative transfer calculations using numerical model forecasts relavent to the
observational environment. The joint probability function for cloudy pixels is obtained from
past 3.7 and 11.0 micrometer imagery. This approach brings in the window brightness
temperature differences used in other approaches. By choosing a probabilility threshold, the
Bayesian method allows one to generate a binary mask which provides optimal
performance for a given application.
The National Aeronautics and Space Administration (NASA) EOS science team has
developed an eloborate cloud detection technique for MODIS data on the Terra and Aqua
polar orbiting satellites. The approach uses a series of threshold tests applied to many of the
36 MODIS channels to identify clouds in indivudual pixels (Ackerman et al. 1998; King et al.
2003). Test thresholds are used for various surface types and at different times of the day.
Each cloud detection test returns a confidence level (1.0 for high confidence and 0.0 for low
confidence) that the pixel is clear. Tests which are capable of detecting similar cloud
conditions are group together and a minimum confidence is determined for each group. The
final cloud mask is then determined from the product of the results from individual groups.
The algorithm uses 10 spectral tests with varying threshold values based upon the
underlying scene. These tests include single channel threshold tests with the 11.0, 13.9, and
6.7 micrometer channels with nominal threshold values of 270, 224, and 220 K respectively.
These tests determine the presence of opaque cold clouds quite well. The algorithm also
uses a test of the 11.0 micrometer window channel against the surface temperature for warm
cloud detection. Four channel difference tests are also used in the technique. The 11.0 – 12.0
micrometer test is the traditional split window test used with AVHRR data. The 11.0 – 3.9
micrometer channel is effective in detecting clouds at night becasue of emissivity variations.
The 3.9 – 12.0 micrometer and 7.2 – 11.0 micrometer tests are useful at high latitudes with
appropriate test thresholds. The 8.6 – 7.2 micrometer channel difference is also used at night
in polar regions. A series of tri-spectral tests with the 8.6, 11.0, and 12.0 micrometer channels
account for variations in ice and water absorption under varying water vapor conditions. A
group of visible and near infrared reflectance tests are also used during the day. These
involve single channel reflectence tests at 0.66, 0.86, and 1.38 micrometers, and ratios of the
0.86 / 0.66 micrometer channels. In all cases a range of values around a scene dependent
threshold is used to bound the confidence in the test.
The bispectral composite threshold (BCT) method (Jedlovec et al. 2008) uses multispectral
channel differences to contrast clear and cloudy regions in satellite imagery. The BCT
method uses the 11.0 and 3.9 micrometer spectral channels in a four-step cloud detection
procedure, comprising of two spatial tests and two spectral tests. The two spatial tests and
one of the spectral tests are applied to the 11.0 and 3.9 micrometer difference imagery, and
the second spectral test is applied to the 11.0 micrometer image. The first test subjects each
pixel in the 11.0 - 3.9 micrometer brightness temperature difference image to an adjacent
pixel test. The variance between adjacent pixels along the scan line in the difference image is
compared to a threshold value to detect a cloud edge. The threshold value is fixed for all
pixels at 7.0 K and was subjectively determined from viewing a large number of difference
image fields. This test is most successful in identifying the edges of clouds during the day.
The second step in the BCT method attempts to fill-in between the cloud edges by analyzing
the one-dimensional spatial variability of the pixels. The difference between two adjacent
pixels is calculated. For a cloud to be detected, this calculated difference value must be less
than the cloudy threshold value if the preceding image location was cloudy, or it must be
either less than the negative of the clear threshold value or greater than two-thirds of the
clear threshold if the preceding image location was clear. In this way the spatial variability
in the difference image corresponding to a cloud free surface versus a cloud is considered.
Threshold values of 3.0 and 0.0 K, for the clear and cloudy regions, respectively, were
derived in a similar fashion as for the first spatial test. The threshold values for these first
two tests can be adjusted (tuned) to tweak algorithm performance for particular applications
and regions. These tests only detect a small percentage of clouds. However, these clouds are
usually on the edge of a larger cloud field, and are diffuse or only partially contaminate a
pixel in the image, and are not easily detected by other means.
The third step in the BCT method is used to detect clouds in regions where the first two
steps do not detect clouds. It utilizes a positive and negative difference image threshold in a
robust 11.0 – 3.9 micrometer difference image test. This test is the most important cloud
determination component in the BCT technique and is used to detect the majority of the
clouds in a satellite image. The positive and negative difference image composites derived
for each observation time (described below) represent the smallest positive and negative
observed difference image values, respectively, for the preceding twenty day period. This
test compares the current difference image value to these composite images. A pixel is
deemed cloudy if the difference image value is smaller than the negative composite value or
if the difference image value, is larger than the positive composite value. The fourth and
final test in the BCT cloud detection method involves using the longwave 11.0 micrometer
channel information. This infrared threshold test uses a twenty day composite of the second
warmest thermal infrared channel values at each pixel location. This product is essentially a
“warm” cloud free thermal infrared image. A pixel in the observed infrared image is
deemed cloudy if its infrared temperature is colder than the composite infrared threshold
value corresponding to its location and time period.
In the bispectral composite threshold technique, the 11.0 and 3.9 micrometer channels are
used to produce a difference image (11.0-3.9 micrometer) used in the cloud-free generation
of the composites. Both positive differences, which mainly occur at low sun angles and at
night, and negative differences that occur at all times, are preserved in the difference image.
From this difference image information, two composite images are created for each
observation time, which represent the smallest negative (values closest to zero) and the
smallest positive difference image values, from the preceding twenty day period. The
premise here is that difference image values close to zero have the highest probability of
representing cloud-free pixels. These composite images serve to provide spatially and
temporally varying thresholds for the BCT method. An additional twenty day composite
image is also generated for each time using the warmest longwave (11.0 micrometer)
brightness temperature for each pixel from the previous twenty day period. These warm
11.0 micrometer brightness temperature composite images are assumed to represent warm
cloud-free thermal images, one for each observation time of day.
The twenty day composite images (positive and negative difference images and the warm
11.0 micrometer images for each hour) used by the BCT algorithm, represent the innovative
aspect of the method and provide both spatially and temporally varying clear-sky threshold
values for comparison to the observed data. By producing these composites, the BCT
approach uses a different threshold value from one pixel to the next and therefore the pixel
location, underlying terrain features, present sun angle and other surface conditions (e.g.
snow cover) are all implicitly taken into account. These composites capture a large variation
in the 11.0 – 3.9 micrometer differences across the region which range from close to zero to
above -15 K. The smaller differences (closest to zero) generally occur over water, and the
larger values are over land. Rivers can be distinguished from the surrounding land in the
composite imagery. The warm infrared brightness temperature composite better represents
the thermal structure of the surface than a single threshold value in the absence of clouds.
These infrared composite images (from the preceding twenty days) contrast the cold surface
of the ocean with warmer land regions during the warm season. The opposite contrast
occurs in the cold season. Temperature variations across the domain are considerable (270 -
310 K) and using the composite allows for a spatially representative surface temperature
value to be used in the detection approach.
3. Validation of cloud detection techniques

The performance of automated cloud detection algorithms is usually evaluated by
comparison to other satellite imagery. However, Schreiner et al. (2001) and Schreiner et al.
(1993) used ground based observations from Automated Surface Observing Station (ASOS)
and aircraft pilot reports (PIREP) to construct a validation data set. Although limited in
coverage and cloud altitude, the use of these data in their evaluation showed that the
NESDIS method correctly detected clouds 71% of the time in a 14 month comparison over
the continental U.S. region. In a more limited study of a dataset from March 2000,
Hawkinson et al. (2001) reported that the method determined the correct sky conditions
(cloudy and clear) 75% of the time.
A few studies have used the subjective determination of the presence of clouds to generate a
“truth” data set for the quantitative evaluation of the accuracy of a particular scheme.
Saunders and Kriebel (1988) determined the presence of a cloud over a region by visually
identifying clouds in the satellite imagery for 250 points over a limited region over Europe.
They used this data set to validate the performance of their algorithm on AVHRR data. For a
day-time scene from April 14, 1985, the algorithm had a hit rate for cloud detection of 98%
but with a false alarm rate of 34% and a skill score of around 65%. These statistics indicate
that the application of the algorithm to this data produced a very conservative cloud mask
for this AVHRR data set. Merchant et al. (2005) developed a quasi-objective “tailored” mask,
using a “widget” in the Interactive Data Language (IDL), which was used as “truth” data to
validate their Bayesian approach. Using a 99%clear sky probability as a threshold for a
binary mask, they reported a hit rate of 97.2%, and a false alarm rate of 23.7% with a skill
score of 73.5 for 43 night-time cases detecting clouds over the western Pacific. This approach
was a significant improvement over the Saunders and Kriebel (1988) method which had a
hit rate of 96.3%, a false alarm rate of 32.3% with a skill score of 64.1% for the same data.
Jedlovec et al. (2008) took a subjective approach to generating a “truth” data set for
validation of the bispectral composite threshold approach to cloud detection with GOES
imagery. In their validation work, satellite meteorologists manually determined sky
conditions for 30 points per time period, per day, during four different months throughout
2004-2005 using GOES visible and infrared imagery. Subjectively determined sky conditions
varied by just a few percent between meteorologists and the individual “truth” values were
averaged between the meteorologists for each point. Their combined results for night and
day applications of the bispectral composite threshold technique applied to GOES data in 4
different seasons produced a hit rate of 84.5%, with a false alarm rate of only 8.1%, with a
skill score of 76.4% for 16,675 validation data points. These combined results indicated good
performance of the algorithm with only minimal over determination of clouds in the
comprehensive data set.
It is difficult to inter-compare the performance of various cloud detection techniques
because they are often used in different settings. Each cloud detection technique may use
different satellite sensors, focus on different geographical regions or even times of the day.
To better understand an algorithm‘s performance relative to another, the algorithms need to
be applied to the same data sets allowing for a direct comparison of the resulting cloud
masks. The example below presents such a comparison for August 2004 over the eastern
two-thirds of the continential U. S. The two algorithms used were the the NASA EOS
science team MODIS cloud mask algorithm and the bispectral composite threshold
technique (BCT) adapted for MODIS data (Haines et al. 2005). The MODIS collection 5 cloud
algorithm (EOS-5; Frey et al. 2008) is similar to that described above with improvements in
performance at night and in polar regions by the inclusion of several additional tests. The
MODIS collection 5 data set was obtained from the EOS Data Active Archive Center
(DAAC). The application of the BCT technique to MODIS data uses five tests applied to the
shortwave and longwave infrared window channels and a similar compositing approach.
The algorithms were applied to all Terra MODIS passes over the eastern two-thirds of the
United States (both day and night) during this case study period. An example of the cloud
masks produced by these algorithms is shown in Figure 4 for a nighttime pass on August 5,
2004. Note that the EOS-5 algorithm assigns confidence levels (i.e., cloudy, possibly cloudy,
probably clear and clear) to each pixel in the cloud mask with corresponding grey shades in
Figure 4 (from white to dark grey, to light grey to transparent, respectively). An appropriate
land type color is assigned to clear regions in both products for display. A qualitative
comparison of these cloud mask products to the corresponding infrared window channel
image indicates that both cloud masks do quite well in capturing the gross cloud features on
this day and time. They both capture the obvious cold clouds observed in the infrared
image, but some differences exist in cloud detection for low (warm) clouds over the ocean
and land regions.
Fig 4. MODIS cloud masks and infrared imagery for August 5, 2004 for the EOS science
team collection 5 cloud detection algorithm (upper left) and the bispectral composite
threshold technique (upper right). The corresponding infrared MODIS image is presented in
the lower panel. Each image is overlaid with grid boxes used in the validation scheme.
To better quantify the performance of these algorithms in detecting clouds over this region,
the determination of the presence of clouds by each algorithm was compared to a set of
“truth“ data subjectively obtained by satellite meteorologists viewing high resolution visible
and infrared imagery from MODIS corresponding to the cloud mask times. A fixed set of 42
comparison points were strategically selected over the region as shown on the images in
Figure 4. The locations were selected based on a fixed grid with adjustments to include a
variety of ocean and land regions (including coastal areas) with a variety of cloud
distributions. This validation approach is similar to that used by McNally and Watts (2003)
and Alliss et al. (2000)
Fig 5. Performance statistics of the bispectral composite threshold (BCT) and the EOS science
team collection 5 (EOS-5) MODIS cloud mask algorithm for August 2004 over the eastern
half of the United States for night over ocean (upper left), night over land (lower left), day
over ocean (upper right), and day over land (lower right). The numbers correspond to the
accuracy (%) of the cloud detection algorithm – either correctly determining the correct sky
conditions (clear or cloudy – blue bar), or the over- (yellow bar) or under- (pink bar)
determination of the cloud conditions.
and nearly identical to that of Jedlovec et al. (2008). For the EOS-5 algorithm, only points
having a probability determination of “cloudy“ were considered to contain clouds. Other
probablistic determinations (possibly cloudy or probably clear) were considered clear. The
statistical results of this comparison were categorized by day and night and land and ocean
locations and are presented in Figure 5. Both algorithms perform nearly as well at night,
correctly detecting clear or cloudy conditons about 75% of the time over the ocean and
around 90% of the time over land. Over the ocean, both schemes over determined the
presence of clouds by nearly 25%. Over land, the EOS-5 algorithm correctly determined
clear or cloud conditions 91.6 percent of the time, with the remaining misses split between
over and under determination. The BCT scheme did nearly as well with 88.2% of the points
correct.
The performance of the two algorithms varys significantly during the day over the ocean,
with significant over determination of clouds in the EOS-5 scheme contributing to its poor
performance. During the day over the ocean, the BCT scheme does extremely well, correctly
determining 93.9% of the points. The performance of the two algorithms are similar over
land during the day with 87.4% of the points being determined correctly as being either
cloudy or clear. The BCT cloud mask has a slightly greater over determination value than
the EOS-5 approach for the points that were incorrectly identified. One should note that the
use of only cloud points having the highest confidence were considered cloudy in with the
EOS-5 data. If a more conservative cloud choice was used (by including “possibly cloudy“
points), the performance statistics for the EOS-5 algorithm would have been worse as a
greater number of points would be considered cloudy increasing the over determination of
clouds at all times and regions (not shown). This qualitative comparison and statistical
analysis indicates that with the proper determination of thresholds, a simple two channel
cloud detection technique, such as the bispectral composite threshold approach (Jedlovec et
al. 2008), can detect clouds in a variety of situations quite well.
4. Summary
Many different approaches have been used to automatically detect clouds in satellite
imagery. Most approaches are deterministic and provide a binary cloud – no cloud product
used in a variety of applications. Some of these applications require the identification of
cloudy pixels for cloud parameter retrieval, while others require only an ability to mask out
clouds for the retrieval of surface or atmospheric parameters in the absence of clouds. A few
approaches estimate a probability of the presence of a cloud at each point in an image. These
probabilities allow a user to select cloud information based on the tolerance of the
application to uncertainty in the estimate. Many automated cloud detection techniques
develop sophisticated tests using a combination of visible and infrared channels to
determine the presence of clouds in both day and night imagery. Visible channels are quite
effective in detecting clouds during the day, as long as test thresholds properly account for
variations in surface features and atmospheric scattering. Cloud detection at night is more
challenging, since only courser resolution infrared measurements are available. A few
schemes use just two infrared channels for day and night cloud detection. The most
influential factor in the success of a particular technique is the determination of the
thresholds for each cloud test. The techniques which perform the best usually have
thresholds that are varied based on the geographic region, time of year, time of day and
solar angle.
5. References
Ackerman, S. A., et al., (1998). Discriminating clear-sky from clouds with MODIS. J.
Geophys. Res. Vol. 103, pp. 32,141-32,158.
Alliss, R. J., et al. (2000). The development of cloud retrieval algorithms applied to GOES
digital data. 10th Conference on Satellite Meteorology and Oceanography, Amer.
Meteor. Soc., Long Beach, 2000, pp. 330-333.
Coakley, J. A., and F. P. Bretherton, (1982). Cloud cover from high-resolution scanner data:
Detecting and allowing for partially filled fields of view, J. Geophys. Res., Vol. 87,
pp. 4917-4932.
Frey, R. A., S. A. Ackerman, Y. Liu, K. I. Strabala, H. Zhang, J. R. Key, and X. Wang, (2008).
Cloud detection with MODIS. Part I. Improvements in the MODIS cloud mask for
collection 5. J. Tech., 25, 1057-1072.
Goodman, A. H., and A. Henderson-Sellers, (1988). Cloud detection and analysis: A review
of recent progress. Atmos. Res., Vol. 21, pp. 203-228.
Haines, S. L., et al., (2005). Spatially Varying Spectral Thresholds for MODIS Cloud
Detection, 13th Conference Satellite Meteorology and Oceanography. Amer. Meteor
Soc. Norfolk, [Online] Available
http://ams.confex.com/ams/13SATMET/techprogram/paper_79134.htm
Hawkinson, J. A. et al. (2001). A validation study of the GOES sounder cloud top pressure
product, 11th Conference on Satellite Meteorology and Oceanography, Amer.
Meteor. Soc., Madison, 2001, pp. 348-350.
Jedlovec, G. J., S. L. Haines, and F. J. LaFontaine, (2008). Spatial and temporal varying
thresholds for cloud detection in GOES imagery. IEEE Trans. Geosci. Rem. Sens.,
Vol 46, No. 6, 1705-1717.
King, M. D., et al., (2003). Cloud and aerosol and water vapor properties, precipitable water,
and profiles of temperature and humidity from MODIS, IEEE Trans. Geosci.
Remote Sensing, Vol. 41, pp. 442-458.
Maturi, E., et al., (2008): NOAA’s sea surface temperature products from opertional
geostationary satellites. Bull. Amer. Meteor. Soc., Vol. 89 pp 1877-1888.
McNally, A. P., and P. D. Watts, (2003): A cloud detection algorithm for high-spectral-
resolution infrared sounders. Q. J. R. Meteorol. Soc. Vol. 129, 3411-3423.
Merchant, C. J., et al., (2005). Probabilistic physically based cloud screening of satellite
infrared imagery for operational sea surface temperature retrieval. Q. J. R.
Meteorol. Soc., Vol. 31, pp. 2735-2755.
Reuter, M., et al., (2009): The CM-SAF and FUB Ccloud detection schemes for SEVIRI:
Validation with synoptic data and initial comparison with MODIS and Calipso. J.
Appl. Meteor. And Climate, Vol. 48, pp 301-316.
Rossow, W. B., and L. C. Garder, (1993). Cloud Detection Using Satellite Measurements of
Infrared and Visible Radiances for ISCCP. J Climate, Vol. 6, pp. 2341–2369.
Saunders, R. W., and K. T. Kriebel, (1988): An improved method for detecting clear sky and
cloudy radainces from AVHRR data. Int. J. Remote Sensing, Vol. 9, pp 123-150.
Schreiner, A. J., et al., (1993). A comparison of ground and satellite observations of cloud
cover. Bull. Amer. Meteor. Soc., Vol. 74, pp. 1851-1862.
Schreiner, A. J., et al., (2001). Observations and trends of clouds based on GOES sounder
data. J. Geophys. Res, Vol. 116, pp. 20349-20363.
Schreiner, A. J., (2001). Derived cloud products from the GOES-M imager," 11th Conference
on Satellite Meteorology and Oceanography, Amer. Meteor. Soc., Madison, (2001),
pp. 420-423.
Simpson, J. J., and J. I. Gobat, (1995). Improved cloud detection in GOES scenes over land,
Remote Sens. of Environ., Vol. 52, pp. 36-54.
Simpson, J. J., and J. I. Gobat, (1995). Improved cloud detection in GOES scenes over the
oceans, Remote Sens. of Environ., Vol. 52, pp. 79-94.
Remote sensing and the disaster management cycle 317
15
X
Remote sensing and the disaster

management cycle
Karen E. Joyce1, Kim C. Wright1, Sergey V. Samsonov1,
and Vincent G. Ambrosia2
1GNS Science
New Zealand
2California State University – Monterey Bay
United States of America
1. Introduction
Disaster management planning is structured around the disaster management cycle
model. The cycle consists of four stages – reduction, readiness, response and recovery.
Remotely sensed data can provide a valuable source of information at each of these
stages, helping to understand spatial phenomena, and providing scientists and authorities
with objective data sources for decision making. The challenge with disaster management
is that the inherent unpredictability and range of hazards does not allow for a single all-
encompassing solution to be developed and explored. Instead, there are a multitude of
different remote sensing platforms and sensors that can and should be employed for
image acquisition. An extensive coverage of each, including optimal processing regimes
for their data would be prohibitively long; instead this chapter aims to give some general
examples of the use of remote sensing in disaster management, while directing the reader
to more specific studies in the literature. The types of data required and information
provision needs for each stage will be discussed including optical, thermal, and synthetic
aperture radar as data sources over a variety of spatial and temporal scales.
Remote sensing can be used to assist risk reduction initiatives through identification of
hazard zones associated with flood plains, coastal inundation and erosion, and active
faults. It can also be used to verify hazard models by measuring the location and
magnitude of actual events. Imagery is widely used by meteorologists for providing
weather forecasting and warnings of potentially severe weather events, providing the
public and emergency responders with information that can assist decision making
around short term readiness. These images are commonly presented in print, television
and on the internet, and they are well accepted by viewers around the world. Imagery of
fires, volcanic eruptions and flooding are often used during the response phase for the
visual impact that they provide. If people in potentially at-risk locations personalise the
risk, they are more likely to take readiness actions such as making emergency plans for
contact and evacuation or assembling emergency kits. Remote sensing images of similar
communities experiencing hazards, or the progress of a hazard such as a fire front, can
assist with this personalisation process. For agencies that respond to emergencies, remote
sensing imagery provides a rapid method of assessing the magnitude of hazard impacts,
areas most affected, and where key transport and other infrastructure links have been
disrupted or destroyed. Remote sensing can also be used to provide an indication of the
rate of recovery in an area post disaster based on indicators such as vegetation regrowth,
debris removal, and reconstruction.
There are few examples where remote sensing is incorporated seamlessly into all stages of
the disaster management cycle for planning purposes. This requires a collaborative effort
from emergency managers, policy planners and remote sensing technical staff that may
not always be co-located, or even working for the same organisation. However, data is
becoming more readily available, and some satellites and constellations are even targeting
at least partially the disaster management / emergency response community in
recognition of the value remotely sensed imagery can provide. If this current trend
continues, integrating remote sensing and emergency management will become
increasingly more commonplace.
2. The disaster management cycle

The traditional approach to hazard risk and disaster management has been one primarily
focussed on response to events as they occur (Gregg & Houghton 2006), managing
residual risk through warning systems and emergency management plans, and more
recently attempting to reduce risk through changing the hazard process or impacts (Board
on Natural Disasters 1999). Examples of attempts at hazard modification include: the use
of stopbanks and levees to provide opportunities to build in areas vulnerable to flood
hazard; building codes for strengthened buildings to allow development in earthquake
prone locations; and building seawalls along coasts to reduce susceptibility to erosion and
coastal inundation. These measures have allowed greater development in hazardous
areas, and are typically designed for protection up to a certain magnitude of event, but
there always exists the potential for design limits to be exceeded (Burby 1998). Because of
reliance on technological solutions, risk is increasing in the developed world as infill and
migration increases in “protected” areas (Mileti 1999). In less-developed nations, risk is
also increasing, although the drivers differ. Reliance on decreasing natural resources,
population increase, poverty, and political drivers push communities into hazardous
areas traditionally left un-settled (Donner & Rodriguez 2008). The body of research into
the evolution of hazards and disaster management now recognises that it is primarily
social drivers that create vulnerability to hazards, and consequently increase the potential
for disasters (Board on Natural Disasters 1999, Cutter & Finch 2008, Donner & Rodriguez
2008, Pertrow et al. 2006, Wisner et al. 2004). The overall focus of emergency management
has shifted to consider disaster management planning as part of a broader system of
planning for sustainable, resilient communities. Whether a hazardous event will become a
disaster - an event that is beyond the capacity of responding agencies, resources, and
community coping capacity (Quarantelli 1985), can be influenced by effective disaster
management planning.
This recognition of the importance of social drivers has brought about a change in how
disaster planning is considered and undertaken. Many nations now plan using a variation
of the Disaster Management Cycle, an integrated, four–phase planning system. Although
the cycle can be considered as a continuum, traditionally the first phase of the cycle is
considered to be reduction, followed by readiness, response, and recovery (Figure 1).
Fig. 1. The disaster management cycle
Reduction incorporates all measures and planning that reduce the likelihood of a disaster
occurring. This is done through the process of risk identification and reduction; either by
modifying the hazard process using traditional structural methods such as stopbanks or
seawalls, or by modifying behaviours and the assets at risk (Gregg & Houghton 2006).
Behaviour modification includes land use planning to: prevent development in hazardous
areas; incorporate good access for response and evacuation; and foster interconnected and
resilient communities (Burby 1998). In theory, land use planning can reduce all risk from
disasters, but centuries of settlement in hazardous locations make this option unrealistic
and impractical. Modifying assets at risk includes such methods as strengthening
buildings and infrastructure and raising floor heights to reduce hazard impacts.
Readiness planning accepts that some residual risk is present for communities and that
measures must be in place to ensure any response to hazards is efficient and reduces
hazard impacts. Readiness planning includes: public education on hazards and their
consequences, and how these consequences can be reduced; training of emergency
planners and responders; installing monitoring and warning systems for hazards;
exercising response plans; and fostering community resilience through increased uptake
in home preparedness such as learning first aid, having an emergency kit and an
evacuation plan (Ronan & Johnston 2005).
The phase of disaster management that has traditionally received the most recognition,
funding and planning effort is Response (Gregg & Houghton 2006). This fact is also
reflected in the remote sensing community, with an overwhelming number of research
papers dedicated to the use of imagery for disaster response, despite the fact that data
often cannot be provided in the timeframe required to be of use for decision makers. The
reality is that most nations do not have the capability to prevent disasters occurring; the
best option for reducing the chance of a disaster is through reducing risk. However,
response capability is important in any disaster as it involves the processes of coordinated
effort to manage resources, including life essentials and personnel, for activities such as
evacuation, relief, search and rescue and needs assessment (Quarantelli 1997).
Recovery, the fourth phase of the cycle has traditionally been focussed on restoration of
lifeline utilities, and building reconstruction. There is now considerable research into
holistic recovery processes, which recognise that for community recovery to be
sustainable, the social, economic, built and natural environments must be considered
(Norman 2004). The four environments are interlinked as communities rely on:
• Natural environment for amenity (recreation, psychological wellbeing), and
resources (to provide opportunities for construction and employment);
• Built environment for lifeline utilities and structures to enable people to live,
work and recreate;
• Economic environment to provide goods, services and livelihoods; and
• Social environment, to provide opportunities for political participation,
community building, networking and psychological wellbeing.
The recovery phase of a disaster can be considered to have several steps, the initial
restoration of lifeline essentials, and the longer term rebuilding of communities. The
recovery phase is often considered to be an optimal time to include measures that will
reduce the risk of future disasters (Becker et al. 2008).
The four phases of the disaster management cycle are not discreet; they are interrelated
and ideally integrated throughout the planning process. Decisions about risk reduction
methods will affect the degree of readiness planning and response that will be required.
Readiness levels of affected communities and responders can determine whether an event
becomes a disaster, as can be seen in the failure to provide evacuation options for the 20%
of the New Orleans population with no vehicle or resources to leave the city prior to
hurricane Katrina’s landfall (Laska & Morrow 2006/7). The effectiveness of the response
phase will play a significant role in how affected communities recover, both physically
and psychologically. Lessons from the response phase can be incorporated into risk
reduction and readiness planning. Finally, the recovery phase can include risk reduction
measures to increase resilience and reduce future vulnerability.
3. Remotely sensed data types

In order to successfully use remote sensing for disaster management, physical indicators
of features or attributes within the disaster management cycle that are measureable in
imagery need to be identified. At that point, selection of the most appropriate remotely
sensed data set is possible by identifying the spatial, spectral, temporal, and radiometric
requirements. The use of a framework for selecting appropriate remotely sensed data has
been demonstrated for mapping and monitoring coastal and tropical wetlands, tropical
rainforests, coastal ecosystems and coral reefs (Phinn 1998, Phinn et al. 2006). This is an
approach that can be modified and applied under many different circumstances and for
various environments. Here we look to apply aspects of the framework to disaster
management. During the reduction, readiness and recovery phases, there may be
sufficient time to develop and apply the framework as the cycle is progressing. However,
as timeliness is a critical factor in the response phase, it is of most use to already have
systems in place to aid with appropriate data selection so that crucial decisions need not
be made under the severe time constraints that are necessitated by rapid response.
Preparation may therefore involve developing a range of scenarios representing potential
situations that require rapid response at a set location, and applying the principles of data
selection and processing in advance. In this way, the decisions regarding remote sensing
in the response phase can actually be made during the readiness phase instead. This
should be done as a collaborative exercise between both remote sensing experts and
emergency management agencies.
The types of satellite and airborne sensors that can be used to support phases of the
disaster management cycle are many and varied. It is most important to consider the
spatial scale of the hazard, in addition to determining the most appropriate data type to
address the problem. For example, geostationary satellites provide data over a large area,
but with minimal spatial detail, and are appropriate for monitoring weather patterns
(readiness) and volcanic ash and gas distribution (response). Conversely, very high spatial
resolution data (e.g. aerial photography, Quickbird, Ikonos, Worldview) are appropriate
for targeting relatively small areas where they can provide a great deal of detail.
Examples of their use include baseline infrastructure mapping for scenario development
and model validation (reduction and readiness), building damage (response), and
observations of debris removal and reconstruction (recovery).
In the disaster reduction phase, the focus for remote sensing is often on mapping
landscape features such as land cover / land use, and the location of potentially
hazardous features or processes to avoid when developing infrastructure (e.g. active
faults, flood plains). During the readiness phase, the emphasis is on monitoring these
features or processes, developing models for forecasting purposes, and using maps and
model for training and education. In the response phase, the timely acquisition of data
and provision of information to emergency services is critical. Much of the attention will
be placed on identifying infrastructure that has been damaged or is likely to be at risk in
the near future (e.g. housing in the path of a bush fire). Finally during the recovery phase,
the focus will shift to long term monitoring of debris removal, vegetation regeneration,
and reconstruction.
3.1 Optical
There are a large number of applications for which optical remotely sensed imagery can
be used to aid the disaster management cycle. Optical data can be of particular use to the
disaster management community as it is generally simple to understand and interpret raw
data, particularly when collected using standard true colour spectral bands (blue, green,
and red). The characteristics of the sensor are important in selecting the most appropriate
data type for use in individual situations. Consideration should be primarily given to the
spatial and temporal resolution of the sensor. These factors will differ depending on the
disaster management activity. For example, during the response phase, rapid acquisition
of data following the event is crucial. During the recovery phase, the speed of acquisition
is less important than repetition on a consistent basis. In the early stages of recovery,
imagery may be useful on a monthly basis, though as time passes, an annual acquisition
may suffice.
Optical data can be used for activities in all stages of the disaster management cycle,
however the greatest potential contributions are for monitoring recovery, and helping to
plan for reduction and readiness. The use of satellite optical data for immediate response
at a local scale is currently hindered by the speed of data acquisition and delivery with
polar orbiting satellites. For large events, a more regional synoptic view is possible using
geostationary satellites; however the amount of detail able to be extracted from these
images is reduced.
The greatest limitation of optical sensors under many hazard or disaster scenarios is the
inability to obtain imagery through clouds, smoke or haze. Events such as wildfires,
volcanic eruptions, and tropical cyclones or other severe storms are characterised by
cloud and smoke, which can effectively obscure damage on the ground both during and
immediately subsequent to an event.
3.2 Thermal
As energy decreases with increasing wavelength, thermal wavelengths have
comparatively low energy levels and consequently thermal image data have a lower
spatial resolution than that capable of being achieved with optical imagery. As yet there
are no very high spatial resolution thermal satellite sensors commercially available.
Nonetheless, thermal imagery provides a valuable source of information about volcanic
eruptions and the location of wildfires. Robust techniques for automatic extraction of
anomalous high temperatures or ‘hotspots’ have been thoroughly tested and considered
operational on a global scale using MODIS, AVHRR or GOES imagery (Wright et al. 2002,
Wright et al. 2004). The University of Hawai’i and Geoscience Australia both apply
automated hotspot detection algorithms for the detection of volcanic activity and
bushfires respectively and serve the information in near real time via the internet. These
algorithms have primarily been developed to detect features above the background or
average temperature values, and to avoid large numbers of false alarms, they are not
sensitive to merely warm features. They are also unable to differentiate between the types
of heat source, so addition spatial information or manual interpretation may be required.
Higher spatial resolution thermal imagery for analysis at local scales can be obtained
using ASTER or Landsat TM/ETM+, though neither of these sensors have the ability to
provide imagery of rapidly changing thermal features, as their orbits only allow them an
overpass frequency of approximately 16 days. Nevertheless, both sensors are useful for
tracking longer term temperature fluctuations, such as the warming and cooling cycles of
volcanic lakes (Joyce et al. 2008b, Oppenheimer 1993, Oppenheimer 1997, Trunk &
Bernard 2008). The higher resolution imagery can also be of use in calibrating and
validating data obtained from the likes of MODIS.
As the temperature of an object increases, the wavelength of peak radiation decreases.
Very hot features can therefore be seen in visible or shortwave infra red (SWIR) imagery
and often become saturated in thermal infra red data if they are sufficiently large with
respect to the pixel size. This relationship has been demonstrated using forest fire size and
the temperature difference between a smouldering and flaming fire that could be of use in
understanding different stages of fire development (Giglio et al. 2008). Unfortunately the
SWIR bands on ASTER were declared non-functional in January 2009 after experiencing
technical difficulties since May 2007. These five SWIR bands fall within a similar spectral
range as Landsat TM/ETM+ band 7 that could be used as an alternative.
3.3 Synthetic Aperture Radar

Synthetic Aperture Radar (SAR) is an active microwave sensor that is capable of acquiring
data in harsh weather and lighting conditions not suitable for optical sensors, such as
dense cloud or smoke coverage (Elachi 1987, Franceschetti & Lanari 1999, Hanssen 2001).
Most modern SAR sensors are designed to acquire data of various ground resolution
elements ranging from 100s of metres to 1-3 metres, but higher spatial resolution images
usually have significantly smaller spatial coverage and are limited by satellite storage and
processing capacities. The incidence angle of SAR sensors can be manipulated in order to
image different areas without changing the satellite orbit, thus decreasing necessary
revisit time.
Both backscatter intensity and the phase of SAR images can be utilised. In most studies
only the relative variability of backscatter intensity within the image is used but absolute
values can be required for some multi-temporal studies. The precise interpretation of
backscatter intensity can be complicated because of its dependence on the dielectric
properties of the reflecting material, surface roughness, and sensor wavelength but at the
same time the variety of useful information still can be easily observed (landslides,
tsunami, flooding, and damage to infrastructure). Phase information of a single SAR
image has no value but comparison of phases from two SAR images acquired at distinct
times are utilised in SAR interferometry or InSAR. InSAR is capable of producing high
resolution ground deformation maps with sub-centimetre accuracy (Rosen et al. 2000).
These maps can then be used for studying the causes of deformation such as earthquakes
or volcanic activity (Massonnet & Feigl 1998). Modern satellite SAR systems are capable of
acquiring simultaneous data with more than one polarisation (e.g. Radarsat-2, ALOS
PALSAR and TerraSAR-X). This information can be used in various studies utilising SAR
polarimetry and POLInSAR techniques, such as land classification, detection of areas
affected by fire or flooding (Cloude & Papathanassiou 1988, Pottier & Ferro-Famil 2008,
van Zyl et al. 1990).
At present, commonly used satellite SAR data is acquired in three wavebands: X (3.1 cm);
C (5.6 cm); and L (23.6 cm). Waveband selection depends on the type of application, land-
cover, time span, and availability. The analysis of backscatter intensity by determining
thresholds associated with certain features can be performed in standard GIS or image
processing software, such as ArcGIS or ERDAS Imagine, but InSAR, SAR polarimetry and
POLInSAR processing require specialised software (or add-on modules to basic packages)
and extensive processing experience. The price of the data greatly varies from a few
dollars per image for purely scientific applications to a few thousand of dollars for
commercial applications. Several recently launched commercial satellites are available to
acquire data of any hazardous event with a very short delay and deliver the data rapidly
to the user, though the cost of priority commissioned data is significantly greater than that
of archived imagery (RADARSAT-2, TerraSAR-X and Cosmo-Skymed).
4. Remote sensing applications

4.1 Reduction
Disasters are social constructs in that social drivers such as migration (forced and
voluntary), conflict, modification of natural buffer systems, reliance on shrinking
resources, private property rights, urban intensification, artificial protection structures,
and economic and political vulnerability are all contributors to people living in hazardous
locations or at levels of vulnerability that make a disaster more likely. Remote sensing
technology can assist with addressing some of these “disaster drivers”, through providing
the data required to assist land use planners, emergency managers, and others tasked
with disaster management. Reduction of risk, and therefore reduction in the probability of
a disaster occurring, is an important part of the disaster management cycle. Remote
sensing can be applied in disaster reduction initiatives through identification and
understanding of hazards (Table 1). This knowledge is then applied to mitigation
activities such as land use planning, engineering structures, building codes and hazard
consequences modelling to determine methods for reducing vulnerability (Gregg &
Houghton 2006). Note that the sensor examples given in Table 1 and subsequent tables are
indicative of current or potential instrument use. Many alternative sensors with similar
characteristics could also be used.
Understanding of hazards, their magnitude, frequency, duration, location, range and
manifestation (e.g. heavy rainfall, tephra, strong winds) has long been accepted as
essential to disaster management. Although it is primarily social factors that amplify a
hazard event into a disaster (Quarantelli 1985, Wisner 2004), improved knowledge of
hazards and their potential consequences is essential for decision making about
modifying hazard characteristics, or modifying vulnerability of people and assets. Remote
sensing can be used directly for hazard identification (e.g. flood plain modelling, slope
stability and landslide susceptibility), but can also be used to derive hazard-independent
information that can be used for disaster reduction (e.g. baseline building, infrastructure,
and topographic mapping). An excellent example of the use of remote sensing for hazard
identification is provided with LiDAR mapping of active fault location (Begg &
Mouslpoulou 2009 in press). Traditionally fault location is conducted using stereo aerial
photography interpretation followed by intensive field survey. However the horizontal
and vertical resolution provided by airborne LiDAR imagery provides the capability for
identifying fault traces and extracting elevation offsets with digital data in an objective
manner. The identification of many previously unknown faults in northern New Zealand
is shown in Figure 2.
Type of information Data required Sensor example Application example

Location of fault High resolution Airborne LiDAR, SAR Use for land use planning around
traces and rupture DEM active faults to reduce risk from
zones future development in fault
hazard locations
Fault displacement Interferometric SAR ERS1/2, ENVISAT Knowledge of fault displacement
ASAR, ALOS rates are used in numerical models
PALSAR in order to forecast the magnitude
of possible earthquakes
Flood plain DEM Airborne LiDAR, Identification of flood plains can
mapping ERS1/2, ENVISAT help inform changes in land use,
ASAR, ALOS and identify areas developing
PALSAR protective measures (e.g.
stopbanks)
Land cover / land Optical and SPOT, ASTER Used for catchment management
use polarimetric SAR RADARSAT-2 planning to reduce flood and
landslide risk
Vegetation change Consistent time SPOT, ASTER Determine drought zones, inform
series of data RADARSAT-2 fire hazard mapping
Determining lahar DEM, high SAR, Airborne Hazard zonation, public
and lava flow paths resolution optical LiDAR, SPOT, awareness, determining location of
imagery AVNIR-2, ASTER safety shelters
Locating potential DEM, Airborne LiDAR, Hazard mapping for infrastructure
and actual unstable Interferometric ERS1/2, ENVISAT planning
slopes SAR, high ASAR, ALOS
resolution stereo PALSAR, aerial
optical imagery photography
Baseline Very high Aerial photography, Assist with hazard mapping to
infrastructure maps resolution optical Quickbird, Ikonos, identify key infrastructure at risk –
imagery Worldview the risk can then be addressed
through mitigation or built in
redundancy. Can also be used for
later damage assessment post-
disaster
Baseline Moderate to high SPOT, AVNIR-2, Hazard modelling
topographic data resolution optical Aerial photography,
imagery Quickbird, Ikonos,
Worldview
Table 1. Examples of information and data requirements during the reduction phase
Advances in Geoscience and Remote Sensing
Fig. 2. Identification of known and new active faults using high resolution airborne LiDAR data acquired in late 2006 (Begg &
Mouslpoulou 2009 in press). Landsat ETM+ false colour composite (5,4,2) acquired in 2001 is inset for a contextual overview of
the site. Of the active fault traces shown here, approximately 85% were unknown before undertaking this study. Also of note
also is the discovery of a large inland area that is below sea level (elevation <0m) and is a potentially hazardous region for
326
tsunami related inundation
Remotely sensed data acquisitions can be used to inform land use planning, a key tool
that authorities and communities employ to avoid or mitigate hazard risk (Burby 1998).
By identifying the location and characteristics of hazards, land use planning methods can
be applied to address the risk these hazards pose. Planning methods include mapping
hazard zones (location and range of hazard impact) and identifying the probability of
occurrence. Hazard maps are applied to developed and green field (undeveloped) land
and options for risk treatment determined. Treatment options can include measures such
as setback zones (no development within the hazard zone, e.g. proximal to active faults or
within coastal erosion or inundation zones), or special building codes (e.g. minimum floor
heights above base flood level) can be introduced to reduce the risk to assets and people
(Godschalk et al. 1998). Understanding of hazard information is one of a number of
critical factors influencing individual and group decision making for risk management
(Paton & Johnston 2001). Where hazard information is readily available to the public in a
variety of forms, including maps, there is a greater likelihood of public support for risk
reduction initiatives introduced through land use planning (Burby 2001).
Other methods for land use planning based on remote sensing data include identifying
changes in land use on flood plains to assist with flood hazard modelling. In the city of
London, Canada, Landsat images taken over a 25 year period have been used to
determine the spread of urban development (Nirupama & Simonovic 2007). The
consequent increase in impermeable surface cover facilitated more rapid runoff and less
natural absorption of rainfall. When compared with flood hydrographs, the rate of land
use change correlates with smaller rainfall events producing flooding. The benefits to
future land use planning are that it can be determined how land use changes affect the
flood hazard risk, and this will guide future development in a way that mitigates the
effects of continued urban sprawl.
Collecting asset data via high resolution remote sensing allows for identification of
infrastructure and buildings in hazardous locations, which can then be targeted for
strengthening or re-location. Asset data is also essential for hazard consequence
modelling, whereby hazard data is combined with asset data and fragility (vulnerability)
information to determine potential losses. Building fragility to hazards is based on such
factors as construction materials (earthquake, volcanic ash fall, tsunami), engineering
design (tsunami, landslide, earthquake), building height (wind), floor areas (earthquake),
proximity of other structures and vegetation (fire) and roof pitch angle (ash fall, snow),
and floor height (flood, tsunami). Remote sensing methods for collecting building and
infrastructure data require high to very high resolution satellite or airborne imagery and
is often completed using manual digitizing or more recently, segmentation and object
oriented classification. Optical imagery is often complemented by LiDAR data, which can
not only aid in detecting building edges, but is also used for calculating building heights.
Incorporation of remotely sensed data into a GIS is vital during this phase for recording
spatial attributes and combining with other data sets.
Remote sensing technology can also be applied to measure the success of risk reduction
initiatives. A common method for addressing flood risk is the construction of stopbanks
to contain flood waters for an event of a given magnitude. Aerial reconnaissance during
major flooding events can identify whether stopbanks are performing to design standard
and identify areas of weakness, overtopping or failure. Monitoring of non-structural risk
reduction initiates is also possible. To address coastal hazard erosion and inundation risk,
many communities choose non-structural options such as beach renourishment and dune
restoration. In Florida, airborne LiDAR captured over time has been applied to measure
coastal erosion from hazards, alongside the success of non-structural beach restoration
methods through determining changes to beach morphology (Shrestha et al. 2005).
Another example of measuring the effects of risk reduction initiatives is analysing post-
disaster images of rainfall induced landslides on land under different vegetation covers
for large events. From analysis of aerial photographs (oblique and vertical) of an event in
2004 which impacted the lower North Island of New Zealand, it was determined that
vegetation cover played an important role in reducing loss of productive soil, and
reducing landslide hazard to assets (Hancox & Wright 2005).
4.2 Readiness
Readiness planning and activities are undertaken in the realisation that residual risk from
hazards has the potential to create emergencies, and in some cases, disasters for affected
populations. Readiness is the identification and development of necessary systems, skills
and resources before hazard events occur. The desired outcome of readiness planning and
activities is that response to hazards is more coordinated and efficient, communities
experience less trauma, and recovery times are reduced (Quarantelli 1997). Examples of
readiness activities include public education, preparedness activities, trainng and exercising,
evacuation planning, developing hazard monitoring and public alerting systems, and putting
in place state, national and international plans and agreements for assistance and aid.
Readiness activities and planning are undertaken at a number of levels to increase resilience
and response capability for individuals, households, organisations, and states or nations. The
provision of good hazard and asset information to assist these activities is essential and
examples where remote sensing can assist this phase are given in Table 2. It is important in this
phase to prepare an archive of and gain familiarity with the most up to date spatial
information including (but not limited to) imagery, DEMs, and vector data. This information is
required to assist with damage assessment during the response and recovery phases.
At the individual and household level there are identified factors that contribute to whether
people will take actions to prepare for disasters. Personalisation of risk is essential (Barnes
2002, Slovic et al. 2000), e.g. “Will it affect me?”, “Do I need to do something about it”, and
“What can I do about it?”. Other factors include belief in the benefits of hazard mitigation
(outcome expectancy) and their belief that what they personally can do will make a
difference (reduce negative outcome expectancy) (Paton 2006). At a community level,
participation in community affairs and projects, and individual’s ability to influence what
happens in their community (empowerment) and the level of trust they have in different
organisations (trust) have also be shown to be key predicators of resilience. Therefore,
communication of risk in a meaningful way is an essential part of preparedness planning.
Remotely sensed data such as LiDAR are used to produce high resolution hazard and risk
maps, which are used by authorities to communicate information about location and range
of hazards to their communities. If individuals believe that a hazard is likely to affect them
detrimentally within an understandable and pertinent timeframe, they are more likely to
take actions to prepare. These actions might include having emergency supplies in the
home, an action plan for evacuation and emergency contact with other household members,
first aid training or training as a civil defence volunteer. The principle of risk perception
aiding preparedness applies to both static and dynamic hazards, e.g. fault trace or flood
plain mapping vs. cyclone or bushfire progression. Remotely sensed images showing the
progression of a bushfire front or the track of a cyclone are commonly used by the media to
inform the public of where hazards are occurring and where they are likely to impact as
they evolve. As community resilience research has shown, awareness of hazards is not the
only factor in triggering actual preparedness actions; however it is one significant driver
(Paton 2006, Paton & Johnston 2001, Ronan & Johnston 2005).

Severe weather RADAR, GOES, NOAA, Provide valuable advanced warning
warnings broadscale visible Meteosat of severe events to the public and
and infra red emergency planners via
imagery meteorologists
Movement and InSAR and PS- ERS-1/2, ENVISAT Rate of movement for slow moving
ground InSAR ASAR, ALOS landslides. Often acceleration of
deformation PALSAR deformation rates means that a large
event is about to follow. Early
detection of deformation in volcanic
regions is used for forecasting of
possible eruptions
Soil moisture Long wavelength SMAP Water shortage leading to drought
SAR and agricultural productivity
decline, ability of soils to retain
water to indicate flood and landslide
potential
Ground Thermal imagery, ASTER, MODIS, Monitoring heating and cooling
temperature or SWIR in the AVHRR cycles of volcanoes to understand
variability case of very hot pre-eruptive characteristics for
features forecasting purposes
Coastal and SONAR, Laser LADS, Topex Tsunami hazard modelling
bathymetric depth ranging Poseidon / Jason
mapping
Display and Moderate to high Aerial photography, For use in public education about
advertisement of resolution optical Quickbird, Ikonos - hazards and risks to foster greater
potential hazards imagery, often usually using black readiness of individuals, households
overlaying a and white or true and organisations
DEM colour composites for
ease of understanding Use in civil defence emergency
management exercises to provide
realistic scenarios that will assist
with staff professional development
and planning
Detecting sea Broad scale MODIS, GOES, Advance warning of severe weather
temperature or thermal imagery, AVHRR approaching to commence
atmospheric geostationary
pressure change in
cyclone/hurricane/
typhoon generating
latitudes
Table 2. Examples of information and data requirements during the readiness phase
At the institutional level, a strong focus is placed on the development of plans and
relationships. A primary way to test the effectiveness of these preparedness plans and
relationship functions is through civil defence emergency management exercises. In order
for exercises to provide an effective learning experience for participants, realistic hazard
scenarios must be developed. Remotely sensed data can assist this process through the
creation of hazard maps, providing realism to exercise injects (new information about
hazards or consequences as the exercise plays out).
At local to national scales, obtaining an overall picture of the hazardscape; identifying at risk
areas, and priority hazards for resources and planning is essential. Granger (2000) discusses
the development of information infrastructure for disaster management in Pacific island
nations, based on remotely sensed data, and GIS interpretation. For countries with limited
budgets, collaboration to purchase remotely sensed data for disaster planning is beneficial
because of cost savings, the opportunities for skill and process sharing, and the consistency
of data for modelling (Granger 2000).
As discussed previously, hazard modelling is important for risk reduction (section 4.1); it is
also important for readiness, as for many hazards residual risk dictates that an effective
emergency response will be the most practical solution for disaster management. For
example, New Zealand has several active volcanoes; Mt Ruapehu is the largest of these.
Ruapehu is a national park and has two commercial ski fields in operation on its slopes.
Depending on the time of year, visitors to the mountain are engaged in a variety of
recreational, educational and scientific activities. The greatest hazards associated with the
volcano are eruptive events and lahar flow (Carrivick et al. 2009). The volcano has a crater
lake at the summit which produces periodic large lahars during eruptions and tephra dam
bursts. These lahars follow channels which are bridged by the main trunk railway line and
State Highway 1, as well as passing through ski field and hiking areas. A lahar event in
1953, before bridges were raised and strengthened, destroyed the Tangiwai rail bridge, and
a passenger train unable to stop was derailed resulting in the death of 151 people. While
bridges have been modified to reduce risk, considerable readiness planning has also been
undertaken to ensure that the events such as the 1953 disaster cannot happen again (Galley
et al. 2004).
Following eruptions in 1995 and 1996 a large tephra dam formed on the crater rim allowing
the lake to fill to higher than normal levels. The volcanic rocks of Crater Lake rim now had a
weakness, a section of the rim comprised of weaker tephra, which would fail when lake
levels reached a certain height. Extensive modelling of potential lahar flow paths and
velocities was undertaken based on high resolution remotely sensed data (Carrivick et al.
2009). The path was verified using aerial photography, LiDAR, ASTER and PALSAR
imagery after the event (Joyce et al. 2009b). The modelling provided the necessary hazard
information for authorities to manage the risk through a suite of preparedness activities. A
bund (levee) has been constructed to prevent lahar flow onto the main highway; and a
comprehensive monitoring and alarm system was constructed to detect lahar break outs. An
integrated response plan involving emergency managers, police, the fire service, road
managers, railways operators, ski field staff, scientists and national park managers, was
developed to stop all trains outside the hazard zone, close the highway, trigger warnings
and response plans at the ski fields (move to ridges away from flow paths) (Leonard et al.
2005), and locate and evacuate any hikers or workers in hazard zones within the national
park. The tephra dam burst early in 2007, and the response based on high quality modelling
went as planned. The lahar was of considerable size but remained within expected channels
and the only significant damage was to an unoccupied public toilet building at the Tangiwai
memorial site.
Lahar flows and eruptions remain an ongoing hazard at Ruapehu. To assist with
preparedness for these hazards, remote sensing is part of the suite of monitoring systems
employed to detect changes in volcanic activity. A combination of synthetic aperture radar,
ASTER thermal imagery (Figure 3), and OMI UV/visible imagery is acquired on a routine
basis for monitoring deformation, Crater Lake temperatures and gaseous emissions
respectively.
Fig. 3. Thermal monitoring of Mt Ruapehu. (a) SPOT-5 image obtained 15 March 2008
demonstrates land cover for contextual purposes; (b) Average temperature image calculated
from night-time ASTER thermal data between 3 September 2007 and 28 September 2008;
and (c) Mt Ruapehu Crater lake subsets using ASTER night-time thermal data. Note the
temperature scale change for illustrative purposes.
The use of remotely sensed data of a previous event can be used in this phase to constrain
geophysical models and help provide realistic scenarios for future events. For example,
InSAR can be used to examine the deformation effects of a single event (such as an
earthquake) by acquiring only two images as close in time as possible, one before and one
after the event. Using this technique, the PALSAR L-band sensor on board the ALOS
satellite was successfully used to map co-seismic deformation of a magnitude 6.7 earthquake
in the vicinity of George Sounds, off the coast of the lower South Island on 16th October 2007
(Petersen et al. 2009 in review). After processing two PALSAR images (22 July and 22
October 2007) displacements were apparent in the coastal region closest to the epicentre
(Figure 4). Landslides were also experienced in the area (though not evidenced in this
figure). The long wavelength L-band is of particular use in this region due to its ability to
penetrate dense vegetation to retrieve the ground signal. The amount and location of
deformation is used in modelling studies to estimate earthquake parameters in order to
learn more about the tectonics of this remote region. As this is an uninhabited area of New
Zealand, there was no observed infrastructure damage that may have otherwise
necessitated acquisition of high resolution optical imagery for response or recovery
purposes.
Fig. 4. Ground deformation following George Sounds earthquake in October 2007.

Background image is a Landsat 7 ETM+ true colour composite scene
Monitoring longer term ground deformation effects such as that produced by ground water
extraction, volcanic activity or slow moving landslides is conducted using multiple SAR
images over a period of time. Using this technique it is possible to detect sub centimetre
scale ground movement over large areas that could otherwise only be monitored or detected
using networks of in-situ GPS. With this method, the C-band sensor on board the ENVISAT
satellite was able to detect sub-centimetre deformation in the Auckland region (Figure 5).
This figure was created using a stack of 117 images, spanning the period 17 July 2003 and 9
November 2007. InSAR is used in this manner for long term monitoring and produces a rate
of change over time. It is believed that most of the observed InSAR signal shown here is
caused by extraction of groundwater; however the link to volcanic activity has also been
investigated (Samsonov et al. 2009 in review).
Fig. 5. Monitoring uplift and subsidence in Auckland. Background image is a green band
grey scale mosaic of SPOT-5 and Landsat 7 ETM+ imagery.
4.3 Response
Response activities are primarily focussed on protecting life and property during disasters.
Activities such as evacuations, search and rescue, sandbagging along riverbanks, evaluating
building safety, establishing immediate emergency shelter, setting up command posts and
other short-term tasks fall into the response phase. Remote sensing can be used here to
provide immediate damage assessment if the data can be provided in a timely manner, and
also to assist evacuation plans through the combination of observing weather patterns and
hazard behaviour (e.g. fire front approaches, water level rises). Other examples of the use of
remote sensing during the response phase are given in Table. 3. Ideally, recovery activities
commence when the response phase begins, to ensure an integrated process for holistic
recovery. This means that damage assessments undertaken via remote sensing during the
response phase will also be integral to the recovery phase.
Type of Data required Sensor example Application example

information
Inundation SAR, optical Radarsat, SPOT, Determine magnitude, location and
ASTER Quickbird, duration of impacts.
Ikonos Use SAR when cloud cover is still
problematic
Widespread SAR, moderate – Radarsat, SPOT, Determine magnitude, location and
storm or high resolution ASTER Quickbird, duration of impacts.
earthquake optical Ikonos
induced
landslides
Volcanic ash and Shortwave infra GOES, TOMS/OMI, Highly temporally variable, so
gases red, thermal MODIS minimum of daily imagery required.
infrared Used for volcanic ash advisories and to
warn airlines of hazardous flight paths
Public High resolution Quickbird, Ikonos Assist those at risk to personalise
information optical imagery hazard threat
during events
Ship location SAR Terra SAR-X, Cosmo Locating ships in the ocean during
Sky-Med storm
Co-seismic and InSAR ERS-1/2, ENVISAT Confirming magnitude of earthquake
post-seismic ASAR, ALOS and forecasting possible aftershocks
deformation PALSAR
Table 3. Examples of information and data requirements during the response phase
During the response phase, the temporal relevancy of remote sensing information is crucial
to allow disaster managers to plan effective mitigation strategies on dynamic situations. In
the case of wildfire events, it is critical to have current and timely intelligence on the fire
location, fire-front, and fuel conditions. Near-real-time information allows the fire
management team to plan fire attack appropriately, consequently saving resources, time and
possibly lives. Concurrently, the information must be of sufficient spatial resolution to allow
detailed tactical assessments and decisions to be made on the wildfire condition, and be
spectrally-relevant to the phenomenon being observed or measured.
Despite the spectacular nature of imagery often captured during a disaster event, the use of
remote sensing during the response phase has experienced mixed levels of success,
particularly in the case of satellite platforms. Regional scale imagery of effects associated
with the development of fire fronts (hot spot detection), volcanic eruptions (gas and ash
emissions), or tropical cyclones (inundation) is generally successful where the area of impact
is sufficiently large. For example, the wildfire management agencies in the United States
currently utilize thermal-infrared (TIR) satellite data provided by MODIS to provide
synoptic, 2-4 times-daily hot-spot detection of fire at continental scales (U.S. Forest Service
2009). The spatial resolution of MODIS is low / moderate (1000 meters), and is used to
derive a regional estimate of fire distribution. Although the temporal frequency of the
MODIS data is sufficient for regional fire assessment, its spatial resolution is insufficient for
more localised events, or for assessing the specific on-ground impact. Conversely, polar
orbiting satellites with appropriately high spatial resolution generally do not have the
overpass frequency or data relay capability to provide imagery quickly enough to be of use
for immediate response. The space science community is attempting to address this issue
with the launch of satellite constellations such as Rapid Eye and the Disaster Monitoring
Constellation (International collaboration between Algeria, China, Nigeria, Turkey and the
UK). There are also avenues for collaboration between international organisations for data
acquisition and provision in the event of disasters, such as the International Charter for
Space Based Disasters (Ito 2005), and Sentinel Asia (Kaku et al. 2006). While potentially
providing a considerable amount of data, neither of these tools can yet be used for
immediate or first response due to the current time delay between requesting and receiving
data. As such, research into airborne platforms has proven to be of greater utility for rapid
data and information provision.
In 2006, 2007 and 2008, the National Aeronautics and Space Administration (NASA) and the
U.S. Forest Service collaborated to evaluate and demonstrate the use of long-duration, large
Unmanned Airborne Systems (UAS), innovative sensing systems, real-time onboard
processing, and data delivery and visualisation technologies to improve the delivery and
usefulness of remote sensing data on wildfire events. The objectives were to demonstrate the
capabilities of providing sensor-derived, GIS-compatible, geo-rectified, processed data on
wildfire conditions to incident management teams within 15-minutes of acquisition from the
sensors on the UAS. The characteristics of this system render it ideal for emergency response
that is not just isolated to wildfire events.
During the 2006, 2007 and 2008 U.S. wildfire season, a series of missions were flown over
wildfires in the western U.S. to demonstrate the integration of the above-mentioned
technologies to provide near-real-time information to disaster managers. The missions were
flown on the NASA Ikhana, a modified General Atomics – Aeronautical Systems, Inc.
Predator-B (MQ-9) Unmanned Aerial Vehicle (UAV), designed specifically for supporting
NASA science missions. The Ikhana is capable of medium / high altitude and long-duration
(24-hours) operations, making it an ideal platform for disaster event monitoring. The Ikhana
UAS flew missions with the NASA AMS-Wildfire sensor onboard, which can be remotely
operated and provides autonomous data processing capabilities (Ambrosia & Wegener
2009).
The use of the Ikhana and accompanying systems has proven successful over a number of
events. In October 2007, four missions were flown over the Santa Ana wildfires in a five-day
period (Figure 6a) and the resultant information was used to deploy fire fighting resources.
In late June 2008, lightning storms in northern California ignited thousands of fires, that
grew together to become over 25 major incidents covering millions of acres of forestlands.
The national airborne remote sensing assets were overwhelmed and with a state emergency
declared, the Ikhana and AMS-Wildfire were requested to support wildfire data collection
operations. During the remainder of the summer, the Ikhana flew four missions in California,
providing near-real-time data on numerous wildfires. The AMS-Wildfire real-time data
were used effectively to locate a major fire surge encroaching on Paradise, California (Figure
6b). The data was used to support the emergency evacuation decision of the entire
population of the community, an effective demonstration of the criticality of near-real-time
remote sensing information supporting disaster management operations.
Fig. 6. (a) Flight routes required to cover 11 major wildfires California over four days in
October 2007; and (b) AMS-Wildfire 3-band graphic image overlay and fire hot-spot detects
(yellow areas) of the Canyon Complex fire approaching Paradise, California. The data was
acquired on 8 July 2008. The hot-spot detect data, showing the fire moving rapidly towards
Paradise, assisted in the evacuation determination for residents in the vicinity. This north-
viewing 3-D data is displayed on Google Earth background information.
One of the key factors to the success of this system is the provision of not only data, but
information that can be ingested and utilised immediately by emergency managers to aid
their decision making. Part of this speed of information delivery is attributed to the
autonomous processing onboard the UAS to create geo-rectified image raster products
(GeoTIFF) and hot-spot detection vector files (.shp files). An emergency situation is not the
time to be experimenting with new algorithms or processing techniques, thus it is necessary
to ensure robust techniques have been thoroughly trialled and considered operational pre-
event (Joyce et al. 2009a). The vector and raster products generated with this system are
transmitted via the Ikhana telemetry system, through a communications satellite to servers
on the ground, where they are automatically processed into Keyhole Markup Language
(KML) files, compatible with Google Earth and made available in near-real-time at NASA
servers. The combination of the near-real-time imagery and the simple Google Earth
visualisation capabilities are a powerful tool that requires minimal (or no) training in its
employment. Embedding a remote sensing specialist within the emergency management
team can further assist with data integration, information understanding, and fielding
specialized requests.
Although the Western States UAS Fire Imaging Missions were focused on demonstrating
remote sensing capabilities to wildfire management entities, they resulted in direct
emergency support to national incidents in all three years. Those missions allowed a
comprehensive assessment of the technologies and resulted in the adaptation and
integration of various components into operational use. The key components to the
“usefulness” of the data were the timeliness of the data (from acquisition to product
delivery) and the simple format which the data was available for visualisation and decision-
making. While these factors are important at all stages of the disaster management cycle,
they become particularly critical during the response phase, where rapid decision making is
most important. The provision of simple hotspot information also means that the emergency
management team is not overwhelmed with too much data or too many visualisation
options. The choice of using Google Earth as a “front-end” display of the data was a careful
decision to provide information in a format and software system that was easily operated
and readily available to the fire management community. Fire Incident Command team
members do not have the time to “learn” new software capabilities or new tools while they
are in the midst of a major wildfire management activity. Google Earth provided a user-
friendly capability to allow quick data integration, zoom capabilities, 3-D visualisation and
ease of use.
The use of UAVs presents opportunities as well as risks. UAVs provide increased range and
flight time and the ability to penetrate environments that might be too hazardous for piloted
aircraft (Henson 2008). Mission and platform costs currently precludes immediate
adaptation of UAS systems by disaster management agencies, but the disaster support
missions we showcased are major steps forward in demonstrating UAS utility and sensor
and processing capabilities available right now! These technologies need not be considered
for use only with unmanned vehicles, but can be adopted for piloted aircraft, and hopefully
for satellite platforms in the future. Autonomous onboard processing has been trialled with
Hyperion for identifying hotspots associated with volcanic eruptions (Davies et al. 2006),
though the challenge remains to progress these techniques to operational status.
4.4 Recovery
The use of remote sensing to aid or monitor disaster recovery is perhaps the least developed
application of this technology. However, this is an area where the remote sensing
community could contribute a great deal through the provision of objective time series
analysis over large areas with both high and medium levels of spatial detail. In other
specialisations, time series analysis of remotely sensed data is an established technique.
Environmental applications such as deforestation and urban sprawl are common targets. In
each case, the monitoring objective is clear. In disaster recovery, there are often some very
clear indicators that can easily be measured and monitored with remote sensing imagery.
Some of these indicators include construction and subsequent removal of medium and long-
term emergency shelters; debris removal; commencement and completion of new
construction or reconstruction (buildings, bridges, roads); vegetation regrowth; and
reduction of siltation from waterways after flooding events (Table 4).
Rate of recovery e.g. Moderate to very Aerial photography, Compare the effectiveness of
debris removal, high resolution Quickbird, different recovery strategies;
vegetation regrowth, imagery in a Worldview, Ikonos Determine if aid funding is being
reconstruction continuous time used appropriately; Wildlife
series habitat recovery (eg after fire);
Identify ‘residual risk’ – areas not
recovered are more vulnerable to
future events
Infrastructure and Very high Aerial photography, Create new baseline maps
facilities locations resolution Quickbird,
imagery Worldview, Ikonos
Revised DEM InSAR, LiDAR ERS-1/2, ENVISAT Necessary after large earthquake
ASAR, ALOS or volcanic eruption if the local
PALSAR and regional elevation changes
Status Quo Very high Aerial photography, Plan areas for funding allocation
resolution Quickbird,
imagery Worldview, Ikonos
Table 4. Examples of information and data requirements during the recovery phase
Using high spatial resolution the amount of housing reconstruction can at least be visually
identified by the presence and absence of blue tarpaulins covering roofs following
Hurricane Katrina (Hill et al. 2006). Conceivably an automated detection method could be
developed to identify these quickly and repeatedly in a time series dataset. The authors also
provide a list of other recovery related features observable over time with Quickbird data. In
Figure 7, the progression of recovery in a small area of New Orleans can be seen with high
resolution data. Notable features in the image acquired a week before the hurricane are a
large car park, sporting fields, and residential housing (Figure 7a). The progression clearly
shows inundation in this area (Figure 7b), and remaining sediment shortly after the water
subsidence. By March 2006, temporary housing is evident in the location of the car park, and
is still visible three years after the event, though the number of roofs covered in blue
tarpaulins has decreased. An analysis of the relative rate of change is given in Figure 7k,
demonstrating that impervious surfaces and lines of communication such as roads moved
towards recovery quite quickly after the event, while mature vegetation takes somewhat
longer. Some roofing damage and a swimming pool appear to remain in an unrepaired state
three years after the event. The key here is that a time series of data is vital to determine if
any change is occurring, and to further extract rates of change.
Recovery rates following a widespread landsliding event in northern New Zealand can also
be seen from a series of SPOT-5 and ALOS AVNIR-2 imagery (Figure 8). Here the
landsliding is apparent as bright scars in the colour infra red imagery acquired four months
after the event (Figure 8b). One year later, recovery of many of the grassy slopes on the
eastern portion of the image can be seen, while the landslides in the western region are also
becoming overgrown (Figure 8c). This recovery becomes even more apparent in the series of
NDVI images, which highlight the contrast between landslides (black) and surrounding
vegetation (various shades of grey) (Figure 8d-f). In an area that was covered with many
thousand landslides (Joyce et al. 2008a), satellite remote sensing is the only time and cost
effective manner of data collection for understanding recovery in the area. Similar
techniques could be used to look at native habitat regeneration following bushfires.
Remote sensing and the disaster management cycle
Fig. 7. Time series of high resolution imagery in New Orleans (a) Before Hurricane Katrina; (b) Soon after the height of the storm;
(c-j) Various time intervals following the recovery process; and (k) Interpreted rate of recovery. Images are © Digital Globe and
Google Earth 2009.
339
Fig. 8. Recovery of vegetation after a widespread landsliding event in northern New

Zealand, July 2007. (a) SPOT-5 CIR obtained before the event; (b) SPOT-5 CIR obtained
shortly after the event; (c) ALOS AVNIR-2 CIR imagery obtained one year later; and (d-e)
NDVI images of the aforementioned data.
Analysis of time series imagery could also help to monitor the effectiveness of different
recovery strategies. By extracting recovery rates from data acquired at appropriate time
intervals, this assessment could help guide recovery plans for future events of a similar
nature. This would also help identify areas of residual risk that require ongoing monitoring
until the physical recovery process completed.
5. Conclusions
Remote sensing can be used to inform many aspects of the disaster management cycle. An
exhaustive coverage of all potential applications would be impossible in a single book
chapter, however we have shown several good examples from which inspiration can be
sought for future use. It is important to consider all aspects of disaster management, rather
than focussing on emergency response. By incorporating remote sensing into reduction and
readiness activities, this can also educate both emergency management staff and the
community about this type of information so that they are familiar with its use under a
response and inherently pressured situation.
The key elements to facilitate the usefulness of remote sensing data in support of the
disaster management community are being able to provide the appropriate information in a
spectrally, temporally, and spatially relevant context. Additionally, one must be aware of
the information requirements of that disaster management community, and tailor the
remote sensing information to meet those needs. That can only come through close
collaborations between the disaster management community and the remote sensing /
geospatial community.
6. Acknowledgements
All SPOT 5 imagery used in this chapter is © CNES. This manuscript incorporates data
which is © Japan Aerospace Exploration Agency ("JAXA") (2008). The data has been used in
this manuscript with the permission of JAXA and the Commonwealth of Australia
(Geoscience Australia) ("the Commonwealth"). JAXA and the Commonwealth have not
evaluated the data as altered and incorporated within the manuscript, and therefore give no
warranty regarding its accuracy, completeness, currency or suitability for any particular
purpose. Environment Bay of Plenty provided the licence to use the LiDAR data. Thank you
to Andy Gray for assistance with graphics and to Phil Glassey and David Johnston for
chapter review.
7. References
Ambrosia V. G. & Wegener S. S. (2009). Unmanned airborne platforms for disaster remote
sensing support. In: Advances in Geoscience and Remote Sensing, 317-346 - this book,
IN-Tech Publishing, Vienna
Barnes P. (2002). Approaches to community safety: risk perception and social meaning.
Australian Journal of Emergency Management, 17, 1, 15 - 23
Becker J., Saunders W., Hopkins L., Wright K., & Kerr J. (2008). Pre-event recovery planning
for use in New Zealand: An updated methodology. GNS Science Report 2008/11.
GNS Science
Begg J. G. & Mouslpoulou V. (2009 in press). Analysis of late Holocene faulting within an
active rift using lidar, Taupo Rift, New Zealand. Journal of Volcanology and
Geothermal Research
Board on Natural Disasters (1999). Mitigation emerges as major strategy for reducing losses
caused by natural disasters. Science, 284, 1943-1947
Burby R. J. (2001). Involving citizens in hazard mitigation planning: Making the right
choices. Australian Journal of Emergency Management, 16, 45 - 51
Burby R. J. (1998). Natural hazards and land use: An introduction. In: Cooperating With
Nature: Confronting Natural Hazards with Land Use Planning for Sustainable
Communities, Burby RJ (Ed.), 1 - 26, Joseph Henry Press, Washington D.C.
Carrivick J. L., Manville V., & Cronin S. (2009). A fluid dynamics approach to modelling the
18th March 2007 lahar at Mt. Ruapehu, New Zealand. Bulletin of Volcanology, 71,
153-169
Cloude S. R. & Papathanassiou K. P. (1988). Polarimetric SAR Interferometry. IEEE
Transactions on Geoscience and Remote Sensing, GE36, 1551 - 1565
Cutter S. L. & Finch C. (2008). Temporal and spatial changes in social vulnerability to
natural hazards. Proceedings of the National Academy of Sciences of the United States of
America, 105, 7, 2301-2306, 00278424
Davies A. G., Chien S., Baker V., Doggett T., Dohm J., Greeley R., Ip F., Castano R., Cichy B.,
Rabideau G., Tran D., & Sherwood R. (2006). Monitoring active volcanism with the
Autonomous Sciencecraft Experiment on EO-1. Remote Sensing of Environment, 101,
4, 427
Donner W. & Rodriguez H. (2008). Population Composition, Migration anti Inequality: The
Influence of Demographic Changes on Disaster Risk and Vulnerability. Social
Forces, 87, 2, 1089-1114, 00377732
Elachi C. (1987). Introduction to the Physics and Techniques of Remote Sensing, John Wiley &
Sons, New York
Franceschetti G. & Lanari R. (1999). Synthetic Aperture Radar Processing, CRC Press, New
York
Galley I., Leonard G. S., Johnston D., Balm R., & Paton D. (2004). The Ruapehu lahar
emergency response plan development process: An analysis. Australian Journal of
Disaster and Trauma Studies, 1
Giglio L., Csiszar I., Restás Á., Morisette J. T., Schroeder W., Morton D., & Justice C. O.
(2008). Active fire detection and characterization with the advanced spaceborne
thermal emission and reflection radiometer (ASTER). Remote Sensing of
Environment, 112, 6, 3055
Godschalk D. R., Kaiser E. J., & Berke P. R. (1998). Integrating hazard mitigation and local
land use planning. In: Cooperating With Nature: Confronting Natural Hazards with
Land Use Planning for Sustainable Communities, Burby RJ (Ed.), pp 85-118, Joseph
Henry Press, Washington D.C.
Granger K. (2000). An information infrastructure for disaster management in Pacific Island
Countries. Australian Journal of Emergency Management, 15, 1, 20 - 32
Gregg C. E. & Houghton B. F. (2006). Natural Hazards. In: Disaster Resilience: An Integrated
Approach, Paton D, Johnston D (Eds.), pp 19-39, Charles C. Thomas Publishers Ltd,
Springfield
Hancox G. T. H. & Wright K. (2005). Analysis of landsliding caused by the 15-17 February
2004 rainstorm in the Wanganui-Manawatu hill country, southern North Island,
New Zealand GNS Science report. Institute of Geological and Nuclear Sciences (64)
Hanssen R. F. (2001). Radar interferometry: data interpretation and error analysis, Kluwer
Academic Publishers
Henson R. (2008). Satellite observations to benefit science and society: recommended missions for the
next decade, National Research Council, National Academies Press, National
Acadamy of Science, 0-309-10904-3
Hill A. A., Keys-Mathews L. D., Adams B. J., & Podolsky D. (2006). Remote sensing and
recovery: A case study on the Gulf Coasts of the United States, Proceedings of the
Fourth International Workshop on Remote Sensing for Post-Disaster Response,
Cambridge, UK
Ito A. (2005). Issues in the implementation of the International Charter on Space and Major
Disasters. Space Policy, 21, 2, 141
Joyce K. E., Belliss S., Samsonov S., McNeill S., & Glassey P. J. (2009a). A review of the status
of satellite remote sensing image processing techniques for mapping natural
hazards and disasters. Progress in Physical Geography, 33, 2, 1 - 25
Joyce K. E., Dellow G. D., & Glassey P. J. (2008a). Assessing image processing techniques for
mapping landslides, Proceedings of the IEEE International Geoscience and Remote
Sensing Symposium, Boston, MA
Joyce K. E., Samonsov S., & Jolly G. (2008b). Satellite remote sensing of volcanic activity in
New Zealand, Proceedings of the 2nd Workshop on the use of remote sensing techniques
for monitoring volcanoes and seismogenic areas, Naples, Italy
Joyce K. E., Samsonov S., Manville V., Jongens R., Graettinger A., & Cronin S. (2009b).
Remote sensing data types and techniques for lahar path detection: A case study at
Mt Ruapehu, New Zealand. Remote Sensing of Environment, 113, 1778 - 1786
Kaku K., Held A. A., Fukui H., & Arakida M. (2006). Sentinel Asia initiative for disaster
management support in the Asia-Pacific region, Proceedings of the SPIE,
Laska S. & Morrow B. H. (2006/7). Social vulnerabilities and Hurricane Katrina: An
unnatural disaster in New Orleans. Marine Technology Society Journal, 40, 4, 16-26
Leonard G. S., Johnston D. M., & Paton D. (2005). Developing effective lahar warning
systems for Ruapehu. Planning Quarterly, 158, 6-9
Massonnet D. & Feigl K. (1998). Radar interferometry and its application to changes in the
Earth surface. Reviews of Geophysics, 36, 4, 441 - 500
Mileti D. S. (1999). Disasters By Design: A reassessment of Natural hazards in the United States,
Joseph Henry Press, Washington D.C.
Nirupama N. & Simonovic S. (2007). Increase of Flood Risk due to Urbanisation; A
Canadian Example. Natural Hazards, 40, 25 - 41
Norman S. (2004). Focus on Recovery: A Holistic Framework for Recovery. In: NZ Recovery
Symposium Proceedings 12-13 July, 2004., Norman S (Ed.), 31-46, Ministry of Civil
Defence and Emergency Management, Wellington
Oppenheimer C. (1993). Infrared surveillance of crater lakes using satellite data. Journal of
Volcanology and Geothermal Research, 55, 117 - 128
Oppenheimer C. (1997). Remote sensing of the colour and temperature of volcanic lakes.
International Journal of Remote Sensing, 18, 1, 5-37
Paton D. (2006). Disaster resilience: Integrating individual, community, institutional and
environmental perspectives. In: Disaster resilience: An integrated approach, Paton D,
Johnston D (Eds.), 320, Charles C Thomas Publisher, 0-398-07663-4, Springfield
Paton D. & Johnston D. (2001). Disaster and communities: vulnerability, resilience and
preparedness. Disaster Prevention and Management, 10, 270-277
Pertrow T., Thieken A., Kreibich H., Merz B., & Bahlburg C. H. (2006). Improvements on
flood alleviation in Germany: Lessons learned form the Elbe flood in August 2002.
Environmental Management, 38, 717-732
Petersen T., Beaven J., Denys P., Denham M., Field B., Francois-Holden C., McCaffrey R.,
Palmer N., Reyners M., Ristau T., Samonsov S., & Team T. G. (2009 in review). The
Mw 6.7 George Sound earthquake of October 16, 2007: response and preliminary
results. Bulletin of New Zealand society for earthquake engineering
Phinn S. R. (1998). A framework for selecting appropriate remotely sensed data dimensions
for environmental monitoring and management. International Journal of Remote
Sensing, 19, 17, 3457 - 3463
Phinn S. R., Joyce K. E., Scarth P. F., & Roelfsema C. M. (2006). The role of integrated
information acquisition and management in the analysis of coastal ecosystem
change. In: Remote sensing of aquatic coastal ecosystem processes: Science and
management applications, Richardson LL, LeDrew EF (Eds.), 325, Springer-Verlag
Pottier E. & Ferro-Famil L. (2008). Advances in SAR polarimetry applications exploiting
polarimetric spaceborne sensors, Proceedings of the 2008 Radar Conference, pp. 1 - 6,
Rome, Italy
Quarantelli E. L. (1997). Ten criteria for evaluating the management of community disasters.
Disasters, 21, 1, 39-56
Quarantelli E. L. (1985). What is Disaster? The Need for Clarification in Definition and
Conceptualisation in Research. In: Disasters and Mental Health Selected Contemporary
Perspectives, Sowder B (Ed.), 41-73, Government Printing Office, Washington
Ronan K. R. & Johnston D. M. (2005). Promoting community resilience in disasters : the role for
schools, youth, and families, Springer, 0387238204, New York
Rosen P., Hensley P., Joughin I., Li F., Madsen S., Rodriguez E., & Goldstein R. (2000).
Synthetic aperture radar interferometry. Proceedings of the IEEE, 88, 3, 333-382
Samsonov S. V., Tiampo K., Gonzalez P., Manville V., & Jolly G. (2009 in review). Modelling
deformation occurring in the city of Auckland, New Zealand mapped by
differential synthetic aperture radar. Journal of Geophysical Research
Shrestha R. L., Carter W. E., Satori M., Luzum B. J., & Slatton K. C. (2005). Airborne Laser
Swath Mapping: Quantifying changes in sandy beaches over time scales of weeks
to years. ISPRS Journal of Photogrammetry & Remote Sensing, 59, 222 - 232
Slovic P., Fischhoff B., & Lichtenstein S. (2000). Cognitive Processes and Societal Risk
Taking. In: The perception of risk, Slovic P (Ed.), 32-50, Earthscan Publications Ltd,
London
Trunk L. & Bernard A. (2008). Investigating crater lake warming using ASTER thermal
imagery: Case studies at Ruapehu, Poás, Kawah Ijen, and Copahué Volcanoes.
Journal of Volcanology and Geothermal Research, 178, 2, 259 - 270
van Zyl J. J., Zebker H. A., & Elachi C. (1990). Polarimetric SAR applications. In: Radar
Polarimetry for Geoscience Applications, 315 - 360, Artech House, Nomood
Wisner B. (2004). Assessment of capability and vulnerability. In: Mapping Vulnerability:
Disasters, Development and People, Bankoff G, Frerks G, Hilhorsdt D (Eds.), 183 - 193,
Earthscan, London
Wisner B., Blaikie P., Cannon T., & Davis I. (2004). The challenge of disasters and our
approach. In: At Risk: Natural Hazards, people's vulnerability and disasters, Wisner B,
Blaikie P, Cannon T, Davis I (Eds.), 3-48, Routledge, London & New York
Wright R., Flynn L., Garbeil H., Harris A., & Pilger E. (2002). Automated volcanic eruption
detection using MODIS. Remote Sensing of Environment, 82, 1, 135
Wright R., Flynn L. P., Garbeil H., Harris A. J. L., & Pilger E. (2004). MODVOLC: near-real-
time thermal monitoring of global volcanism. Journal of Volcanology and Geothermal
Research, 135, 1-2, 29-49
8. Glossary
Note that the satellite sensors listed here and within the text are simply examples of the
types of instruments that can be used, rather than being a complete listing of all possibilities.
ALOS AVNIR-2 Japanese Space Agency (JAXA) Advanced Land Observing Satellite
Advanced Visible and Near Infrared sensor. Useful for local to
regional scale mapping and monitoring
ALOS PALSAR Japanese Space Agency (JAXA) Advanced Land Observing Satellite L
Band SAR satellite. Useful for deformation monitoring in regions of
dense vegetation
ASAR C Band Advanced Synthetic Aperture RADAR
ASTER Advanced Spaceborne Thermal Emission and Reflection Radiometer
on board NASA’s Terra satellite. Useful for monitoring volcanic
activity
AVHRR Advanced Very High Resolution Radiometer (NOAA – National
Oceanic and Atmospheric Administration). Useful for regional to
national scale applications
CIR Colour Infrared – three band standard display of green, red and near
infrared light displayed as blue, green and red respectively
DEM Digital elevation model

DMC Disaster Monitoring Constellation. International collaboration
between space agencies in Algeria, China, Nigeria, Turkey and the
UK for regional scale mapping (optical)
ERS European Space Agency Satellite with a suite of SAR and optical
sensors
ENVISAT European Space Agency Satellite with a suite of SAR and optical
sensors
GOES Geostationary Operational Environmental Satellites – used for
metrological applications
Ikonos Very high spatial resolution commercial satellite (GeoEye). Useful for
local scale mapping and monitoring (e.g. buildings and assets)
InSAR Interferometric Synthetic Aperture RADAR – technique used for
measuring surface deformation
KML Keyhole Markup Language – native language for Google Earth files
LADS Laser Airborne Depth Sounder
Landsat ETM+ Enhanced Thematic Mapper plus. Useful for long term regional scale
mapping and monitoring, though technical malfunctioning limits
data coverage
Landsat TM Thematic Mapper. Useful for long term regional scale mapping and
monitoring
LiDAR Light Detection and Ranging. Used for creating very high spatial
resolution DEMs
Meteosat European geostationary meteorological satellite
MODIS Moderate Resolution Imaging Spectrometer. Used for hotspot
monitoring of fires and volcanic activity on a regional to continental
scale
NASA National Aeronautics and Space Administration
OMI Ozone Monitoring Instrument – used for monitoring volcanic gas
emissions
POLInSAR Polarimetric Interferometric Synthetic Aperture RADAR
PS-InSAR Permanent Scatterers Interferometric Synthetic Aperture RADAR
Quickbird Very high spatial resolution commercial satellite (Digital Globe).
Useful for local scale mapping and monitoring (e.g. buildings and
assets)
RADARSAT Canadian Space Agency C Band SAR satellite
RapidEye Constellation of five high resolution optical satellites, the
combination of which provides a daily revisit capability
SAR Synthetic Aperture RADAR. Active sensor capable of capturing data
through clouds, smoke and haze
SMAP Soil Moisture Active Passive – scheduled for launch in 2012
SPOT Satellite Pour l’Observation de la Terre – French Space Agency
(CNES – Centre National d’Etudes Spatiales) colour infrared and
panchromatic earth observation satellite
SWIR Short Wave Infrared. Used for volcanic ash and gas monitoring and
also vegetation applications
Terra SAR-X X-band SAR satellite
TIR Thermal Infrared, used for fire and volcanic activity monitoring
Topex Poseidon Joint CNES / NASA satellite altimetry mission, used for studying sea
level, ocean bathymetry, tides and ocean currents (now succeeded by
Jason)
UAS Unmanned Airborne System
UAV Unmanned Aerial Vehicle
UV Ultra Violet – non-visible short wavelength radiation, useful for
volcanic gas estimation
Worldview Very high spatial resolution commercial satellite (Digital Globe).
Useful for local scale mapping and monitoring (e.g. buildings and
assets). Currently only panchromatic.
Prediction of Volumetric Shrinkage in Expansive Soils (Role of Remote Sensing) 347
16
X
Prediction of Volumetric Shrinkage in

Expansive Soils (Role of Remote Sensing)
Fekerte Arega Yitagesu, Freek van der Meer and Harald van der Werff
International Institute for Geo-information Science and Earth Observation, ITC
The Netherlands
1. Introduction
Life time, performance and environmental compatibility of civil engineering infrastructure
largely depend on the quality of geotechnical investigations. Apart from being the basis for
much of the costing (in terms of time and money both at the construction and maintenance
stages) of engineering works, safety of structures lies on information from a geotechnical
survey. Existence of expansive soils in construction sites is a great problem that need due
consideration in geotechnical investigations.
Expansive soils are weak and unstable when subjected to moisture content fluctuations
either due to seasonal climatic variations (cyclic dry and wet periods) or artificial causes.
Their presence is one of the crucial factors that can significantly impact engineering costs for
it may cause major deterioration and distresses on lightweight and shallowly founded
structures. Primary characteristics of expansive soils are their potential to swelling and
shrinkage in response to moisture content increase and decrease respectively. These
properties are adverse in civil construction works for they pose a huge damage especially on
lightweight infrastructures (small buildings, roads and airport runway pavements, pipelines
and sewerage systems etc). An increase in volume (swelling) from expansive soils can
exceed the downward pressure exerted from lightweight structures and hence cause
deformations and development of cracks. Substantial decrease in volume (shrinkage) on the
other hand is responsible for uneven settlement. Expansive soils awe their characteristics to
their mineralogical assemblage; that is presence of active clay minerals and their amount.
Damages due to volume changes of expansive soils in form of swelling and shrinkage that
cost billions of dollars are reported in various parts of the world (Al-Rawas, 1999; Erguler
and Ulusay, 2003; Gourley et al., 1993; Nelson and Miller, 1992; Ramana, 1993; Shi et al.,
2002). Hence expansive soils should be identified and sufficiently characterized at early
stages of geotechnical applications in order to guide a detailed design survey so as to avoid
or minimize unnecessary expenses and delays in the construction and maintenance of
structures.
There are a number of direct and indirect in-situ and laboratory testing procedures available
to identify expansive soils and quantify the magnitude of volume change expected.
Expansive soils can be identified by surface manifestations in the field, for they show cracks
of polygonal pattern in dry seasons (Figure 1) as a result of appreciable volume decrease.
The depth of desiccation cracks are important in indicating the magnitude of volume change
since this depth represents the thickness in which moisture deficiency exists upon drying.
Various mineralogical identification methods such as X-ray diffraction (XRD), Transmission
Electron Micrograph (TEM), Scanning Electron Micrograph (SEM), Differential Thermal
Analysis (DTA), dye absorption and chemical analyses are important in research
laboratories for exploring the basic properties of clays. But they are costly and hence are not
commonly used in soil mechanics laboratories. Direct measurements include determination
of swelling and shrinkage potential of expansive soils. Use of consolidation apparatus and
triaxial methods are famous for measuring the swelling pressure that is required to
counteract the soil swell and swelling potential that can be exerted by the soil expansion.
These testing should be done in a sophisticated and controlled conditions with anticipated
environmental conditions fulfilled. The techniques give an opportunity of directly observing
the effects of soil expansion on different scale or magnitude of loadings that resembles the
actual conditions. Suction method can also be used to measure soil swell potential.
Volumetric shrinkage determination on the other hand provides with a measure of the
magnitude of shrinkage that the soil can undergo upon severe drying. These are the most
common and useful swell and heave prediction testing methods (Nelson and Miller, 1992).
However, it might take several days and loading steps before the swell pressure is
determined even for a single sample which in turn makes the methods expensive and
laboursome. Indirect means involve use of index parameters to identify and estimate
magnitude of swelling and shrinkage in expansive soils. Atterberg limits (Liquid limit,
plastic limit, plasticity indices, shrinkage limit and shrinkage indices) are the most popular
and frequently used index tests. Due to the simplicity of these tests and the good correlation
that they show with soil swell and shrinkage potentials, Atterberg limits are used in the
identification and classification of cohesive soils; as well as directly used in construction
specifications and standards (e.g. American Society of Testing Materials (ASTM), British
Standard Institution Specification etc) for quality controlling of materials that will be used in
fill, embankment and subbase constructions. Cation exchange capacity (CEC), Free swell,
Linear shrinkage, Coefficient of linear extensibility (COLE), expansion index (EI), California
bearing ratio swell (CBR swell) etc are also some of the index parameters. The more soil
testing that is done before hand, the easier it is to reduce risk in the design of infrastructure
and produce economically feasible as well as environmentally compatible structures.
However it is quite impractical to attempt collecting many samples over short distances and
analyze them for it is costly and time consuming.
Fig. 1. Typical expansive soil (black cotton soil of Ethiopia) with polygonal pattern cracks as
a manifestation of considerable shrinkage in dry season. These cracks are wide (on the order
of tens of centimeters) and also deep (on the order of meters).
Considerable amount of work has been done in the past to support geotechnical
investigations of expansive soils with remote sensing techniques (Chabrillat et al., 2002;
Goetz et al., 2001; Kariuki et al., 2004; Yitagesu et al., 2009). While Chabrillat et al., (2002)
and Goetz et al., (2001) demonstrated potential use of optical remote sensing data for
mapping abundances of clay minerals (the three clay species which are smectite, illite and
kaolinite that are important with respect to soil swell-shrink potential) responsible for soil
swell-shrink characteristics; Kariuki et al., (2004) established one to one relationships
between selected engineering parameters of expansive soils and absorption feature
parameters. Yitagesu et al., (2009) on the other hand found relationships between
engineering parameters (Atterberg limits, cation exchange capacity and free swell) and
laboratory acquired spectral reflectance of expansive soils; and indicated wavelength
regions to look into in attempting to extrapolate the approach to image datasets for
quantitative mapping of soil swell-shrink characteristics.
In this chapter potential use of remote sensing data in modeling volumetric shrinkage and
related index properties of expansive soils is illustrated. We presented:
 Possibility of classifying expansive soils with respect to dominant clay mineralogy which
is a primary controlling factor in soil swell-shrink characteristics, i.e. to qualitatively
characterizing expansive soils.
 Relationships of soil reflectance spectra with geotechnical properties of expansive soils

(shrinkage, plasticity and compaction characteristics; grading particularly clay content; and
strength or bearing capacity as evaluated by California Bearing Ration (CBR)). Mainly we
dealt with prediction of volumetric shrinkage of expansive soils from their respective
reflectance spectra.
 A multivariate calibration technique, partial least squares regression (PLSR) model for
estimating magnitude of volumetric shrinkage from laboratory acquired (ASD fullrange
spectrometer) soil reflectance spectra.
We demonstrated means of identifying soils susceptible to considerable swell and shrink,
and propose a simple way of estimating their volumetric shrinkage. Partial least squares
regression analysis method is used based on the assumption that volumetric shrinkage and
spectral signatures of soils are both a function of clay type and concentration in soil
specimens. Empirical relationship between soil volumetric shrinkage and spectral
reflectance is presented.
2. Methodology
2.1 Study area
The study area (Figure 2) is located in the central part of Ethiopia, in the upper valley of the
Awash River which drains the northern part of the Rift Valley.
Topography ranges from a relatively plain to hilly, undulating and steep mountainous
terrain; with elevation ranging from 1500 to 2500 meters above sea level. Conical-shaped
isolated hills of scoria which are products of gas rich mafic lava formed during the late
stages of volcanism are common in the study area.
Climate is moderate to wet with mean annual rainfall of 1200 mille meters, and temperature
ranging from 25 degree centigrade to 8 degree centigrade. While the temperature is high in
January, and from March to May, rainfall is heavy in July and August.
Geology (Abebe et al., 1999) at the start (around TuluDimtu) of the study area is influenced
by tertiary to quaternary volcanic formation which includes alkaline basalts, spatter and
cinder cones, ignimbrites, rhyolitic flows and domes, and trachyte. Near DebreZeyt the
formation is dominated by alluvial and lacustrine deposits which include sand, silt, clay,
diatomite and limestone. From DebreZeyt to Mojo town lacustrine deposits, and after modjo
town fall and poorly welded pyroclastic deposits dominate with ryolitic and trachytic
formations in between.
Soils in the study area are classified into vertisols, luvisols, leptosols, phaeozems and
andosols (Figure 1). According to FAO (1998) definitions vertisols are clay rich (smectitic)
expanding soils that swell and shrink with fluctuation in moisture content. Luvisols are
common soil types in flat or gently sloping land, derived from a variety of unconsolidated
material including alluvial, colluvial and eolian deposits. Leptosols are very shallow soils
over hard rock or in unconsolidated gravely material, and are most common in
mountainous areas. Phaeozems are soils that are predominantly derived from basic material
and are rich in organic matter. Andosols are young soils in volcanic regions that are usually
associated with pyroclastic parent materials.
Fig. 2. location map of the study area showing spatial distribution of various soil types,
sampling spots are shown following the new highway route to be constructed.
From engineering perspective, soil that are predominately black, highly plastic and
expansive clay (vertisol family which commonly termed as black cotton soil) are found
during the section from Addis Ababa to Modjo town covering extensive area with varying
thicknesses. Thick layers of black cotton soils, which make their susceptibility to volume
changes larger, are reported. In these soils prominent desiccation cracks (as shown in Figure
1) are evident in dry periods. While this is the case in the flat to rolling sections, the hilly to
mountainous terrains are mainly covered by fresh to partially weathered basalt with minor
rhyolitic composition.
Natural vegetation cover is in general poor since most of the area is farmland. An extensive
area is farmland with built-up areas following the existing road alignment from Addis
Ababa to Nazareth, being the major ones; Kaliti and Akaki (at the outskirts of Addis),
Dukem, DebreZeyt, Modjo and Nazareth itself. Deeply incised drainage patterns and gully
erosions are common features in the area particularly between Modjo and Nazareth.
2.2 Sampling and laboratory testing
Expansive soils mainly distribution of black cotton soils in the study area is so areally
extensive thus sufficient detailed characterization is required to aid in formulating proper
design and construction techniques. Particularly the amount of volume change (in the form
of swelling or shrinkage) should be predicted to eliminate or minimize its detrimental effect
on the highway subgrade and associated adverse environmental impacts. Therefore the
sampling and laboratory testing strategy was tailored to achieve this aim.
Soil samples were collected following the new express highway from its starting point near
TuluDimtu to its end at Nazareth town. The sampling was part of a comprehensive
investigation and testing scheme for assessing suitability and quality of subgrade material
for the newly proposed road alignment connecting Addis Ababa with Nazareth town.
Samples were taken from open pits of one to three meters depth which is commonly the
depth at which shallowly founded structures are laid. A total of thirty exploratory pits were
excavated.
2.3 Determination of geotechnical characteristics

Atterberg limits (liquid limits, plasticity limits and plasticity indices) were determined in
accordance with ASTM D4318-05 Standard test method for liquid limit, plastic limit and
plasticity indices of soils.
Volumetric shrinkages were determined in accordance with ASTM D4943-02 standard test
method for shrinkage factors of soils by the wax method.
Maximum dry density (MDD) and optimum moisture content (OMC) were determined in
accordance with ASTM D 689 standard test method for laboratory compaction
characteristics of soil using standard effort (12400ft-1bf/ft3 (600KN-m/m3)).
The strength or bearing value of the soil specimen was determined by measuring their
California bearing ratio (CBR). Accompanying CBR swell was also measured. The ASTM
D1883-05 Standard test method for measuring CBR of laboratory compacted soils is used.
Particle size distributions were determined in accordance with ASTM D6913-04e1 standard
test method for particle size distribution (gradation) of soils using sieve analysis (for the
fraction passing 2mm, 0.425mm and 0.075 mm ASTM sieve openings); and ASTM D422-
63(2007) standard test method for particle size analysis of soils (hydrometer analysis of
percent passing 2 μm ASTM sieve opening which is a clay fraction of the soils).
2.4 Soil reflectance measurement

ASD fieldspec full range spectrometer (http://www.asdi.com) that covers the 350
nanometer to 2500 nanometer wavelength region of the electromagnetic spectrum was used
to collect soil reflectance spectra. The measurement was done in a laboratory, using a
contact probe method.
Spectral resolution of the ASD fieldspec FR spectrometer is 3 nanometer for the 350
nanometer - 1000 nanometer region and 10 nanometer for the 1000 nanometer - 2500
nanometer region; whereas its spectral sampling interval is 1 nanometer (ASD, 1995). The
spectrometer has three separate detectors, one in the visible near infrared region (VNIR) and
the two are in the short wave infrared (SWIR) region. Since major absorption features
associated with clay minerals are found in the SWIR wavelength region of the
electromagnetic spectrum, clay minerals are among the mineral groups that are suitable for
ASD analysis. Additional spectral information can also be obtained from the VNIR
wavelength regions.
2.5 X-Ray Diffraction (XRD) analysis

Mineralogical composition of soil samples were examined using X-ray diffractometer (XRD).
The instrument used is Siemens D5000 diffractometer. Bulk (to determine the overall
constituents) as well as clay fractions (to quantify the major, minor and trace composition of
clay species) of the soil samples were analyzed.
2.6 A multivariate analysis Partial least squares regression (PLSR)

Partial least squares regression (PLSR) is regression by means of projections to latent
variables (Martens and Naes, 1989; Wold et al., 2001). The method was first proposed in the
1970’s, and is currently very popular in various disciplines, among which is spectroscopy. In
visible near infrared spectroscopy, PLSR has become a widely spread technology for
qualitative as well as quantitative analysis. This includes routine quality control activities; in
chemical, pharmaceutical and agro-industries, for it is found to be a fast, cheap, simple and
non-destructive technique with little or no sample preparation requirements.
PLSR is particularly important when dealing with a large number of variables that express
common information to avoid multicollinearity problems (Martens and Naes, 1989). It
reduces the impact of irrelevant X variation in the calibration modeling by balancing the
information in the X and Y spaces. This is especially the case when one acquires a large data
set using modern instrumentation like spectrometers, where, apart from having numerous
X-variables, there is also a tendency of these variables for being correlated, sometimes being
noisy and incomplete (Wold et al., 2001). On the other hand, the need to use PLSR analysis
method can arise from difficulty to obtain measurements only of the specific parameters that
one is interested in. Martens and Naes (1989) discussed problem of selectivity while trying
to take measurements of specific properties from inhomogeneous materials and presented a
multivariate calibration method, PLSR as a solution. PLSR combine principal component
analysis (PCA) and multiple linear regressions (MLR). In this technique X-variables are first
decomposed into set of orthogonal factors named latent variables. During the
decomposition the common structures between predictors and response is captured. Unlike
PCA which decomposes the X variables to eliminate multi-collinearity problems and extract
components that explain X, PLS finds components from X that are also relevant to Y. Since
PLSR considers the variation in Y when calibrating the model, the covariance structure
between the predictor and response variables is reflected (Martens and Naes, 1989; Wold et
al., 2001). This is achieved by projecting the X and Y-spaces into new coordinates T and U-
scores respectively that summarize the common structure in X and Y. Thus resulting latent
variables have the best predictive power in explaining the response. Then as in multiple
linear regression it builds a linear model Y=XB+E, where Y is an n cases by m variables
response matrix, X is an n cases by p variable predictor matrix, B is a q by m regression
coefficient matrix, and E is a noise term for the model which has the same dimensions as Y
(Wold et al., 2001).
Prior to considering the calibrated models for practical applications, i.e. the prediction and
subsequent understanding of new data set or samples, models should be validated (Wold et
al., 2001). This is a crucial step in PLSR modeling, for it gives an indication on how well the
models will perform in the future and the degree of certainty that one might expect while
using the models to solve practical problems. Different types of validation methods were
discussed and presented in various literature (Martens and Naes, 1989; Wold et al., 2001). Of
which, cross validation method has found its application in cases where the data set that one
is working on is small and hence a separate or independent and representative validation
data set is unavailable (Martens and Naes, 1989). Kooistra (2005) used a full cross validation
method, which is based on a leave one out principle where one sample will be left out at a
time and the model is calibrated on the remaining samples (CAMO, 2005). This will be
repeated N times until every sample is left out once and the model is computed on the
remaining samples and the left out sample is predicted.
To avoid erroneous calibration and deterioration of models prediction ability, it is important
to detect outliers and remove or replace them by accurate values (Hocking, 2003; Martens
and Naes, 1989). Outliers are abnormal observations that show significant deviation from
the rest of the dataset in the population. They might arise both due to error in the
experiment or instrument, or represent different information other than the material of
interest and hence irrelevant. Presence of outliers in the data set is known to influence both
the calibration and validation of PLSR models. Different methods are developed to detect
sample outliers in PLSR modeling. Martens and Naes (1989) presented outlier detection
criteria based on the analysis of residuals and leverages. In PLSR modeling, residuals are of
diagnostic interest. It is possible to examine the residual variances (the variation that is left
unexplained) in the X as well as the Y-spaces. Wold et al. (2001) demonstrated that large Y-
residuals indicate that the model is poor. Normal probability plots of the residuals of a
single Y-variable are also useful for identifying outliers in the relationship between T and Y.
The X-residuals (part of X that is not used in modeling Y) are also useful for identifying
outliers in the X-space, i.e., observations that do not fit the model. In addition, uncertainty
tests e.g. Martens uncertainty limit tests (CAMO, 2005; Martens and Naes, 1989) can be used
for testing which variables are causing perturbations in the model.
3. Results and Discussions

3.1 Spectral analysis
In the spectral interpretations, spectral libraries of different sources (e.g. TSG (the spectral
geologist), ENVI and PIMA view built-in mineral libraries, USGS and JPL mineral libraries)
that are developed upon experimental investigations on minerals and verified with variety
of conventional testing methods (e.g. X-ray diffraction (XRD), Transmission Electron
Micrograph (TEM), Scanning Electron Micrograph (SEM) etc) are used.
Differences in spectral characteristics among spectra of different soil samples were used for
differentiating various clay mineral types that are present in the soil samples. Position of
absorption features, their shapes, types and number, depth intensity and asymmetry; shape
of spectral curves, differences in slopes of spectral curves and variations in reflectance
intensity of spectra were some of the important qualitative parameters that helped in
identifying spectrally dominant clay mineral from the soil reflectance spectra (Figure 3).
Some spectra show a sharp rise in slopes and variable reflectance intensity throughout the
whole wavelength region of the electromagnetic spectrum. Others show lower reflectance
intensity throughout the whole wavelength range and were on overall dark. The later also
exhibited monotonously rising convex slopes in the VNIR (visible near infrared) wavelength
region and less variable reflectance intensity in the SWIR. Some show moderate rise in
slopes and also moderate increase in reflectance intensity from the VNIR to the SWIR
wavelength regions.
In the visible near infrared portion of the spectra changes in slopes were the prominent
features that were recognized coupled with changes in reflectance intensities. The
absorption features that are apparent on the VNIR region are relatively few, broad, wider
and less intense. Whereas, in the SWIR region the main absorption features of clay minerals
were observed with variable intensity being the prominent ones at ~ 1400 nanometer, ~ 1900
nanometer and ~ 2200 nanometer (Clark, 1999; Van der Meer, 1999).
Fig. 3. Spectral reflectance curves of some soil samples (no offset). Note spectral
characteristic differences, reflectance and slope variability in the VNIR wavelength region
(350 nanometer – 1000 nanometer) and absorption features in the SWIR (1000 nanometer –
2500 nanometer).
3.2 Geotechnical Characteristics

As presented in the plasticity chart (Figure 4) and particle size distribution of soil samples
(Figure 6), the soil samples exhibit high plasticity and are finer grained with high
percentages of clay fraction. Generally the more plastic and finer grained soils are the
greater swell-shrink potential that they are susceptible to, though swell-shrink potential is
dictated by mineralogy and other factors as well. Majority of the soil samples are plotted
above the A-line spanning from CI to CE zones. The soils that fall in the CI zone are of
intermediate plasticity behavior, while those falling in the CH, CV and CE are of high, very
high and extremely high plasticity nature. The higher the plasticity the larger will be the
susceptibility of the soil to significant volume change characteristics. Accordingly swell-
shrink potential of soil samples in the CH, CV, and CE zones are of high, very high and
extremely high. Few samples fall below the A-line in the MV zone; these soils have high
inherent expansion potential. The ‘A’ line is an empirical boundary separating inorganic
clays from silty and organic soils. Soils of the same geological origin usually plot on the
plasticity chart as straight lines parallel to the ‘A’ line. "Fat" or plastic clays plot above the
line. Organic soils, silts and clays containing a large portion of "rock flour" (finely ground
non-clay minerals) plot below it.
70
CE
U line PI=0.90*LL-0.8
60 (Upper boundary for
experimental data)
Plasticity Index (PI)
50 A line PI=0.73*LL-20
CV (Emperical divsion
between clayey and
silty soils)
40 ME
CH
30 MV
CI
Plasticity/Expansion
20 potential
MH Ext high
CL Very high
High
10 MI Intermediate
20 30 40 50 60 70 80 90 100 110 120

Liquid Llimit (LL)
Fig. 4. Plasticity chart showing classes of expansion potential of soil samples.
The scattering of soil specimen in the plasticity chart show a wide range of variability in
their swell-shrink characteristics (low to extreme cases). This is important in ensuring
representativeness of samples for the intended PLSR modeling since it affects model
prediction ability (Martens and Naes, 1989). Calibration in narrow range can bring about a
risk of inability to extrapolate model into observations spanning a wider range. That is for
instance bad prediction ability in case of failing to cover a total range of variability in a new
dataset.
Proctor test results of some soil samples depicting the maximum dry density (MDD) and
optimum moisture content (OMC) are shown in Figure 5. As presented in the graph the soil
samples are also labeled with mineralogical groups obtained upon interpretation of spectral
reflectance curves of respective soil samples.
1.8
S4
1.7 S5
Maximum dry density (MDD)
S1
S3
1.6 S6
S2
1.5
1.4
1.3
1.2
0 10 20 30 40 50 60
Optimum moisture content (OMC)
Fig. 5. Compaction curves of selected soil samples from proctor test labeled with
mineralogical classes of soil from spectral interpretation.
Smectite clay species dominated soils (S4 and S5) exhibited low dry densities (MDD) with
large moisture intake as indicated by their optimum moisture contents (Figure 5). As the
grading curves (Figure 6) depicted S4 is finer than S5 with higher clay content. Accordingly
S4 shows higher magnitudes of volumetric shrinkage and plasticity character (LL, PI) than
its similar species S5 (Table 1).
Halloysite dominated soils (S2, S3 and S6) are characterized by dry densities and moisture
intakes that seem to vary with their clay contents. While S6 has highest clay content of the
three soils followed by S3 and S2 (Figure 6), the magnitude of volumetric shrinkage and
plasticity that it exhibited is also high followed by S3 and S2 (Table 1). Dry density of S2 is the
highest followed by S3 and S6 respectively (Figure 5) in the halloysite clay dominated sols.
The quartz dominated soil (S1) attained the highest dry density with lowest moisture intake
as compared with soils dominated by halloysite and smectite clay varieties (Figure 5). This
soil is also coarser than the other soils (Figure 6) and is non-expansive (Table 1).
100
90
80
70
60
% Pass
S2
50
S3
40
S1
30
S4
20
S5
10 S6
0
100 10 1 0.1 0.01 0.001
Particle Size, mm
Fig. 6. Particle size distributions of selected soil samples labeled with mineralogical classes
of the soil specimen obtained from spectral interpretation.
Even though the clay content of halloysite dominated soil sample S6 is higher than the clay
content of smectite dominated soil sample S5, the later is characterized by higher
magnitudes of volumetric shrinkage and plasticity, and exhibited low dry density (MDD)
coupled with high moisture intake (OMC). This observation can be attributed to the fact that
clay mineralogy dominantly controls swell-shrink characteristics of expansive soils. On the
other hand within similar clay mineralogy, amount of clay fraction seems to govern the
magnitude of volumetric shrinkage, plasticity (LL and PI) and compaction characteristics
(MDD and OMC).
Sample ID Spectrally dominating mineral Volumetric shrinkage Liquid limit Plasticity index
S4 smectite 137 103 58
S5 smectite 118 84 43
S6 halloysite 99 72 34
S3 halloysite 48 48 22
S2 halloysite 32.8 46 22
S1 quartz - Np Np
Table 1. summary of geotechnical properties of selected soil samples. NP: non plastic
material
Variations in CBR and CBR swell of the six soil samples from different mineralogical groups
are presented in Figure 7. The CBR value is an indicator of soils strength or bearing capacity,
which is directly used in the design of subgrade, subbase and base material for pavement.
Highest CBR and lowest CBR swell values are attained by the quartz dominated soil sample,
S1. For the smectite dominated soils (S4 and S5) lowest CBR and highest CBR swell values
are recorded. CBR of S4 and S5 are below standard (e.g. to lay an embankment directly over
or to be used as subgrade material in road construction) and their CBR swell is much higher
than allowable CBR swell values. Halloysite dominated soils (S2, S3 and S6) exhibited CBR
and CBR swell values which seems to vary according to their clay fractions. Among the
halloysitic soils S6 show low CBR which is below standard and higher CBR swell which is
higher than the allowable CBR swell value. S3 on the other hand exhibited marginal CBR
value with high though within the allowable range in different standard specifications.
LOAD(KN)
LOAD(KN)
PENETRATION(mm) PENETRATION(mm)
ID Penetration load Standard load CBR CBR Swell ID Penetration load Standard load CBR CBR Swell
(mm) (KN) (KN) (%) (%) (mm) (KN) (KN) (%) (%)
S1 2.5 1.98 13.24 14.95 0.31 S2 2.5 1.16 13.24 8.76 0.94
5 4.36 20 21.80 5 1.67 20 8.35
LOAD(KN)
LOAD(KN)
ID P en e tr atio n Lo a d S ta nd a rd lo a d C BR C B R S w ell
ID Penetration load Standard load CBR CBR Swell (m m ) ( KN) (K N ) (% ) (% )
(mm) (KN) (KN) (%) (%)
S3 2.5 0.63 13.24 4.76 1.89 S4 2 .5 0.1 4 1 3.2 4 1 .06 14
5 1.02 20 5.10 5 0.1 9 20 0 .95
LOAD(KN)
LOAD(KN)
ID Penetration Load Standard load CBR CBR Swell ID Penetration load Standard load CBR CBR Swell
(mm) ( KN) (KN) (%) (%) (mm) (KN) (KN) (%) (%)
S6 2.5 0.44 13.24 3.32 4.29
S5 2.5 0.44 13.24 3.32 10 5 0.68 20 3.40
5 0.54 20 2.70
Fig. 7. CBR and CBR swell results, graphs with accompanying tables.
XRD test results show that the soil samples contain clay minerals (Figure 8) such as smectite
(mainly montmorillonite and nontronite), illite and kaolinite (halloysite and kaolinite) which
significantly influence engineering behavior of expansive soil due to their high activity; and
original minerals such as quartz, feldspar and mica are which are common constituents of
expansive soils but do not contribute to the expansiveness of soils due to their low activity.
Qualitative XRD analysis results are summarized in Table 2 and chemical analysis results
are resented in Table 3.
d=22.77507
1400
1300
1200
d=15.40376
1100
1000
900
Lin (C ounts)
800
d=9.93849
700
600
500
400
d=3.33473
d=3.23691
d=7.16025
d= 3.57443
300
d=4.98949
200
100
3 10 20 30 40
2-Theta - Scale
File: JUL4509-9-UNTRD.raw
File: JUL4509-9-GLC.raw
File: JUL4509-9-400C.raw
1000
900
800
d=15.37182
d=17.68018
700
600
Lin (Counts)
500
d=13.26454
400
d=10.05282
d=3.34135
300
d=3.24903
d=7.15563
d=9.99100
200
d=3.57426
d=3.02598
d=5.00512
d=4.25701
100
3 10 20 30 40
2-Theta - Scale
500
d=7.16733
400
d=3.57301
d=16.99007
d=3.34220
300
Lin (Counts)
d=10.06157
200
d=3.24641
d=4.25687
d=5.00644
100
d=2.98899
d=4.46364
3 10 20 30 40
2-Theta - Scale
Fig. 8. XRD patterns of clay fractions (< 2µ) of some soil samples (Table 2 and 3 present the
mineralogical assemblage and chemical results of these three patterns in descending order).
XRD mineralogy assignment
ID
Major Moderate Minor Trace Spectrally dominant mineralogy
kaolinite, *brookite, *rutile,
Bulk Montmorillonite quartz, calcite, plagioclase, *goethite
Nontronite potassium feldspar,
Illite Smectite
1
Clay Fraction Montmorillonite - illite, kaolinite,
(quartz), (calcite), -
(potassium feldspar)
quartz, kaolinite, *brookite, *rutile,

Bulk Montmorillonite Nontronite plagioclase, *goethite
potassium feldspar,
Smectite
Illite
2
Clay Fraction I/M mixture Nontronite illite, kaolinite, -
(quartz), (calcite),
(potassium feldspar)
quartz, potassium felspar,

kaolinite, illite, plagioclase, *brookite, *pyrite,
- halloysite montmorillonite, *goethite
Bulk nontronite, I/M
3
Halloysite
mixed
(quartz), illite, *montmorillonite,

Clay Fraction Halloysite - I/M mixed *(potassium
feldspar)
Table 2. Summary of Qualitative XRD results showing abundance of major (>30%),

moderate (10 -30 %), minor (2 – 10 %) and trace (< 2 %) mineral constituent of soil samples.
ID Sio2 Al2O3 Fe2O3 MgO CaO Na2O K2O TiO2 P2O5 MnO Cr2O3 V2O5 LOI Sum
% % % % % % % % % % % % % %
1 46 13.8 7.14 2.29 5.66 0.79 1.46 1.14 0.08 0.18 0.02 0.02 21.4 99.98
2 50.8 16.9 8.46 1.5 1.33 2.16 2.53 1.03 0.04 0.24 0.01 0.02 14.5 99.52
3 52.9 19 7.95 1.63 0.3 0.77 1.86 1.22 0.15 0.17 0.01 0.02 13.2 99.18
Table 3. Chemical analysis results.
The information obtained on the clay mineralogical assemblage of the soil samples from X-
ray diffraction analysis is in conformity with spectrally dominant mineralogical group
assignments from interpreting reflectance spectra of respective soil samples.
Major clay mineral that is responsible for swelling and shrinkage of soils in the study area is
smectite (montmorillonite and nontronite) as identified from soil spectral reflectance and
confirmed by the X-ray diffraction analysis; and mixed layer combination of
montmorillonite and illite. The hydrous variety of kaolinite group (halloysite) shows
variable, that is low to appreciable swell-shrink character. Halloysite show low bulk density
than kaolinite coupled with high porosity; its hydraulic conductivity is also reported to be
higher (West et al., 2004). Geotechnical character of halloysite dominated soils seems to vary
according to their particle size distribution (clay to coarser fraction proportions).
Montmorillonite, (Ca, Na) 0.67Al 4(Si, Al) 8 O20(OH)4·nH2O, is a product of weathering of
iron and magnesium rich parent materials and is one of the most common smectite minerals
(AusSpec International, 2005; Fitzpatrick, 1980) found in soils. It also form from the
weathering of volcanic ash or primary silicate minerals such as feldspars, pyroxenes, or
amphiboles under conditions of insufficient leaching of soil profile due to low permeability
and excessive evaporation (Snethen, 1975). As indicated in the brief summary of the
Geology of the study which is covered by rocks of volcanic origin where volcanic debris and
mafic rocks like basalt are common and abundant, occurrence of montmorillonite can be
favored by the environment. Nontronite, (Ca, Na) 0.66 Fe3+4(Si, Al) 8O 20(OH) 4·nH2O, is
also a common smectite mineral found in soils and weathered bedrock. Its formation is
favored by alkaline to neutral pH environments, as well as by the availability of iron and
calcium minerals (AusSpec International, 2005; Fitzpatrick, 1980). Thus its formation is also
favored by the geology of the study area. Additional prevailing environmental conditions
such as alkaline conditions coupled with high evaporation exceeding precipitation and poor
leaching which facilitate retention of magnesium and calcium in the soils contribute to the
development of smectite clays in the area.
Kaolinite, Al2Si2O5 (OH)4 is another commonly occurring clay mineral in soils. It can be
derived from almost all silicate minerals (AusSpec International, 2005), hence its formation
in the study area can be favored by the environmental conditions. Halloysite,
Al2Si2O5(OH)4·4H2O, occurs in soils and the uppermost weathered part of bedrock
(AusSpec International, 2005) and is a common constituent of many volcanic soils
(Takahashi et al., 2001). It is a kaolinite group clay mineral formed as a result of weathering
of aluminum rich minerals that are also abundant in the study area and its surroundings.
Illites on the other hand are commonly seen in soils, and form by weathering of silicates
primarily feldspars. Generally their formation is favored by alkaline environment and high
concentrations of Aluminum and potassium (Fitzpatrick, 1980); which are conditions
fulfilled in the study area. It is common that illites appear with smectitic interlayer clays
(AusSpec International, 2005).
Some spectra show presence of iron oxides in the soil samples. Different kinds of iron
oxides, for example, goethite (αFeO+3(OH)) is present (Figure 3 and Tables 2 and 3) in the
soil samples as a result of weathering product of iron-bearing minerals. Since volcanic rocks
that are rich in iron minerals are abundant in the study area; which is coupled with the
action of chemical weathering in the humid atmosphere of the local tropical climatic
conditions; presence of goethite in the soil samples can be favored by the environmental
conditions. Spectra of some soil samples showing diagnostic features of iron oxides similar
to the spectral characteristics of goethite (Crowley et al., 2003) can be seen in Figure 3.
Figure 9 presents relationships between magnitudes of soil volumetric shrinkage; soil
mineralogical group resulted from spectral reflectance interpretation and plasticity classes
according to the plasticity chart. Mean volumetric shrinkage of smectite dominated soil
samples is higher than that of the halloysite dominated ones. The two mineralogical classes
that the soil samples are grouped into show clear separation. The halloysite dominated soils
fall into the intermediate and high plasticity classes which denoted intermediate and high
swell-shrink susceptibility. Smectite dominated soils on the other hand fall within the
extremely high and very high plasticity classes with few samples falling in the high
plasticity class signifying extremely high, very high and high swell-shrink potential
respectively. Even though few smectitic soils are noted to belong in the high plasticity class,
mean volumetric shrinkage of these soils is higher (~ 120) than the mean volumetric
shrinkage of halloysitic soil samples which is about 80. The whiskers of boxes denoting high
plasticity class of halloysite rich soils and very high class of smectite rich soils suggesting
some overlap on the magnitude of volumetric shrinkage of the two classes though fall
within different mineralogical categories. This might be related with high clay content of
halloysite rich soil samples and low clay fraction of those of smectite rich soils. Takashi
(2001) reported halloysite rich volcanic soils with appreciable clay fraction exhibiting high
cation exchange capacity indicating high swell-shrink potential.
Plasticity
15 0.00 Ext high
Very high
High
Volumetric shrinakge
12 5.00
Intermediate
10 0.00
75 .00
50 .00
S H
Spectral mineralogy
Fig. 9. Relationships between volumetric shrinkage, mineralogy and plasticity of the soil
samples. Note an increase in magnitude of volumetric shrinkage as mineralogy changes to
highly expanding clay species (smectite) dominated soils, coupled with increase in mean
volumetric shrinkage as plasticity changes from intermediate to high, very high and
extremely high classes.
Smectite and halloysite rich soils fall into two distinct clusters while plotting the values of
liquid limit (LL) in the X-axis against the magnitude of volumetric shrinkage in the Y-axis
(Figure 9). K-means clustering (which is a procedure that attempt to identify and cluster
homogeneous groups based on distance from specified or computed cluster centers) also
gave these two clusters with LL of 65 and volumetric shrinkage of 100 as boundaries
separating the two clusters. Same kind of clustering is noted in the plots showing the
relationship between Plasticity indices (PI) versus volumetric shrinkage. Sample labeling 1
and 2 in each graph are cluster designations obtained from the K-means clustering analysis.
2 2 2
2 2
2 2
150 2
2
2 2
2 2
2
Volumetric shrinakge 2 2
2 2
2 2
2
2
2
100
1
1
1
1
1 1
50 spectral mineralogy
1
H
S
1
20 40 60 80 100 120
LL
2 2
2
2
2
150 2 2 2
2
2 2
2 2
2
Volumetric shrinakge
2 2
2 2
2
2 2
2
2
100
1
1
1 1
1
1
50 Spectral mineralogy
1
H
S
1
10 20 30 40 50 60 70
PI
Fig. 10. Scatter plots showing the relationships between volumetric shrinkage versus liquid
limit, and volumetric shrinkage versus plasticity index with samples labeled with
mineralogical classes. Sample labeling 1 and 2 are cluster designation obtained from K-
means clustering analysis.
2.5 Partial Least Squares Regression (PLSR) models

PLS1 analysis (predicting a single parameter at a time) method implemented in The
Unscrambler software (CAMO, 2005) was used for the multivariate calibration and
validation. Outlier detection was performed, since presence of outliers in the input dataset
can deteriorate models prediction ability and also their reliability. Outliers were mainly
identified manually instead of using the automatic outlier detection method. The automatic
detection tends to include many of the extreme values (either from the lower or higher
extremes) in the outliers list. In general outlier detection was done based on using a
combination of different tests rather any single method. We used Martens uncertainty limit
tests to test the stability of all variables in the model. The stability of each predictor with
respect to the samples was examined by using stability plots.
A sufficient number of PLS factors were used, that is three PLS factors with optimal
predictive ability. Including too many PLS factors was observed to often lead to over fitting
problems (Kooistra, 2004) for it incorporates irrelevant information or noise. The number of
PLS factors used in the models were based on different tests intended for testing the
significance of PLS factors in the prediction. Among the tests are simultaneous examinations
of the residual variances and the root mean square errors (RMSE) of each factor coupled
with significance tests through cross validation. Cross validation is a practical and reliable
way of testing the predictive significance of PLS factors analysis (Martens and Naes, 1989)
that has become a standard in PLS (Wold et al., 2001). Martens and Naes (1989)
demonstrated the application of cross validation methods for determining number of PLS
factors that should be included in a model for proper explanation of the phenomena of
interest that one would like to model.
In deciding which variables, among the predictors, should be retained in the model,
Martens uncertainty limit test was used. This test is significance test in PLS analysis which is
useful in testing whether the regression coefficients used in a model are significantly
contributing to the model. Regression coefficients that are significant were then identified
and those that have got an uncertainty limits passing the origin were left out. Another
uncertainty test that was applied on the score and loading plots was stability calculation
which can be visualized in The Unscrambler as stability plots. It shows the influence of each
variable and hence its significance (CAMO, 2005) in the model. Samples far from the center
have more influence on the model than those that are near, and the uncertainty is larger on
those with larger spread implying that these variables are not significant (CAMO, 2005).
Fig. 11. Result of PLSR modeling for volumetric shrinkage showing: Scores principal
components 1 versus 2 with samples in the 95% confidence ellipse showing that there is no
particular grouping of samples, but rather a random pattern (one population) suggesting a
single model can fit the data.
Fig. 12. Result of PLSR modeling for volumetric shrinkage showing: Regression coefficients
or statistically significant wavelength regions in predicting volumetric shrinkage from the
laboratory soil reflectance spectra.
Fig. 13. Result of PLSR modeling for volumetric shrinkage showing: Residual variances
(blue during calibration and red during prediction) exhibiting the remaining variation that
is not taken into account by the model is minimum after fitting three PLS components
suggesting much of the variability in volumetric shrinkage is explained by the model. The
upper graph shows Y-variance and the lower graph shows the X-variances.
Fig. 14. Result of PLSR modeling for volumetric shrinkage showing: Regression overview
showing predicted and measured volumetric shrinkage.
Note that there are no outlying samples lying outside the rest of the cloud (Figure 11). Only
one sample lies outside the ellipse; under normal situation it is expected that about 5 % of
the samples to lie outside the ellipse.
As shown in Figure 12 Relevant X variables (wavelengths from soil spectra) in predicting
volumetric shrinkage fall in the VNIR and SWIR regions. Wavelengths in the SWIR bands
are related with clay mineral diagnostic features (Clark, 1999). Wavelengths from the VNIR
can be organic matter and iron oxides related (Ben-Dor and Banin, 1994; Wan et al., 2002)
and sand or quartz related spectral features (Viscarra et al., 2006).
Figure 13 shows that most of the Y and X variances are accounted for by the three PLS
factors, reaching a minimum level (close to zero) of unexplained variances at the third PLS
factor. The blue and the red lines represent residual variances during the calibration and
validation stages respectively. Note that the two lines do not significantly differ indicating
that the calibration data are well fitted and that the model describes the validation data
equally well.
In the regression overview (Figure 14) comparing laboratory measured and corresponding
predicted (from spectral reflectance of soil samples using PLSR analysis) soil volumetric
shrinkage magnitudes, calibration and prediction points lie very close to each other
suggesting the model fitted to the calibration data set described the prediction data set as
good as possible. Note that an internal calibration, full cross validation leave one out at a
time method was used.
Apart from high coefficient of determination (~0.91), the model gives low estimation errors
during the calibration (RMSEC = 0.05) as well as the validation (RMSEP = 0.06) stages.
Standard errors of calibration (SEC = 0.05) and prediction (SEP = 0.06) which are indicators
of precision of the calibration and prediction respectively are also small. The bias which is
the average value of the difference between the predicted and measured values, is also small
(both in the calibration and prediction stages) for the given number of PLS factors indicating
that the effect of bias is negligible (Martens and Naes, 1989). The values of offsets in the
calibration and validation stages are small showing minor deviation from an ideal line
where the measured and predicted volumetric shrinkage magnitudes are expected to be
equal. Model performance indices are presented in Table 4.
Calibration Validation
Correlation Correlation
coefficient RMSEC SEC Bias Offset coefficient RMSEP SEP Bias Offset
0.965 0.05 0.05 0.005 0.044 0.952 0.586 0.0596 0.001 0.051
Table 4. Summary of model performance indices resulted from PLSR modeling of soil
volumetric shrinkage.
4. Conclusions
In this chapter we demonstrated that it is possible to estimate soil volumetric shrinkage
from spectral reflectance of soil. The approach can be of potential geotechnical utility
contributing to the quality of geotechnical investigations of expansive soil, playing
prominent role in planning and designing of infrastructure. This is particularly in
minimizing uncertainties through identifying geotechnically problematic areas and
estimation of critical parameters (in this case volumetric shrinkage), hence providing better
basis for decision making. This in turn present relevant information that can be useful in site
selection, route planning and search for construction materials (borrow, subbase etc)
especially in the reconnaissance and preliminary design stages of construction projects. In
summary;
 The expansive soils in this study are identified by their constituent dominant clay
mineral type and accordingly classified into two clusters. Those soils comprised of active
clay mineral smectite exhibit high swell-shrink potential as suggested by the magnitude of
volumetric shrinkage, plasticity and other related geotechnical characteristics. They also
exhibit low maximum dry densities while their optimal moisture contents are high.
Halloysite clay mineral dominated soils on the other hand show less swell-shrink
susceptibility, and variable maximum dry densities with optimum moisture contents lower
than those exhibited by the smectite dominated soil samples. Some halloysite bearing soils
can be weak and show high swell-shrink susceptibility as indicated by their geotechnical
properties. Note their CBR strengths presented in Figure 7 for the wettest likely condition
likely and the associated CBR swell that they exhibited coupled with values of volumetric
shrinkage and plasticity characters that they show. Quartz dominated soil samples on the
other hand show higher maximum dry density with low moisture intake. These soils also
attain high CBR strength and show negligible CBR swell.
 Differences in geotechnical characteristics show dependency on clay mineralogy
among different species (smectite, halloysite), and on clay fraction among similar varieties in
the studied soils. The fact that clay mineralogy is a crucial factor dictating a number of soil
geotechnical behaviors laid emphasis on the importance of qualitative clay mineralogical
assemblage analysis of expansive soils. Since spectroscopy is a cheap, rapid and non-
destructive way of analyzing soil mineralogy, it can be used for such kind of routine
analysis in order to complement quantitative analyses.
 Coefficient of determination obtained in the PLSR analysis show that soil

volumetric shrinkage is highly correlated with spectral parameters. The low estimation and
interference errors, negligible bias and small offsets indicate spectroscopic techniques
potential to be used in routine quantitative analyses of soil shrinkage potential. The
presented empirical relations with the added valuable information on the content of major
clay minerals (from spectral interpretations) establish a simple way of characterizing
expansive soils. Despite the complex nature (comprised of various constituents other than
clay materials which are responsible for their expansive characteristics) of soil the results
proved that a remarkable amount of information about soil properties can be obtained from
their reflectance spectra.
 The current study gave an outlook for future application of optical remote sensing
to map soils susceptible to swell-shrink and variation in the magnitude of expansion and
shrinkage, provided with the availability of sufficiently high resolution (both spatially and
spectrally in order to resolve vital spectral details ) data.
5. References
Abebe, Franscesco, Fabrizio and Piero, 1999. Geological Map of Debreziet Area (Ethiopia).
GSE, Addis Ababa.
Al-Rawas, A.A., 1999. The factors controlling the expansive nature of the soils and rocks of
northern Oman. Engineering Geology, 53: 327-350.
ASD, 1995. Analytical Spectral Devices (ASD): Technical Guide, Portable
Spectroradiometers, Analytical Spectral Devices, Inc.
AusSpec International, 2005. The Spectral Geologist software. AusSpec Interbational,
Sydney.
Ben-Dor, E. and Banin, A., 1994. Visible and Near-Infrared (0.4-1.1µm) analysis of Arid and
Semiarid Soils Remote Sensing of Environment, 48: 261 - 274.
CAMO, 2005. The unscarmbler 9.2.
Chabrillat, S., Goetz, A.F.H., Krosley, L. and Olson, H.W., 2002. Use of hyperspectral images
in the identification and mapping of expansive clay soils and the role of spatial
resolution. Remote Sensing of Environment, 82: 431 - 445.
Clark, R.N., 1999. Spectroscopy of Rocks and Minerals, and Principles of Spectroscopy. In:
A. Rencz (Editor), Chapter 1 in: Manual of Remote Sensing. John Wiley and Sons,
Inc, New York.
Crowley, J.K. et al., 2003. Spectral reflectance properties (0.4-2.5 µm) of secondary Fe-oxides,
Fe-hydroxides, and Fe-sulphate-hydrate minerals associated with sulphide -bearing
mine wastes. Geochemistry; Exploration, Environment, Analysis, 3: 219-228.
Erguler, Z.A. and Ulusay, R., 2003. A simple test and predictive models for assessing swell
potential of Ankara (Turkey) clay. Engineering Geology, 67: 331 - 352.
FAO, 1998. World reference base for soil resources. International Society of Soil Science,
Rome.
Fitzpatrick, E.A., 1980. SOILS; their formation, classification and distribution. Longman Inc.,
London, NY, 354 pp.
Goetz, A.F.H., Chabrillat, S. and Lu, Z., 2001. Field Reflectance Spectrometery for Detection
of Swelling Clays at Construction Sites. Field Analytical Chemistry and
Technology, 5(3): 143-155.
Gourley, C., Newill, D. and Schreiner, H.D., 1993. Expansive soils: TRL's research strategy,
First International Symposium on Engineering Characteristics of Arid Soils.
Overseas Centre, Transport Research Laboratory, Crowthorne, Berkshire, UK., City
University, London.
Hocking, R.R., 2003. Methods and Applications of Linear Models Regression and the
analysis of variance. Wiley series in probability and statistics.
Kariuki, P.C., Woldai, T. and Van der Meer, F.D., 2004. Effectiveness of spectroscopy in
identification of swelling indicator clay minerals. International Journal of Remote
Sensing, 25(2): 455 - 469.
Kooistra, L., 2004. Incorporating spatial variability in ecological risk assessment of
contaminated river floodplains. PhD Thesis, Katholieke Universiteit Nijmegen,
Nijmegen, The Netherlands.
Martens, H. and Naes, T., 1989. Multivariate Calibration. John Wiley & Sons, INC.
Nelson, J.D. and Miller, D.J., 1992. Expansive Soils Problems and Practice in Foundation and
Pavement Engineering. John Wiley & Sons, Inc.
Ramana, K.V., 1993. Humid tropical expansive soils of Trinidad: Their geotechnical
properties and areal distribution. Engineering Geology, 34: 27 - 44.
Shi, B., Jiang, H., Liu, Z. and Fang, H.Y., 2002. Engineering geological characteristics of
expansive soils in China. Engineering Geology, 67: 63 - 71.
Snethen, D.R., 1975. A Review of Engineering Experiences with Expansive Soils in Highway
Subgrades, Fedral Highway Administration Office of Research and Debvelopment,
Washington D.C.
Takahashi, T., Dahlgren, R.A., Theng, B.K.G., Whitton, J.S. and Soma, M., 2001. Potassium-
Selective, Halloysite-Rich Soils Formed in Volcanic Materials from Northern
California. Soil Science Socaitey of America, 65: 516 - 526.
Van der Meer, F.D., 1999. Can we map swelling clay with remote sensing? International
Journal of Applied Earth Observation & Geoinformation (JAG), 1: 27-35.
Viscarra, R.A.R., Walvoort, D.J.J., McBratney, A.B., Janik, L.J. and Skjemstad, J.O., 2006.
Visible, near infrared, mid infrared or combined diffuse reflectance spectroscopy
for simultaneous assessment of various soil properties. Geoderma, 131: 59 - 75.
Wan, Y., Kwong, J., Brandes, H.G. and Jones, R.C., 2002. Influence of Amorphous Clay - Size
Materials on Soil Plasticity and Shrink - Swell Behaviour. Journal of Geotechnical
and Geoenvironmental Engineering, 128: 1026 - 1031.
West, S.L., Juo, A.S.R., White, G.N., McInnes, K.j. and Dixon, J.B., 2004. Kaolinite, Halloysite,
and Iron Oxide Influence on Physical Behavior of Formulated Soils. Soil Science
Socaitey of America, 68: 1452 - 1460.
Wold, S., Sjostrom, M. and Eriksson, L., 2001. PLS-regression: a basic tool of chemometrics.
Chemometrics and Intelligent Laboratory Systems, 58: 109-130.
Yitagesu, F.A., Van der Meer, F.D. and Van der Werff, H., 2009. Quantifying engineering
parameters of expansive soils from their reflectance spectra. Engineering Geology,
105: 151 - 160.
Monitoring Soil Moisture from Spaceborne Passive Microwave
Radiometers: Algorithm Developments and Applications to AMSR-E and SSM/I 371
17
X
Monitoring Soil Moisture from Spaceborne

Passive Microwave Radiometers: Algorithm
Developments and Applications
to AMSR-E and SSM/I
Hui Lu1*, Toshio Koike1, Tetsu Ohta1, David Ndegwa Kuria2, Kun Yang3,
Hideyuki Fujii4, Hiroyuki Tsutsui1 and Katsunori Tamagawa1
1The
University of Tokyo, Japan
2Jomo Kenyatta University of Agriculture and Technology, Kenya
3Institute of Tibetan Plateau Research, Chinese Academy of Sciences, China
4Japan Aerospace Exploration Agency, Japan
1. Introduction
Soil moisture patterns, both spatial and temporal, are the key to understanding the spatial
variability and scale problems that are paramount in scientific hydrology, meteorology and
climatology. Soil moisture controls the ratio of runoff and infiltration (Delworth & Manabe,
1988; Wagner et al., 2003), decides the energy fluxes (Entekabi et al., 1996; Prigent et al.,
2005) and influents vegetation development and then carbon cycle. A long term soil
moisture data set on a region scale therefore could provide valuable information for
researches such as climate change and global warming (Seneviratne et al., 2006), and then
improve the weather forecasting (Beljaars et al., 1996; Schar et al. , 1999) and water resources
management.
Soil moisture profile can be observed at point scale by using gravimetric sampling or some
automatic probes, such as Time Domain Reflectometry (TDR), Neutron Probe (NP), etc.
These methods are commonly used to provide accurate and continuous soil moisture
information and adopted by the meteorology, hydrology and agriculture stations. But these
point information are not enough for the regional research and application, and are also not
available in the remote areas where difficult to access and to maintain such stations. On the
other hand, satellite remote sensing offers a possibility to measure surface soil moisture at
regional, continental and even global scales.
Although surface soil moisture can be estimated indirectly from visible/infrared remote
sensing data (Verstraeten et al., 2006), it failed to produce routinely soil moisture map
mainly due to factors inherent in optical remote sensing, such as atmosphere effects, cloud
masking effects and vegetation cover masking effects. Fortunately, microwave remote
sensing offers a possibility to observe area-averaged surface soil moisture regularly in the
global scale, by directly measuring to the soil dielectric properties which are strongly related
*Corrosponding author and address: Dr. Hui LU, River and Envi. Eng. Lab., Dept. of Civil Eng., Univ.
of Tokyo, Hongo 7-3-1, Bunkyo-ku, Tokyo, 113-8656, Japan. Tel. 81-3-5841-6109 Fax: 81-3-5841-6130
E-mail: lu@hydra.t.u-tokyo.ac.jp
to the liquid moisture content (Hipp, 1974). Moreover, extra advantages of microwave
remote sensing include: (1) long wavelength in microwave region which enable the low
frequency microwave signals to penetrate clouds and to provide physical information of the
land surface; and (2) independent of illumination source which enables the spaceborne
sensors to observe earth all-day with all-weather coverage.
There are two approaches through which microwave remote sensing estimating surface soil
moisture: active ways by Radar and/or Synthetic Aperture Radar (SAR) with high spatial
resolution (in the order of ten to hundred meters) and long revisiting period (about 1
month), passive ways by radiometers with coarse resolution (~ order of 10 km) and frequent
temporal coverage (daily or bi-daily). Considering the temporal resolution requirement of
the meteorological and hydrological modeling, passive ways are more suitable for the
application in these fields and have been wildly used in recent 30 years.
Although it was recognized early that microwave sensors operated at L-band (1-2GHz)
provide the best surface soil moisture observation (Schmugge et al. 1988), L-band
radiometers are not equipped on any satellites. This situation is mainly due to the limitation
of our current technical capability, which is a significant challenge to built L-band antenna
big enough to provide reasonable resolution. So all current available passive microwave
remote sensing data are observed by radiometers operating with higher frequencies, such as
the Scanning Multichannel Microwave radiometer (SMMR; 6.6, 10.7, 18.0, 21 and 37 GHz) on
board Nimbus-7 Pathfinder (Gloersen & Barath, 1977), the Special Sensor Microwave Imager
(SSM/I; 19.35, 22.2, 37.0 and 85.5 GHz) on board Defense Meteorological Satellite Program
(DMSP) (Hollinger et al., 1990), and the Advanced Microwave Scanning Radiometer
(AMSR-E; 6.925, 10.65, 18.7, 23.8, 36.5 and 89 GHz) of the Earth Observing System (EOS) on
board Aqua (Kawanishi et al., 2003). Among them, AMSR-E is the only passive systems
which including surface soil moisture as a target product. In terms of continuous
observation, SSM/I series, starting from 1987, is highly expected to provide long-term global
soil moisture estimation.
A number of techniques have been used to estimate surface soil moisture information from
microwave remote sensed data, such as statistical inversion (Njoku & Kong, 1977), artificial
neural networks (Said et al., 2008), and genetic algorithm (Singh & Kathpalia, 2007).
However, only the radiative transfer-based methods are considered as the true retrieval, for
both passive and active techniques. The radiative transfer models adopted in those methods
generally are consists of three parts: a dielectric model (Wang & Schmugge, 1980; Dobson et
al., 1985; Mironov et al., 2004) which relating surface soil moisture content to the dielectric
constants; a surface roughness model (Choudhury et al., 1979; Fung et al., 1992; Wegmuller
& Matzler, 1999; Chen, et al. 2003; Shi, et al. 2005) which accounting for surface scattering
effects; and a vegetation layer model (Ulaby et al., 1983; Paloscia & Pampaloni, 1988; Jackson
& Schmugge, 1991) which accounting for the vegetation masking effects.
Algorithms considering various other factors have also been proposed to retrieve soil
moisture content from passive microwave remote sensed data. T. Jackson (Jackson, 1993)
developed a so-called single channel algorithm (SCA), in which the brightness temperature
of the 6.9 GHz horizontal polarization channel was used. In this algorithm, ancillary data
such as air temperature, land cover, Normalized Difference Vegetation Index (NDVI),
surface roughness, and soil texture and porosity are needed. The algorithm of Njoku et al
(Njoku & Entekhabi, 1996; Njoku et al., 2003) is a multiple channel iterative retrieval
algorithm. It uses the brightness temperature observed by the lowest six channels of AMSR-
E. Using their algorithm, the surface temperature, the vegetation opacity and the soil
moisture are estimated simultaneously. The algorithm proposed by Paloscia (Paloscia et al.,
2001; 2006) is an experiment-based linear regression retrieval, in which soil moisture is
estimated by using both the Polarization Index (PI) at 10.7GHz and the brightness
temperature at 6.9GHz.
After more than 20 years effort, good results were obtained and several global and
continental scale soil moisture datasets (e.g. Njoku et al. 2003; Owe et al. 2008) were
generated. But both the quality and application region of these algorithms can be further
improved. For example, Shibata et al. (2003) pointed out that the soil moisture in desert
regions retrieved from AMSR-E soil moisture algorithms indicate very wet areas. To solve
such problem, the forward model, viz. Radiative Transfer Model (RTM) should be improved
firstly.
In this study, we present a new soil moisture retrieval algorithm developed at the
University of Tokyo. This algorithm is based on a modified radiative transfer model (Lu et
al., 2006), in which the volume scattering inside soil layers is calculated through dense
media radiative transfer theory (DMRT) (Wen et al., 1990; Tsang & Kong, 2001) and the
surface roughness effect is simulated by Advanced Integration Equation Model (AIEM)
(Chen et al., 2003). The optimal values of forward model parameters are estimated using in
situ observation data and lower frequency brightness temperature data. And with those
optimized parameters, we run the forward model to generate a lookup table, which relates
the variables of interest, such as soil moisture content, soil physical temperature, vegetation
water content and atmosphere optical thickness, to the brightness temperature or some
indexes calculated from brightness temperature data. Finally, soil moisture content is
estimated by linearly interpolating the brightness temperature or index into the inversed
lookup table. The algorithm was validated by using the AMSR-E match up data set at
Mongolia region. Moreover, the capability of our algorithm to retrieve soil moisture from
SSM/I was also checked at the same region.
The paper is organized as follows. In Section 2, we present our physically-based radiative
transfer model, emphasizing the soil RTM, so-called DMRT-AIEM model. In Section 3 we
describe the structure of our algorithm. Section 4 and 5 discusses the application of our new
algorithm on AMSR-E and SSM/I data, respectively. Section 6 contains some concluding
remarks.
2. The Forward Model: A Fully Physically-based Radiative Transfer Model

Our algorithm is based on a look up table, which is a database of brightness temperature
simulated by a radiative transfer model for various possible conditions. The quality of
retrieved soil moisture, therefore, is heavily dependent on the performance of the radiative
transfer model. So, the main task of our algorithm development was to develop a
physically-based soil moisture retrieval algorithm, which is able to estimate soil moisture
content from low frequency passive microwave remote sensing data and to overcome the
misrepresent problems occurred in dry areas.
2.1 The radiative transfer process for land surface remote sensing
For the land surface remote sensing by spaceborne microwave radiometers, the radiative
transfer process from land to space can be divided into as four stages as follows:
(1) Radiative transfer inside soil media.

The initial incident energy is treated as the one starting from the deep soil layer, which
propagates through many soil layers, attenuating by the soil absorption effects (dominative
at wet cases) and volume scattering effects (dominative at dry cases), experiencing multi-
reflection effects between the interfaces of soil layers, finally reaching the soil/air interface.
(2) Surface scattering process at soil/air interface.
At the soil/air interface, the surface scattering influences this upward initial radiation by
changing its direction, magnitude and polarization status. At the same time, the downward
radiation from the cosmic background, atmosphere, precipitation and canopy are reflected
by the air/soil interface, and parts of the reflected radiation propagate along the same
direction as that emitted from the soil layers.
The upward radiation just above the soil/air interface, therefore, is not only the product of
soil medium but also the product of downward radiation.
(3) Radiative transfer inside vegetation layers.
After leaving the soil/air interface, the upward radiation propagates through the canopy
layer (if there are vegetations), experiences the volume scattering effects from the leaves and
stems of vegetations and the multi-reflection effects between canopy/air and soil/air
interfaces. At the same time, parts of the upward radiation from vegetations join our target
radiation.
(4) Radiative transfer inside atmosphere layers.
After transmitting from vegetation layer, the radiation continues its way, traversing the
cloud and precipitation layers, affected by the absorptive atmosphere gases, scattered by
precipitation drops, incorporating the emission from surroundings, finally detected by the
sensors boarded on satellites.
The story of radiative transfer is so complicate that make it necessary to simplify the process
to make it computable. In microwave region, the reflectivity of the air/soil interface is
generally small. The downward radiation from vegetation and rainfall, which is reflected by
the soil surface, therefore, is neglected. Moreover, for the lower frequencies region of
microwave, the atmosphere is transparent. Finally, after neglecting all the downward
radiation and parts of upward radiation from surroundings, the radiative transfer model is
written as:
Tb = Tbs e −τ c e −τ r + (1 − ωc )(1 − e −τ c )Tc e −τ r + ∫ (1 − ωr ( R ))(1 − e −τ r ( R ) )Tr ( R )dR (1)
where Tbs is the emission of the soil layer, Tc is the vegetation temperature, Tr is the
temperature of precipitation droplets, τc and ωc are the vegetation opacity and single
scattering albedo, and τr and ωr are the opacity and single scattering albedo of precipitation.
For the frequencies less than 18GHz, equation (1) can be even simplified by omitting the
precipitation layer, as:
Tb = Tbs e −τ c + (1 − ω c )(1 − e −τ c )Tc (2)
2.2 Radiative process inside soil media: profile effects and volume scatting effects
Microwave can penetrate into soil media, especially for dry cases, in which the penetration
depth of C-band is about several centimeters. The soil moisture observed by microwave
remote sensing, therefore, is inside a soil media with a volume of several centimeters depth.
The radiative transfer process inside a soil media includes various effects, such as moisture
and temperature profile effects and the volume scattering effects of dry soil particles. To
simulate these effects, the dielectric constant model should be addressed at first.
(1) Dielectric constant model of soil
In the view of microwave, soil is a multi-phase mixture, with a dielectric constant decided
by moisture content, bulk density, soil textural composition, soil temperature and salinity.
In our algorithm, the dielectric constant of soil is calculated using Dobson model (Dobson et
al., 1985):
α ρb α (3)
ε soil = 1+ (ε ss − 1) + mvβ ε αfw − mv
ρ ss
where ρb is the bulk density of soil, ρss = 2.65 is the density of solid soil particle; εss = (4.7, 0.0)
is the dielectric constant of soil particle; mv is the volumetric water content; εfw is the
dielectric constant of free water, can be calculated by the model proposed by Ray (Ray,
1972); α=0.65 is an empirical parameter; and β is a soil texture dependent parameter as
follows:
β = 1.09 − 0.11S + 0.18C (4)
where S and C are the sand and clay fraction of the soil, respectively.
(2) Profile effects of soil media
The heterogeneity inside soil media causes the so-called profile effects. The profile effects
can be accounted for by using the simple zero-order noncoherent model proposed by
Schmugge and Choudhury (1981) or by more complicate first-order noncoherent model
given by Burke et al. (1979). The volume scattering effects inside soil media are not included
in both models.
In order to include the volume scattering effects, a more complicate model was adopted in
our algorithm. We assumed that the soil has a multi-layer structure and is composed of
many plane-parallel and azimuthally symmetric soil slabs with spherical scattering
particles. The radiative transfer process in a plane-parallel and azimuthally symmetric soil
slab with spherical scattering particles can be expressed as (Tsang & Kong, 1977)：
d  I v (τ , µ )   I v (τ , µ )  1 ω 0 1
 PVV PVH   I v (τ , µ ′)  (5)
µ  =
dτ  I h (τ , µ )  I h (τ , µ )
 − (1 − ω 0 ) B(τ ) 1 −
 2
∫  P
−1 HV PHH   I h (τ , µ ′)
dµ ′
where IP(τ,μ) is the radiance at optical depth τ (dτ=Kedz, with extinction coefficient Ke and
layer depth dz) in direction μ for polarization status P (horizontal or vertical), ω0 is the single
scattering albedo of a soil particle, B(τ) is the Planck function and Pij (i,j=H or V) is the
scattering phase function. The 4-stream fast model proposed by Liu (Liu, 1998) solves (5) by
using the discrete ordinate method and assuming that no cross-polarization exist. The
Henyey-Greenstein formula (Henyey & Greenstein, 1941) is used to express the scattering
phase function.
(3) Volume scattering effects of dry soil particles
With considering the facts that the soil particles are densely compacted, the multi-scattering
effects of soil particles should be accounted for. In our algorithm, this volume scattering
effect were calculated by the so-called dense media radiative transfer theory (DMRT) under
Quasi Crystalline Approximation with Coherent Potential (QCA-CP) (Wen et al., 1990;
Tsang & Kong, 2001). Dense Media radiative transfer theory was derived from Dyson’s
equation under the quasi-crystalline approximation with coherent potential (QCA-CP) and
the Bethe-Salpeter equation under the ladder approximation of correlated scatterers.
By using the DMRT, the extinction coefficient Ke and albedo ω used in equation (5) were
calculated. And then the radiance of each soil slab was calculated by the 4-stream fast
model. The radiance just below the soil/air interface was obtained by integrating the
radiance from bottom layer to the top layer. Finally, the apparent emission of soil media, Tbs
in equation (1) and (2), was obtained.
2.3 Surface roughness effects

When an electromagnetic wave reaches the air/soil interface, it suffers the reflection and
refraction due to the dielectric constant changing in the two sides of the interface. The
roughness of the interface divides the reflected wave into two parts, one is reflected in the
specular direction and another is scattered in all directions. Generally, the specular
component is often referred to as the coherent scattering component. And the scattered
component is known as the diffuse or noncoherent component, which consists of power
scattered in all directions but with a smaller magnitude than that of the coherent
component. Qualitatively, surface roughness increases the apparent emissivity of natural
surfaces, which is caused by increased scattering due to the increase in surface area of the
emitting surfaces. And it was demonstrated by many researches that the surface roughness
has a nonnegligible effects on the accuracy of soil moisture retrieval by spaceborne
microwave sensors (Oh & Key, 1998; Singh et al., 2003). In general, the surface roughness
effects are simulated by two ways: semi-empirical models and fully physical-based models.
(1) Semi-empirical models
The semi-empirical models are simply and do not cost too much computation efforts. The
parameters used in semi-empirical models are often derived from field observations.
Depending on the parameters involved, there are three different semi-empirical models: Q-
H model (Choudhury et al., 1979; Wang & Choudhury, 1981), Hp model (Mo & Schmugge,
1987; Wegmuller & Matzler, 1999; Wigneron et al., 2001) and Qp model (Shi et al., 2005).
(2) Fully physical-based model
In our algorithm, we simulated the land surface roughness effect using the Advanced
Integral Equation Model (AIEM) (Chen et al., 2003). AIEM is a physically-based model with
only two parameters: standard deviation of the height variations s (or rms height) and
surface correlation length l . AIEM is an extension of the integral Equation Model (IEM)
(Fung et al., 1992). It has been demonstrated that IEM has a much wider application range
for surface roughness conditions than other models such as the Small Perturbation Model
(SPM), Physical Optics Model (POM) and Geometric Optics Model (GOM). AIEM improves
the calculation accuracy of the scattering coefficient compared with IEM by retaining the
absolute phase term in the Green’s function.
By coupling AIEM with DMRT (DMRT-AIEM), this radiative transfer model for soil media
is fully physically-based. As a result, the parameters of DMRT-AIEM, such as the rms height,
correlation length and soil particle size, have clear physical meanings and their values can
be obtained either from field measurement or theoretical calculation.
2.4 Vegetation masking effects

The existence of canopy layers complicates the electromagnetic radiation which is originally
emitted solely by soil layers. The vegetation may absorb or scatter the radiation, but it will
also emit its own radiation. The effects of a vegetation layer depend on the vegetation
opacity τc and the single scattering albedo of vegetation ωc (Schmugge & Jackson, 1992). The
vegetation opacity in turn is strongly affected by the vegetation columnar water content Wc.
The relationship can be expressed as (Jackson & Schmugge, 1991):
τc =
b′λχWc (6)
cosθ
where λ is the wavelength, θ the incident angle, Wc the vegetation water content. χ and b'
are parameters determined by vegetation type.
The single scattering albedo, ωc, describes the scattering of the emitted radiation by the
vegetation. It is a function of plant geometry, and consequently varies according to plant
species and associations. The value of it is small in the low frequency microwave region
(Palosica & Pampaloni, 1988; Jackson & Oneill, 1990). In our algorithm, ωc is calculated by
ωc = ω0 ⋅ Wc (7)
The value of albedo parameter ω0 is decided empirically in current researches.Experimental

data for this parameter are limited, and values for selected crops have been found to vary
from 0.04 to about 0.12.
By combing the Tbs solved by equation (5), the surface reflectivity calculated by AIEM, the
vegetation opacity τc calculated by equation (6), and the vegetation single scattering albedo
ωc estimated by equation (7), a physical-based radiative transfer model was developed.
3. The Algorithm
The basis of our algorithm is a data base of brightness temperature and/or some indexes
calculated from brightness temperature. By searching the data base (or look up table) with
the satellite observation as the input, soil moisture and other related variables of interest can
be estimated quickly. Such high searching speed is the main reason why we adopt the look up
table method for soil moisture retrieval. The implementation of our algorithm consists of three
steps: (1) fixing the parameters used in the forward model; (2) generating a look up table by
running forward model; and (3) retrieving soil moisture by searching the look up table.
3.1 Parameterization
As in other physically-based algorithms, such as that developed by Njoku et al. (2003) and
the SCA developed by Jackson (1993), the parameters used in our algorithm have clear
physical meanings. This advantage derives from the strength of the forward radiative
transfer model. Before running the forward RTM to generate look up table, the parameters
should be confirmed at first. The parameters to be confirmed include rms height (s),
correlation length (l), soil particle sizes (r) and vegetation parameters such as χ and b′.
Currently, we can obtain these parameters through two methods: best-fitting method and a
parameter optimization method supported by a Land Data Assimilation System developed

in the University of Tokyo (LDAS-UT).
(1) Best-fitting method
For the region where in-situ soil moisture and temperature observation are available and
when such observation are also representative, we can use a best-fitting way to optimize
parameters. In order to simplify the calculation, low frequencies simulation and observation
were used. These parameters are optimized by minimizing the cost function:
n
J ( s , l , r , b′, χ ,) = ∑∑ ∑ ABS [TBsim (i , f , p ) − TBobs (i , f , p )] (8)
i =1 f p = H ,V
where the subscript sim denotes the model simulated value and obs is the observed value. n
is the number of samples used in the optimization. p denotes the polarization status: H for
horizontal polarization and V for vertical. f is some frequency in the long wavelength region
where the atmospheric effect may be ignored, such as 6.9, 10.7 and 18.7 GHz of AMSR-E, 1.4
GHz of SMOS and 19GHz of SSM/I.
(2) Parameter optimization by LDAS-UT
For most remote regions, in-situ representative observation is not available. A more general
parameter optimization method is proposed by Yang et al. (2007). In this method, long term
(around 2 months) meteorological field was used to drive a land surface model (Simple
Biosphere model, SiB2) to generate time series of soil moisture and temperature data set.
And then corresponding TB was simulated with our RTM. Since the land surface parameter
set (soil texture, porosity, particle size, roughness, etc.) was also used in SiB2, simulated soil
moisture and temperature and corresponding TB were varying as the parameter set
changes. By minimizing the difference between simulated TB and that of satellite
observation, the best parameter set can be obtained. The optimized parameters by LDAS-
UT, therefore, are depended on models and also influenced by the quality of forcing data.
The detail of this method can be found from Yang et al. (2009).
3.2 Look up table generation

After Step 1, the optimal parameter values are then stored in the forward RTM. We then run
the forward model by inputting all possible values of variables used in Equation (1), such as
soil moisture content, soil temperature, vegetation water content and atmosphere optical
thickness. A family of brightness temperatures is then generated. Based on this brightness
temperature database, we select brightness temperatures of special frequencies and
polarization to compile a lookup table or to calculate some indices to compile a lookup table.
For example, in order to partly remove the influences of physical temperature, the ratio of
TB at different frequencies and polarizations can be used. For instant, we can compile a look
up table by using the index of soil wetness (ISW) (Koike et al., 1996; Lu et al., 2009), and
Polarization Index (PI) (Paloscia & Pampaloni, 1988).
3.3 Soil moisture estimation

The lookup table generated in Step 2 is reversed to give a relationship which maps the
brightness temperature or indices obtained from satellite remote sensing data to the
variables of interest (such as soil moisture, soil temperature and vegetation water content).
Finally, we estimate soil moisture content by linear interpolation of the brightness

temperature or indices into the inverted lookup table.
4. Application to AMSR-E Data Set

We tested our algorithm by retrieving soil moisture and temperature from AMSR-E TB data
at a Coordinate Enhanced Observing Period (CEOP) (Koike, 2004) reference site in the
Mongolian Gobi. The results were validated by comparing with in situ measurements
byAutomatic Stations for Soil Hydrology (ASSH) and Automatic Weather Stations (AWS).
4.1 CEOP Mongolia reference site and AMSR-E match-up data set
The application region of this research is the AMPEX (ADEOS II Mongolian Plateau
EXperiment for Ground Truth) area. AMPEX has joined the CEOP as the Mongolia reference
site. AMPEX is designed to validate the AMSR and AMSR-E soil moisture algorithm(s). In
this area, meteorological and land hydrological factors are measured with very densely
installed instruments. AMPEX is located in the Mongolian Plateau, 235 km south of Ulan
Bator. The area stretches 160km in the longitudinal direction (106°E~108°30 ′E) and 120 km in
the latitudinal direction (45°30′N~47°N) on the Mandalgobi, where 6 AW Ss and 12 ASSHs
were installed. Figure 1 illustrates the distribution of observation sites in this area. For more
details of AMPEX, please visit the following website: http://home.hiroshima-
u.ac.jp/~ampex/hm/index-e.htm.
6
5
4
3 AWS
2
ASSH
1
0
A B C D E F G H I J
Fig. 1. Distribution of ASSH and AWS in AMPEX study area
The ASSHs provide soil moisture and temperature profile measured at two depths, 3 cm
and 10 cm below the surface. The AWSs provide soil moisture and temperature profiles
measured at four depths: 3 cm, 10 cm, 40 cm and 100 cm below the surface. The soil
moisture measurements used TDRs, and the soil temperature was measured by platinum
resistance thermometers.
Through matching the AMSR-E footprints to in situ stations, we generated a match-up data
set consisting of brightness temperature data observed by AMSR-E and in situ data
measured by SMTMS and AWS. The coverage of this data set is 2.5 by 2.5 degrees, with a
resolution 0.05 degrees for all frequencies. The in situ data consists of soil moisture and soil
temperature data. It is in the form of an image type and a text type. The text files record
AMSR-E brightness temperature and in situ data at each ground station. The in situ data
include observations made within 12 hours of the AMSR-E observation. In this research, the
mach-up in situ data at each AMSR-E satellite over passing is calculated by interpolating the
in situ data on the hour.
Based on the AMSR 2002 field experiment results, the soil bulk density in this region is 1.258
g•cm-3. The soil texture is obtained from the Net Primary Productivity (NPP) Database
(Chuluun & Ojima, 1996): a sand fraction of 0.6, a silt fraction of 0.2 and a clay fraction of 0.2.
There is sparse vegetation in study area. The vegetation water content was measured in June
and August, 2003. Based on this in situ observation, we found that the maximum vegetation
water content in our study area was 0.11 kg•m-2. It is a small value reflecting the sparse
vegetation coverage.
4.2 Best-fitting parameters for AMSR-E

With the AMSR-E match-up data, the land surface parameters can be obtained easily by
using the best-fitting methods. AMSR-E TB data obtained from low frequency channels
(6.925, 10.65 and 18.7 GHz) were used to optimize model parameters. Since the wavelength
of those channels is generally much larger than the diameter of atmospheric particles, the
atmospheric effect is negligible for the data measured with those channels.
As reported in the literature, it is reasonable to assume that there is little or no volume
scattering for soil moisture levels over 10% (Ulaby et al., 1986). So, we first used the data
observed on wet days to estimate the roughness parameters, rms height and correlation
length, in a best-fitting way. The particle size parameter could then be obtained by running
a coupled DMRT-AIEM model to best fit the data observed on dry days.
In order to run DMRT, we used uniform soil moisture and temperature vertical profiles with
the value observed at 3 cm depth. The bottom of the soil medium was set to be 1.0m (layer
thickness is 1cm) and the brightness temperature at the bottom was assumed to be the soil
physical temperature at that level, that is, the emissivity was equal to one. The downward
radiation from each soil layer, reflected at the bottom boundary, was not considered in this
study. The interactions at the boundaries between neighboring soil layers were also
neglected because of the vertically uniform soil moisture and temperature profiles.
First, we used the AIEM model to best fit several wet day observations by changing rms
height (s) and correlation length (l). Second, employing this set of s and l, we could obtain
the surface emissivity for all observations. Third, with some dry day observations, we could
best fit the particle size parameter using the DMRT-AIEM model. Finally, we calculated
brightness temperature from April 10th 2003 to April 30th 2004 with best-fitting parameters.
Here, we use the A3 station as an example, to introduce the whole procedure and the result.
Information about the data we used to calibrate the model is listed in following table.
cases Number of days Mv range (%) T(3cm) range Period

Wet 32 10~20 275.26~291.25 May. 12th ~ Aug. 20th ,03
Dry 40 1~7 273.1~293.8 Apr.14th~Jul.13th,03
ALL 254 1~20 270~293.8 Apr. 12th,03 ~Apr, 30th, 04
Table 1. Data used for parameter optimization
The calibrated parameter values of AIEM with consideration of shadowing effects are: s =
0.46 cm; l = 1.03 cm. Then, with this set of s and l, using data for 40 dry days, we best fit the
particle size parameter as in table 2:
F(GHz) 6.925 10.65 18.7 23.8 36.5

Wave Length λ (cm) 1.997 1.298 0.739 0.581 0.379
R(cm) 0.45 0.307 0.165 0.126 0.084
R/λ 0.104 0.109 0.103 0.100 0.102
Table 2. Best-fitting particle size parameters in Mongolia
As in Table 2, the best-fit particle sizes change at different frequencies: longer wavelengths
are matched with larger particle sizes. However, the ratio between the best-fit radius and the
wavelength in the sand is nearly constant. Therefore, we call the best-fit radius the effective
radius. The effective radii are generally larger than the physical values, consistent with
similar results reported by Kendra and Sarabandi (1999).
4.3 AMSR-E Look up table generation

Based on the best-fitted parameter sets, we build a lookup table composed of the soil
physical temperature, soil moisture content, brightness temperature at 10.65GHz vertical
polarization and an index dTB calculated as follows:
dTB = TB (18.7, H ) − TB (10.65, H ) (9)
Look UP Table of 18 and 10 GHz

40
35
30
25
18H-10H
20 Mv=40%
15
T=303
10
T=273
T=270
5
Mv=2%
0
210 220 230 240 250 260 270 280 290 300
10v
t=270 mv=10 mv=30

t=303 mv=15 mv=40
mv=2 mv=20 a3obs
mv=5 mv=25
Fig. 2. Lookup table for the AMSR-E soil moisture retrieval algorithm
The lookup table of our AMSR-E algorithm is shown in Figure 2. It covers a region in which
soil moisture content varies from 2% to 40%, and soil physical temperature varies from 270K
to 303K. Compared with in situ observation values, this range is large enough to include all
of the actual soil moisture and temperature states in Mongolia.
Since the one-to-one relationship in our lookup table is very clear, it becomes simple to reverse
the lookup table, so that the soil moisture can easily be estimated from the AMSR-E data set.
4.4 Retrieval soil moisture from AMSR-E

In this study, we retrieved soil moisture data for the period from July to August, 2003. The
estimation is shown in Figure 3 for (a) time variation and (b) accuracy comparison.
(a) Time variation of soil moisture (ASSH stations' average)

30 0
5
25
10
20
15
Rain(mm)
Mv(%)
15 20
25
10
30
5
35
0 40
7-1 7-11 7-22 8-2 8-12 8-23
rainfall Estimated Observed
Fig. 3(a). Time series of retrieved soil moisture, observed soil moisture and precipitation.
(b) Comparison of SMC estimation (ASSH stations' average)

25
2
R = 0.3953
20
Estimated (%)
15
10
0
0 5 10 15 20 25
Observed (% )
Fig. 3(b). Comparison of retrieval results with in situ observation
It is clear from figure 3 that the algorithm gives a reliable soil moisture content estimate in
both tendency and amplitude. The value of R-square is 0.3953, and the Standard Error of the
Estimate (SEE) is 3.8%. From figure 3(a), we find some overestimation around Aug. 4, 14
and 20, when moderate rainfall (5~10 mm) occurred. Such errors can be attributed partly to
the difference between the TDR sensor depth and the penetration depth of the X band and
Ku band. Moderate rainfall makes the soil surface much wetter than the soil 3cm below the
surface where the TDR sensors were located. Such vertical heterogeneity of soil moisture in
the first 3cm of soil was not considered in our algorithm. On the other hand, the wet surface
situation decreases the penetration depth dramatically. The combination of these reasons
makes our algorithm estimate higher soil moisture content than the in situ observations for
moderate rainfall periods.
One advantage of our proposed algorithm is that it estimates soil physical temperature and
soil moisture simultaneously. This is important for studies involving energy and water
budget, such as studies of land surface processes and of weather forecasting.
The average retrieved physical temperature for ASSH stations is shown in figure 4(a) and
figure 4(b). As in soil moisture comparisons, the algorithm effectively retrieved physical
temperature on average for 10 stations. The value of R-square is 0.5458, and the value of SEE
is 4.4K. As with our soil moisture analysis, the overestimation of daily temperature variation
can also be explained partly as the results of different observation depths.
Physcial temperature for ASSH stations' average

315
DMRT-AIEM Obs
310
305
300
T(K)
295
290
285
280
7/1 7/11 7/21 7/31 8/10 8/20 8/30
Time
Fig. 4(a). Time series of retrieved and in situ observations of soil physical temperature.
Comparison of physical temperature for ASSH stations'

average
305
300
R2 = 0.5458
295
Est(K)
290
285
280
280 285 290 295 300 305
Obs(K)
Fig. 4(b) Comparison of algorithm-estimated soil physical temperature with in situ

observation
5. Application to SSM/I
Since the parameters used in our RTM have clear physical meaning, they are independent
on the configuration of radiometers. The parameters used in AMSR-E soil moisture retrieval
therefore can be directly used to the SSM/I data set, in the same region. In this test, we first
checked the accuracy of TB simulation of our DMRT-AIEM model. And then a look up table
for SSM/I data was generated and soil moisture was retrieved.
5.1 SSM/I TB Validation

Using the parameters optimized by AMSR-E math-up data set, with the in-situ observed soil
moisture and temperature as input, we run the DMRT-AIEM model to generate TB at 19.35
and 37.0 GHz, two frequencies operated by SSM/I. The SSM/I TB validation results were
shown in figure 5, for the A3 station, during the period from Jul. 1st to Jul. 30th, 2003.
Simulation of TB at 19GHz for SSM/I A3 station V-P Simulation of TB at 19GHz for SSM/I A3 station H-P
295 285
2 2
R = 0.6602 R = 0.382
290 280
285 275
Sim
Sim
280 270
275 265
270 260
265 255
265 270 275 280 285 290 295 255 260 265 270 275 280 285
Obs Obs
Simulation of TB at 37GHz for SSM/I A3 station V-P Simulation of TB at 37GHz for SSM/I A3 station H-P
295 285
2
R = 0.6285 2
280 R = 0.331
290
285 275
Sim
Sim
280 270
275 265
270 260
265 255
265 270 275 280 285 290 295 255 260 265 270 275 280 285
Obs Obs
Fig. 5. Comparison of simulated brightness temperature with the one observed by SSM/I
From figure 5, it is clear that, for the vertical polarization, the TB simulated by DMRT-AIEM
is in a good correlation with the SSM/I observation, with slight underestimation for 19 GHz
and overestimation for 37 GHz. For the horizontal polarization, the performance of our RTM
is not so good. Statistically, the Average Absolute Error (AAE, see equation (10)) and the
square of correlation coefficient between observed brightness temperature and simulated
one are listed in table 3.
Channel 19V 19H 37V 37H

AAE (K) 4.34 4.41 3.45 3.85
R2 0.66 0.38 0.63 0.33
Table 3. AAE and correlation coefficient of DMRT-AIEM model for SSM/I data
n
AAE = {∑ [ ABS (TBS i ,V − TBOi ,V ) + ABS (TBSi ,H − TBOi ,H )]} /(n * 2)
(10)
i =1
where, TBS is simulated brightness temperature, TBO is observed brightness temperature

by spaceborne sensor; n is number of samples.
5.2 Look up table of SSM/I

Through the TB validation, it was confirmed that our DMRT-AIEM model was able to
produce reasonable TB at vertical polarization channels of SSM/I. But there was some gaps
between the simulated TB with the one observed by SSM/I, especially for horizontal
polarization channels. Moreover, as we know, the atmosphere effects should be considered
for 37 GHz. All of these make it difficult to build a look up table with the same way used for
AMSR-E. Authors proposed a simple solution by nudging the SSM/I TB data to fit the
simulation and by using PI and ISW indexes to generate a look up table. Detail of the TB
adjustments and look up table generation can be found from (Ohta et al., 2007). Figure 6
shows the look up table for SSM/I, in which the PI calculated from 19 GHz and the ISW
calculated from the horizontal polarization of 37 and 19 GHz were used. The black points
represent the PI and ISW values calculated from corresponding TB data observed by SSM/I.
0.25
Mv = 0
Mv = 2
0.20 Mv = 5
Mv = 10
Mv = 15
0.15 Mv = 20
Wc = 0.0
ISW (37H, 19H)
Wc = 0.1
0.10 SSM/ I
0.05
0.00
- 0.05
0.00 0.02 0.04 0.06 0.08 0.10 0.12 0.14 0.16
PI (19V, 19H)
Fig. 6. Diagram of the Look Up Table for SSM/I soil moisture algorithm
5.3 Soil moisture retrieval from SSM/I
30
Obs. SSM/ I
25
20
Mv [%]
15
10
0
07/ 1 07/ 11 07/ 21 07/ 31 08/ 10 08/ 20 08/ 30
Fig. 7. Comparison of soil moisture retrieved from SSM/I with in-situ observation
By using the PI-ISW look up table, soil moisture was estimated from SSM/I TB data. The
results are shown in figure 7, for the period from July to August, 2003. The line represents
the in-situ soil moisture observation; the open cycles are SSM/I soil moisture estimate. From
this figure, it is clear that the performance of SSM/I soil moisture retrieval algorithm is very
good. So, it is feasible to get reasonable soil moisture estimation form SSM/I with the helps
from AMSR-E. But we must keep it in mind that the TB adjustment was applied to the
SSM/I data. Such good performance of SSM/I algorithm is therefore just for the special
cases where the in-situ observation are available and the appropriate TB adjustment are
possible.
6. Conclusions
Spatial distributed soil moisture information is an essential parameter for hydrological,
meteorological and ecological studies. This paper presents the structure and contents of a
soil moisture retrieval algorithm for the spaceborne passive microwave remote sensing. This
algorithm was validated by using the AMSR-E match-up data set at CEOP Mongolia
reference site. Comparing to the in-situ observation, reliable surface soil moisture was
retrieved by the algorithm.
The transferability of our algorithm was also tested by using SSM/I TB data. At first, it was
demonstrated that the forward RTM of our algorithm was capable to represent the SSM/I
TB data only after the parameters were calibrated by AMSR-E data set. And then, with some
adjustments to the SSM/I TB data, reasonable surface soil moisture was also retrieved from
SSM/I data by our algorithm.
The results presented in this paper clearly show that we had built a bridge between the
parameters retrieved from AMSR-E and those for SSM/I. With some further consideration
about the difference between AMSR-E and SSM/I, e.g. the footprint size and the observation
patterns, it is believed that our algorithm could provide a possibility to use the long
historical global data observed by SSM/I. Moreover, it is possible to extend our algorithm to
other available radiometers. And then, we can merge multi-sensor or/and multi-satellite
observations to generate a long term global historical soil moisture product. Such a long term
historic data set should be much useful for large scale hydrological and climatologic studies.
As mentioned in section 1, the retrieval of surface soil moisture is physically limited by the
current satellite instruments which are operating at high frequencies. The low frequency, i.e.
L-band, passive microwave soil moisture observation will firstly be available through the
launch of Soil Moisture Ocean Salinity (SMOS) mission of ESA (Kerr, et al., 2001). NASA
will provide a combined L-band radiometer and L-band radar observation through the Soil
Moisture Active and Passive (SMAP) mission (Entekhabi, et al., 2008). Since the
configuration of our algorithm was not specified to any sensors, it is also possible to apply
our algorithm to the incoming L-band radiometers. We hope this algorithm will be helpful
for these future soil moisture missions and for connecting current available C-band and X-
band observations to the L-band observations.
7. Acknowledgements
This study was carried out as part of the Coordinated Enhanced Observing Period (CEOP)
and Verification Experiment for AMSR/AMSR-E funded by the Japanese Science and
Technology Corporation for Promoting Science and Technology Japan and by the Japan
Aerospace Exploration Agency (JAXA). The authors express their great gratitude to them.
8. References
Beljaars, A.C.M.; Viterbo P.; Miller M.j. & Betts A.J. (1996). The anomalous rainfall over the
United States during July 1993: sensitivity to land surface parameterization and soil
moisture anomalies. Monthly Weather Review, 124, 362-383.
Burke, W.J.; Schmugge, T.J. & Paris, J.F. (1979). Comparison of 2.8 and 21 cm microwave
radiometer observations over soils with emission model calculation. Journal of
Geophysics Research, 84, 287-294.
Chen, K.S.; Wu, T.D.; Tsang, L.; Li, Q.; Shi, J.C. & Fung, A.K. (2003). Emission of rough
surfaces calculated by the integral equation method with comparison to three-
dimensional moment method simulations. IEEE Transactions on Geoscience and
Remote Sensing, 41, 90-101.
Choudhury, B. J.; Schmugge, T.J.; Chang, ATC & Newton N.R. (1979). Effect of surface
roughness on the microwave emission from soils. Journal of Geophysical Research, 84,
5699-5706.
Chuluun, Togtohyn & Ojima, D.S. (1996). Simulation studies of grazing in the Mongolian
steppe, Proceedings of the fifth International Rangeland Congress, Society of Range
Management, Denver, CO, USA.
Dobson, M.C.; Ulaby, F.T.; Hallikainen, M.T. & Elrayes, M.A. (1985). Microwave dielectric
behavior of wet soil .2. dielectric mixing models. IEEE Transactions on Geoscience and
Remote Sensing 23, 35-46.
Delworth, T. & Manabe, S. (1988). The influence of potential evaporation on the variability
of simulated soil wetness and climate. Journal of Climate, 13, 2900−2922.
Entekhabi, D.; Rodrigues-Iturbe I. & Castelli F. (1996). Mutual interaction of soil moisture
state and atmospheric processes. Journal of Hydrology, 184, 3-17.
Entekhabi, D.; Njoku, E.; Oneill, P. E.; Spencer, M.; Jackson, T.; Entin, J.; Im, E. & Kellogg, K.
(2008). The Siol Active Passive Mission (SMAP). Proceedings of the IEEE 2008
International Geoscience and Remote Sensing Symposium (IGARSS’08), 3, 1–4, July,
2008, Boston, Doi: 10.1109/IGARSS.2008.4779267.
Fung, A.K.; Li, Z.Q. & Chen, K.S. (1992). Backscattering from a randomly rough dielectric
surface. IEEE Transactions on Geoscience and Remote Sensing, 30, 356-369.
Gloersen, P. & Barath F. T. (1977). A Scanning Multichannel Microwave Radiometer for
Nimbus-G and SeaSat-A. IEEE Journal of Oceanic Engineering, 2:172-178.
Henyey, L. C. & Greenstein, J. L. (1941). Diffuse radiation in the galaxy, The Astrophysics
Journal, 93, 70–83.
Hipp J.E. (1974). Soil electromagnetic parameters as functions of frequency, soil density, and
soil moisture, Proceedings of the IEEE, 62:98–103
Hollinger, J.P.; Peirce, J.L. & Poe, G.A. (1990). SSM/I Instrument Evaluation. IEEE
Transactions on Geoscience and Remote Sensing, 28(5), 781-790.
Jackson, T. J. & Oneill, P. E. (1990). Attenuation of Soil Microwave Emission by Corn and
Soybeans at 1.4 Ghz and 5 Ghz. IEEE Transactions on Geoscience and Remote Sensing,
28(5), 978-980.
Jackson, T.J. & Schmugge, T.J. (1991). Vegetation effects on the microwave emission of soils.
Remote Sensing of Environment, 36, 203-212.
Jackson, T.J. (1993). Measuring surface soil moisture using passive microwave remote
sensing. Hydrology Processes, 7, 139-152.
Kawanishi,T.; Sezai, T.;Ito Y.; Imaoka, K.; Takeshima, T.; Ishido, Y.; Shibata, A.; Miura, M.;
Inahata, H. & Spencer Roy W. (2003). The advanced microwave scanning
radiometer for the earth observing system (AMSR-E), NASDA’s contribution to the
EOS for global energy and water cycle studies. IEEE Transactions on Geoscience and
Remote Sensing, 41, 184-194.
Kendra, J. R. & Sarabandi, K. (1999). A hybrid experimental theoretical scattering model for
dense random media. IEEE Transactions on Geoscience and Remote Sensing, 37, 21-35.
Kerr, Y. H.; Waldteufel, P.; Wigneron, J. P.; Martinuzzi, J. M.; Font, J. & Berger, M. (2001).
Soil moisture retrieval from space: The Soil Moisture and Ocean Salinity (SMOS)
mission. IEEE Transactions on Geoscience and Remote Sensing, 39, 1729-1735
Koike, T.; Tsukamoto, T.; Kumakura, T. & Lu, M. (1996). Spatial and seasonal distribution of
surface wetness derived from satellite data, Proceedings of International workshop on
macro-scale hydrological modeling, 87-96.
Koike, T. (2004). The Coordinated Enhanced Observing Period – an initial step for integrated
global water cycle observation, WMO Bulletin, 53(2), 1-8.
Liu, G. (1998). A fast and accurate model for microwave radiance calculations, Journal of
Meteorology Society of Japan, 76(2), 335–343.
Lu, H.; Koike, T.; Hirose, N.; Morita, M.; Fujii, H.; Kuria, D. N.; Graf, T. and Tsutsui, H.
(2006). A basic study on soil moisture algorithm using ground-based observations
under dry condition. Annual Journal of Hydraulic Engineering, Japan Society of Civil
Engineer, 50, 7-12.
Lu, H.; Koike, T.; Fujii, H.; Ohta, T. & Tamagawa, K. (2009). Development of a Physically-
based Soil Moisture Retrival Algorithm for Spaceborne Passive Microwave
Radiometers and its Application to AMSR-E, Journal of The Remote Sensing Society of
Japan, 29(1), 253-261.
Mironov, V.L.; Dobson, C.; Kaupp, V.H.; Komarov, V.A. & Kleshchenko, V.N. (2004).
Generalized refractive mixing dielectric model for moist soils. IEEE Transactions on
Geosciences and Remote Sensing, 42:773–785. doi:10.1109/TGRS.2003.823288
Mo, T. & Schmugge, T. J. (1987). A Parameterization of the Effect of Surface-Roughness on
Microwave Emission. IEEE Transactions on Geoscience and Remote Sensing, 25(4), 481-
486
Njoku, E.G. & Kong, J.A. (1977). Theory for passive microwave remote sensing of near-
surface soil moisture. Journal of Geophysical Research, 82, 3108-3118.
Njoku, E.G. & Entekhabi, D. (1996). Passive microwave remote sensing of soil moisture.
Journal of Hydrology, 184, 101-129.
Njoku, E.G.; Jackson, T.J.; Lakshmi, V.; Chan, T.K. & Nghiem, S.V. (2003). Soil moisture
retrieval from AMSR-E. IEEE Transactions on Geoscience and Remote Sensing, 41, 215-
229.
Oh, Y. & Kay, Y. C. (1998). Condition for precise measurement of soil surface roughness.
IEEE Transactions on Geoscience and Remote Sensing, 36(2), 691-695.
Ohta, T; Koike, T.; Lu, H.; Kuria, D.N.; Tsutsui, H.; Graf, T.; Kaihotsu, I; Davaa, G. &
Matsuura, N. (2007). Development of a special sensor microwave imager (SSM/I)
algorithm for long term monitoring of soil moisture, Annual Journal of Hydraulic
Engineering, Japan Society of Civil Engineer, 51, 205-210.
Owe, M.; De Jeu, R.A.M. & Holmes, T.R.H. (2008). Multi-sensor historical climatology of
satellite-derived global land surface moisture. Journal of Geophysics Research, 113,
doi:10.1029/2007JF000769.
Paloscia, S. & Pampaloni, P. (1988). Microwave polarization index for monitoring vegetation
growth. IEEE Transactions on Geoscience and Remote Sensing, 26, 617-621.
Paloscia, S.; Macelloni, G.; Santi, E. & Koike, T. (2001). A multifrequency algorithm for the
retrieval of soil moisture on a large scale using microwave data from SMMR and
SSM/I satellites, IEEE Transactions on Geoscience and Remote Sensing, 39, 1655~1661.
Paloscia, S.; Macelloni, G. & Santi E. (2006). Soil moisture estimates from AMSR-E brightness
temperatures by using a dual-frequency algorithm, IEEE Transactions on Geoscience
and Remote Sensing, 41(11), 3135-3144.
Prigent, C.; Aires, F.; Rossow, W. B. & Robock, A. (2005). Sensitivity of satellite microwave
and infrared observations to soil moisture at a global scale: Relationship of satellite
observations to in situ soil moisture measurements. Journal of Geophysical Research,
110, D07110. doi:10.1029/2004JD005087
Ray, P. S. (1972). Broadband Complex Refractive Indices of Ice and Water. Applied Optics, 11
(8), 1836-1844.
Said, S., Kothyari, U.C. & Arora, M.K. (2008). ANN-based soil moisture retrieval over bare
and vegetated areas using ERS-2 SAR data. Journal of Hydrologic Engineering, 13,
461-475.
Schar, C.; Luthi, D.; Beyerle, U. & Heise, E. (1999). The soil-precipitation feedback: A process
study with a regional climate model. Journal of Climate, 12, 722−741.
Schmugge, T.J. & Choudhury B.J. (1981). A comparison of radiative transfer models for
predicting the microwave emission from soils. Radio Science, 16, 927-938.
Schmugge, T.J.; Wang, J.R. & Asrar G. (1988). Results from the Push Broom Microwave
Radiometer flights over the Konza Prairie in 1985. IEEE Transactions on Geoscience
and Remote Sensing, 26(5), 590–597.
Schmugge, T. J. & Jackson, T. J. (1992). A Dielectric Model of the Vegetation Effects on the
Microwave Emission from Soils. IEEE Transactions on Geoscience and Remote Sensing,
30(4), 757-760.
Seneviratne, S. I.; Luthi, D.; Litschi, M. & Schar, C. (2006). Land–atmosphere coupling and
climate change in Europe. Nature, 443(7108), 205−209.
Shi, J. C.; Jiang, L.M.; Zhang, L.X.; Chen, K.S.; Wigneron, J.P. & Chanzy, A. (2005). A
parameterized multifrequency-polarization surface emission model, IEEE
Transactions on Geoscience and Remote Sensing, 43, 2831-2841.
Shibata, A.; Imaoka, K. & Koike, T. (2003). AMSR/AMSR-E level 2 and 3 algorithm
developments and data validation plans of NASDA. IEEE Transactions on Geoscience
and Remote Sensing, 41, 195-203
Singh, D.; Sing, K. P.; Herlin, I. & Sharma, S.K. (2003). Ground-based scatterometer
measurements of periodic surface roughness and correlation length for remote
sensing. Advances in Space Research, 32(11), 2281-2286.
Singh, D. & Kathpalia, A. (2007). An efficient modeling with GA approach to retrieve soil
texture, moisture and roughness from ERS-2 SAR data. Progress in Electromagnetics
Research-Pier, 77, 121-136.
Tsang, L. & Kong, J. A (1977). Theory for thermal microwave emission from a bounded
medium containing spherical scatterers, Journal of Applied Physics, 48, 3593-3599.
Tsang, L. & Kong, J. A (2001). Scattering of Electomagnetic Waves: Advanced Topics, Wiley, New
York.
Ulaby, F.T.; Razani, M. & Dobson, M.C. (1983). Effects of vegetation cover on the microwave
radiometric sensitivity to soil moisture. IEEE Transactions on Geoscience and Remote
Sensing, 21, 51-61.
Ulaby, F.T. ; Moore, R.K & Fung A.K. (1986). Microwave Remote Sensing: Active and Passive,
Artech House, Norwood, MA.
Verstraeten, W.W.; Veroustraete, F.; van der Sande, C.J.; Grootaers, I. & Feyen, J. (2006). Soil
moisture retrieval using thermal inertia, determined with visible and thermal
spaceborne data, validated for European forests. Remote Sensing of Environment, 101,
299-314.
Wagner, W.; Scipal, K.; Pathe, C.; Gerten, D.; Lucht, W. & Rudolf, B. (2003). Evaluation of the
agreement between the first global remotely sensed soil moisture data with model
and precipitation data. Journal of Geophysical Research, 108(D19), 4611- ,
doi:10.1029/2003JD003663
Wang, J.R. & Schmugge, T.J. (1980). An empirical model for the complex dielectric
permittivity of soils as a function of water content. IEEE Transactions on Geoscience
and Remote Sensing, 18, 288-295.
Wang, J. R. & Choudhury, B. J. (1981). Remote-Sensing of Soil-Moisture Content over Bare
Field at 1.4 Ghz Frequency. Journal of Geophysical Research-Oceans and Atmospheres,
86 (Nc6), 5277-5282.
Wegmuller, U. & Matzler, C. (1999). Rough bare soil reflectivity model. IEEE Transactions on
Geoscience and Remote Sensing, 37, 1391-1395.
Wen, B.; Tsang, L.; Winebrenner, D. P. and Ishimura, A. (1990). Dense media radiative
transfer theory: comparison with experiment and application to microwave remote
sensing and polarimetry, IEEE Transactions on Geoscience and Remote Sensing, 28, 46-
59.
Wigneron, J. P.; Laguerre, L. & Kerr, Y. (2001). A simple parameterization of the L-band
microwave emission from rough agricultural soils. IEEE Transactions on Geoscience
and Remote Sensing, 39(8), 1697-1707.
Yang, K.; Watanabe, T.; Koike, T.; Li, X.; Fujii, H.; Tamagawa, K.; Ma Y. & Ishikawa H.
(2007). An auto-calibration system to assimilate AMSR-E data into a land surface
model for estimating soil moisture and surface energy budget. Journal of
Meteorology Society of Japan, 85A, 229-242.
Yang, K.; Koike, T.; Kaihotsu, I. & Qin, J. (2009). Validation of a Dual-pass Microwave Land
Data Assimilation System for Estimating Surface Soil Moisture in Semi-arid
Regions, Journal of Hydrometeorology, 10(3), 780-794.
Multiwavelength Polarimetric Lidar for Foliage Obscured Man-Made Target Detection 391
18
X
Multiwavelength Polarimetric Lidar for Foliage

Obscured Man-Made Target Detection
Songxin Tan1 and Jason Stoker2
1South
Dakota State University
2USGS Earth Resources Observation and Science (EROS)
USA
1. Introduction
LIDAR is an acronym for LIght Detection And Ranging. It is also called optical radar, laser
radar or ladar under different application scenarios. The fundamental principle of lidar is
similar to that of the microwave radar. However, lidar operates in optical frequency while
radar works in microwave frequency. There are many unique ways that light interacts with
matter, hence lidar differs in many respects from radar. The interaction between light and
matter has different physical mechanisms, which include Rayleigh scattering, Mie
scattering, Raman scattering, resonance scattering, fluorescence, absorption, and differential
absorption and scattering, etc. (Measures, 1992). The interaction is detected by lidar, and the
information is used to obtain the target characteristics. Lidar can be used to characterize a
wide variety of targets, from a hard target such as a vehicle or a tree, to a pure phase object
such as the atmosphere or the cloud. It has been used in many remote sensing applications,
including vegetation and forest monitoring, land-use and land-cover detection, land
management, cloud and aerosol detection, sea shore bathymetry, global ice sheet
monitoring, among many others. The information provided by lidar improves our under-
standing of the environment, the ecological system, and the biocomplexity of the planet
Earth.
A typical lidar system consists of a laser source, transmitting optics and receiving optics,
photodetector, analog–to-digital (A/D) converter, signal and data processor, and output
device. A block diagram of a lidar system is illustrated in Figure 1. The sequence of lidar
operation is described as follows. The laser beam is sent out through the transmitter towards
the target, and the transmitted light interacts with the media and is reflected/scattered back
toward the receiving optics. The photodetector then converts the received light signal into
an analog electrical signal, which is further transformed into a digital signal by an A/D
converter. The signal is analyzed by the signal and data processor, and the final result is sent
to an output device for display.
While most existing commercial lidar systems employ one single laser wavelength and lack
the polarimetric measurement capability, a new type of lidar, the multiwavelength
polarimetric lidar, was introduced in vegetation remote sensing (Tan & Narayanan, 2004).
The multiwavelength lidar is able to provide more spectral information about the target
under study. For example, it is well known that both coniferous and deciduous trees have
similar reflectivity in visible wavelengths, but deciduous trees have distinctively higher
reflectivity at near-infrared compared to coniferous trees. As a result, such spectral
information at several laser wavelengths can be used for tree species classification
(Jaaskelainen et al., 1994). On the other hand, polarimetric lidar provides polarization state
of the backscattered laser which contains information such as target composition, surface
roughness, water content, etc. The information is essential in order to understand the target
characteristics. For example, a previous study found that deciduous trees had distinct
depolarization signature at 1064-nm while coniferous trees had such signature at 532-nm
(Kalshoven & Dabley, 1993). At this time, the understanding of polarimetric scattering
property of vegetation at optical wavelengths is very limited. Both theoretical and
experimental studies (Ma et al., 1990) are needed in order to fill the gap. As a new concept
lidar sensor, the potential of multiwavelength polarimetric lidar is yet to be explored.
Beam expander
Laser Transmitting optics
Target
Receiving
telescope
Photo-
detector
Output device
Signal and data

A/D
processor
Fig. 1. Illustration of a typical lidar system.
2. Foliage Obscured Target Detection

Many remote sensing applications demand for the foliage obscured man-made objects
detection. These applications may range from civilian applications, such as search and
rescue missions in densely forested areas, to military applications, such as detecting
camouflaged enemy vehicles. Traditional passive remote sensing has very limited capability
in foliage obscured target detection. Active sensors such as the foliage penetration (FOPEN)
radar has been proven successful in many applications and has become the primary tool
when foliage obscured target detection is needed. Studies on foliage obscured target
detection using radar technology have been reported in the literature (Nashashibi & Ulaby,
2005; Xu & Narayanan, 2003). The reason is that microwave has longer wavelength and can
penetrate the vegetation canopy. The penetration ability of microwave or lightwave is
related to its wavelength. Generally speaking, the longer the wavelength, the better it
penetrates. Lidar is not traditionally considered when discussing foliage penetration

(FOPEN) applications. This is due to the fact that the laser wavelength is much shorter than
the microwave wavelength. Therefore, a laser beam is considered as incapable of
penetrating through the vegetation canopy. However, multiwavelength polarimetric lidar
can be a viable alternative even though it has received little attention.
Nevertheless, gaps do naturally occur inside a forest. These include the gaps inside a single
tree crown and gaps in between different tree canopies. Consequently, the target is usually
not 100% covered by the foliage. If the laser beam or even a portion of the laser beam is able
to go through the gaps and hit the target and return to the detector, then it is still possible to
detect the hidden target using a lidar sensor. For example, a previous study demonstrated
that a scanning lidar system could be used to detect the hard obstacles on the ground which
were non-drivable and the foliage area which was drivable for a robotic vehicle (Castano &
Matthies, 2003). It demonstrated that lidar does have certain foliage penetration capability.
With its inherent shorter wavelength, lidar provides better range resolution and better
spectral purity in comparison with radar. Therefore, it can offer advantage in certain target
detection and identification applications.
In this study, a multiwavelength polarimetric lidar was used to detect a man-made target
obscured by trees. The multiwavelength polarimetric lidar provides several advantages in
target detection. Firstly, the spectral reflectance of a man-made target is usually different
from that of the foliage. Even when certain type of camouflage cover or paint is applied, it is
extremely difficult to match the natural foliage reflectance pattern at several different
wavelengths. Secondly, the polarimetric scattering properties of the man-made target and
the foliage are generally quite different. Since the polarimetric scattering property at an
optical wavelength is determined by many factors such as the material composition, target
shape, surface roughness, etc., it is extremely difficult to produce a camouflage that has the
same polarimetric scattering feature as the foliage. Therefore, the detection rate and
identification success can be improved using the combined information from spectral and
polarimetric features.
3. Lidar System Description

3.1 MAPL lidar system
The lidar system built and used in this study is the Multiwavelength Airborne Polarimetric
Lidar (MAPL) system. The MAPL is composed of several sub-systems: the laser source, the
optical receiver assembly and the data acquisition and processing hardware and software.
The MAPL system block diagram is shown in Figure 2.
The MAPL system functions as follows. A computer controls a digital delay/pulse generator
through an RS-232 interface. TTL signals are then generated by the delay generator. One
TTL signal is sent to trigger the Q-switch of the laser; the other TTL signal is delayed and
sent to trigger the digitizer. The delay time is determined by the range of the target. The
laser simultaneously sends out two laser pulses at 1064-nm and 532-nm. On top of the laser
head, there is a silicon PIN detector attached to monitor the relative laser output. The laser
pulses are backscattered by the target, and received by four photomultiplier tube (PMT)
detectors. These detectors enable detection at both co-polarized and cross-polarized
direction at the two wavelengths. Outputs from the PMT detectors are digitized by a high-
speed analog to digital (A/D) converter and then sent to the computer through a PCI
interface card. The ranging ability is achieved by precisely timing the transmission and
reception of the laser pulse. Data are stored in a hard drive and post-processed by the
computer to extract information.
Fig. 2 Block Diagram of the MAPL system.
3.1.1 Laser
The MAPL employs a Big Sky CFR-400 Nd:YAG laser. This is a multiwavelength pulsed
laser system, and is designed to be compact, rigid and stable, and is thus suitable for field
applications. The laser has two outputs, one at 1064-nm and one at 532-nm, emitting from
two separate apertures. Special optical wedge assembly is provided inside the laser module
to ensure that the two beams are parallel. The two beams are linearly polarized in the
horizontal direction. The polarization purity is greater than 100:1 at both wavelengths. Each
beam has a divergence angle of about 4 mrad, which will form a laser footprint with a
diameter of about 4 m at a distance of 1000 m. The highest pulse repetition rate is 10 Hz. The
laser output energy is adjustable and typically 30 mJ/pulse at each wavelength is used. An
attenuator can also be mounted in front of the laser head to decrease the laser output energy
for safety considerations. The near field beam diameter is about 6 mm at each wavelength.
The pulse width is 10 ns, which yields a range resolution of 1.5 m. To ensure accurate
timing, trigger of flash lamp and Q-switch are separately controllable.
A Si PIN detector from Thorlabs is sealed using an O-ring and attached externally on top of
the laser head, and is used to monitor the relative output laser energy. This serves as a
reference in order to calibrate the received laser return. The power supply unit and cooling
group unit are integrated into a Mini-ICE (integrated cooler and electronics) subsystem. It is
a compact and rugged sub-system designed to operate in a harsh outdoor environment. A
remote control box is provided to run all the controls on the laser. An RS-232 interface is also
provided for computer controls.
3.1.2 Receiving Optics

A total number of four receiving channels are employed in the MAPL system. There are two
channels, one co-polarized and one cross-polarized, for the 1064-nm light; and the same
configuration is used for the 532-nm light. Adding more receiving channels can enable the
measurement of the Stokes parameters, which are used to fully describe the polarization
state of the backscattered light. However, the circularly polarized light from the vegetation
and the ground is very weak under linearly polarized laser illumination (Kalshoven &
Dabney, 1993). Useful information is extremely difficult to retrieve from the circular
polarization channel because of the low signal to noise ratio (Tan, Narayanan & Kalshoven,
2001). The other limiting factor comes from the system size limit. More receiving channels
require a larger system which is impractical for field deployment. Thus, fully polarimetric
ability is not currently integrated in the MAPL system.
The optical receiver apertures are 25 mm in diameter. There is a linear polarizer, an optical
interference filter, and a single plano-convex lens with a focal length of 65 mm, inside each
receiver, as shown in Figure 3. The received light is diffused onto the PMT detection plane,
instead of imaging onto it. The advantage of the so-called straight-through structure is to
minimize the reflection and refraction optics inside the receivers, thus reducing possible
modifications to the polarization state of the backscattered light (Kalshoven & Dabney,
1993). The 1064-nm polarizers inside the receivers are from Corning Polarcor and the 532-
nm polarizers are from Polaroid, both of them have polarization extinction ratios higher
than 1,000:1. The system also uses narrowband interference filters from Andover. These
filters have full-width-at-half-maximum (FWHM) bandwidths of 1 nm. The narrowband
filters serve to reduce the background solar radiation. The peak transmission for 1064-nm
filter is ~38%, and for 532-nm filter is ~47%. For each receiver there is a field stop at the focal
plane of the lens, which defines the detector FOV to be 6 mrad, slightly larger than the laser
beam divergence angle.
Fig. 3. The receiver optics.
Highly sensitive and ultrafast PMT detectors from Hamamatsu are used as photodetectors.
Hamamatsu R632-01 is used to detect 1064-nm light, and R1464 is used for 532-nm, making
good use of their spectrum response.
One important parameter for a photodetector is the minimum detectable signal. Usually this
is specified by the noise equivalent power (NEP). This quantity refers to the radiant flux in
watts to produce a signal to noise ratio of unity, at the output end of the detector under dark
current only condition. This is also called the photon limited signal to noise ratio. In dark
current limited case, the NEP is defined as (Budde, 1983)
NEP = 2e ⋅ idark ⋅ G ⋅ ∆f / S , (1)
where S is the PMT current sensitivity at peak response wavelength, e is the electron’s
charge, is the dark current, G is the detector gain, and ∆f is the bandwidth of the
electronic system. At a bandwidth of 1 Hz, the NEP is watts for the R632-01
detector at 1064-nm, and watts for the R1464 detector at 532-nm.
From the backscattered laser signal, the cross-polarization ratio is used to quantitatively
describe the polarimetric scattering property of the target surface. It is defined as the ratio
between the cross-polarized return and the co-polarized return, i.e.,
I⊥
δ= , (2)
I //
where I // represents the co-polarized laser return and I ⊥ represents the cross-polarized
return received at the PMT detectors. The value of the cross-polarization ratio is usually
within the range of 0 to 1.0. However, for some targets, it was observed that this ratio can
exceed 1.0 (Kalshoven & Dabney, 1993).
3.1.3 Data Acquisition and Processing

A Berkeley Nucleonics model 555 digital delay/pulse generator is integrated in the system
as a timing source. It is used to generate the control TTL signals to trigger the laser and the
digitizer. The timing accuracy of the delay generator is 1 ns. For A/D conversion, two 8-bit
digitizing cards, DC256 and DC110, from Acqiris are selected. The DC256 has four input
channels, and a sampling rate of 500 MS/s, and is used to digitize the outputs from the PMT
detectors. The DC110 has one channel, also samples at 500 MS/s, and is used to digitize the
reference signal. Both cards are housed inside a modular CompactPCI crate, and connected
to the computer through a PCI interface card with 100 Mb/s transfer rate. The Acqiris
digitizers employ an internal time trigger interpolator with 5 ps resolution, which is used to
assist timing calibration and trigger positioning. The accurate timing ability of the digitizer
improves the range measurement accuracy, and helps obtain accurate lidar waveforms
which record the entire history when the laser propagates and interacts with the target.
The system software is programmed using Labview. Functions of the Labview program
include control of the digital delay generator, the laser, the digitizer, data recording and real
time waveform display. Data are recorded and post-processed.
The photographs of the MAPL lidar system are shown in Figure 4, where the photo at the
left side shows the laser head and the optical receiver assembly and the photo at the right
side shows the electronic devices mounted inside a standard 19-inch rack. This is the
configuration for ground experiments. The system needs to be repackaged in order to be
installed inside an aircraft for future airborne missions.
(b)
(a)
Fig. 4. Photographs of the MAPL system, where (a) is the optical assembly and (b) is the
electronic subsystem.
3.2 Multiwavelength polarimetric lidar calibration

To account for the laser power output instability, photodetector gain variation and other
potential error sources, the lidar data must be calibrated. Calibration is always a very
difficult task. This is especially true for a polarimetric lidar where the polarimetric
reflectance data calibration requires special consideration. In general, an ideal Lambertian
reflection surface with known reflectivity needs to be used as a calibration standard. But an
ideal Lambertian surface is extremely difficult to obtain even for laboratory measurement,
not to mention the field calibration.
Measurement of materials with potential to be used as calibration standards was done
inside an optical laboratory, in search of a better field calibration standard. Table 1 lists all
six materials used. The Stokes parameters of these materials at both 1064-nm and 532-nm
were measured to quantify their polarimetric reflectance properties. The measurement
results are shown in Figure 5. We can see that the tarp and the concrete (Figure 5 (a) and (d))
resemble an ideal diffusing scatterer at both wavelengths. Their backscattered light
intensities do not show strong dependence on the incident angle. Thus, it is concluded that
the backscatter properties of the tarp and the concrete are close to an ideal Lambertian
surface. Both of them could be used as a calibration standard in field experiments.
Canvas tarpaulin (Gosport Manufacture Inc.)

White paper (Great White® 86700)
Plywood (APA Champion 329)
Concrete (with surface polished)
Machined aluminum plate
Black anodized aluminum plate
Table 1. Materials used for Stokes parameter measurement.
In practice, the canvas tarp is used as a calibration standard due to its portability.
Laboratory measurement of the polarimetric reflectance at co-polarized and cross-polarized
directions from the canvas tarp is shown in Figure 6, where solid blue lines are for 1064-nm
and dashed red lines are for 532-nm. The data shown here are used in the calibration for the
MAPL system. The standard calibration procedure for ground experiments is described as
follows. Before measuring each target, the calibration standard (i.e., the canvas tarp) is set
up side-by-side with the target, and the lidar reflectance data are collected from the tarp.
The high voltage supply to each PMT detector is individually adjusted to make sure that the
signal has enough amplitude and is not saturated after A/D sampling. The data are
compared with the laboratory measurement data as shown in Figure 6. Finally, the ratio
between the averaged field calibration data and the laboratory results from the same tarp
was used as a calibration constant to correct for any measurement variations. Once the
calibration constant is obtained, the calibration procedure is done. Then the MAPL system is
directed to the target under study and multiple measurements are recorded as raw data.
These raw data from the target are then multiplied with the calibration constant to obtain
the calibrated data. All future data analysis should be based on the calibrated data.
80 80
70
(a) 70 (b)
60 60
Stokes parameters
Stokes parameters
50 50
40 40
30 30
20 20
10 10
0 0
-10 -10
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
Incident angle (Degree) Incident angle (Degree)

80 80
70
(c) 70
(d)
60 60
Stokes parameters
Stokes parameters
50 50
40 40
30 30
20 20
10 10
0 0
-10 -10
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50

300 300
250
(e) 250 (f)
Stokes parameters
Stokes parameters
200 200
150 150
100 100
50 50
0 0
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50

Fig. 5. Stokes parameters measurement of backscattered laser light from various materials.
(a). Canvas Tarp. (b). White paper. (c). Plywood. (d). Concrete. (e). Aluminum plate.
(f). Black anodized aluminum plate. (‘+’=I, ‘*’=Q, ‘◊ ’=U and ‘o’=V. Solid blue lines are for
1064-nm and dashed red lines are for 532-nm.).
Incident angle (degree)
Fig. 6. Laboratory measurement of the polarimetric reflectance of a canvas at 1064-nm (solid

blue lines) and 532-nm (dashed red lines) for calibration purpose.
4. Field test and data analysis

4.1 Description of the test site
The field test site was selected at Oakwood Lakes State Park in Eastern South Dakota. There
were plenty of trees and shrubs inside this park. The tree species observed at the test site
were primarily cottonwood, ash, oak and willow. A silver Dodge minivan was used in this
experiment as a man-made target. The vehicle was parked and hidden behind several trees
as shown in Figure 7. The MAPL system was set up across the lake at a distance of about
1000 m. The laser beams were shot horizontally across the lake toward the test scene.
Scan 1
Scan 2
Scan 3
Scan 4
Scan 5
Scan 6
Fig. 7. Photograph of the test site, where a vehicle was hidden inside the trees. The vehicle
was ~1000 m away from the MAPL sensor. The approximate positions of the laser scans are
marked by red arrows here.
The laser scanned over the test area from the left to the right. The scanning pattern is
demonstrated in Figure 7. A total number of six separate scans were made during the
experiment. Only one of the scans, scan number 4, coincided with the vehicle’s location.
4.2 Experimental results

From the lidar data, it is observed that the vehicle paint strongly reflects near-infrared laser
at both co-polarized and cross-polarized directions. The co-polarized return at 1064-nm is
shown in Figure 8, where a white arrow is pointing at the vehicle position. The reflectivity of
the vehicle paint at near-infrared is so strong that the returned signal is saturated after A/D
conversion as seen from scan number 4. From scan lines 1, 2, 3 and 5, it is seen that the
reflectance of the foliage at 1064-nm is much lower compared with the vehicle paint. The
relative return of the vegetation is typically below 55 after A/D conversion. It is also
observed that the top of the vegetation (scan number 1) is sparse so the returned signal
strength is weaker than that at the bottom of the trees (scan number 5) where the vegetation
becomes more dense. In general, it is observed that as the scan line gets lower, the returned
signal becomes stronger due to more dense foliage. However, at scan number 6, the laser
beam finally reaches the lake water and the returned signal becomes almost invisible. Two
combined effects cause the backscattered signal level to be very low in this case. The first
one is that water strongly absorbs the near-infrared laser, in comparison with foliage and
vehicle paint. In addition to that, the water surface also acts as a specular surface and
reflects the laser energy in the forward direction away from the detector due to the very
large incident angle. As a result, there is little signal observed from scan number 6 as shown
in the bottom of Figure 8. For better clarity, Figure 9 shows the image from scan number 4
separately. An arrow is used to highlight the vehicle inside the trees. All the weak scatterers
in this image represent the vegetation; while the strongest reflector represents the hidden
vehicle.
The direct return data from the co-polarized and cross-polarized green channels at 532-nm
reveal no observable difference between the foliage and the vehicle. This is demonstrated in
Figure 10 at the co-polarization channel. The result shows that the vehicle paint has a very
similar reflectivity at 532-nm compared with the vegetation. Therefore, using the direct
return information from a single wavelength at 532-nm and a single polarization, it is
impossible to detect the foliage-obscured vehicle. Conversely, the polarization information

can provide enhanced capability in detection and identification. After computing the cross-
polarization ratio following Eq. (2), the difference between the vehicle and the vegetation is
revealed. The cross-polarization of the vehicle surface is significantly lower than that of the
vegetation. This is partially explained by the fact that the surface of the vehicle paint is much
smoother compared with the surface of vegetation leaves and barks at optical wavelength.
As a result, the vehicle surface can better maintain the original linear polarization state of
the incident laser thus leads to low cross-polarization ratio. Moreover, the random
orientation and distribution of the tree leaves, i.e., the vegetation canopy structure, also
contributes to higher depolarization during the multiple scattering process. As observed in
Figure 11, the cross-polarization ratio of the vehicle (as indicated by an arrow) is typical
below 0.2. On the other hand, the cross-polarization ratio of the vegetation is consistently
above 0.4. At certain vegetated areas, this ratio can reach a value of larger than 0.8. This may
be related to trees with very rough leaf surface and trees with very dense canopy structure.
Figure 12 and Figure 13 show the scatter plot of the returned signal at 1064-nm and 532-nm,
respectively. It is clear that the classification boundary between the vegetation and the
vehicle can be easily obtained in this case. From Figure 12 it is seen that the reflectivity of the
vehicle paint at 1064-nm is consistently higher at both co-polarized and cross-polarized
directions compared with the vegetation. The vehicle signal is so strong that it is almost
always saturated. From Figure 13 it is observed that although the vehicle paint at 532-nm
has comparable reflectivity in comparison with the vegetation, the cross-polarization ratio of
the vehicle is much lower than that of the vegetation. The cross-polarization ratio at 1064-
nm was not computed since the return data from the vehicle are saturated at both the co-
polarized and cross-polarized channels. However, a logarithmic amplifier may be used in
the lidar system to reduce the dynamic range of the signal. In that case, it is possible to
obtain the cross-polarization ratio at 1064-nm. With the additional polarization information
at 1064-nm, better detection and identification should be available for more targets.
Fig. 8. Scanning result of the relative lidar return from the co-polarization channel at 1064-
nm. The arrow is pointing to the vehicle inside the trees.
Fig. 9. Scan number 4 of the relative lidar return from 1064-nm at the co-polarization
channel. The arrow is pointing to the vehicle (the strong scatterer) inside the trees (the weak
scatterers).
Fig. 10. Scan number 4 of the relative lidar return from 532-nm at the co-polarization
channel. The vehicle could not be identified in this image.
Fig. 11. The cross-polarization ratio at the 532-nm from scan number 4. The arrow is
pointing to the vehicle with low cross-polarization ratio inside the trees.
Fig. 12. Scatterplot of the lidar return at 1064-nm, where the x-axis is the co-polarization
channel and the y-axis is the cross-polarization channel.
Fig. 13. Scatterplot of the lidar return at 532-nm, where the x-axis is from the co-polarization
channel and the y-axis is from the cross-polarization channel.
5. Conclusions and future work

The experimental study demonstrates that the multiwavelength polarimetric lidar provides
an enhanced capability in the detection of a foliage-obscured vehicle. It is revealed that the
reflectivity of the vehicle and the reflectivity of the foliage are very close at 532-nm.
However, at 1064-nm the reflectivity of the vehicle is significantly higher than that of the
foliage. Using multiwavelength lidar in this situation demonstrates improvement in target
detection. It is also shown that the cross-polarization ratio of the tree canopies is much
higher than that of the vehicle at 532-nm. If a logarithmic amplifier was employed in the
lidar system, the cross-polarization ratio at 1064-nm could be obtained, which can further
improve the detection and identification accuracy. It is evident from the experiment that the
polarimetric information provides enhanced target detection and identification capability.
The multiwavelength polarimetric lidar certainly has its advantage over the existing
commercial lidar systems, which are single wavelength and non-polarimetric, in target
detection and identification.
In this study, only ground test data were collected. The airborne lidar data are expensive
and were not collected at the time. However, it is known that more openings of the forest
canopies are observed looking from above the ground. This means that the laser beam has a
better chance to penetrate the tree canopies from airborne sensors. As a result, there is a
better chance to detect the foliage obscured target from an airborne lidar sensor than from a
ground based sensor.
This study also demonstrates the validity of the technique to detect and identify other
foliage-obscured man-made targets using a multiwavelength polarimetric lidar. As long as
we can find the difference in the spectral reflectance or polarimetric reflectance between the
man-made target and the vegetation, we can always employ this technique. The selection of
appropriate laser wavelengths and polarization combinations should be able to maximize
the differences between the vegetation and the man-made target, hence enable the detection
of many other types of man-made targets.
Further research is needed in order to use this multiwavelength polarimetric lidar detection
technique in practical applications. From the lidar hardware side, many improvements
could be made. For example, a tunable laser could be used in the system to provide a wide
range of laser wavelengths for better detection. Furthermore, we can introduce a logarithm
amplifier or a 12-bit (or higher) A/D converter instead of an 8-bit one in the lidar system to
make sure that the system can handle signals at a larger dynamic range without signal
saturation. Polarimetric lidar calibration is another challenge and better calibration strategy
still needs to be explored. A polarimetric calibration standard that has a constant
polarimetric reflectance property over repeated use and is easy to deploy in the field is
desired. Even more importantly, efforts need to be made for the multiwavelength
polarimetric lidar operation and data processing to be simple enough so that even a person
with limited engineering background can operate such a sensor. From the software side, an
algorithm that can automatically detect and identify the targets needs be developed. Both
artificial neural networks (NN) and support vector machine (SVM) seem to be good
candidates for automated data processing. The classification accuracy will be improved if
we understand the target polarimetric scattering characteristics. Such understanding can
greatly facilitate the development of classification algorithm. At this time, however, there is
only limited research available. Consequently, more studies on the polarimetric scattering
properties of both man-made targets and vegetation are in urgent need (Tan, Narayanan &
Helder, 2005). In addition, more ground experiments are also needed. For example, man-
made targets under camouflage conditions need to be tested. Camouflage presents a
challenge in target detection. However, multiple-wavelength and polarization should
provide advantage over existing lidar technology for camouflaged target detection.
Camouflage is usually designed to match the spectral reflectance of the background. It is
usually not intended to match the polarimetric reflectance of the background. Even if
anybody wanted to do so, it would be extremely difficult to control the polarimetric
reflectance of the camouflage. Therefore, a polarimetric lidar have a better chance to detect
and identify the camouflaged target.
In summary, the multiwavelength polarimetric lidar is a new concept lidar in foliage
obscured man-made target detection. Even though more studies are needed in this new
technology, the experimental study clearly serves the purpose to demonstrate that the
multiwavelength polarimetric lidar can provide better target detection and identification
over the traditional single wavelength and non-polarimetric lidar system.
6. References
Budde, W. (1983). Physical detectors of optical radiation, In : Optical Radiation Measurements,
Grum, F. & Bartleson, C. (Ed.), Academic, New York.
Castano, A. & Matthies, L. (2003). Folidage discrimination using a rotating ladar, IEEE
Internation Conference on Robotics and Automation, Vol. 1, pp. 1-6.
Jaaskelainen, T.; Silvennoinen, R.; Hiltunen, J.; & Parkkinen, P. (1994). Classification of the
reflectance spectra of pine, spruce and birch, Applied Optics, Vol. 33, No. 12, pp.
2356-2362.
Kalshoven, J. & Dabley, P. (1993). Remote sensing of the Earth’s surface with an airborne
polarimetric laser, IEEE Trans. Geoscience and Remote Sensing, Vol. 31, pp. 438-446.
Ma, Q.; Ishimazu, A.; Phu P.; & Kuga, Y, (1990). Transmission, reflection, and depolarization
of an optical wave for a single leaf, IEEE Trans. Geoscience and Remote Sensing, Vol.
28, pp. 865-872.
Measures, R. (1992), Laser Remote Sensing: Fundamentals and Applications, Krieder Publishing,
Malabar, FL.
Nashashibi, A. & Ulaby, F. (2005). Detection of stationary foliage-obscured targets by
polarimetric millimeter-wave radar, IEEE Trans. Geoscience and Remote Sensing, Vol.
43, pp. 13-23.
Tan, S.; Narayanan, R. & Kalshoven J. (2001). Measurement of Stokes parameters of
materials at 1064-nm and 532-nm wavelengths, Proc. SPIE, laser radar technology and
applications VI, Vol. 4377, pp. 263-271.
Tan, S. & Narayanan, R. (2004). Design and performance of a multiwavelength airborne
polarimertric lidar for vegetation remote sensing, Applied Optics, Vol. 43, pp. 2360-
2368.
Tan, S.; Narayanan, R. & Helder, D. (2005). Polarimetric reflectance and depolarization ratio
from several tree species using a multiwavelength polarimetric lidar, Proc. SPIE,
Polarization Science and Remote Sensing II, Vol. 5888, 5888-23.
Xu, X. & Narayanan, R. (2003). FOPEN SAR imaging using UWB step-frequency and
random noise waveforms, IEEE Trans. Aerospace and Electronic Systems, Vol. 37, pp. 1287-
1300.
Unmixing Based Landsat ETM+ and ASTER
Image Fusion for Hybrid Multispectral Image Analysis 407
19
X

Image Fusion for Hybrid Multispectral
Image Analysis
Nouha Mezned1, Sâadi Abdeljaoued1
and Mohamed Rached Boussema2
1Faculté des Sciences de Tunis,
2 Ecole Nationale d'Ingénieurs de Tunis
TUNISIA
1. Introduction
Monitoring the environmental risks associated with mine tailings in a quick and timely
fashion is the first step towards mitigating their impact. Mine tailings may have a
widespread geographical distribution; their location and extent may also vary along time,
due to reprocessing and disposal activities. For these reasons, the characterization of mine
tailings using traditional field work alone is both costly and inefficient. Remote sensing
techniques have been proven extremely valuable in the inventory, characterization, and
remediation of mine tailings elsewhere (Peters & Hauff, 2000). Determining the location,
extent, and geochemistry of mine tailings is the first step towards remediation and hence
avoidance of negative health and environmental consequences.
In this chapter we focus on mine tailing cartography and monitoring using multispectral
data (Landsat ETM+ and ASTER). These data are well available from much passed years at
regular intervals. We are particularly interest in the case study of zinc and lead mines in the
north of Tunisia (North Africa). Particularly, the mine of Jebel (Hill) Hallouf-Bouaouane
(36°42'N 9°0'5''E), which is forsaken since 1986, is among several types of mines in the
Mejerda river watershed (Mansouri, 1980). The Mejerda River is the most important river
which is exploited for the agriculture irrigation and drinking water alimentation of the north
of Tunisia. In this site, mine tailings cause the environment degradation, so it had polluted
soils, vegetation and water quality (Souissi, 2007). Mine tailing cartography becomes
fundamental in order to follow the environmental changes and pollution quickly.
The approach was organized in two steps; a coarse cartography based on the Landsat
(Enhanced Thematic Mapper Plus) ETM+ spectral unmixing using image derived
endmember and a detailed cartography based on a multispectral inter-images fusion using a
simplified version of Multisensor Multiresolution Technique (MMT). The analysis of the
resulting hybrid image allowed to tailing cartography refinement. Detailed fraction maps
were generated.
(a) (b)
(c)
Fig. 1. The three endmember spectra selected from first three PC analyses.
The chapter is structured as follows. Section 2 is a tailing cartography approach description.

In section 3, we consider a coarse and refinement cartography of the mine tailings using
hybrid Multispectral data and demonstrate that its important special case. In section 4, we
present results and discussion. Section 5 concludes the chapter.
2. Mine tailing cartography

The adopted cartography approach includes two main parts: The first one deal with the
classification of the ETM+ data which results give the distribution and the abundance image
of mine tailing surface cover. The resulting map represents a coarse mine tailing
cartography. The second part, focus on the ETM+ and ASTER data fusion and analysis
(Mezned et al., 2007). We used The (MS) and (Pan) ETM+ data acquired on May 3, 2000 and
Aster Level 2B SWIR reflectance surface product acquired on June 26, 2000. The resulting
hybrid image have the high spatial resolution of the (MS)/(Pan) Landsat ETM+ image (15m)
and its relative high spectral resolution combined with ASTER SWIR image (10 bands).
2.1 Multispectral image classification

The first classification was performed on Landsat ETM+ MS/Pan fused image (196x196
pixels) with 6 bands using constrained spectral unmixing method. We used the Principal
Component (PC) Spectral Sharpening technique to merge (MS) and (PAN) Landsat ETM+
data. Spectral information’s are conserved as these two images were acquired by the same
sensor. The used images were georeferenced, co-registered, topographically and
atmospherically corrected using the FLAASH (Cooley, 2002) software before performing the
sharpening. Indeed, (Pan) data was only topographically corrected.
The first step of the analysis identifies the pure materials that are present in the scene.
Principal Component (PC) analysis method was used for optimal endmember extraction
(Boardman, 1993); (Carreiras and al., 2002). It allows one to reduce the dimensionality of the
data and to know the number of endmembers present in the scene as explained in (Keshava
& Mustard, 2002). Endmembers lay correctly at the corners of the triangular shaped
diagram, indicating the purest pixel spectra (Boardman, 1993). Figure 1 show the three
endmember, which were selected using the first three principal components (contain more
than 95 % of the information), corresponding to tailings, vegetation and soil. Their
identification was based on their typical spectra and their location.
(a) (b)
(c)
Fig. 2. Fraction maps (196x196 pixels) resulting from Landsat ETM+ coarse classification: (a)
tailings, (b) vegetation and (c) soils.
The fraction maps which estimated by the constrained least-squares algorithm (Singer,
1981); (Shimabukuro & Smith, 1991) for the three pure materials are represented in Figure 2.
Note that a black (respectively, white) pixel in the map indicates a large (respectively, small)
value of the abundance. The accuracy of the endmember fraction maps are evaluated using
the Root Mean Square (RMS) error. The most of RMS error (96 %) are lower than 1.35 10-5,
which shows a good result.
2.2 Unmixing based multispectral Image Fusion

The proposed unmixing based image fusion method consists on a two data set fusion: the
first set has a high spatial resolution (and low spectral resolution), called classifying
instrument (CI) and the second one have a high spectral resolution (and low spatial
resolution), called measuring instrument (MI). The unmixing method uses the spatial high
resolution data classification (CI) to unmix the lower spatial resolution image (MI).
Furthermore, the unmixing can be performed only relative to the recognized classes in the
(MI) image.
The main contribution consists on considering a fusion of two different multispectral
sensors. Thus, we chose the (MS) Landsat ETM+ data combined with the corresponding
(PAN) band (Chavez et al., 1991) as the (CI) image (Figure 3 a) and the ASTER SWIR
product as the (MI) image (Figure 3 b). These images were co-registered before fusion. The
used fusion technique is based on:
- classification of the high spatial resolution image (CI). In our case, the classification
of the CI Landsat ETM+ MS/Pan fused image (only 4 bands were used, the last
two bands were removed due to its overlapping with the ASTER SWIR bands
which represent better spectral resolution) was performed by the ISODATA
unsupervised technique (Richards, 1986) with K0 = 16 classes (gives minimal error
during the constrained unmixing).
- definition of class contributions to the signal of the low spatial resolution of MI
pixels. It is based on the resulting high-resolution classification map k (m, n) which
represents different class areas. The contribution of class is given by the following
equation:
ci (l , s; k0 ) = ∑ ρ (l , s; m, n)
k ( m , n )∈k 0
i
where ci (l, s; k0) is the contribution of class k0 to the signal of low spatial resolution
of MI pixel (l, s) in different band i, and ρi (l, s; m, n) is a discrete approximation for
the sensor PSF (Point spread function) which sum over a high resolution pixel (m,
n) is assumed to be normalized to 1. The discrete PSF includes the registration
component (MI image is co-registered with CI image) and the atmospheric
component (CI image was corrected to eliminate the atmospheric and illumination
effects before the classification and thus, these effects will be also removed during
the unmixing).
- we proposed a simplified version of the constrained unmixing algorithm (Hu et al.,
1999). In this algorithm, we make the hypothesis that the reflectance of MI pixels
equal to the sum of the mean reflectance of each class in the window.
(a) (b)
(c)
Fig. 3. Landsat ETM+ MS/Pan (196x196) with 15 m pixel size acquired in May 2000 shown
in gray scale at wavelength λ=0.66 µm (a), ASTER SWIR (196x196) with 30 m pixel size
acquired in June 2000 shown in gray scale at wavelength λ=2.4 µm (b) and Hybrid fused
image (196x196) with 15 m pixel size shown in gray scale at wavelength λ=0.66 µm (c).
The unmixing of the MI pixels was performed in a 5 x 5 window that was moved
with the step of 1 MI-pixel size. The central MI pixel in each window is unmixed by
an inversion of a system of linear mixture equations that are written as following
for all pixels in the window:
K _
Ri (l , s ) = ∑ ci (l , s, k ) Ri (k ) + ε i (l , s )
k =1
_
where Ri (l, s) is the reflectance of MI pixel (l, s) in the window, R i (k) is the mean
MI-reflectance for class k in the window and εi (l, s) is the model error. The
numerical inversion of the linear system given in (Zhukov et al., 1999); (Zurita
Milla et al., 2008) was done with the matlab software using the least-square
function independently for each MI band (Matlab, 2004) This function returns the
_ _ _
vector R that minimizes norm (C* R - R ) subject to R ≥ 0 .
- restoration of the unmixed lower resolution channels image was performed by
assigning the estimated mean class reflectance to the corresponding high resolution
pixels of the classification map.
The resulting hybrid image (Figure 3 c) present the high spatial resolution of the (MS)/(Pan)
Landsat ETM+ image (15 m) and its relative high spectral resolution (4 bands) combined
with ASTER SWIR bands (6 bands).
3. Tailing cartography refinement

The tailing cartography refinement approach includes the generation of tailing mask needed
for unmixing the hybrid image. The mask was generated from the first tailing coarse
cartography. It was validated before being used in the hybrid image classification process.
The classification was performed using the constrained spectral unmixing method. The
resulting maps represent detailed mine tailing cartography. The laboratory analyses were
used as field truth for tailing maps validation.
3.1. Generation and validation of the tailing mask

The tailings fraction map represents their abundance and their spatial distribution in the
study area. The mean value for tailing proportions is 0.32. This value supposes that the
study area is covered by a considerable quantity of tailings. Those for vegetation and soil are
estimated by 0.26 and 0.41 respectively. A tailing mask (Figure 4. a) was generated
(extracted from fractional abundance map) to highlight the mine tailing area and masking
the surrounding zones (vegetation and soils). The spatial distribution of the mine tailing was
validated using high resolution data. The validation of tailing mask was based on aerial
photo. It consists on the comparison of the tailing distribution (indicated by red lines)
detected on the mask as well as aerial photo (Figure 4. c).
Indeed, certain zones are inaccessible or difficult of access. A spatial interpretation will be
easier and faster. So, we refer to the interpretation of the aerial photo to analyze the
distribution of the tailings in the Jebel Hallouf-Bouaouane region. The comparison reveals
the presence of an individual quantity of tailings transported eastward and deposited near
the Kassab wad. It is detected on the mask as well as on the aerial photo.
2
II
I
III 1
(a)
(b)
(c)
Fig. 4. Figure (a) indicates the generated and filtered tailing mask (196x196 pixels). Figure (b)
shows the aerial photo (scale: 1/6250) scanned at 100 dpi of the corresponding Jebel Hallouf-
Bouaouane region. Blue roman numbers locate tailing deposits: (I) the first dyke of Jebel
Hallouf, (II) the second dyke of Jebel Hallouf and (III) the dykes of Bouaouane. Orange
numbers locate the eroded and transported tailings towards: (1) kassab wad and (2) swamps
in the north-east. Figure (c) shows the validation of the region of interest with 171 km2
(indicated with red rectangle). The contours lines (red lines) which correspond to tailing
zones were superposed on the georeferenced aerial photo.
3.2. Hybrid Multispectral data analysis

The classification of hybrid multispectral data, which is based on the constrained linear
spectral unmixing using JPL mineral library spectra, generates mineral detailed maps. Only
mine tailing pixels are used in the analysis. We used the tailing mask generated from the
coarse classification to exclude the undesired zones and to analyze minerals in the tailing
area.
The used endmembers are: calcite (with additionally cerusite, quartz and barite), gypsum,
oxides (hematite and goethite), clays (kaolinite and illite) and pyrite. This mineralogical
composition was revealed from samples by
(a) (b)
(c) (d)
(e)
Fig. 5. Detailed classification: Endmember fraction maps (196*196 pixels) derived from
Hybrid constrained linear unmixing: (a) calcite, (b) clays (kaolinite and illite), (c) gypsum,
(d) oxides (hematite and goethite) and (e) pyrite.
calcimetry, X-ray diffraction and polished section analysis. Tailing samples were collected
from deposit location areas (considered as pollution source). Mineral library spectra (Grove,
1992) were resampled to hybrid band passes before classification. 96 % of the classified
pixels show an error lower than 1.9 10-5. This classification is then acceptable. The resulting
endmember fraction maps show an interesting contribution of the hybrid image for mineral
detection.
The results of linear constrained spectral unmixing show different fraction maps (Figure 5),
which show the comparison of the variability’s of the endmember distribution in a spatial
context. Among the 5 endmembers, only calcite and gypsum are of most interest. Indeed,
this is not a surprising result since the magazine rocks are carbonates. The unmixing images
suggest a relative significant concentration of Hematite mineral within the tailing deposit
area (resulting from pyrite oxidizing exposure surfaces) and in the soils (polluted soils by
alluvium) (due to the lithologic nature of soils and the eroded tailings). This result can
explain the reddish color of deposit area and soils. The clays fraction map shows also a less
significant concentration within deposit than in the soils (polluted soils by alluvium). For
pyrite concentration, is more detected within deposit, particularly in Bouaouane dykes, than
soils. This would suggest that this mineral is almost completely oxidized in the deposit
surface which is exposed to the climatic effects. It is presence in the soils (polluted soils by
alluvium), can be explicated by the wind and hydrous erosion effects which transport and
depose tailings far from deposit.
4. Results analysis and discussion

In this section, we compare the abundance of minerals considered by both multi-scale
methods. Indeed, the results of hybrid classification are then compared to the abundances of
minerals (endmembers) calculated from laboratory analysis. This comparison allowed
validating the results of the adopted tailing cartography approach. The statistical regression
analysis consists on calculation of the coefficient of determination (R2). The analysis was
conducted for the whole samples.
The comparison was revealed the existence of positive linear regression with variable
intensities. According to the results obtained and exposed in the figure 6, we note, that the
coefficient of determination is 0.85 (figure 6 b) for clay minerals. The slope of the regression
equation is near unity (0.98), and the intercept is 1.14. This indicates that the hybrid image
data can well quantify the abundance of clays. For pyrite, the coefficient of determination is
also important (0.62) (figure 6 e). The slope of the regression equation is 0.77 and the
intercept is 0.7. The pyrite can be quantified by hybrid image. For gypsum, the coefficient of
regression is also important (0.6) (figure 6 c). However, the slope of the regression equation
is weak (0.19) and the intercept close to 0 (0.06). The abundances estimated by laboratory
analysis are not equal to those quantified by hybrid image, and thus the regression slope
does not approach a value of 1. For oxyhydroxides, the coefficient of determination is also
above 0.5 (0.506) (figure 6 d). The slope of the regression equation is 0.67 and the intercept
close to 0 (0.07). This indicates that hybrid image can quantify relatively hematite and
goethite. For calcite, the coefficient of determination is 0.504 (figure 6 a). The slope of the
regression equation is 0.73 and the intercept is 52. This raised value of the intercept is
probably due to underestimation of the abundance from hybrid image. Furthermore, the
(a) (b)
(c) (d)
(e)
Fig. 6. Comparison between the abundance of minerals estimated from hybrid unmixing
results and the laboratory analysis results.
values of R2 showed that the regression for particularly clays and pyrite had the higher
goodness of fit than those for calcite, gypsum and oxyhydroxides .
The above results of regression analysis imply that positive attitudes are related to concrete
complementarities of hybrid unmixing and laboratory analysis results. The experiment was
conducted only for the relatively few used samples. Thus, we should note that there is a
limit for a rigorous interpretation of the regression results.
4. Conclusion
In this chapter we have described a tailing cartography approach which is very useful in the
case of multispectral multisensory study. A simplified version of the Multisensor
Multiresolution Technique (MMT) for image fusion was also presented. Moreover, we have
tested this technique on two different multispectral sensors; Landsat MS/Pan image and
ASTER SWIR image. The unmixing results’ has shown that the hybrid multispectral data
can be used successfully to provide information on refining mine tailing cartography.
Furthermore, multispectral hybrid image shows a complementary result with laboratory
analysis for selected endmember minerals.
The classification of hybrid multispectral data which is based on the constrained linear
spectral unmixing using JPL mineral library spectra generated mineral detailed maps. Only
mine tailing pixels were used in the analysis. In fact, the first coarse classification was used
to highlight the mine tailing area and masking the surrounding zone in the hybrid image.
The mask was validated with high resolution image before being used in the classification
process.
The resulting mineral fraction maps show the comparison of the variability of their
distribution in a spatial context for classification method. These results show an interesting
contribution of the hybrid image for mineral detection. Furthermore, multispectral hybrid
image classification shows complementary results with laboratory analysis, used as field
truth, for selected endmember minerals.
Furthermore, future refinement of the detailed tailing maps will result in presence of high
spectral and spatial resolution data. Moreover, we propose a strict sampling in the mine
tailing deposit for more significant mineralogy quantification.
5. Acknowledgments
The authors would like to thank the NASA Land Processes Distributed Active Archive
Center and the User Services USGS Earth Resources Observation and Science (EROS) for
providing numerous remotely sensed data.
6. References
Boardman, J. W. (1993). Automatic spectral unmixing of AVIRIS data using convex
geometry concepts, Summaries 4th Annual JPL Airborne Geoscience Workshop, October
25-29, Vol. 1 AVIRIS Workshop.
Carreiras, J. M. B.; Shimabukuro, Y. E. & Pereira, J. M. C. (2002). Fraction images derived
from SPOT-4 VEGETATION data to asses land-cover change over the State of Mato
Grosso, Brazil, Int. J. Remote Sens., vol 23, n°. 23, pp. 4979-4983.
Chavez, P. S.; Sides, S. C. & Anderson, J. A. (1991). Comparaison of three different methods
to merge multiresolution and multispectral data: Landsat TM and SPOT
panchromatic, Photogramm. Eng. Remote Sensing, vol. 57, N° 3.
Cooley, T.G.; Anderson, P.; Felde, G.W.; Hoke, M.L.; Ratkowski, A.J.; Chetwynd, J.H.;
Gardner, J.A.; Adler-Golden, S.M.; Matthew, M.W.; Berk, A.; Bernstein, L.S.;
Acharya, P.K. & Miller, D. (2002). FLAASH, a MODTRAN4-based atmospheric
correction algorithm, its application and validation, International Geosciences and
Remote Sensing Symposium IGARSS'02, vol.3, 2002, pp. 1414 -1418.
Grove, C. I., Hook, S. J., and Paylor II, E. D. (1992). Laboratory Reflectance Spectra of 160
Minerals, 0.4 to 2.5 Micrometers, Jet Propulsion Laboratory Publication 92-2.
Haertel, V. F. & Shimabukuro Y. E. (2005). Spectral linear mixing model in low spatial
resolution image data, IEEE Trans. Geosci. Remote Sensing, vol. 43, N° 11.
Hu, Y. H.; Lee, H. B.; & Scarpace, F. L. (1999). Optimal linear spectral unmixing, IEEE Trans.
Geosci. Remote Sens., vol. 37, N°. 1, 1999, pp. 639-644.
Keshava, N. & Mustard, J. F. (2002). Spectral unmixing, IEEE Signal Process. Mag., vol. 19, N°.
1, 2002, pp. 44-57.
Mansouri, A. (1980). Gisements de Pb-Zn et Karstification en milieu continental: le district
du Jbel Hallouf- Sidi Bou Aouane (Tunisie Septentrionale), Thèse de doctorat, Univ.
de Pierre et Marie Curie, Paris VI, 199 p.
Matlab (2004). MATLAB Users Guide, MATLAB 7.0 Release 14.
Mezned, N.; Abdeljaouad, S. & Boussema, M.R. (2007). Unmixing Based Landsat ETM+ and
ASTER Image Fusion for Hybrid Multispectral Image Analysis, IEEE International
on Geoscience and Remote Sensing Symposium IGARSS’07.
Peters, D. C., and Hauff, P. L. (2000). Multispectral remote sensing to characterize mine
waste, Springer-Verlag, Berlin, pp. 113-164.
Richards, J. A. (1986). Remote sensing digital image analyzes. An introduction, Berlin,
Germany. Springer-Verlag.
Shimabukuro, Y. E., and Smith, J. A. (1991). The least-squares mixing models to generate
fraction images derived from remote sensing multispectral data, IEEE Trans. Geosci.
Remote Sens., vol. 29, n°. 1, pp. 16-20.
Singer, (1981). Near-infrared spectral reflectance of mineral mixtures: systematic
combinations of pyroxenes, olivine, and iron oxides. Journal of geophysical Research,
86, pp. 7967-7982.
Souissi, F. (2007). Minéralogie et géochimie des gîtes minéraux, des séries sédimentaires
anciennes et de l'environnement (Synthèse des travaux de recherche). Habilitation
universitaire, Faculté des Sciences de Tunis FST, 83p.
Zhukov, B.; Oertel, D.; Lanzi, F. & Reinhäckel G. (1999). Unmixing-Based Multisensor
Multiresolution Image Fusion, IEEE Trans. Geosci. Remote Sensing, vol. 37, N° 3.
Zurita Milla, R.; Clevers, J.G.P.W. and Schaepman. M.E. (2008). Unmixing-based Landsat
TM and MERIS FR data fusion, IEEE Geoscience and Remote Sensing Letters 5, pp.
453-457.
InSAR phase analysis: Phase unwrapping for noisy SAR interferograms 419
0
20
InSAR phase analysis: Phase unwrapping for

noisy SAR interferograms
Riadh ABDELFATTAH1,2
1 École Supérieure des Communications de Tunis (URISA/Sup’Com)
Tunisia
2 ITI/Telecom Bretagne
France
1. Introduction
Interferometric Synthetic Aperture Radar (InSAR) exploits the phase difference between two
complex radar signals for extracting information about the earth surface. Some significant ap-
plication fields concerned by InSAR techniques are Digital Elevation Model (DEM) generation
(Graham, 1974), geophysical hazard analysis (Massonet & Feigl, 1998), desertification (Bodart
et al., 2009), deforestation (Balzter, 2001), glacier velocity measurements (Mattar et al., 1998)
and land use classification (Wegmullerl & Werner, 1997). They exploit the interferometric
phase, which is proportional to the difference of the path lengths between the radar positions
and a target on the earth surface (Gabriel & Goldstein, 1988), and the interferometric corre-
lation (Abdelfattah & Nicolas, 2006), which depends on target characteristics (Abdelfattah &
Nicolas, 2002).
The phase of complex InSAR pair, called the interferogram, may be used in the reconstructing
of high resolution terrain digital elevation models (DEM), in measuring earth displacement
and many other topographic applications. However, before the reconstruction applications,
the interferometric phase have to be unwrapped. the phase that is measured (observed) is
only modulo 2π, the so called principal phase value or wrapped phase. Moreover, it is well
known that the speckle effects in the InSAR data generate noise in the InSAR phase, which is
also noisy due to various decorrelation effects (thermal noise, temporal, geometrical, etc.) be-
tween the two images used in the interferogram generation. The phase unwrapping consists
in estimating the true regular phase profile (multiples of 2π must be added or substracted)
from the observed interferogram, which is noisy and wrapped, to make the phase image as
smooth as possible. Formally, we have for each pixel (i, j):
φij = ϕij + 2kπ (1)
where φij is the true phase value (unwrapped phase), ϕij is the observed phase (wrapped),
and k ∈ Z is an integer accounting for the number of 2π multiplies (Ghiglia & Pritt, 1998).
Generally, the phase noise have to be reduced from the wrapped phase before the computa-
tion of the exact multiplies number. This will produce a regular fringe patterns and reducing
the number of irregularities (phase discontinuities). Thus, the phase unwrapping step would
be easier after the filtering step. However, the filtering process may cause a loss of resolution,
which will affect the accuracy of the extracted information. Moreover, areas with high coher-
ence don’t need really to be filtered. So, the interferogram global filtering should be avoided,
and an adaptive filtering approach should be considered.
The phase unwrapping process have to respect the trade-off between the spatial resolution
and the accuracy of the interferometric phase. Two groups of phase unwrapping algorithms
have been proposed in the literature: the first group uses local approach (Goldstein et al.,
1988) and the second one uses global approach (Ghiglia & Romero, 1994).
• The local approaches perform the summation of the phase differences along a selected
path from pixel to pixel across the interferogram. They are path dependent and requires
the phase residues connection using a phase residue map which could be directly ob-
tained from the interferogram. A phase residue represents a potential origin of phase
unwrapping errors. These approaches perform well in areas with sufficient high coher-
ence. However, they fail when the interferogram is noisy due to the residues created by
the noise.
• The global approaches are path independent and referred to as the least squares method
(Ghiglia & Romero, 1994). The unwrapped phase is then obtained as the minimizer,
in the mean-square sense, of a dependent (on unwrapped phase and phase difference
fields) functional. These approaches are powerful in the local reconstruction of the un-
wrapped phase, but propagates the unwrapping errors from decorellated parts of the
image into high coherent areas.
In the literature, it is shown that these methods are optimized when the extraction of the
true phase values is performed on a filtered interferogram. Unfortunately, noise reduction
is achieved at the expense of a loss in spatial resolution. Moreover, these approaches are
incapable to recover inconsistencies due to the InSAR acquisition system: they cannot extract
the phase profile across regions of invalid information (such as shadow or layover regions).
In this chapter, we propose a modified filtering algorithm to the López and Fabregas (López-
Martı́nez & Fabregas, 2002) noise reduction algorithm for the interferometric phase noise in
SAR interferometry using a multiresolution approach. Our contribution to the existing algo-
rithm consists on the exploitation of the InSAR coherence map in order to generate a more
adaptive mask for each decomposition level. Moreover, an improved Markov Random Field
(MRF) path independent phase unwrapping method is proposed. It is based on the Rodriguez
and Servin (Rodriguez-Vera & Servin, 1995) framework which is well adapted for phase un-
wrapping of locally inconsistent and noisy principal value phase fields that may contain re-
gions of invalid information. This framework is based on Bayesian estimation theory with
the use of Markov random field models to construct the prior distribution, so that the solu-
tion to the unwrapping problem is characterized as the minimizer of a piecewise-quadratic
functional. Our contribution is then the definition of an adaptive (and more realistic) energy
functional for the optimization of the random field.
In the next section, we will introduce the InSAR phase characteristics: it includes the statistics,
the decorellation effects and the noise model. In the third section, InSAR phase noise reduction
using the multiresolution approach is presented and validated through simulated and real
interferograms. The InSAR phase unwrapping based on MRF model is detailed in the forth
section.
2. InSAR phase characteristics

The InSAR phase is affected by the presence of several decorrelation effects, which collectively
result in an interferogram phase noise.
2.1 InSAR phase statistics

It is known that, under the traditional assumptions of fully developed speckle, single look
complex SAR images follow the ”circular Gaussian model”, i.e., the phase follows a uniform
law between 0 and 2π, and the real and the imaginary intensities follow Gamma laws with the
same variance. An interferometric pair is composed of two single look complex SAR images:
each pixel is represented by a vector Z of dimension 2 whose components z1 and z2 are the
values of pixel on each image of the interferometric pair, where each component follows the
circular Gaussian model. Then, the circular Gaussian model with 2 dimensions applies to this
vector Z, and its probability density function (pdf) is given by (Maı̂tre, 2001):
1
pz (Z) = exp −Z∗T Cz−1 Z (2)
π2 det (Cz )
where Cz is the covariance matrix of Z, ∗ the complex conjugate operator and T signifies
hermetian transpose. With the assumption that in each of the correlated signals, z1 and z2 ,
the quadrature components of each sample are zero-mean Gaussian processes and indepen-
dently distributed, the complex Wishart distribution for the covariance matrix estimate can be
derived (Wishart, 1928). The covariance matrix is given by:

σ12 σ1 σ2 D e jβ
Cz =
σ1 σ2 D e− jβ σ22
where D and β are the coherence magnitude and the interferometric phase. σ12 and σ22 are
the individual signal variances. An estimate of Cz could be calculated by the N-look sample
covariance matrix, as the average of several independent samples, and it is given as (Touzi &
Lopez, 1996)

1 N T ∗ 1 N z21 (k) z1 (k )z2∗ (k)
Cz =
Zk Zk =
N k∑
=1
N k∑
=1
z1∗ (k)z2 (k) z22 (k)
where N is the number of looks. Thus, the off-diagonal elements describe the complex mul-
tilook interferogram, where the noise problem is reduced by averaging neighboring pixels.
Integrating the complex Wishart-distribution with respect to the diagonal elements will leads
to the joint density function of the interferometric magnitude and phase given by (full ex-
pressions are given in (Touzi & Lopez, 1996)). Integrating this joint pdf with respect to the
magnitude leads to the marginal phase density function (Lee et al., 1994):
Γ( N + 12 )(1 − d2 ) N d cos( ϕ − β) (1 − d2 ) N 1 2 2
p’ ( ϕ ) = √ 1 + 2 F1 ( N, 1; ; d cos ( ϕ − β ))
2 2
2 π Γ( N )(1 − d cos ( ϕ − β)) 2 N + 2π 2
(3)
An example of the marginal pdf (3) is plotted in Fig. 1 for N = 9 and different coherence
values, varying between 0.9 and 0.3. The peak of the distribution is located at ϕ = β which
is considered to be nul in the example of Fig. 1. It can be shown that the phase is uniformly
distributed when D = 0 and becomes a Dirac delta function when D = 1. It is also obvious
that multi-look processing improves the phase accuracy, which lead to the decreasing of the
phase variance for a larger number of looks (Lee et al., 1994).
D = 0.9
6
4
Phase pdf
3 D = 0.7
2 D = 0.5
D = 0.3
1
0
−1.5 −1 −0.5 0 0.5 1 1.5
Phase [rd]
Fig. 1. Theoretical marginal multi-look pdf of the interferometric phase with N = 9 and β = 0.
2.2 Decorrelation effects

The interferometric phase can be affected by mainly three decorrelation factors: thermal
(ρ̂thermal ), temporal (ρ̂temporal ) and geometrical (ρ̂ geom ) decorrelation (Hanssen, 2001). The
Doppler centroid decorrelation and the processing induced decorrelation could be avoided
or neglected (Franceschetti & Lanari, 1999). The considered decorrelation effects increase the
interferometric phase noise, which will complicate more the unwrapping phase operation.
When considered together, the above decorrelation factors are multiplicative so the approxi-
mate total decorrelation value ρ̂total can be estimated as (Zebker & Villasenor, 1992):
ρ̂total ρ̂thermal · ρ̂temporal · ρ̂ geom (4)

where :
• ρ̂thermal is the decorrelation value induced by the temperature of the sensors (thermal
noise) on the interferometric phase during acquisition. It can be expressed by the signal
to noise (SNR) of the specific sensor by (Zebker & Villasenor, 1992):
1
ρ̂thermal = (5)
1 + SNR−1
The value of the SNR ratio is defined as:
PS
SNR =
PN
where Ps is the signal power and PN the noise power.

• ρ̂temporal is the temporal decorrelation which indicates all the physical changes occur-
ring on the surface between acquisitions, mainly in the case of repeat-pass interferome-
try. It includes changes of soil moisture content, surface roughness and vegetation. An
analytical model of the temporal decorrelation is given in (Zebker & Villasenor, 1992)
by:
2
1 4π
ρ̂temporal = exp − (σx2 sin2 θ + σz2 cos2 θ ) (6)
2 λ
where σx and σz are the average standard deviation, respectively, in azimuth and in
range of a random displacement of the reflectors inside a resolution cell. λ is the radar
wavelength and θ is its average look angle.
• ρ̂ geom represents the effects of the geometric decorrelation on the interferometric phase
due to the interferometric acquisition geometry (baseline, registration, target rotation
with respect to the sight line). Indeed, the same ground resolution cell is imaged from
two slightly different looking angle. This means that the sensors (for both the interfer-
ometric acquisitions) don’t look at the target with the same incidence angle. The nu-
merical assessments of these effects depend on the length of the baseline, the incidence
angle and the spatial resolution. In (Zebker & Villasenor, 1992), the authors provide
the geometric baseline decorrelation function as the result of the phase offset due to the
difference in the incidence angle between the two InSAR acquisitions:
2B⊥ l g cos θ
ρ̂ geom = 1 − (7)
λR0
where B⊥ is the orthogonal component of the baseline to the radar look direction, l g the
ground range resolution, and R0 the slant range. A modified version of (7), given by
Lee and Liu (Lee & Liu, 1999), includes the terrain slope α is given by:
cB⊥ | cos(θ0 − α) |
ρ̂ geom = 1 − (8)
λR0 Bw
where c is the speed of light, θ0 the nominal incidence angle on the ellipsoidal earth (23◦
for ERS-1 and ERS-2) and Bw the frequency bandwidth of the transmitted signal.
2.3 InSAR phase noise model

The InSAR phase quality is measured with the absolute coherence (Abdelfattah & Nicolas,
2006). Coherence interference due to reflection by random scatterers degrades the complex
image. The amplitude is corrupted by multiplicative noise, while the phase is corrupted by
additive noise. Indeed, based on (3), Lee et al. proved in (Lee et al., 1998) that: Since ϕ
distribution is symmetrical about β, β is the mean. The standard deviation is independent of
β. Consequently, ϕ can be characterized by an additive noise model in the real domain:
ϕz = ϕ x + ν (9)
where ϕz is the measured value, ϕ x is the original phase to be estimated and ν is the zero-mean
noise with the standard deviation σν . In the real domain, this phase noise model presents the
problem of phase jumps. It is due to the complex representation of the interferometric phase
(a wrapped phase, modulo 2π). A real phase value of (−π) could be equal in a complex rep-
resentation to (−π + 2π = +π). Thus, the phase jumps from (−π) to (+π). In order to avoid
the filtering of phase jumps, before the unwrapping process, it is more convenient to consider
the complex domain for the interferometric phase noise reduction, where phase jumps are not
present. Moreover, the interferometric phase is represented as a given point on the unit circle.
Lets consider
e jϕz = cos( ϕz ) + j sin( ϕz ) (10)

Then, the complex InSAR phase noise model for (10) could be derived easily from (9) for each
term (real and imaginary part). Then, the complex interferometric phase noise model is given
by:

 cos( ϕz ) = cos( ϕ x ) cos (ν) − sin( ϕ x ) sin(ν)

ν1 ν2 (11)

sin( ϕz ) = ν1 sin( ϕ x ) + ν2 cos( ϕ x )
where the original phase ϕ x is separated from the noise contribution ν. In (López-Martı́nez
& Fabregas, 2002), Lopez and Fabregas had proved and validated this model using simulated
and real interferometric phases. They deduced a more practical complex model given by:

cos( ϕz ) = Nc cos( ϕ x ) + νc
(12)
sin( ϕz ) = Nc sin( ϕ x ) + νs
where Nc is the mean value of the noise contribution ν1 , which can be represented by the
addition of its mean plus zero-mean random variable (ν1 )
ν1 = cos(ν) = Nc + ν1 (13)
The same representation can be used for the noise contribution ν2
ν2 = sin(ν) = Ns + ν2 (14)

where Ns = 0 (López-Martı́nez & Fabregas, 2002) is the mean value of ν2 and (ν2 ) the zero-
mean random variable. Thus, the expressions of νc and νs are given by:

νc = ν1 cos( ϕ x ) − ν2 sin( ϕ x )
(15)
νs = ν1 sin( ϕ x ) + ν2 cos( ϕ x )
3. InSAR phase noise reduction using a multiresolution approach

The multi-look process reduces some noise. The remaining noise has to be reduced by a fil-
tering process such as the adaptive Suksmono filtering (Suksmono & Hirose, 2002), adaptive
Lee filtering (Lee et al., 1998) or Goldstein-Werner (G-W) filtering (Goldstein & Werner, 1998).
In this chapter, we consider a multiresolution filtering approach.
3.1 Multiresolution image filtering in the wavelet domain

The wavelet transform (WT) decomposes a signal into its low frequency components (ap-
proximations or residues) and high frequency components (details) (Mallat, 1998). The re-
sultant set of Coefficients is a pyramidal multiresolution representation of the original signal.
The most approximated component is located at the lowest resolution level and other levels
consist of corresponding detail components. The original signal is recovered in the recon-
struction process by synthesizing these approximated and detail components. In the discrete
wavelet transform (DWT), decomposition and reconstruction are conducted by filtering pro-
cesses, called the wavelet filter banks. Ideal filters for the DWT are needed in order to preserve
the linear characteristics of the WT. The ideal assumption holds if the DWT is computed using
the Shannon wavelet (Vidakovic, 1999).
Filtering in the wavelet domain allows one to perform signal processing operations that are
localized in both space and frequency (Hess-Nielsen & Wickerhauser, 1996). This can be ad-
vantageous in many non-stationary problems, such as radar imaging (Odegard et al., 1995)
and particularly SAR interferometry (Zhou et al., 1999). Furthermore, empirical evidence
shows that wavelet bases generally provide more efficient representations of real-world data
than pixel or frequency domain representations. Because wavelets are able to concisely rep-
resent complicated signal structure, filtering techniques based on the wavelet decomposition
are much better at separating signals from noise than classical approaches based in the pixel
or frequency domain. From a theoretical perspective, it has been shown that, because wavelet
bases are unconditional bases for wide classes of signals, the wavelet transform is optimal for
noise reduction (Donoho & Johnstone, 1995).
In this chapter we consider the InSAR phase noise reduction in the wavelet domain. This
choice is argued by the facts that:
• Filtering the InSAR phase noise is necessary for increasing the precision of practical
measurements (extracted from the InSAR phase topography). Thus, The practical ap-
plication of InSAR requires that the position of the fringes in the interferogram can’t be
shifted after being filtered.
• The fringe pattern in the interferogram have to be preserved after the filtering pro-
cess. The localization of the wavelet functions is particularly to preserve fine details
like edges (Donoho & Johnstone, 1995).
• Filtering in the wavelet domain don’t need a windowing process such for the classical
filters. Indeed, these filters are based on a windowing process of the original signal.
Consequently, the filter performance depends highly on the window dimensions.
3.2 A complex wavelet model for InSAR phase filtering

Using the above mentioned interferometric complex phase noise model (§ 2.3), Lopez and
Fabregas proposed in (López-Martı́nez & Fabregas, 2002) an equivalent model adapted in the
wavelet domain. Assuming ideal filters for the two dimensions DWT (DWT2D ), the noise
model equations, for a complex interferometric phase image in the wavelet domain, have the
following expressions (López-Martı́nez & Fabregas, 2002):

DWT2D (cos(φz )) = 2i Nc cos(φxw ) + νcw .
(16)
DWT2D (sin(φz )) = 2i Nc sin(φxw ) + νsw .
where νcw and νsw represent the DWT of the noise terms νc and νs , respectively, at the scale
2i . φxw represents the phase information within the complex wavelet domain. The main char-
acteristics of this model, as stated and proved in (López-Martı́nez & Fabregas, 2002), can be
summarized as follows:
• Each time a wavelet scale is calculated, the DWT increases the original signal compo-
nent by a factor 2, without varying the power’s noise. Indeed, the variance of the noise
component in the wavelet domain has the same value as its variance in the spatial do-
main (σν2cw = σν2c and σν2sw = σν2s ). This will allow a larger signal improvement for each
new wavelet decomposition level.
• The wavelet coefficient intensity, computed using the defined real and imaginary parts,
is directly related to the mean value of the noise contribution ν1 , Nc . The mean of the
intensity for a wavelet coefficient has the expression
2
jφz
E DWT2D e = 22i Nc2 + σν2cw + σν2sw (17)
When the wavelet coefficient contains only noise (i.e., Nc = 0, as Nc is monotone increas-
ing with the coherence), its value is reduced to (σν2cw + σν2sw ). However, for Nc = 0, and
considering three wavelet decomposition levels, Lopez and Fabregas (López-Martı́nez
& Fabregas, 2002) show that, σν2cw + σν2sw is negligible compared with 26 Nc , and then
2

E DWT2D e jφz 22i Nc2 (18)
• For the phase of the wavelet coefficients φzw , it has the following expression

2i Nc sin(φxw ) + νsw
jφz
arg DWT2D e = arctang (19)
2i Nc cos(φxw ) + νcw
Where, arg[], represents the argument of a complex quantity. When the wavelet coeffi-
cient contains only noise, its phase is given by
w
ν
arg DWT2D e jφz = arctang sw (20)
νc
Moreover, when the wavelet coefficients are computed for a large number of decompo-
sition levels (≥ 3), and for Nc = 0, the interferometric wavelet phase is given as

2i Nc sin(φxw )
jφz
arg DWT2D e = arctang φxw (21)
2i Nc cos(φxw )
Thus, we deduce the original (without noise) wrapped interferometric phase.
Considering the above properties, of the presented interferometric noise model in the wavelet
domain, Lopez and Fabregas had developed a noise reduction algorithm, which flow diagram
is illustrated in Fig. 2. The used filtering algorithm is then based on the analysis of the wavelet
residue bands (signal bands) using the discret packet transform and it is composed of six steps.
In the following, we resume these steps considering the paper of Lopez and Fabregas (López-
Martı́nez & Fabregas, 2002).
• Step 1: A three wavelet scales transform is applied to the complex interferometric phase.
In the third decomposition level, the wavelet filter is applied to all the bands (signals
(a2 ) + details of the second decomposition level (d2 )), using the Discrete Wavelet Packet
Transform (DWPT). This is in order to obtain a constant amplification by a factor 23
of the signal wavelet coefficient intensities of the first decomposition level. A detailed
description of the nomenclature of the DWPT is depicted in Fig. 3.
Fig. 2. Noise reduction algorithm flow diagram, proposed by Lopez and Fabregas (López-
• Step 2: A mask, indicating the pixel locations of the signal coefficients is generated.
This is done using a two defined parameters: The signal quality (Γsig ) and the threshold
(Thw ) parameter. (Γsig ) is given by the following expression
Iw − 22i σw2
Γsig = (22)
Iw
where Iw represents the intensity for those wavelet coefficients in the low frequency
sub-band (a2 and all the d2 ) and σw2 is the noise variance in the corresponding spatial
area described by Iw in the noise bands (d1 ). To detect the wavelet signal coefficients, a
threshold (thw ) is applied to Γsig
Γsig ≥ thw ⇒ Signal coefficient

Γsig < thw ⇒ Noise coefficient.
The threshold value considered by Lopez and Fabregas in (López-Martı́nez & Fabregas,
2002) is equal for all the 16 signal bands. It defines up to which coherence level the
Fig. 3. Wavelet transformation process employed to transform interferometric complex phase

images. Relations between pixels (black boxes) in different scales are shown (López-Martı́nez
& Fabregas, 2002).
signal is processed. In the resulting mask, those isolated coefficients that are detected
as signal coefficients are removed in order to reduce noise effects in the mask.
• Step 3: The real and imaginary parts of those coefficients detected as signal in the pre-
vious step are multiplied by two. For the noise coefficients the real and imaginary parts
are maintained.
• Step 4: The inverse DWT is applied but only reducing one wavelet scale.
• Step 5: To obtain a mask locating signal coefficients for previous scales (higher fre-
quency scales), a new mask is derived from the one generated in Step 2. Each four bands
in a scale 2i are derived from a single band in the previous wavelet scale 2i−1 , where
a 1-to-4 space relation is established between wavelet coefficients. First, the masks of
the four bands at the scale 2i are merged through a logical OR operation. Then, the
dimensions of the merged mask are doubled to fit the 2i−1 scale band dimensions. In
this case, if a pixel of the merged mask is classified as signal, the four pixels referring
to the same spatial area, but in the band of the scale 2i−1 , are also classified as signal;
otherwise they are classified as noise. This sequence of mask growing allows to obtain
a mask locating useful signals in the original domain.
Then, we iterate (dashed lines in Fig. 2) the sequence of Steps 3-5 three times in order
to obtain the complex signal in the original domain.
• Step 6: From the complex output of the algorithm, the phase is calculated, which is the
estimation of the original interferometric phase.
3.3 The proposed filtering algorithm

We propose in this section a modified version of the Lopez and Fabregas algorithm for the
SAR interferometric phase noise reduction in the wavelet domain. We aim by this modifi-
cation to overcome the inconsistencies in the resulting filtered interferogram caused by the
extrapolation in the mask growing step. Compared to this method, our proposal explained
below has two different features:
• The inverse discret wavelet transform (DWT) is applied considering an adaptive mask
extracted from the InSAR coherence map, sub-sampled to the convenient resolution of
the corresponding wavelet decomposition level (to be processed).
• The threshold is adaptive and it is computed based on the equivalent 4 signal bands,
and not all the signal bands.
Fig. 4. Noise reduction algorithm flow diagram, proposed by Lopez and Fabregas (López-
Figure 4 is the flow diagram of the developed algorithm named FAMM (for Filtrage par Ap-
proche Multi-échelle Modifiée). Both algorithms, the FAMM and the Lopez ones, consist of the
same six steps. However, the processing in Step 2 and Step 5 is different. In fact:
• In Step 2, when generating the mask, the Lopez algorithm consider the same threshold
for all the 16 (4 sub-bands from a2 and 4*3 sub-bands from d2 ) signal bands, where
we do not. The reason is that the a2 and the d2 are not the resulting coefficients of the
same process, and so they don’t have necessary the same signal dynamic. This could
be confirmed in (Buccigrossi & Simoncelli, 2001), where we find that the d2 sub-bands
have a generalized gaussian distribution (GGD) and the a2 coefficient has not. Thus, we
define two thresholds, thwa and thwd , for each sub-band category. The values of these
thresholds are computed with respect to the mean of the sub-band dynamics.
• In Step 5, when doubling the dimensions of the merged mask to fit the 2i−1 scale band
dimensions, we do not systematically classify the four pixels, in the band of the scale
2i−1 , to the same class. We propose to generate a sub-sampled coherence map, from the
initial InSAR coherence one, having the same dimensions as the the band of the scale
2i−1 . Then, the mask growing will depend on the coherence values of the four consid-
ered pixels in the band of the scale 2i−1 . Figure 5 shows an illustrative example of the
mask growing step for a given (m, n) pixel in the band of the scale 2i and its correspond-
ing four pixels in the band of the scale 2i−1 , {(k, l ); (k, l + 1); (k + 1, l ); (k + 1, l + 1)}. The
decision rule will depend on:
1. If the pixel (m, n) in the mask is a signal coefficient, so
|ρ(k, l ) − ρ p | ≤ ε c ⇒ p is a signal coefficient

|ρ(k, l ) − ρ p | > ε c ⇒ p is a noise coefficient
where p ∈ {(k, l + 1); (k + 1, l ); (k + 1, l + 1)}
2. If the pixel (m, n) in the mask is a noise coefficient, so
|ρ(k, l ) − ρ p | ≤ ε c ⇒ p is a noise coefficient

|ρ(k, l ) − ρ p | > ε c ⇒ p is a signal coefficient
Fig. 5. Mask growing taking into account, the InSAR coherence map.
3.4 Noise reduction on simulated and real interferograms

In order to validate the proposed noise reduction algorithm, synthetic and real interferograms
were used. The accuracy of the denoising process was estimated using two different ap-
proaches:
• The PSNR, computed between the filtered and the original simulated interferograms.
It is given by:

(2π )2
PSNR = 10log10
MSE
where MSE is the Mean Square Error between the filtered and the original simulated
interferograms.
• The difference image distribution, computed between the filtered and the original real
interferograms.
Moreover, a quantitative comparison (using the PSNR ratio) with alternative interferometric
phase filters addressed in the literature is reported.
Table 1. Interferometric phase filtering PSNR for the cone and pyramid interferograms, with
noise variance of 10−2 .
Interferogram Gaussian White Uniform

noise (dB) noise (dB) noise (dB)
Lopez cone 39.9 39.85 41.5
(López-Martı́nez & Fabregas, 2002) pyramid 39.51 39.77 40.89
FAMM cone 41.36 41.29 42.78
pyramid 41.13 40.81 41.67
Lee cone 37.18 37.52 37.22
(Lee & Liu, 1999) pyramid 36.74 37.03 37.28
The simulated data were a 512x512 pixel interferograms, representing a cone and a pyramid,
relatives to single look complexe images, with coherence varying from 0.9 to 0.45. More de-
tails could be found in (Abdelfattah & Bouzid, 2008) about the simulation process. The exper-
imental results given in (Abdelfattah & Bouzid, 2008) show that the Daubechies 20-coefficient
wavelet transform gives the best PSNR for both cone and pyramid interferograms. Table 1
shows the numerical comparison for these simulated interferograms. The used empirical pa-
rameters for the thresholds were thwa = −0.4, thwd = −0.2 and ε c = 6.10−4 . The results in Table
1 confirm that the FAMM filtering algorithm allows a gain more than 2 dB compared to the
one of Lopez and Fabregas (López-Martı́nez & Fabregas, 2002) and more than 4 dB compared
to the Lee one (Lee & Liu, 1999).
In order to test the spatial resolution maintenance properties of the different above mentioned
filters, we simulated a pyramid with phase noise coherence equal to 0.45 and a ten-pixel fringe
period. The critical point is to maintain the pyramid edges. Fig. 6 shows the noisy phase
image and the result obtained after applying the different filters. The pyramid edges are
clearly maintained for both the FAMM and the Lopez Filters.
We also tested the developed noise reduction algorithm (FAMM) on a real experimental ERS2
interferogram of Capbon, in the north of Tunisia. This single look complex interferometric
phase has a 2048x2048 pixel dimension with an approximate spatial resolution of 25x25 m.
This phase image has been filtered by the FAMM filter using the same parameters as those
used for the simulated interferograms. From Fig. 9 (c), one can notice that the algorithm, like
the Lopez one (López-Martı́nez & Fabregas, 2002), is able to process, in the same efficient way,
areas with smooth or steep slopes at the same time. Figure 9 (d) shows the difference between
the original and the filtered phases, which has a mean equal to zero and does not contain any
image detail, demonstrating that the proposed filter preserves the topographic information.
4. InSAR phase unwrapping based on MRF model

The aim of this section is to present a general framework for the phase unwrapping problem,
based on Markov random fields (MRF(s)) models (Griffeath, 1976). The proposed algorithm
exploit Bayesian estimation theory to perform energy minimization in order to reconstruct the
original phase field. Our contribution consists in the exploitation of the Interferometric phase
distribution (Abdelfattah & Nicolas, 2006) in order to develop an adaptive energy functional
for the optimization of the MRF to be retrieved.
6 6
50 50
100 5 100 5
150 150
4 4
200 200
250 250
3 3
300 300
350 2 350 2
400 400
1 1
450 450
500 500
100 200 300 400 500 100 200 300 400 500
(a) (b)
6 6
50 50
100 5 100 5
150 150
4 4
200 200
250 250
3 3
300 300
350 2 350 2
400 400
1 1
450 450
500 500
100 200 300 400 500 100 200 300 400 500
(c) (d)
Fig. 6. Interferometric phase filtering results with an interferogram representing a pyramid

with |ρ| = 0.45 and a variance Gaussian noise 10−2 . (a) Original noisy image, (b) FAMM filter,
(c) Lopez filter and (d) Lee filter.
4.1 General context

The problem of recovering the absolute phase field from observations is ill-posed and can be
solved if one first regularizes it by introducing further information (prior knowledge) about
the behavior of the solution. To introduce this additional information, we exploit a statistical
framework, which consist in considering that the searched field φ is a Markov random one
(Griffeath, 1976). Thanks to the Hammersly-Clifford theorem (Griffeath, 1976), which estab-
lishes a correspondence between MRF and Gibbs distribution, the distribution of φ field is
expressed with the following representation (Griffeath, 1976):
1
P(φ) = exp(−U (φ)) (23)
Z
where Z is a normalizing constant, and U (φ) is a function which encodes the local character-
istics of the phase behavior being modelled. The true phase field φ is related to the wrapped
phase field ϕ through the equation given by:
φ = ϕ + 2kπ (24)
6
200 200 0.9
400 5 400 0.8
600 600 0.7

4
800 800 0.6
1000 1000 0.5
3
1200 1200 0.4
1400 2 1400 0.3
1600 1600
0.2
1
1800 1800
0.1
2000 2000
500 1000 1500 2000 500 1000 1500 2000
(a) (b)
6 6
200 200
400 5 400 5
600 600
4 4
800 800
1000 1000
3 3
1200 1200
1400 2 1400 2
1600 1600
1 1
1800 1800
2000 2000
500 1000 1500 2000 500 1000 1500 2000
(c) (d)
Fig. 7. Filtering result with the FAMM filter of (a) the original interferometric phase of Capbon
(ERS2 interferogram), (b) the corresponding coherence map, (c) The filtered interferogram and
(d) the difference image (original interferogram - filtered interferogram).
where k is the integer to determine for each (i, j) pixel in the ϕ field, in order to solve (24)
and recover φ. Considering the MRF model in two dimensional processing, each pixel (i, j)
of an image is defined as a site s in a set of regular lattice of nodes S. Let us assume that the
observed data obey to the equation, related to each site s given by (Rodriguez-Vera & Servin,
1995):
φs = ϕs + 2πk s + q (25)
where φs the phase to be retrieved at the site s, ϕs is the observed phase at the site s and q is
a constant. We introduce a new field f as a correction field, defined by the elements f s , where
s ∈ S and given by (Rodriguez-Vera & Servin, 1995):
q
fs = ks + (26)
2π
Note that the values of each element f s of the field f , are not necessarily integers ( f s ∈ R), but
the differences are, so that for any two neighboring sites {s, t} ∈ S, we have (Rodriguez-Vera
& Servin, 1995):
f s − f t = r( f s − f t ) (27)
where f s and f t are phase values related to two neighboring sites and r ( x ) is the closet integer
to x. Thus, (25) yields (Rodriguez-Vera & Servin, 1995):
φs = ϕs + 2π f s (28)
Assuming (28), that introduce the correction field f , is verified everywhere in the lattice struc-
ture of the phase field φ, then we obtain the conditional probability, P( ϕ, f |φ), of the observed
phase ϕ and f given φ, that takes the form (Rodriguez-Vera & Servin, 1995):

1 if φs = ϕs + 2π f s ∀ s ∈ S
P( ϕ, f |φ) = (29)
0 otherwise
we note that the likelihood distribution P( ϕ, f |φ) corresponds to Dirac impulse response, and
can be expressed by (Rodriguez-Vera & Servin, 1995):
1 2
P( ϕ, f |φ) = exp −α ∑ ϕs − φs + 2π f s (30)
Z1 s∈S
where Z1 is a normalizing constant and α is a weight parameter.
4.2 Prior and posterior model

To define a probabilistic model that describes the real interferometric phase information, it is
necessary to define an energy functional:
U (φ) = ∑ Uc (φ) (31)
c∈C |s∈c
where the Uc (.) are called potentials functions, indexed by the clique c in the set of cliques C
of the neighborhood system of a given site s. Note that each potential Uc depends only on the
values taken on the clique sites xs , s ∈ c, and, therefore, accounts only for local interactions be-
tween neighboring pixels. As consequence, local dependencies in the field φ can be modelled
by defining suitable potentials. In the literature, a number of models have been proposed for
various applications, for further comprehension refer to (Li, 2001). To enforce global smooth-
ness of reconstructed phase, a quadratic potential model could be adopted (Rodriguez-Vera
& Servin, 1995):
Uc (φ) = (φs − φt )2 (32)
The equations (31) and (32) gives:
U (φ) = ∑ (φs − φt )2 (33)

c∈C |s∈c
Probabilistic description of the current estimate of the field (φ, f ), given the observed data ϕ,
is calculated using Bayes’rule (Rodriguez-Vera & Servin, 1995):
P( ϕ|φ, f ) P(φ, f )
P(φ, f | ϕ) = (34)
P( ϕ)
Assuming that P( ϕ) is a constant, and that the correction field f obey to the prior model
defined by (23), we obtain:
P(φ, f | ϕ) ∝ P( ϕ|φ, f ) P(φ, f ) (35)
1
∝ exp −U (φ, f ) (36)
Z2
where Z2 is a normalizing constant, and U (φ, f ) is the posterior energy function, given by
merging (23), (30) and (14) (Rodriguez-Vera & Servin, 1995):
2
U (φ, f ) = ∑ Uc (φ) + α ∑ ϕs − φs + 2π f s (37)
c∈C |s∈c s∈S
The maximum a posteriori (MAP) estimator for φ may now be obtained by the minimization
2
of (37). Using (28), and requiring that f s − f t − r ( f s − f t ) be small for nearest-neighbor pairs
of sites, the minimization problem involving two unknown fields φ and f , may be simplified
by expressing the energy in term of the field f only. This allows to absorb the noise and
interpolate the missing data. The new expression of the posterior energy becomes (Rodriguez-
Vera & Servin, 1995):
2
U ( f ) = ∑ Uc ( ϕ + 2π f ) + α ∑ f s − f t − r ( f s − f t ) (38)
c∈C |s∈c t∈Vs
where Vs is a set of sites neighbouring the site s and α is a regularization parameter that
controls the smoothness of f .
4.3 Minimization algorithm-Gradient descent method

The gradient descent optimization method could be adopted to restor the field f that minimize
the energy function denoted by (38), and hence to unwrap ϕ. this iterative method is useful
to prevent the initialization problem related to the application of the MRF models for phase
unwrapping procedure (Rodriguez-Vera & Servin, 1995). to start the decent algorithm with
the appropriate initial state, we must set the initial value as follow:
ϕs
fs = − (39)
2π
The implementation of the minimization algorithm, in which the gradient descent method is
performed by updating the value of each variable f s of the field f , in parallel programming
way, according to the following equation (Rodriguez-Vera & Servin, 1995):
( n +1) (n) ∂U [ f (n) ]

fs = fs −h (40)
∂ fs
where n denotes the iteration number, and h is the step size of the gradient descent. The
gradient of U is expressed by (Rodriguez-Vera & Servin, 1995):
∂U ( f (n) ) ∂Uc ( ϕ + 2π f (n) )
= ∑
∂ fs c∈C |s∈c
∂ fs
(41)
(n) (n) (n) (n)
+ 2λ ∑ f s − f t − r ( f s − f t )
t∈Vs
For the regular lattice S, we consider the first-order system, so we have C = C1 ∪ C2 , with
C1 = {s|s ∈ S} and C2 = {{s, t} |t ∈ Vs , s ∈ S}. Using (32) the developed form of (40) and then
the gradient descent automaton is given by (Rodriguez-Vera & Servin, 1995):

( n +1) (n) (n) (n)
fs = f s − 2h ∑ 2π ( gs − gt ) + 4π 2 ( f s − f t )
t∈Vs
(42)
(n) (n) (n) (n)
+ λ f s − f t − r( f s − f t )
where Vs = t : s − t = 1 and where we set ϕs = 0 and f s = 0 ∀ s ∈

/ S.
pi
50 pi/2
pi/4
100
150
−\pi/2
200
−pi/4
250 −pi
50 100 150 200 250
(a) (b)
Fig. 8. (a) Interferogram generated from a Gaussian DEM with a maximum height of 100 m
and standard deviations (in rows (σx ) and columns (σy )) of σx = σy = 2.10−4 . (b) Reconstructed
elevation data obtained with the gradient descent algorithm..
Figure 8 displays the interferogram generated from a synthetic Gaussian DEM with a maxi-
mum height of 100 m and standard deviations (in rows (σx ) and columns (σy )) of σx = σy =
2.10−4 . The algorithm described above works well with highly coherent interferograms (for
data affected by low noise level). However, gradient descent method does not give satisfactory
results in the case of noisy interferograms. In fact, the chosen a priori model does not inte-
grate a noise model. Thus, it is necessary to accommodate the potential function to recover the
discontinuities (noise) present in the observed data. A modification of the prior potentials is
proposed in (Rodriguez-Vera & Servin, 1995), making the phase jumps, that overcomes a cer-
tain threshold a, contribute with a fixed amount to the energy, independently of their size. A
truncated quadratic potentials, are then proposed and the (32) is replaced by (Rodriguez-Vera
& Servin, 1995):
(φs − φt )2 if |φs − φt | < a,
Uc (φ) = (43)
a2 otherwise
where a is a positive parameter.
Figure 9 (a) illustrates a noisy (mean = 0.6 and standard deviation = 0.9) interferogram gen-
erated from a two nearby Gaussians elevation data of height 100 m and 50 m, with local
inconsistencies (nul data) marked by blue squares. However, as we can see on Fig. 9 (b), the
reconstructed elevation data obtained with the gradient descent algorithm is not perfect, in
spite of the correction of the inconsistencies. This have lead us to think for an adaptive energy
functional modelling.
4.4 Our contribution

In order to enforce the global smoothness of the algorithm, the above considered quadratic
potentials is given by:
Vi,j (zi , z j ) = (zi − z j )2 . (44)
20
40
60
80
100
120
140
160
180
200
20 40 60 80 100 120 140 160 180 200
(a) (b)
Fig. 9. (a) Interferogram generated from a two nearby Gaussians elevation data of height 100
m and 50 m, with local inconsistencies marked by blue squares. (b) Reconstructed elevation
data obtained with the gradient descent algorithm..
Such a potential function makes this algorithm work well for smooth surface. However it fails
to recover discontinuous surfaces as well. Our main contribution in this work is to propose
a modification to the above prior potential so that the jumps contribution to the energy are
mitigated by weighting them by the probability of their occurrence which decrease when they
increase, the proposed potential function is then equivalent to (Elmzoughi et al., 2008):
Vi,j (zi , z j ) = (zi − z j )2 PZ (zi − z j ). (45)
In our case, Z corresponds to the absolute phase φ which pdf can be written as an expres-
sion implying hypergeometric functions. However, A good approximation can be given by a
Gaussian model:
1 (φ − φ)2
PΦ (φ) = √ exp(− ), (46)
2πσφ σφ2
where σφ and φ are respectively the variance and the mean of the absolute phase φ. Theses
parameters can be easily estimated from the theoretical hypergeometric pdf. By considering
the transfert theorem, the gradient descent automation in equation (38) could be given by:
√ t t 2

f it+1 = f it − σ2πh × [ g − g + 2π ( f t − f t )] exp [ gi − g j +2π ( f i − f j )]
φ
∑ j∈Vi i j i j 4σφ2
, (47)
× 1 − 4σ1 2 [ gi − g j + 2π ( f it − f jt )] + λ[ f it − f jt − r ( f it − f jt )]
φ
where Vi is the set of neighbors of the considered site i. In (Elmzoughi et al., 2008), it is
shown that both algorithms seems to be equivalent for smooth surfaces. However, using the
proposed potential function allowed to ameliorate results, especially in terms of mean square
errot and in terms of SNR, for surfaces with discontinuities. This could be explained by a
better preservation of these discontinuities in the reconstructed phase.
5. Conclusion
In this chapter, we proposed a modified filtering algorithm to the López and Fabregas (López-
Martı́nez & Fabregas, 2002) noise reduction algorithm for the interferometric phase noise in
SAR interferometry using a multiresolution approach. Our contribution to the existing algo-
rithm consists on the exploitation of the InSAR coherence map in order to generate a more
adaptive mask for each decomposition level.
Moreover, we presented a general probabilistic framework to phase unwrapping problem
based on the work of Rodriguez and Servin (Rodriguez-Vera & Servin, 1995). Both MRF and
Bayesian estimator were applied to recover the desired phase, as the optimal field solution of
the maximum a posteriori (MAP) estimation criterion. An iterative method that minimizes a
general energy function is proposed, and a parallel algorithm based on gradient descent opti-
mization is designed to perform this task. The proposed solution overcomes some important
limitations of most of the phase unwrapping procedures, and the results show robustness,
and stability.
These results give us new ideas for the applications of the InSAR unwrapping phase MRF
algorithm for unwrapping interferometric synthetic Aperture Sonar (InSAS) (Bonifant et al.,
2000) phase. Both two phases present similar statistics. However, if the ground truth could be
used for the InSAR result validations, it not the case for the InSAS results. In fact, the ability to
produce full coverage bathymetric maps and generate accurate measurements of the seafloor
height, is limited. So, an adaptive unwrapping procedure could be interesting.
6. References
Abdelfattah, R. & Bouzid, A. (2008). Sar interferogram filtering in the wavelet domain using
a coherence map mask, Proceedings of the 15th IEEE International Conference on Image
Processing, ICIP 2008, San Diego, CA, pp. 1888–1891.
Abdelfattah, R. & Nicolas, J. (2002). Topographic sar interferometry formulation for high-
precision dem generation, IEEE Transaction on Geoscience and Remote Sensing Vol.
40(No. 11): pp. 2415–2426.
Abdelfattah, R. & Nicolas, J. (2006). Interferometric sar coherence magnitude estimation using
second kind statistics, IEEE Transaction on Geoscience and Remote Sensing Vol. 44(No.
7): pp. 1942–1953.
Balzter, H. (2001). Forest mapping and monitoring with interferometric synthetic aperture
radar (insar), Progress in physical geography Vol. 25(No. 2): pp. 159–177.
Bodart, C., Gassani, J., Salmon, M. & Ozer, A. (2009). Contribution of sar interferometry (from
ers1/2) in the study of aeolian transport processes: The cases of niger, mauritania
and morocco, in A. Marini & M. Talbi (eds), Desertification and Risk Analysis Using
High and Medium Resolution Satellite Data Training Workshop on Mapping Desertification,
Springer, pp. 129–136.
Bonifant, W., Richards, M. & McClellan, J. (2000). Interferometric height estimation of the
seafloor via synthetic aperture sonar in the presence of motion errors, IEE Proceedings
- Radar, Sonar and Navigation Vol. 147(No. 6): pp. 322–330.
Buccigrossi, R. & Simoncelli, E. (2001). Image compression via joint statistical characterization
in the wavelet domain, IEEE Transactions on Image Processing Vol. 8(No. 12): pp. 1688–
1701.
Donoho, D. & Johnstone, I. (1995). The zπm algorithm: a method for interferometric image
reconstruction in sar/sas, Journal of the American Statistical Association Vol. 90(No.
432): pp. 1200–1224.
Elmzoughi, A., Maalej, A., Abdelfattah, R. & Belhadj, Z. (2008). Insar phase unwrapping based
on a combination of markov random fields and hypergeometric phase pdf models,
Proceedings of the IEEE International Geoscience and Remote Sensing Symposium, IGARSS
2008, Boston, Massachusetts, pp. 538–541.
Franceschetti, G. & Lanari, R. (1999). Synthetic Aperture Radar processing, CRC Press.
Gabriel, A. & Goldstein, R. (1988). Retrieval of vegetation parameters with sar interferometry,
International Journal of Remote Sensing Vol. 9(No. 4): pp. 857–872.
Ghiglia, D. & Pritt, M. (1998). Two-dimensional phase unwrapping: Theory, algorithms and software,
John Wiley & Sons Inc.
Ghiglia, D. & Romero, L. (1994). Robust two-dimensional weighted and unweighted phase
unwrapping that uses fast transforms and iterative methods, Journal of the Optical
Society of America A Vol. 11(No. 1): pp. 107–117.
Goldstein, R. & Werner, C. (1998). Radar interferogram filtering for geophysical applications,
Geophysical Research Letters Vol. 25(No. 21): pp. 4035–4038.
Goldstein, R., Zebker, H. & Werner, C. (1988). Satellite radar interferometry: two-dimensional
phase unwrapping, Radio Science Vol. 23(No. 4): pp. 713–720.
Graham, L. (1974). Synthetic interferometer radar for topographic mapping, Proceedings of the
IEEE Vol. 62(No. 6): pp. 763–768.
Griffeath, D. (1976). Introduction to markov random fields, in J. Kemeny, J. Snell & A. Knapp
(eds), Denumerable Markov Chains, Springer-Verlag, pp. 424–457.
Hanssen, R. (2001). Radar Interferometry: Data Interpretation and Error Analysis, Kluwer Aca-
demic Publishers.
Hess-Nielsen, N. & Wickerhauser, M. (1996). Wavelets and time frequency analysis, Proceed-
ings of the IEEE Vol. 82(No. 4): pp. 523–540.
Lee, H. & Liu, J. G. (1999). Spatial decorrelation due to topography in the interferometric sar
coherence imagery, Proceedings of the IEEE International Geoscience and Remote Sensing
Symposium IGARSS’99, Hamburg, Germany, pp. 485–487.
Lee, J., Hoppel, W. & Mango, S. (1994). Intensity and phase statistics of multilook polarimetric
and interferometric sar imagery, IEEE Transaction on Geoscience and Remote Sensing
Vol. 32(No. 5): pp. 1017–1027.
Lee, J., Papathanassiou, K., Ainsworth, T., Grunes, M. & Reigber, A. (1998). A new technique
for phase noise filtering of sar interferometric phase images, IEEE Transaction on Geo-
science and Remote Sensing Vol. 36(No. 5): pp. 1456–1465.
Li, S. (2001). Markov Random Field Modeling in Image Analysis, Springer-Verlag.
López-Martı́nez, C. & Fabregas, X. (2002). Modeling and reduction of sar interferometric
phase noise in the wavelet domain, IEEE Transaction on Geoscience and Remote Sensing
Vol. 40(No. 12): pp. 2553–2566.
Maı̂tre, H. (2001). Traitement des images de RSO, Hermes.
Mallat, S. (1998). A wavelet tour of signal processing, Academic Press.
Massonet, D. & Feigl, K. (1998). Radar interferometry and its application to changes in the
earth’s surface, Reviews of Geophysics Vol. 36(No. 4): pp. 441–500.
Mattar, K., Vachon, P., Geudtner, D., Gray, A., Cumming, I. & Brugman, M. (1998). Vali-
dation of alpine glacier velocity measurements using ers tandem-mission sar data,
IEEE Transaction on Geoscience and Remote Sensing Vol. 36(No. 3): pp. 974–984.
Odegard, J., Guo, H., Lang, M., Burrus, C., Jr., R. . W., Novak, L. & Hiett, M. (1995). Wavelet
based sar speckle reduction and image compression, Proceedings of the SPIE Sympo-
sium on OE/Aerospace Sensing and Dual Use Photonics, Orlando, Florida, pp. 17–21.
Rodriguez-Vera, R. & Servin, M. (1995). Modeling and reduction of sar interferometric phase
noise in the wavelet domain, Journal of the Optical Society of America A Vol. 12(No.
12): pp. 2578–2585.
Suksmono, A. & Hirose, A. (2002). Adaptive noise reduction of insar images based on a
complex-valued mrf model and its application to phase unwrapping problem, IEEE
Transaction on Geoscience and Remote Sensing Vol. 40(No. 3): pp. 699–709.
Touzi, R. & Lopez, A. (1996). Statistics of the stokes parametersand of the complex coherence
parameters in one-look and multilook speckle fields, IEEE Transaction on Geoscience
and Remote Sensing Vol. 34(No. 2): pp. 519–531.
Vidakovic, B. (1999). Statistical Modeling by Wavelets, Wiley-interscience.
Wegmullerl, U. & Werner, C. (1997). Retrieval of vegetation parameters with sar interferome-
try, IEEE Transaction on Geoscience and Remote Sensing Vol. 35(No. 1): pp. 18–24.
Wishart, J. (1928). The generalized product moment distribution in samples from a normal
multivariate population, Biometrika Vol. 20A(No. 1/2): pp. 32–52.
Zebker, H. & Villasenor, J. (1992). Decorrelation in interferometric radar echoes, IEEE Transac-
tion on Geoscience and Remote Sensing Vol. 30(No. 5): pp. 950–959.
Zhou, Y., Li, D. & Genderen, J. V. (1999). Wavelet based sar speckle reduction and image
compression, Proceedings of the SPIE Symposium Vol. 3723, Wavelet Applications VI, Or-
lando, Florida, pp. 208–214.
Soil moisture estimation using L-band radiometry 441
21
X
Soil moisture estimation using

L-band radiometry
Alessandra Monerris1 and Thomas Schmugge2
1Remote Sensing Laboratory, TSC-UPC, and SMOS-BEC, Barcelona, Spain
2Physical Sciences Laboratory, New Mexico State University, Las Cruces, NM, USA
1. Introduction
About 70% of the Earth is covered with water, 97% of which is part of oceans. Heating of
oceans by the Sun keeps the Earth's water in a continuous circulation from the atmosphere
to the Earth and back to the atmosphere through condensation, evapotranspiration, and
precipitation processes. This continuous water motion is called water cycle. The
groundwater in the unsaturated or vadose zone of land is usually referred to as soil
moisture. The thickness of this zone extends from soil surface to a few metres below the
surface in humid regions, and to 300 m or more below surface in arid regions. Although the
water held by soils is a small fraction of the Earth's water budget, soil moisture plays an
important role in the water cycle since it controls the proportion of rainfall that percolates,
runs off, or evaporates from the land, influences plants growth and transpiration, and is
related to the precipitation variability within a region (Koster et al., 2004).
Although soil moisture is one of the main parameters used in climate models, because of the
large temporal and spatial variations of soil moisture sparse in-situ measurements are
inadequate to be of much use in these models. Remote sensing with sufficient accuracy
would provide meaningful soil moisture data over large regions. The maximum sensitivity
to soil moisture is achieved in the lower range of microwave frequencies, which goes from 1
to 5 GHz (wavelength from 30 cm to 5 cm, respectively). The justification to this behaviour
can be found if the dielectric properties of a target are analysed, since they have a large
influence on its microwave brightness temperature. Land surfaces can be considered as a
mixture of soil, water, and air particles and thus soil moisture might be estimated from
measurements if the contrast between the water and soil particles were large enough. At
microwave frequencies the real part of dry soil and water dielectric constants are
approximately 4 and 80, respectively (Jackson & Schmugge, 1989. The scientific literature
concluded that microwave radiometry at L-band (1.4-1.427 GHz) is optimal to estimate soil
moisture, not only because it is very sensitive to soil moisture, but also because provides all-
weather coverage, since the atmosphere at microwave frequencies may be considered nearly
transparent, and vegetation is semi-transparent, which allows observations of the
underlying layers (Eagleman, 1976; Wang & Choudhury, 1981; Jackson & Schmugge 1991).
This specific band was chosen because it is a radio astronomy quiet band.
1.1. Satellite missions for soil moisture estimation using passive microwave sensors
While L-band microwave sensors provide maximum sensitivity to soil moisture their long
wavelength implies the need for large antennae to achieve useful spatial resolution. For
example, to obtain a ground resolution of 50 km or less using classical solutions on low-orbit
satellites implies an antenna size of up to 20 m. At present, different scientific groups are
developing new techniques to face this problem, and two space missions have been
proposed to measure soil moisture at global scale: ESA's SMOS and NASA's SMAP.
The Soil Moisture and Ocean Salinity mission

The Soil Moisture and Ocean Salinity (SMOS) mission will observe soil moisture over land
and salinity over oceans, and will measure snow and ice areas contributing to studies of the
cryosphere (Kerr et al., 2001; Barré et al., 2008). Scheduled for launch in autumn 2009, it is
the second Earth Explorer mission and part of ESA's Living Planet Programme. SMOS was
thought of as a cost-effective, demonstrator mission with a nominal (extended) lifetime of 3
(5) years. The orbit is quasi-circular, sun-synchronous and dawn-dusk, and will be in the
low-Earth range, at 763 km.
The SMOS mission provides a completely new approach in the field of remote sensing at L-
band. Its payload MIRAS (Microwave Radiometer by Aperture Synthesis) is the first-ever
two-dimensional interferometric radiometer in space. MIRAS consists of a Y-shaped
antenna array with three arms, each arm having an approximate length of 4.5 m and 21
dual-polarisation L-band antennae or LICEF (Light Cost Effective Front-end) spaced 0.875
times the wavelength. Each LICEF is a total power radiometer in its own. Nine redundant
antenna and three full-polarimetric noise injection radiometers are located in a fixed
structure in the centre of the array, while the arms are divided in three segments to be
folded during the launch (McMullan et al., 2008). If classical solutions would have been
used, with this antenna size and orbit altitude the field of view (FOV) would be of near 3000
km in diameter. However, because of the microwave interferometry technique, the
instrument Y-shape and the antenna spacing, the resulting FOV is a hexagonal-like area of
less than 1000 km (see Figure 1). The spatial resolution varies from 35 km at the FOV centre
to 50 km at the border. MIRAS will provide observations of a single azimuth at various
incidence angles (from 0° to 55°) and radiometric resolutions depending on its position
within the field of view during a satellite overpass. This fact will much improve the retrieval
algorithms because a lot of independent information of each pixel will be registered, and
will permit the estimation of soil roughness, vegetation opacity and albedo, etc. together
with soil moisture.
To satisfy the scientific requirements, SMOS aims at providing global maps of soil moisture
every 3 days with a ground resolution better than 50 km, and a 0.04 m3/m3 volumetric
humidity. This soil moisture accuracy is referred to pixels outside mountainous, urban, and
partially frozen or snow-covered areas. For sea salinity, maps with accuracy better than 1.2
psu and 200 km ground resolution will be acquired every 30 days.
(a) (b)
Fig. 1. (a) SMOS field of view. Blue lines indicate the observation angle, black ellipses
indicate the pixel size depending on the position within the snap-shot, and red lines show
how a single pixel can be observed under different viewing positions during a satellite
overpass. (b) MIRAS during the IVT campaign at ESA-ESTEC Maxwell anechoic chamber
premises (picture courtesy of EADS-CASA Espacio)
The Soil Moisture Active and Passive mission

The Hydrosphere State (HYDROS) mission was proposed by NASA to enhance the
understanding of the land hydrosphere state and improve the climate prediction models
(Entekhabi et al., 2004). The mission had a circular, polar, sun-synchronous, 6 am/pm
equator crossing orbit located at 670 km altitude. Hydros aimed at providing soil moisture
estimates with a 4% volumetric accuracy in the top 2-5 cm and capturing freeze/thaw state
transitions in integrated vegetation-soil continuum at the spatial scale of landscape
variability (3 km). The payload consisted of both active and passive sensors. An L-band
radiometer measured the first, second, and third Stokes parameters with a 40 km spatial
resolution, and 1 K relative accuracy, while an L-band radar acquired VV, HH, and HV
polarisations with a 10 km resolution, and 0.5 dB accuracy for VV and HH. Observations at
a constant incidence angle between 35° and 50°. Both the radar and the radiometer share a 6
m diameter reflector antenna which rotates about its nadir axis. The HYDROS mission was
cancelled in 2005, but many HYDROS science and technology issues are being reviewed to
be implemented in the future Soil Moisture Active and Passive (SMAP) mission.
1.2. Scope of this chapter

Space missions for Earth observation require a huge amount of previous work to validate
the existing science and technology. This is accomplished through airborne and ground-
based field experiments. This chapter reviews the basics of microwave radiometry and the
existing models of soil emissivity and soil dielectric constant. The field experiments over
land carried out during the last three decades are summarised, paying special attention to
the recent experiments carried out by the Remote Sensing Laboratory from the Universitat
Politècnica de Catalunya (UPC) as part of the preparatory activities for the SMOS mission
over land.
2. L-band emission of land covers

Microwave remote sensing is based on the measurement of the thermal radiation or
brightness temperature of a target, which is determined by its physical temperature and
emissivity. The emissivity of land covers depends on soil moisture, but also on soil
temperature (Choudhury et al., 1982; Wigneron et al., 2001), soil surface roughness (Mo &
Schmugee, 1987; Escorihuela et al., 2007), vegetation canopy (Jackson & Schmugee, 1991;
Ferrazzoli et al., 2002), snow cover (Schwank et al, 2004), relief (Mätzler & Standley, 2000;
Talone et al., 2007), etc.
2.1. Soil dielectric constant model

The dielectric constant determines the response of the soil to an incident electromagnetic
wave. This response is composed of two parts, real and imaginary (ε = ε' + jε'') which
determine the wave velocity and energy losses respectively. In a non-homogeneous medium
such as soils, the dielectric properties have a strong impact on its microwave emission.
However, the relationship between the soil dielectric constant and the soil physical
properties is not straightforward. A large number of studies have been performed during
the last decades to find out this relationship since it plays an important part in the soil
moisture retrieval algorithms from remote sensing data (Hipp, 1974; Wang & Schmugge,
1980; Topp et al., 1980; Hallikainen et al., 1985; Dobson et al., 1985; Roth et al., 1992; Mironov
et al., 2004). Some of these models are simple empirical models in which data is fitted by a
curve unique for all soils; others propose semi-empirical approaches which take into
account some soil physical properties. The dielectric constant of dry soils is almost
independent of temperature (Topp et al., 1980) and frequency. On the contrary, wet soils
show a complex behaviour depending on the interaction between soil, water, and air
particles. Hallikainen et al. (1985) performed a series of dielectric constant measurements of
five soils with different texture composition at frequencies between 1.4 and 18 GHz and
found out that texture has a strong effect on the dielectric behaviour which is specially
pronounced at frequencies below 5 GHz.
In the dielectric-mixing model by Roth et al. (1992), differences in soil texture and bound
water and free water are ignored altogether. Other models use a semi-empirical approach
that contains a model of the complex dielectric constant and the volume fraction of each of
the soil components. This kind of approach was followed by two widely used models
presented in Wang & Schmugge (1980) and Dobson et al. (1985). The starting point of both
of them is the dielectric mixing model by (Birchak et al., 1974),
ε α = Vsε α + Vaε aα + V fw ε fwα + Vbwε fwα , (1)
where Vs (εs), Va (εa), Vfw (εfw) and Vbw (εbw) are the volume fraction (dielectric constant) of
solid phase, air, free water, and bound water in the soil, respectively. The expression in (1)
can be rewritten as a function of the bulk density ρb, the particle density ρs, and the
volumetric moisture mv as
ρb α
εα = 1 + (ε s - 1) + V fw ε fwα + Vbwε fwα - m v . (2)
ρs
The Wang & Schmugge (1980) model was proposed for 1.4 and 5 GHz frequencies and starts
from (2) with α=1. This model provides separate dielectric constant equations for volumetric
water content lower than or greater than the transition moisture wt. The transition moisture
is the moisture content at which the free water phase begins to dominate the soil hydraulics,
and it is strongly dependent on texture. The soil dielectric constant is then estimated as
m vε x + (P - m v )ε a + (1 - P)ε r , m v < w t
ε= . (3)
w tε x + (m v - w t )ε fw + (P - m v )ε a + (1 - P)ε r , m v > w t
with
 m
ε i + (ε fw - ε i ) γ , m v < wt
v
εx =  wt . (4)
ε + (ε - ε )γ , m > w
 i fw i v t
and γ = -0.57wt + 0.481, εi, εa, εfw, and εr the dielectric constants of ice, air, free water, and
rock, respectively, εx the dielectric constant of the initially adsorbed water, and P the soil
porosity. The variable εfw is estimated using the Debye equation.
On the other hand, the Dobson et al. (1985) model starts from (2) and assumes that there is
not distinction between bound and free water. Taking this into account, the new expression
for the dielectric constant is
1 /α
 ρb α 
ε = 1 + (ε s - 1) + m βv ε fwα - m v  . (5)
 ρs 
The real and imaginary parts of the dielectric constant of soils are obtained separately using
the percentage of sand S and clay C in the soil, and β =(127.48-0.519S-0.152C)/100 for the
real part, and β =(1.33979-0.603S-0.166C)/100 for the imaginary part. For further information
on dielectric constant models refer to Behari (2005) and Chukhlantsev (2006).
2.2. Emission from bare soils

The emissivity of bare soils es is given by
Γs,p = 1 − e s,p (6)
where Γs,p is the total reflectivity at p polarisation in a chosen observation direction, which can
be expressed in terms of the bistatic scattering coefficients σpp and σpq as follows (Fung, 1994):
Γs, p (θ i , ϕ i ) =
1
[ ]
∫ σ pp (θ i , ϕ i , θ r , ϕ r ) + σ pq (θ i , ϕ i , θ r , ϕ r ) dΩ r .
4π 4π
(7)
Subscripts i and r stand for the incident and scattered radiation, and θ and ϕ are the
incidence and azimuth angles, respectively. The coefficient σpp, often called coherent
component, considers the scattering on the observation direction and polarisation, while σpq,
or incoherent component, considers the scattering in whatever other direction and
polarisation. The integration over the entire half space and errors in the modelling of the
bistatic scattering coefficients make the computation of (7) unfeasible or prone to errors. For
this reason, the uniform half space model with smooth surface is commonly adopted. This
model is simpler and thus suitable to be used in soil moisture retrieval algorithms from
remotely sensed data. In this case, the total reflectivity in (7) equals the reflection coefficient
given by the Fresnel formulation.
Soil brightness temperature is related to soil emission through the soil effective temperature,
Teff, TBp = epTeff. The theoretical effective temperature of a soil profile can be estimated as
(Ulaby et al., 1986):
∞
[ x
Teff = ∫0 T( z)α ( z) exp − ∫0 α ( z' ) dz' dz ,] (8)
where T(z) is the thermodynamic temperature at a depth z, and α is the attenuation

coefficient, which is a function of the soil dielectric constant and of the microwave emission
wavelength λ. Several simple formulations have been developed to estimate the soil
effective temperature from soil properties, and soil moisture and temperature profiles.
Choudhury et al. (1982) proposed a parameterisation of Teff based on the soil temperature at
deep soil (T∞) corresponding to a depth between 50 cm and 1m, and on a surface
temperature (Tsurf) corresponding to a depth of 0 to 5 cm:
Teff = T∞ + C (Tsurf − T∞ ). (9)
The coefficient C was considered constant for a given frequency, and equal to 0.246 at L-
band. On the other hand, Chanzy et al. (1997) presented a model for the soil effective
temperature at L- and C-bands based on the air temperature, a deep soil temperature, and
the brightness temperature measured at X-band and V-pol. Wigneron et al. (2001) proposed
a parameterisation based on (9), but with a coefficient C dependent on the volumetric water
content ws, and two semi-empirical parameters. Another formulation using the soil
dielectric constant instead of the volumetric water content was proposed by Holmes et al.
(2006). The performance of some of these approaches is analysed in Wigneron et al. (2008).
The effect of soil surface roughness on the brightness temperature has been an issue widely
addressed in the literature (Choudhury et al., 1979; Wang & Choudhury, 1981; Mo &
Schmugge, 1987; Wang, 1983; Wegmüller & Mätzler, 1999; Wigneron et al., 2001;
Schneeberger et al., 2004; Escorihuela et al., 2007). Fung (1994) proposes a theoretical
physical model based on surface characteristics derived from the measured soil height
profile. A simple empirical roughness model which takes into account only the coherent
term of the scattering was reported in Choudhury et al. (1979),
[
Γs,p = Γs*,p exp − ( 2 kσ s cos θ )2 ,] (11)
where Γs*, p is the Fresnel specular reflectivity of soil, k is the electromagnetic wave number,
σs is the standard deviation of the surface height, and θ is the incidence angle. This model
was reviewed, and another formulation was proposed in Wang & Choudhury (1981):
[ ]
Γs, p = (1 − Qs ) Γs*, p + Qs Γs*, q exp( − h s cos θ n ). (12)
In this case, two semi-empirical parameters (hs and n) were included to model the effects of
the polarisation mixing Qs, and surface roughness. The dependence of these parameters on
surface properties such as correlation length or standard deviation of height is not yet clear.
Mo & Schmugge (1987) and Wigneron et al. (2001) conclude that the n=2 dependence in (12)
is too strong for L-band. A value of n=0 at both polarisations was found to be consistent
with measurements in Wigneron et al. (2001), while Escorihuela et al. (2007) distinguishes n
values for both polarisations (1 at horizontal and -1 at vertical). Similarly, there are
discrepancies on the value of roughness parameter hs. Some authors obtain hs from
experimental data by best-fit (Wigneron et al., 2001) while others propose expressions for hs
as a function of geophysical parameters. Mo & Schmugge (1987) obtain good results with
two parameterisations of hs as a function of σs and lc. Finally, there is a general agreement on
the value of the cross-polarisation parameter Qs, which has been found to be very small (0 to
0.12) at L-band.
Apart from these considerations, the effects of frequency and incidence angle on the
roughness parameters have not been studied thoroughly. Mo & Schmugge (1987) and Shi et
al. (2002) pointed out that the roughness effects depend on both the frequency and the
incidence angle. The later suggested a parameterisation of the surface reflectivity derived
from data simulated for a wide range of soil water content and roughness properties using
the integral equation model (Fung, 1994). The surface reflectivity model of Shi et al. (2002) was
tested in Schneeberger et al. (2004) and found not to be capable of explaining discrepancies
between the ground truth and remotely sensed data. As a consequence, a new model was
developed for describing the influence of the topsoil structure on the L-band emission as an
impedance matching between the dielectric constants of soil and air (Mätzler, 2006).
If the remote sensor is placed above a canopy looking downwards, the measured brightness
temperature will contain not only information on the soil, but also on vegetation, since
vegetation radiates its own energy and, moreover, attenuates and scatters the soil radiation.
Chukhlantsev (2006) revises the theory and conducted experimental research over vegetated
areas. Although the modelling of the land emission involves analytical solutions of the
radiative transfer equation (Ferrazzoli et al., 2002), this approach is not easy to use with
experimental data. Hence, the common practice is to use approximate formulas or semi-
empirical models in which the different components of the brightness temperature could be
differentiated. The brightness temperature of a soil covered by vegetation is usually
estimated as the contribution of three terms: (i) the radiation from the soil that is attenuated
by the overlying vegetation, (ii) the upward radiation from the vegetation, and (iii) the
downward radiation from the vegetation, reflected by the soil, and attenuated by the canopy
(Ulaby et al., 1986):
 1 − es,p  1  es,p
TB,model
p = 1 +
  1−
 L  (1 − ω ) Tveg + Ts , (13)
 L veg  veg  L veg
where Tveg and Ts are the physical temperatures of the vegetation and soil, respectively,
Lveg=exp(τ secθ ) is the attenuation due to the vegetation cover, τ=bVWC is the optical
thickness, b is the b-factor (Van de Griend & Wigneron, 2004), VWC is the vegetation water
content, and ω is the single scattering albedo. This formulation is known as the τ−ω model
and it is based on the single scattering approach proposed in Kirdiashev et al. (1979). The
optical depth is related to the vegetation density and the frequency. The single scattering
albedo describes the scattering of the emitted radiation by the vegetation, and is a function
of plant geometry.
3. Soil moisture retrieval algorithms

The brightness temperature of land covers is influenced by many variables, the most
important being soil moisture and temperature, and vegetation characteristics. The
challenge is to reconstruct the environmental parameters from the measured signal by using
a minimum of ancillary data. To do this, different soil moisture retrieval algorithms have
been developed. The first one is based on the experimental relationship between the
geophysical variables and the radiative transfer equation using a regression technique. This
approach has limited applicability, since often the regression is valid only for the test sites
where they were obtained. The second approach is based on the use of neural networks.
These algorithms have been used with satisfactory results in the retrieval of agricultural
parameters from radiometric data (Del Frate et al., 2003), but need a training phase that is
not always feasible. The third type of algorithms is widely used and is based on the
inversion of radiative transfer models. Obviously, this approach has also disadvantages,
since errors of the model lead to errors in the retrieval. The soil moisture models are used as
forward models, and the geophysical variables are retrieved by minimisation of a cost
function of the type (Wigneron et al., 2003; Pardé et al., 2004; Saleh et al., 2006)
F=∑
(T
Bh − TBmeas
h ) + (TBv − TBmeas
2
v )
2
+∑
(P
ini
N − PN )
2
. (14)
2
σT 2
σP
θ B N N
The simulated brightness temperature TBp is computed using (13), TBmeas

p
is the measured
brightness temperature, and PN is any of the parameters on which TBp depends. A first-guess
value of the parameter PN ( PNini ) with associated standard deviation σPN can be also
considered in the cost function.
4. Soil moisture field experiments

Space missions for Earth observation require a huge amount of previous work to validate the
existing science and technology. This is accomplished through airborne or ground-based field
experiments. Table 1 summarises some of the campaigns that have been conducted in the last
years over land surfaces using L-band radiometers. Most of these radiometers have been
designed within the context of preparatory activities previous to a space mission launch.
In the early 70’s NASA was getting ready to launch its first microwave radiometers for Earth
observations on the Nimbus 5 spacecraft and it needed to learn more about the capabilities of
these sensors for land observations. To do this several field campaigns were conducted in the
period 1971 to 1973. For the remote sensing of soil moisture irrigated agriculture test areas
were selected in the southwestern part of the United States. These experiments demonstrated
the sensitivity of microwave radiometers to surface soil moisture and that the longer
wavelength sensors worked best (Schmugge et al. 1974; Burke et al., 1979). These data were
used to study the effects of surface roughness (Choudhury et al., 1979), soil temperature
(Choudhury et al., 1982) and soil texture (Schmugge, 1980). These experiments defined the
basic sensitivity of microwave radiometers for soil moisture sensing and some of the factors
affecting this sensitivity. Since that time much has been done to refine and improve this
understanding. Some of these later experiments will be decribed in the following sections. A
description of the MOUSE 2004, T-REX 2004/2006, and SMOS REFLEX 2003/2006 campaigns
over land carried out by UPC and of their results is provided in the next sections. Radiometric
measurements were acquired using the UPC L-band Automatic Radiometer (LAURA), which
has a working frequency of 1.4135 GHz, the same as the SMOS payload MIRAS (Villarino,
2004). A view of the experiment sites and the LAURA radiometer is shown in Figure 2.
(a) T-REX 2006 (b) SMOS REFLEX 2003 (c) MOUSE 2004
Fig. 2. Picture of the MOUSE 2004, T-REX 2006, and SMOS REFLEX 2003 experiment sites
Year Campaign Site Surface Instrument

1971/1975 Phoenix,AZ Irrigated Multi-frequency
Imperial valley, agriculture Microwave
CA (USA) Radiometers
1976/1978 Joint Hand County, SD Dryland PMIS, MFMR
US/USSR (USA) agriculture
exchange
1985/1989 FIFE Konza Prairie, Short grass, PBMR
Kansas (USA) prairie
1990/1991 Monsoon 90 Walnut Gulch AZ Desert PBMR, ESTAR
(USA)
1992 HAPEX- Niger, West Africa Semi-arid PBMR
Sahel savannah
1992/1994 Washita Oklahoma (USA) Agricultural ESTAR, AIRSAR,
PBMR
1997/1999 SGP Oklahoma (USA) Agricultural ESTAR, PSR, PAL
1999 ESSC Reading (UK) Bare soil SWaMP-L
2001 INRA Avignon (FR) Bare, corn EMIRAD
2001 EuroSTARSS Toulouse (FR), Various STARSS
Utiel (SP)
2002 ETH Eschikon (CH) Bare ELBARA
2002 SMEX Iowa (USA) Various ESTAR, PSR,
PALS, AIRSAR
2003/2006 SMOS Utiel (SP) Vines LAURA
REFLEX
2003 ELBARA Wageningen (NL) Bare, grass ELBARA
2003
2003-2006 SMOSREX Toulouse (FR) Bare, fallow LEWIS
2003 SMEX OK,AL,GA(USA) Various 2DSTAR, PSR,
AIRSAR
2004 SMEX AR, NM (USA) Various 2DSTAR, PSR
2004 MOUSE Ispra (IT) Bare soil LAURA

2004/2006 T-REX Agramunt (SP) Bare soil LAURA
2004 ETH Eschikon (CH) Grass ELBARA
2004 Bray Les Landes (FR) Forest EMIRAD
2005 ETH Julich (GE) Forest ELBARA
2005 SMEX Iowa (USA) Various APMIR
2005 NAFE Upper Hunter Grass, forest EMIRAD, PLMR
(AU)
2006 TuRTLE El Brull (SP) Mountain LAURA
2006/2007 MELBEX Utiel (SP) Shrub, vines EMIRAD
2006 NAFE Yanco (AU) Various PSR, PALS
2007 CLASIC Oklahoma (USA) Various PSR, PALS
2008-2010 GRAJO Vadillo de la Grass, barley, LAURA,SMIGOL
Guareña (SP) bare soil
Table 1. Summary of L-band field campaigns over land performed since 1970. Acronyms
stand for USA (United States of America), UK (United Kingdom), FR (France), SP (Spain),
NL (The Netherlands), CH (Switzerland), GE (Germany), IT (Italy), and AU (Australia).
5. Texture effects on the L-band emissivity of bare soils: the MOUSE 2004
experiment
The Monitoring Underground Soil Experiment (MOUSE) 2004 was performed from day of
year (DoY) 159 to 184, 2004, at the Joint Research Centre (JRC) complex of the European
Union in Ispra, north of Italy (45°48N, 8°37E; 213 m altitude). The test lane consisted of six
6 m long by 6 m wide by 1 m depth bare plots of different soil types which are distributed
along a lane with West-East orientation. The LAURA radiometer, an infra-red video-camera,
and the control rack were installed on a metallic gantry guided by a rail on one side and
freely rolling on the other. A picture of the setup is shown in Figure 2(c). Soil temperature
was measured using 107 probes from Campbell Scientifc installed just below the surface,
and at 5 cm and 10 cm depth. An infra-red video camera was mounted by LAURA to
acquire soil surface temperature images concurrently with the radiometric measurements.
Most probably, satellite-borne radiometers will have to rely on this kind of auxiliary sensors to
estimate soil moisture from radiometric observations. On the other hand, soil moisture was
automatically acquired every 30 min by seven ML2x Thetaprobes per plot installed below the
surface (x2), and at 5 cm (x2), 10 cm (x2), and 15 cm (x1). Plots were irrigated three times
during the experiment, except RS which was kept dry and protected from water by a sliding
tent. Soils at the experiment site were loam (LO), natural sand (SA), sieved sand used in the
construction industry (RS), clay from rice fields (CL), a mixture of loam and of commercial
products for gardening (OR), and ferromagnetic crushed volcanic soil from Naples (FE).
Figure 3 represents the sensitivity of the H-pol emissivity to the percentage of field capacity,
which depends on soil texture. Measurements at 25° (blue), 35° (red), and 45° (green)
incidence angles have been plotted. LO, RS, and FE emissivity has a linear dependence on
field capacity. The R2 estimator at 35° for LO is 71%, for RS 70%, and for FE is 85%. Clay has
the worst linear fit, with R2 lower than 20%. Soil samples of each plot were acquired and
their dielectric constant measured at laboratory (Vall-llossera et al., 2005a).
Soil moisture was retrieved from radiometric measurements using the cost function in (14)
with different constraints. The minimum soil moisture RMSE obtained considering different
depths as ground-truth data and three different dielectric constant models (Wang &
Schmugge, 1980; Dobson et al., 1985; Vall-llossera et al., 2005a) is represented in Figure 4.
Best results for LO and FE are achieved for the soil moisture in the 0 to 5 cm layer, but few
discrepancies between RMSE values at different depths were found out. This is mainly due
to their almost linear soil moisture profile. Best results for SA were obtained for the soil
moisture in the 0-15 cm layer. As sandy soils dry faster, the penetration depth is higher. On
the other hand, as RS plot was kept dry throughout the experiment, no important
differences in the RMSE have been found in the comparison. Clay had a high water content
at 10 cm (>0.22 m3/m3) and hence the main signal contribution comes from the first 10 cm of
soil. Similar RMSE values have been found for OR considering either the 0 to 5 cm layer or
the 0 to 10 cm layer as ground-truth. Similarities are because the water content at 5 cm and
10 cm depth is almost the same during the experiment. On the other side, in Figure 4 it can
be noted the performance of each dielectric constant models for each soil type. Results for
Wang & Schmugge (1980) model are good for loamy and ferromagnetic soils, but their
performance for sandy soils is worse than Dobson et al. (1985) and Vall-llossera et al.
(2005a). The model in Dobson et al. (1985) offers best results for sandy soils (SA and RS)
which registered lower soil moisture values and dried faster than the others. The
expressions derived from UPC laboratory experiments show good performance for most of
the fields. These facts confirm that the selected dielectric constant model has an impact on soil
moisture estimation error and can increase it between 3 to 5% depending on the soil type.
(a) Loam (b) Sand (c) Reference sand
(d) Clay (e) Organic (f) Ferromagnetic

Fig. 3. Variation of the emissivity with the field capacity at H-pol. Colours indicate the
incidence angle: 25° (blue), 35° (red), and 45° (green).
(a) Wang & Schumugge (b) Dobson et al. (c) Lab. measurements
Fig. 4. Minimum root mean square error of the soil moisture retrieval. Results for a soil
dielectric constant model and different soil moisture ground-truth data are represented.
6. Emission of bare rough soils: the T-REX experiments

Several experiments have been performed in the past years to study the L-band emission
from bare soils, both at small and large scales. Although small scale measurements are not
representative of onboard sensor data, they can be useful to evaluate the various approaches
to the land emission modelling. The Terrain-Roughness EXperiments (T-REX) were carried
out in Agramunt, Spain (41°48N, 1°7W; 356 m altitude) from DoY 333 to DoY 337, 2004, and
then again from DoY 143 to DoY 160, 2006. The sites consisted of bare soil plots with four
different ploughing, but the same textural composition (21% clay, 21.3% sand). This section
presents some results from the T-REX 2006 experiment. A picture of the site has been shown
in Figure 2(a). Each parcel was 6 m x 15 m and the same ploughing was applied to two
different plots to observe spatial diversity effects. LAURA was installed on a trailer towed
by a car. Radiometric measurements were acquired at incidence angles from 40° to 65° in
T-REX 2006, in 5° steps. Every plough type was measured twice each day of experiment.
Parameter Plot label

(mm) A B C D
lc 126 100 190 96
σs 17 16.6 8 33
σλ 7 6.8 2.5 17
hair-soil 31 26 13 46
Table 2. Roughness characteristics of the T-REX 2006 ploughs
Roughness measurements were performed in the along and across directions of LAURA's
field of view. The standard deviation of height (σs) and the correlation length (lc) of the
ploughs at the T-REX 2006 site are summarised in Table 2. Two additional parameters have
also been included: (i) hair-soil, which is the transition layer thickness used to evaluate the
small-scale soil roughness in the model proposed in Schneeberger et al. (2004), Mätlzer
(2006, section 4.7) and (ii) σλ, which is the standard deviation of height assuming a
windowing of length λ. Soil temperature was measured at 0, 5, 10, 15, and 20 cm depth
using thermometers. On the other hand, soil moisture was measured with ML2x ThetaProbe
sensors at various positions within the radiometer's FOV. As the purpose of this experiment
was to study the impact of soil roughness, the site was not irrigated to keep as constant a value
of soil moisture as possible. The mean soil moisture during T-REX 2006 was of 4% to 6%.
Figure 5 shows the variation of the emissivity measured during the T-REX 2006 as a function
of the incidence angle. Colours indicate the polarisation, H-pol (red) and V-pol (blue), while
line styles distinguish the ploughing. While soil emission at H-pol decreases almost linearly
with increasing incidence angle (and the trend is more significative at incidence angles
above 55°), the emission at V-pol slightly increases with incidence angle until an inflexion
point at 55°, when it begins decreasing. Ploughs C and D, which correspond to the smoother
and rougher plots in the site, have the highest (lowest) range of variation of the emissivity.
The comparison between radiometric measurements and numerical simulations obtained
with a numerical approach by means of the integral equation method (IEM; Fung (1994)) are
shown in Figure 6. Icons stand for the mean value of LAURA measurements while error
bars indicate their standard deviation during the whole T-REX 2006 experiment. Although
IEM predicts an increase of the vertical emission with incidence angle, dry soil
measurements decrease as incidence angle increases. A good matching between simulations
and measurements is observed at horizontal polarisation. The effects of taking into account
the incoherent component are only relevant in the case of very rough ploughs (see Figure
6(c), plough D).
Fig. 5. Variation of the emissivity measured during the T-REX 2006 as a function of the
incidence angle for the four ploughs. Colours indicate the polarisation: H-pol (red) and V-
pol (blue). Line style and icons distinguish the ploughing.
a) Plough C b) Plough B c) Plough D

Fig. 6. Variation with time of the emissivity measured during the T-REX 2006 experiment as
a function of the incidence angle. The mean value of LAURA measurements (black icons)
and their standard deviation during the whole experiment (bars) have been represented
together with IEM simulations.
a) Plough C, V-pol b) Plough B, V-pol c) Plough D, V-pol
d) Plough C, H-pol e) Plough B, H-pol f) Plough D, H-pol

Fig. 7. Comparison between T-REX 2006 measurements (black icons) and the models in
Wang & Choudhury (1981), Mätzler (2006, section 4.7), Schneeberger et al. (2004), Mo &
Schmugge (1987), and Escorihuela et al. (2007). Solid lines indicate results using the
standard deviation of the surface height profile while dashed lines indicates results using
the λ-average standard deviation of height
T-REX 2006 measurements were also compared to the semi-empirical soil reflectivity models
in Wang & Choudhury (1981), Mo & Schmugge (1987), Escorihuela et al. (2007), and the air-
to-soil transition model proposed in Schneeberger et al. (2004). These models propose
different expressions for the effective roughness parameter hs in (12). Results are shown in
Figure 7. Upper plots correspond to vertical polarisation while bottom plots correspond to
horizontal polarisation. Colours indicate the soil reflectivity model, while line types indicate
whether lc and σs of the height profile (solid lines) or their equivalents taking into account
the λ-windowing (dashed lines) have been used. Both the model in Wang & Choudhury
(1981) using as inputs the λ -windowing soil roughness characteristics and the air-to-soil
transition model from Schneeberger et al. (2004) fit data. However, Schneeberger et al. (2004)
does not fit T-REX 2006 measurements of wet soil, which suggests that inhomogeneities in
soil moisture or the presence of dew make it difficult to properly characterise the upper soil
layer, which is needed for the air-to-soil transition model.
7. Radiometric observations of fully developed vines

Several papers are available that present experimental studies of brightness temperatures
from vegetation canopies (Jackson et al. 1982; Jackson et al., 1991; Ferrazzoli et al., 2002; Van
de Griend & Wigneron, 2004; Saleh et al., 2006). A look at the field experiments conducted
during the past decades over vegetated sites shows how they focused their attention on
crops such as alfalfa, wheat corn, or soybean, on grass, on forests, and on bushes, but few
studies had been conducted over vineyards. In this context, two ground-based field
experiments were performed to study the variations on the emissivity due to vines
development. The SMOS REFerence pixel L-band EXperiments (REFLEX) were carried out
in July 2003 and then again from July to November 2006 at two vineyards within the
València Anchor Station (VAS; Figure 2(b)), Spain (Vall-llossera et al., 2005b). The VAS has
been selected as a SMOS Cal/Val site because of its almost homogeneous land cover
(vineyards) and its size, which is approximately that of a SMOS pixel. The goal of the SMOS
REFLEX experiments was to assess the impact of grapevines on the radiometric emission
and on the sensitivity of the sensor to soil moisture. Results from the SMOS REFLEX 2003
experiment are presented hereafter.
(a) (b)
Fig. 8. Mean value and standard deviation of the emissivity and of the ground-truth soil
moisture as a function of (a) the incidence angle and for all azimuth angles, and (b) the
azimuth angle and for all incidence angles. DoY 185 (solid line, wet soil) and DoY 181
(dashed-dotted line, dry soil) have been represented. Blue and red indicate vertical and
horizontal polarisations, respectively.
During SMOS REFLEX 2003 radiometric measurements were acquired from DoY 181 to 191
at nine incidence angles (from 25° to 65° in 5° steps) and seven azimuth angles. Pictures at
every look direction were taken to estimate the percentage of FOV covered by plants. The
site was irrigated until saturation on DoY 182 and 185, and then was let to dry out. Even
though the field was irrigated until saturation, different soil moisture values were measured
depending on the location of the test point within the vineyard. This is most probably
because of surface inhomogeneities and variations on the compactness of the terrain due to
plough and roots distribution. This large spatial variability in soil moisture, which in this
case varies from 3% up to 35% in a 20 m x 20 m area, is one of the main problems soil
moisture retrieval algorithms must face since it complicates the comparison between
ground-truth and estimated ws at the plot scale. Volumetric soil moisture, soil temperature,
and soil roughness were measured, and plants were fully characterised.
Figure 8(a) represents the mean value and standard deviation of the emissivity and of the
ground-truth soil moisture as a function of the incidence angle and for all azimuth angles.
Measurements from DoY 185 (solid line) and DoY 181 (dashed-dotted line) have been
represented. The mean value of both polarisations seems independent of incidence angle for
completely dry soils, which is the case of DoY 181, being the std. of measurements lower
than 0.005 (0.01) in emissivity -1.5 K (3 K) in brightness temperature- for vertical (horizontal)
polarisations. In contrast, for wet soil an ascending trend is observed for V-pol emissivity up
to 45°, when V-pol emissivity seems to saturate, while H-pol emissivity decreases as
incidence angle increases. The std. of measurements for wet soils is 0.025 (0.03) for vertical
(horizontal) polarisations.
The mean value and standard deviation of the emissivity from DoY 181 and DoY 185 as a
function of the azimuth angle and for all incidence angles is represented in Figure 8(b). For
dry soils no dependence on the azimuth angle is observed, being the standard deviation of
measurements lower than 1.5 K at both polarisations. In the case of wet soils, and since the
contribution from vines is supposed to be the same for specular azimuth angles with respect
to zero, the inhomogeneity in soil moisture content, which is higher for positive azimuth
angles, leads to an asymmetry in the emission measurements of DoY 185 (solid lines), which
as expected are higher for the lowest soil moisture. For a given azimuth angle, the standard
deviation of the measured emissivity at V-pol goes from 0.014 (4 K) at 0° down to 0.008 (2 K)
at 65°, whereas the largest standard deviation at H-pol goes from 0.007 (2.6 K) at 0° up to
0.022 (7 K) at 45°.
Figure 9(a) shows the variation of the retrieved albedo with respect to the incidence angle
for all azimuth angles. Icons indicate the day of experiment, while the solid line is the mean
value of the whole experiment at each incidence angle. The albedo is smaller than 0.15,
having its maximum for an incidence angle of 35° and then decreasing as the incidence
angle increases. On the other hand, the opacity tends to increase with incidence angle as
shown in Figure 9(b) which is in accordance to the higher fraction of area covered by
vegetation as the incidence angle. Good agreement between measured and computed values
is appreciated for incidence angles below 55°. For 55° the algorithm does not always
converge and at 60° and 65°, the convergence is rarely achieved, pointing out a problem in
the emissivity model function for large incidence angles. This discrepancy can be due to the
fact that as the incidence angle increases, the signal from the soil suffers from more
attenuation and scattering in the canopy. The τ−ω model is a valid and simple forward
model to be used in optimal estimation approaches, but improvements are required if it is to
be used at high incidence angles.
(a) Albedo (b) Opacity (c) Soil moisture

Fig. 9. Dependence of the retrieved vegetation albedo (a), and opacity (b) on the incidence
angle. Icons represent the value estimated from LAURA measurements for each day of
experiment and all azimuth angles. The mean value is represented with a black solid line. (c)
Ground truth soil moisture vs. estimates from radiometric measurements.
In order to retrieve the soil water content from the LAURA's measurements, an iterative
algorithm, based on the least squares minimisation procedure, was applied. A better
performance was found out when the albedo, opacity and soil moisture were
simultaneously retrieved using (14). The first guesses for the opacity and the albedo are the
estimated values obtained in the iterative algorithm described before. The algorithm
converges for incidence angles equal to or smaller than 50°, only for some day of
experiments at 55°, and rarely for 60° or 65° due to the increasing discrepancy between the
model and the data as the incidence angle increases. When the incidence angle increases,
the canopy influence becomes more important (higher attenuation on the signal and higher
scattering contribution), so a higher order approximation for the direct model becomes
necessary. In Figure 9(c), the soil moisture estimated from LAURA data is compared to
ground-truth soil moisture data. The slope of the computed linear regression is 0.9 (close to
1), which shows a good agreement between both values. The error between estimations and
measurements is about 2.3%, which is smaller than the 4% accuracy required for SMOS.
8. Conclusions
The basics of land emission modelling have been summarised in this chapter, and an
overview of the field experiments over land carried out during the last three decades has
been given to present to the reader the state of the art. Some results of the recent SMOS
REFLEX 2003, MOUSE 2004, and T-REX 2006 experiments have been reported here. These
are, among others, part of the preparatory experiments for the SMOS mission, which in the
near future will provide soil moisture maps at global scale.
Soil texture influences the electromagnetic behaviour of soils since it affects their dielectric
constant. During MOUSE 2004, six types of bare at soils with different soil moisture contents
were measured at five incidence angles. As expected, variations on the moisture profile
were noted depending on texture. The emission of loams and sands showed a linear
dependence on the percentage of field capacity, especially at H-pol. In this case, the R-
squared estimator at 35° was found to be between 70% and 85%. On the contrary, clay
showed no linear trend which, however, can not be a general conclusion since clay data had
a large scattering.
The modified Wang & Choudhury (1981) soil emission model, with Qs=0 and n=0, was used
as forward model in the soil moisture retrieval algorithm. Best results were obtained when
no constraints were applied to soil moisture and temperature. The estimated soil moisture
was compared with the soil moisture measured at different depths. For loam and
ferromagnetic soils best results were achieved considering as ground-truth data the soil
moisture in the 0 to 5 cm layer, 0 to15 cm layer for sands, and 0 to 10 cm layer for clays. On
the other hand, when results for every dielectric constant model were compared, it was
found out that the Wang & Schmugge (1980) model had a better performance in the retrieval
algorithm of loam and ferromagnetic soils, while the Dobson et al. (1985) model offered best
results for sands.
Results from MOUSE 2004 suggest that the soil dielectric constant model should be selected
as a function of texture, since the root mean squared error can vary from 2% to 7%
depending on the soil texture, and that soil moisture estimates will be the averaged value in
a soil layer which will depend on texture.
Soil roughness has a strong impact on land brightness temperature. This effect is more
noticeable in the case of dry soils. Data from bare soils with a standard deviation of height
varying from 8 mm to 33 mm was acquired during the T-REX experiments, and was
compared to predictions from numerical and semi-empirical soil emission models. The
goodness of these models is of key importance for an accurate soil moisture estimation from
satellite data. As expected, dry soil emissivity at H-pol decreased almost linearly with
increasing incidence angle, being the decrement inversely proportional to soil roughness.
Although similar emission was measured at H-pol for all plough at low incidence angles, it
was noted that emission decreases slowly for rough soils than for smooth soils. Above an
incidence angle of 50°, the decreasing slope of H-pol emission was higher for all plough. The
sensitivity of V-pol to roughness is lower than that of H-pol. The trend of V-pol emission
was found to be different depending on whether the soil was wet or dry. When the soil is
wet the emission increased with the incidence angle, which is in accordance with most
model predictions. However, the trend for dry soils is decreasing with increasing incidence
angle and roughness. This may suggest the existence of a relationship between soil moisture
and the effective roughness.
The integral equation method (IEM) predicts an increase in V-pol emission with incidence
angle, which is not in accordance with measurements of dry soils. On the other hand, the
predicted descending trend at H-pol with increasing incidence angle is in accordance to
measurements, although IEM underestimates the soil emission both for dry and wet soils. In
a randomly ploughed field, without a significant tillage direction, the impact of choosing an
exponential or gaussian height probability density function in the IEM model is minimum,
whereas it was noted that the incoherent term of the reflectivity must be considered for
rough soils.
In general, all semi-empirical land emission models follow the trend of dry soils
measurements at H-pol, whereas discrepancies exist for wet soils. Neither the semi-
empirical models nor the IEM describe the trend of dry soils V-pol measurements, being the
error at this polarisation larger as the incidence angle increases. The lowest error between
predictions and measurements for H-pol was obtained using the Wang & Choudhury (1981)
model, but with roughness statistics averaged as a function of the measurements
wavelength. The dependence on the incidence angle proposed by Wang & Choudhury (a
squared cosine) had been discussed by other authors who considered it to be too much
strong at L-band. However, this simple formulation has been tested with T-REX data with
good results if the averaged standard deviation in transects equal to the wavelength is used.
The SMOS REFLEX 2003/2006 experiments site was in the Valencia Anchor Station, a
selected area for the SMOS calibration and validation activities. No previous studies over
vines were reported in the literature prior to these campaigns. In the first experiment, fully
developed vines were characterised during two weeks, while controlled irrigations
moistened the field. The second experiment was planned to monitor changes in the L-band
emission of vineyards during different stages of plants development.
Since vines do not have a predominant vertical nor horizontal structure, the opacity and the
albedo were found to be independent on the polarisation. The Wang & Choudhury (1981)
model was used as the soil emission forward model in the retrieval algorithm, while the
dielectric constant model was the one from Wang & Schmugge (1980). Good results were
obtained for incidence angles up to 55°, but the convergence of the algorithm was rarely
achieved above that value. The error between ground-truth and estimated soil moisture was
2.3%, better than the 4% required for SMOS. For incidence angles above 55° the convergence
of the algorithm was rarely achieved, probably due to the larger effect of the vegetation at
large incidence angles, not accurately described by the model. Higher order models should
be accounted for at these angles.
Although the radiometric behaviour varies from one canopy to the other, at a larger scale
such as that of SMOS it is very likely that the vegetation types can be averaged over the
footprint and that it is not necessary to account for an accurate distinction between canopies.
9. Acknowledgements
A. Monerris would like to thank Spanish Ministry of Science and Education projects
MIDAS-2 ESP2002-11648-E, MIDAS-3 ESP2004-00671, and MIDAS-4 ESP2005-06823-C05-02,
for funding the SMOS REFLEX, T-REX and MOUSE 2004 field experiments, and the FPU
AP2003-1567 grant. Special thanks are also given to M. Vall-llossera and A. Camps, from
UPC, for their valuable comments during the analysis of the experimental data.
10. References
Barré, H., Duesmann, B. & Kerr, Y.H. (2008). SMOS: The mission and the system. IEEE
Trans. Geosci. and Remote Sens., 46, 587-593
Behari, J. (2005). Microwave dielectric behaviour of wet soils (Remote Sensing and Digital Image
Processing), Springer, ISBN-10:1402032714 , The Netherlands
Brady, N.C. & Weil, R.R. (2003). Elements of the nature and properties of soils, 2nd ed., Prentice
Hall, ISBN-10:013048038X , NJ, USA
Burke, W. J., Schmugge, T. J. & Paris, J.F. (1979). Comparison of 2.8 and 21 cm microwave
radiometer observations over soils with emission model calculations, J. Geophys.
Res., 84, 287-294.
Chanzy, A., Raju, S. & Wigneron, J.-P. (1997). Estimation of soil microwave effective
temperature at L and C bands. IEEE Trans. Geosci. and Remote Sens., 35, 3, 570-580
Choudhury, B.J., Schmugge, T.J., Chang, A.T.C. & Newton, R.W. (1979). Effect of surface
roughness on the microwave emission of soils. J. Geophys. Res., 84, 5699-5705
Choudhury, B.J., Schmugge, T.J. & Mo, T. (1982). A simple parameterization of effective soil
temperature for microwave emission. J. Geophys. Res., Vol., No., (month and year of
the edition) 1301-1304
Chukhlantsev, A.A. (2006). Microwave radiometry of vegetation canopies, Springer, ISBN-10 1-
4020-4681-2, The Netherlands
Del Frate, F., Ferrazzoli, P. & Schiavon, G. (2003). Retrieving soil moisture and agricultural
variables by microwave radiometry using neural networks. Rem. Sens. Env., 84, 2,
174-183
Dobson, M. C., Ulaby, F. T., Hallikainen, M. T. & Elrayes, M. A. (1985). Microwave dielectric
behaviour of wet soils 2. Dielectric mixing models. IEEE Trans. Geosci. and Remote
Sens., 23, 35-46
Eagleman, J.R. & Lin, W.C. (1976). Remote sensing of soil moisture by a 21 cm passive
radiometer. J. Geophys. Res., 81, 3660-3666
Entekhabi, D., Njoku, E., Houser, P., Spencer, M., Doiron, T., Smith, J., Girard, R., Belair, S.,
Crow, W. ,Jackson, T., Kerr, Y., Kimball, J., Koster, R., McDonald, K., O'Neill, P.,
Pultz, T., Runnig, S., Shi, J., Wood, E. & van Zyl, J. (2004). The Hydrosphere State
(HYDROS) mission concept: An Earth system pathfinder for global mapping of soil
moisture and land freeze/thaw. IEEE Trans. Geosci. and Remote Sens., 42, 2184-2195
Escorihuela, M.J., Kerr, Y.H., de Rosnay, P., Wigneron, J.-P., Calvet, J.-C. & Lemaitre, F.
(2007). A simple model of the bare soil microwave emission at L-band, IEEE Trans.
Geosci. and Remote Sens., 45, 1978-1987
Ferrazzoli, P., Guerriero, L. & Wigneron, J.-P. (2002). Simulating L-band emission of forests
in view of future satellite applications. IEEE Trans. Geosci. and Remote Sens., 40,
2700-2708
Fung, A.K. (1994). Microwave Scattering and Emission. Models and their applications, Artech
House, ISBN-10:0890065233, Norwood, MA.
Hallikainen, M.T., Ulaby, F.T., Dobson, M.C., El-Rayes, M.A.& Wu, L. (1985). Microwave
dielectric behaviour of wet soil - Part 1: Empirical models and experimental
observations. IEEE Trans. Geosci. and Remote Sens., GE-23, 25-34
Hipp, J.E. (1974). Soil electromagnetic parameters as a function of frequency, soil density
and soil moisture, Proceedings of IEEE, 98-103
Holmes, T., de Rosnay, P., de Jeu, R., Wigneron, J.-P., Kerr, Y., Calvet, J.-C., Escorihuela,
M.J., Saleh, K. & Lemaitre, F. (2006). A new parameterization of the effective
temperature for L- band radiometry. Geophys. Res. Lett., 33, 7405
Jackson, T.J., Schmugge, T.J. & Wang, J.R. (1982). Passive microwave sensing of soil moisture
under vegetation canopies. Water Resources Research, 18, 1137-1142
Jackson, T.J. & Schmugge, T.J. (1989). Passive microwave remote sensing system for soil
moisture: Some supporting research. IEEE Trans. Geosci. and Remote Sens., 27, 2, 225-
235
Jackson, T. J. & Schmugge, T. J. (1991). Vegetation effects on the microwave emission from
soils. Remote Sens. Environ., 36, 203-212
Kerr, Y.H., Waldteufel, P., Wigneron, J.-P., Martinuzzi, J.M., Font, J. & Berger, M. (2001). Soil
moisture retrieval from space: the Soil Moisture and Ocean Salinity (SMOS)
mission. IEEE Trans. Geosci. and Remote Sens., 39, 1729-1735
Kirdiashev, K.P., Chukhlantsev, A.A., & Shutko, A.M. (1979). Microwave radiation of the
Earth's surface in the presence of vegetation cover. Radio Eng. Electron. Phys., 2, 37-
56.
Koster, R.D., Dirmeyer, P.A., Guo, Z., Bonan, G., Chan, E., et al. (2004). Regions of strong
coupling between soil moisture and precipitation. Science, 305, 5687, 1138-1140
Mätlzler, C. & Standley, A. (2000). Relief effects for passive microwave remote sensing, J.
Geophys. Res., 21, 2403-2412
Mätzler, C. (2006). Thermal Microwave Radiation: Applications for Remote Sensing, IET Electro-
magnetic Waves Series 52, ISBN-10:0863415733, London, UK
McMullan, K.D., Brown, M.A., Martín-Neira, M., Rits, W., Martí, J. & Lemanczyk, J. (2008).
SMOS: The payload. IEEE Trans. Geosci. and Remote Sens., 46, 3, 594-605
Mironov, V.L., Dobson, M.C., Kaupp, V.H., Komarov, S.A. & Kleshchenko, V.N. (2004).
Generalized refractive mixing dielectric model for moist soils. IEEE Trans. Geosci.
and Remote Sens., 42, 4, 773-785
Mo, T. & Schmugge, T. (1987). A parameterization ot the effect of surface roughness on
microwave emission, IEEE Trans. Geosci. and Remote Sens., 25, 47-54
Monerris, A., Benedicto, P., Vall-llossera, M., Santanach, E., Piles, M. & Prehn, R. (2008).
Assessment of the topography impact on microwave radiometry at L-band. J.
Geophys. Res., 113(B12202)
Pardé, M., Wigneron, J.-P., Waldteufel, P., Kerr, Y.H., Chanzy, A., Sobjaerg, S.S. & Skou, N.
(2004). N-parameter retrievals from L-band microwave observations acquired over
a variety of crop fields. IEEE Trans. Geosci. and Remote Sens., 42, 6, 1168-1178
Roth, C.H., Malicki, M.A. & Plagge, R. (1992). Empirical evaluation of the relationship
between soil dielectric constant and volumetric water content as the basis for
calibrating soil moisture measurements by TDR. J. of Soil Sci., 43, 1-3
Saleh, K., Wigneron, J.-P., de Rosnay, P., Calvet, J.-C., Escorihuela, M.J., Kerr, Y. &
Waldteufel, P. (2006). Impact of rain interception by vegetation and mulch on the L-
band emission of natural grass. Remote Sens. Environ., 101, 127-139
Schmugge T., P. Gloersen, T. T. Wilheit, and F. Geiger, 1974. Remote Sensing of Soil
Moisture with Microwave Radiometers, NASA/GSFC X-652-72-305, Jour. Geophys.
Res., 79(2), pp. 317-323.
Schmugge, T. J. 1980. Effect of Texture on the Microwave Emission from Soils, NASA Tech.
Memo. 80632, IEEE Trans. on Geoscience and Remote Sensing, GE-18, pp. 353-361.
Schneeberger, K., Schwank, M., Stamm, Ch., de Rosnay, P., Mätzler, Ch. & Flühler, H.
(2004). Topsoil structure influencing soil water retrieval by microwave radiometry.
Vadose Zone Journal, 3, 1169-1179
Schwank, M., Stähli, M., Wydler, H., Leuenberger, J., Mätzler, C. & Flühler, H. (2004).
Microwave L-band emission of freezing soil, IEEE Trans. Geosci. and Remote Sens.,
42, 1252-1261
Talone, M., Camps, A., Monerris, A., Vall-llossera, M., Ferrazzoli, P. & Piles, M. (2007).
Surface topography and mixed pixel effects on the simulated L-band brightness
temperatures, IEEE Trans. Geosci. and Remote Sens.,42, 786-794
Topp, G.C., Davies, J.L. & Annan, A.P. (1980). Electromagnetic determination of soil water
content: Measurements in coaxial transmission lines. Water Resour. Res., 16, 574-582
Ulaby, F.T., Moore, R.K. & Fung, A.K. (1986). Microwave Remote Sensing: Active and Passiv,
vol. II, Radar remote sensing and surface scattering and emission theory, Addison-
Wesley, ISBN-10:0890061912, NY
Van de Griend, A. A. & Wigneron, J. P. (2004). The b-factor as a function of frequency and
canopy type at H-polarization. IEEE Trans. Geosci. and Remote Sens., 42, 4, 786-794
Vall-llossera, M., Cardona, M., Blanch, S., Camps, A., Monerris, A., Corbella, I., Torres, F. &
Duffo, N. (2005a). L-band dielectric properties of different soil types collected
during the MOUSE 2004 field experiment. Proceedings IGARSS, 2, 1109-1112
Vall-llossera, M., Camps, A., Corbella, I., Torres, F., Duffo, N., Monerris, A., Sabia, R., Selva,
D., Antolín, C., López-Baeza, E., Ferrer, J.F. & Saleh, K. (2005b). SMOS REFLEX
2003: L-band emissivity characterization of vineyards. IEEE Trans. Geosci. and
Remote Sens., 43, 5, 973-982
Villarino, R. (2004). Empirical Determination of the Sea Surface Emissivity at L-band: A
contribution to ESA's SMOS Earth Explorer Mission. Ph.D. thesis, RSLab, UPC.
[Online] http://www.tesisenxarxa.net/TDX-1019104-170028/index.html
Wang, J.R. & Schmugge, T.J. (1980). An empirical model for the complex dielectric
permittivity of soils as a function of water content. IEEE Trans. Geosci. and Remote
Sens., 18, 288-295
Wang, J.R. (1983). Passive microwave sensing of soil moisture content: The effects of soil
bulk density and surface roughness. Remote Sens. Environ., 13, 329-344
Wang, J.R. & Choudhury, B.J. (1981). Remote sensing of soil moisture content over bare field
at 1.4 GHz frequency. J. Geophys. Res., 86, 5277-5282
Wegmüller, U. & Mätzler, C. (1999). Rough bare soil reflectivity model. IEEE Trans. Geosci.
and Remote Sens., 37, 1391-1395
Wigneron, J., Laguerre, L. & Kerr, Y.H. (2001). A simple parameterization of the L-band
microwave emission from rough agricultural soils. IEEE Trans. Geosci. and Remote
Sens., 39, 1697-1707
Wigneron, J.-P., Calvet, J.-C., Pellarin, T., de Griend, A.A. Van, Berger, M. & Ferrazzoli, P.
(2003). Retrieving near-surface soil moisture from microwave radiometric
observations: current status and future plans. Remote Sens. Environ., 85, 489-506
Wigneron, J.-P., Chanzy, A., de Rosnay, P., Räudiger, C. & Calvet, J.-C. (2008). Estimating
the effective soil temperature at L-band as a function of soil properties. IEEE Trans.
Geosci. And Remote Sens., 46, 3, 797-807
AggieAir: Towards Low-cost Cooperative
Multispectral Remote Sensing Using Small Unmanned Aircraft Systems 463
1
22

Multispectral Remote Sensing Using Small
Unmanned Aircraft Systems
a Haiyang Chao, a,b Austin M. Jensen, a Yiding Han, a YangQuan Chen
and b Mac McKee
a Center for Self Organizing and Intelligent Systems (CSOIS),Utah State University
haiyang.chao@aggiemail.usu.edu, yqchen@ieee.org
b Utah Water Research Laboratory (UWRL), Utah State University
austin.jensen@aggiemail.usu.edu
United States of America
1. Introduction
This chapter focuses on using small low-cost unmanned aircraft systems (UAS) for remote
sensing of meteorological and related conditions over agricultural fields or environmentally
important land areas. Small UAS, including unmanned aerial vehicle (UAV) and ground
devices, have many advantages in remote sensing applications over traditional aircraft- or
satellite-based platforms or ground-based probes for many applications. This is because small
UAVs are easy to manipulate, cheap to maintain, and remove the need for human pilots to per-
form tedious or dangerous jobs. Multiple small UAVs can be flown in a group and complete
challenging tasks such as real-time mapping of large-scale agriculture areas.
The purpose of remote sensing is to acquire information about the Earth’s surface without
coming into contact with it. One objective of remote sensing is to characterize the electromag-
netic radiation emitted by objects (James, 2006). Typical divisions of the electromagnetic spec-
trum include the visible light band (380 − 720nm), near infrared (NIR) band (0.72 − 1.30µm),
and mid-infrared (MIR) band (1.30 − 3.00µm). Band-reconfigurable imagers can generate sev-
eral images from different bands ranging from visible spectra to infra-red or thermal based for
various applications. The advantage of an ability to examine different bands is that different
combinations of spectral bands can have different purposes. For example, the combination
of red-infrared can be used to detect vegetation and camouflage and the combination of red
slope can be used to estimate the percent of vegetation cover (Johnson et al., 2004). Different
bands of images acquired remotely through UAS could be used in scenarios like water man-
agement and irrigation control. In fact, it is difficult to sense and estimate the state of water
systems because most water systems are large-scale and need monitoring of many factors in-
cluding the quality, quantity, and location of water, soil and vegetations. For the mission of
accurate sensing of a water system, ground probe stations are expensive to build and can only
provide data with very limited sensing range (at specific positions and second level temporal
resolution). Satellite photos can cover a large area, but have a low resolution and a slow up-
date rate (30-250 meter or lower spatial resolution and week level temporal resolution). Small
UAVs cost less money but can provide more accurate information (meter or centimeter spatial
resolution and hour-level temporal resolution) from low altitudes with less interference from
clouds. Small UAVs combined with ground and orbital sensors can even form a multi-scale
remote sensing system.
UAVs equipped with imagers have been used in several agricultural remote sensing applica-
tions for collecting aerial images. High resolution red-green-blue (RGB) aerial photos can be
used to determine the best harvest time of wine grapes [Johnson et al. 2003]. Multispectral
images are also shown to be potentially useful for monitoring the ripeness of coffee [John-
son et al. 2004]. Water management is still a new area for UAVs, but it has more exact re-
quirements than other remote sensing applications: real-time management of water systems
requires more and more precise information on water, soil and plant conditions, for example,
than most surveillance applications. Most current UAV remote sensing applications use large
and expensive UAVs with heavy cameras (in the range of a kilogram). Images from reconfig-
urable bands taken simultaneously can increase the final information content of the imagery
and significantly improve the flexibility of the remote sensing process.
Motivated by the above remote sensing problem, AggieAir, a band-configurable small UAS-
based remote sensing system has been developed in steps at Center for Self Organizing and
Intelligent Systems (CSOIS) together with Utah Water Research Lab (UWRL), Utah State Uni-
versity. The objective of this chapter is to present an overview of the ongoing research on this
topic.
The chapter first presents a brief overview of the unmanned aircraft systems focusing on the
base of the whole system: autopilots. The common UAS structure is introduced. The hard-
ware and software aspects of the autopilot control system are then explained. Different types
of available sensor sets and autopilot control techniques are summarized. Several typical
commercial off-the-shelf and open source autopilot packages are compared in detail, includ-
ing the Kestrel autopilot from Procerus, Piccolo autopilot from CloudCap, and the Paparazzi
open source autopilot etc.
The chapter then introduces AggieAir, a small and low-cost UAS for remote sensing. Ag-
gieAir comprises of a flying-wing airframe as the test bed, the OSAM-Paparazzi autopilot
for autonomous navigation, the Ghost Foto image system for image capture, the Paparazzi
ground control station (GCS) for real time monitoring, and the gRAID software for image
processing. AggieAir is fully autonomous, easy to manipulate, and independent of a runway.
AggieAir can carry embedded cameras with different wavelength bands, which are low-cost
but have high spatial resolution. These imagers mounted on UAVs can form a camera array
to perform multi-spectral imaging with reconfigurable bands, depending on the objectives of
the mission. Developments of essential subsystems, such as the UAV autopilot, imaging pay-
load subsystem, and image processing subsystem, are introduced in detail together with some
experimental results to show the orthorectification accuracy.
Several typical example missions together with real UAV flight test results are focused in Sec.3
including land survey, water area survey, riparian applications and remote data collection.
Aerial images and stitched maps showed the effectiveness of the whole system. The future di-
rection is more accurate orthorectification method and band-reconfigurable multi-UAV-based
cooperative remote sensing for real-time water management and distributed irrigation con-
trol.
2. Small UAS Overview

In this paper, the acronym UAV (Unmanned Aerial Vehicle) is used to represent a power-
driven, reusable airplane operated without a human pilot on board. The UAS (Unmanned
Aircraft System) is defined as an UAV and its associated elements which may include ground
control stations, data communication links, support equipment, payloads, flight termination
systems, and launch/recovery equipments (Tarbert et al., 2009). Small UAS (sUAS) could be
categorized into five groups based on gross take-off weight by the sUAS aviation rule making
committee, as shown in Tab. 1. Group i include those constructed in a frangible manner that
would minimize injury and damages if there is a collision, compared with group ii.
Group Gross Take-off Weight

i ≤ 4.4 lbs or 2 kgs
ii ≤ 4.4 lbs or 2 kgs
iii ≤ 19.8 lbs or 9 kgs
iv ≤ 55 lbs or 25 kgs
v lighter than air (LTA) only
Table 1. Small UAS Categories
The topic of small UAS is quite active in the past few years (Chao et al., 2009). A lot of small
fixed-wing or rotary-wing UAVs are flying in the air under the guidance from the autopilot
systems for different applications like forest first monitoring, coffee field survey, search and
rescue, etc. A typical UAS includes:
(1) Autopilot: an autopilot is a MEMS system used to guide the UAV without assistance
from human operators, consisting of both hardware and its supporting software. The
autopilot is the base for all the other functions of the UAS platform.
(2) Airframe: the airframe is where all the other devices are mounted including the
frame body, which could be made from wood, foam or composite materials. The air-
frame also includes the flight control surfaces, which could be a combination of either
aileron/elevator/rudder, or elevator/rudder, or elevons.
(3) Payload: the payload of UAS could be different bands of cameras, or other emission
devices like Lidar mostly for intelligence, surveillance, and reconnaissance uses.
(4) Communication system: most UAS have more than one wireless link supported. For
example, RC link for safety pilot, wifi link for large data sharing, and wireless serial
modem.
(5) Ground control station: ground control station is used for real-time flight status moni-
toring and flight plan changing.
(6) Launch and recovery devices: some UAS may need special launching devices like a
hydraulic launcher or landing devices like a net.
The whole UAS structure is shown in Fig 1. The minimal UAS onboard system requires the
airframe for bedding all the devices, the autopilot for sensing and navigation, the basic imag-
ing payload for aerial images, and the communication systems for data link with the ground.
The left section concentrates on the autopilot overview since it is the base for the UAS plat-
form and it also needs to provide accurate orientation and position data for each data set the
UAS collects.
Autopilot systems are now widely used in modern aircrafts and ships. The objective of UAV
autopilot systems is to consistently guide UAVs to follow reference paths, or navigate through
some waypoints. A powerful UAV autopilot system can guide UAVs in all stages including
Fig. 1. UAS Structure.
take-off, ascent, descent, trajectory following, and landing. The autopilot needs also to com-
municate with ground station for control mode switch, to receive broadcast from GPS satellite
for position updates and to send out control inputs to the servo motors on UAVs.
An UAV autopilot system is a close-loop control system with two fundamental functions: state
estimation and control signal generation based on the reference paths and the current states.
The most common state observer is the inertial measurement unit (IMU) including gyros,
accelerometers, and magnetic sensors. There are also other attitude determination devices
available like infrared or vision based ones. The sensor readings combined with the GPS
information can be passed to a filter to generate the estimates of the current states for later
control uses. Based on different control strategies, the UAV autopilots can be categorized to
PID based autopilots, fuzzy based autopilots, neutral network (NN) based autopilots, etc. A
typical off-the-shelf UAV autopilot system comprises of the GPS receiver, the IMU, and the
onboard processor (state estimator and flight controller) as illustrated in Fig. 2.
Fig. 2. Functional Structure of the UAV Autopilot.
Due to the limited size and payload of the small UAVs, the physical features like size, weight
and power consumption are the primary issues that the autopilot must take into consider-
ation. A good autopilot should be small, light and have a long endurance life. It is not so
hard to design the hardware to fulfill the autopilot requirements. The current bottleneck for
autopilot systems lies more in the software side.
2.1 Autopilot Hardware

A minimal autopilot system includes sensor packages for state determination and onboard
processors for estimation & control uses, and peripheral circuits for servo & modem commu-
nications. Due to the physical limitations of small UAVs, the autopilot hardware needs to be of
small sizes, light weights and low power consumptions. The accurate flight control of UAVs
demands a precise observation of the UAV attitude in the air. Moreover, the sensor packages
should also guarantee a good performance, especially in a mobile and temperature-varying
environment.
2.1.1 MEMS Inertial Sensors

Inertial sensors are used to measure the 3-D position and attitude information in the inertial
frame. The current MEMS technology makes it possible to use tiny and light sensors on small
or micro UAVs. Available MEMS inertial sensors include:
(1) GPS receiver: to measure the positions (pn , pe , h) and ground velocities (vn , ve , vd ).
(2) Rate or gyro: to measure the angular rates (p, q, r).
(3) Acceleration: to measure the accelerations (a x , ay , az ).
(4) Magnetic: to measure the magnetic field for the heading correction (ψ).
(5) Pressure: to measure the air speed (the relative pressure) and the altitude (h).
(6) Ultrasonic sensor or SONAR: to measure the relative height above the ground.
(7) Infrared sensor: to measure the attitude angles (φ,θ).
(8) RGB camera or other image sensors: to replace one or several of the above sensors.
GPS plays an indispensable role in the autonomous control of UAVs because it provides an
absolute position measurement. A known bounded error between GPS measurement and the
real position can be guaranteed as long as there is a valid 3-D lock. For instance, u-blox 5
GPS receiver could achieve a three meter 3-D accuracy (PACC) in the best case for civilian
applications in the United States. There are also differential GPS units which could achieve
centimeter level accuracy. The disadvantage of GPS is its vulnerability to weather factors and
its relatively low updating frequency (commonly 4Hz), which may not be enough for flight
control applications.
2.1.2 Possible Sensor Configurations

Given all the above inertial sensors, several sensor combinations could be chosen for different
types of UAVs to achieve the basic autonomous waypoints navigation task. Most current
outdoor UAVs have GPS receivers onboard to provide the absolute position feedback. The
main difference is the attitude measurement solution, which could be inertial measurement
unit (IMU), infrared sensor or image sensor etc.
2.1.2.1 Inertial Measurement Unit (IMU)

A typical IMU includes 3-axis gyro rate and acceleration sensors, which could be filtered to
generate an estimation of the attitude (φ, θ, ψ). A straightforward sensor solution for small
UAVs is to use the micro IGS, which can provide a complete set of sensor readings. Microstrain
Gx2 is this kind of micro IMU with an update rate up to 100 Hz for inertial sensors. It has 3-
axis magnetic, gyro and acceleration sensors (Microstrain Inc., Accessed 2008).
2.1.2.2 Infrared Sensor

Another solution for attitude sensing is using infrared thermopiles. The basic idea of infrared
attitude sensor is to measure the heat difference between two sensors on one axis to determine
the angle of the UAV because the Earth emits more IR than the sky. Paparazzi Open Source
Autopilot group used this kind of infrared sensors as their primary attitude sensor (Paparazzi
Forum, Accessed 2008) (Egan, 2006). The infrared sensors can be used for UAV stabilization
and RC plane training since it can work as a leveler. However, it is not that accurate for later
georeferecing.
2.1.2.3 Vision Sensor

Vision sensor could also be used to estimate the attitude by itself or combined with other in-
ertial measurements (Roberts et al., 2005). The pseudo roll and pitch can be decided from the
onboard video or image streams (Damien et al., 2007). Experiments on vision only based nav-
igation and obstacle avoidance have been achieved on small rotary wing UAVs (Calise et al.,
2003). In addition, vision based navigation has potentials to replace the GPS in providing po-
sition measurements especially in task oriented and feature based applications. Vision based
navigation for small fixed wing UAVs is still an undergoing topic and a lot of work are still
needed for mature commercial autopilots.
2.2 Autopilot Software

All the inertial measurements from sensors will be sent to the onboard processor for further
filter and control processing. Autopilot could subscribe services from the available sensors
based on different control objectives.
2.2.1 Autopilot Control Objectives

Most UAVs can be treated as mobile platforms for all kinds of sensors. The basic UAV way-
points tracking task could be decomposed into several subtasks including:
(1) Pitch attitude hold.
(2) Altitude hold.
(3) Speed hold.
(4) Automatic take-off and landing.
(5) Roll-Angle hold.
(6) Turn coordination.
(7) Heading hold.
There are two basic controllers for the UAV flight control: altitude controller, velocity and
heading controller. Altitude controller is to drive the UAV to fly at a desired altitude including
the landing and take-off stages. The heading and velocity controller is to guide the UAV to
fly through the desired waypoints. Most commercial autopilots use PID controllers, and the
control parameters could be tuned off-line first and re-tuned during the flight.
2.3 Typical Autopilots for Small UAVs

In this section, several available autopilots, including both commercial and open source ones,
are introduced and compared in terms of sensor configurations, state estimations and con-
troller strengths. Most commercial UAV autopilots have sensors, processors and peripheral
circuits integrated into one single board to account for size and weight constraints. The advan-
tage of the open source autopilots is its flexibility in both hardware and software. Researchers
can easily modify the autopilot based on their own special requirements.
2.3.1 Procerus Kestrel Autopilot

Procerus Kestrel Autopilot is specially designed for small or micro UAVs weighing only 16.7
grams (modem and GPS receiver not included), shown in Fig. 3. The specifications are shown
in Table 2. Kestrel 2.2 includes a complete inertial sensor set including: 3-axis accelerometers,
3-axis angular rate sensors, 2-axis magnetometers, one static pressure sensor (altitude) and
one dynamic pressure sensor (airspeed). With the special temperature compensations for sen-
sors, it can estimate the UAV attitude (φ and θ) and the wind speed pretty accurately (Beard
et al., 2005).
Fig. 3. Procerus Kestrel Autopilot (Beard et al., 2005).
Kestrel has a 29MHz Rabbit 3000 onboard processor with 512K RAM for onboard data log-
ging. It has the built-in ability for autonomous take-off and landing, waypoint navigation,
speed and altitude hold. The flight control algorithm is based on the traditional PID control.
The autopilot has elevator controller, throttle controller and aileron controller separately. El-
evator control is used for longitude and airspeed stability of the UAV. Throttle control is for
controlling airspeed during level flight. Aileron control is used for lateral stability of the UAV
(Beard et al., 2005). Procerus provides in-flight PID gain tuning with real-time performance
graph. The preflight sensor checking and failsafe protections are also integrated to the autopi-
lot software package. Multiple UAV functions are also supported by Kestrel.
2.3.2 Cloud Cap Piccolo

Piccolo family of UAV autopilots from Cloud Cap Company provide several packages for
different applications. PiccoloPlus is a full featured autopilot for fixed-wing UAVs. Piccolo I I
is an autopilot with user payload interface added. Piccolo LT is a size optimized one for small
electric UAVs as shown in Fig. 4. It includes inertial and air data sensors, GPS, processing,
RF data link, and flight termination, all in a shielded enclosure (CloudCap Inc., Accessed
2008). The sensor package includes three gyros and accelerometers, one dynamic pressure
sensor and one barometric pressure sensor. Piccolo has special sensor configuration sections
to correct errors like IMU to GPS antenna offset, avionics orientation with respect to the UAV
body frame.
Piccolo LT has a 40M Hz MPC555 onboard microcontroller. Piccolo provides a universal con-
troller with different user configurations including legacy fixed wing controller, neutral net
helicopter controller, fixed wing generation 2 controller, and PID helicopter controller. Fixed
Fig. 4. PICCOLO LT Autopilot (CloudCap Inc., Accessed 2008).
wing generation 2 controller is the most commonly used flight controller for conventional
fixed wing UAVs. It includes support for altitude, bank, flaps, heading & vertical rate hold,
and auto take-off and landing. Piccolo autopilot supports one ground station controlling mul-
tiple autopilots and it also has a hardware-in-the-loop simulation.
2.3.3 Paparazzi Autopilot

Paprazzi autopilot is a pretty popular project first developed by researchers from ENAC uni-
versity, France. Infrared sensors combined with GPS are used as the default sensing unit.
Although Infrared sensors can only provide a rough estimation of the attitude, it is enough
for a steady flight control once tuned well. Tiny 13 is the autopilot hardware with the GPS
receiver integrated, shown in Fig. 5. Paparazzi also has Tiny Twog autopilot with two open
serial ports, which could be used to connect with IMU and modem. One Kalman filter is
running on the autopilot to provide a faster position estimation based on GPS updates.
(a) Tiny 13 (b) Paparazzi GCS
Fig. 5. Paparazzi Autopilot System(Brisset et al., 2006).
Paparazzi uses LPC 2148 ARM7 chip as the central processor. For the software, it could
achieve waypoints tracking, auto-takeoff & landing, and altitude hold. The flight controller
could also be configured if gyro rate is used for roll and pitch tracking control especially for
micro UAVs. However, paparazzi doesnt have a good speed hold and changing function cur-
rently since no air speed sensor reading is considered in the controller part. Paparazzi is also a
truly autonomous autopilot without any rely on the ground control station (GCS). It also has
a lot of safety considerations in conditions like RC signal lost, out of predefined range, GPS
lost, etc.
2.3.4 Specification Comparisons

The physical specifications of the autopilots are important since small UAVs demand as fewer
space, payload and power as possible. The size, weight, and power consumption issues are
shown in Table 2. Both the Crossbow MNAV and Procerus Kestrel have a bias compensa-
tion to correct the inertial sensor measurement under different temperatures. The functional
specifications of these three typical autopilot are listed in detail in Table 3.
Size Weight (g) Power Price DC In CPU Memory

(cm) w/o radio Consumption (k USD) (V) (K)
Kestrel 2.2 5.08*3.5*1.2 16.7 500mA 5 6-16.5 29MHz 512
(3.3 or 5V)
Piccolo LT 11.94*5.72*1.78 45 5W - 4.8-24 40MHz 448
(w.modem)
Pprz Twog 4.02*3.05*1 8 N/A 0.125 6.1-18 32 bit ARM7 512
Table 2. Comparison of Physical Specifications of Autopilots (Chao et al., 2009)
Kestrel Picocolo LT Paparazzi

Waypints Navigation Y Y Y
Auto-takeoff & landing Y Y Y
Altitude Hold Y Y Y
Air Speed Hold Y Y N
Multi-UAV Support Y Y Y
Attitude Control Loop - - 20/60 Hz
Servo Control Rate - - 20/60 Hz
Telemetry Rate - 25Hz or faster Configurable
Onboard Log Rate ≤100Hz - N
Table 3. Comparison of Autopilot Functions (Chao et al., 2009)
3. AggieAir UAS Platform

Although most current autopilot systems for UAVs have the ability to autonomously navigate
through waypoints, it is actually not enough for the real remote sensing applications since the
end users need aerial images with certain spatial and temporal resolution requirements from
different bands of cameras. More importantly, most civilian remote sensing users want the
UAV platform to be inexpensive. AggieAir UAS platform is developed considering all these
remote sensing requirements. AggieAir is a small and low-cost UAV remote sensing platform,
which includes the flying-wing airframe, the OSAM-Paparazzi autopilot, the GhostFoto im-
age capture subsystem, the Paparazzi ground control station (GCS), and the gRAID software
for aerial image processing. All the subsystems are introduced in detail in this section together
with a method to help improve the orthorectification accuracy and calibrate the aircraft sen-
sors using ground references.
3.1 Remote Sensing Requirements

Let Ω ⊂ R2 be a polytope including the interior, which can be either convex or nonconvex. A
series of band density functions ηrgb , ηnir , ηmir . . . are defined as ηi (q, t) ∈ [0, ∞) ∀q ∈ Ω. ηrgb
can also be treated as three bands ηr , ηg , ηb , which represent RED, GREEN and BLUE band
values of a pixel. The goal of remote sensing is to make a mapping from Ω to η1 , η2 , η3 . . . with
a preset spatial and temporal resolution for any q ∈ Ω and any t ∈ [t1 , t2 ] (Chao et al., 2008).
With the above remote sensing requirements, several specific characteristics need to be con-
sidered to get accurate georeferenced aerial images aside from an autonomous flying vehicle:
• Expense: most civilian applications require inexpensive UAS platforms instead of ex-
pensive military unmanned vehicles. However, most commercial-off-the-shelf (COTS)
autopilots cost more than $6000, let alone the camera and the air frame.
• Orientation Data: the orientation information when the image is taken is critical to the
image georeferencing. But many open source UAS autopilot systems don’t have good
supports to the accurate sensors. For example, Paparazzi uses IR sensors as the main
sensing unit by default.
• Image Synchronization: some COTS UAV could send videos down to the base station
and record them on the ground computer. But there is a problem that the picture may
not match up perfectly with the UAV data on the data log. The images may not syn-
chronize perfectly with the orientation data from the autopilot.
• Band configurable ability: a lot of remote sensing applications require more than one
band of aerial images like vegetation mapping and some of them may require RGB, NIR
and thermal images simultaneously.
3.2 AggieAir System Structure

AggieAir UAS includes the following subsystems:
(1) The flying-wing airframe: Unicorn wings with optional 48”, 60” and 72” wingspans are
used as the frame bed to fit in all the electronic parts. The control inputs include elevons
and a throttle motor.
(2) The OSAM-Paparazzi autopilot: the open source Paparazzi autopilot is modified by re-
placing the IR sensors with the IMU as the main sensing unit. Advanced navigation
routines like the survey of a random polygon are also added to support image acquisi-
tion of an area with a more general shape.
(3) The GhostFoto imaging payload subsystem: a high resolution camera system with both
the RGB and NIR band is developed. More importantly, the image system could guar-
antee an accurate synchronization with the current autopilot software.
(4) The communication subsystem: AggieAir has a 900MHz data link for GCS monitoring,
a 72MHz RC link for safety pilot backup, and an optional 2.4GHz wifi link for real time
image transmission.
(5) The Paparazzi ground control station (GCS): Paparazzi open source ground station is
used for the real-time UAS health monitoring and flight supervising.
(6) The gRAID software: a new World Wind plug-in named gRAID is developed for aerial
image processing including correcting, georeferencing and displaying the images on a
3D map of the world.
The physical structure of AggieAir is shown in Fig. 6, with the specifications in Tab 4 and the
airborne layout in Fig. 7.
AggieAir has the following advantages over other UAS platforms for remote sensing mis-
sions:
(1) Low costs: AggieAir airborne vehicles are built from scratches including the airframes
and all the onboard electronics. The total hardware cost is around $3500.
Fig. 6. AggieAir UAS Physical Structure.
(a) Aircraft Layout (b) Main Bay Layout
Fig. 7. AggieAir UAS Airborne Layout.
Specifications
Weight up to 8 lbs
Wingspan 72”
Flight Time ≤ 1 hour
Cruise Sped 15 m/s
Imaging Payload RGB/NIR/thermal camera
Operational Range up to 5 miles
Table 4. AggieAir UAS Specifications
(2) Full autonomy: AggieAir uses the Paparazzi autopilot, which supports the total auton-
omy of the air vehicle even without the ground station.
(3) Easy manipulation: only two people are required to launch, manipulate and land the
vehicle.
(4) Run-way free capability: the bungee launching system supports take-off and landing
basically at any soft field with only one launching operator.
(5) High spatial and temporal resolution: the image system could achieve submeter level
ground resolution and hour level time accuracy.
(6) Multiple bands for cameras: AggieAir supports RGB, and NIR bands for current image
subsystems. More band configurable imagers are also under developments.
3.3 OSAM-Paparazzi Autopilot Subsystem

It is clear that Kestrel and Piccolo autopilots are small, light and powerful. But their prices
are relatively high and most of their onboard software is not accessible to users, which is a
main disadvantage when georeferencing the aerial image after the flight (Jensen et al., 2008).
Paparazzi UAV project provides a cheap, robust and open source autopilot solution including
both hardware and software. But it uses Infra-red sensors for the attitude measurement, which
is not accurate enough when compared with most commercial UAV autopilots above.
To achieve an accurate image georeferencing with a fair price, our team choose to add an in-
ertial measurement unit (IMU) to the Paparazzi autopilot replacing the IR sensors. Paparazzi
Tiny WithOut GPS (TWOG) board is used together with the 900MHz Maxtream data modem
for real time communication to the GCS. Microstrain Gx2 IMU and Ublox Lea-5H GPS module
serve as the attitude and position sensors, respectively. Due to the limits from the IO ports, the
gumstix microprocessor is used as a bridge to connect IMU and GPS to the TWOG board. The
cascaded PID flight controller then converts all the sensor information into PWM signals for
the elevon and throttle motor to guide the vehicle for preplanned navigation. There is also a
72MHz RC receiver on board so that the human safety pilot could serve as the backup for the
autopilot in case of extreme conditions like strong winds. The physical layout of the airborne
system is shown in Fig. 8.
Fig. 8. AggieAir Airborne System Structure.
3.4 GhostFoto Image Subsystem

GhostFoto image subsystem is the second generation remote controlled digital camera system
developed at CSOIS (Han, Jensen & Dou, 2009). The hardware includes the Canon CCD cam-
era for image capture and the gumstix microprocessor for payload control and georeferencing
logging. Canon PowerShot SX100 IS CCD camera is used, illustrated in Fig. 9. This camera
has the remote capturing capability, an 8 mega pixel CCD panel supporting up to 3264 x 2448
pixels size and a 10x optical zoom lens with optical image stabilizer. The compact size and
relatively light weight (265g) of this camera make it easy to fit on small UAVs. Besides the
Fig. 9. Camera Body (left) and its CCD Sensor (right).
commonly used RGB channels, the camera could also support near infrared spectra by replac-
ing the visible light filter with an NIR filter. Our 72” airframe can carry two or three of these
imagers with different bands after removing unnecessary parts.
GFoto cameras are remotely controlled by the gumstix through USB 1.1 interface with Ghost-
Eye image capture software. GhostEye is based on libgphoto2 (gPhoto, Accessed 2008), which
is an open source portable digital camera library of C functions for UNIX-like operating sys-
tems. With libgphoto2 library functions, GhostEye is able to remotely control and configure
multiple cameras simultaneously through a Picture Transfer Protocol (PTP) driver. PTP is
a widely supported protocol developed by the International Imaging Industry Association
for transfer of images from digital cameras to computers (Picture Transfer Protocol, Accessed
2008). GhostEye also provides the communication link between the payload and the UAV
system. Messages can be reported from GhostEye to the ground station. The messages can be
shared even with other UAVs with the same protocol. Meanwhile, messages from the UAV
system can trigger the imagers. For example, after the altitude of the aircraft reaches a certain
level, the plane is able to command the imager to activate or deactivate capturing. The geo-
referencing data is logged by GhostEye in XML format to import the images into the gRAID.
3.5 gRAID Image Georeference Subsystem

The Geospatial Real-Time Aerial Image Display (gRAID) is a plug-in for NASA World Wind,
a 3D interactive open source world viewer (Jensen, 2009). gRAID takes the raw aerial images,
makes corrections for the camera radial distortion, and then overlays the images on the 3D
earth based upon the position and orientation data collected when they are captured. This
process can be done either in real-time while the plane is flying or after the flight. Human-
in-the-loop feature based image stitching can be done with conventional GIS software after
gRAID exports the image to a world file. gRAID could also create a gray scale image from
a single RGB channel. The images can be converted into world files and loaded into con-
ventional GIS software for further, advanced image processing. The detailed georeferecing
procedure is described as below.
To georeference the aerial images, several coordinate systems must first be defined, shown in
Figure 10.
• The body frame: the origin is defined at the center of gravity (CG), with the x axis
pointing through the nose, the y axis pointing to the right wing and the z axis pointing
down.
• The camera frame: the origin is located at the focal point of the camera. The axes of the
camera frame are rotated by φc , θc and ψc with respect to the body frame.
• The inertial frame: the origin is usually defined on the ground with the x, y, z axes
pointing towards the north, east and down, respectively. The orientation of the UAV
with respect to the NED frame is given by φ, θ and ψ.
• The earth centered earth fixed (ECEF) frame: the z axis passes through the north pole,
the x axis passes through the equator at the prime meridian and the y axis passes
through the equator at 90◦ longitude.
Fig. 10. Aircraft Coordinate Systems
Any point in an image can be rotated from the camera frame to the ECEF coordinate system in
order to find where it is located on the earth. However, it is only necessary to find the location
of the four corners of the image in order to georeference it. Assuming the origin is at the focal
point and the image is on the image plane, equation 1 can be used to find the four corners
of the image. As defined in figure 11, FOVx is the FOV around the x axis, FOVy is the FOV
around the y axis and f is the focal length.
Focal Point
(0,0)
V1 V2
Top
V3 V4
Image Plane
Fig. 11. Definition of Initial Image Corners

v1c = f tan( FOVy /2) − f tan( FOVx /2) f (1)

v2c = f tan( FOVy /2) f tan( FOVx /2) f (2)

v3c = − f tan( FOVy /2) − f tan( FOVx /2) f (3)

v4c = − f tan( FOVy /2) f tan( FOVx /2) f (4)
To rotate the corners into the navigation frame, they first need to be rotated into the body
frame. The Euler angles with respect to the body frame are given by φc , θc and ψc , and can be
used to create a clock-wise rotation matrix Rbc which rotates a vector in the body frame to the
camera frame.
Rbc = R xyz (φc , θc , ψc ) (5)

To rotate from the camera frame to the body frame, the transpose of Rbc is used.
Rcb = ( Rbc ) T = Rzyx (−θc , −ψc , −φc ) (6)

The same rotation matrix is used, with φ, θ and ψ, to rotate from the body into the navigation
frame.
Rbn = ( Rnb ) T = Rzyx (−θ, −ψ, −φ) (7)

Now each corner is rotated from the camera frame into the navigation frame using equation
8.
vin = Rbn Rcb vic (8)

Now that the corners are in the NED coordinate system, they are scaled to the ground to find
their appropriate magnitude (assuming flat earth) where h is the height of the UAV above
ground and vin (z) is the z component of vin .
h
vin = vin (9)
vin (z)
The next step is to rotate the image corners into the ECEF coordinate system. This is done
with another rotation matrix and the latitude (λ) and longitude (α) of the UAV.
π
Rnw = Rzyy (−α, , λ) (10)
2
viw = Rnw vin (11)
After the corners are rotated into the ECEF coordinate system, they are located in the center of
the earth and need to be translated up to the position of the UAV in cartesian coordinates (p).
viw = viw + p (12)

Now viwrepresents the position of each of the image corners, in cartesian coordinates, pro-
jected on the earth.
3.6 Image Orthorectification

Even though small, low-cost unmanned aerial vehicles (UAVs) make good remote sensing
platforms by reducing the cost and making imagery easier to obtain, there are also some
tradeoffs. The low altitude, small image footprint and high number of images make it difficult
and tedious to georeference the images based on features. Auto-orthorectification techniques
based on the position and attitude of the UAV would work well except the inherent errors
in the UAV sensors reduce the accuracy of the orthorectification significantly. The orthorec-
tification accuracy can be improved by calibrating the UAV sensors. This is done by inverse
orthorectifing the images to find the actual position and attitude of the UAV using ground
references setup in a square. Actual data from a test flight is used to validate this method
(Jensen, Han & Chen, 2009).
As detailed above, a point in the image plane (pi ) can be transformed into Earth-Centered
Earth-Fixed (ECEF) coordinates ( pw ) using equation 13 where uw is the position of the UAV
in ECEF, Rcb is the rotation matrix from the camera frame to the body frame, Rbn is the rota-
tion matrix from the body frame to the navigation frame, Rnw is the rotation matrix from the
navigation frame to ECEF, and h is the height above ground of the UAV.
pw = aRnw Rbn Rcb pi + uw (13)

h
a=
vz T Rbn Rcb pi
 
0
vz =  0 
1
There is a possibility that equation 13 could be used directly to find the position and attitude
of the UAV given multiple known ground control points ( pw ) and their positions on an image
(pi ). However, this could prove to be very complicated. The method presented here will take a
simple, indirect approach by setting up the ground control points in a square (Figure 12). The
properties of this square, where the locations of the corners are measured, can be compared to
the properties of another square where the corner positions are estimated using equation 13.
By changing the position and attitude of the UAV, the properties of the estimated square can
be adjusted to match the properties of the measured square. The correct position and attitude
of the UAV is found when the properties of the measured and estimated squares match. For
example, the difference between the areas of each square reflects the measured and actual
altitude of the UAV above ground. If the measured square has an area greater than the area of
the estimated square, the altitude of the UAV needs to be increased. The estimated square is
then recalculated using equation 13 and the areas are compared again. Once the areas match,
the correct altitude is found.
Fig. 12. Ground Targets in Square
The position and yaw of the UAV are easier to find than the altitude. This is because the
difference in the position and orientation of the squares are directly related to the difference
between the measured and actual position and yaw of the UAV. Therefore, the difference
between the position and orientation of the squares can simply be added to the measured
values of the position and yaw of the UAV to find the actual position and yaw.
Finding a property of the square related to roll and pitch is more complicated than the other
properties. The shape, the length of each side and the length of the diagonals could all have a
relationship to roll and pitch. However, this relationship all depends on the orientation of the
square relative to the image. More work will need to be done in order to find the actual roll
and pitch with this method.
This method was tested by collecting 40 images of the ground references at various heights
and headings. Without any correction, the position of the ground references had errors of up
to 45m. Correcting the altitude did not show significant improvement, however correcting the
yaw decreased the error to below 20m. Correcting for the position also had a profound effect
and reduced the error to below 5m of error.
Linear relationships were also found between the actual and measured position and attitude
of the UAV from the experiment. The altitude had a small bias of 4m and a slope of 0.93 which
shows that the error of the altitude gets worse as the altitude increases. The yaw was also very
linear with a one-to-one relationship and a bias of 13 degrees. A relationship between the
actual and measured position was more unclear but still showed that as the altitude increases,
the magnitude of the position error also increases. The unclear relationship in the position
error is probably because the roll and pitch were not compensated for. The direction of the
position error, however, had a linear relationship with the heading of the aircraft with a bias
of 64 degrees. This could be due to a misalignment between the cameras and the body axis of
the aircraft.
4. Sample Applications for AggieAir

The typical sample application of using the AggieAir UAS for remote sensing missions could
be defined as follows (Chao et al., 2008). Given a random area Ω, UAVs with functions of
altitude and speed maintenance and waypoint navigation: speed v ∈ [v1 , v2 ], possible flight
height h ∈ [h1 , h2 ], camera with specification: focal length F, image sensor pixel size: PSh ×
PSv , image sensor pixel pitch PPh × PPv , the interval between images acquired by the camera
(the “camera shooting interval”) tshoot , the minimal shooting time tshootmin , the desired aerial
image resolution res, the control objective is:
min t f light = g(Ω, h, v, {q1 , . . . , qi }, tshoot , res), (14)
s.t.v ∈ [v1 , v2 ], h ∈ [h1 , h2 ], tshoot = k × tshootmin . where t f light is the flight time of the UAV for
effective coverage, g(Ω, h, v, tshoot ) is the function to determine the flight path and flight time
for effective coverage, k is a positive integer. In other words, the UAS is required to make a
full coverage map of the given area, which could also be called the coverage control problem.
The control inputs of the coverage controller include bounded velocity v, bounded flight
height H, a set of preset UAV waypoints {q1 , q2 , . . . , qi } and the camera shooting interval tshoot .
The system states are the real UAV trajectory {q̄t1 , . . . , q̄t2 } and the system output is a series of
aerial images or a video stream taken between t1 and t2 .
Assume the imager is mounted with its lens vertically pointing down towards the earth; its
footprint (shown in Fig. 13) can be calculated as:
h × PPh × PSh h × PPv × PSv

FPh = , FPv = .
F F
Most UAVs can maintain a certain altitude while taking pictures so the UAV flight height h
can be determined first based on camera and resolution requirements. Assuming that different
flight altitudes have no effect on the flight speed, we get
√
res × F
h= . (15)
max ( PPh , PPv )
Given the flight height h and the area of interest Ω, the flight path, cruise speed and camera
shooting interval must also to be determined. Without loss of generality, Ω is assumed to be
a rectangular since most other polygons can be approximated by several smaller rectangles.
The most intuitive flight path for the UAV flight can be obtained by dividing the area into
strips based on the group spatial resolution, shown in Fig. 14(a). The images taken during
Fig. 13. Footprint Calculation.
UAV turning are not usable because they may have bad resolutions. Due to the limitation
from the autopilot, GPS accuracy and wind, the UAV cannot fly perfectly straight along the
preset waypoints. To compensate the overlapping percentage between two adjacent sweeps
o must also be determined before flight; this compensation is based on experience from the
later image stitching as shown in Fig. 14(b).
(a) Ideal Flight Path (b) Flight Path with Overlap
Fig. 14. UAV Flight Path.
Given the overlapping percentage o% between sweeps, the ground overlapping o g can be
determined by:
o g = (1 − o%) × FPh . (16)
The minimal camera shooting interval can be computed as:
(1 − o%) × FPv
tshootmin = . (17)
v
This open-loop solution is intuitive, robust to all the polygons and requires little computation.
However, this method requires that many parameters, especially the overlapping percentage
o% to be set up based on experience; it cannot provide an optimal solution. More work on a

close-loop real-time solution is needed for an optimal solution.
Preliminary experimental results are shown in this section to demonstrate the effectiveness
of the whole UAV remote sensing system on both the hardware and software levels. Three
sample applications are introduced in detail. A remote data collection application is offered to
demonstrate the feasibility of using UAVs to collect data from ground-based sensors through
wireless modems. Acquisition of photographic data over Desert Lake, Utah, illustrates an
application involving the use of the RGB and NIR imagers. The data collected by UAVs from
ground sensors can also be used for comparison and calibration with information from UAV
images. Finally, we present a farmland coverage test with image stitching in our regular test
flight site.
4.1 Farmland Coverage

The goal of irrigation control is to minimize the water consumption while sustaining the agri-
culture production and human needs (Fedro et al., 1993). This optimization problem requires
remote sensing to provide real-time feedback from the farmland field including:
• Water: water quantity and quality with temporal and spatial information, for example
water level of a canal or lake.
• Soil: soil moisture and type with temporal and spatial information.
• Vegetation: vegetation index, quantity and quality with temporal and spatial informa-
tion, for example the stage of growth of the crop.
The “real-time” here means daily or weekly temporal resolution based on different applica-
tions. AggieAir is currently involved in a large scale agricultural project which will use visual
and NIR images from AggieAir to measure the soil moisture of the area to help save water
(Jensen, Chen, Hardy & McKee, 2009). The project includes 30 square miles of different types
of crops where 87 wireless soil moisture ground probes are placed. The ground probes sample
and send the data every hour through wireless to a base station where the data is displayed
on the Internet. The ground probes and Landsat data will be used to calibrate the images
from AggieAir using the downscaling techniques described in (Kaheil, Gill, McKee, Bastidas
& Rosero, 2008) and (Kaheil, Rosero, Gill, McKee & Bastidas, 2008). After calibration, these
images should be able to measure soil moisture and evapotranspiration for water managers
and farmers whenever it is needed. However, AggieAir has not yet been flown for this project
due to the large area and the high altitude AggieAir will need to fly at. Approval from the
Federal Aviation Administration (FAA) is currently being sought to ensure the safety of all
airborne vehicles before AggieAir is flown for this project. A research farm (one square mile)
coverage map is provided in Fig. 15 to show the capability of AggieAir.
4.2 Road Surveying

AggieAir UAS could also provide low-cost aerial images for road and highway construction
and maintenance. Figure 16 shows a highway intersection located in Logan Canyon, which
was recently rebuilt for better safety to turn onto the main road. The Utah Department of
Transportation (UDOT) usually needs to photograph the area to be built or altered before con-
struction with a manned aircraft. However no imagery is available during or after construc-
tion due to the cost and availability of the imagery. UDOT is not only interested in AggieAir
to lower the cost and increase the availability of imagery for construction, but also to update
their inventory of signs, culverts, traffic lines, etc. The aerial images acquired by manned
aircraft and by AggieAir are shown in Fig. 16.
Fig. 15. Cache Junction Farm Coverage Map
(a) Before Construction (b) After Construction (AggieAir)
Fig. 16. Beaver Resort Intersection

4.3 Water Area Coverage

Water areas include wetlands, lakes, or ponds, etc. Water areas could provide lots of informa-
tion to ecological environment changes, flood damage predictions, and water balance man-
agement. Desert Lake coverage mission is a typical example, which lies in west-central Utah
(latitude: 39◦ 22 5 N, longitude: 110◦ 46 52 W). It is formed from return flows from irrigated
farms in that area. It is also a waterfowl management area. This proposes a potential problem
because the irrigation return flows can cause the lake to have high concentrations of min-
eral salts, which can affect the waterfowl that utilize the lake. Managers of the Desert Lake
resource are interested in the affect of salinity control measures that have been recently con-
structed by irrigators in the area. This requires estimation of evaporation rates from the Desert
Lake area, including differential rates from open water, wetland areas, and dry areas. Estima-
tion of these rates requires data on areas of open water, wetland, and dry lands, which, due to
the relatively small size and complicated geometry of the ponds and wetlands of Desert Lake,
are not available from satellite images. A UAV can provide a better solution for the problem
of acquiring periodic information about areas of open water, etc., since it can be flown more
frequently and at little cost.
The whole Desert Lake area is about 2 ∗ 2 miles. It is comprised of four ponds and some
wetland areas. The early version of AggieAir imaging payload, GFDV, is used in this mission
together with the Procerus UAV with real-time, simultaneous RGB and NIR videos. Both the
RGB and NIR videos are transmitted back to the ground station in real time. The photos are
stitched using gRAID, shown in Fig 17.
(a) RGB Image (b) NIR Image
Fig. 17. Desert Lake Coverage Map
4.4 Riparian Surveillance

Riparian buffer surveillance is becoming increasingly more important since it is challenging
to maintain stream ecosystem integrity and water quality with the current rapidly changing
land use (Goetz, 2006). AggieAir UAS could be used in several applications including river
tracking, vegetation mapping and hydraulic modeling, etc.
4.4.1 River Tracking

The path and flow of a river might constantly change due to drought, flood or other natu-
ral calamities. Because of this, the aerial images of the river path could be outdated or in
low quality, making it difficult to perform studies of the changed river and the variations of
its nearby ecological system. AggieAir UAS platform with the high resolution multi-spectral
camera system could present a real-time low-cost solution to the river tracking problem Han,
Dou & Chen (2009). A flight plan with 3D waypoints could be formed by integrating flow
line data from NHDPlus (National Hydrography Dataset Plus) and DEM (Digital Elevation
Model) from USGS (U.S. Geological Survey). The images captured by the cameras are pro-
cessed in real time. Based on the information derived from these images, waypoints are dy-
namically generated for the autonomous navigation so that the UAV can exactly follow the
changed river path and the focus of each image from the camera system is on the center of the
river. The actual flight results collected in several flying experiments along a river verify the
effectiveness of AggieAir System, shown in Fig 18. Example RGB and NIR pictures acquired
by AggieAir is also shown in Fig 19.
Fig. 18. River Tracking Map after Stitching.
4.4.2 Vegetation Mapping & Hydraulic Modeling

Figure 20 shows some imagery taken with AggieAir of a small section of the Oneida Narrows
near Preston Idaho. A team of engineers used this imagery to map the substrate and vegeta-
tion for 2D hydraulic and habitat modeling. Normally, the team uses low resolution, outdated
imagery to map rivers. This can be difficult when the vegetation, the path and the flow of the
river are always changing. The imagery from AggieAir, however, was up-to-date (within
a week) and had high resolution (5 cm), which made mapping the river quick and easy. Not
only could different types of vegetation be distinguished from the imagery, but different types
of sediment, like sand piles, could also easily be distinguished.
4.5 Remote Data Collection

Many agricultural and environmental applications require deployment of sensors for mea-
surement of the interested field. However, it is not always easy or inexpensive to collect all
(a) RGB Image (b) NIR Image
Fig. 19. Sample Picture for River Tracking
(a) Substrate Map (b) Vegetation Map
Fig. 20. Oneida Narrows Imagery
the data from remote data loggers for further processing. Many applications still require hu-
mans to get close to the ground-based sensors to retrieve the data from their data loggers.
Wireless sensor networks and satellite networks are used in environmental data collection
applications, but wireless communication can be expensive and vulnerable to changing envi-
ronmental conditions (such as loss of line-of-sight due to vegetation growth). This problem is
especially difficult when the sensors are deployed sparsely over a large geographic area where
transportation might be limited by terrain conditions. UAVs can fly into such areas without
affecting the vegetation on the ground; they can spare humans from having to enter danger-
ous or difficult areas; and they might be able to operate at lower costs that might be required
for approaches involving direct human access to the data. Moreover, UAVs can achieve better
wireless communication since the signal can be transmitted more dependably in the air than
near ground level.
One typical example of remote data collection is fish tracking. In order to understand fish
habitats, radio transmitters are planted in fish in order to locate and track their movements.
Human operators are needed to drive a boat around a lake or down a river following the
periodic beacon sent from the transmitters. The beacon is heard through a radio receiver with
a directional antenna and its strength is highly dependent on the distance and the direction
the antenna is pointed at. AggieAir could be employed here with an onboard self-designed
device to catch the signal from the transmitter and record its strength. Thus, the location of
the fish can be found and recorded much easier and faster since the wireless signal transmits
much better in the air. Hardware developments and real experiments are still undergoing.
5. Towards Band Reconfigurable Multiple UAV based Remote Sensing

A single AggieAir system could have many applications as mentioned above, but some irriga-
tion applications may require remote sensing of a large land area (more than 30 square miles)
within a short time (less than one hour). Acquisition of imagery on this geographic scale is
difficult for a single UAV. However, groups of UAVs (which we call ”covens”) can solve this
problem because they can provide images from more spectral bands in a shorter time than a
single UAV.
The following missions will need multiple UAVs (covens) operating cooperatively for remote
sensing:
• Measure η1 , η2 , η3 . . . simultaneously.
• Measure ηi (q, t) within a short time.
To fulfill the above requirements, UAVs equipped with imagers having different wavelength
bands must fly in some formation to acquire the largest number of images simultaneously.
The reason for this requirement is that electromagnetic radiation may change significantly,
even over a period of minutes, which in turn may affect the final product of remote sensing.
The “V” or “—” formation, keeping algorithm similar to the axial alignment (Ren et al., 2008),
can be used here since the only difference is that the axis is moving:
q̇dm = − ∑ [(qm − qn ) − (δm − δn )], (18)

n∈Jm (t)
where qdm is the preset desired waypoints, Jm (t) represents the UAV group, δm = [δmx , δmy ] T
can be chosen to guarantee that the UAVs align on a horizontal line with a certain distance in
between.
Based on the theoretical analysis and our preliminary results, more effort is needed to achieve
the final band reconfigurable multi-UAV based cooperative remote sensing.
• (1) Multiple UAVs: the preliminary results have shown that the sensing range of the Ag-
gieAir UAS is about 2.5 × 2.5 miles, given the current battery energy density. However,
light reflection varies a great deal in one day and accurate NIR images require at most a
one-hour acquisition time for capture of the entire composite image. This motivates the
use of covens, or multiple UAVs, for this type of application, with each UAV carrying
one imager with a certain band.
• (2) More robust control: the current UAV platform requires winds of less than 10 m/s.
However, the UAV needs to have the ability to deal with wind gusts, which is problem-
atic for UAV flight.
• (3) Accurate real-time image stitching and registration: the current software requires
manual post processing for accurate georeferencing results. For the application to be at-
tractive to managers of real irrigation systems, manual post processing is unacceptable
because of its cost and requirement of technically trained individuals.
Acknowledgement
This work is supported in part by the Utah Water Research Laboratory (UWRL) MLF Seed
Grant (2006-2009) on “Development of Inexpensive UAV Capability for High-Resolution Re-
mote Sensing of Land Surface Hydrologic Processes: Evapotranspiration and Soil Moisture.”
This work is also partly supported by NSF Grants #0552758 and #0851709. The authors would
like to thank Professor Raymond L. Cartee for providing the Utah State University research
farm at Cache Junction as the UAV flight test field, and Calvin Coopmans, Long Di, Daniel
Morgan for their technical support and help to the flight tests. The authors would also like to
thank Shannon Clemens for georeferencing and generating mosaics.
6. References
Beard, R., Kingston, D., Quigley, M., Snyder, D., Christiansen, R., Johnson, W., Mclain, T. &
Goodrich, M. (2005). Autonomous vehicle technologies for small fixed wing UAVs,
AIAA J. Aerospace Computing, Information, and Communication 5(1): 92–108.
Brisset, P., Drouin, A., Gorraz, M., Huard, P. S. & Tyler, J. (2006). The paparazzi solution, MAV
2006, Sandestin, Florida, USA.
Calise, A. J., Johnson, E. N., Johnson, M. D., & Corban, J. E. (2003). Applications of adaptive
neural-network control to unmanned aerial vehicles, in Proc. AIAA/ICAS Int. Air and
Space Symp. and Exposition: The Next 100 Years, Dayton, Ohio, USA.
Chao, H., Baumann, M., Jensen, A. M., Chen, Y. Q., Cao, Y., Ren, W. & McKee, M. (2008).
Band-reconfigurable multi-UAV-based cooperative remote sensing for real-time wa-
ter management and distributed irrigation control, in Proc. IFAC World Congress,
Seoul, Korea, pp. 11744–11749.
Chao, H., Cao, Y. & Chen, Y. Q. (2009). Autopilots for small fixed wing unmanned aerial
vehicles: a survey, Int. J. Control, Automation and Systems . Accepted to appear.
CloudCap Inc. (Accessed 2008).
URL http://www.cloudcaptech.com.
Damien, D., Wageeh, B. & Rodney, W. (2007). Fixed-wing attitude estimation using computer
vision based horizon detection, in Proc. Australian Int. Aerospace Congress, Melbourne,
Australia, pp. 1–19.
Egan, G. K. (2006). The use of infrared sensors for absolute attitude determination of un-
manned aerial vehicles, Tech. Rep. MECSE-22-2006, Monash University.
Fedro, S. Z., Allen, G. S. & Gary, A. C. (1993). Irrigation system controllers, Series of the Agri-
cultural and Biological Engineering Department, Florida Cooperative Extension Service, In-
stitute of Food and Agricultural Sciences, University of Florida .
URL: http://edis.ifas.ufl.edu/AE077
Goetz, S. J. (2006). Remote sensing of riparian buffers: Past progress and future prospects,
Journal of the American Water Resources Association 42(11): 133–143.
gPhoto (Accessed 2008).
URL http://www.gphoto.org/.
Han, Y., Dou, H. & Chen, Y. Q. (2009). Mapping river changes using low cost autonomous
unmanned aerial vehicles, in Proc. AWRA Spring Specialty Conf. Managing Water Re-
sources Development in a Changing Climate, Anchorage, Alaska, USA.
Han, Y., Jensen, A. M. & Dou, H. (2009). Programmable multispectral imager development
as light weight payload for low cost fixed wing unmanned aerial vehicles, in Proc.
ASME Design Engineering Technical Conf. Computers and Information in Engineering,
number MESA-87741, San Diego, California, USA.
James, B. (2006). Introduction to Remote Sensing, 4th edn, Guilford Press.
Jensen, A. M. (2009). gRAID: A geospatial real-time aerial image display for a low-cost autonomous
multispectral remote sensing platform, M.S. Thesis, Utah State Univeristy.
Jensen, A. M., Baumann, M. & Chen, Y. Q. (2008). Low-cost multispectral aerial imaging
using autonomous runway-free small flying wing vehicles, in Proc. IEEE Int. Conf.
Geoscience and Remote Sensing Symp., Boston, Massachusetts, USA, pp. 506–509.
Jensen, A. M., Chen, Y. Q., Hardy, T. & McKee, M. (2009). AggieAir - a low-cost autonomous
multispectral remote sensing platform: New developments and applications, in Proc.
IEEE Int. Conf. Geoscience and Remote Sensing Symp., Cape Town, South Africa, pp. –.
Jensen, A. M., Han, Y. & Chen, Y. Q. (2009). Using aerial images to calibrate inertial sensors
of a low-cost multispectral autonomous remote sensing platform (AggieAir), in Proc.
IEEE Int. Conf. Geoscience and Remote Sensing Symp., Cape Town, South Africa, pp. –.
Johnson, L., Herwitz, S., Dunagan, S., Lobitz, B., Sullivan, D. & Slye, R. (2003). Collection of
ultra high spatial and spectral resolution image data over California vineyards with a
small UAV, in Proc. Int. Symp. Remote Sensing of Environment, Honolulu, Hawai, USA.
Johnson, L., Herwitz, S., Lobitz, B. & Dunagan, S. (2004). Feasibility of monitoring coffee
field ripeness with airborne multispectral imagery, Applied Engineering in Agriculture
20: 845–849.
Kaheil, Y. H., Gill, M. K., McKee, M., Bastidas, L. A. & Rosero, E. (2008). Downscaling and
assimilation of surface soil moisture using ground truth measurements, IEEE Trans.
Geoscience and Remote Sensing 46(5): 1375–1384.
Kaheil, Y. H., Rosero, E., Gill, M. K., McKee, M. & Bastidas, L. A. (2008). Downscaling and
forecasting of evapotranspiration using a synthetic model of wavelets and support
vector machines, IEEE Trans. Geoscience and Remote Sensing 46(9): 2692–2707.
Microstrain Inc. (Accessed 2008).
URL http://www.mirostrain.com.
Paparazzi Forum (Accessed 2008).
URL http://www.recherche.enac.fr/paparazzi/.
Picture Transfer Protocol (Accessed 2008).
URL http://en.wikipedia.org/wiki/Picture_Transfer_Protocol.
Ren, W., Chao, H., Bourgeous, W., Sorensen, N. & Chen., Y. Q. (2008). Experimental validation
of consensus algorithms for multi-vehicle cooperative control, IEEE Trans. Control
Systems Technology 16(4): 745–752.
Roberts, P. J., Walker, R. A. & O’Shea, P. J. (2005). Fixed wing UAV navigation and control
through integrated GNSS and vision, in Proc. AIAA Guidance, Navigation, and Control
Conf. and Exhibit, number AIAA 2005-5867, San Francisco, California, USA.
Tarbert, B., Wierzbanowski, T., Chernoff, E. & Egan, P. (2009). Comprehensive set of rec-
ommendations for sUAS regulatory development, Tech. Rep., Small UAS Aviation
Rulemaking Committee.
Possibilistic and fuzzy multi-sensor fusion for humanitarian mine action 491
23
X
Possibilistic and fuzzy multi-sensor fusion

for humanitarian mine action
Nada Milisavljević1 and Isabelle Bloch2
1Signal
and Image Centre
Royal Military Academy
Brussels, Belgium
2TELECOM ParisTech (ENST)
CNRS UMR 5141 LTCI

Paris, France
1. Introduction
Some multi-sensor data fusion applications to humanitarian mine action are presented in
this chapter. Data fusion techniques may be useful to two main humanitarian mine action
types: close range antipersonnel mine detection and remote sensing mined area reduction.
Close range antipersonnel mine detection refers to detecting surface and subsurface
anomalies that may be related to the presence of mines (e.g., detection of differences in
dielectric constant, thus in the type of material, using a ground-penetrating radar) and/or
detection of explosive materials. Area reduction refers to identifying the mine-free areas out
of the mine-suspected areas. Data fusion for these two applications is discussed here.
In case of both humanitarian mine action types, efficient modelling and fusion of extracted
features can increase the overall performance of single-sensor based processing. Yet, a wide
variety of scenarios and conditions typically exists between different minefields and within
a minefield itself. Therefore, a high-quality performance of humanitarian mine action tools
can only be obtained using multi-sensor and data fusion approaches. In addition, as the
sensors used are, actually, detectors of different anomalies that may be related to mines (and
not of mines themselves), combinations of these complementary pieces of information might
improve the detection and classification results. Finally, in order to consider the inter- and
intra-minefield variability, uncertainty, partial knowledge and ambiguity, knowledge-based
theories prove to be useful, such as belief functions in the framework of the Dempster-
Shafer evidence theory (Shafer, 1976; Smets, 1994; Smets & Kennes, 1994) or possibilistic and
fuzzy theory (Dubois & Prade 2006; Zadeh, 1965). We focus on the latter in this chapter.
Namely, as the possibilistic and fuzzy theory allows for the flexibility in the choice of
combination operators, we exploit this fact in order to account for the different
characteristics of the sensors to be combined.
For close range detection, modeling and fusion of extracted features are shown, based on the
possibilistic and fuzzy theory. The approaches are discussed for the case of using real data
coming from three complementary sensors (ground-penetrating radar - GPR, infrared
sensor – IR, and metal detector - MD), collected within the Dutch project HOM-2000. These
results are obtained within two Belgian humanitarian demining projects, HUDEM and
BEMAT. In case of mined area reduction, a multi-sensor data fusion method based on fuzzy
logic is presented and discussed in the case of using real data of synthetic-aperture radar
and multi-spectral sensors, gathered within the EU project on space and airborne mined
area reduction tools (SMART). The influence of various knowledge sources on the
performance of the overall system is demonstrated too.
2. Fuzzy and possibilistic fusion

In the framework of fuzzy sets and possibility theory (Zadeh, 1965; Dubois & Prade, 2004),
the modeling step consists in defining a membership function to each class or hypothesis in
each source, or a possibility distribution over the set of hypotheses in each source. Such
models explicitly represent imprecision in the information, as well as possible ambiguity
between classes or decisions.
For the combination step in the fusion process, the advantages of fuzzy sets and possibilities
rely in the variety of combination operators, which may deal with heterogeneous
information (Dubois & Prade, 1985). Among the main operators, we find t-norms, t-
conorms, mean operators, symmetrical sums, and operators taking into account conflict
between sources or reliability of the sources.
We can classify these operators with respect to their behaviour (conjunctive, disjunctive,
compromise (Dubois & Prade, 1985)), the possible control of this behaviour, their properties
and their decisiveness, which proved to be useful for several applications (Bloch, 1996). On
the contrary to other data fusion theories (e.g., Bayesian or Dempster-Shafer combination),
fuzzy sets provide a great flexibility in the choice of the operator, that can be adapted to any
particular situation. The theory of belief functions has also been endowed with a similar
flexibility in some recent works (Denoeux, 2008).
An advantage of this approach is that it is able to combine heterogeneous information,
which is usually the case in multi-source fusion (as in both examples given in the chapter),
and to avoid to define a more or less arbitrary and questionable metric between pieces of
information issued from these images, since each piece of information is converted in
membership functions or possibility distributions over the same decision space.
Decision is usually taken from the maximum of membership or possibility values after the
combination step. Constraints can be added to this decision, typically for checking for the
reliability of the decision (is the obtained value high enough?) or for the discrimination
power of the fusion (is the difference between the two highest values high enough?).
A main difference between fuzzy classification and possibilistic classification is that classes
are generally considered as fuzzy sets in the first case and as crisp ones in the second case.
In the following sections, we illustrate these two types of modelling.
In Section 3, we consider the first application (close-range detection), where the decision
classes are crisp ones. We model the uncertainty and imprecision in the features as
possibility distributions of the set of classes. Then in Section 4, we go one step further by
modeling the imprecision attached to the measurements themselves, or the derived
possibility values, leading to more robustness.
In Section 5, we consider the second application (mined area reduction), and model the
classes (here landcover classes) as fuzzy sets, accounting for the imprecision attached to the
images and the definition of classes.
3. Possibilistic approach to close-range detection

We define the problem of close-range detection in the context of humanitarian demining as
the assignment of some measures (provided by different sensors) to one of two classes:
mines or friendly objects. In this section, we summarize previous results (Milisavljević &
Bloch, 2003; Milisavljević & Bloch, 2008) obtained using possibilistic models, by considering
the two classes as crisp ones. A number of measures are extracted from the data of the
sensors used:
• the area and the shape (elongation and ellipse fitting) of the object observed using the
IR sensor,
• the size of the metallic area in MD data,
• the propagation velocity (i.e., the type of material), the burial depth of the object
observed using GPR, and the ratio between object size and its scattering function.
3.1 Possibilistic modelling of the measures

In the frame of the possibility theory, we model information provided by these measures in
terms of possibility distributions. We have two spaces for each function: the set of
hypotheses, which is the same for each function, and the feature space, which depends on
the measurements from which the function is derived. Each function for one hypothesis is
therefore not a value but a function over the feature space (a function of depth value, of
shape measure, etc.). The specific shape of the functions and their parameters are derived
from the knowledge, indicating their general behaviour.
IR measures. Elongation and ellipse fitting measures provide information mainly on shape
regularity (Milisavljević & Bloch, 2003; Milisavljević & Bloch, 2008). The possibility degrees
of being a regular-shaped mine (MR), derived from these two measures, are denoted by
π1I(MR) and π2I(MR), while π1I(MI) and π2I(MI) represent the possibility degrees of being an
irregular-shaped mine (MI). Possibility degrees of being a regular-shaped non-dangerous
(friendly) object (FR) and irregularly shaped friendly object (FI) are denoted by π1I(FR) and
π1I(FI) for elongation and by π2I(FR) and π2I(FI) for ellipse fitting measure.
In case of elongation, we find r1 as the ratio between minimum and maximum distance of
bordering pixels from the center of gravity (on thresholded images) and r2 as the ratio of
minor and major axis obtained from second moment calculation (Haralick & Shapiro, 1992;
Milisavljević & Bloch, 2003). Then we define:
, (1)
. (2)
Regarding ellipse fitting, we calculate Aoe as the part of object area that belongs to the fitted
ellipse as well, Ao is the object area, and Ae is the ellipse area. Using these values, the
following possibility degrees are derived (the subtraction of 5 pixels is related to the limit
case - minimum 5 points needed to define an ellipse):
, (3)
. (4)
Finally, the area directly provides a degree π3I(M) (where M=MR∪MI) of being a mine -
since the range of possible anti-personnel (AP) mine sizes is approximately known, the
degree of possibility of being a mine is derived as a function of the measured size:
, (5)
where aI is the actual object area on the IR image, while the approximate range of expectable
mine areas is between a1 and a2 . For AP mine detection, we set a1 = 15 cm2 and a2 = 225 cm2
(Milisavljević & Bloch, 2003; Fischer et al., 2007). Friendly objects can be of any size,
meaning that the measured size is uninformative about the possibility of being a friendly
object (F=FR∪FI):
. (6)
MD measures. Under some conditions (Thonnard & Milisavljević, 1999; Das, 2006;
Milisavljević & Bloch, 2008), it is possible to extract the object shape and area as seen by MD,
as well as the burial depth. In reality, it can be difficult to adjust the sensitivity so that all the
low-metal content mines are detected without causing the data saturation for high-metal
content objects. In addition, in order to speed up the scanning time in large minefields, the
data gathering resolution in the cross-scanning direction can be very poor. Thus, the MD
information consists typically of only one measure, which is the width of the region in the
scanning direction, w [cm]. As friendly objects can contain metal of any size, we define:
. (7)
For most of AP mines, the range of the expected sizes of metal in mines is between 5 cm and
15 cm), so we can assign possibilities to mines as:
. (8)
GPR measures. All three GPR measures provide information about mines (Milisavljević et
al., 2003).
Regarding burial depth information (D), friendly objects can be found at any depth. On the
contrary, mainly due to their activation principles, there exists a maximum depth (Dmax,
typically 25 cm) up to which AP mines can be found. Therefore, possibility distributions for
mines, π1G(M), and friendly objects, π1G(F), for this measure can be modeled as follows:
, (9)
. (10)
In case of the ratio d/k between object size seen in the scanning direction, d, and its scattering
function, k (which is directly related to the object shape (Capineri et al., 1998)), friendly
objects can have any value of this measure. For mines, there is a range of values that mines
can have, and outside that range, the object is quite certainly not a mine:
, (11)
, (12)
where md is the d/k value at which the possibility distribution reaches its maximum value
(here, md = 700, chosen based on (Capineri et al., 1998)), and p is the width of the exponential
function (here, p = 400).
The third GPR measure, propagation velocity, v, can provide information about object
identity. For this measure, we extract depth information on a different way than in the case
of the burial depth measure 0 and we preserve the sign of the extracted depth. This
information indicates whether a potential object is above the surface, in which case the
extracted propagation velocity should be close to c = 3⋅108 m/s, the propagation velocity in
vacuum. If the sign indicates that the object is below the soil surface, the value of v should
be around the values for the corresponding medium, e.g., from 5.5⋅107 m/s to 1.73⋅108 m/s
(Capineri et al., 1998) in case of sand:
, (13)
with vt being the most typical velocity for the medium (for sand, it is 0.5⋅ (5.5⋅107 +1.73⋅108) =
1.14⋅108 m/s, and for air, it is equal to c), and h is the width of the exponential function (here,
h = 6⋅107 m/s). Again, friendly objects can have any value of the velocity:
. (14)
Note that all these functions have formal expressions that have been chosen in order to
achieve suitable shapes and behaviours. However, other functions sharing the same
properties could be used as well and we experienced that a fine tuning is not necessary.
3.2 Combination of possibility degrees

The combination is performed in two steps: the first one applies to all measures derived
from one sensor, while the second one combines results obtained in the first step for all three
sensors. Here, only the combination rules related to mines are considered. The issue of
combination rules for friendly objects is discussed in detail in (Milisavljević & Bloch, 2008).
Let us first discuss the first step for each sensor. For IR, since mines can be regular or
irregular, the information about regularity on the level of each shape measure is combined
using a disjunctive operator:
, (15)
. (16)
The choice of the maximum (smallest disjunction and idempotent operator) as a t-conorm is
related to the fact that the measures cannot be considered as completely independent from
each other, so there is no reason to reinforce the measures by using a larger t-conorm, and
the idempotent one is preferable in such situations. These two shape constraints should be
both satisfied to have a high degree of possibility of being a mine (Milisavljević & Bloch,
2008). Thus, they are combined in a conjunctive way (here using a product). Finally, the
object is possibly a mine if it has a size in the expected range, or if it is not in the expected
range, but satisfies the shape constraint, hence the final combination for IR is:
. (17)
In case of MD, as there is just one measure used, there is no first combination step and the
possibility degrees obtained using (7) and (8) are directly used.
Regarding GPR, it is possible to have a mine if the object is at shallow depths and its
dimensions resemble a mine and the extracted propagation velocity is appropriate for the
medium. Therefore, the combination of the obtained possibilities for mines is performed
using a t-norm, expressing the conjunction of all criteria. Here the product t-norm is used:
. (18)
Finally, since the final possibility should be high if at least one sensor provides a high
possibility and since it is better to assign a friendly object to the mine class than to miss a
mine, the second combination step is performed using the algebraic sum, which is chosen
for its strong disjunctive behaviour:
. (19)
3.3 Decision
The final decision can be simply obtained by thresholding the fusion result for M and
providing the corresponding possibility degree as the confidence degree. As presented in
(Milisavljević & Bloch, 2008), an alternative for the final decision making is to derive the
combination rule for F as well, compare the final values for M and F and derive an adequate
decision rule.
4. Fuzzy approach to close-range detection

In this section, we propose to introduce the fuzziness in the proposed approach. Fuzziness
may be indeed attached to the measurements. An example where classes are prone to
fuzziness is presented in the next section.
4.1 Fuzzy modelling of the measures

In our previous work (Section 3), all measures derived from the sensor data were considered
as crisp numbers. However, they may suffer from imprecision, due to the procedures used
to extract these measures. As a typical example, a threshold is applied on the IR images and
measured are derived from the resulting binary regions. Although the threshold values are
chosen with care, varying them by a few units may lead to variations in the derived
measures.
As an original feature of the present work, we propose to model this imprecision in the
fuzzy set framework and consider all measures as fuzzy numbers and not crisp ones. We
detail the modeling step in this section, while introducing the fuzzy numbers in the fusion
methods will be described in the next one.
The imprecision can be introduced at several levels, directly on the measures, or on the
derived possibility values.
Modeling the measures as fuzzy numbers. Let us consider for instance the elongation
measures. The location of the bordering pixels and the center of gravity may be subject to
imprecision (for example, depending on the threshold). Let a be the minimum distance and
b the maximum distance. We could estimate the typical imprecision (e.g. 2 pixels), e.g., by
varying the threshold and measuring the induced variation on a and b. This could then
define the 0.5-cut of a fuzzy number representing a (respectively b), as illustrated in Figure 1.
We can choose simple shapes for membership functions, such as triangles, which are
sufficient to represent the imprecision and only require two values to be defined: the actual
measured value (a0) in this example, and the typical imprecision ε. In this example, this leads
to the following definition of the fuzzy number representing a:
(20)
Note that we should restrict the support to admissible values for the considered measure.
For instance if the measure has to be in [0, 1], then the support of the fuzzy number should
be included in [0, 1]. Otherwise we could have values with non-zero membership values (i.e.
considered as possible), while they are actually not admissible.
Fig. 1. Example of fuzzy number representing a

A similar model can be estimated for b, leading to a fuzzy number µb.

Then r1 can be derived using the extension principle (Zadeh, 1975).
Let us recall the extension principle in the case of an operation applying on fuzzy numbers.
Let µ1 and µ2 be the membership functions defining two fuzzy numbers (i.e., functions from
the real line into [0, 1], such that all α-cuts are closed intervals, and having a unique modal
value). Let  be an operation between two numbers to be extended to fuzzy numbers. The
resulting fuzzy number µ is defined as:
(21)
Other extensions principles can be defined, using for instance the product instead of the
minimum (Dubois & Prade, 1980).
In our particular case, we can derive a fuzzy model for r1 from µa and µb as follows:
(22)
Alternatively, we can directly model r1 as a fuzzy number. Indeed, the variations of a and b
depending on the thresholds are probably correlated (for instance if a increases because the
region has been expanded by the thresholding, then b increases too), and this is not taken
into account in the previous model.
We reason in a similar way for all measured quantities, in order to obtain fuzzy numbers for
all of them.
Summarizing, a general procedure for modeling a fuzzy number corresponding to a
measure x derived from sensor data is as follows:
1. Let x0 the actually measured value.
2. Let εx the typical imprecision on this measure.
3. Define a fuzzy number µx as a triangular fuzzy set on the real line with
support [x0 - εx, x0 +εx] and modal value x0 .
4. If needed, reduce the support to guarantee its inclusion in the set of
admissible values for x.
This model is simple but robust enough. The most important feature is the ability to account
for imprecision. The shape of the membership function does not need to be precisely
estimated.
Introducing imprecision directly in the possibility values. An alternative way for

introducing the imprecision is to model directly the possibility values as fuzzy numbers.
This approach avoids frequent calls to the extension principle and leads to simpler modeling
and faster computation.
Based on the imprecision estimated for each measure, it is easy to estimate the derived
imprecision on the possibility values, even in a rough way (again this will be sufficient to
achieve a good robustness). For instance the imprecision on π1I(MR) will be of the same
order of magnitude as the imprecision on r1 and r2. So this possibility degree can be directly
modeled as a fuzzy number of triangular shape with min(r1, r2) as modal value and a
support of length 2ε, where ε is derived from the variability on r1 and r2 when the threshold
value is varied, as explained above. For other possibility values, the imprecision should be
computed by applying the same functions as for the measures (e.g. exponential, cosh, etc.).
Then the modeling steps are as follows:

1. Compute the modal values for each possibility exactly as in Subsection 3.1.
2. Introduce imprecision by defining triangular shaped membership functions, using
these modal values and the typical imprecision attached to each measure derived
from the data.
4.2 Fusion based on fuzzy input

For the fusion step, we propose to rely on the same approach as in Subsection 3.2, but
extending it to account for fuzzy input values. Then all π and m values become fuzzy values,
by using the extension principle, and their combination as well.
For instance, if r1 and r2 are fuzzy numbers defined by membership functions and ,
then Eq. 1 becomes:
(23)
Note that with this definition, the minimum of two fuzzy numbers is not necessarily one of
the two initial fuzzy numbers, as illustrated in Figure 2.
Fig. 2. Minimum of two fuzzy numbers, the result is in red.
However, a nice and useful property of this operation is that the support of the result is
always included in the union of the supports of the two fuzzy numbers. In this particular
example, where a possibility degree is derived from the minimum of two values in [0, 1], we
get a fuzzy possibility degree, having a support in [0, 1], which is consistent.
Let us now consider Eq. 2. For computing 1 minus a fuzzy value, we will consider a
membership function defining the crisp value 1 as follows: µ1(1) = 1, µ1(λ) = 0 ∀λ ≠ 1. Then
Eq. 2 can be extended using the extension principle applied to µ1 and π1I(MR). Here it takes
the following simple form:
(24)
All other combinations can be defined in a similar way.
4.3 Decision
The same decision rules can be applied as in the original model (summarized in Subsection
3.3). Decision consists of both an answer on whether the considered object is a mine or a
friendly object, and a confidence value. Here this confidence value can either be kept fuzzy
or a crisp value can be obtained by defuzzifying the resulting possibility value (which are
fuzzy numbers), for instance by using the center of gravity of these fuzzy quantities. Let µ
the membership function of a resulting value, having a support [λ1, λ2]. Its defuzzification
using the center of gravity is expressed as:
(25)
Alternative methods are possible as well (e.g. taking the value having the highest
membership value, etc.).
Note that a defuzzification step could be performed at any stage of the process. Considering
fuzziness all through the process avoids taking crisp decisions on the possibility values at
too early stages, but also leads to heavier computation.
5. Fuzzy approach to remote sensing mined area reduction

5.1 SMART system
The goal of area reduction is to determine which mine-suspected areas do not contain
mines. This task is recognized as a mine action activity that should result in reduction in
time and resources. Several well-known methods are in use to perform area reduction,
especially using mechanical means. Most of the time, these expensive methods change and
damage the environment and the ecosystem. In order to avoid this, some approaches have
been developed that acquire the necessary information remotely, from air or space, using
appropriate sensors associated with context information collected from the field and
integrated in a geographical information system (GIS). One of these projects is the SMART
project (Yvinec, 2005), funded by the European Commission/DG/INFSO and applied to
Croatia. The aim of this project is to provide the human analyst with the SMART system.
This system is GIS-based, augmented with dedicated tools and methods designed to use
multispectral and radar data in order to assist in human analyst’s interpretation of the
possibly mined scene during the area reduction process. The usefulness of such image
processing tools in helping photo-interpretation is in the possibility to process automatically
a large amount of data and help a visual analysis (SMART consortium, 2004). The use of
SMART includes a field survey and an archive analysis in order to collect knowledge about
the site, a satellite data collection, a flight campaign to record the data and the exploitation
of the SMART tools by an operator to detect indicators of presence or absence of mine-
suspected areas. With the help of a data fusion module based on belief functions and fuzzy
sets, the operator prepares thematic maps synthesizing all the knowledge gathered with
these indicators. These maps of indicators can be transformed into risk maps showing how
dangerous an area may be according to the location of known indicators and into priority
maps indicating which areas designed to help the mined area reduction process.
Figure 3 illustrates the global SMART approach. We focus here on the fusion step, which
provides an intermediary result in SMART, consisting in improved land-cover classification
maps, along with confidence values. This result is exploited by the deminers together with
the final result.
5.2. Data and their specificities in SMART

The available images include SAR, multispectral, high resolution optical and satellite data.
SAR data were collected with the E-SAR system of the German Aerospace Centre (DLR) in
fully polarimetric P- and L-band and in vv-polarization (waves are vertically transmitted
and received) X- and C-band. Multispectral Daedalus data were collected with a spatial
resolution of 1 m and in 12 channels, ranging from visible blue to thermal infrared. SAR and
Daedalus data were geocoded. DLR also provided a complete set of RMK photographic
aerial views recorded with a colored infrared film at a resolution of 3 cm. This non-
geocoded data set is used as evidence to control the processing tools and for qualitative
interpretation by photo-interpreters. Finally, geocoded KVR-1000 black-and-white satellite
images with a resolution of 2 m, recorded before the war in Croatia, were purchased in
order to assess the changes in the landscape due to the war.
The legend (expected classes in the images), derived based on the existing and gathered
knowledge about the mined areas, is given in Table 1. Ground truth was provided as a set of
regions (training regions and validation regions). In the fusion module, training regions are
used for estimating the parameters of some of the proposed methods; validation regions are
used for the evaluation of the results.
Table 2 summarizes the inputs of the fusion module. More information about these inputs
can be found in (Bloch, I. et al., 2007).
Contextual info (anomaly description, legend, training set…)

Contextual info (sensors,
agricultural data…)
Data Contextual info
Det1 (topography, MIS…)
…
DetM Over-
Fusion view Danger
Pre- image
Cl1 map
Processing …
/ Pre-
classification ClN SMART
operator
ChD Decision
Output: confidence map for
each class and potential indicator
Fig 3. The global SMART approach; Det – detector, Cl - classifier, M – number of detectors,
N – number of classifiers, ChD – change detection, MIS - mine information system.
Class no. Legend

1 Abandoned agricultural land
2 Agricultural land in use
3 Asphalted roads
4 Rangeland
5 Residential areas
6 Trees and shrubs
7 Shadow
8 Water
Table 1. Expected classes in the images
Data type Type of result

SAR Classification with confidence images per class
SAR & Daedalus Detection of hedges, trees, shadows, rivers, with confidence
degrees, sometimes discounted
Daedalus Supervised classification, result as a decision image
Daedalus Region-based classification with confidence images per class
Daedalus Belief function classification with confidence images per class
SAR & Daedalus Binary detection of roads
SAR River detection (binary)
Daedalus & KVR Change detection (binary image)
Table 2. Summary of the input of the fusion module
5.3. Fuzzy fusion approach in SMART

The computations here are performed at pixel level. A final regularization step is then
applied. Different fusion strategies have been developed (Bloch et al., 2007) and we present
here the most promising one, a fuzzy approach.
In this approach, we choose for each class the best classifiers, based on the confusion matrix
for each classifier or detector, by comparing the diagonal elements in all matrices for each
class. Then, for each class, we combine the best classifiers with a maximum operator
(possibly with some weights). Finally, the decision is made according to a maximum rule.
This approach is very fast. It uses only a part of the information, which could also be a
drawback if this part is not chosen appropriately. Some weights have to be tuned, which
may need some user interaction in some cases. Although it may sound somewhat ad hoc, it is
interesting to show what we can get by using the best parts of all classifiers.
After this first combination step, the next step which follows is knowledge inclusion. It is
one of the main powers of our algorithms with respect to the commercial ones. This aspect
has led to a lot of work in SMART, at different levels. Note that knowledge on the classifiers,
their behaviors, etc. is already included in the previous steps. At this step, we use only the
pieces of knowledge that directly provide information on the landcover classification. Other
pieces of knowledge such as mine reports, etc. are not directly related to classes of interest,
but rather to the dangerous areas, and are thus included in the danger map construction,
which follows the fusion. Several pieces of knowledge proved to be very useful at this step.
For example, some detectors are available for roads and rivers, which provide areas or lines
that surely belong to these classes. There is almost no confusion, but some parts can be
missing. These detections can be imposed on the classification results. As for roads,
additional knowledge is used, namely on the width of the roads (based on observations
from the field missions). Since the detectors provide only lines, they are dilated by the
appropriate size, taking into account both the actual road width and the resolution of the
images. Another type of knowledge is also very useful: the detection of changes between
images taken during the project and KVR images obtained earlier. The results of the change
detection processing provide mainly information about class 1, since they exhibit the fields
which were previously cultivated, and which are now abandoned. These results do not
show all regions belonging to class 1, but the detected areas surely belong to that class, so
these results can also be imposed to the classification results.
The final step is regularization, as it is very unlikely that isolated pixels of one class can
appear in another class. Several local filters were tested, such as a majority filter, a median
filter, or morphological filters, applied on the decision image. A Markovian regularization
approach on local neighborhoods was tested too.
The results of this fusion in its basic version are already very good (Bloch et al., 2007), due to
the fact that not all information provided by the classifiers is used, but only the best part of
them. Further improvements are obtained by knowledge inclusion and regularization.
6. Conclusion
Two main humanitarian action types, close-range antipersonnel mine detection and remote
sensing mined area reduction, are very sensitive tasks as they both deal with human lives,
which puts high demands to the quality of the sensors involved, their efficiency, robustness,
detection rate... Due to the variety of situations in which mines and minefields can be found,
referring to different types of terrain, of mines, of mine laying, of weather conditions etc., it
is hardly possible that one sensor can cover all these situations while reaching the high
detection performance needed. Thus, a solution is being sought in multi-sensor fusion
systems. Taking into account that typically there is a good knowledge about these two
problems, while it is difficult or even impossible to come up with some statistically relevant
training data, due to a high variability of the conditions in which mines and minefields can
be found, knowledge-based theories are useful. Among these theories, possibilistic and
fuzzy theory provides additional advantages, discussed in this chapter, such as different
fusion operators depending on the information and its characteristics. This flexibility turns
out to be very helpful for the two humanitarian action types due to different characteristics
of the measures and sensors that are combined.
The application of fuzzy and possibility theory to humanitarian mine action presented in
this chapter also covers a wide range of possible cases, starting from crisp classes
(mines/friendly objects), where we have some uncertainty/imprecision in deriving a class
from the measures, modelled as possibility discributions (Section 3), via the case when
measures themselves are prone to imprecision so fuzzy measures are introduced in the
possibilistic model (Section 4), up to the case where classes are fuzzy so we apply a fuzzy
model (Section 5).
Although the presented modeling is related to sensors used in some of our projects (which
are also sensors that are typically used for these problems worldwide), it is flexible enough
to be easily adapted to new pieces of information about the types of objects and their
characteristics, as well as to new sensors.
7. References
Bloch, I. (1996). Information combination operators for data fusion: a comparative review
with classification. IEEE Trans. Systens, Man and Cybernetics, Vol. 26, pp. 52-67.
Bloch, I., Milisavljević, N. & Acheroy, M. (2007). Multisensor Data Fusion for Spaceborne
and Airborne Reduction of Mine Suspected Areas. XI International Journal of
Advanced Robotic Systems, pp. 173-186, Vol. 4, No. 2.
Capineri, L., Grande, P. & Temple, J. A. G. (1998). Advanced image-processing technique

for real-time interpretation of ground penetrating radar images. Int. J. Imaging Syst.
Technol., Vol. 9, pp. 51-59.
Das, Y. (2006). Effects of soil electromagnetic properties on metal detectors. IEEE Trans.
Geosci. Remote Sens., Vol. 44, No. 6, pp. 1444-1453.
Denoeux, T. (2008). Conjunctive and Disjunctive Combination of Belief Functions Induced
by Non Distinct Bodies of Evidence. Artificial Intelligence, Vol. 172, pp. 234–264.
Dubois, D. & Prade, H. (1980). Fuzzy Sets and Systems : Theory and Applications. Academic
Press, New York.
Dubois, D. & Prade, H. (1985). A review of fuzzy set aggregation connectives. Information
Sciences, Vol. 36, pp. 85-121.
Dubois, D. & Prade, H. (2004). Possibilistic logic: a retrospective and prospective view. Fuzzy
Sets and Systems, Vol. 144, pp. 3-23.
Dubois, D. & Prade, H. (2006). Possibility theory and its applications: a retrospective and
prospective view. In: Decision theory and multi-agent planning (CISM International
Centre for Mechanical Sciences), G. Della Riccia et al. (Eds.), Springer, pp. 89-100.
Fischer, C. et al. (2007). Detection of antipersonnel mines by using the factorization method
on multistatic ground-penetrating radar measurements. IEEE Trans. Geosci. Remote
Sens., Vol. 45, No. 1, pp. 85-91.
Haralick, R.M. & Shapiro, L.G. (1992). Computer and Robot Vision. Boston, MA: Addison-
Wesley Longman Publishing Co., Inc.
Milisavljević, N. & Bloch, I. (2003). Sensor fusion in anti-personnel mine detection using a
two-level belief function model. IEEE Trans. Systems, Man and Cybernetics, Part C,
Vol. 33, No. 2, pp. 269-283.
Milisavljević, N.; Bloch, I.; van den Broek, S.P. & Acheroy, M. (2003). Improving mine
recognition through processing and Dempster-Shafer fusion of ground-penetrating
radar data. Pattern Recognition, Vol. 36, No. 5, pp. 1233-1250.
Milisavljević, N. & Bloch, I. (2008). Possibilistic vs. Belief Function Fusion for Anti-Personnel
Mine Detection. IEEE Trans. Geosci. Remote Sens., Vol. 46, No. 5, pp. 1488-1498.
Shafer, G. (1976). A Mathematical Theory of Evidence. Princeton, NJ: Princeton Univ. Press.
SMART consortium (2004). SMART - final report.
Smets, P. (1994). What is Dempster-Shafer’s model?. In: Advances in the Dempster-Shafer
Theory of Evidence, R. R. Yager, M. Fedrizzi and J. Kacprzyk (Eds). New York: Wiley,
pp. 5-34.
Smets, P. & Kennes, R. (1994). The transferable belief model. Artificial Intelligence, Vol. 66, pp.
191-234.
Thonnard, O. & Milisavljević, N. (1999). Metallic shape detection and recognition with a
metal detector. in Proc. European Workshop on Photonics applied to Mechanics and
Environmental Testing Engineering (PHOTOMEC'99 - ETE'99), Liège, Belgium, pp.
99-103.
Yvinec, Y. (2005). A Validated Method to Help Area Reduction in Mine Action with Remote
Sensing Data, Proc. of the 4th Int’l Symposium on Image and Signal Processing and
Analysis (ISPA 2005), Zagreb, Croatia.
Zadeh, L. (1965). Fuzzy sets. Information and Control, Vol. 8, No. 3, pp. 338-353.
Zadeh, L.A. (1975). The Concept of a Linguistic Variable and its Application to Approximate
Reasoning. Information Sciences, Vol. 8, pp. 199–249.
Non specular reflection and depolarisation due to walls under oblique incidence 505
24
X
Non specular reflection and depolarisation due

to walls under oblique incidence
Iñigo Cuiñas, Manuel G. Sánchez and Ana V. Alejos
Universidade de Vigo
Spain
1. Introduction
Bi static radar remote sensing has became a large source of data intended to analyse the
composition and the behaviour of targets placed very far away from the sensor. The
possibility of determining the nature of the surface in which the electromagnetic waves have
impacted, and have been reflected by, is directly correlated to the knowledge of high
precision scattering models. In most of the situations, the Fresnel coefficients for reflection
have been the only procedure used to obtain the electromagnetic characteristics of the target
material from the reflection this surface induced in the propagating wave. That
electromagnetic characterisation is quite important in remote sensing techniques, as it
depends on the frequency and, so that, it could help in defining the spectral signature of
each target material.
Historically, several models have been proposed in order to improve the performance of the
simple Fresnel coefficients to obtain the dielectric constant of an obstacle. The aim of such
models is to define the outcomes of the incidence of a propagating wave on a constructive
surface (Balanis, 1989; Burnside & Burgener, 1983; Correia & Françês, 1995; Sato et al., 1997).
When the reflective phenomenon is limited to the specular direction, it is commonly known
as specular reflection; whereas when several directions are taken into account, the reflection
concept (or, in general terms, scattering) is extended to all directions in the incidence region.
Any obstacle generates its own reflection pattern, which depends on the electromagnetic
characteristics of the material, the surface roughness, the frequency, and the angle of
incidence. Published measured results show that there may be several scattering directions
as significant as the main specular reflection one. This result leads to conclude that the
received power in the specular direction could not represent the main contribution among
the reflections on the surface under study. Thus, computing the dielectric constant of the
surface material from just the data reflected towards the specular direction could lead to
large mistakes.
The specular reflection on flat or almost flat obstacles can be considered as a well studied
propagation mechanism, and several models have been enunciated to describe its
performance. The interest of specular direction resides in the need of exact models and fast
algorithms that are used in the radio electric planning tools. The non-specular or general
reflection, although less taken into account in practice, is also a topic that focuses several
works (Beckmann & Spizzichino, 1963-1987; Cuiñas et al., 2007). These results have capital
importance when the application is remote sensing and the objective is to detect and even to
identify possible targets by means of their scattering patterns.
A phenomenon associated to reflection, the depolarisation that could be generated when a
wave beats a flat obstacle, appears to be not so fine defined and modelled. The reason of this
lack of interest is probably because the typical application of reflection models has been the
radio planning tools. These tools were designed for frequencies bellow millimetric bands,
assigned to cellular phone or television broadcasting, and the typical obstacles (walls) are
electrically flat enough to provide strong specular reflections at these frequencies. At higher
frequencies, the electrical size of a given obstacle becomes larger, and the specular
reflections could not be so dominant among the complete scattering arc angles.
Although there are different published research works on depolarisation, they are
commonly centred at lower frequencies, and they often treat on complete radio channels,
instead of the analysis of isolated obstacles. Several of these works reports the polarisation
diversity gain measured at indoor radio channels: up to 15 dB at UHF band (Sánchez &
Sánchez, 2000), and from 4 to 10.5 dB at 2.05 GHz (Dietrich et al., 2001); or even in outdoor
environments: between 11 to 5.2 dB at 1800 MHz (Turkmani et al., 1995), depending on the
type of environment: urban, suburban or rural. The contents of this chapter are not
comparable to the previous work, as they are oriented to analyse the radio channel,
including multipath propagation, whereas the aim of the presented work is the study of
isolated flat obstacles.
This chapter summarises the theoretical approaches to the study of scattering and
depolarisation generated by flat isolated obstacles, and it also outlines measurement results
obtained by the authors at 5.8 GHz, involving three different constructive materials and
taking into account two orthogonal polarisations for the incident wave, which comes from
several angles of incidence to the obstacle.
The chapter is organised as follows. The contents of the section 2 are the theoretical basis of
this work: an overview of specular reflection models, which are used to extract the
electromagnetic parameters of each considered material, and the exposition of the Physical
Optics model used to compute the reflection patterns.
The depolarisation indexes computation procedure is introduced in section 3. This section
contains the discussion of a methodology to determine the amount of cross-polarised wave
that could be generated by reflection on the surface of an obstacle in the radio channel. The
proposed matrix formulation, which is inspired in the polarimetric matrix (Mott, 1986),
allows the separation between the depolarisation due to the antennas and that produced by
the obstacle. The depolarisation indexes indicate the fraction of the signal power that is
depolarised after beating the obstacle.
The section 4 shows the results, comparing the behaviour of the Physical Optics model
application when it simulates reflection patterns in similar geometrical conditions that
previously measured. Finally, the section 5 contains the conclusions.
2. Scattering coefficients
The specular reflection is just a simplification of the complete, and complex, effect of walls
over propagating waves. When a wave beats a wall, a scattering phenomenon is generated
towards all space directions, defining a reflection pattern. However, specular reflection
models are very useful to obtain an initial approach to the characteristic parameters, as
permittivity and conductivity, which drive the electromagnetic behaviour of the material.
These parameters could be used as input data to those more complex scattering models.
The scattering due to surfaces of any kind is a classical area of study that has its biggest
impetus with the advent of radar, in the middle of 20th century. There are several models to
predict the behaviour of a surface when an incident wave beats it (Ruck et al., 1970). The
strategy to model their effects on the propagating wave depends on the type of roughness.
Besides, the scattering by perfectly conductive surfaces is a problem analysed by several
reflector antenna methods. These problems are tried to be solved by employing high
frequency methods, which are valid when the reflector size (in our case, the obstacle size) is
large in terms of wavelength (Scott, 1990). A combination of both radar ideas and antenna
analysis methods has been applied in the work summarised at this chapter to model the
reflection due to constructive walls. The formulation, based on Physical Optics, is intended
to compute scattering patterns due to flat and rough surfaces, both dielectric and
conductive.
2.1 Specular reflection coefficients

The behaviour of a wave when it beats an obstacle mainly depends on its permittivity. This
behaviour has been modelled by different methods (Landron et al., 1993, 1997; Lähteenmäki
&Karttavi, 1996; Cuiñas et al., 2001), most of them based on classical Fresnel formulation.
This Fresnel theory works consistently when the obstacle thickness is electrically large
and/or the material losses are huge. Typical walls are not infinite thickness, and this is the
reason because several models have been enunciated to take into account this situation.
Among the models based on Fresnel method, the internal successive reflections (ISR) model
must be mentioned: it tries to explain transmission and reflection phenomena as a result of
the coherent sum of several multipath components, generated in both boundaries between
the obstacle and the free space (Burnside & Burgener, 1983). Although this model assumes
that the slab surface is infinite, the results are adequate if the material sample is wide
enough to contain the first Fresnel ellipsoid of the radio link.
2.2 Physical Optics model

A good characterisation of scattering due to rough or slightly rough surfaces must include
all-direction effects. The Physical Optics formulation fits this condition, and it is a classical
method to characterise conductive surfaces (Beckmann & Spizzichino, 1963-1987).
The electric field strength, as well as its derivative respect to surface normal, could be
estimated by using Kirchhoff approximations. Thus, it can be assumed that the total field in
a point on the surface is the same as in a tangent plane to the surface at the same point. The
larger the roughness curvature radius is, the better the approximation is.
Assuming an electrically large obstacle, the Physical Optics model leads to a single equation
to define the reflection coefficient, the general formulation of the scattering coefficient.
1 L
ρ=
4 L cos θ i ∫
−L
(aξ '−b) ⋅ e jvx x+ jvzξ dx (1)
There is no general, exact and explicit solution for finite conductivity surfaces, but some
approximations can be applied, which performance is better when the rough surface
presents soft slopes, or large radii of curvature: in other words, when we manage rough, but
smooth enough surfaces.
As a general conclusion, it can be said that scattering due to non-conductive surfaces is
affected by finite conductivity only when local reflection coefficients are more influenced by
its local incident angle, than by the electromagnetic properties of the scatterer.
2.3 Application of the model

The application of this formulation is suggested for three different strategies. Successive
proposals have growing computational cost and complexity, but they give better
concordance to actual situations (Cuiñas et al., 2007). Figure 1 depicts the three strategies.
Fig. 1. Strategies to apply the Physical Optics formulation: A, direct; B, by segments; C, 3D
The first option focuses on the direct application of the Physical Optics formulation
considering an incidence field on a flat surface, with an incidence angle determined by
direct propagation path between the transmitter antenna and the centre of illumination. This
case does not consider the radiation pattern of the antenna.
As the field strength over the illuminated area is not uniform, but varies with the angle from
the centre of the beam according to the gain pattern of the antenna, a second option is
proposed. This strategy consists in dividing the surface into several parallel segments of the
same width and assumed infinite length. The local angle of incidence at each segment is
considered when applying Physical Optics method. Thus, part of the effect of the antenna
radiation pattern can be considered, as local situations are taken into account. This method
may be a good choice for flat or for one-dimensional rough or periodic surfaces.
The third option consists of determining the illuminated surface on the obstacle, taking into
account the radiation pattern of the transmitter antenna. That radiation pattern determines a
footprint on the obstacle surface, with elliptic shape. This footprint on the surface is divided
into several patches. The Physical Optics formulation is applied on each patch, considering
the local incidence angle at each of them. In this situation, the application of the algorithm
needs a reformulation, as the local scattered electric field has to be computed at each patch,
and then all these contributions have to be coherently combined to compute the complete
scattered field at any scattering direction. This idea is based on antenna analysis techniques,
as (Arias et al., 1996).
3. Depolarisation indexes
The depolarisation index, for any material, at any angle of incidence and any polarisation of
the transmitted waves could be defined as the fraction of the power of this wave that is
received in the orthogonal polarisation. From this definition, depolarisation indexes may be
computed by means of a matrix procedure. The proposed method provides a set of
depolarisation indexes that characterise the reflection mechanism generated when a wave
reach an obstacle with oblique incidence (Cuiñas et al., 2009).
The definition of the matrix model begins in the characterisation of the radio channel
following the Friis formula, but taking into account two orthogonal polarisations,
considering the presence of an obstacle in the radio path, and assuming that both
transmitting and receiving antennas are pointed to the same spot on the reflecting surface.
Then, the transmitting antenna is illuminating the reflection point with its maximum gain,
and the receiving antenna is getting the waves from the obstacle by its maximum gain
direction. The free space attenuation coefficient could be computed by substitution of the
obstacle by a perfectly conductive surface.
With the model equation, the definition of the obstacle behaviour is completely independent
from the effect of the free space propagation in the open links between the transmitter and
the obstacle, and between the obstacle and the receiver. The matrix AOBS contains the data
corresponding to the obstacle effect on the complete radio channel. Figure 2 depicts the
physical significance of each element in the matrix.
aVV
OBS
(θinc ,θ obs ) aVH
OBS
(θinc ,θ obs )
A OBS
(θinc ,θ obs ) =  OBS  (2)
aHV (θ inc ,θ obs ) aHH (θ inc ,θ obs )
OBS
The computation of depolarisation indexes is difficult, in general terms, because of the

separation between the depolarisation due to the obstacle and the depolarisation due to the
antennas is needed. Nevertheless, this problem could avoided with the matrix formulation
that has been commented, as the elements of the obstacle matrix (AOBS) are only related to
the obstacle, and they do not include other effects. The depolarisation indexes can be
defined from the obstacle matrix elements as:
DI H (θ inc ,θ obs ) =
OBS
aHV (θinc ,θ obs )
aHV (θ inc ,θ obs ) + aHH
OBS OBS
(θinc ,θ obs ) (3)
aVH (θ inc ,θ obs )
OBS
DIV (θ inc ,θ obs ) = OBS
aVH (θ inc ,θ obs ) + aVV
OBS
(θinc ,θ obs )
where the sub index indicates vertical or horizontal incidence (transmission). As these
indexes are computed from AOBS elements, each pair of θinc and θobs would define their
associated pair of depolarisation indexes.
Fig. 2. Definition of matrix elements: A, vertical incidence; B: horizontal incidence

4. Results
The wellness of the proposed methods has been checked by comparing to measurement
outcomes performed at 5.8 GHz (Cuiñas et al., 2007). The exposition of the results begins
with the Physical Optics formulation: measured scattering patterns are compared with those
computed by applying Physical Optics formulation, following the three described strategies.
Then, the depolarisation indexes are computed.
4.1. Reflection pattern

The Physical Optics formulation has been applied following the three strategies previously
mentioned. Direct application of the model provides reflection patterns with a very narrow
main lobe and many smaller side lobes, clearly different from the measured patterns.
When the surface is perfectly conductive, the application of the Physical Optics model by
segments leads to a simulation error (compared to measurement outcomes) similar to the
direct use of Beckmann formulation, whereas the use of the elliptic patching (3D) model
provides an improvement of around 57% in the simulated results.
The results for dielectric obstacles seem to indicate better fitting when the application takes
into account the geometry of the problem; i.e. when the illuminated surface is patched in
concentric rings and local contributions at each patch are coherently added at each
scattering direction. This appreciation is supported by the reduction percentage in the
relative simulation error obtained by applying the Physical Optics theory by segments or by
elliptic patching compared to that obtained by the direct application of the Physical Optics
formulation. The 3D application of the integral equation led to an improvement of the
relative error of 1.5% for a brick wall obstacle, and 67% for a highly reflective smooth chip
wood panel, whereas the segment application improvements were 0.2% and 62%,
respectively. This result confirms the visual appreciation: the more specularly reflective the
material is, the better the improvement of Physical Optics 3D method is.
4.2. Depolarisation indexes

The difference between co polar and cross polar relative powers presents strong variations
in measurement results: in the vicinity of the specular direction it is commonly larger than
in other directions of observation, in which the cross polar component could be even
stronger than the co polar. This appreciation confirms that the non-specular directions could
not be ignored when planning the network.
The difference between co polar and cross polar received power, observed towards the
specular direction could be up to 58 dB, for brick walls, 28 dB for chip wood panels, and 31
dB for stone walls, at 5.8 GHz (Cuiñas et al., 2009).
Once computed the depolarisation indexes, towards the specular direction, they resulted to
be up to 0.75% for the brick wall, up to 9.8% for the chip wood panel, and up to 9.27% for
the stone wall. This indicates that brick wall provides reduced depolarised waves compared
to the co polar reflected waves in the specular direction.
But in a general case, all scattering directions have to be considered, and not just specular
one, as the reflector could be randomly located and oriented. With this aim, median
depolarisation indexes for each material could be useful, being 23% and 30% for brick wall,
18% and 18.5% for chip wood panel, and 4.5% and 4% for stone wall, in horizontally and
vertically polarised incident wave, respectively, at 5.8 GHz. Once several angles of
observation are introduced, not just the specular ones, the depolarisation indexes grow, and
differences between incident polarisations appear in the brick wall case. The brick wall is the
more non isotropic material among the considered, as it presents a clearly oriented
structure, whereas the chip wood panel and the stone wall are the result of the solidification
of a mass, which is expected to present a more isotropic behaviour. The large median values
of depolarisation indexes indicate that high depolarised waves could be generated when
several scatterers are present in an environment, which is the case of indoor scenarios.
5. Conclusions
Two methods to characterise an obstacle in terms of reflection and depolarisation, and the
different strategies to apply these formulations, have been summarised along this chapter,
providing the references to in depth study both formulation. The behaviour of such
methods have been also compared to measurement results at 5.8 GHz, using three different
flat obstacles, with more or less rough surfaces, and two orthogonal linear polarisations in
the incident wave.
The modelling of scattering patterns generated by flat conductive and dielectric obstacles is
proposed, based on the Physical Optics formulation. Among three possible strategies of
implementation, that based on elliptic patching the illuminated area on the obstacle surface
appears to provide more accurate results, since it is better adapted to the real problem: it
defines the finite surface in a more precise way and it takes into account the radiation
pattern of the antennas employed in the experimental work. The comparison between
simulation results and actual situation measurements shows the good behaviour of the
algorithm. The flatter and more conductive the material is, the better the algorithm works.
When observing the depolarisation, the measured results reflect differences between co
polar and cross polar received powers from 17 to 58 dB in specular directions of observation,
depending on the obstacle and the angle of incidence. But the most interesting observations
are in the non-specular direction, in which the cross polar received power could be even
stronger than the co polar, and with amplitudes that could not be ignored. These situations
indicated that up to more than 50% of the incident wave could be scattered as a cross
polarised wave.
The results of this experimental work were processed by means of a matrix model to
compute the depolarisation indexes. Indexes up to 10 % have been obtained in the specular
direction, and median values up to 30% are reached when considering the entire scattering
region.
6. References
A.M. Arias, M.E. Lorenzo, A.García-Pino, “A novel fast algorithm for Physical Optics
analysis of single and dual reflector antennas”. IEEE Transactions on Magnetics,
vol.32, pp. 910-913, May 1996.
C.A. Balanis, “Advanced engineering electromagnetics”, John Willey & sons, 1989.
P. Beckmann, A. Spizzichino, The scattering of electromagnetic waves from rough surfaces, Artech
House, 1963-1987.
W.D. Burnside, K.W. Burgener, “High frequency scattering by a thin lossless dielectric slab”,
IEEE Transactions on Antennas and Propagation, vol. Ap-31, no. 1, pp. 104-110,
January 1983.
L.M. Correia, P.P. Françês, "Transmission and isolation of signals in buildings at 60 GHz",
In: International symposium on personal, indoor and mobile radio communications,
Toronto (Canada), September 1995.
I. Cuiñas, J.-P. Pugliese, A. Hammoudeh, M. García sánchez, “Frequency dependence of
dielectric constant of construction materials at microwave and millimeter wave
bands”, Microwave and Optical Technology Letters, vol. 30, no. 2, pp. 123-124, July
2001.
I. Cuiñas, D. Martínez, M. García Sánchez, A. Vázquez Alejos, “Modelling and measuring
reflection due to flat dielectric surfaces at 5.8 GHz”, IEEE Transactions on Antennas
and Propagation, vol. 55, no. 4, pp. 1139-1147, April 2007.
I. Cuiñas, M. García Sánchez, A. Vázquez Alejos, “Depolarization due to scattering on walls
in the 5 GHz band”, IEEE Transactions on Antennas and Propagation, June 2009.
C.B. Dietrich, K. Dietze, J. Randall Nealy, W.L. Stutzman, “Spatial, polarisation, and pattern
diversity for wireless handheld antennas”, IEEE Transactions on Antennas and
Propagation, vol. 49, no. 9, pp. 1271-1281, September 2001.
A.K.Fung, “Scattering and depolarisation of EM waves from a rough surface”, Proceedings of
the IEEE, pp. 395-396, March 1966.
T. Kurner, D.J. Cichon, W. Wiesbeck, "Concepts and results for 3D digital terrain-based
wave propagation models: an overview", IEEE Journal on Selected Areas in
Communications, vol. 11, no. 7, September 1993.
J. Lähteenmäki, T. Karttavi, “Measurements of dielectric parameters of wall materials at 60
GHz band”, Electronics Letters, vol. 32, no. 16, pp. 1442-1444, August 1996.
O. Landron, M.J. Feuerstein, T.S. Rappaport, “In situ microwave reflection measurements
for smooth and rough exterior wall surfaces”, In: IEEE 43rd Vehicular Technology
Conference, pp. 77-80, May 1993.
O. Landron, M.J. Feuerstein, T.S. Rappaport, “A comparison of theoretical and empirical
reflection coefficients for typical exterior wall surfaces in a mobile radio
environment”, IEEE Transactions on Antennas and Propagation, vol. 44, no. 3, pp. 341-
351, March 1996.
M. Lebherz, W. Wiesbeck, W. Krank, "A versatile wave propagation model for the
VHF/UHF range considering three-dimensional terrain", IEEE Transactions on
Antennas and Propagation, vol. 40, no. 10, pp. 1121-1131, October 1992.
H. Mott, Polarisation in Antennas and Radar, John Wiley & Sons, New York, 1986.
G.T. Ruck, D.E. Barrick, W.D. Stuart, C.K. Krichbaum, Chapter 9: “Rough Surfaces”, Radar
cross section handbook, vol. 2 Plenum Press, New York, 1970.
Sato, Katsuyoshi, et al, "Measurements of reflection and transmission characteristics of
interior structures of office building in the 60-GHz band", IEEE Transactions on
Antennas and Propagation, vol. 45, no. 12, pp. 1783-1791, December 1997.
M. Sánchez, M.G. Sánchez, “Analysis of polarisation diversity at digital TV indoor
receivers”, IEEE Transactions on Broadcasting, vol. 46, no. 4, pp. 233-239. December
2000.
C. Scott, Modern methods of reflector antenna analysis and design, Artech House, 1990.
A.M.D. Turkmani, A. Arowjolu, P.A. Jefford, C.J. Kellett “An experimental evaluation of the
performance of two-branch space and polarisation diversity schemes at 1800
MHz”, IEEE Transactions on Vehicular Technology, vol. 44, no. 2, pp. 318-326, May
1995.
Dynamical Enhancement Technique for Geophysical Analysis of Remote Sensing Imagery 515
25
X
Dynamical Enhancement Technique

for Geophysical Analysis of Remote
Sensing Imagery
Iván E. Villalón-Turrubiates
University of Guadalajara Campus Los Valles
Guadalajara–Ameca Highway Km. 45.5, 46600 Ameca Jalisco, Mexico
Telephone (+52 37) 5758-0148 ext. 7267, E-Mail: villalon@ieee.org
1. Introduction
Considerable progress has been made generally in the application of remote sensing tech-
niques to both research and operational problems for urban planning and natural resource
management. Modern applied theory of image processing is now a mature and well devel-
oped research field, presented and detailed in many works.
Although the existing theory offers a manifold of statistical techniques to tackle with the
particular environmental monitoring problems, in many applications areas there still remain
some crucial theoretical and data processing problems. One of them is particularly related to
the extraction and dynamical analysis of physical characteristics (e.g., water, land cover,
vegetation, soil, humid content, and dry content) for implementation in natural resources
management (modeling and planning).
The extraction of environmental physical characteristics from a particular geographical
region through remote sensing data processing allows the generation of electronic signature
maps, which are the basis to create a high-resolution collection atlas processed in time for a
particular geographical zone. This can be achieved using a systematical tool for supervised
segmentation and classification of the environmental remote sensing signatures that em-
ploys multispectral remote sensing imagery based on pixel statistics for the class descrip-
tion. Moreover, the analysis of a dynamical model of environmental characteristics extracted
from a geographic region generates useful information for natural resource management;
using the signatures map extracted from the remote sensing imagery for a particular geo-
graphic zone in discrete time the evolution study of the environmental characteristics is
performed to obtain the dynamical model of the physical variables. This provides a back-
ground for understanding the future trends of the multispectral image. This chapter ex-
plores the implementation possibilities of the multispectral image classification technique
with the dynamic analysis for natural resources management applications.
1.1 Remote Sensing

The goal of science is to discover universal truths that are the same yesterday, today and
tomorrow. Hopefully, the knowledge obtained can be used to protect the environment and
improve human quality of life. To identify these universal truths, scientists observe and
make measurements about (a) the physical world (e.g., the atmosphere, water, soil, rock), (b)
its living inhabitants (e.g., Homo sapiens, flora, fauna), and (c) the process at work (e.g.,
mass wasting, deforestation, urban sprawl).
Scientists formulate hypotheses and then attempt to accept or reject them in a systematic,
unbiased fashion. The data necessary to accept or reject a hypothesis may be collected di-
rectly in the field, often referred to as in situ or in place data collection. This can be a time-
consuming, expensive, and inaccurate process. Therefore, considerable research during the
past century has gone into the development of aerial platforms (e.g., suborbital aircraft,
satellites, unmanned aerial vehicles) and sensors (e.g., cameras, detectors) that can collect
information some remote distance from the subject (e.g., from 10,000 meters above ground
level). This process is called remote sensing (RS) of the environment (Jensen, 2005).
The remote sensor data can be stored in an analog format (e.g., a hardcopy 9x9 in. vertical
aerial photograph) or in a digital format (e.g., remote sensing imagery consisting up to sev-
en registered matrices of brightness values). The analog and digital remote sensing data can
be analyzed using analog (visual) and/or digital image processing techniques.
A science is defined as the broad field of human knowledge concerned with facts held to-
gether by principles (rules). Scientists discover and test facts and principles by the scientific
method, and orderly system of solving problems. Scientists generally feel that any subject
that humans can study by using the scientific method and other special rules of thinking
may be called a science. The science includes (a) mathematics and logic, (b) the physical
sciences, such as physics and chemistry, (c) the biological sciences, such as botany and zool-
ogy, and (d) the social sciences, such as geography, sociology, anthropology, etc. Interes-
tingly, some persons do not consider mathematics and logic to be sciences. But the fields of
knowledge associated with mathematics and logic are such valuable tools for science that
cannot be ignored.
Remote sensing is a tool or technique similar to mathematics. Using sensors to measure the
amount of electromagnetic radiation (EMR) existing an object or geographic area from a
distance and the extracting valuable information from the data using mathematically and
statistically based algorithms is a scientific activity (Jensen, 2005). It functions in harmony
with other spatial data collections techniques or tools of the mapping sciences, including
geographic information systems (GIS) (Fussel et al., 1986).
The process of visual photo or image interpretation brings to bear not only scientific know-
ledge but all of the background that a person has obtained in his lifetime. Such learning
cannot be measured, programmed or completely understood. The synergism of combining
scientific knowledge with real-world analyst experience allows the interpreter to develop
heuristic rules of thumb to extract information from the imagery. Some image analysts are
superior to other image analysts because they (a) understand the scientific principles better,
(b) are more widely traveled and have seen many landscape objects and geographic areas,
and/or (c) have the ability to synthesize principles and real-world knowledge to reach logi-
cal and correct conclusions. The remote sensing image interpretation is both and art and a
science (Jensen, 2005).
Sensors can be used to obtain very specific information about and object (e.g., the diameter
of a cottonwood tree’s crowns) or the geographic extent of a phenomenon (e.g., the poly-
gonal boundary of a cottonwood stand). The EMR reflected, emitted, or back-scattered from
an object or geographical area is used as a surrogate for the actual property under investiga-
tion. The electromagnetic energy measurements must be calibrated and turned into informa-
tion using visual and/or digital image processing techniques.
Urban planners (e.g., land use, transportation, utility) and natural resource management
(e.g., wetland, forest, grassland, rangeland) recognize that spatially distributed information
is essential for ecological modeling and planning. Unfortunately, it is very difficult to obtain
such information using in situ measurements. Therefore, public agencies and scientists have
expanded significant resources in developing methods to obtain the required information
using remote sensing science (Shkvarko & Villalon, 2007).
1.2 Image Interpretation

The classical process of classification consists of two stages (Smith, 2000). The first is the
recognition of categories of real-world objects (segmentation). In the context of multispectral
remote sensing (MRS) of the land surface these categories could include, for example, wood-
lands, water bodies, grassland and other land cover types, depending on the geographical
scale and nature of the study. The second stage is the labeling of the entities (classification)
that are to be classified.
In digital image classification these labels are numerical, so that a pixel that is recognized as
belonging to the class ‘water’ may be given the label ‘1’, ‘woodland’ may be labeled ‘2’, and
so on. The process of image classification requires the user to perform the following steps:
(a) determine a-priori the number and nature of the categories in terms of which environ-
mental remote sensing signatures (RSS) are to be described, and (b) assign numerical labels
to the pixels on the basis of their RSS properties using a decision-making procedure, usually
termed a classification rule or a decision rule.
Clustering is a kind of exploratory procedure, the aim of which is to determine the number
(but not initially the identity) of distinct RSS present in the area covered by the image, and
to allocate pixels to these categories in terms of the nature of the RSS types is a separate
stage that follows the clustering procedure. Several clusters may correspond to a single RSS
type (Smith, 2000).
These two approaches to pixel labeling are known in the RS literature as supervised and
unsupervised classification procedures, respectively (Smith, 2000). They can be used to
segment an image into regions with similar attributes. In the simplest case, a pixel is charac-
terized by a vector whose elements are its grey levels in each spectral band. This vector
represents the spectral properties of that pixel.
In a supervised classification, the identity and location of some of the RSS types (e.g., urban,
agriculture or wetland) are known a priori through a combination of field work, interpreta-
tion of aerial photography, map analysis or personal experience. The analyst attempts to
locate specific sites in the MRS data that represents homogeneous examples of these known
RSS types. These areas are commonly referred to as training sites because the spectral cha-
racteristics of these known areas are used to train the classification algorithm for eventual
RSS mapping of the remainder of the image. Multivariate statistical parameters (means,
standard deviations, covariance matrices, correlation matrices, etc.) are calculated for each
training site. Every pixel both within and outside the training sites is then evaluated and
assigned to the class of which it has the highest likelihood of being a member (Jensen, 2005).
In an unsupervised classification, the identities of RSS types to be specified as classes within
a scene are not generally known a priori because ground reference information is lacking or
surface features within the scene are not well defined. The computer is required to group
pixels with similar spectral characteristics into unique clusters according to some statistical
determined criteria. The analyst then re-labels and combines the RSS spectral clusters into
formation classes (Jensen, 2005).
The term dynamic refers to phenomena that produce time-changing pattern, the characteris-
tics of the pattern at one time being interrelated with those at other times. The term is nearly
synonymous with time-evolution or pattern of change (Luenberger, 1979). Nearly all ob-
served phenomena in our daily lives or in scientific investigation have important dynamic
aspects. Scientific examples may arise in (a) a physical system, such as a signal traveling
through the space, a home heating system, or in the mining of a mineral deposit, (b) a social
system, such as the movement within an organization hierarchy, the evolution of a tribal
class system, or the behavior of an economic structure, and (c) a life system, such as that of
genetic transference, ecological decay, or population growth.
Many dynamic systems can be understood and analyzed intuitively, without resort to ma-
thematics and without development of a general theory of dynamics. However, in order to
approach unfamiliar complex situations efficiently, it is necessary to proceed systematically.
Mathematics can provide the required economy of language and conceptual framework.
The term dynamics takes a dual meaning. It is a term for the time-evolutionary phenomena
in the world about us, and a term for that part of mathematical science that is used for the
representation and analysis of such phenomena (Luenberger, 1979). Dynamic systems are
represented mathematically in terms of either differential or difference equations. These
equations provide the structure for representing time linkages among variables. The use of
either differential or difference equations to represent dynamic behavior corresponds, re-
spectively, to whether the behavior is viewed as occurring in continuous or discrete time.
Continuous time corresponds to our usual conception, where time is regarded as a continuous
variable and is often viewed as flowing smoothly past us (Luenberger, 1979). In mathematical
terms, continuous time of this sort is quantified in terms of the continuum of real numbers.
Discrete time consists of an ordered sequence of points rather than a continuum. In terms of
applications, it is convenient to introduce this kind of time when events and consequences
either occurring are accounted for only at discrete time periods, such as daily, monthly, or
yearly. Accordingly dynamic behavior viewed in discrete time is usually described by equa-
tions relating the value of a variable at one time to the values at adjacent times. Such equa-
tions are called difference equations.
2. Multispectral Image Classification

Multispectral imaging is a technology originally developed for space-based imaging. The
multispectral images are the main type of images acquired by MRS radiometers. Usually,
MRS systems have from 3 to 7 radiometers; each one acquires one digital image (also called
scene) in a small band of visible spectra, ranging 450 nm to 690 nm, called red-green-blue
(RGB) regions (Villalon, 2008).
For different purposes, combinations of spectral bands can be used. They are usually
represented with red (R), green (G) and blue (B) channels. This is referred to as True-Color
RS imagery (Villalon, 2008). The wavelengths for the spectral bands are as follows, where
the values are approximated, exact values depends on the particular RS instruments (Math-
er, 2004): (a) blue: 450-520 nm, (b) green: 520-600 nm, and (c) red: 600-690 nm.
2.1 Weighted Order Statistics method

The WOS method is a generalization of the median filter (Perry et al., 2002), and is charac-
terrized by a weight vector and a threshold value. The order statistics (OS) filtering metho-
dology (Yli-Harja et al., 1991) shifts a n×n window W (with cardinality n×n, i.e., |W| = n×n)
over an input RS image frame and, at each position of the frame, takes the n×n inputs (w11,
w12,…, wij, ..., wnn) under Wij and then outputs the r-th element of the sorted input.
Fig. 1. Weighted order statistics filtering method
The WOS method is a generalization of the OS filter that is characterized by a weight vector
ij = (1, 2, …, n×n) of n×n positive weight thresholds w, 0 ≤ w ≤ 255 (gray-level threshold).
To compute the output, each input w is duplicated to the number of corresponding weight
, then they are sorted and the w-th order element is chosen as the output, expressed as
WOSij  median( Υ ij ) , (1)
where WOSij is the weighted order of the (i, j)-th pixel of the image. The decision rule for
classification based on the WOS filter determines that, based on the a priori information for
class segmentation (number of RSS to be classified and their respective thresholds), the
WOS value for each image pixel is compared with the a priori thresholds (gray-level) and
classified according to the most proximal value. Fig. 1 shows the process structure of the
WOS filter.
2.2 Minimum Distance to Means method

The MDM decision rule is computationally simple and can result in classification accuracy
comparable to other more computationally intensive algorithms (Jensen, 2005). It is charac-
terized by the mean values of the RSS classes and the Euclidean distances based on the
Pythagorean Theorem. An important aspect of this method is that it is applied to the MRS
imagery. The a priori information for class segmentation (number of RSS to be classified and
their respective mean values) conform the means matrix  (c×b size) that contains the mean
values cb: (0 ≤ cb ≤ 255, gray-level) of the RSS classes for every MRS band. Here, c is the
number of RSS classes to be classified, and b is the number of spectral bands.
Fig. 2. Minimum distance to means method
The input is defined by the vector ij, which contains the (i, j)-th image pixel values ijb
employed for every spectral band. To compute the output of the classifier, the distance be-
tween each input ij and the means matrix  is calculated using the Euclidean distance
based on the Pythagorean Theorem. This is expressed as
      
2 2 2
Dij ,c  ij ,1  c 1 ij ,2  c 2     ij , b  cb , (2)
where Dij,c is a vector ordered by multi-index (ij,c) of c×1 size that contains the distances
between the (i, j)-th image pixel value and the c-class value for each band b.
The decision rule for classification based on the MDM filter determines that, based on the a
priori information for class segmentation (number of RSS to classify and their respective
mean values), each image pixel generates an ordered distance vector Dij, and therefore, the
pixel is classified according to the minimum value on the vector. Fig. 2 shows the processing
structure of the MDM classifier.
2.3 Weighted Pixel Statistics method

The Weighted Pixel Statistics (WPS) classificatory rule is computationally simple and can
result in classification accuracy comparable to other more computationally intensive algo-
rithms (WOS and MDM methods). It is characterized by the mean and variance values of
the RSS signatures (classes) and the Euclidean distances based on the Pythagorean Theorem.
An important aspect of this method is that it is applied to the MRS imagery.
The training data for class segmentation requires the number of RSS to be classified (c); the
means matrix M (c×c size) that contains the mean values cc: (0 ≤ cc ≤ 255, gray-level) of the
RSS classes for each RGB bands; and the variances matrix V (c×c size) that contains the va-
riances of the RSS classes for each RGB bands. The matrix M and V represents the weights
of the classification process. Next, the image is separated in the spectral bands (R, G and B)
and each (i, j)-th pixel is statistically analyzed calculating the means and variances from a
neighborhood set of 5x5 pixels for each RGB band, respectively. To compute the output of
the classifier, the distances between the pixel statistics and the training data is calculated
using Euclidean distances based on the Pythagorean Theorem for means and variances,
respectively. The decision rule used by the WPS method is based on the minimum distances
gained between the weighted training data and the pixel statistics.
The WPS techniques provide a high level of RSS segmentation and classification. Figure 3
shows the detailed processing structure of the WPS classifier.
3. Dynamic Model of Remote Sensing Signatures
This is a new innovative paradigm for mathematical analysis of the space-time dynamic
evolution of the particular environmental RSS extracted from MRS images in evolution time.
This is performed via the Geophysical Dynamic Laboratory (GDL) method, which unifies
the RSS mapping scheme with its dynamic analysis to provide the high-resolution mapping
of the RSS in evolution time. If the attributes of interest of a system are changing with time,
then it is referred to a dynamic system. A RSS process is the evolution over time of such a
dynamic system (Grewal & Andrews, 2001).
3.1 Mathematical model of RSS in continuous time

The model of a RSS of interest treated as a linear dynamic system (LDS) is presented in its
state variables (Falkovich et al., 1989) described over the continuous evolution time (CET)
domain (   ) and the discrete evolution time (DET) domain (   ). A LDS is
represented in CET and DET as shown in Fig. 4(a) and 4(b), respectively, where () and
{()} are the inputs to the linear system, () and {()} are the outputs of the linear system,
respectively.
Fig. 3. Computational structure of the WPS method
The model of the equation of observation (EO) in CET is represented (Falkovich et al., 1989)
as Σ( )  S( Λ( ))   ( ) , where () is the white observation Gaussian noise and    , start-
ing at 0 (initial moment of continuous evolution time), and the linear amplitude-modulated
(Falkovich et al., 1989) model is S( Λ( ))  Λ( )S0 ( ) , where S0() represents the deterministic
“carrier” RS image frame of a given model, and () is the unknown stochastic information
process to be estimated via processing of the MRS image observation data frame ().
a. Continuous evolution time (CET) do- b. Discrete evolution time (DET) domain:
main: -continuous time argument -discrete time argument
Fig. 4. Representation of linear dynamic systems
Is assumed that () satisfies the dynamical model specified by the following N-th order
linear differential equation (Villalon & Shkvarko, 2008)
d N Λ( ) d N 1Λ( ) d N 1 ( )

  N 1  ...   0 Λ ( )   N 1  ...   0 ( ) . (3)
d N d N 1 d N 1
where  and  are the constant coefficients of the dynamical system model for evolution of
the RSS (). This stochastic model can be redefined as follows: the differential equation (3)
may be transformed into a system of linear differential equations of the first order via per-
forming the following replacement of variables
 z1 ( )  Λ( ),
 dz 
 z2 ( )  1 )   N 1 z1 ( )   N 1 ( ),
(
 d
                                               
 (4)
 z ( )  dz N 1 ( )   z ( )    ( ),
 N
d
1 1 1

 dz N ( )
  0 z1 ( )   0 ( )
 d
where
z ( )   z1 ( ) z2 ( ) ... z N ( )  .
T
(5)
Based on the replacement of variables specified by (4), the dynamic differential equation
model (3) can be now represented in a canonical vector-matrix form as follows
dz ( )
 Fz ( )  Gξ ( ) , Λ ( )  Cz ( ) . (6)
d
where
  N 1 1 0 ... 0
    N 1 
 N  2 0 1 ... 0   
F   ... ... ... ... ... , G   N 2  , C  1 0 ... 0 . (7)
   ... 
 1 0 0 ... 1  
  0 0 ... 0   0 
 0
The representation in the form of (6) is referred to as a canonical equation of linear dynamic
system in state variables in continuous time. Here, z() is the state vector, the vector C de-
fines a linear operator that introduces the relation between the RSS map in continuous time
and the state vector z(), F is a transition matrix, G is a transition vector, and () represents
the white model generation noise vector characterized by the statistics, ξ ( )  0 and
ξ ( )ξT ( ')  Pξ ( ) (   ') , respectively. Here, P() is referred to as state model disperse
matrix that characterizes the dynamics of the state variances developed in continuous time
  0    starting from the initial instant . The dynamic model equation in the conti-
nuous time states the relation between the RSS map  extracted from the MRS scene, thus
the desired dynamical RSS map  can be represented as follows (Villalon & Shkvarko,
2008)
Σ( )  S 0 ( )C   z ( )  ν ( )  H ( )z ( )  ν ( ) , where H ( )  S 0 ( )C   . (8)
The stochastic differential model of equations (6) and (8) allows applying the theory of dy-
namical filtration to reconstruct the desired RSS map in continuous time incorporating the a
priori model of dynamical information about the RSS. The aim of the dynamic filtration is to
find an optimal estimate of the desired RSS, Λ ˆ ( )  Czˆ ( ) , developed in continuous time
  0    via processing the RSS maps Σ( ) extracted from the MRS scenes taking into
considerations the a priori dynamic model of the desired RSS map specified through the
state equation (6). The optimal dynamic filter when applied to the RSS maps Σ( ) specified
by the dynamic image model (8) must provide the optimal estimation of the desired RSS
map Λ ˆ ( )  Czˆ ( ) , in which the state vector estimate z() satisfies the a priori dynamic beha-
vior modeled by the stochastic dynamic state equation (6).
3.2 Mathematical model of RSS in discrete time

The canonical discrete-time solution to equation (6) in state variables for discrete time  is
expressed as follows (Falkovich et al., 1989)
z (  1)  Φ( )z ( )  Γ( )ξ ( ) , Λ ( )  C( )z ( ) , (9)
where Φ( )  F(  )  I ; Γ( )  G (  ) , and  represents the continuous time sam-
pling interval for dynamical modeling of the RSS map in discrete time. The statistical cha-
racteristics of the a priori information are as follows
1. Generating model noise {}:
ξ ( )  0 ; ξ ( )ξT ( ')  Pξ ( ,  ') . (10)
2. Observation (RSS map) noise {}:

ν ( )  0 ; ν ( ) νT ( ')  Pν ( ,  ') . (11)
3. State vector {z()}
z (0)  m z (0) ; z (0)zT (0)  Pz (0) . (12)
where 0 argument implies the initial state for initial time instant (= 0). The disperse matrix
Pz(0) satisfies the following disperse dynamic equation (Villalon & Shkvarko, 2008)
Pz (  1)  z (  1)zT (  1)  Φ( ) Pz ( )ΦT ( )  Γ( ) Pξ ( ) ΓT ( ) . (13)
3.3 Optimal dynamic RSS filtering technique

The strategy is to design an optimal decision procedure that, when applied to all RSS obser-
vations will provide an optimal solution to the state vector z() subjected to its prior defined
dynamic model given by the stochastic dynamic equation (9). The estimate of the state vec-
tor optimally defined in the sense of the Bayesian minimum risk strategy (Shkvarko, 2004)
in discrete time  can be represented in the conditioned form
zˆ( )  z ( ) Σ(0),Σ(1),...,Σ( ) , (14)

opt
were  represents an ensemble averaging operator. For discrete time, the design procedure
is based on the concept of mathematical induction (Falkovich et al., 1989). This is a supposi-
tion that after  observations Σ(0),Σ(1),...,Σ( ) the desired optimal estimate is produced,
defined at the ultimate step as
zˆ ( )  zˆ ( ) . (15)
opt
In order to use the estimate zˆ( ) it is necessary to design an algorithm that produces the
opt
optimal estimate z(+1) incorporating new measurements (+1) according to the state
dynamic equation (9). This is, we have to design an optimal decision procedure (optimal
filter) that, when applied to all reconstructed RSS maps Σ( ) ordered in discrete time
  0    , provides an optimal reconstruction of the desired RSS map represented via the
estimate of the state vector z() subject to the numerical dynamic model (9). To proceed with
derivation of such a filter, we first represent the state dynamic equation (9) in discrete time 
as follows
z (  1)  Φ( )z ( )  Γ( )ξ ( ) . (16)

3.4 Dynamic RSS map reconstruction

According to the dynamical model of equation (16), the anticipated mean value for the state
vector can be expressed as (Villalon & Shkvarko, 2008)
m z (  1)  z (  1)  z (  1) zˆ ( ) . (17)
The mz(+1) is considered as the a priori conditional mean value of the state vector for the
(+1)-st estimation step, thus, from equations (16) and (17) we obtain
m z (  1)  Φ z ( ) Σ(0), Σ(1),..., Σ( )  Γ ξ ( )  Φzˆ ( ) , (18)
hence, the prognosis of the mean value becomes
m z (  1)  Φzˆ ( ) . (19)
From the analysis of equations (16) thru (19), it is possible to deduce that given the fact that
the particular RSS map Σ( ) is treated at discrete time, it makes the previous reconstruc-
tions Σ(0), Σ(1),..., Σ( ) irrelevant; hence, the optimal filtering strategy is reduced to the
dynamical one step predictor described by the equation (16). Using these derivations, the
dynamical estimation strategy can be modified to the one step optimal prediction procedure
(Villalon & Shkvarko, 2008)
zˆ (  1)  z (  1) Σ(0), Σ(1),..., Σ( ), Σ(  1)  z (  1) zˆ ( ); Σ(  1) ;

(20)
zˆ(  1)  z (  1) Σ(  1); m z (  1) .
For the current (+1)-st discrete time estimation/prediction step, the dynamical RSS map
estimate of the equation (8) in discrete time becomes
Σ(  1)  H (  1)z (  1)  ν (  1) , (21)
with the a priori predicted mean calculated by the equation (17) for the desired state vector
given by (16). Applying the Wiener minimum risk strategy (Shkvarko, 2004) to solve the
equation (21) with respect to the state vector z() and taking into account the a priori infor-
mation summarized by the equations (10) thru (12), we obtain the dynamic solution for the
RSS map state vector
zˆ(  1)  m z (  1)  Θ(  1)  Σ(  1)  H(  1)m z (  1) , (22)
where the desired dynamic filter operator Θ(  1) is defined as (Villalon & Shkvarko, 2008)
Θ(  1)  K Θ (  1)HT (  1)Pν1 (  1) , (23)

and
1
K Θ (  1)   Ψ Θ (  1)  Pz1 (  1)  , (24)
Ψ Θ (  1)  HT (  1)Pν1 (  1)H (  1) . (25)
Finally, using the derived filter equations (22) thru (25) and the initial RSS map state model
of equation (9), the optimal filtering procedure for the dynamic reconstruction of the desired
RSS map can be represented in discrete time  as
ˆ (  1)  Φ( )zˆ ( )  Θ(  1)  Σ(  1)  H (  1)Φ( ) zˆ ( )  ;

Λ  =0,1, , (26)
4. Geophysical Dynamic Laboratory

The described technique provides the dynamical RSS map based on the atlas of RSS maps
extracted from MRS scenes.
The GDL method is defined in the form of equation (26) based on the atlas of RSS maps in
discrete time as follows
ˆ
GDL  Λ (  1)  Φ ( ) z ( )  Θ (  1)  Σ (  1)  H (  1)Φ ( ) z ( )  .
Λ ˆ ˆ ˆ (27)
Here, the observation vector Σ is formed by the threshold values of the same (i, j)-th pixel
through the different RSS maps of the atlas in the discrete time .
The estimate vector z is formed by the estimation values  one step prior in the same cur-
rent discrete time. Fig. 5 shows the detailed computational structure of the GDL method for
the environmental RSS extraction from MRS imagery.
5. Simulation Experiments for Hydrological Signatures
In the reported here simulation results, hydrological RSS electronic maps are extracted from
high-resolution MRS images. Three level RSS are selected for this particular simulation
processes, moreover, unclassified zones are also considered (2-bit classification) and are
described as follows.
Fig. 5. Computational structure of the GDL method

██ – Black regions represents the RSS relative to the wet zones of the MRS image.
██ – Heavy-gray regions represents the RSS relative to the humid zones of the MRS.
██ – Light-gray regions represents the RSS relative to the dry zones of the MRS.
██ – White regions represent the unclassified zones of the RSS map.
5.1 Multispectral Image Classification

To analyze the overall performance of the WPS technique, a set of four high-resolution
(1024x1024-pixels) MRS scenes in TIFF format are used, borrowed from diverse zones in
Mexico.
A comparison with the results obtained with the classical WOS and MDM methods is pro-
vided. Figs. 6(a), 7(a), 8(a) and 9(a) show the four different MRS scenes, respectively.
To perform the qualitative study, Figs. 6(b), 7(b), 8(b) and 9(b) show the results obtained
with the WOS method for each scene, respectively.
Figs. 6(c), 7(c), 8(c) and 9(c) show the results obtained with the MDM method for each scene,
respectively.
Figs. 6(d), 7(d), 8(d) and 9(d) show the results obtained with the WPS method for each
scene, respectively.
The quantitative study is performed calculating the classified percentage obtained with the
WOS, MDM and WPS methods, respectively, and compared with the original class quanti-
ties from the original MRS scenes. Tables 1, 2, 3 and 4 show the quantitative results.
The theory of the WOS method defines that the classification is performed only using one
band (Jensen, 2005), for this simulation the G band was used. The resulting RSS map shows
a large unclassified zone, this is due to the color gradient present on the original MRS image
and the lack of supervised data.
The MDM method uses the three RGB bands (Jensen, 2005). The WPS method also uses the
three RGB bands to analyze the pixel-level means and variances to perform a more accurate
segmentation and classification; therefore, using the statistical pixel-based information the
RSS map obtained shows a high-accurate classification without unclassified zones. From the
details shown in Figures 6 thru 9, the WPS method performs a more accurate and less
smoothed identification of the classes.
Tables 1 to 4 show the quantitative performances. From this analysis, the WPS classified
image provides a lower percentage difference from the original MRS scenes than the WOS
or MDM classified images. Moreover, the WOS and MDM reveal some unclassified zones
due to their respective decision rules (Johannsen et. al, 2003); the WPS method classifies all
the pixels due to the use of pixel-based statistical training data.
These qualitative and quantitative results probe the overall performance of the developed
WPS technique.
a. Original high-resolution MRS scene b. RSS map extracted with the WOS method
c. RSS map extracted with MDM method d. RSS map extracted with the WPS method
Fig. 6. Simulation results for hydrological RSS map extraction from the first MRS scene
Method → Original WOS method MDM method WPS method

Base [%] % Diff. % Diff. % Diff.
Wet 34.34 35.37 -1.03 31.93 +2.41 50.09 -15.74
Humid 32.60 26.75 +5.85 17.00 +15.61 18.37 +14.3
Dry 33.06 9.45 +23.61 47.18 -14.12 31.54 +1.51
Unclass. ----- 28.43 +28.43 3.90 +3.90 0.00 +0.00
Percentage Points
58.92% 36.04% 31.5%
Difference →
Table 1. Comparative table of the hydrological RSS percentages obtained with the classifica-
tion methods from the first MRS scene
Fig. 7. Simulation results for hydrological RSS map extraction from the second MRS scene.

Wet 33.11 23.66 +9.45 29.53 +5.57 51.22 -18.12
Humid 33.11 21.78 +11.32 21.24 +13.87 19.79 +13.3
Dry 33.79 27.03 +6.76 50.06 -16.28 28.99 +4.83
Unclass. ----- 27.53 +27.53 1.17 +1.17 0.00 +0.00
Percentage Points
55.05% 36.88% 36.2%
Difference →
tion methods from the second MRS scene.
Fig. 8. Simulation results for hydrological RSS map extraction from the third MRS scene.

Wet 33.02 6.87 +26.15 40.98 -7.96 21.31 +11.7
Humid 33.09 33.30 -0.22 35.38 -2.29 36.38 -3.31
Dry 33.89 23.07 +10.82 20.66 +13.23 42.31 -8.43
Unclass. ----- 36.75 +36.75 2.98 +2.98 0.00 +0.00
Percentage Points
73.94% 26.45% 23.4%
Difference →
tion methods from the third MRS scene.
Fig. 9. Simulation results for hydrological RSS map extraction from the fourth MRS scene.

Wet 32.41 65.35 -32.94 19.06 +13.35 20.62 +11.8
Humid 34.13 25.52 +8.61 60.14 -26.00 58.95 -24.82
Dry 33.46 9.03 +24.42 18.22 +15.23 20.43 +13.1
Unclass. ----- 0.10 +0.10 2.58 +2.58 0.00 +0.00
Percentage Points
66.07% 57.16% 49.6%
Difference →
tion methods from the fourth MRS scene.
5.2 Dynamical Analysis

A set of hydrological RSS electronic maps were extracted from 40 MRS high-resolution
images of a particular scene obtained with the same time interval (discrete time).
The GDL dynamic post-processing method is applied to the high-resolution collection of
RSS map (Shkvarko & Villalon, 2007) based on the computational structure described in Fig.
5. First, the collection of 40 RSS maps (Marple,1987) collected in different time of the same
scene is set for the simulation. Therefore, the discrete time  = 40. Second, the pixel evolu-
tion vector ij is defined for this simulation as

ˆ
Σij  Σ ij ,1
ˆ
Σ ij ,2
ˆ
 Σ 
ij ,32 , (28)
where Σ̂ represents the threshold values of the same (i, j)-th pixel from the 40 RSS maps.
This is the observation signal to be post-processed with the dynamic post-processing me-
thod.
Third, the measurement matrix H and the state transition matrix are simplified to I be-
cause the equation of observation (9) and the stochastic dynamic state equation (21) are
supposed to be ideal (noiseless, because the observation vector is directly extracted from the
RSS maps). The dynamic filter operator (gain matrix)  determines the variance evolution of
the observation values (28) of the dynamically reconstructed RSS. The initial conditions are
ˆ (0)   Σ(0) .
the initial observation value Σ(0) and its initial estimation Λ
The GDL method specified by equation (25) is applied to estimate the ultimate value Λ̂ that
is the next ( + 1)-st evolution time step of the observation vector ij. This represents the
dynamic filtration of the desired RSS from the reconstructed observation data and can be
expressed as
ˆ Λ
Λ ˆ   1 . (29)
ij
This process is performed through all the {(i, j)} pixels of the 40 RSS maps to obtain a single
aggregated RSS map Λ ˆ . The simulation results of application of the developed GDL
GDL
method are presented in Figs. 10 and 11.

Figs. 10(a) thru 10(e) show the first five high-resolution (1024x1024-pixel) hydrologic RSS
maps extracted from the first five MRS scenes (corresponding to the Banderas Bay of Puerto
Vallarta in Mexico) in different evolution time (1, 2, 3, 4 and 5), respectively.
Fig. 10(f) shows the dynamic RSS map reconstructed with the application of the GDL me-
thod for the time step () specified by the computational structure described in Fig.
5.
Fig. 11(a) shows the first original high-resolution (1024x1024-pixel) MRS scene (1). Fig.
11(b) shows the dynamic MRS map reconstructed with the application of the GDL method
for the time step.
The RSS maps were reconstructed in discrete time , therefore, the GDL method produces
the desired dynamic RSS prediction of the RSS map for the next time step (+ 1); where  =
0, 1, ….
a. RSS map extracted from the MRS scene b. RSS map extracted from the MRS scene
for a  = 1 discrete time step for a  = 2 discrete time step
c. RSS map extracted from the MRS scene d. RSS map extracted from the MRS scene
for a  = 3 discrete time step for a  = 4 discrete time step
e. RSS map extracted from the MRS scene f. RSS dynamic prediction obtained with
for a  = 5 discrete time step GDL for the  = 41 discrete time step
Fig. 10. Simulation results for dynamic RSS map analysis.

a. Original high-resolution MRS scene for b. High-resolution MRS scene predicted for
the  = 1 discrete time step the  = 41 discrete time step
Fig. 11. Simulation results for dynamic MRS map analysis.
6. Summary of Computational Algorithms

The detailed stages of the computational algorithm for the WPS and GDL methodologies are
summarized as follows.
6.1 Weighted Pixel Statistics method

1. Set the number of RSS to classify.
2. Select one point on the MRS image for each class to be classified.
3. Separate the spectral RGB band from the true-color MRS image.
4. The selected points determine the training weights that consist of the means matrix M
and the variances matrix V. These matrixes contain the mean and variance of each point in
the R, G and B bands, respectively.
5. For each (i, j)-th pixel in the R, G and B bands, respectively, perform the following
process
- Set a 5x5 pixel neighbourhood shift window W.
- Determine the mean of the shift window W.
- Determine the variance of the shift window W.
- Calculate the Euclidean distances between the means and the training means for each
band and for each class (Fig. 3).
- Calculate the Euclidean distances between the variance and the training variances for
each band and for each class (Fig. 3).
- Select the minimum class distance for the means.
- Select the minimum class distance for the variances.
- Perform a comparison between the class distance for the mean and the class distance
for the variance, and classify the pixel according to the minimum value and the class
from which is obtained.
6.2 Geophysical Dynamic Laboratory method

1. Set a collection of discrete time ( RSS maps from the electronic atlas extracted from the
MRS imagery for a particular scene.
2. For each (i, j)-th pixel on the RSS maps perform the following process:
- Set the pixel-based evolution vector ij, which contains the threshold RSS values for
the pixel in discrete evolution time .
- Apply the GDL method to the vector ij (Fig. 5) to obtain the dynamic prediction
ˆ (  1) , which conform the matrix Λ
Λ ˆ .
ij
ˆ
ˆ matrix conform the Λ
3. The reconstructed Λ ij GDL dynamic RSS image in discrete time .
7. Conclusion
The extraction of remote sensing signatures from a particular geographical region allows the
generation of electronic signature maps, which are the basis to create a high-resolution sig-
natures atlas processed in discrete time, and moreover, perform its dynamical analysis. This
chapter analyzed the implementation possibilities of the WPS and GDL methods for hydro-
logical resources management based geophysical applications. The extraction of hydrologi-
cal RSS from high-resolution MRS imagery was reported to probe the efficiency of the de-
veloped techniques.
7.1 Multispectral Image Classification

From the simulation results one may deduce that the WOS classifier generates several un-
classified zones; while the MDM classifier is more accurate because it uses more robust
information in the processing (several image spectral bands), nevertheless, despite the fact
that few zones are unclassified, the results have considerable density of unity pixels (suffi-
cient for decision making based on these extracted RSS). The developed WPS method pro-
vides the high resolution environmental RSS electronic map with a high-accurate classifica-
tion and without unclassified zones. This is achieved because the WPS method uses the
three RGB bands to analyze the pixel-level means and variances to perform a more accurate
segmentation and classification; therefore, using the statistical pixel-based information the
RSS map obtained shows a high-accurate classification without unclassified zones. The
resulting RSS map ensures better results in the classification achieved with the developed
WPS method. This is probed by the RSS percentages obtained with the WPS method, which
manifest the lowest percentage difference to those obtained with the WOS and MDM classi-
fication techniques. Also, the WPS method for RSS extraction can be applied to several MRS
images of a particular geographical region obtained in different moments of time (discrete
time), to generate a RSS atlas of environmental electronic maps. This process is a powerful
tool for geophysical resource management.
7.2 Dynamical Analysis

The GDL method provides a possibility to perform the high-resolution intelligent analysis
of the dynamic behavior or the desired environmental RSS map model with a high-accurate
classification of the particular RSS map evolution.
This is achieved because the GDL algorithm aggregates the RSS map atlas information for a
particular MRS scene in discrete evolution time and employs more detailed robust a priori
information from the original MRS scene. The resulting dynamic RSS prediction map en-
sures high-accurate estimation results in the classification achieved with the developed GDL
method. The reported here simulation results shows the qualitative and quantitative analy-
sis of the overall performance of the WPS and GDL methods for remote sensing signatures
analysis. The application as an auxiliary tool in Geophysical information retrieval and data
interpretation for land use management and analysis are a matter of further studies.
8. References
Falkovich S.; Ponomaryov V. & Shkvarko Y. (1989). Optimal Spatial-Temporal Signal
Processing for Spread Radio Channels, Radio and Communication Press, U.S.S.R.
Fussell J.; Rundquist D. & Harrington J. (1986). On defining remote sensing. Journal of Photo-
grammetric Engineering & Remote Sensing, Vol. 52, 1507-1511
Grewal M. & Andrews A. (2001). Kalman Filtering: Theory and Practice using Matlab, John
Wiley & Sons, U.S.A.
Jensen J. (2005). Introductory Digital Image Processing: A Remote Sensing Perspective, Prentice-
Hall, U.S.A.
Johannsen C.; Petersen G.; Carter P. & Morgan M. (2003). Remote sensing: changing natural
resource management. Journal of Soil & Water Conservation, Vol. 58, 42-45
Luenberger D. (1979). Introduction to Dynamic Systems: Theory, Models and Applications, John
Wiley & Sons, U.S.A.
Marple S. (1987). Digital Spectral Analysis, Prentice-Hall, U.S.A.
Mather P. (2004). Computer Processing of Remotely-Sensed Images, John Wiley & Sons, U.S.A.
Perry S.; Wong H. & Guan L. (2002). Adaptive Image Processing: A Computational Intelligence
Perspective, CRC Press, U.S.A.
Shkvarko Y. (2004). Unifying regularization and bayesian estimation methods for enhanced
imaging with remotely sensed data. Part I - theory. IEEE Transactions on Geos-
cience and Remote Sensing, Vol. 42, 923-931
Shkvarko Y. & Villalon I. (2007). Remote sensing imagery and signature field reconstruction
via aggregation of robust regularization with neural computing, In: Lecture Notes in
Computer Sciences, Blanc-Talon J.; Philips W.; Popescu D. & Scheunders P. (Ed.), pp.
235-246, Springer Verlag, Germany
Smith S. (2000). The Scientist and Engineer’s Guide to Digital Signal Processing, Web Edition,
U.S.A.
Villalon I. & Shkvarko Y. (2008). Comparative study of the descriptive experiment design
and robust fused bayesian regularization techniques for high-resolution radar im-
aging. Scientific and Technical Journal on Radioelectronics and Informatics, Vol. 1, 34-48
Villalon I. (2008). Weighted pixel statistics for multispectral image classification of remote
sensing signatures: performance study, Proceedings of the 5th IEEE International Con-
ference on Electrical Engineering, Computing Science and Automatic Control, pp. 534-
539, IEEE Press, Mexico City
Yli-Harja O.; Astola J. & Neuvo Y. (1991). Analysis of the properties of median and
weighted order median filters using threshold logic and stack filter representations.
IEEE Transactions on Signal Processing, Vol. 39, 395-410
Forest Inventory using Optical and Radar Remote Sensing 539
26
X
Forest Inventory using Optical

and Radar Remote Sensing
Yolanda Fernández-Ordóñez1, Jesús Soria-Ruiz2 and Brigitte Leblon3
1Postgraduate College in Agricultural Sciences (COLPOS), Mexico
2National Institute of Research for Forestry, Agriculture and Livestock (INIFAP), Mexico
3Faculty of Forestry and Environmental Management, University of New Brunswick,
Canada
1. Introduction
Forests cover approximately a third of the planet’s land area, and range from undisturbed
primary forests to forests managed and used for a variety of purposes. Using forests as a
resource has an impact on the environment and on the economies of nations. Forests have a
history of being exploited either adequately or abusively, but more efforts are now being
made towards their sustainable use. The assessment of forests in terms of their extent,
condition, use, and value is periodically realized at both global and country levels (Ridder,
2007). Applications relating to the monitoring of forest status and to forest management
require updated and accurate inventories summarizing knowledge of land use changes
related to forests, including the rates and patterns of deforestation and afforestation. This
knowledge must bear more detail for use at regional and local scales. Inventories are
produced in most countries at various levels and various update periodicities. The survey
methods and technologies applied in each case vary according to the intended use, inherent
costs, timeliness, and desired accuracy of the inventory. Sustainable Forest Management
(SFM) aims to manage forest land in order to obtain products and services while
simultaneously minimising any undesirable effects on the social and physical environments.
Major issues concerning SFM include establishing the location, extent, volume, and status of
different forest types such as permanent forest estates managed for timber, areas susceptible
to reconversion, and protected areas where logging and extracting activities are to be strictly
controlled (ITTO, 2006). Forest volume inventories provide valuable data for estimating
above-ground biomass density and carbon stores (Zheng et al., 2007). Good estimates of
forest biomass are relevant in monitoring deforestation and forest degradation, which
impact the global carbon cycle (Haripriya, 2000; DeFries et al., 2006).
Traditionally aerial photographs have been used for forest inventories, but their availability
and area coverage can be limited. A promising alternative is the use of images acquired by
space borne sensors. Satellite images used in forest inventory were first provided by optical
sensors, but their availability is still limited because their acquisition depends on solar
radiation. More recently, synthetic aperture radar (SAR) imaging systems were developed.
They are active sensors that produce their own incident radiation and thus are able of
acquiring images whatever the weather and illumination conditions. The wavelength they
used is also longer than the one of the optical sensors, allowing measuring different
characteristics of the imaged area.
Thus optical and radar remote sensing (RS) provide different but complementary
capabilities for estimates of general forest status such as crown height, density and shape,
tree basal area, timber volume and biomass. For forest inventory some variables would be
more useful per se than others where composition at the species level becomes relevant.
Therefore, approaches that combine information obtained from multiple sources offer a
promising approach towards establishing more accurate and detailed land cover
inventories, particularly at regional and local levels.
This chapter is concerned with forest inventory mapping using combinations of optical and
radar RS techniques and intervening field studies at cartographic scales (1:20,000) applicable
to regional and local forest management. These themes are examined with respect to
providing insights into the use of combined RS techniques for more comprehensive, cost-
effective, and accurate forest inventories particularly in Mexico, a mega-diverse country
with an increasing concern for the conservation of its biodiversity and sustainable natural
resources management (Conservation International, 2008).
In the following two sections, a review of optical and radar RS as they pertain to forest
inventory is presented. Then follows a section on the combination and fusion of optical and
radar RS data. Ongoing and future work which aim at applying these techniques using
medium and high resolution optical imagery and radar imagery in Central Mexico are
described next. A paragraph on conclusions and foreseeable developments closes this
chapter. The list of references is included.
2. Optical RS
Global and regional forest mapping are increasingly relying on optical RS because of the
lower cost of gathering synoptic forest data. Traditional sampling methods can be costly or
inapplicable due to terrain topography or field work resources and costs (Shiver and
Borders, 1996). Optical sensors use wavelengths from 0.4 to 2.5 μm to measure reflected
solar radiation. The range of the visible radiation, from 0.4 to 0.7 μm, also termed
photosynthetically active radiation (PAR), is useful for the detection of plant photosynthesis.
Visible, near-red, and infra-red wavelengths used for data imaging are of the same order of
magnitude as the cell components and prevailing pigments of plants, so a variety of
techniques have been established for determining tree canopy biochemistry and for deriving
plant condition.
Space-borne optical sensors are passive or active, depending on the solar radiation. Data
acquisition is affected by illumination conditions and is subject to atmospheric absorption
and scattering, due to the applicable wavelength which is of the same order of magnitude as
the size of atmospheric constituents (Gerstl, 1990). Over the years, research has provided
knowledge on how various factors, such as tree foliage type and understorey vegetation,
impact reflectance data in the visible, near-infrared, and infrared bands (Myneni et al., 1995).
The passive satellite-borne sensor applications have focused mainly on mapping biophysical
and biogeochemical information from multi-spectral images. More recently airborne optical
active sensors called LiDAR that send their own incident energy have been developed. They
are increasingly being used for forest applications mainly for estimating tree height, crown
dimensions, canopy height of stands and canopy cover. LiDAR is not further discussed here
because there is not yet an operational space-borne LiDAR sensor, and the reader is referred
to recent review works which show the potential advantages of this type of sensor (Sun &
Ranson, 2000); (Packalén and Mantamo, 2008).
Passive optical RS applications for forestry inventories are based on the sensitivity of optical
radiation to the chlorophyllian content of the stands as detailed in Leblon (1997). However
several factors can interfere such as the optical properties of the soil background,
illumination and viewing geometries as well as from meteorological factors (wind, cloud).
To reduce the effect of these factors on the forest canopy spectral response, single-band
reflectances measured by the sensors are generally combined into vegetation indices. At
least fifty different vegetation indices exist (Bannari et al., 1995). The most commonly used
vegetation indices are ratios of single-band or linear-combined reflectances. Ratioing allows
removal of the disturbances affecting, in the same way, reflectance in each band. As a
spectral band, most ratio-based vegetation indices use either the red band, which is related
to the chlorophyll light absorption or the near-infrared band, which is related to the green
vegetation density. Indeed, both bands contain more than 90% of the spectral information of
a plant canopy. Also, in red and near-infrared bands, the contrast between vegetation and
bare soil is at a maximum. Other indices based on the short-wave infrared have been
developed that are sensitive more to the moisture content of the canopy than to its
chlorophyll content. It has been shown that these indices better estimated the density of
managed conifer forests than those sensitive to chlorophyll (Aguirre-Salado et al., 2009).
When using optical RS images in forest inventory mapping, one should pay attention to
characteristics of the phenological cycle which are directly related to both vegetation type
and species diversity and thus indirectly to small-scale heterogeneity of climatic and
topographic conditions in the corresponding study region. Therefore a key component for
forest attribute determination is the time-series analysis of repetitive multispectral and even
hyperspectral imagery (Vuolo et al., 2008; Schowengerdt, 1997).
In addition to this time dimension, one also needs to consider the spatial dimension. High
spatial resolution images such as Ikonos and Quickbird provide better estimates for tree
crown cover over medium resolution images from SPOT and Landsat sensors in the case of
pine forests. A small scale experiment has shown that there is confusion with tree shadows
when using medium resolution optical imagery, which is a drawback for forested areas in
ravines or over steep slopes (Valdez-Lazalde et al., 2009). Approaches to overcome this and
other drawbacks include multi-angle optical RS whereby separating sunlit and shaded
leaves and using canopy architectural parameters is possible (Chen et al., 2003).
Species distribution modelling which is commonly based on ecological models has been
shown to improve when incorporating spatio-temporal information derived from optical RS
(Pearson, 2007). Recent work in this area (Cord et al., 2009) has shown that analytical
techniques that exploit data beyond their spectral dimension to gather knowledge about
phenological cycles, seasonal, and latitudinal variations over time can contribute to more
accurate vegetation type and species diversity determination. This is important for
inventories over natural forested areas at regional and local levels. Optical imagery covering
the targeted inventory zones should be available over time. Cloud cover can impair this
requirement, while radar sensors provide data even under this condition in addition to
being potentially able to provide complementary information about forest structure and
vegetation water content because of their different physics nature.
3. SAR RS
Microwave RS instruments are synthetic aperture radar (SAR) systems that image an area
independently of solar illumination and are not affected by cloudy or dusty conditions as
are optical instruments (Freeman, 1996). They operate in longer wavelengths in the optical
regions of the electromagnetic spectrum giving different information on the tree canopies.
The SAR sensors image the area by transmitting electromagnetic pulses in frequency bands
between 1 and 100 GHz. The corresponding wavelength (λ) ranges between 1 mm and 1 m.
It is of the same order of magnitude as the sizes of tree leaves, branches and canopies. By
comparison, the wavelength of optical instruments ranges from 0.4 to 2.5 μm, and thus is of
the same order of magnitude as the cell components and prevailing pigments of plants
(Leblon & LaRocque 2008). Thereby, optical imagery is more suitable for determining tree
canopy biochemistry and inferring vegetation condition, while radar images are better for
determining canopy structural parameters
SAR sensors were first developed for military applications. The current satellite civil SAR
sensors use one of the following bands: X-band: 12.5-8 GHz (λ: 2.4 to 4.8 cm) (TerraSAR-X);
C-band: 8.0-5.0 GHz (λ: 4.8-7.5) (ERS1/2, ENVISAT, RADARSAT-1/2); and L-band: 2.0-1.0
GHz (λ: 15.0-30.0) (JERS and ALOS-PALSAR). A satellite SAR sensor called BIOMASS
operating in the P-band (1.0-0.3 GHz or λ: 30.0-100.0) is in preparation by the European
Space Agency. It is designed for forestry applications. As detailed in Leblon and LaRocque
(2008), microwave radiations which encounter natural surfaces are either:
(i) diffusely scattered towards the sensor to produce the backward scattering called radar
backscatter or radar return (σ0 sigma nought ) which is the only scattering measured by
the sensor;
(ii) diffusely transmitted to lower layers through forward scattering;
(iii) specularly reflected in the case of smooth surfaces.
The interactions between microwave radiations and natural surfaces depend on: (i)
geometrical factors like the surface roughness as well as the incidence angle and the
wavelength of the incident radiation and (ii) dielectric factors which are related to the nature
and the moisture content of the target.
Because of the longer wavelength than the one for optical sensors, microwave radiation
produces not only surface scattering, but also volume scattering. The intensities of both
scatterings depend on the surface roughness, the incidence angle and the wavelength of the
incident radiation. Volume scattering also depends on the penetration depth -a function of
the wavelength- and occurs when the wave encounters dielectric discontinuities within the
material of the target (Leblon & La Rocque 2008).
Radar backscatter is rather complex on tree canopies. Ulaby et al., (1990) identified eleven
components of the radar backscatter of tree canopies. Fig. 1 shows nine major backscattering
sources for tree canopies (Piwovar, 1997): 1- Diffuse scattering from the ground; 2- and 3-
Direct scattering from various vegetation components; 4- Double- bounce vegetation-
ground interaction; 5- Corner reflector between tree trunks and ground; 6- Direct
backscatter from the forest canopy; 7- Volume scattering from within the forest canopy; 8-
Diffuse scattering from the ground; 9- Shadowing by parts of the forest canopy of other
parts of the canopy or ground.
The importance of each of the above components on the total scattering of a tree canopy
depends mostly on the microwave wavelength (Leblon & LaRocque, 2008). Radar
backscatter in shorter wavelengths, like X- and C-band, is mainly due to smaller tree
elements (leaves or needles and small branches of the upper tree crown level) (Hoekman
1993); (Mougin et al., 1993); (Yatabe & Leckie, 1995); (Ahern et al., 1995). Therefore,
shortwave bands (X- or C-band) can be used to discriminate between forest species
(Drieman et al., 1989); (Hoekman, 1990); (Dobson et al., 1992a); (Mougin et al., 1993); (Rignot
et al., 1994a); (Yatabe & Leckie, 1995); (Ahern et al., 1993, 1995). In particular, the C-band
backscatter magnitude of red pine, jack pine and black spruce was found to be inversely
related to needle length (Ahern et al., 1993); (Yatabe & Leckie, 1995). Radar backscatter of
longer wavelengths (P- and L-bands), which are more penetrating, originates from the major
branches of the crown, from the trunks, the double bounce scattering from the tree trunk or
crown to ground and from the ground (Richards et al., 1987); (Le Toan et al., 1992);
(Hoekman 1993); (Ahern et al,. 1995).
1 Diffuse scattering from the ground; 2 and 3 Direct scattering from various vegetation components; 4 Double-
bounce vegetation-ground interaction; 5 Corner reflector between tree trunks and ground; 6 Direct backscatter
from the forest canopy; 7 Volume scattering from within the forest canopy; 8 Diffuse scattering from the
ground; and 9 Shadowing by parts of the forest canopy of other parts of the canopy or the ground.
Fig. 1. Surface and volume scattering of a SAR beam for trees (Piwowar, 1997).
According to Leblon et al. (2002), radar backscatter (σ0) measurements of forested areas
depend on (i) vegetation type, species, and structure (Leckie, 1990); (Dobson et al., 1992a;
1992b; 1995), (ii) vegetation biomass (Kasischke et al., 1995); (Harrell et al., 1995; 1997);
(Pulliainen et al., 1996); (iii) topography and surface roughness (Shi et al., 1997) and canopy
height (Riom & Le Toan, 1981); (Dobson et al., 1992b); (Harrell et al., 1997); (iv) flooding and
the presence/absence of standing water (Hess et al., 1990; (Kasischke & Bourgeau-Chavez,
1997), and (iv) near-surface soil moisture or fuel moisture (French et al., 1996); (Leblon et al.,
2002); (Abbott et al., 2007).
Discrimination between forest and non forest areas as well as estimation of total above-
ground forest biomass or stem volume is better achieved using long wave bands (P-band)
(Le Toan et al., 1992); (Dobson et al., 1992b); (Koch et al., 1992); (Rignot et al., 1994b); (Rauste
et al., 1994); (Ranson & Sun, 1994); but the relationship has a logarithmic shape showing a
saturation of radar backscatter beyond a certain value of biomass.
SAR sensors acquire images in one single band. By contrast, optical sensors like Landsat-TM
acquire images in a multispectral mode, i.e., several images at different wavelengths at the
same time. However, radar sensors can acquire images under various incidence angles
which might highlight differences. The most recent radar sensors like ENVISAT-1, ALOS-
PALSAR TerraSAR-X, and RADARSAT-2 also allow the acquisition of images under
different polarizations.
This will also allow the production of colour composites like for multispectral optical
images. The polarization of the transmitted and returned radiation is an important property,
which can be used as a target separator in SAR image analyses. Polarization refers to the
position of the locus of the electric field vector in the plane perpendicular to the direction of
propagation of the radar microwave radiation. Current SAR systems transmit horizontally
polarized (H) or vertically polarized (V) EM waves, which upon interacting with a target
generate a backscattered wave with a horizontal polarization (H) or a vertical polarization
(V).
The possible combinations are (transmit polarization given first): HH, VV, HV and VH.
With multi-polarization sensors, the following imaging capabilities are thus possible: single
polarization (HH or VV), dual polarization (HH and HV, VV and VH, or HH and VV); and
quad-polarization (HH, VV, HV, and VH). Some multi-polarization sensors (RADARSAT-2,
ALOS-PALSAR and TerraSAR-X) have also a polarimetric mode, that provides not only the
intensity information in each polarization (HH, VV, HV, and VH), but also the phase
information. As a result, the complexity of SAR images processing increases with the level
of complexity of the polarization information as it is shown in Fig. 2 in the case of the
RADARSAT-2 sensor.
The information provided by multi-polarized or polarimetric SAR images has been
investigated, either directly with airborne and satellite radar data, or through simulation
models, with respect to their ability to discriminate land cover or forest types (Freeman &
Durden, 1998); (Touzi et al., 2004; (Freeman, 2007), to estimate forest growth (e.g., Izzawati
et al., 2006), biomass (e.g., Le Toan et al., 1992) or forest architecture (Lucas et al., 2006).
Although the anticipated potential of the polarimetric SAR images like those acquired by
RADARSAT-2 for mapping forest inventory, clear cuts, fire scars and biomass has been
rated as moderate, merging polarimetric SAR data with other data sources is a promising
investigation approach (van der sanden et. al., 2005). This is discussed in the following
sections of this chapter.
4. Forest Inventory Research in Mexico Using Optical and SAR Data

National forest inventories are labour-intensive and costly endeavours in need of advanced
technologies to reduce costs and time schedules and to produce information at the levels
required by all involved actors. The National Forest Inventory (NFI) in Mexico pertains to
the federal government and its general terms are regulated by law. The cartographic scale
used in the last inventory, NFI 2000, is 1:250000.
Fig. 2. Information sequence contained in a Radarsat-2 image (courtesy of G. Staples, MDA)
There is, however, a need for establishing more specific forest inventories at the state level,
and is the responsibility of state and municipal governments. For the State of Mexico the
most recent forest inventory was completed in 1994. The state is located in central Mexico
and it contains 75% of the country’s population, and although the physical environment has
been modified in recent years there is no up-to-date, reliable, official assessment of forest
resources. Information about land cover change has been obtained recently but it is not
oriented toward forest inventory (Escalona-Maurice, 2006). Many regions have been
transformed by deforestation resulting from various activities including illegal logging, the
opening of new agricultural areas, and the development of upper-class residential areas
within or near forested regions.
Since 2004, the responsibility for producing an NFI has lain with two government agencies,
the National Forestry Commission (CONAFOR) and the Ministry of Environment and
National Resources (SEMARNAT). The 2009 inventory will provide valuable forest and
soils information but with insufficient detail to adequately support regional or local
protection, conservation, and afforestation activities. To our knowledge, there are currently
no operational applications of radar RS within Mexican government agencies oriented
towards forest inventory or forest management. We consider that SAR and optical RS
technologies used together will contribute to satisfying the time requirements, accuracy, and
finer levels of detail that are required for forest inventory.
4.1 Forest Inventory

In Mexico, as in most countries, forest inventories at different levels of scale are a current
concern. There is a worldwide public awareness of the need for sustainable development.
However, in developing countries this need is not so easy to meet when faced with other
immediate priorities and with a scarcity of economic resources. Even so, periodic national
information requested by FAO for global forest resource assessments must be provided. At
a global level the requirement is for a picture of existing forests, statistics, and derived
trends. At the national level the request includes both the country as a whole and the major
administrative subdivisions of the territory. Information items include land use and the
changes in forested areas, estimates of existence and growth, and diagnosis of the health
status of forests. Information about the conservation of natural ecosystems and of
biodiversity is also sought (Toledo & Ordoñez, 1993).
Mexico, with an area of nearly two million km2, is one of three mega-diverse countries
(along with the United States and Colombia) with coastlines on the Atlantic and the Pacific.
However a growing population and ensuing changes in land use pose many threats to the
permanence of this biodiversity, particularly in forested regions. In spite of inadequate
funding and technical difficulties due to the extent and varied topography of the country,
there has been considerable progress in the construction of the NFI. The last NFI dates back
to the year 2000 (Palacio et al., 2000). The cartography was based on a classification scheme
based on visual interpretation of Landsat ETM+ colour composite images using the national
digital base map. The accuracy assessment was supported with high resolution digital aerial
photography. The analysis of deforestation was based on a classification which included
four forest classes: (i) primary and secondary tropical and temperate; (ii) scrubland; (iii)
man-made covers (grazing land, agricultural cropland, and human settlements); and (iv)
other covers (Mas et al., 2002). The NFI 2000 provides the current frame of reference for
lower-scale inventories which will be used for the 2009 inventory.
The 2009 inventory, which is 80% complete in the stratified sample stage and as yet
unpublished, will use optical images (AVHRR, MODIS, LANDSAT and SPOT (Sandoval
Uribe, 2007). The national mapping scale is being maintained at 1:250000 but a new scale of
1:50000 for reporting at the state level is being adopted. Information will also be reported at
the ecological region level, using the official ecological zonification scheme. The results are
to be included in a web information system (CONAFOR, 2009). The study area for the
investigation reported in this chapter is located approximately within the bounding
rectangle of Region 13 Temperate Mountain Areas, which includes the ecosystems of (i)
Forest and (ii) High and Medium Jungle, with reported national volume/ha values for these
ecosystems of 41.86 and 39.81, respectively (Fig. 3). The preceding discussion motivates and
justifies the research project reported in section 4.2.
4.2 Research Project

The general research objective of our current research project is to produce a methodology
to improve forest inventory assessment and results through the combined use of different
sources of data. The methodology should be amenable to translation for use at regional and
state levels elsewhere in the country. Particular objectives include the following: (i) to assess
the benefits of using polarimetric SAR data in this region, which includes forested areas in
the mountain chain “Eje Neovolcánico Transversal”; and (ii) to evaluate techniques for the
fusion of optical and SAR data. Part of this project is included in the RADARSAT SOAR
#2755 project. The main objective of the Canadian Space Agency SOAR program is to assess
and explore operationally viable solutions to current problems and issues, including those
related to the environment. Initiatives include fostering research and development of new
applications using RADARSAT-2 data, particularly those of fully polarimetric modes (MDA
et al., 2009).
Fig. 3. National Forest Inventory Region 13 Temperate Mountain. (CONAFOR, Mexico)
Although the concept of data fusion is intuitively understood, its exact meaning varies
among researchers and among scientific and technological domains. Therefore the term
should be formally defined and its various aspects discussed. Fusion has been taken as a
synonym of merge, combine, integrate, classify, and synergize. The range of application
objectives, methods, and techniques in RS is very diverse, so it is unlikely that a single axis
(whether wavelength, acquisition means, image, spatial resolution, mathematical tools, or
application domains) can be favoured upon which a fusion definition can be based.
Increased quality of the information is often the goal when using data and information from
different sources, but quality in RS depends on the particular objectives to be attained and
on whom the end-consumer of the results is. Quality can refer to the enhancement of
previous results, of method efficiency, or of results accuracy, all of which are clearly tied to
the application and its intended users (Fasbender et al., 2009). Hall and Llinas (1997) (cited
by Wald, 1999) consider information quality in terms of “data fusion techniques (that…)
combine data from multiple sensors, and related information from associated databases, to
achieve improved accuracy and more specific inferences that could be achieved by the use
of a single sensor alone.”
The notion of the pixel has frequently been used to address data fusion “at the pixel level”,
but it is often required to consider the merging of information at levels other than the pixel.
The information content of a pixel is not comparable in optical and SAR images. Wald (1999)
proposed a definition which suits the objectives of this chapter and of our research, since he
contends that fusion is a framework, not to be confined to the sensors’ signals, nor restricted
to the peculiarities of the sensor systems: “data fusion is a formal framework for the alliance
of data originating from different sources. It aims at obtaining information of greater
quality; the exact definition of ‘greater quality’ will depend upon the application.” The
processing tools and methods thus comprise no part of this definition of fusion as a
framework, even though they are relevant to a particular implementation or methodology of
a data fusion effort. The terms “merge” and “combine” are intentionally considered to be
looser concepts, allowing for the description of processes and methods in a general way
without getting into details. The reader is referred to the ample literature for the ongoing
discussion on the definition of the fusion concept and its many and fruitful uses (Pohl & van
Genderen, 1998); (Nikolakopoulos, 2004); (Wang et al., 2005).
Our current work is part of a broader on-going action to map land cover and to assess land
cover changes in the State of Mexico (Soria-Ruiz et al., 2009). Forest inventory is an
emerging research line within this broader action. Previous related work combined Landsat-
ETM and SPOT producing a map atlas (Soria-Ruiz et al., 2005). The selected study area for
this stage of the project is the Zinacantepec Municipality in the State of Mexico. This area is
well known from extensive field surveys containing over 60 control points, making it
suitable for analyzing natural and deciduous stands (mixed and homogeneous species) and
other types of vegetation considered by the national forest inventory. The terrain includes
flatland and part of the “Nevado de Toluca” volcano.
Previous tests using Radarsat-1 data and Landsat-ETM in Central Mexico for land cover in
Central Mexico indicated a global map accuracy with optical imagery alone of 78%, whilst
fusing optical and SAR resulted in an accuracy of 89%, with a kappa coefficient
improvement from 0.72 to 0.86%. So the fusion of Landsat ETM+ and RADARSAT-1
encouraged more tests of fusion for forest classification and for other land covers as well
(Soria-Ruiz & Fernandez-Ordonez, 2004).
Partial classification results, for four classes, using Landsat ETM images from 2000-2003 are
shown in Fig. 4.
The main goals set out for the period 2009-10 using other optical and SAR data are:
(i) Continued update of the land cover map base with high form SPOT, Ikonos, and
Quickbird;
(ii) Integration of multi-polarization Radarsat-2 C-band images over the summer and
winter seasons to investigate the applicability in this area of HH, VV, and HV
polarizations. At least six scenes will be used, including polarimetric, fine quad-pol HH
and VV, and HV and VH;
(iii) Derivation of polarization signatures and relating them to forest parameters for the
genera Abies, Pinus, and Quercus, which are largely represented in the area and are of
importance for forestry;
(iv) Accuracy assessment of the classifications; and
(v) Assessment of other polarimetric techniques (like interferometry) for different types of
stands, including commercial/natural, homogeneous/mixed, and older/younger.
Recent work by Viergever et al., (2007) has investigated biomass estimation for tropical
forests in Central America using backscatter and fully polarimetric data from AIRSAR. C-,
L-, and P-bands were used; vegetation heights were underestimated due to the
heterogeneity of the vegetation cover. In another case, confusion was observed in
backscatter values for P and L bands for the same type of vegetation (Freeman and Durden
1998, Freeman 2007). SAR and polarimetric SAR allow using interferometry techniques that
are suitable for tree height and density (Papathanassiou & Cloude 2001). For local forest
inventory a high accuracy of classification (on the order of 80 % or better) is desirable. Our
work will explore these technical possibilities, and assess their cost-effectiveness, although
the long repeat cycle for the satellite overpass leads to de-correlation that restricts the use of
such techniques. Since different wavelengths have different penetration properties in forest
targets data from other radar sensors such as TerraSAR-X and ALOS-PALSAR (L-band) will
also be used.
Fig. 4. Forest inventory in the Zinacantepec Municipality, State of Mexico, 2008.
5. Conclusion
State and local governments in Mexico have expressed a need for assessment of forest
resources at state and municipal level. Bio-conservation and ecological research have data
requirements at even smaller cartographic scales, less than 1:10000. Currently, techniques
for gathering information at this level are performed on a per-project basis under various
limitations and at great cost especially in mountain regions.
The avenue of investigation undertaken in our work is how best to combine optical and SAR
RS techniques profiting from the different but complementary information they provide as
it was shown in the current chapter. Our long term objective is to determine whether
polarimetric radar RS in complementary use with optical RS offer conveniences for forest
inventory and ecosystem studies in Central Mexico. The study will be based on SPOT,
Quickbird and Ikonos images and RADARSAT-2 C-band polarimetric SAR images. These
images offer new capabilities in forest inventories, because they are acquired in four
polarizations and have the phase information. SAR RF for forests is a challenging research
field with many unsolved issues pertaining to forest inventories in mega-diverse areas. Our
long term objective is to determine what contribution SAR RS is able to make with respect to
forest inventory and ecosystem studies in Central Mexico. However, it has been shown that
an even better SAR sensor for forestry applications should operate in the P-band, like the
BIOMASS sensor that is currently designed by the European Space Agency. Further projects
will be undertaken once the images from this new satellite will become available. Also,
interferometry and polarimetric interferometry have been shown to be well suitable for tree
height and canopy density measurements from SAR images. Such techniques will be more
efficient once images from the new German satellite system called TANDEM-X will be
available. The system allows acquiring two images on the same area in almost the same time
allowing computing interferograms without temporal decorrelations
6. References
Abbott, K.N.; Leblon, B; Staples, G.C.; Maclean, D.A. & Alexander, M.E. (2007). Fire danger
monitoring using RADARSAT-1 over northern boreal forests. International Journal of
Remote Sensing. Vol. 28 , Issue 6 (March 2007). 1317-1338. ISSN 0143-1161.
Ahern, F.J.; D.G.; Leckie & J.A. Drieman, J.A. (1993). Seasonal changes in relative C-band
backscatter of Northern forest cover types, IEEE Transactions on Geoscience
and Remote Sensing, Vol. 31, Issue. 3, (May 1993). 668-680, ISSN 0196-2892.
Ahern, F.J.; Landry, R.; Paterson, J.S.; Boucher, D. & McKirdy, I. (1995). Forest land cover
information content of multi-frequency multi-polarized SAR data of a boreal forest,
Proc. 17th Canadian Symposium on Remote Sensing, June 1995, Saskatoon,
Saskatchewan, Canada, 537-549. ISSN 1712-7971.
Aguirre-Salado, C.A.; Valdez-Lazalde, J.R.; Angeles-Perez, G.; De Los Santos-Posadas, H.M.;
Haapanen, R. & Aguirre-Salado, A.I. (2009). Mapping aboveground tree carbon in
managed patula pine forests in Hidalgo, Mexico. Agrociencia, Vol 43, (February-
March 2009), 209-220, ISSN 1405-3195.
Bannari, A.; Morin, D.R.; Huete A.R. & Bonn, F. (1995). A review of vegetation indices,
Remote sensing reviews, No. 13, 95-120. ISSN 0275-7257.
Chen, J.M.; Liu, J.; Sylvain G.; Leblanc, S.G.; Lacaze, R. & Roujean, J.L. (2003). Multi-angular
optical remote sensing for assessing vegetation structure and carbon absorption.
Remote Sensing of Environment, Vol. 84, Issue 4 (April 2003) 516–525, ISSN 0034-4257.
CONAFOR, Mexican National Forestry Commission. (2009). National Forest Inventory
System SNIF. Accesed 19.02.09
http://148.223.105.188:2222/gif/snif_portal/index.php?option=com_content&task
=view&id=4&Itemid=59
Conservation International (Content Partner). Mark McGinley (Topic Editor). (2008).
Biological diversity in the Madrean pine-oak woodlands.0020 In: Encyclopedia of Earth.
Eds. Cutler J. Cleveland (Washington, D.C.: Environmental Information Coalition,
National Council for Science and the Environment). First published in the
Encyclopedia of Earth May 3, 2007; Last revised August 22, 2008; Accessed 02.01.09
http://www.eoearth.org/article/Biological_diversity_in_the_Madrean_pine-
oak_woodlands
Cord, A.; Schmidt, M. & Dech, S. (2009). Potential and limitations of multi-temporal Earth
observation data to improve modeled results of tree species distribution in Mexico.
Proceedings of 33rd International Symposium on Remote Sensing of the Environment,
May 4-8, 2009, Stresa, Lago Maggiore, Italy.
DeFries, R.; Achard, F.; Brown, S.; Herold, M.; Murdiyarso, D.; Schlamadinger, B. & De
Souza, C. Jr. (2006). Reducing Greenhouse Gas Emissions from Deforestation in
Developing Countries: Considerations for Monitoring and Measuring. Global
Observation of Forest and Land Cover Dynamics. Report Series GOFC-GOLD-34,
(August 2006). Vol. 29, No. 5 (October 2003) 633–649, ISSN 1712-7971.
Dobson, M.C.; Pierce, L.; Sarabandi, K.; Ulaby, F.T. & Sharik, T. (1992a). Preliminary
analysis of ERS-1 SAR for forest ecosystem studies, IEEE Transactions on Geoscience
and Remote Sensing, Vol. 30 Issue 2, (March 1992) 203-210. ISN 0196-2892.
Dobson, M.C.; Ulaby, F.T.; Le Toan, T.; Beaudoin, A. & Kasischke, E.S. (1992b). Dependence
of radar backscatter on coniferous forest biomass, IEEE Transactions on Geoscience
and Remote Sensing, Vol. 30, Issue 2 (March 1992) 412-416. ISSN 0196-2892.
Dobson, M. C.; F. T. Ulaby, & L. E. Pierce, (1995). Land-Cover Classification and Estimation
of Terrain Attributes Using Synthetic Aperture Radar. Remote Sensing of
Environment, Vol. 51, No. 1, (January 1995) 199-214, ISSN 0034-4257.
Drieman, J.A.; Leckie, D.G. & Ahern, F.J. (1989). Multitemporal C-sar for forest typing in
Eastern Ontario, Proceedings IEEE International Geoscience and Remote Sensing
Symposium IGARSS’89 & 12th Canadian Symposium on Remote Sensing, Vol. 3, 1376 –
1378, Vancouver, British Columbia, Canada, July 1989.
Drieman, J.A.; Ahern, F.J. & Corns, I.G.W. (1989). Visual interpretation of multipolarization
C-sar imagery of Alberta Boreal Forest, Proceedings IEEE International Geoscience and
Remote Sensing Symposium IGARSS’89 & 12th Canadian Symposium on Remote
Sensing, Vol. 3, 1401-1405, Vancouver, British Columbia, Canada, July 1989.
Escalona-Maurice, M.J. (2006). Landscape of the Texcoco Municipality: morphological and
functional analysis in a cartographic model framework. State of Mexico. Mexican
United States. (In Spanish: El Paisaje del Municipio de Texcoco: análisis
morfológico y funcional en el marco de un modelo cartográfico. Estado de México.
Estados Unidos Mexicanos). Doctoral thesis. Departamento de Geografía,
Universidad de Alcalá, España. 529 p.
Freeman, A. (2007) Fitting a two-component scattering model to polarimetric SAR data from
forests. IEEE Transactions on Geoscience and Remote Sensing, Vol. 45, Issue 8, (August)
2583 – 2592. ISSN 0196-2892.
Harrell, P.A.; Bourgeau-Chavez, L.L.; Kasischke, E.S.; French, N. H. F. & Christensen, N.L.
(1995). Sensitivity of ERS-1 and JERS-1 radar data to biomass and stand structure in
Alaskan boreal forest, Remote Sensing of Environment, Vol. 54, Issue 3 (December
1995) 247-260, ISSN 0034-4257.
Harrell, P.A.; Kasischke, E.S.; Bourgeau-Chavez, L.L.; Haney, E. & Christensen, N.L. (1997).
Comparison of approaches to estimate of aboveground biomass in southern pine
forests using SIR-C data, Remote Sensing Environment. Vol. 59, Issue 3 (February
1997) 223-233. ISSN 0034-4257.
Hess, L.; Melack, J.M. & Simonett, D.S. (1990). Radar detection of flooding beneath the
forest canopy, International Journal of Remote Sensing Vol. 11, No. 7, 1313-1325. ISSN
0143-1161.
Hoekman, D.H. (1990). Radar remote sensing data for applications in forestry, Ph.D. thesis,
Wageningen Agricultural University, The Netherlands, 277 p.
Hoekman, D.H. (1993) Radar signature and forest vegetation, In: Land Observation by Remote
Sensing: Theory and Applications, J.G.P.W. Clevers and H.J. Buiten and (Eds), Current
Topics in Remote Sensing, vol.3, Gordon and Breach Science Pub. Ch. 11, 219-235.
ISBN 2-88124-939-6, New York, NY, USA.
Kasischke, E. & Bourgeau-Chavez, L. (1997). Monitoring South Florida wetlands using ERS-
1 SAR imagery. Photogrammetric Engineering and Remote Sensing, Vol. 63 No. 3,
(MONTH) 281-291. ISBN 0099-1112
Kasischke, E.S.; Christensen, N.L. Jr. & Bourgeau-Chavez, L.L. (1995). Correlating radar
backscatter with components of biomass in loblolly pine forests, IEEE Transactions
on Geoscience and Remote Sensing, Vol. 33, Issue 3 (May 1995) 643-659, 1995. ISSN
0196-2892.
Koch, B.; Foerster, B. & Ammer, U. (1992). The use of polarimetric radar remote sensing for
application in forestry, Proc. European Int. Space Year Conf. "Remote Sensing for
Environmental Monitoring and Resource Management", Munich (Germany), April
1992, ESA SP-341, 789-794.
Leblon, B. 1997. Soil and plant optical properties, The Remote Sensing Core Curriculum, vol.
4, module 9, http://www.r-s-c-c.org/rscc/Volume4/Leblon/leblon.htm
Leblon, B.; Kasischke, E. S.; Alexander, M. E.; Doyle, M. & Abbott, M. (2002) Fire danger
monitoring using ERS-1 SAR images in the case of northern boreal forests.. Natural
Hazards, 27, pp. 231-255.
Leblon, B. & LaRocque, A. (2008). Radar Polarimetry, online course, 11 chapters Accessed
26.09.08 http://extend.unb.ca/oalp/courses/for4304.php
Leckie, D.G. (1990). Advanced in remote sensing technologies for forest surveys and
management. Canadian Journal of Forestry Research. 20: 464-483. ISSN 0045-5067.
Le Toan, T.; Beaudoin, A.; Riom, J. & Guyon, D. (1992). Relating Forest Biomass to SAR
Data. IEEE Transactions on Geoscience and Remote Sensing, Vol. 30, Issue 2, (March
1992) 403-411, ISSN 0196-2892.
Lucas, R.M.; Lee, A.C & Williams, M.L. (2006). Enhanced Simulation of Radar Backscatter
From Forests Using LiDAR and Optical Data. IEEE Transactions on Geoscience and
Remote Sensing, Vol. 44, Issue 10, (October 2006) 2736-2754. ISSN 0196-2892.
Mas, J.F.; Velázquez, A.; Palacio-Prieto, J.L.; Bocco, G.; Peralta, A. & Prado, J. (2002).
Assessing forest resources in Mexico: Wall-to-wall land use/cover mapping.
Photogrammetric Engineering and Remote Sensing, Vol. 68, No 10, (October 2002)
966-968, Accessed 07.04.09
http://www.asprs.org/publications/pers/2002journal/october/
MDA, CSA & CCRS. (2009). The Science and Operational Applications Research for
RADARSAT-2 Program (SOAR). Canada.
Accessed 22.03.09 http://www.radarsat2.info/outreach/soar/index.asp
Mougin, E.; Lopes, A.; Karam, M.A. & Fung, A.K. (1993). Effect of tree structure on X-band
microwave signature of conifers. IEEE Transactions on Geoscience and Remote Sensing,
Vol. 31, Issue 3, (May 1993) 655-667. ISSN 0196-2892.
Myneni, R.B.; Hall, F.G.; Sellers, P.J.; Marshak, A.L. (1995). The interpretation of spectral
vegetation indices. IEEE Transactions on Geoscience and Remote Sensing, Vol. 33, Issue
2, (March 1995) 481 – 486, ISSN 0196-2892.
Nikolakopoulos, K.G. (2004). Comparison of four different fusion techniques for IKONOS
data. (2004) 1401-1405. Proceedings of IEEE International Geoscience and Remote
Sensing Symposium, 2004. IGARSS’04, Vol. 4, 2534 – 2537, ISBN 0-7803-8742-2,
Anchorage, Alaska, USA. September 2004.
Packalén, P. & Maltamo, M. (2008). Estimation of species-specific diameter distributions
using airborne laser scanning and aerial photographs. Can. J. For. Res. Vol. 38. (May
2008). 1750–1760, ISSN 1712-7971.
Palacio P. J.L.; Bocco, G; Velázquez, A.; Mas, J.F.; Takaki, F.; Victoria, A.; Luna, L.; Gómez,
G.; López, J.; Palma, M.; Trejo, I.; Peralta, A.; Prado, J.; Rodríguez, A.; Mayorga, R.
& González, F. (2000). The Current Condition of Forest Resources in Mexico:
Results From the National Forest Inventory (In Spanish). La Condición Actual de
los Recursos Forestales en México: Resultados del Inventario Forestal Nacional
2000. Investigaciones Geográficas, Boletín del Instituto de Geografía, UNAM. No. 43:
183-203. ISSN 0188-4611.
Papathanassiou, K.P. & Cloude, S.R. (2001). Single-baseline polarimetric SAR
interferometry, IEEE Transactions on Geoscience and Remote Sensing, Vol. 39, Issue 11
(November 2001) 2352–2363. ISSN 0196-2892.
Pearson, R.G. (2007). Species’ Distribution Modeling for Conservation Educators and
Practitioners. Synthesis. American Museum of Natural History. Accessed 07.01.08
http://ncep.amnh.org.
Piwowar, J.M.. (1997). Radar Image Characteristics 22. Complex Scattering Examples.
Geomatics International, RADARSAT Distance Learning Program.
uregina.ca/piwowarj/geog309/21-RadarImagery.pps
Pohl, C. & van Genderen, J. L. (1998). Multisensor image fusion in remote sensing: concepts,
methods and applications, International Journal of Remote Sensing, Vol. 19, No. 5,
(March 1998), 823-854. ISSN 0143-1161.
Pulliainen, J.T.; Mikkelä, P.J.; Hallikainen, M. T. & Ikonen, J.P. (1996). Seasonal dynamics of
C-band backscatter with applications to biomasss and soil moisture estimation.
IEEE Transactions on Geoscience and Remote Sensing, Vol. 34, Issue 3, (May 1996) 758-
770. ISSN 0196-2892.
Ranson, K.J. & Sun, G. (1994). Mapping biomass of a Northern forest using multifrequency
SAR data, IEEE Transactions on Geoscience and Remote Sensing, Vol. 32, Issue 2,
(March 1994) 388-396. ISSN 0196-2892.
Rauste, Y.; Häme, T.; Pulliainen, J.; Heiska, K. & Hallikainen, M. (1994). Radar-based forest
biomass estimation, International Journal of Remote Sensing, Vol. 15, No. 14,
(September 1994) 2797-2808. ISSN 0143-1161.
Richards, J.A.; Sun, G.Q. & Simonett, D.S. (1987). L-band radar backscatter modeling of
forest stands, IEEE Transactions on Geoscience and Remote Sensing, Vol. 25, Issue 4,
(July 1987) 487-498. ISSN 0196-2892.
Ridder, R. M. (2007). Global Forest Resources Assessment 2010- Options and

recommendations for a global remote sensing survey of forests. Forest Resources
Assessment Programme, Working paper 141, Rome. (March 2007). Accessed 02.02.09
http://www.fao.org/forestry/fra2010-remotesensing/en/
Rignot, E.; Williams, C.; Way, J. & Viereck, L. (1994a). Mapping of forest types in Alaskan
boreal forests using SAR imagery, IEEE Transactions on Geoscience and Remote
Sensing, Vol. 32, Issue 5 (September 1994) 1051-1059. ISSN 0196-2892.
Rignot, E.; Way, J.; Williams, C. & Viereck, L. (1994b). Radar estimates of above ground
biomass in boreal forests of Interior Alaska, IEEE Transactions on Geoscience and
Remote Sensing, Vol. 32, Issue 5, (September 1994). 1117-1125. ISSN 0196-2892.
Riom J. & Le Toan, T. (1981). Relations entre des types de fôrets de pins maritimes et la
retrodiffusion radar en bande L, Proceedings of Spectral Signatures of Objects in
Remote Sensing Symposium, 455-465, Avignon, France.
Sandoval Uribe, A. (2007). Advances of the National Forest and Soil Inventory 2004-2009. (In
Spanish: Avances del Inventario Nacional Forestal y de Suelos 2004 – 2009),
Proceedings Convención Nacional de Geografía 2007, 18-20 (April 2007), Guadalajara,
Jalisco, Mexico.
Schowengerdt, R.A. Remote Sensing, 2nd ed. (1997). Academic Press, ISBN 0-12-628981-6, San
Diego CA., USA.
Shi, J.; Wang, J.; Hsu, A.Y.; O’Neill, P.E & Engmann, T. (1997). Estimation of bare surface
soil moisture and surface roughness parameter using L-Band SAR image data, IEEE
Transactions on Geoscience and Remote Sensing, Vol. 35, Issue. 5, (September) 1254–
1265. ISSN 0196-2892.
Shiver, B. D. & Borders, B.E. (1996). Sampling Techniques for Forest Resource Inventory. John
Wiley & Sons. ISBN-10: 0471109401; ISBN-13: 978-0471109402. New York.
Soria-Ruiz, J.; Fernández-Ordóñez, Y. & Gonzalez, E. (2005). Atlas Estatal de Uso del Suelo y
Vegetación del Estado de México. INIFAP. Mexico. .
Soria-Ruiz, J.; Y. Fernandez-Ordoñez & I. Woodhouse. Land cover classification using radar
and optical images: a case study in Central Mexico. International Journal of Remote
Sensing, ISSN 0143-1161. (Accepted in 2008).
Soria-Ruiz, J. & Y. Fernández-Ordoñez. (2004). Improving Land-use/cover classification in
México. Fusing optical and radar microwave images. GIM International, No. 8, Vol.
18, (April 2004) 33-35 ISSN 1566-9076.
Staples, G.; Hornsby, J.; Branson, W.; O’Neill, K. & Rolland P. (2004). Turning the
scientifically possible into the operationally practical: RADARSAT-2
commercialization plan. Technical note. Canadian Journal of Remote Sensing, Vol. 30,
No. 3, (June 2004), 408–414. ISSN 1712-7971.
Sun, G. & Ranson, K.J. (2000). Modeling lidar returns from forest canopies. IEEE Transactions
on Geoscience and Remote Sensing, Vol. 38, Issue 6, (November 2000) 2617 – 2626.
ISSN 0196-2892.
Toledo V. M. & Ordóñez, M. J. (1993). The Biodiversity scenario of Mexico: a review of
terrestrial habitat. In: Biological diversity of Mexico: origin and distributions. T.P.
Ramamoorthy, R. Bye, A. Lott, J. Fa (Ed). 757-777, Oxford University Press, ISBN-
10: 019506674X, USA.
Touzi, R.; Landry, R. & Charbonneau, F.J. (2004). Forest type discrimination using
calibrated C-band polarimetric SAR data. Can. J. Remote Sensing, Vol. 30, No. 3, pp.
543–551, ISSN 1712-7971.
Valdez-Lazalde, J.R.; González-Guillén, M. de J. & De los Santos-Posadas H. M. (2009). Tree
crown cover estimation using high resolution multispectral satellite images.
Agrociencia, Vol 43, (May-June), 383-394, ISSN 1405-3195.
van der Sanden, J.J.; Thomas, S.J.; Lukowski, T.I.; Charbonneau, F.J.; DeAbreu, R.; Hawkins,
R.K.; Livingstone, C.E.; McNairn, H.; Scheuchl, B.; Singhroy, V.; Toutin, T.; Touzi,
R.; Vachon, P.W. (2005). Applications potential of RADARSAT-2 – update.
Proceedings of IEEE International Geoscience and Remote Sensing Symposium, 2005.
IGARSS’05, Vol. 1, 25-29, ISBN: 0-7803-9050-4, Seoul, South Korea.
Viergever, K.M.; Woodhouse, I.H. & Stuart, N. (2007). Backscatter and Interferometry for
Estimating Above- ground Biomass in Tropical Savanna Woodland. Proceedings of
IEEE International Geoscience and Remote Sensing Symposium 2007 IGARSS’07, 2346-
2349. ISBN 978-1-4244-1211-2, Barcelona, Spain.
Vuolo, F.; Dini, L. & D’Urso, G. (2008). Recent Advances in Quantitative Remote Sensing:
Second International Symposium, 25th-29th September 2006, Torrent, Spain.
International Journal of Remote Sensing, Vol. 29, Issue 17-18 (September 2008) Pages
5063-5072. ISSN 0143-1161.
Wald, L. Some Terms of Reference in Data Fusion. (1999). IEEE Transactions on Geoscience and
Remote Sensing, Vol. 37, No. 3, (May 1999) 1190-1193. ISSN 0196-2892.
Wang, Z.; Ziou, D.; Armenakis, C.; Li, D. & Li, Q. (2005). A comparative analysis of image
fusion methods, IEEE Transactions on Geoscience and Remote Sensing, Vol. 43, Issue 6,
(June 2005) 391- 1402, ISSN 0196-2892.
Yatabe, S.M. & Leckie, D.G. (1995). Clearcut and forest-type discrimination in satellite SAR
imagery, Can. J. Remote Sensing, 21(4): 455-467.
Zheng, G.; Chen, J.M.; Tian, Q.J.; Ju, W.M. & Xia, X.Q. (2007). Combining remote sensing
imagery and forest age inventory for biomass mapping. Journal of Environmental
Management 85, Issue 3 (November), 616–623, ISSN 0301-4797.
Application of Multi-Frequency Synthetic Aperture Radar (SAR) in Crop Classification 557
27
X
Application of Multi-Frequency Synthetic

Aperture Radar (SAR) in Crop Classification
Jiali Shang, Heather McNairn, Catherine Champagne and Xianfeng Jiao
Research Branch, Agriculture and Agri-Food Canada
Canada
1. Introduction
The application of remote sensing to agriculture has traditionally focused on the use of data
from optical sensors such as Landsat Thematic Mapper (TM) and SPOT. Due to cloud and
haze interference, however, optical images are not always available at phenological stages
important for crop discrimination. When gaps in data acquisition occur during critical
growth periods, classification accuracies using optical data are often inadequate (Jewell,
1989; McNairn et al., 2002; Blaes et al., 2005). Mid to late season optical images are essential
to achieve accurate crop classification, and this dependency on late-season data reduces the
ability to deliver early-season crop acreage estimates (McNairn et al., 2008a, b; Shang et al.,
2006, 2008). These constraints seriously impede the use of optical data for operational annual
crop mapping. Unlike visible and infrared wavelengths which are sensitive primarily to
plant biochemical properties, longer-wavelength microwave energy responds to the large-
scale structural attributes of vegetation, including the size, shape and orientation of the
leaves, stems, and fruits. The dielectric properties of the vegetation canopy also influence
the magnitude of the radar backscatter. These diverse sensitivities suggest that the
integration of data from optical and radar sensors will generate a synergistic effect. Recent
research has found that this complementarity, in most cases, provides enough information
to separate a wide variety of crop types when an integrated optical-radar dataset is used
(McNairn et al., 2008a; Shang et al., 2006, 2008).
1.1 Airborne multi-frequency SAR applications to agriculture

Although an integrated optical-radar approach can consistently discriminate crops, the
continued dependency on optical data, particularly in cloud-prone regions, is less than ideal
for operational delivery of crop information. A radar-only approach to crop discrimination
and acreage estimation would provide an operational advantage. Until recently, however,
the successful development of a radar-only approach to crop classification has been
hindered because of the availability of radar data with only limited dimensionality. Single
frequencies, and for some sensors single polarizations, do not provide enough information
for accurate discrimination even when multi-temporal acquisitions are exploited (Shang et
al, 2006, 2008). Research campaigns based on airborne SAR acquisitions have explored the
benefits of multi-frequency SAR for crop discrimination. The Jet Propulsion Laboratory’s
AirSAR is a fully polarimetric SAR sensor operating at P- (0.45 GHz), L- (1.25 GHz), and C-
(5.31 GHz) bands (Lee and Pottier, 2009). The value of multi-frequency fully polarimetric
data has been demonstrated for many land applications. For example, Rao et al. (1993)
studied multi-frequency (P-, L-, C-band) polarimetric AirSAR data over corn fields. It found
that the mean polarization phase difference increases with increasing wavelength. Lemonie
and associates (1994) used the AirSAR data to study the contribution of multi frequency
radar to increased agricultural class separabilities. The study by Baronti and associates
(1995) carried out a three-frequency (P-, L-, and C-band) AirSAR data analysis. It found that
P-band data are effective only in discriminating broad classes of agricultural landscapes.
The integration of L- and C-band helps reveal finer class details.
Much research on the advantages of multi-frequency SAR has also been conducted with
radar scatterometers. For example, Snoeij et al. (1990) used C- and X-band airborne SAR
data to study the general behaviour of the radar signature of different European test sites as
a function of frequency. The study conducted by van Leeuwen (1992) used six-frequency (L-
band at 1.2 GHz; S-band at 3.2 GHz; C-band at 5.3 GHz; Ku1-band at 13.7 GHz; and Ku2-
band at 17.3 GHz) radar scatterometer data over beet and wheat fields to examine the
physical meaning of radar model (CLOUD-model: Attema & Ulaby, 1978) parameters in
relation to crops. More recently Inoue et al. (2002) studied multi-frequency (Ka-, Ku-, X-, C-
& L-band) radar backscattering signatures over paddy rice fields and their relationship with
rice canopy growth variables. This research demonstrated that the backscatter coefficients of
higher frequency bands (Ka and Ku) are highly correlated with the weight of heads. Lower
frequency bands such as L-band, are better correlated with fresh biomass while C-band is
better correlated with leaf area index.
1.2 Space-borne multi-frequency SAR applications to agriculture

Airborne SAR systems enabled radar scientists to develop methods to derive information of
interest from multi-frequency and multi/full-polarization SAR, often under controlled
experimental conditions. Airborne sensors provide a far greater signal to noise ratio and much
higher spatial resolution compared with spaceborne sensors. However, these airborne
platforms are not suited for large scale operational campaigns. With their wide swaths and
repeat orbits, spaceborne sensors provide a cost effective solution for operational activities.
Since the launch of the first spaceborne radar system in 1965, Radar Evaluation Pod (REP),
many spaceborne radar sensors have been launched (Lacomme et al., 2001). Earth Resources
Satellite (ERS-1 launched in 1987 and ERS-2 launched in 1995) data have been used to
develop many applications in agriculture, wetlands and forestry (Ban & Howarth, 1999;
Bouman et al., 1992; Engdahl et al., 2001; Kohl et al., 1994; Michelson et al., 2000; Paudyal et
al., 1995; Wang et al., 1998). The frequency-polarization (C-HH) of RADARDAT-1 (launched
in 1995) was selected to maximize information for marine applications including sea ice and
ocean features. Nevertheless, scientists developed the use of these data for a wide range of
land applications including agriculture (McNairn et al., 1998a, b, c; Phoompanich et al., 2005;
Ribbes & Toan, 1999; Shang et al., 2006).
The next generation European C-Band sensor, ASAR, has further advanced the use of SAR
for agricultural mapping. The availability of dual like-polarizations (HH-VV) or dual like-
cross polarizations (HH-HV or VV-VH) with ASAR, assists in providing more information
on vegetation type and condition. Radar backscatter varies from one polarization to another
since interaction is dictated by the transmitting polarization relative to the horizontal and
vertical structure of the canopy. Consequently sensors which have polarization diversity
provide more information on both crop structure and crop condition. The launch of ALOS
PALSAR (L-band: 1.27 GHz) in 2006, followed by TerraSAR-X (X-band: 9.6 GHz) and
RADARSAT-2 (C-band: 5.405 GHz) in 2007, marked the beginning of the multi-frequency
spaceborne SAR era. Availability of data from this suite of satellites has accelerated the use
of SAR for land applications. When used together, a multi-frequency dataset from multiple
SAR platforms holds the promise to provide exceptional information for agriculture. Using
a Canadian example, this chapter demonstrates the application of integrated L-, C-, and X-
band SAR for crop mapping.
2. Study sites and data collection

2.1 Study sites
In 2006 Agriculture and Agri-Food Canada (AAFC) established two research sites close to
their Ottawa (Ontario, Canada) research station. The Canadian Food Inspection Agency
(CFIA) site is a controlled experimental site within the city of Ottawa (centred at 45°13’N,
75°46’W). The second site, Casselman, is a region of private land ownership (centred at
45°37’N, 75°01’W) approximately 50 km east of Ottawa. Both sites support non-irrigated dry
land farming with one crop grown during the relatively short May to September growing
season. The size of the fields in this part of Canada tends to be relatively small, 20 ha on
average. These sites are typical of the crop mix found in this part of Canada, with
production acreages primarily consisting of corn, soybean, cereal and pasture-forage.
2.2 Satellite data collection

(a) CFIA site
Satellite data were acquired from optical sensors (Landsat-5) as well as SAR sensors
(RADARSAT-1, Envisat-ASAR, and ALOS PALSAR) throughout the 2006 growing season.
Acquisitions were targeted to capture crop growth stages of importance for crop
discrimination using optical and SAR sensors (Table 1).
Date Resolution (m) Mode Polarization Incidence Angle

Landsat-5 TM
June 5 30
July 7 30
August 22 30
Envisat ASAR
May 27 30 IS3 VV, VH 25.8° - 31.2°
June 9 30 IS1 VV, VH 14.5° - 22.1°
July 1 30 IS3 VV, VH 25.8° - 31.2°
July 14 30 IS1 VV, VH 14.5° - 22.1°
August 5 30 IS3 VV, VH 25.8° - 31.2°
September 18 30 IS4 VV, VH 30.8° - 36.1°
RADARSAT-1
May 18 30 S1 HH 24° - 31°
July 5 30 S1 HH 24° - 31°
August 22 30 S1 HH 24° - 31°
ALOS PALSAR
May 19 10 PLR Polarimetric 21.5°
July 4 10 PLR Polarimetric 21.5°
August 19 10 PLR Polarimetric 21.5°
Table 1. Satellite data acquired over CFIA during the 2006 growing season
(b) Casselman site

Optical and radar satellite data were collected over the Casselman site during the 2008
growing season. No cloud-free (less than 20% cloud cover) Landsat data were available due
to poor weather conditions throughout the summer of 2008. The SPOT sensor was
programmed in two week windows through the entire 2008 season. This acquisition
strategy yielded four SPOT-4 images. Six TerraSAR-X scenes were also acquired. Due to the
late start of the TerraSAR-X project, X-band data acquisition did not commence until mid
July. Table 2 gives the details of each satellite acquisition.
Date Resolution (m) Mode Polarization Incidence Angle
SPOT-4
June 5 20
July 7 20
August 22 20
RADARSAT-2
May 27 10 FQ19 quad-pol 38.3° - 39.8°
June 9 10 FQ19 quad-pol 38.3° - 39.8°
July 1 10 FQ19 quad-pol 38.3° - 39.8°
July 14 10 FQ19 quad-pol 38.3° - 39.8°
TerraSAR-X
July 19 6 Stripmap VV, VH 43.6° - 44.6°
July 30 6 Stripmap VV, VH 43.6° - 44.6°
August 10 6 Stripmap VV, VH 43.6° - 44.6°
August 21 6 Stripmap VV, VH 43.6° - 44.6°
August 26 6 Spotlight HH, VV 53.9°
September 1 6 Stripmap VV, VH 43.6° - 44.6°
Table 2. Satellite data acquired over Casselman during the 2008 growing season
2.3 Ground data collection

For both sites, ground truth observations were collected twice over the growing season,
once in early July and once in mid August. The second visit provided an opportunity to
check for errors which might have occurred during the first field visit. During the second
visit, variations in crop growth condition, harvesting, and emergence of under seeded crops
were also noted. Underseeding is a cropping system where a primary species is seeded with
a successive species that emerges at the end of the primary species growth cycle, for
instance, where annual cereal crops are underseeded with a perennial forage crop such as
alfalfa. Wheat in this region is usually harvested between late July and mid August
depending on the planting date. Underseeded wheat fields are thus characterized by rapid
growth of forage after wheat harvest. Harvesting of corn and soybean typically occurs near
the end of October.
In 2006, a total of 240 fields were visited. Table 3 gives details on the number of fields
surveyed per crop.
Crop Type Number of Training Fields Number of Testing Fields

Cereal 16 17
Corn 35 35
Soybean 30 30
Forage/Pasture 38 39
Table 3. Ground truth data used for CFIA 2006 crop classification
For the Casselman site, a total of 247 fields were surveyed during the 2008 growing season.
The distribution of field surveyed is documented in Table 4.
Crop Type Number of Training Fields Number of Testing Fields

Cereal 17 18
Corn 45 45
Soybean 41 41
Forage/Pasture 33 34
Table 4. Ground truth data used for Casselman 2008 crop classification
3. Data pre-processing
3.1 Atmospheric correction of optical data
Atmospheric correction was applied to all optical data to retrieve the at-surface reflectance
using the Atcor algorithm in PCI software (Richter, 2004). The Atcor algorithm uses the
MODTRAN 4.2 radiative-transfer code for the radiance to reflectance conversion
(Champagne et al., 2005).
3.2 Speckle filtering of SAR data

Speckle is an inherent phenomenon for coherent systems such as SARs. To suppress speckle,
adaptive radar filters should be applied prior to classification of SAR data. All ALOS
PALSAR, RADARSAT-2, and TerraSAR-X data were speckle filtered, using a 5 X 5 Gamma-
MAP filter.
3.3 Geometric correction and co-registration

For the purposes of integrating the various data sources, and in order to facilitate
comparisons with ground data, all of the Landsat, SPOT, PALSAR, TerraSAR-X and
RADARSAT-2 data were geocorrected and registered to the same coordinate system (UTM).
A nearest neighbor re-sampling method was adopted with an output resolution of 10 m.
3.4 Trainning and validation site selection

To reduce bias, the training and validation pixels were selected from different fields. For
each crop type, training and validation fields were selected randomly from the total ground
truth data set in ArcGIS (Figure 1). As a first step, a 10m buffer was applied to each field
boundary. The pixels within this boundary buffer zone were excluded from training and
validation to reduce contamination from headlands and mixed pixels. Half of the fields
surveyed were randomly selected and used for training the classification algorithm. The
remaining fields were reserved for quantifying the accuracy of the classification. As a result,
there was no overlap between training and validation pixels.
Field boundary for corn
Training site for corn
Validation site for corn
Fig. 1. An example showing the spatial arrangement of the training and validation fields.
4. Methodology
The type of classification methods used can greatly impact the classification results. When
adequate ground truth data are available, supervised classification approaches generally
produce better results relative to unsupervised classifications. Consequently for this study, a
supervised classification was selected. The choice of classification algorithm is influenced by
many factors, including data requirement, sensitivity to variation of training data, and
computational requirements. This study adopted a supervised decision tree (DT) classifier
(McNairn et al., 2008a). DT takes a sequential classification approach (Pal & Marther, 2003).
This non-parametric classifier is appropriate for use with SAR data, which typically are not
normally distributed. A DT classifier can also handle data gaps which are commonly
encountered when cloud masking is applied to optical data.
AAFC developed an in-house DT graphical user interface (GUI) which integrates PCI
Geomatica and the See5 softwares (Rulequest Research, 2008). The DT classifier was run
using boosting over 5 trials with a global pruning of the model of 25%. All classifications
were performed on a per pixel basis without a null class.
Pixel-based classifications often result in a salt-pepper appearance, especially when radar

data are used. Therefore a post-classification filter was applied to the resultant maps. For
this study, spatial filtering was accomplished using segments created witin eCognition and
assigning the mode class to each segment. The filtered maps are visually more consistent
and exhibit increased classification accuracies.
5. Results and discussion

5.1 Single-frequency classification performance comparison
The classification accuracies for single frequency imagery are being discussed in this section.
DT classifications were run using single frequencies (L- and C-band) to assess which radar
wavelength provides the highest accuracies. To facilitate this comparison, only L- and C-
band data collected close in time were used.
For the CFIA site, three pairs of data collected in 2006 were compared using data with the
same polarization. Comparisons were restricted to pairs of data acquired within a seven-day
window to avoid significant variations caused by plant growth between the two acquisition
dates.
Polarization
Pasture/ Overall
Sensors Frequency Used for Date Soybean Corn Wheat
Forage Accuracy
Comparison
PALSAR L-band VV/VH May 20 24.1 57.5 84.1 1.4 49.7
ASAR C-band VV/VH May 27 60.0 50.7 81.7 4.9 55.7
PALSAR L-band VV/VH July 5 18.0 67.0 86.7 11.8 54.0
ASAR C-band VV/VH July 1 70.6 65.2 88.8 33.0 68.1
May 20
PALSAR L-band HH 55.8 49.3 83.8 9.2 54.8
July 5
May 18
RSAT-1 C-band HH 79.8 54.0 61.0 8.1 52.8
July 5
Table 5. Comparison of single- and two-date PALSAR, ASAR, and RADARSAT-1 2006 crop
classification accuracies (producer’s) over the CFIA site
For the two acquisition windows (late May and early July) and considering overall accuracy,
C-band data performed better than the L-band data using VV and VH polarizations. For
larger biomass crop such as corn, the two frequencies (VV/VH) are comparable. For lower
biomass crops, such as forage, the shorter wavelength C-band performs better.
When two dates of HH SAR data are used (one in May and one in July), L-band produced
an overall accuracy of 54.8%, slightly higher than C-band’s 52.8%. For larger biomass corn
crops, L-band performs significantly better than C-band with accuracies of 83.8% and 61.0%,
respectively. For lower biomass crops, such as cereal and pasture-forage, L-band was less
effective. With lower biomass and a less random vegetation structure, greater penetration
into the crop canopy due to the longer wavelength can be expected, which would include
greater contribution from the underlying soil, as well as from vegetation-soil interactions
(Freeman & Durden, ; Hill et al., ). C-band outperforms L-band for lower biomass crops.
For the Casselman study site, two pairs of TerraSAR-X (TSX) and RADARSAT-2 (RSAT-2)
imagery collected close in time were selected for comparison. Table 6 documents the
classification results derived using data acquired over Casselman during the 2008 growing
season.
Polarization
Pasture/ Overall
Sensors Frequency Used for Date Soybean Corn Wheat
Forage Accuracy
Comparison
TSX X-band VV/VH July 16 69.1 58.2 48.0 85.2 59.9
RSAT-2 C-band VV/VH July 19 47.1 64.7 50.7 42.8 54.2
TSX X-band VV/VH August 9 59.1 79.4 71.0 61.8 71.0
RSAT-2 C-band VV/VH August 10 39.6 73.8 56.2 36.8 57.4
Table 6. Comparison of single-date TerraSAR-X and RADARSAT-2 crop classification
accuracy (producer’s) over the Casselman site from the 2008 growing season
For both acquisition windows (mid July, early August), X-band data outperformed the C-
band data. When comparing overall accuracies, the mid July X-band data produced a crop
map with an accuracy 5.7% higher than for C-band data acquired only three days later.
Comparing data acquired in early August, X-band provided significantly better accuracies -
an overall accuracy of 71% or 13.6% higher than the C-band data. In August the X- and C-
Band data were acquired only one day apart. Examining the individual class accuracies, X-
band performed better in identifying all crop types later in the season. Among all crop
types, X-band provided dramatically higher accuracies for wheat. A 42.4%increase for mid
July and 25% increase for early August are noted for the wheat class, when X-band results
are compared with those of C-band. For the same wheat class, X-band data also performed
better than C-band later in the growing season. At mid season (mid July), results derived
from X- and C-band are similar for corn, with a difference of less than 3%.
5.2 Multi-frequency classification comparison

To evaluate the benefits of a multi-frequency SAR approach for crop classification, four
datasets from the CFIA site were analyzed. Table 7 provides the classification accuracies
derived using single-frequency (L- or C-band) and two-frequency (L- and C-band)
approaches.
Polarization
Pasture/ Overall
Sensors (Date) Frequency Used for Soybean Corn Wheat
Forage Accuracy
Comparison
1 ALOS (May20) L-band VV/VH 24.1 57.5 84.1 1.4 49.7
1 ASAR (May27) C-band VV/VH 60.0 50.7 81.7 4.9 55.7
1 ASAR (May27) C- & L-
VV/VH 60.9 61.9 70.8 40.8 60.6
+ 1 ALOS (May20) band
1 ALOS (July 5) L-band VV/VH 18.0 67.0 86.7 11.8 54.0
1 ASAR (July 1) C-band VV/VH 70.6 65.2 88.8 33.0 68.1
1 ASAR (July 1) C- & L-
VV/VH 61.1 77.6 92.4 48.4 73.9
+ 1 ALOS (July 5) band
2 ALOS (May20, Jul5) L-band HH 55.8 49.3 83.8 9.2 54.8

2 RS1 (May18, Jul5) C-band HH 79.8 54.0 61.0 8.1 52.8
2 ALOS (May20, Jul5) C- & L-
HH 87.5 73.2 82.1 24.8 69.8
+ 2 RS1 (May18, Jul5) band
2 ALOS (May20, Jul5) L-band VV/VH 54.2 67.4 83.6 37.5 64.7
2 ASAR (May27, Jul1) C-band VV/VH 74.1 61.2 91.6 52.6 72.2
2 ASAR (May 27, Jul 1) C- & L-
VV/VH 72.4 84.5 95.7 69.6 82.9
+ 2 ALOS (May20, Jul 5) band
Table 7. Producer’s accuracies using multi-frequency SAR classifications from the 2006 data
acquired over CFIA
Results in Table 7 clearly confirm the benefits of a multi-frequency solution for crop
identification. For a single-date dual-polarization (VV/VH) comparison, there is an increase
of 4.9% in overall accuracy in late May compared to result derived using ASAR alone. When
compared with result derived from May ALOS, there is an increase of 10.9%. When early
July L- and C-band data are integrated together in the classifier, a similar improvement in
accuracy (5.8%) was observed. When multiple dates (one in May and one in July) of L- and
C-band were used a 15% gain in accuracy was observed, even though only a single
polarization (HH) was used. Two dates of dual-frequency and dual-polarization SAR from
PALSAR (L-band) and ASAR (C-band) produced a map with an overall accuracy of 82.9%.
For the Casselman site, comparisons were made between classifications using a single
frequency (C- or X-band) and results achieved by integrating these two frequencies (Table
8). The multi-temporal X-band data on its own was capable of identifying crops with an
overall accuracy of 84.9%. Consequently, adding C-band SAR to the classification brought
only modest improvements in overall accuracy. Nevertheless C-band did assist in boosting
accuracies for most individual crop classes.
Polarization
Sensors Pasture/ Overall
Frequency Used for Soybean Corn Wheat
(Date) Forage Accuracy
Comparison
4 RSAT-2 C-band VV/VH 66.2 82.9 76.1 64.0 75.4
5 TerraSAR-X X-band VV/VH 83.2 83.6 87.3 84.1 84.9
4 RSAT-2 + 5 C- & X-
VV/VH 84.1 86.8 89.9 85.6 87.3
TerraSAR-X band
Table 8. Producer’s accuracies of multi-frequency SAR classification from 2008 growing
season over Casselman
6. Conclusions and future research

A multi-year and multi-site study by Agriculture and Agri-Food Canada demonstrated the
improvements brought by integrating multiple frequencies (L-, C-, and X-band) of SAR data
for crop classification. Penetration into the crop canopy is dependent upon SAR frequency
and results indicate that the differences in this depth between frequencies are advantageous
for crop identification. The case study presented here concludes that when multi-temporal
multi-frequency SAR data are used, satisfactory crop classification (above 85% accuracy) can
be achieved using a SAR-only dataset.
Even with these promising results, further improvements in accuracy would be desirable
prior to implementing a radar-alone solution for crop classifications. The acquisition
planning associated with the datasets used in the research was limited by several factors.
TerraSAR-X data collection did not begin until mid season due to a late start in the project.
In future growing seasons, a more complete data set will be collected. This study also did
not permit comparisons among all three frequencies as TerraSAR-X, ASAR, RADARSAT
and PALSAR data were not all collected over either site. The programming of PALSAR in
concert with TerraSAR-X and RADARSAT-2 was not successful. In future growing seasons,
all three sensors have been programmed over the Casselman site. These methods will also
be evaluated in future growing seasons over a third site in the Canadian prairies, which will
represent a more complex cropping system with a greater variety of crops. Lastly,
acquisitions of data in RADARSAT-2’s polarimetric mode will permit assessment of
polarimetric parameters derived from multi-frequency SAR for improved crop
classification.
7. Acknowledgments
Funding of this research project was provided by the Canadian Space Agency’s Government
Related Initiatives Project (GRIP) and AAFC Research Branch’s A-base project. The
TerraSAR-X data were provided under the TerraSAR-X AO project LAN0337. The authors
wish to acknowledge the assistance of Dr. Ridha Touzi (Canada Centre for Remote Sensing)
in sharing his expertise on processing of the PALSAR data. We would like to thank JAXA
for having provided the ALOS data under the PI project 228.
8. References
Attema, E.P.W. & Ulaby, F.T. (1978). Vegetation model as a water cloud. Radio Science, Vol.
13, pp. 357-364.
Ban, Y. & Howarth, P. J. (1999). Multitemporal ERS-1 SAR data for crop classification: a
sequential-masking approach. Canadian Journal of Remote Sensing, Vol. 25, pp. 438-
447.
Berger, Z.; Irving, R. & Clark, J. (2009). Exploration applications of RADARSAT imagery in
the Foothills and Western Canadian Basin, Intermap whitepapers:
<http://www.intermap.com/uploads/1170701249.pdf>.
Blaes, X., Vanhalle, L, and Defourny, P. (2005). Efficiency of crop identification based on
optical and SAR image time series, Remote Sensing of Environment 96: 352-365.
Bouman, B.A.M. & Uenk, D. (1992). Crop classification possibilities with radar in ERS-1 and
JERS-1 configuration. Remote Sensing of Environment, Vol. 40, pp. 1-13.
Champagne, C., Shang, J., and McNairn, H. (2005). Exploiting spectral variation from crop
phenology for agricultural land-use classification, Proceedings of SPIE Optics and
Photonics, Vol. 5884, San Diego, CA, USA, July–August, 2005 (CD ROM).
Engdahl, M.E.; Borgeaud, M. & Rast, M. (2001). The use of ERS-1/2 Tandem interferometric
coherence in the estimation of agricultural crop heights. IEEE Transactions on
Geoscience and Remote Sensing, Vol. 39, No. 8, pp. 1799-1806.
Freeman, A. & Durden, S.L. (1998). A three-component scattering model for polarimetric
SAR data. IEEE Transactions on Geoscience and Remote Sensing, Vol. 36, pp. 963-
973.
Hill, M.J.; Ticehurst, C.J.; Lee, J.S.; Grunes, M.R.; Donald, G.E. & Henry, D. (2005).
Integration of optical and radar classifications for mapping pasture type in western
Australia. IEEE Transactions on Geoscience and Remote Sensing, Vol. 43, pp.1665-
1681.
Inoue, Y.; Kurosu, T.; Maeno, H.; Uratsuka, S.; Kozu, T.; Dabrowsk-Zielinska, K. & Qi, J.
(2002). Season-long daily measurements of multifrequency (Ka, Ku, X, C, and L)
and full-polarization backscatter signatures over paddy rice field and their
relationship with biological variables. Remote Sensing of Environment, Vol. 81, pp.
192-204.
Kohl, H.G.; Nezry, E.N.; Mróz, M. & De Groof, H. (1994). Towards the integration of ERS
SAR data in an operational system for rapid estimates of acreage at the level of the
European Union, Proceedings of the 1st Workshop on ERS-1 Pilot Projects, pp.433-
441, Toledo, Span, June1994.
Lacomme, P.; Marchais, J-C. & Normant, E. (2001). Air and Spaceborne Radar systems: an
introduction. William Andrew Publishing, Norwich, NY, USA, 504p.
Lee, J & Pottier, E. (2009). Polarimetric Radar Imaging: From Basics to Applications, 433p.,
9781420054972, CRC/Taylor & Francis Group, New York, NY, USA.
Lemonie, G.G.; de Grandi, G.F. & Sieber, A.J. (1994). Polarimetric contrast classification of
agricultureal fields using MASTRO1 AirSAR data. International Journal of Remote
Sensing, Vol. 15, pp. 2851-2869.
McNairn, H.; Wood, D.; Gwyn, Q.H.J.; Brown, R.J. & Charbonneau, F. (1998a). Mapping
tillage and crop residue management practices with RADARSAT, Canadian Journal
of Remote Sensing, Vol. 24, No. 1, pp. 28-35.
McNairn, H.; Brown, R.J. & Wood, D. (1998b). Incidence angle considerations for crop
mapping using multi-temporal RADARSAT data, Proceedings of the 20th Canadian
Symposium on Remote Sensing, pp. 211-214, Calgary, AB, Canada, May 1998.
McNairn, H.; Wood, D. & Brown, R.J. (1998c). Mapping crop characteristics using
multitemporal RADARSAT images, Proceedings of 1st International Conference:
Geospatial Information in Agriculture and Forestry, Vol. 2, pp. 501-507, Orlando,
USA, June 1998.
McNairn, H.; van der Sanden, J.J.; Brown, R. & Ellis, J. (2000). The potential of RADARSAT-2
for crop mapping and assessing crop condition, Proceedings of 2nd International
conference on Geospatial Information in Agriculture and Forestry, Vol. 2, pp. 81-88,
Lake Buena Vista, Florida, USA.
McNairn, H, Ellis, J., van der Sanden, J J, Hirose, T. and Brown, R J. 2002. Providing crop
information using RADARSAT-1 and satellite optical imagery. International
Journal of Remote Sensing 23:851-870.
McNairn, H.; Champagne, C., Shang, J., Holmstrom, D. and Reichert, G. (2008a). Integration
of optical and Synthetic Aperture Radar (SAR) imagery for delivering operational
annual crop inventories. Photogrammetry and Remote Sensing (in press:
doi:10.1016/j.isprsjprs.2008.07.006)
McNairn, H., Shang, J., Jiao, X., and Champagne, C. (2008b). The Contribution of ALOS
PALSAR Multi polarization and Polarimetric Data to Crop Classification, IEEE
Transactions on Geoscience and Remote Sensing, accepted with modifications.
Michelson, D.B.; Liljeberg, b.M. & Pilesjo, P. (2000). Comparison of algorithms for classifying
Swedish landcover using Landsat TM and ERS-1 SAR data. Remote Sensing of
Environment, Vol. 71, No. 1, pp. 1-15.
Pal, M. & Mather, P.M. (2003). An assessment of the effectiveness of decision tree methods
for land cover classification. Remote Sensing of Environment, Vol. 86, pp. 554-565.
Paudyal, D.R.; Eiumnoh, A. & Aschbacher, J. (1995). Multitemporal analysis of SAR images
over the tropics for agriculture applications, Proceedings of the International
Geoscience and Remote Sensing Symposium, pp. 1219-1221, 0-7803-2567-2, Firenze,
Italy, July 1995, IEEE.
Phoompanich, S.; Anan, T. & Phongam, S. (2005). Multi-temporal RADARSAT for crop
monitoring, Proceedings of 26th Asian Conference on Remote Sensing, Hanoi,
Vietnam, November 2005
< http://www.aars-acrs.org/acrs/proceeding/ACRS2005/Papers/AGC2-3.pdf>.
Rao, K.S.; Rao, Y.S. & Gurusamy, R. (1993). Frequency dependence of polarization phase
difference and polarization index for vegetation covered fields using polarimetric
AirSAR data, Proceedings of the International Geoscience and Remote Sensing
Symposium, pp. 37-39, Tokyo, Japan, August 1993, IEEE, Piscataway, N.J., U.S.A.
Ribbes, F. & Toan, T.L. (1999). Rice field mapping and monitoring with RADARSAT data.
International Journal of Remote Sensing, Vol. 20, No. 4, pp. 745-765.
Rulequest Research (2008). Data Mining Tools See5 and C5.0 [Online]. Available by
Rulequest Research http://www.rulequest.com/see5-info.html.
Shang, J., McNairn, H., Champagne, C. and Jiao, X. (2008). Contribution of multi-frequency,
multi-sensor, and multi-temporal radar data to operational annual crop mapping.
International Geoscience and Remote Sensing Symposium, Boston (Massachusetts),
7-11 July 2008, 4 pp. Invited paper.
Shang, J., Champagne, C., and McNairn, H. (2006). Agriculture land use using multi-sensor
and multi-temporal Earth Observation data. In Proceedings of the MAPPS/ASPRS
2006 Fall Specialty conference, San Antonio, Texas, USA.
Snoeij, P.; Swart, P.J.F. & Attema, E.P.W. (1990). The general behavior of the radar signature
of different European test sites as a function of frequency and polarization,
Proceedings of the 10th International Geoscience and Remote Sensing Symposium, pp.
2315-2318, May 1990, College Park, Maryland, USA.
Skriver, H.; Svendsen, M.T. & Thomsen, A.G. (1999). Multitemporal C- and L-band
polarimetric signatures of crops. IEEE Transactions on Geoscience and Remote
Sensing, Vol. 24, No. 5, pp. 2413-2429.
Van Leeuwen, H.J.C. (1991). Multifrequency and multitemporal analysis of scatterrometer
radar data with respect to agricultural crops using the Cloud model, Proceedings of
the 11th International Geoscience and Remote Sensing Symposium, Vol. 4, pp. 1893-1897,
Helsinki, Finland.
Wang, J.; Shang, J.; Brisco, B. & Brown, J. (1998). Evaluation of multidata ERS-1 and
multispectral Landsat imagery for wetland detection in southern Ontario. Canadian
Journal of Remote Sensing, Vol. 24, No. 1, pp. 60-68.
Multispectal Image Classification Using Rough Set Theory and Particle Swam Optimization 569
28
X
Multispectal Image Classification Using Rough

Set Theory and Particle Swam Optimization
Chih-Cheng Hung, Hendri Purnawan, Bor-Chen Kuo* and Scott Letkeman
Southern Polytechnic State University, Marietta, GA 30060 USA
*National Taichung University, Taichung, Taiwan
Abstract
This chapter provides an exploration of the rough set theory and particle swarm
optimization for multispectral image classification. The rough set theory, particle swarm
optimization and K-means clustering algorithm are briefly described. Two multispectral
image classification algorithms based on the rough set theory and particle swarm
optimization are proposed: Algorithm 1 is a multispectral image classification approach
based on the rough set theory which uses upper and lower bounds for the class description,
and Algorithm 2 is a hybrid rough K-means algorithm for image classification. The rough
set theory is used to extract classification rules and establish the lower and upper bounds for
data clustering with the K-means algorithm. This algorithm is able to deal with vagueness in
image date, but since its ability to determine some key parameters is limited, partial swarm
optimization must subsequently be used to locate optimal values for those parameters.
Experimental results show that the proposed algorithms perform well and improve the
classification in the blurred and vague areas of the image. A comparison of Algorithm 1
with the parallelepiped classifier, where the former uses the concept of cuts and the later
uses the maximum and minimum values, is performed. Preliminary experimental results
show that the proposed classifiers are effective for multispectral image classification.
1. Introduction
In the real world, data representation is most often imperfect, in the sense that the data may
be either incomplete or redundant. Philosophers, logicians and mathematicians have dealt
with this problem for a long time. In recent years, propelled by the advent of the computer,
the problem of imperfect knowledge has been becoming an important topic for computer
scientists engaged in artificial intelligence research, especially those involved with
knowledge discovery from databases, expert systems, and pattern recognition.
Our research is focused on rough set as a tool for image processing, or more precisely, for
image segmentation. Many techniques for image segmentation have been developed over
time. There are clustering, edge detection, region growing and even more advanced
techniques that use neural networks. In general, image segmentation techniques can be
categorized as supervised or unsupervised. Supervised techniques require previously

known truth data for training purposes, while unsupervised techniques have no such
requirement
In this chapter, the classical rough set theory is reviewed in section 2. Particle swarm
optimization is then introduced in section 3. The Davies-Bouldin measure for cluster
validity is also described in this section. The K-means algorithm is briefly sketched in
section 4. Multispectral image classification using rough set theory is discussed in section 5.
A hybrid algorithm which combines the K-means algorithm, rough set and particle swarm
optimization is given in section 6. Experimental results are shown in section 7. The
conclusion and future work then follow.
2. Rough Set Theory

Rough set theory [5] is a mathematical tool that deals with the uncertainty of the data. The
theory consists of finite sets, equivalence relations and cardinality concepts. As the theory
matures and more applications reap the benefits of the concept, an abundance of related
theorems and algorithms are being incorporated to extend rough sets theory.
It was introduced by Pawlak in the early 1980’s and has been argued to overlap with other
theories, such as statistics, evidence theory and fuzzy set. Furthermore, rough set is said to
complement fuzzy set, a theory introduced by Zadeh in the early period. Rough set and
fuzzy set were both introduced to deal with imprecise information however; fuzzy set deals
with vagueness, while rough set deals with coarseness. Rough set does not need as much
preliminary knowledge about the data where as fuzzy set requires knowledge of the
possible values in advance. Basically, when using rough set, the data itself is used to come
up with the approximation in order to deal with the imprecision within. It can therefore be
considered a self-sufficient discipline.
Rough set mainly deals with data analysis in table format. The approach is generally to pre-
process the data in the table and then to analyze them. Reducts are extracted with an
algorithm and finally rules are generated based on the reducts. Rough set does not support
analog values in the table attributes; therefore discretization must be performed in advance
in order to evaluate the table. The following subsections will use a simple example to
illustrate the concept of rough set theory.
2.1 Information Systems

In essence, an information system is a set of objects represented in a data table (attribute -
value system). Each row contains an object and each column represents a measurable
attribute for each object. Formally, an information system is a pair A = (U, A) where U is a
non- empty finite set of objects representing the universe and A is a non-empty finite set of
attributes such that a: U→Va for every a ∈ A. The set Va is the set of values for a.
Object Index Salary Age

u1 80 30
u2 30 23
u3 80 40
u4 50 45
u5 80 55
u6 50 45
u7 30 60
u8 100 35
Table 1. shows an information system which is a collection of salary and age attributes.
2.2 Decision Systems

If an information system has an additional attribute, namely a decision attribute, then it
becomes a decision system. The decision attribute is associated with the object classification
outcome, and it may depend on several other attributes. Formally, a decision system is a
piece of information whose form is A = (U, A ∪ {d}), where d ∈A is the decision attribute. A
decision attribute called “Class” has been added as shown in Table 2, where M and E denote
Manager and Employee, respectively. The table was modified from the original [12].
Object Index Salary Age Class

u1 80 30 M
u2 30 23 M
u3 80 40 E
u4 50 45 M
u5 80 55 E
u6 50 45 E
u7 30 60 M
u8 100 35 E
Table 2. A decision system where each row is classified into a class. Salary and Age are the
condition attributes, while class is the decision attribute.
2.3 Indiscernibility
Objects in information and decision systems may be indistinguishable from one another
based on a set of attributes B that belongs to A (B ⊆ A). A set of objects is indiscernible or
equivalent when their attributes are related by an equivalence relation. An equivalence
relation is a relation on a set B when it is:
1. Reflexive (if a R a, then R is reflexive).

2. Symmetric (if a R b then b R a, then R is symmetric).
3. Transitive (if a R b and b R c, then a R c, thus R is transitive).
For an information system A = (U, Α), there is an equivalence relation for any of the sets B ⊆
A. The equivalence relation can be formalized as
IND A ( B) = {( x, x' ) ∈ U 2 | ∀a ∈ B, a ( x) = a ( x' )} .

Referring to table 4, the IND relations for Salary can be written as shown.
IND( Salary ) = {{u1 , u3 , u5 , u8 },{u 2 , u 4 , u 6 , u 7 }}
It is impossible to write the IND relations for Salary until discretization is completed.
2.4 Discretization
Discretization is not directly related to rough set theory. It is simply a preprocessing
technique. Discretization is associated with information loss. In general, when it is too
coarse (i.e. longer interval), there is too much information loss or noise in the data.
However, it is better for the classification capability of unseen objects. When the
discretization is more fine (i.e. shorter interval), less noise exists in data, but classification
capability of unseen objects may be impaired.
In our decision system table, both Salary and Age need to be discretized. The set of
possible Salary and Age values, respectively referred to as s and a from here on, is given
by
Vs = [15, 120)
Va = [18, 65)
The lower and upper bounds of the attribute’s interval are extended to cover possible
values. For example, the Age attribute is extended to include likely working ages from
age 18 through 65.
The set of values of s and a in U is
s(U) = {30,50,80,100}
a(U) = {23, 30, 35, 40, 45, 55, 60}
The intervals obtained for s are
[30, 50); [50, 80); [80, 100)
The intervals obtained for a are
[23, 30); [30, 35); [35, 40);

[40, 45); [45, 55); [55, 60).
Boundary intervals such as [15, 30) and [100, 120) should not be used since one can not
discern anything for this data set.
The intervals introduce a set of cuts, which are defined as (s, c) where c ∈Vs and (a, c) where
c ∈ Va. If the cut is taken based on the mid-point of each interval, the set of cuts P obtained
for s and a are respectively
(s, 40); (s, 65); (s, 90);

(a, 26.5); (a, 32.5); (a, 37.5);
(a, 42.5); (a, 50); (a, 57.5)
The next step is to find the set of minimal cuts that can discern all of the objects that are
needed. It turns out that the problem of finding the irreducible set of cuts P in the decision
system is NP-complete while the effort to find the optimal set of cuts P in a decision system
is NP-hard [5].
However, there are heuristics that can be used to find the optimal set of cuts P in practical
time. One of them is the Maximal Discernability heuristic [1], [5], which is demonstrated
here. The algorithm to construct table A* from A is listed in the following steps:
1. Each column in table A* is a Boolean variable of the corresponding column in A. If each

pair of objects can be discerned by the Boolean variable, then assign value 1, else assign
0.
2. Choose a column from A* that has a maximal number of 1’s and delete all the rows
which contain a 1 in the selected column.
3. Repeat step 2 and continue until all columns and rows are consumed.
A* p1s p2s p3s p1a p2a p3a p4a p5a p6a d

u1, u3 0 0 0 0 1 1 0 0 0 1
u1, u5 0 0 0 0 1 1 1 1 0 1
u1, u6 0 1 0 0 1 1 1 0 0 1
u1, u8 0 0 1 0 1 0 0 0 0 1
u2, u3 1 1 0 1 1 1 0 0 0 1
u2, u5 1 1 0 1 1 1 1 1 0 1
u2, u6 1 0 0 1 1 1 1 0 0 1
u2, u8 1 1 1 1 1 0 0 0 0 1
u3, u4 0 1 0 0 0 0 1 0 0 1
u3, u7 1 1 0 0 0 0 1 1 1 1
u4, u5 0 1 0 0 0 0 0 1 0 1
u4, u6 0 0 0 0 0 0 0 0 0 1
u4, u8 0 1 1 0 0 1 1 0 0 1
u5, u7 1 1 0 0 0 0 0 0 1 1
u6, u7 1 0 0 0 0 0 0 1 1 1
u7, u8 1 1 1 0 0 1 1 1 1 1
new 0 0 0 0 0 0 0 0 0 0
Table 3. Decision System A*.
The following example clarifies the process of constructing table A* from A mentioned in
step 1 of the algorithm. Each cut previously obtained is assigned a Boolean variable, which
in turn is used as a condition attribute in table A*.
s
For example, (s, 40) is assigned Boolean variable p1 . Each object pair in A*, is derived from
table A by looking at the decision attribute. Objects that do not have the same decision
attribute should be paired up. For example, u1 and u3 are paired up since they have different
decision values (i.e. M and E).
The resulting table A* created from A is shown in Table 3.
The optimal cut chosen is (s, 65), (a, 32.5) and (a, 50). These cuts are then used to discretize
the decision table 2. The rank can be assigned using the following rules:
1 If s < 65, value 0 is assigned to s, else assign value of 1.
2 If a < 32.5, value 0 is assigned to a.
3 If (32.5 ≤ a < 50), value 1 is assigned to a.
4 If a ≥ 50, value 2 is assigned to a.
A discretized table can be produced by applying the condition to each analog value in the
table.
Index Salary Age Class
u1 1 0 M
u2 0 0 M
u3 1 1 E
u4 0 1 M
u5 1 2 E
u6 0 1 E
u7 0 2 M
u8 1 1 E
Table 4. Discretized Decision System.
2.5 Lower and Upper Approximations in Rough Set and Accuracy

Let U be the non-empty finite set and R be an equivalence relation. The pair A = (U, R) is an
approximation space. The equivalence relation R on U leads to a partition of the objects in
the universe U. The idea here is to partition the objects that have the same outcome, or in
other words, to partition objects that have the same decision attribute. However, this may
not always be as easy as stated. There will be objects with the same condition attributes (in
the same equivalence class), but different decision attributes. Therefore one can not define
every set precisely.
In cases where the set can not be defined precisely, it can be approximated. This is where
rough set emerges. Let us assume that there is an information system A = (U, Α), a set of
attributes B ⊆ A, and a set of objects X ⊆ U. Using the set of attributes B, one can
approximate the objects X into:
1. Lower Approximation: the set of objects that can be classified as member X with
certainty. Formally stated as
B ( X ) =  {E i ∈ U B : E i ⊆ X }
(2.1)
2. Upper Approximation: the set of objects that can possibly be classified as a member X.
Formally stated as
B( X ) =  {Ei ∈ U B : Ei ∩ X ≠ {}}
(2.2)
Between the lower and upper approximation, one can define the set of objects that cannot be
classified into X decisively. This set is also known as the B-boundary region of X.
There is a coefficient that reflects the accuracy of approximation,
B( X )
αB (X ) = (2.3)
B( X )
where |X| denotes the cardinality of X ≠ ∅. When α B = 1, the X is crisp with respect to B,
otherwise, X is rough with respect to B.
For our example, the boundary region would be for object u4 and u6 since they can not be
discerned. The lower and upper approximations can be written as
B ( M ) = {u1 , u 2 , u 4 , u 7 }
B( M ) = {u1 , u 2 , u 4 , u 6 , u 7 }
B( M ) 4
α B (M ) = = = 0.8
B( M ) 5
In general, the value of α reflects the accuracy of decision rules obtained.
2.6 Reducts
One way to increase computation efficiency is to reduce the size of data by reducing
attributes that need to be taken into account. Only attributes that do not contribute to the
classification result can be omitted such that the indiscernibility relation remains intact. The
set of remaining attributes is the minimal set and is called a reduct.
Although finding the equivalence class is a relatively straightforward computation process,

finding reducts with minimal attributes is known to be NP-hard. Fortunately, there are
heuristics that allow minimal reducts to be computed in reasonable time.
2.7 Discernibility Matrix

Computing the reducts of an information system A = (U, A) can be started by creating the
indiscernibility matrix. This matrix is a symmetric n x n matrix where each entry cij is
defined as
cij = {a ∈ A | a ( xi ) ≠ a ( x j )} for i, j = 1,..., n
(2.4)
Each cell in the matrix holds the set of attributes where objects xi and xj are discernable.
The cell would have empty set when:
1 xi = x j , that is case for diagonal cells.
2 For decision systems, when the decision attribute of objects xi and xj are equal, or
formally, d ( xi ) = d ( x j ) .
2.8 Discernibility Functions

Based on the discernibility matrix, a discernibility function can be immediately obtained. It
is constructed using Boolean expressions from the discernibility matrix, defined as
f A (a1* ,..., a m* ) = ∧ {∨ cij* | 1 ≤ j ≤ i ≤ n, cij ≠ {}}

(2.5)
where a1* ,..., am* are related to attributes a1 ,..., am . The attributes may be transformed during
* *
discretization process. cij = {a | a ∈ cij } is the set of Boolean variables.
Once the discernibility function f A is formed, it can be further developed using Boolean
algebra simplification.
2.9 Decision Rules

When applying rough set for supervised learning, we need to construct a set of rules from
the training data, such that new or unseen objects can be separated into known classes.
A basic method for forming the decision rules is begun by finding the reducts of the
decision table. Then for each reducts R = {r1 ,..., rn } , we generate the decision rule by taking
the conjunction (r1 = r1 (u )) ∧ .... ∧ (r2 = r2 (u )) as the predecessor. Next we take the decision
attribute d with value d (u ) as the successor and format the predecessor and successor
values as
(r1 = r1 (u )) ∧ .... ∧ (r2 = r2 (u )) ⇒ d = d (u ) (2.6)
Rule induction is about deciding which attributes should be included in the predecessor of
the rule. Rules obtained can always be minimized, but it will introduce noise and may
poorly classify the unseen objects.
Once the rules are obtained, they can be used to classify the objects that were unseen before.
The basic steps involved can be outlined as follows [1].
1. Apply the existing rules to the new objects so that it can determine which rules
actually are a fit to the new objects.
2. If none of the rules are matched, then fallback a must be chosen, or the objects
would be classified as undefined.
3. If more than one rule is applicable, then a negotiation among the rules must be
performed to decide which one to be used.
For the discernibility function extracted from the decision table 2, we obtain the following
sets of decision rules by:
1. If (a < 32.5), then d = M

2. If (s > 65) and (32.5 ≤ a < 50), then d = E
3. If (s < 65) and (32.5 ≤ a < 50), then d = M
4. If (s > 65) and (a ≥ 50), then d = E
5. If (s < 65) and (32.5 ≤ a < 50), then d = E
6. If (s < 65) and (a ≥ 50), then d = M
3. Particle Swarm Optimization

PSO was originally introduced by Kennedy and Eberhart [21]. The algorithm was inspired
by a sociological observation of a flock of birds behavior while searching for food. Each
member of the flock moves with a direction and speed influence by its own previous state
and that of the as a whole flock.
PSO consists of a swarm (collection) of particles searching through the solution space. Each
particle holds information that can potentially become the solution. Each particle has a
position and velocity that are mutually affecting those of other particles. Each particle will
adjust its parameter according to the swarm’s best outcome, while still considering its own
experience. Therefore, at any instance, the following information is maintained by each
particle.
• xi, the current position of the particle;

• vi, the current velocity of the particle; and
• yi, the personal best position of the particle (pbest); the best position visited so far by the
particle.
• ŷ, the global best position of the swarm (gbest); the best position visited so far by the
entire swarm.
The search performed by the swarm is either to maximize or minimize the objective function
f(x). The personal best position (pbest) is obtained by evaluating the following.
 y (t ) if f ( xi (t + 1)) ≥ f ( yi (t ))
yi (t + 1) =  i (3.1)
 xi (t + 1) if f ( xi (t + 1)) < f ( yi (t ))
The global best position (gbest) is obtained by using
yˆ (t ) ∈ { y0 , y1 , ..., ys } = min{ f ( y0 (t )), f ( y1 (t )), ..., f ( ys (t ))} (3.2)
After each iteration, the current position (xi) and velocity (vi) are recalculated using
vi (t + 1) = ω vi (t ) + c1 r1 (t )( yi (t ) − xi ) + c2 r2 (t )( yˆ (t ) − xi (t )) (3.3)
xi (t + 1) = xi (t ) + vi (t + 1) (3.4)
where ω is the inertia weight which reflects the memory of previous velocities. yi(t) – xi
(cognitive component) represents the particle’s own experience as to where the best solution
is. ŷ(t)–xi (social component) represents the direction of the entire swarm towards the best
solution. The c1 and c2 are acceleration constants. r1(t) , r2(t) are in the distribution of U(0,1)
which will be a random number between 0 and 1.
In image classification, the PSO algorithm is used to optimize the objective functions that are
mainly to:
• Minimize the distance between pixels and cluster means for each cluster.
• Maximize the distance between clusters.
In unsupervised training, there is no prior knowledge of the number of clusters. Therefore

the cluster validity is determined by the objective functions. In the algorithm, the Davies-
Bouldin index is used as the means to evaluate the result of each iteration.
3.1 Cluster Validity – Davies-Bouldin Index.

The accuracy or validity of the classification results need to be measured using certain
criteria. As a prerequisite, a set of objects needs to possess a natural group structure. In our
image classification algorithm outlined in section 5, the Davies-Bouldin (DB) index is used
as the aid in parameter tuning. Our objective function is to minimize the DB index, since a
smaller index value indicates compact and well-separated clusters. The similarity index
between two clusters Ci and Cj can be expressed as [17]
si + s j (3.5)
Rij =
d ij
where si and sj are a measure of distance within a cluster, and dij is the distance between
cluster i and j. The si is defined as [17]
1
∑
r
si = ( x − m i )1/r (3.6)
ni x∈Ci
where ni is the number of pixels in the cluster Ci .The distance between two clusters dij is
defined as [17]
l q
d ij = ( ∑ mik − m jk )1/q (3.7)

k =l
where l is the number of clusters and m represents the mean distance.

Let Ri be defined as [17]
Ri = j =1,..., m , j ≠ i max Rij , i = 1,..., m (3.8)
Then the DB index is defined as [17]

1 m
DBm =
m
∑R i
(3.9)
i =1
4. The K-means algorithm for multispectral image classification

The K-means algorithm is one of the simplest and most efficient unsupervised learning
algorithms to solve clustering problems in image segmentation. In this algorithm, random
cluster means are assigned and repeatedly modified throughout the process in order to
minimize the squared error function. Suppose there are N pixels in an image to be
classified into m clusters. Each pixel vi , where 1 ≤ i ≤ N , is assigned to one of the
clusters cj , where 1 ≤ j ≤ m , based on squared Euclidean distance of each pixel to each
cluster mean.
N m
d ( v , c ) = ∑∑ ( vi − c j )2 (3.10)
i =1 j =1
Upon the completion of the assignment, each new cluster mean is calculated using
∑v i
cj = v∈c (3.11)
n
where 1 ≤ i ≤ N , and n is the number of pixels in cluster c j . The process ends when c j
stabilizes.
The weakness of K-means is that it is dependent on the initial selection of the cluster means
and it may be trapped into locally optimal results. However, running the algorithm
repeatedly and randomly selecting different sets of cluster means may offset the problem. In
a paper by Hung and Germany [19] it is shown that the local optimal results may also be
avoided by assigning the cluster means based on distribution of patterns in histogram of an
image.
5. Multispectral Image Classification using Rough Set Theory

Multi-spectral images can be analyzed using rough set theory. However, since all the
attribute values are analog, the discretization process is required. Multispectral images
contain multiple bands, for example the RGB color band.
Object Index R G B Class

u1 149 148 143 1
u2 154 155 150 1
u3 159 160 155 1
u4 174 171 164 2
u5 164 161 154 2
u6 179 183 186 3
u7 159 165 163 3
u8 178 184 182 3
Table 5. Decision System Table for multi-spectral image.
The values of the condition attributes are obtained from the image data shown in figure 1,
while the values of the decision attributes are obtained from 'ground truth' data..
Each object has three condition attributes, Red (R), Green (G) and Blue (B) which are
associated with a decision attribute. The decision attributes signify the following:
1 Class 1 represents land
2 Class 2 represents village
3 Class 3 represents water.
The value of each attribute ranges from 0 to 255, hence the training data from Table 5 can be
expressed as:
VR={0, 149, 154, 159, 164, 174, 178, 179, 255}

VG={0, 148, 155, 160, 161, 165, 171, 183, 184, 255}
VB={0, 143, 150, 154, 155, 163, 164, 182, 186, 255}
Based on the above intervals, the following set of cuts are obtained.
For the R attribute:

(r, 151.5); (r, 156.5); (r, 161.5); (r, 169); (r,176); (r,178.5)
For the G attribute:
(g, 151.5); (g, 157.5); (g, 160.5); (g, 163); (g, 168); (g, 177); (g, 183.5)
For the B attribute:
(b, 146.5); (b, 152); (b, 154.4); (b, 159); (b,163.5); (b, 173); (b, 184)
The optimal set of cuts needs to be selected now. There are many ways to perform the
selection. For decision table A = (U, A ∪ {d}), a local method can be used as [1]:
Input: The consistent decision table A.

Output: The semi-minimal set of cuts D consistent with A.
Method: Initialize the binary tree variable T with the empty tree. Label the root by the set of
all objects U and fix the status of the root to be unready.
while there is a leaf marked by unready do

begin for any unready leave N of the tree T
begin
if objects labeling N have the same decision value then
begin
Replace the object set at N by its common decision
Change the status of N to ready.
end
else
begin
Compute the value WN (a,c) for all cuts from CA and search for cut
(a*, c*) maximizing the function WN(.) i.e.
(a*, c*) = arg ma x W N (a, c)
( a ,c )
Replace the label of N by (a*, c*) and mark it as ready;

Create two new nodes N1 and N2 with status unready as the left and right
subtrees of N, where:
N1 = {u ∈ N : a * (u ) < c*} and
N 2 = {u ∈ N : a * (u ) ≥ c*}
end
end
end
return T
By applying the algorithm above to the image data as shown in Table 6, the following
details are derived. For each cut of the R, G and B attributes, we find the cut that yields the
maximum number of pairs. The search gives us (g, 160.5) as the optimal solution which
yields 15 pairs.
The cut (g, 160.5) divides the set into two, X1 = {u1, u2, u3} and X2 = {u4, u5, u6, u7, u8}. Notice
that X1 actually consists of objects of the same class, so the search ends. The search continues
for X2. Three sets of cuts are found from the R, G and B attributes for X2. All of the cuts, (r,
176), (g, 177) and (b, 173) yield the same number of objects (4 pairs). We only need to select
one, and the one chosen is (r, 176). Again, this cut divides the set into two, Y1 = {u4, u5, u7}
and Y2 = {u6, u8}. Y2 consists of objects of the same class, so the search ends. The search
continues for Y1. The cut that can discern the most from Y1 is (r, 161.5).
The cut (r, 161.5) divides Y1 into two sets, Z1 = {u4, u5} and Z2 = {u7}. The search ends since
both sets contain objects of the same class.
The set of cuts selected are:
(r, 161.5); (r, 176.0)

(g, 160.5)
It appears that our data set only requires two attributes to be fully discerned. Note that
different discretization methods will obtain different results. For example, if a naïve
algorithm was used, the B attribute will be considered in generating the cuts.
Using the cuts, a discretized table is subsequently generated. The asterisk in the B column
indicates that it is not needed to discern the classes. This, however, will not be the case when
the training set grows larger.
Object Index R G B Class

u1 0 0 * 1
u2 0 0 * 1
u3 0 0 * 1
u4 1 1 * 2
u5 1 1 * 2
u6 2 1 * 3
u7 0 1 * 3
u8 2 1 * 3
Table 6. Discretized Decision System Table for multispectral image.
Based on Table 6, the following rules are generated:

1 If (g < 160.5), then d = 1.
2 If (161.5 ≤ r < 176), then d = 2.
3 If (r ≥ 176), then d = 3.
4 If (r < 161.5) and (g ≥ 160.5), then d = 3.
6. The Hybrid Rough K-means Algorithm and Particle Swarm Optimization for
Multispectral Image Classification
The K-means clustering method is categorized as a hard clustering method. Using K-means
to classify images that have obscured or blurred boundaries will not bring a satisfactory
result. There are many methods proposed to deal with this. The fuzzy C-means [22] and
genetic K-means [23] algorithms are two examples.
Rough K-means is a recently proposed method that deals with the coarseness of the
information. In gray image classification the challenge is on segmenting the blurred
boundaries between clusters. Using rough sets theory, an image can be represented as sets
of lower and upper approximation. The rough K-means model for our proposed image
segmentation algorithm is adapted from [20].
 ∑ v∈ A ( x ) v j + ∑ v∈( A( x )− A( x )) v j ,
 wlower * wupper * if A( x ) − A( x ) ≠ ∅
 A( x ) A( x ) − A( x )
cj =  (6.1)
 ∑ v∈ A ( x ) v j ,
 w lower * otherwise
 A( x )
Each image pixel can be classified into lower or upper approximations. Following basic
rough set properties:
• A pixel can be part of only one lower approximation
• If a pixel is part of a lower approximation, then it is also part of the upper

approximation
• If a pixel does not belong to any lower approximation, then it belongs to two or more
upper approximations.
Applying rough set into K-means requires the formula to include lower and upper
approximations. The formula, as shown below, includes the weighing factor wlower and
wupper . Let v be a pixel vector and d (v, ci ) be the distance between the pixel and the mean of
cluster i. Let
d ( v , ci ) = min1≤ j ≤ k d ( v , c j ) (6.2)
and
T = { j : d ( v , ci ) − d ( v , c j ) ≤ threshold and i ≠ j } (6.3)
In order to correctly classify a pixel, the following classification criteria are used:
1. If T is not an empty set, then the pixel is classified as an upper approximation of both
clusters i and j.
2. If T is an empty set, the pixel is classified as a lower approximation for cluster i. It will
also be classified as an upper approximation for cluster i.
To summarize, the following are steps to perform the rough K-means algorithm [26]:
1. Initialize K clusters randomly.
2. Select wlower and a threshold value.
3. For each cluster, find d using Equation 6.2 and T using Equation 6.3.
4. Classify the pixel using the classification criteria.
5. Calculate the new cluster center (mean) using Equation 6.1.
6. If every cluster converges, then stop. Otherwise, repeat step 3.
The parameters involved are wlower, wupper and the threshold. The sum of wlower and wupper
will always be one. These parameters are set manually by trial and error. Since it is not
trivial to come up with good parameter values, this is the major disadvantage for this
method. In order to adjust these parameters automatically, this algorithm needs to be
improved using automatic tuning mechanism. The PSO algorithm alleviates the limitation
by automatically searching and modifying the parameters during the image segmentation
process.
The proposed algorithm that combines rough K-means and PSO algorithm is outlined as
follows [26]:
1. Initialize the mean of each cluster.

2. Initialize a number of particles where each of the particles is randomly assigned with
wlower and the threshold.
3. Find the minimum pair of distance of x to all clusters, d(x-ci). Then assign the pixel
according to the following criteria.
• If the difference of the distance d(x-ci) – d(x-cj) is less than the threshold, then
the pixel belongs to upper approximation of both clusters ci and cj.
• Otherwise, the pixel x belongs to lower approximation of cluster ci.
4. Calculate the DB index of each particle. Save the DB index of each particle and compare
them with those of other particles. Find the global best index and tune the lower
approximation and thresholds of each particle according to the following guidelines.
• If the personal best DB index equals the global best DB index, then lower the
threshold so that it includes only the pixels that are definitely in the lower
approximation.
• If the personal best DB index is greater than the global best DB index, then
adjust the wlower and the threshold toward the particle with the global best DB
index.
5. Calculate the new mean for each cluster.
6. Repeat steps 3, 4 and 5 until all particles converge.
7. Experimental Results
To test the effectiveness of the proposed algorithms, multispectral and artificial images were
used in our experiments. The original image is processed to obtain the multispectral
information. Then the Rough Set Exploration System (RSES) software was used to process
the image data [24]. A selected percentage of the image pixels were sampled for training
purpose. Finally MATLAB was used to make the results viewable as an image.
Experimental results are described in section 7.1. The experiment on the rough K-means
algorithm is intended to show the effect of parameter selection on the results of the
classification. Experimental results on the algorithm are shown in section 7.2.
7.1 Experimental Results on the Rough Set Theory

Due to the size of the table in our training sample (in the range of over 80,000 pixels), we
need to resort to the decomposition tree feature of the RSES. This feature allows us to break
the table into sections no larger than a predefined size. In this case, a size of 500 samples is
selected as the maximum size of each leaf in the decomposition tree. These methods are
further elaborated in [3] and [4]. During the decomposition process, the table is also
discretized. A local method like the one outlined in Section 2 is chosen as the method for
selecting the optimal cuts. Each leaf of the decomposition tree contains a set of rules that
was dynamically created. The rules are then used to classify the unseen objects. Using RSES,
there are two formats of output that the user can select: confusion matrix or classification
results in table format.
After applying the rules to the pixels and obtaining the classification result, the reverse
process is done using MATLAB to get the classified image. All pixels, including the
unclassified ones, are assigned a specific color for visualization. The original image as
shown in Figure 1(a) is a terrain image that has land, water and village. After the
classification using rough set theory, the classified result is obtained in Figure 1(b). The
confusion matrix with an average accuracy of 0.79 is shown in Table 7.
(a) (b)
Fig. 1. (a) An original satellite image, and (b) the classified result using the rough set theory.
The color blue represent land, the red represents water and green represents village.
Undefined objects are left as black pixels.
Classified
Actual 1 2 3
1 181,838 2,264 1,052
2 960 1,631 90
3 2,083 124 67,381
True Pos Rates 0.98 0.41 0.98
Table 7. Rough set classification accuracy assessment.
With the parallelepiped classification algorithm [25], the ordering of the classes affects the
final result. The experimental results are shown in Figure 2. First we show the result of
classification, where the order of classes are 1, 2, and 3 (respectively land, village and water).
It is apparent that the RGB spectral signatures for village and water overlap. Since the order
of analysis begins with village (2), most pixels of water (3) were classified as village. The
confusion matrix for the results in Figure 2 is shown in Table 8 with an average accuracy of 0.6.
Fig. 2. A classification result of Figure 1(a) using the parallelepiped method.

Ordering of the classification for classes 1, 2, 3 (land, village and water)
Classified
Actual 1 2 3
1 188,043 38 0
2 2,113 1,305 0
3 361 40,661 29,626
Table 8. Parallelepiped classification accuracy assessment.
The classification ordering is class 1, 2, and 3 (land, village, and water)
The experiment is repeated for the parallelepiped classifier. The ordering is now started
with classes 1, 3 and 2 (respectively land, water and village). Contrary to the result in Figure
2, now most pixels of the village area are classified as water. The confusion matrix for the
result in Figure 3 is shown in Table 9 with an average accuracy of 0.72. The increase in the
average accuracy, because of the misclassification of village, is mitigated by the number of
its pixels overall.
Fig. 3. A classification result of Figure 1(a) using the parallelepiped method. ordering 1,3 ,2
Classified
Actual 1 3 2
1 188,043 38 0
3 361 70,237 0
2 2,113 689 616
Table 9. Parallelepiped classification accuracy assessment.
The classification ordering is class 1, 3, and 2 (land, water, and village)
The following experiment requires ground truth data for accuracy assessment. The remote
image sensing truth data was obtained from Dr. Su in the National Central University in
Taiwan, while the ground truth data for the artificial images were created using custom
software written in Java. The decision rules, which are required for classification of the
image, are facilitated by RSES [24]. The process for remotely sensed images begins by
sampling 30% of the image pixels as training data to create decision rules. The process to
create decision rules follows the outline in Section 2. After obtaining the rules with RSES,
they are used to classify the image. The image consists of approximately 262,000 pixels.
Referring to Figure 4(a), class 1 is land, class 2 village and class 3 water. The average
accuracy for the classification is 79%. The confusion matrix is shown in Table 10. The most
difficult pixels to classify are the village pixels, as indicated by the small value of its true
positive rate. The ground truth data and classified result are shown in Figure (b) and (c).
(a) (b) (c)

Fig. 4. (a) An original remote sensing image, (b) Ground Truth Data, and (c) Classification
result.
Classified
Actual 1 2 3
1 181,838 2,264 1,052
2 960 1,631 90
3 2,083 124 67,381
True Positive Rates 0.98 0.41 0.98
Table 10. Rough set classification accuracy assessment for remote sensing image.
The other experiment is performed on the artificial image that consists of several shapes,
namely, a cube, a serpentine, two airbrush shapes and a round shape (Figure 5). Similarly,
30 % of the pixels in the image are used for training. After obtaining the decision rules, the
image is classified. The artificial image has a total of 10000 pixels. Referring to Table 11, class
1 is the cube, class 2 is the connector of the airbrush images, class 3 is the airbrush images,
class 4 is the round shape and class 5 is the background. Some difficulties occur while trying
to obtain the ground truth, due to the inherent limitations of the image processing software.
The results however, indicate that class 3, the airbrush shapes, has the most incorrectly
identified pixels. The total accuracy is still about 99% as shown in Table 11.
(a) (b) (c)

Fig. 5. (a) Original artificial shapes (b) Image truth (c) Classification result
Classified
Actual 1 2 3 4 5
1 833 0 0 0 1
2 5 884 3 7 3
3 0 13 907 3 2
4 0 3 0 1,434 8
5 0 0 2 3 5,878
True Positive Rate 0.99 0.98 0.99 0.99 1
Table 11. Rough set classification accuracy assessment for artificial shapes.
7.2 Experimental Results of the Hybrid Rough K-Means and PSO

In the experiment shown in Figure 6.1 (b), the parameter wlower is set to 0.55 and the
threshold 0.45. The first parameter weights how much the previous calculated mean will
affect the new mean and the second parameter adjusts the boundary region. In other words,
the second parameter is the criteria limiting whether a pixel should be included in the upper
approximation of a class. The higher the value of the threshold, the more less the criteria is
constrained.
The experiment was done for several different combinations of wlower and the threshold
value. After careful inspection on the results shown in Figure 6(b) through Figure 6(f), it
turns out that a monotonic increase or decrease of wlower and the threshold does not
guarantee improvement in the classification results. From Figure 6(b) to (c), the accuracy
decreased. Although the threshold was reduced, the boundary area between land, village
and river actually turns blurred. In the result of Figure 6(d) the accuracy improves. Also,
from Figure 6(d) to (e) the accuracy decreases again, although not as badly as between
Figure 6(b) to (c). The accuracy improves again in the results of Figure 6(f). These are strong
indications that varying the parameters (wlower and the threshold) do not guarantee that the
best results can be predicted easily. As a matter of fact, the most optimal parameters can
only be found empirically. This is exactly the shortcoming of the rough K-means algorithm
and the problem is addressed using PSO to tune the parameters.
(a) (b) (c)
(d) (e) (f)

Fig. 6. Rough K-Means for remote sensing image. (a) Ground Truth, (b) wlower= 0.55 and
Threshold = 0.45 Accuracy = 44.95 %, (c) wlower = 0.65 and Threshold = 0.35 Accuracy =
38.54 %, (d) wlower = 0.75 and Threshold = 0.25 Accuracy = 49.18 %, (e) wlower = 0.85 and
Threshold = 0.15 Accuracy = 46.62 %, and (f) wlower = 0.95 and Threshold = 0.05 Accuracy =
51.42 %
Results with similar consistency are obtained for the image in Figure 7. From Figure 7(b) to
7(c), we can see improvement visually. As the parameters change in one direction, the
accuracy drops as shown in Figure 7(d). Finally the best value for the experiment is shown
in figure 7(f). Looking at those classified results, it may lead us into thinking that increasing
the wlower and decreasing the threshold gives a better result. That is not necessarily the case,
since doing so means that we are counting on the lower bound more and reducing the
threshold value, while at the same time discounting the upper bound. At the extreme, where
wlower is almost 1 and threshold is almost zero, roughness is actually removed, and the set
becomes crisp. This is also formulated in Equations 6.1, 6.2 and 6.3 earlier.
(a) (b) (c)
(d) (e) (f)

Fig. 7. Rough K-Means for artificial image. (a) Ground Truth, (b) wlower = 0.55 and Threshold
= 0.45 Accuracy = 68.17 % (c) wlower = 0.65 and Threshold = 0.35 Accuracy = 82.83 %, (d)
wlower = 0.75 and Threshold = 0.25 Accuracy = 78.92 % (e) wlower = 0.85 and Threshold = 0.15,
Accuracy = 85.11 %, and (f) wlower = 0.95 and Threshold = 0.05, Accuracy = 89.54 %.
Experiments using the K-means and rough K-means PSO algorithms are performed. For the
comparison, the number of iteration is limited to 50 and the tolerance is set to 0.001. The
result shown in Figure 8 is selected from the best outcome of 20 runs of the K-means and
rough K-means algorithms. For the rough K-means PSO, 10 particles are used to explore the
search space. Comparing the results of the K-means, rough K-means and rough K-means
PSO algorithms, it reveals that although the improvement can be made, it is in the order of
more or less 5 %. It is not very significant, but we should note that the rough K-means PSO
achieve the optimal results independent of initial mean selections.
(a) (b) (c)

Fig. 8. Comparison of the results. (a) K-Means Accuracy = 56.26 %, (b) Rough K-Means wlower
= 0.95 and Threshold = 0.05, Accuracy = 51.42 %, and (c) Rough K-Means PSO, Accuracy =
59.96 %.
While running, the algorithm is tuned by keeping track of the DB index and adjusting the
PSO particle accordingly to calculate the new mean. Figure 9 shows the DB index tracking
for Figure 8(a) and 8(c).
Rem ote Sensing Im age
0.5
0.4
DB Index
0.3 RKM PSO

0.2 K-Means
0.1
0
1
15
22
29
36
43
50
Iterations
Fig. 9. K-means and rough K-means PSO DB index tracking for Figure 8(a) and 8(c).
Referring to Figure 9, it is apparent that the K-means algorithm eventually converges and
locks into a certain mean value. The rough K-means PSO shows a better capability to search
for solutions, because there are about 10 particles to keep track of global best and adjust the
velocity towards the best solution in every iteration. Similar results are obtained from the
remaining tests performed. The resulting improvement, however, is not as obvious as those
shown in the artificial image (Figure 10). Part of the reason is because the artificial image
does not have enough roughness. Hence, it is not difficult for K-means to perform well in
this case. Figure 11 shows the tracking of the DB index.
(a) (b) (c)

Fig. 10. Comparison of the results. (a) K-means Accuracy 90.12 %, (b) Rough K-means wlower
= 0.95 and Threshold = 0.05, Accuracy = 89.55 %, and (c) Rough K-means PSO, Accuracy =
90.65 %
Artificial Im age
0.5
0.4
DB Index
0.3 RKM PSO

0.2 K-Means
0.1
0
1
15
22
29
36
43
50
Iterations
Fig. 11. The K-means and rough K-means PSO DB index tracking for the artificial image
shown in Figure 10.
For the planet image, there is no ground truth data available. However, the visual inspection
reveals improvement. Based on the results shown in Figure 12, we can see that the K-means
algorithm actually has some difficulty in the segmentation of the blurred or rough
boundaries. The rough K-means PSO however, appears to be able to discern the rough
boundaries, and therefore comes up with a much more rounded shape for the planet. The
outer shape of the planet appears sharper, more rounded and less distorted. Figure 13
shows the DB index tracking of the planet image.
(a) (b) (c)

Fig. 12. Experimental results for planet image, (a) K-means, (b) rough K-means wlower = 0.95
Threshold = 0.05, and (c) rough K-means PSO.
Planet
0.3
0.25
DB Index
0.2
RKM PSO
0.15
K-Means
0.1
0.05
0
1
7
13
19
25
31
37
43
49
Iterations
Fig. 13. The K-means and rough K-means PSO DB index tracking for the planet image
shown in Figure 12.
8. Conclusions and Future Work

Image classification and segmentation by applying rough set theory may be approached
from two different perspectives: unsupervised or supervised methods. From the
experiment, it is generally shown that a supervised classification achieves better results as
compared with the unsupervised methods. However, it should be noted that unsupervised
classification may be preferred because it requires less prior knowledge. The K-means
algorithm can be enhanced by using the rough set theory for image classification, however it
has a practical limitation by itself, since the parameters (wlower and the threshold value) are
difficult to tune manually. To solve this problem, the PSO is used for tuning these
parameters. The algorithm is tested on several and in general its improved noise immunity
can be seen in the results especially when the images have rough boundaries or noisy
details. For future work, the ant colony optimization and differential evolution algorithms
will be explored for tuning the parameters in the rough K-Means algorithm.
9. References
[1] Bazan, J., H.S. Nguyen, S.H. Nguyen, P. Synak, and J. Wroblweski, “Rough Set Algorithms
in Classification Proble,” In L. Polkowski, S. Tsumoto, and T. Lin (Editors), Rough Set
Methods and Applications. New York, Physica-Verlag Heidelberg New York (2000):
pp 49-88.
[2] Bazan, J. and M. Szczuka, “The Rough Set Exploration System,” In J. Peter and A. Skowron
(Editors), Transactions on Rough Sets III. Springer (2005): pp 37-55.
[3] Bazan, J.and M. Szczuka, “RSES and RSESlib - A collection of tools for Rough Set
Computations,” Lecture Notes in Artificial Intelligence 3066, Heidelberg, Springer
(2000): pp. 592-601.
[4] Kim, D., “Data classification based on tolerant rough set.” Pattern Recognition, 34 (2001):
pp. 1613-1624.
[5] Komorowski, J., Z. Pawlak, L. Polowski, and A. Skowron, “Rough Sets: A Tutorial,” In S. K.
Pal and A. Skowron (Editors), Rough Fuzzy Hybridization. Springer (1999): pp 3-98.
[6] Lillesand, T. M. and R. W. Kiffer, Remote Sensing and Image Interpretation (3rd ed), John
Wiley & Sons, Inc., 1994.
[7] Lingras, P., “Unsupervised Rough Set Classification Using GAs,” Journal of Intelligent
Information Systems, 16. (2001): pp 215-228.
[8] Nguyen, H. S., Data Regularity analysis and applications in data mining, Ph.D thesis,
Supervisor B. Chlebus, Warsaw University, 1999.
[9] Pawlak, Z., “Rough Set,” International Journal of Information and Computer Science, 11.
(1982): pp 341-356.
[10]Pawlak, Z., J. Grzymala-Busse, R. Slowinski and W. Ziarko, “Rough Sets,” Communication
of ACM, 38 (1995): pp 89-95.
[11] Schowengerdt, R. A., Remote Sensing: Models and Methods for Image Processing (2nd ed),
Academic Press, 1997.
[12] Qin, Z., G. Wang, Y. Wu, and X. R. Xue, „A Scalable Rough Set Knowledge Reduction
Algorithm,” Heidelberg , Springer - Verlag, (2004): pp 445-454.
[13] Petrosino, A. and G. Salvi, “Rough Fuzzy set based scale space transforms and their use in
image analysis.” International Journal of Approximate Reasoning, 41 (2006): pp 212-
228.
[14] Seul M., L. O’Gorman, and M. J. Sammon, Practical Algorithms for Image Analysis,
Cambridge University Press, 2000.
[15] Sonka M., V. Hlavac, and R. Boyle, Image Processing, Analysis, and Machine Vision (2nd
ed), Brooks/Cole Publishing Company, 1999.
[16] Das S., A. Abraham and S. K. Sakar, “A Hybrid Rough Set – Particle Swarm Algorithm for
Image Classification.” Hybrid Intelligent System 6th Conference, 6 (2006): pp 26.
[17] Theodoris S. and K. Koutroumbas, Pattern Recognition (3rd Ed), Academic Press, 2006.
[18] Jang Q. and S. S. R. Abidi, “A hybrid of conceptual clusters, rough sets and attribute
oriented induction for inducing symbolic rules,” Machine Learning and Cybernetics
Conference, 9 (2005): pp 5573 – 5578.
[19] Hung, C. C. and G. Germany, "K-means and Iterative Selection Algorithms in Image
Segmentation,” in proceedings of IEEE Southeast Conference, Session 1: Software
Development, Jamaica, West Indies, April 4 - 6, 2003.
[20] Lingras, P and C. West, “Interval Set Clustering of Web Users with Rough K-Means,”
Journal of Intelligent Information Systems, 23.1 (2004): pp 5-16.
[21] Kennedy, J and R. C. Eberhar, “Particle Swarm Optimization,” Proceedings of IEEE

International Conference on Neural Networks, Perth, 1995.
[22] Pal, N. R., K. Pal, J. M. Keller and J. C. Bezdek, “A possibilistic fuzzy c-Means clustering
algorithm,” IEEE Transaction on Fuzzy Systems, 13. 4 (2005): pp 517-530.
[23] Dariusz, M. and T. W. Sławomir, “Standard and Genetic k-means Clustering Techniques
in Image Segmentation,” The 6th International Conference on Computer Information
Systems and Industrial Management Applications, (2007): pp 299-304.
[24] Bazan, J., R. Latkowski, S. H. Nguyen, P. Synak, A. Wojna, M. Wojnarski and J.
Wróblewski, Rough Set Exploration System Software. Version 2.2.
http://logic.mimuw.edu.pl/~rses
[25] Hung, C. C., H. Purnawan, and B.-C. Kuo, “Multispectral Image Classification Using
Rough Set Theory and the Comparison with Parallelepiped Classifier,” The IEEE
International Conference on Geosciences and Remote Sensing (IGARSS), Barcelona,
Spain, July 23-27, 2007 (CD).
[26] Hung, C. C. and H. Purnawan, “A ybrid Rough K-means Algorithm and Particle Swarm
Optimization for Image Classification,” Springer-Verlag, Lecture Notes in Computer
Science (LNAI 5317 pp. 585-593), MICAI 2008: Advances in Artificial Intelligence.
Derivative analysis of hyperspectral oceanographic data 597
29
X
Derivative analysis of hyperspectral

oceanographic data
Elena Torrecilla and Jaume Piera
Marine Technology Unit, Spanish National Research Council (UTM - CSIC)
Spain
Meritxell Vilaseca
Centre for Sensors, Instruments and Systems Development,
Technical University of Catalonia (CD6 - UPC)
Spain
1. Introduction
It has long been recognized that transmission of light in sea water is essential to the
productivity of the oceans. It provides the energy necessary for ocean currents, and the
majority of marine life is supported by the thin layer of warm water near the ocean's surface.
Light plays a decisive role in the primary formation of biomass by oceanic chlorophyll-
bearing marine plants, through the process of photosynthesis, which is the basis of the entire
marine food chain. Light transmission is therefore a key factor in the ecology of the upper
ocean and biogeochemical cycling, since it has a strong influence on the dynamics of the
chemical compounds. In addition, variability of light in the sea is strongly influenced by the
distribution of the components in the water column, which varies across horizontal and
vertical space and time scales.
The subdiscipline of ocean optics, which concerns the assessment of the propagation of light
through the oceanic water column and surface, has become fundamental for understanding
water dynamics and composition. Over the last decades, there has been a noticeable increase
in the number of coastal and open-ocean studies using in situ and remote sensing optical
measurements. Recent advances in monitoring surface and underwater optical properties
have been specially related to the progressive shift from using multispectral to high spectral
resolution (hyperspectral) acquisition systems. Hyperspectral technology has opened the
possibility for optical oceanographers to more accurately characterize complex oceanic
environments (Chang et al., 2004). A broad range of hyperspectral sensors, covering from
several hundreds to thousands of contiguous spectral bands, have recently been developed
for different monitoring applications. The use of high sample rate hyperspectral systems
may provide the potential for mapping phytoplankton functional types, including the
detection of harmful algal blooms (Nair et al., 2008; Craig et al., 2006).
The higher spectral resolution provides the opportunity to better perform spectral shape
analysis of oceanic optical measurements. Techniques such as derivative spectroscopy,
which serves to enhance subtle features in spectra, have been extensively used to assess
information regarding optically significant water constituents. Derivative of hyperspectral
data can yield more information than traditional analysis based on ratios of discrete spectral
bands (multispectral approaches). However, when hyperspectral measurements are to be
used for further analysis, the uncertainties of the measurement system must be taken into
account. Appropriate calibration strategies must be followed, since the accuracy of the
measurements relies on the associated uncertainties of the measurement system. A well-
calibrated spectroradiometer includes a characterization process, describing the instrument
behavior in terms of the spectral sensitivity, the signal-to-noise ratio, the dark current, the
wavelength calibration, the nonlinearity, the temperature dependence of the measurements
and the spectral scattering, or what is called the spectral stray-light (Brown et al., 2006). The
stray-light of a spectroradiometer is described as the unwanted background radiation,
scattered due to imperfections in the dispersive device and other internal optical elements.
In ocean optics applications, special attention must be paid to corrections for the inherent
distortions of the hyperspectral sensors (e.g. noise, spectral stray-light), because the errors in
the measured radiance distributions may be potentially significant and lead to inaccurate
retrievals of water properties. This issue becomes even more important when commonly
derivative spectroscopy is used to explore subtle features in shape of hyperspectral data
(Torrecilla et al., 2008a). Derivative analysis is notoriously sensitive to noise and smoothing
techniques must be used to overcome this problem. Therefore, in order to make an optimal
application of the derivative analysis, it is worth noting that a suitable selection of the
smoothing and derivative parameters (filter size and band separation) must be done
according to the resolution of each type of hyperspectral data (Torrecilla et al., 2007). An
effort must be made to determine the best trade-off between denoising and the ability to
resolve spectral features of interest.
There are two major goals in this chapter. The first one is to offer a brief review of relevant
advances that have been achieved using hyperspectral technology in oceanography and a
description of some important issues that must be considered when derivative spectroscopy
is applied to data acquired by hyperspectral sensors. The second goal is to provide some
results demonstrating the feasibility of applying derivative spectroscopy to hyperspectral
measurements of remote-sensing reflectance in open-ocean environments, in particular to
identify phytoplankton pigment assemblages. In this part, a simulation-based framework is
proposed to address this issue and special attention is given to the role of involved
parameters when derivative analysis of hyperspectral data is performed. For instance, the
effect of spectral sensitivity of each high spectral resolution instrument considered on the
results of the derivative-based approach is also discussed.
2. Ocean Optics: an overview

In order to understand light propagation in the ocean, which is an optically complex
environment, the radiative transfer theory was defined as the theoretical framework (Kirk,
1994; Mobley, 1994). Optical properties of sea water are classified into inherent and apparent
optical properties (IOPs, AOPs) and are related via the radiative transfer equation (RTE).
Most of the quantities are shown in Figure 1 and have wavelength dependence.
Fig. 1. Schematic diagram illustrating the relationships among the ocean optics quantities,
based on Figure 3.27 of Mobley, 1994.
Spectral absorption and scattering coefficients are inherent optical properties (IOPs), i.e.
they are only affected by the constituents of the aquatic medium and are independent of the
ambient illumination conditions. IOPs are additive and their variation is generally attributed
to variability in four constituents of the aquatic medium in natural oceanic waters: pure sea
water, phytoplankton, detritus and gelbstoff or colored dissolved organic matter (CDOM).
In contrast, the apparent optical properties (AOPs), such as spectral radiance and irradiance,
are not additive and depend both on the medium and on the directional structure of the
ambient light field. Several types of models have been reported to predict AOPs given field
measurements of IOPs and environmental conditions (“direct models”) or to estimate IOPs
given experimental measurements of AOPs (“inverse models”) (IOCCG, 2000).
In situ and remote sensed optical observations of IOPs and AOPs are two complementary
oceanographic applications that offer the potential of being direct proxies for important
biological and biogeochemical variables in the ocean (Chang et al., 2006). Remote sensing
provides global-scale optical data, whereas in situ measurements are useful for calibration
of remote sensing measurements and for obtaining higher temporal resolution data. Both
types of observations of the ocean are important for providing relevant information
regarding the relative concentrations of optically significant constituents in the water
column. Several studies have traditionally focused on the development of bio-optical
algorithms linking measurable optical properties to the primary pigment in phytoplankton,
chlorophyll-a, a proxy for the phytoplankton biomass (Bricaud et al., 1998; Reynolds et al.,
2001; O'Reilly et al., 2000), and on the detection of phenomena such as marine harmful algal
blooms (HABs) (Cullen et al., 1997). However, other approaches have been more recently
applied to estimate the surface concentration of particulate organic carbon (POC) from
optical measurements (Stramski et al., 2008), to derive the vertical distribution of
phytoplankton communities in the open-ocean based on the near-surface chlorophyll-a
content (Uitz et al., 2006), and to map phytoplankton functional types (PFTs) from ocean-
color data (Nair et al., 2008), which serve as a contribution to the mapping of biodiversity in
marine phytoplankton on the global scale.
3. Hyperspectral Data Acquisition Systems in Oceanography

Technological advances and innovations in instrumentation have given rise during the last
few decades to a broad range of optical sensors for different monitoring applications in
oceanography, including active and passive systems. Passive radiometric systems have been
widely used to measure properties such as radiance and irradiance distributions, and solar-
induced fluorescence. They can operate only with sunlight and are restricted to
measurements within the euphotic depth (i.e. to a depth where light intensity falls to 1% of
that at the surface). On the other hand, active radiometric systems use an internal light
source to measure other optical properties such as absorption, scattering and fluorescence.
These systems are able to provide optical measurements at deeper depths and at
wavelengths that have reduced penetration in the ocean (e.g. the UV portion of the
electromagnetic spectrum). However, their use involves a higher power consumption.
The rapid maturing of optical instrumentation has led to the most significant recent advance
which is the development of hyperspectral sensors. The ability to measure surface and
underwater light fields at hundreds of narrow and closely spaced wavelength bands, with a
resolution better than 10 nm, has become one of the most powerful and fastest growing
technologies in the field of ocean optics. Relative to multispectral measurements,
hyperspectral sensors potentially enable the extraction of more detailed spectral information
than is possible with any other type of in situ and remote sensing optical data. Figure 2
shows examples of phytoplankton absorption (IOP) and remote-sensing reflectance spectra
(AOP), gathered by multi- and hyperspectral systems. Hyperspectral patterns of absorption
or reflectance across wavelengths provide more information about spectral singularities.
The advantage of a hyperspectral over a multispectral inversion for phytoplankton species

identification stems from the fact that more accurate spectral information, such as spectral
features related to characteristic phytoplankton pigment absorption peaks, is resolved. In
the past, analyses of single band-ratios obtained in discrete multispectral bands were
employed to resolve a variety of water column properties (e.g. chlorophyll-a concentration
(O’Reilly et al., 2000), the presence of specific species and other water constituents
(Sathyendranath et al., 2004), etc.). However, the advent of hyperspectral sensors provided
ample spectral information and facilitated the identification of water column constituents,
such as phytoplankton species, which are spectrally unique. For instance, Craig et al. (2006)
assessed the feasibility of remote detection and monitoring of the toxic dinoflagellate,
Karenia brevis, applying two numerical methods to in situ hyperspectral measurements of
remote-sensing reflectance. Bracher et al. (2008) recently adapted the differential optical
absorption spectroscopy (DOAS) technique for retrieving the absorption and biomass of two
major phytoplankton groups of cyanobacteria and diatoms from high spectrally resolved
data of the satellite sensor SCIAMACHY (Scanning Imaging Absorption Spectrometer for
Atmospheric Chartography). Lohrenz et al. (2008) directly explored the hyperspectral
patterns of remote-sensing reflectance to better characterize water mass properties in coastal
areas, and Gagnon et al. (2008) performed sea-bed mapping of benthic assemblages in
optically complex shallow waters.
Fig. 2. (Left) Hyperspectral (solid continuous line, 1 nm resolution) and multispectral (closed
circles) phytoplankton absorption spectra. (Right) Hyperspectral (solid line, 1 nm
resolution) and multispectral (closed circles) remote-sensing reflectance spectra. (Absorption
and multispectral remote-sensing reflectance data were provided by Dariusz Stramski and
Rick Reynolds, Scripps Institution of Oceanography, from field measurements during the
ANT-XXIII/1 expedition on R/V Polarstern in the eastern Atlantic in 2005 (Stramski et al.,
2008)).
3.1 Hyperspectral Sensors

Hyperspectral sensors generally use diffraction gratings or linear variable optical filters as a
dispersive element to separate light into specific wave bands centered on desired
wavelengths, and different scanning mechanisms to generate one- or two-dimensional high-
resolution spectra or images. Mechanical-scanning spectrometers have been traditionally
employed to obtain a spectrum by using a single photo-detector and rotating the dispersive
element over a given specific spectral region after passing through the sample. However,
more recent spectrometers use a fixed grating and multi-element array detectors that allow
an entire spectrum to be acquired over some finite spectral region simultaneously.
Array spectrometers are being widely used as a tool for rapid measurements of spectral
distributions in oceanographic applications (Figure 3) in which acquisition speed is an
important issue. Spectra can be measured using three types of arrays: charge-coupled device
(CCD), photodiode or complementary metal-oxide semiconductor (CMOS). High
performance is available in all technologies today when they are designed properly, and
each have their own strengths and weaknesses. For instance, CCD arrays show a slightly
higher dynamic range (sensitivity) and a lower system noise than CMOS ones, whereas
CMOS ones offer more integration, smaller system sizes and higher speeds. The proper
selection of the type of array depends on many parameters (e.g. pixel dimensions,
sensitivity, spectral range coverage, dynamic range, saturation exposure, integration time)
and on the specific application. Other important advantages of array spectrometers are the
non moving parts, the robustness and the low production costs. A key challenge has also
been to construct small size array spectrometers without compromising performance.
However, they involve several drawbacks compared with mechanical-scanning
spectrometers, such as the fixed wavelength resolution, the lower sensitivity and the higher
stray-light radiation (described in more detail below), due to the lack of an output slit and
the integral illumination over the full wavelength range.
Fig. 3. Example of a miniature hyperspectral CMOS-array spectrometer designed for

oceanographic monitoring purposes and its interior view (Pons et al., 2007). The concave
grating splits the light collected through the input fiber and images each component in one
detector in the array.
New technologies and the miniaturization of electro-optical components have permitted the
development of accurate, low-cost and energy-efficient hyperspectral sensors designed to
measure oceanic high spectral resolution IOPs and AOPs (absorption, total upwelling or
downwelling radiance and irradiances, etc.). In the design and development of an
instrument for hyperspectral measurements in the ocean, it is necessary to consider several
optimal characteristics required of its detection system. The wavelength coverage should be
broad, ideally from 350 to 750 nm. A low-light-level detector should be selected with a large
dynamic range. Therefore, its responsivity should be high over the whole spectral region
covered, especially in the blue and red spectral portions, and the dark signal should be low.
Furthermore, in order to minimize a variety of external perturbations during measurement
(e.g. changing sky conditions, surface wave noise producing high-frequency fluctuations in
irradiance, ship shadow effect), the scan time of the system should be fast. The fastest
devices are multi-element array spectrometers with a fixed grating system.
3.2 Oceanographic Platforms for Hyperspectral Sensors

Currently available miniature spectrometers based on linear array detectors represent best
possible compromises between size, cost, resolution, range and performance. As a result,
their use is no longer limited to bench-top applications but they are being deployed on fully
fledged sea-going observational platforms. New interdisciplinary ocean observing platforms
incorporating hyperspectral sensors are being deployed (Dickey et al., 2006). Various
satellite-based hyperspectral imagers are currently available: NASA's Hyperion sensor on
the EO-1 satellite, ESA's CHRIS sensor on PROBA satellite or the U.S. Air Force Research
Lab's FTHSI sensor on the MightySat II satellite. In addition, examples of existing
hyperspectral airborne sensors providing high spatial resolution images of specific areas are
the Ocean Portable Hyperspectral Imager for Low-Light Spectroscopy (PHILLS, Davis et al.,
2002), the Compact Airborne Spectrographic Imager (CASI), the Airborne Visible Infrared
Imaging Spectrometer (AVIRIS), etc. Other in situ ocean observing platforms for which
hyperspectral sensors may be suitable are ships and moored buoys (Kuwahara et al., 2007).
These platforms offer continuous measurements at a high spatial and temporal resolution
delivering data even under cloudy conditions, thereby complementing discrete observations
by satellites and aircrafts and providing the data essential for calibration and validation
purposes. Furthermore, Lagrangian platforms that follow a particular water mass (e.g. floats
and drifters) can provide hyperspectral measurements in oceanic regions not normally
accessible by satellites or oceanographic research vessels. The goal of these research
initiatives is to determine the material composition of the water mass under investigation
through the accurate analysis of simultaneous optical and hydrographic measurements.
The number of newly emerging ocean observing platforms capable of incorporating

hyperspectral instrumentation is continuously growing. For instance, new autonomous
underwater vehicles (AUVs) have been developed recently with integrated equipment for
adaptive sampling and the ability to perform a wide-range of preprogrammed monitoring
surveys over extended periods of time (Perry & Rudnick, 2003). Another member of this
family of instruments are the special AUVs called gliders that propel themselves through
the water by changing their buoyancy with a minimal expenditure of energy. The glider
provides a wide spatial coverage measuring – among others - the optical properties during
periods of three or four weeks (Woods, 2009). Another approach to monitor the variability
of optical properties in the ocean is the use of profiling systems which can, for instance, be
part of some of the emerging cabled observatories. Instrument platforms such as the Vertical
Profiler System (VPS), part of Neptune Canada (Neptune, 2009), the world’s first regional-
scale underwater ocean observatory plugged directly into the Internet, offer great power
and bandwidth for gathering large quantities of hyperspectral measurements.
The amount of hyperspectral data sets available collected using all these types of platforms,
covering a wide range of temporal and spatial scales, will increase in the near future. New
interdisciplinary research initiatives are being successfully carried out such as the
Hyperspectral Coastal Ocean Dynamics Experiment (HyCODE, 2009) conceived to exploit
the new capabilities of hyperspectral ocean color sensors. Within this framework, several
collaborative short-term and long-term field experiments, involving the simultaneous use of
hyperspectral devices deployed in situ and remotely, are being conducted for calibrating,
groundtruthing and relating subsurface optical properties to remote sensing ocean color
measurements. These investigations are essential to develop and validate optical models
and to further our understanding of the processes that contribute to the temporal and spatial
variability of IOPs and AOPs in the ocean. One of the most important challenges of
HyCODE experiment is the standardization of the quality control protocols, including the
definition of procedures to perform detailed analysis of the reliability of measurements and
an evaluation of the impact of instrumentations inaccuracies.
4. Hyperspectral Data Pre- and Post-Processing Methods in Oceanography

4.1 Data Quality Assurance
When hyperspectral measurements can be performed and are to be used for further
analysis, the uncertainties of the measurement system must always be taken into account.
Appropriate calibration and pre-processing strategies must be undertaken, since the
accuracy of the measurements relies on the associated uncertainties of the measurement
system. The clearest example in optical oceanography is perhaps the in situ radiometric
measurements carried out during several field campaigns, which are still currently essential
to provide a proper calibration and validation of remote sensing measurements collected by
ocean color satellites and airborne platforms (Zibordi et al., 2001; McClain et al., 2004; Clark
et al., 2003).
A well-calibrated high spectral resolution array-based spectrometer, for both in situ and
remote sensing applications, should include likewise its characterization and an evaluation
of all meaningful sources of uncertainty (Lewis, 2008; Voss et al., 2008). It is therefore always
necessary to describe the instrument's behavior in terms of responsivity, signal-to-noise
ratio (SNR), dark current, nonlinearity, temperature dependence of measurements and
spectral scattering, or what is called the spectral stray-light of an instrument.
Changes in the thermal response of these systems are due to the silicon used for the arrays
of detectors. For instance, within the framework of the SORTIE (Spectral Ocean Radiance
Transfer Investigation Experiment) project, corrections to in situ hyperspectral radiometric
measurements were performed and demonstrated to be very repeatable (McLean, 2008). The
thermal characterization was carried out using shutters and installing a thermistor on the
diode array. The spectral scattering or stray-light radiation of an array-based spectrometer is
described as the unwanted background radiation that has been scattered due to
imperfections in the fixed dispersive element and other optical elements, such as higher-
order diffraction gratings, surfaces and internal baffles. The characterization of an
instrument's response for measurement errors arising from its spectral stray-light can be
performed using a set of monochromatic spectral line sources covering the entire
instrument's operational spectral range. This method is based on computing the ratio of the
spectral stray-light signal to the total signal within the bandpass of an array-based
spectrometer (Brown et al., 2006). Figure 4 shows an example of the effectiveness of this
spectral stray-light correction method applied to the hyperspectral CMOS-array
spectrometer from Figure 3 (Torrecilla et al., 2008a). The correction method has been proved
to be effective using a calibration broadband light source and a green absorption bandpass
filter. The stray-light signals outside the filter's bandpass region are clearly reduced by more
than two orders of magnitude, to a level of 10-4.
Fig. 4. Spectral stray-light correction applied to the miniature hyperspectral CMOS-array

spectrometer in Figure 3, using a broadband source with a bandpass filter. Normalized
measured and corrected signals. The y-axis is a logarithmic scale.
A data pre-processing and corrections for the spectral stray-light must be given special
attention for underwater optical measurement purposes, in which very weak signals are
collected and the errors in the measured radiometric distributions may be potentially
significant, leading to inaccurate retrievals of water properties. For instance, Figure 5 (left
panel) depicts measured and spectral stray-light corrected signals corresponding to field
underwater light measurements gathered by a miniature hyperspectral CCD-array
spectrometer at two different depths at a test site in the Alfacs Bay (Ebre Delta, NW
Mediterranean). It is worth noting that the difference in percentage (Figure 5, right panel)
between measured and spectral stray-light-corrected data at each depth has a spectral
dependency (i.e. it is not the same for each spectral band). Therefore, the amplitude and
shape of signals is modified when the spectral stray-light correction is applied. This is an
issue that becomes very important when spectral shape analysis techniques such as the
central topic of this chapter derivative spectroscopy, which is described below are used to
explore subtle features in hyperspectral data.
Fig. 5. (Left) Spectral stray-light correction applied to field underwater light measurements
at two depths, acquired by a miniature hyperspectral CCD-array spectrometer. (Right)
Difference in percentage between measured and corrected signals at each depth.
In addition to the above-mentioned corrections, new spectrometric devices and deployment

techniques are now being developed, in order to optimize spectrometer's throughput and
accuracy. Some spectrometers are designed with a two-dimensional area detector. They
offer a significant improvement in the signal-to-noise ratio by averaging a vertical row of
pixels to obtain each value of the whole spectrum at each wave band. Another solution is a
new deployment technique called the multicast technique, which consists in performing and
processing several casts in order to achieve the highest accuracy when underwater
radiometric measurements are collected by hyperspectral devices (McLean, 2008). These are
critical issues that should be tackled in addition to the radiometric and wavelength
calibrations that must be carried out periodically using standard and highly characterized
lamps (e.g. at the National Institute of Standards and Technology [NIST]).
4.2 Derivative Analysis

The emerging hyperspectral technology in remote sensing and in situ optical oceanography
leads to a need for ongoing evaluation and improvement of hyperspectral processing
methods. The higher spectral resolution provides the opportunity to develop and evaluate
advanced methods of spectral shape analysis, such as derivative spectroscopy, that better
distinguish subtle features in spectra and may be critical for discriminating optically
significant water constituents.
Derivative spectroscopy has been commonly used in the analysis of hyperspectral data
using different computation algorithms (Tsai & Philpot, 1998; Ruffin et al., 2008). The
process of estimating derivative spectra can be addressed using a finite divided difference
algorithm, named “finite approximation”, which consists in computing the changes in
curvature of a given spectrum over a sampling interval ( ∆λ ) or band separation ( BS ),
defined as ∆λ = λ j − λi , where λ j > λi . The first and the nth derivative are obtained using
Eqs. 1 and 2, respectively:
ds s (λi ) − s (λ j ) (1)
≈
dλ i ∆λ
d ns d  d ( n −1) s  (2)
=  
dλ n j
dλ  dλ( n −1) 
Derivative analysis can be applied to hyperspectral measurements of both inherent and

apparent oceanographic optical properties (e.g. absorption, irradiance, remote-sensing
reflectance). It is a useful tool for enhancing spectral features, which are often relevant, for
instance, because they can be related to absorption bands of pigments present in the
considered water samples (Figure 6, left panels). Derivatives of hyperspectral data are less
affected by possible spectral fluctuations of skylight conditions. However, successful
extraction of spectral details of interest through derivative spectroscopy depends on the
band separation chosen. The importance of selecting a suitable band separation ( ∆λ = BS ) in
each case stems from the fact that spectral data features of interest with a smaller scale than
the band separation will not be preserved in the derivative results. Figure 6 (right panel)
shows different derivatives of the spectrum shown in the left panel, computed according to
different finite band resolutions or band separations ( BS ).
Fig. 6. (Left) Original spectrum and its second derivative of a water sample containing a
unique phytoplankton algal culture (Alexandrium minutum), acquired by a miniature
hyperspectral CCD-array spectrometer. (Right) Second derivatives computed for several
values of band separation ( BS ), each of them leading to spectral features at different scales.
Noise level in hyperspectral data can be considerable, as the small amount of energy
gathered by the narrow bandwidth may be exceeded by the intrinsic sensor's noise. In order
to make an optimal application of derivative analysis, which is a technique clearly sensitive
to noise, smoothing techniques must be applied previously to derivative computation of
hyperspectral data (Vaiphasa, 2006). A number of smoothing algorithms have been
developed within the last few decades (e.g. Savitzky-Golay, Kawata-Minami or mean-filter
smoothing). In all approaches the smoothing level applied depends on the size of the filter
window ( WS ). It is therefore worth noting that an appropriate selection of the smoothing
and derivative parameters (filter size and band separation) must be done according to the
resolution of each type of hyperspectral data (Torrecilla et al., 2007). An important effort
must be made to determine the best compromise between denoising and the ability to
resolve fine spectral details of interest.
The advantages offered by hyperspectral measurements of remote-sensing reflectance

(Rrs(λ)) in combination with derivative spectroscopy have been recently exploited for
various purposes in optical oceanography. For example, Louchard et al. (2002) assessed
qualitative and quantitative information regarding major sediment pigments of benthic
substrates from derivatives of shallow marine hyperspectral Rrs(λ). In addition, a
bathymetric algorithm was suggested through the analysis of the magnitude of some
derivative peaks. Craig et al. (2006) detected a toxic algal bloom from the analysis of the
fourth derivative of phytoplankton absorption spectra, estimated from in situ hyperspectral
measurements of Rrs(λ) using a quasi-analytical inversion algorithm. The monitoring of
these harmful algal blooms is possible because some accessory pigments are unique to
individual phytoplankton taxa and can be better differentiated in hyperspectral absorption
spectra than in multispectral spectra with a limited number of wavelengths. The advantage
of a hyperespectral over a multispectral inversion for identification of other algal blooms
was also confirmed by Lubac et al. (2008). In this case, the inversion and quantitative
assessment are directly based on the analysis of the position of the maxima and minima of
the second derivative of Rrs(λ).
4.3 A model-based approach to evaluating the efficacy of processing methods

Most of the studies dedicated to derivative analysis of hyperspectral remote-sensing
reflectance have traditionally focused on the use of ratios of discrete values of derivative
spectra to characterize or classify oceanic environments. However, more advanced and
contemporary methods are beginning to take advantage of the information contained in the
whole derivative spectrum. For example, Filippi (2007) proposes a combination of derivative
spectroscopy and artificial neural network (ANN) algorithms. The derivative-neural
approach has been proven to be effective for providing bathymetry, bottom type and
constituent concentration estimations from remote-sensing reflectance (Rrs(λ))
measurements.
In order to test the effectiveness of such a methodology, or the one described in this chapter,
it would be desirable to have a large number of observed oceanographic hyperspectral data
sets, covering a wide range of environmental conditions. However, because valid
hyperspectral data sets are still difficult to obtain and often unavailable, oceanic radiative
transfer (RT) models are used. Hydrolight is an example of a radiative transfer numerical
model (Mobley, 1994; Mobley & Sundman, 2008), which computes radiance distributions
and derived quantities (e.g. Rrs(λ)) given water column IOPs and other oceanographic
environmental conditions. The Hydrolight code employs mathematically sophisticated
invariant imbedding techniques to solve the radiative transfer equation (RTE, Figure 1) and
offers the possibility of performing numerical simulations in controlled environments
(Albert & Mobley, 2003; Kempeneers et al., 2005). For instance, Figure 7 depicts several
Hydrolight-generated hyperspectral Rrs(λ) spectra, corresponding to different open-ocean
scenarios, each dominated by a single phytoplankton group.
Fig. 7. Results of Hydrolight radiative transfer simulations, showing four remote-sensing

reflectances Rrs(λ) (1 nm resolution) obtained under different composition conditions. Labels
indicate each dominant phytoplankton group considered in each case.
One of the logical steps for improving the Hydrolight-based approaches, when simulated
hyperspectral Rrs(λ) are to be used as a basis for further validation of some processing
techniques, is to conduct sensor experiments. With the goal of exploring the potential of any
processing technique, the response of the sensor with which the Rrs(λ) acquired would be
acquired must be taken into account. More accurate and realistic retrievals will be carried
out from hyperspectral oceanographic data if, as has been stated as a priority in Section 4.1,
the effect of the sensor in terms of noise, sensitivity, spectral resolution, stray-light, etc. is
included in the simulation-based approach. This is one of the main issues that will be
discussed in the next section.
5. Experimental Results
5.1 Experimental Design
In the following study, we investigate the potential offered by hyperspectral sensors and
derivative spectroscopy to identify phytoplankton assemblages from hyperspectral
measurements of remote-sensing reflectance (Rrs(λ)) in open-ocean waters. A simulation-
based framework was used to achieve this goal, which includes the use of the Hydrolight-
Ecolight version 5 radiative transfer model (Mobley & Sundman, 2008). The analysis
methodology of this research is presented schematically in Figure 8.
Fig. 8. Flowchart showing the methodology followed for testing the automatic identification
of phytoplankton assemblages from hyperspectral Rrs(λ) and derivative spectroscopy.
The generation of a set of Hydrolight-simulated hyperspectral Rrs(λ) spectra (1 nm

resolution) takes into account the constituents present in the water column and their
inherent optical properties (absorption, scattering, etc.). The sensor's behavior can also be
included in the Rrs(λ) computation (see dashed box in Figure 8). The next step defined in the
diagram is the computation of the second derivative of normalized Rrs(λ) spectra (i.e. Rrs(λ)
normalized by the Rrs(λ) obtained at λ=555nm). A normalized Rrs(λ) is used as an input in
order to emphasize the shape singularities of each spectrum before derivative spectroscopy.
Previous to derivative analysis, a mean filter smoothing type is also applied to normalized
Rrs(λ) data. It consists of a simple average of points within the chosen filter window. As
stated in Section 4.2, the efficient way to accomplish the smoothing and derivative is to
carefully adapt the filter size and the sampling interval (BS) to better match the scale of the
spectral features of interest in each case.
Finally, the goal of validating the potential of hyperspectral Rrs(λ) measurements for
identifying phytoplankton assemblages is addressed through the comparison of second
derivative spectra instead of discrete values of Rrs(λ) spectra themselves. To make such a
comparison, an approach based on hierarchical cluster analysis (HCA) is used (Jain et al.,
1999). HCA is a common methodology consisting in creating a hierarchical cluster tree to
partition a data set into subsets (clusters) using a single linkage algorithm. The linkage
algorithm is based on a previously calculated pairwise distance between observations (i.e.
each second derivative of a normalized Rrs(λ) spectrum). The selected distance measure
determines how the similarity of two spectra is calculated. In this case, one minus the cosine
of the included angle between two vectors was used as a distance measure (cosine distance).
As a linkage algorithm, the shortest distance between vectors, also called the nearest
neighbor (single linkage), was selected. The traditional representation of this hierarchical
tree is a dendrogram, with individual elements at one end and a single cluster containing
every element at the other. Note that the smaller the cosine distance between two
observations, the more similar are the features of the two compared derivatives of
normalized Rrs(λ) spectra. Therefore, spectra corresponding to observations with a similar
phytoplankton composition are expected to appear closer in the dendrogram than those
having a very different phytoplankton composition. The feasibility of applying derivative
spectroscopy to hyperspectral measurements of Rrs(λ) to identify phytoplankton
assemblages will be assessed by analyzing how close spectra belonging to similar
phytoplankton assemblages appear in the computed cluster tree (or dendrogram).
5.2 Automatic Identification of Phytoplankton Assemblages

The results shown in this subsection provide the first detailed demonstration of the
advantages and limitations of integrating hyperspectral Rrs(λ) data and derivative
spectroscopy for the automatic identification of phytoplankton assemblages in open-ocean
waters. A hyperspectral Rrs(λ) data set, covering a range of environmental conditions, was
created using absorption and scattering properties of six different phytoplankton groups
(Figure 9, left panel) (Kim and Philpot, 2006).
Fig. 9. (Left) Specific absorption spectra of six phytoplankton groups. (Right) Simulated
hyperspectral Rrs(λ) spectra (1 nm resolution) corresponding to different dominating
phytoplankton groups at different concentrations (0.01, 0.03, 0.05, 0.07 and 0.09 mg/m3).
A total of thirty hyperspectral Rrs(λ) spectra were simulated, each of them dominated by a
single phytoplankton group along the water column, which was assumed to be
homogeneous. The spectra were generated from combining six different dominating
phytoplankton groups and five different concentration values (Figure 9, right panel)
corresponding to typical concentrations encountered under non-bloom sea conditions. It
must be noted that Rrs(λ) spectra display great variability in both magnitude and spectral
shape, in correspondence with the variable phytoplankton composition and concentration
rates.
Spectral derivatives and clustering techniques (HCA) were used independently to identify
and group similar phytoplankton assemblages from hyperspectral Rrs(λ) spectra. Figure 10
shows the results of cluster analysis applied to the raw hyperspectral normalized Rrs(λ)
spectra (top panel) and to the second derivative of hyperspectral normalized Rrs(λ) spectra
(central panel). Each simulated Rrs(λ) is identified with a specific label, consisting of the
name of the dominating phytoplankton group and the concentration value. For instance, if
diatoms are the dominating phytoplankton group, with a concentration of 0.05 mg/m3
along the homogeneous water column, the label identifying that case will be Diat_0.05. It is
worth noting that when the raw normalized Rrs(λ) is used for cluster analysis (Figure 10, top
panel), the only group of Rrs(λ) spectra which is clustered satisfactorily is the one
corresponding to the Prasinophyceae phytoplankton group (see labels in cyan color). The
remaining cases, from different phytoplankton groups, are mixed. However, when second
derivatives of normalized Rrs(λ) are used, the majority of phytoplankton groups at different
concentration rates are identified and connected by closer dendrites in the dendrogram.
Therefore, derivatives of spectra from the same phytoplankton group (and color) are located
closer in the dendrogram. The only cases which are not well grouped are those
corresponding to the lowest concentration rates (i.e. 0.01, or in some case 0.03 mg/m3). The
method does not resolve different phytoplankton assemblages at such a low concentration
and groups these cases together in the same cluster (see the multicolor cluster at the bottom
of the central panel in Figure 10).
The results confirm the potential of using all the information contained in the derivative of
hyperspectral normalized Rrs(λ) in comparison with the use of multispectral measurements
or band ratios of discrete spectral values, and are in concordance with the results of
Torrecilla et al. (2008b). The HCA cluster analysis based on the hyperspectral input
information, with variable optical conditions, was able to automatically bring together
assemblages corresponding to the same phytoplankton groups. Furthermore, the
performance of this derivative analysis tool is dependent on the examination of the spectral
data and on the selection of suitable parameters (i.e. smoothing filter size and band
separation) for each particular data set and for each specific purpose of the analysis. If
smoothing and derivative analysis parameters are, for instance, selected too coarse, spectral
features of interest will be lost and worse results will be achieved in the automatic
identification of phytoplankton assemblages using clustering analysis of derivatives of Rrs(λ)
spectra. This is the case shown in Figure 10 (bottom panel), where it can be seen that
Prymnesiophyceae and diatoms are grouped together, and spectra corresponding to the
group of Cyanophyceae (in blue) appear mixed with other phytoplankton groups.
Fig. 10. Cluster analyses based on: (top) raw hyperspectral normalized Rrs(λ), (centre) second
derivative of hyperspectral normalized Rrs(λ) spectra and (bottom) second derivative of
hyperspectral normalized Rrs(λ) spectra using analyzing parameters that are too coarse (i.e.
derivative band separation and smoothing filter size).
5.3 Impact of the sensor’s response on the proposed automatic identification

A sensor sensitivity analysis was also conducted to test the effect of the sensor’s response,
with which the radiometric measurements would hypothetically be acquired in the
proposed methodology (Figure 8) to identify different phytoplankton assemblages from
derivatives of hyperspectral Rrs(λ) spectra.
The response of a miniature hyperspectral CMOS-array spectrometer (Figure 3) designed for

oceanographic monitoring purposes (Pons et al., 2007; Torrecilla et al., 2008a) was
characterized in terms of sensitivity. The spectral sensitivity curve obtained was applied to
both simulated hyperspectral water-leaving radiance (Lw(λ)) and downwelling irradiance
(Ed(λ)), the ratio of which was used to estimate the new data set of Rrs(λ) spectra. Each new
Rrs(λ) spectrum (Figure 11, red curve in the left panel) now has a spectral resolution of 3 nm
approximately, in contrast to the 1 nm resolution of Hydrolight-modeled Rrs(λ), since it has
been adapted to the sensor’s spectral resolution. The magnitude of Rrs(λ) spectra has not
been essentially modified after the consideration of the sensor’s spectral sensitivity. This is
because remote-sensing reflectance is an apparent optical property obtained as the ratio of
two radiometric properties. This rationing serves as an effective way to remove the effects of
the magnitudes of Ed(λ), Lw(λ), and other possible external light fluctuations. However, the
small differences between Rrs(λ) spectra with 1 and 3 nm spectral resolution become more
evident on the results of derivative analysis (Figure 11, right panel) and will play an
important role on the further clustering analysis.
Fig. 11. (Left) An example of a simulated hyperspectral remote-sensing reflectance, Rrs(λ),

before and after applying the modeled sensor’s response (black and red curves,
respectively). (Right) Second derivatives of spectra in the left panel.
Analogously to the previous subsection, derivative and clustering analysis was carried out
with this new data set of hyperspectral-sensor Rrs(λ). As can be seen in Figure 12, a proper
identification and clustering of similar phytoplankton assemblages was only possible when
derivative spectra were considered (right panel). Again, a cluster was also created by
grouping all those cases corresponding to low concentrations (0.01 mg/m3).
Fig. 12. Results of analog derivative and cluster analysis when the hyperspectral sensor’s
response in terms of sensitivity is considered based on: (left) raw hyperspectral-sensor
normalized Rrs(λ) and (right) second derivative of hyperspectral-sensor normalized Rrs(λ)
spectra.
Though similar results were obtained, even including the effect of the sensor’s response in
terms of sensitivity, it should be pointed out that the applied derivative spectroscopy
needed to be adapted. Suitable values of smoothing filter size and derivative band
separation were selected because of the different spectral resolution of the new Rrs(λ) data
set. Furthermore, based on the results from this experiment, the selected low-cost and
miniature hyperspectral CMOS-array spectrometer (Pons et al., 2007) has been confirmed as
a potential tool for water component detection and monitoring.
6. Conclusion and Future Work

Derivative spectroscopy has been satisfactorily applied to numerical simulations of
hyperspectral remote-sensing reflectance (Rrs(λ)) corresponding to different open-ocean
environments. Its feasibility for identifying phytoplankton pigment assemblages in
comparison with the use of raw Rrs(λ) spectra or traditional ratios of discrete spectral bands
(Torrecilla et al., 2008b) has been confirmed using a validation approach based on
hierarchical cluster analysis (HCA). Simulation-based experiments have also been
performed, including a model of the response in terms of sensitivity of a hyperspectral
sensor, with which radiometric measurements would hypothetically be acquired. The aim
was to analyze the sensor’s effect on the proposed methodology of analysis. For each
simulated hyperspectral data set, a proper adaptation of parameters involved in derivative
analysis (i.e. smoothing filter size and derivative band separation) was necessary according
to the characteristics of the signals and particularly their spectral resolution.
The experiments yield promising results when all the information contained in the second
derivative of hyperspectral Rrs(λ) spectra is considered. This method can therefore provide a
means for optical oceanographers to better characterize complex oceanic waters, detect
harmful algal blooms or map phytoplankton functional types from hyperspectral
oceanographic information. The recent advances in hyperspectral technology (e.g.
miniaturization and power supply reduction) have given rise to a great number of sensor
configurations that are suitable for incorporating in a large number of in situ and remote
sensing platforms of oceanographic observing systems (e.g. satellites, gliders), which will
allow these challenges to be overcome once hyperspectral data become more available.
Future research will focus on experiments based on larger field hyperspectral data sets,
since the potential of integrating hyperspectral Rrs(λ) measurements and derivative
spectroscopy has been emphasized in this work. However, the design of a more realistic
approach when simulated-based experiments are employed would be useful for better
validating the efficacy of the proposed method. An effort of detailed consideration of several
the above-mentioned factors involved in the analysis process is suggested. For instance, the
possibility of having some variability of the inherent optical properties (IOPs) used as an
input in the radiative transfer modeling should be explored. Furthermore, the accuracy of
the derivative-based method for identifying different phytoplankton compositions would be
improved if distortion experiments, caused by the sensor (noise, spectral stray-light, thermal
effects, etc.) were included in the simulation-based approach.
7. Acknowledgements
This study was supported by the projects HIDRA (PIE06-301102) and ANERIS (PIF08-015)
funded by the Spanish Ministry of Science and Innovation.
8. References
Albert, A. & Mobley, C. D. (2003). An analytical model for subsurface irradiance and remote
sensing reflectance in deep and shallow case-2 waters. Opt. Express., 11(22), 2873–
2890.
Bracher, A.; Vountas, M.; Dinter, T.; Burrows, J.P.; Röttgers, R. & Peeken, I. (2008).
Quantitative observation of cyanobacteria and diatoms from space using
PhytoDOAS on SCIAMACHY data. Biogeosciences Discussions, 5, 4559-4590.
Bricaud, A.; Morel, A.; Babin, M.; Allali, K. & Claustre, H. (1998). Variations of light
absorption by suspended particles with chlorophyll a concentration in oceanic (case
1) waters: Analysis and implications for bio-optical models. J. Geophys. Res., 103,
31,033–31,044.
Brown, S. W.; Eppeldauer, G. P. & Lykke, K. R. (2006). Facility for spectral irradiance and
radiance responsivity calibrations using uniform sources. Applied Optics, Vol. 45,
No. 32, pp. 8218-8237.
Chang, G. C.; Mahoney, K.; Briggs-Whitmire, A.; Kohler, D.; Mobley, C.; Moline, M.; Lewis,
M.; Boss, E.; Kim, M.; Philpot, W. & Dickey, T. (2004). The New Age of
Hyperspectral Oceanography. Oceanography, 17(2), 22-29.
Chang, G. C.; Dickey, T. & Lewis, M. (2006). Toward a global ocean system for
measurements of optical properties using remote sensing and in situ observations.
In: Remote Sensing of the Marine Environment: Manual of Remote Sensing, Volume 6,
Chapter 9, edited by J. Gower, 285-326.
Clark, D. K.; Yarbrough, M. A.; Feinholz, M. E.; Flora, S.; Broenkow, W.; Kim, Y. S.; Johnson,
B. C.; Brown, S. W.; Yuen, M. & Mueller, J. L. (2003). MOBY, a radiometric buoy for
performance monitoring and vicarious calibration of satellite ocean color sensors:

measurement and data analysis protocols. Ocean Optics Protocols for Satellite Ocean
Color Sensor Validation, NASA Technical Report Series, Revision 4, Volume 6, Mueller, J.
L., Fargion, G. S. and McClain, C. R., Greenbelt, MD, NASA Goddard Space Flight
Center, NASA/TM-2003-211621, 3-34.
Craig, S. E.; Lohrenz, S. E.; Lee, Z.; Mahoney, K. L.; Kirkpatrick, G. J.; Schofield, O. M. &
Steward, R. G. (2006). Use of hyperspectral remote sensing reflectance for detection
and assessment of the harmful alga, Karenia brevis. Applied Optics, Vol. 45, No. 21,
pp. 5414-5425.
Cullen, J.J.; Ciotti, A.M.; Davis, R.F. & Lewis, M.R. (1997). Optical detection and assessment
of algal blooms. Limnol. Oceanogr., 42 (5), 1223–1239.
Davis, C. O.; Bowles, J.; Leathers, R.A.; Korwan, D.; Downes, T.V.; Snyder, W.A.; Rhea, W.J.;
Chen, W.; Fisher, J.; Bissett W.P. & Reisse, R.A. (2002). Ocean PHILLS hyperspectral
imager: Design, characterization, and calibration. Opt. Express, 10(4), 210–221.
Dickey, T.; Lewis, M. & Chang, G. (2006). Optical oceanography: recent advances and future
directions using global remote sensing and in situ observations. Reviews of
Geophysics, 44, RG1001, 1-39.
Filippi, A. M. (2007) Derivative-Neural Spectroscopy for Hyperspectral Bathymetry
Inversion. The Professional Geographer, 59 (2), 236-255.
Gagnon, P.; Scheibling, R.E.; Jones, W. & Tully, D. (2008). The role of digital bathymetry in
mapping shallow marine vegetation from hyperspectral image data. Int. Journal of
Remote Sensing, 29, 3, 879–904.
HyCODE (2009). Retrieved on 1 July 2009. <http://www.opl.ucsb.edu/hycode.html>
IOCCG, International Ocean-Colour Coordinating Group (2000). Remote sensing of ocean
colour in coastal, and other optically-complex waters, edited by S. Sathyendranath,
IOCCG Rep. 3, pp 140, Dartmouth, N. S., Canada.
Jain, A.K.; Murty, M. N. & Flynn, P. J. (1999). Data clustering: A review. ACM Comput.
Surveys, 31, 3, 264-323.
Kempeneers, P.; Sterckx, S.; Debruyn, W.; De Backer, S. ; Scheunders, P.; Park, Y. & Ruddick,
K. (2005). Retrieval of oceanic constituents from ocean color using simulated
annealing, Proceedings of the 25th International Geoscience and Remote Sensing
Symposium (IGARSS), Seoul.
Kim, M. & Philpot, W. (2006) Ocean Optical Phytoplankton Simulator. Retrieved on 12 June
2009. <http://ceeserver.cee.cornell.edu/wdp2/OOPS/oops_manual.html>
Kirk, J. T. O. (1994). Light and Photosynthesis in Aquatic Ecosystems, 2nd ed., pp. 509,
Cambridge Univ. Press, New York.
Kuwahara, V.S.; Chang, G. & Dickey, T.D. (2007). Innovations in ocean optics for coastal and
open ocean mooring applications, Proceedings of IEEE/OEE Oceans Conference and
Exhibition, OCEANS'07 Europe, Aberdeen.
Lewis, M. (2008) Spectral Ocean Radiance Transfer Investigation Experiment (SORTIE), Pan
Ocean Remote Sensing Conference, Guangzhou, China.
Lohrenz, S.E.; Cai, W., Chen, X. & Tuel, M. (2008). Characterizing water mass properties in
river dominated coastal waters using underway hyperspectral remote sensing
reflectance, Proceedings of Ocean Optics XIX Conference, Barga, Italy.
Louchard, E. M.; Reid, R. P.; Stephens, C. F.; Davis, C. O.; Leathers, R. A.; Downes, T. V. &
Maffione, R. (2002). Derivative Analysis of Absorption Features in Hyperspectral
Remote Sensing Data of Carbonate Sediments. Optics Express, 10(26),1573–1584.
Lubac, B.; Loisel, H.; Guiselin, N.; Astoreca, R.; Felipe Artigas, L. & Mériaux, X. (2008)
Hyperspectral and multispectral ocean color inversions to detect Phaeocystis
globosa blooms in coastal waters. J. Geophys. Res., 113, C06026.
McClain, C. R.; Feldman, G. C. & Hooker, S. B. (2004). An overview of the SeaWiFS project
and strategies for producing a climate research quality global ocean bio-optical
time series. Deep Sea Research II, 51, 5-42.
McLean, S. (2008). Radiometric Data Processing Course, Ocean Optics XIX Conference, Barga,
Italy.
Mobley, C. D. (1994) Light and water: radiative transfer in natural waters, pp. 592, Academic
Press, San Diego, CA.
Mobley, C. D. & Sundman, L. (2008). Hydrolight 5.0 Users’ guide. Sequoia Scientific, Inc.,
Bellevue, WA.
Nair, A.; Sathyendranath, S.; Platt, T.; Morales, J.; Stuart, V.; Forget, M.; Devred, E. &
Bouman, H. (2008). Remote sensing of phytoplankton functional types. Remote
Sensing of Environment, 112, 3366−3375.
Neptune (2009). Retrieved on 26 June 2009. <http://neptunecanada.ca>
O’Reilly, J. E.; Maritorena, S.; Siegel, D. A. et al (2000). Ocean Color Chlorophyll a
Algorithms for SeaWiFS, OC2, and OC4: Version 4. SeaWiFS Postlaunch Technical
Report Series, Volume. 11, SeaWiFS Postlaunch Calibration and Validation Analyses, Part
3, edited by Hooker, S. B. and Firestone, E. R., Greenbelt, Maryland, NASA/TM-
2000-206892, 9–27.
Perry, M. J. & Rudnick, D.L. (2003). Observing the oceans with autonomous and Lagrangian
platforms and sensors: The role of ALPS in sustained ocean observing systems.
Oceanography, 16(4), 31–36.
Pons, S.; Aymerich, I.F.; Torrecilla, E. and Piera, J. (2007). Monolithic spectrometer for
environmental monitoring applications, Proceedings of IEEE/OEE Oceans Conference
and Exhibition, OCEANS'07 Europe, Aberdeen.
Reynolds, R. A.; Stramski D. & Mitchell, B. G. (2001). A chlorophyll-dependent
semianalytical model derived from field measurements of absorption and
backscattering coefficients within the Southern Ocean. J. Geophys. Res., 106, 7125–
7138.
Ruffin, C.; King, R.L. & Younan, N.H. (2008). A combined derivative spectroscopy and
Savitzky-Golay filtering method for the analysis of hyperspectral data. GIScience
and Remote Sensing, 45 (1), 1-15.
Sathyendranath, S.; Watts, L.; Devred, E.; Platt, T.; Caverhill, C. & Maass, H. (2004).
Discrimination of diatoms from other phytoplankton using ocean colour data. Mar.
Ecol. Prog. Ser., 272, 59– 68.
Stramski, D.; Reynolds, R.A.; Babin, M.; Kaczmarek, S.; Lewis, M.R.; Röttgers, R.; Sciandra,
A.; Stramska, M.; Twardowski, M.S.; Franz, B.A. & Claustre, H. (2008).
Relationships between the surface concentration of particulate organic carbon and
optical properties in the eastern South Pacific and eastern Atlantic Oceans.
Biogeosciences 5: 171-201.
Torrecilla, E.; Aymerich, I.F.; Pons, S. & Piera, J. (2007). Effect of spectral resolution in
hyperspectral data analysis. Proceedings of International Geoscience and Remote
Sensing Symposium, pp. 910-913, Barcelona.
Torrecilla, E.; Pons, S.; Vilaseca, M.; Piera, J. & Pujol, J. (2008a). Stray-light correction of in-
water array spectroradiometers. Effects on underwater optical measurements.
Proceedings of IEEE/OEE Oceans Conference and Exhibition, Quebec.
Torrecilla, E.; Stramski, D.; Reynolds, R. A.; Piera, J. & Millan Nunez, E. (2008b).
Identification of phytoplankton pigment assemblages using spectral shape analysis
of hyperspectral remote-sensing reflectances, Proceedings of Ocean Optics XIX
Conference, Barga, Italy.
Tsai, F. & Philpot, W.D. (1998). Derivative analysis of hyperspectral data. Remote Sensing of
Environment, 66, 1, 41-51.
Uitz, J.; Claustre, H.; Morel, A. & Hooker, S. (2006). Vertical distribution of phytoplankton
communities in open-ocean: An assessment based on surface chlorophyll. Journal of
Geophysical Research, 111, CO8005, doi:10.1029/2005JC003207.
Vaiphasa, C. (2006). Consideration of smoothing techniques for hyperspectral remote
sensing. Journal of Photogrammetry and Remote Sensing, 60, 2, 91-99.
Voss, K.; Gordon, H.; Lewis, M.; Johnson, C.; Yarbrough, M.; Flora, S.; Feinholz, M., & Trees,
C. (2008). Radiometry and Uncertainties from SORTIE (Spectral Ocean Radiance
Transfer Investigation and Experiment), NASA Carbon Cycle and Ecosystems Joint
Science Workshop.
Wood, S. (2009). Autonomous underwater gliders. In: Underwater Vehicles, Chapter 26, edited
by A. V. Inzartsev, 499-524, In-Tech, Vienna, Austria.
Zibordi, G.; D’Alimonte, D.; van der Linde, D.; Berthon, J.-F.; Hooker, S.B.; Mueller, J.L.;
Lazin, G. & McLean, S. (2002). The Eighth SeaWiFS Intercalibration Round-Robin
Experiment (SIRREX-8), September–December 2001. NASA Tech. Memo. 2002-206892,
Volume 21., edited by Hooker, S.B. & Firestone, E.R., pp. 39, NASA Goddard Space
Flight Center, Greenbelt, MD.
Electromagnetic Scattering Analysis of Simple Targets
Embedded in Planar Multilayer Structures: Remote Sensing Applications 619
30
X
Electromagnetic Scattering Analysis of Simple

Targets Embedded in Planar Multilayer
Structures: Remote Sensing Applications
Sidnei J. S. Sant’Anna(1,2), J. C. da S. Lacava(2) and David Fernandes(2)
(1)Instituto Nacional de Pesquisas Espaciais
(2)Instituto Tecnológico de Aeronáutica
Brazil
1. Introduction
The potential and usefulness of SAR (Synthetic Aperture Radar) data for land uses and land
cover change applications have been demonstrated over the past few decades. More
recently, polarimetric SAR data have been widely used in electromagnetic modeling and
inverse problems. Moreover, the combined use of numerical modeling and inversion
methods has enabled computerized quantitative information extraction of the polarimetric
data. In particular, the electromagnetic modeling can be used to increase the knowledge of
non-metallic target reflectivity. The problem of detecting the many types of non-metallic
land mines in the presence of widely varying soil background is an example of one of the
most difficult subsurface sensing problems.
The target electromagnetic scattering characterization (van Zyl et al., 1987; Cloude &
Pottier, 1996; Freeman & Durden, 1998; Sant’Anna et al., 2007a) might influence some
remote sensing practical applications, as SAR image classification, sensor calibration among
others. Consequently, understanding the target scattering mechanism plays a valuable role
in the analysis and interpretation of polarimetric SAR data. In order to improve the insight
into scattering mechanism subject, several electromagnetic models have been proposed in
literature (Lin & Sarabandi, 1999; Israelsson et al., 2000; Zurk et al., 2001). Usually, in these
models complex targets are modeled by a collection of targets having simple and canonical
geometries. For instance, a three can have its elementary parts as trunk, branches and leaves,
built through a set of dielectric cylinders and discs. Besides, depending on the modeling the
simple targets have different sizes, shapes, orientations and dielectric properties. Therefore,
the study of electromagnetic scattering of simple targets is fundamental to enhance the
comprehension of the electromagnetic characteristics of more complex ones.
Keeping this issue in mind, the purpose of this work is to show the potentiality and
applicability of an electromagnetic model within remote sensing subject. This modeling is
based on scattering elements embedded in a multilayer planar structure, which is
illuminated by an elliptically polarized plane wave at oblique incidence. Therefore, the
chapter is organized as follow. In Section 2, the multilayer structure under analysis is
described jointly with an explanation of the electromagnetic modeling employed to solve

the scattering problem. The electromagnetic model usefulness is illustrated in Section 3,
through some remote sensing applications. The Section 4 is devoted to draw the conclusions
concerning the obtained results.
2. Electromagnetic Model
The electromagnetic model is based on the determination of the electromagnetic fields scattered
by a multilayer planar structure that is excited by plane waves. The structure under analysis is
composed of N+2 isotropic, linear and homogenous layers stacked up in z direction. The layers
are assumed to be unbounded along the x and y directions. The lower layer, having complex
permittivity g and complex permeability g, is denoted as ground layer and occupies the
negative-z region. The next N layers are characterized by thickness ℓn, complex permittivity n
and complex permeability n, where 1  n  N. The planar interface z = dN separates the N-th
layer from free space (the upper layer). Elements supporting electric (J) or magnetic (M) surface
current densities are printed at arbitrary positions on each one of the N+1 interfaces of the
structure. These elements will behave as scattering elements. The development is based on a
global right-handed rectangular coordinate system located on the top of the ground layer
(interface z = 0) and lying on the xy-plane. The geometry of the planar multilayer structure is
depicted in Figure 1.
Free Space
0 , 0)
Tx
Rx
z
i s
z = dN
JN Layer N
N MN
N , N)
z = dN-1
JN-1 MN-1
Layer N-1
 N 1 N-1 , N-1)
z = dN-2
JN-2 MN-2
z = d2
J2 M2
Layer 2
2 2 , 2)
z = d1
J1 Layer 1 M1
1 1 , 1)
z=0
Jg Ground Layer Mg x
g , g)
Fig. 1 – Geometry of the planar structure with N+2 layers (lateral view).
This electromagnetic modeling for multilayer structures is completely described in

(Sant’Anna, 2009) and it can be synthetically summarized in the flowchart illustrated on
Figure 2. In this figure it is shown step by step the entire approach to obtain the far
electromagnetic fields scattered by the multilayer structure. Starting from the wave equation
in each layer, the methodological flowchart highlights the Green’s functions acquirement
and the integral equations formulation, which are solved applying the method of moments.
Double Fourier Differential Boundary Linear Sistem

Wave Equation Transform Equation Condition 4N+4 equations
(in each layer)
(kx,ky,z) =  {E(x,y,z)} System (at each interface) 4N+4 unknowns
Spectral Domain
Green’s Functions
Spectral Fields Method (kx,ky,z)
Linear Expansion Integral
in any Point of of
(Basis Functions) Equations
the Structure Moments
Structure
Excitation Field
Inverse Double Far Electromagnetic

Fourier Transform Stationary Phase Fields Scattered
Method by the Structure
E(x,y,z) = -1 {(kx,ky,z)}
Fig. 2 – Methodological flowchart.
2.1 Electromagnetic Fields in the Structure

The electromagnetic fields in a multilayer structure can be determined through the
methodological approach described in (Lacava et al., 2002). According to this methodology,
which employs the spectral domain full-wave technique, the structure is treated as a
boundary value problem, where the surface current densities induced on the printed
elements (electric or magnetic) are the virtual sources of the scattered fields. Since the layers
of the structure are free of sources, assuming time dependence of the form e it , the wave
equations for the n-th layer is written as
 2 E n (x , y , z)  kn2 E n (x , y , z)  0 , (1)
 2 H n ( x , y , z)  k n2 H n (x , y , z)  0 , (2)
where for free space n = 0 and for ground layer n = g, k n2   2  n n gives the wave number
in the n-th layer,  is the angular frequency, and the vectors En(x, y, z) and Hn(x, y, z) denote
the complex electric and magnetic fields, respectively (bold face letters represent vectors).
The wave equations can be solved in the spectral domain using the double Fourier
transform. In this work, the Fourier transform pair is defined as

i ( kx x  k y y )
F ( k x , k y , z)    F ( x , y , z) e dx dy , (3)


1 i ( kx x  k y y )
F ( x , y , z) 
4 2 F (kx , k y , z) e dk x dk y , (4)

where the function F(x, y, z) represents the fields En(x, y, z) and Hn(x, y, z). Applying the
double Fourier transform to (1) and (2) yields a differential equation system whose solution,
in terms of the field components, is given by
E n ( k x , k y , z)   n ( k x , k y ) e i n z , (5)

i  n z
H n ( kx , ky , z)  n ( kx , ky ) e , (6)
with
 n  (1) kn2  ( k x2  k y2 ) , Im ( n )  0 (7)
where  n ( k x , k y ) and n ( k x , k y ) are the amplitudes of the transformed field
components, kx and ky are the spectral variables,  n is the -th propagation constant in the
n-th layer,  = x, y or z, and Im( ) means the imaginary-part of a complex function. The
variable , which defines the wave propagation direction, can assume values 1 or 2. Only the
former value, representing propagation in the positive-z direction, occurs in the upper layer
(free space). In this case the wave propagation constant will be denoted by  0 . For the
ground layer, on the other hand,  is equal to 2, i.e., a wave propagating in the negative-z
direction and the propagation constant is represented by  g . For the inner layers, however,
both values of  will occur, resulting in  n 1   n 2 and the wave propagation constant is
denoted by  n .
Interesting relations among the amplitudes of the transformed fields are derived by
introducing the inverse Fourier transform of (5) and (6) in the Maxwell’s curl equations.
Such that the amplitudes of the transversal components (x and y directions) are written as
functions of the amplitude of the longitudinal ones (z direction). By enforcing the boundary
conditions for the electromagnetic fields at each interface a set of 4N+4 equations with equal
number of unknowns is obtained. The analytical solution of this system leads to the spectral
domain Green’s functions. These functions, jointly with the transformed surface current
densities, allow the determination of the transformed fields at any point of the multilayer
structure. The transformed electromagnetic field components are expressed by
E n ( k x , k y ,z)  Gx ( k x , k y , z) jx ( k x , k y )   Gy ( k x , k y , z) jy ( k x , k y ) 

(n) (n)
 
Q (n)x (kx , k y , z) mx (k x , k y )  Q (n)y (k x , k y , z) my ( kx , k y ) , (8)

 
H n ( k x , k y ,z)   P x ( k x , k y , z) jx ( k x , k y )   Py ( k x , k y , z) jy ( k x , k y ) 

(n) (n)
 
 (n)
R x ( k x , k y , z) mx ( k x , k y )   R 
(n)
y ( k x , k y , z) my ( k x , k y ) , (9)
 
(n) (n)
where G ( kx , k y , z) and Q  ( k x , k y , z) represent, respectively, the spectral domain Green’s
functions of electric type in the n-th layer, which relate the  ( = x or y) components of the
electric field to the transformed surface current densities jς(kx, ky) and mς(kx, ky) located at
(n) (n)
the interface d, with   {g, 1, 2, …, N}, and P ( k x , k y , z) and R  ( k x , k y , z) are the
respective spectral domain Green’s functions of magnetic type. Note that dg = 0 for the
structure defined in Figure 1.
In order to demonstrate the mathematical aspect of the spectral domain Green’s functions, it
is shown in equations (10) and (11) examples of electric and magnetic function types,
respectively, for free-space region. These equations are obtained when the structure is
composed by four layers (N = 2).
2
4 2  k x2  0  2 k y  0  2  i 0 z
G x( 02)x ( k x , k y , z)     e , (10)
( k x2  k y2 )   1 2 
2
4 2  k y  0  5 k x2  0  5  i 0 z
R x( 02)x ( k x , k y , z)      e , (11)
( k x2  k y2 )   1 2 
where
1   4 3 e  i 0 {1 2 ( 2 0 cos 2  i 0 2 sin 2 )( g  1cos1  i1 g sin1 ) 
 2 1 ( 0 2 cos 2  i 2 0 sin 2 )(1 g cos1  i g  1sin1 )} , (12)
3  i 0
 2  4 e {1 2 ( 2 0 cos 2  i 0 2 sin 2 )(  g  1cos1  i 1 g sin1 ) 
2 1 ( 0 2 cos 2  i 2 0 sin 2 )( 1 g cos1  i  g  1sin1 )} , (13)
 2   2{1 2 (1 g sin1  i  g  1cos1 ) sin 2 
 2 1 (1 g cos1  i g  1sin1 ) cos 2 } , (14)
5   2 {1 2 ( g  1cos1  i1 g sin1 ) cos 2 
 2 1 (  g  1sin1  i1 g cos1 ) sin 2 } , (15)
 2   2 {1 2 (  g  1cos1  i 1 g sin1 ) cos 2 
2 1 (  g  1sin1  i 1 g cos1 ) sin 2 } , (16)
 5   2 {2 1 (1 g cos1  i  g  1sin1 ) cos 2 
1 2 (1 g sin1  i  g  1cos1 ) sin 2 } , (17)
with 0 = 0 d2, 1 = 1 d1, 2 = 2 (d2−d1), d1 = ℓ1 and d2 = ℓ1 + ℓ2. It can be noted that the
Green’s functions carry information of electromagnetic parameters from all structure layers.
2.2 Method of Moments (MoM)

Once the Green’s functions are derived, the next step is to set up integral equations
constrained to the required boundary conditions. The integral equations arise as a combination
of the Green’s functions and the structure excitation fields, as can be seen from the dashed box
in Figure 2. The boundary conditions are related to the total electric and magnetic tangential
fields to each respective scattering element. The subsequent development will be described
only when electric elements are present in the structure. However, it could be also applied to
magnetic elements, taking into account its specific nature. Therefore, the integral
equation is a statement of the boundary condition requiring that the total electric field
tangential to the each of the electric element is zero (Newman & Forrai, 1987). That is,
z  [ E i (x , y , d )  E r (x , y , d )]  z  E s (x , y , d ) , on S, (18)
where S is the support region of the electric element, Es(x, y, d) denotes the scattered
electric field excited by the current on S, Ei(x, y, d) stands for the incident electric field and
Er(x, y, d) identifies the field that are reflected by the multilayer structure in the absence of
electric elements. The incident and reflected fields define the excitation mechanism of the
structure, which in this analysis is due to an elliptically polarized plane wave at an arbitrary
incidence angle.
The currents induced on the electric elements by these fields are unknown. To solve the
integral equations (18), with the unknown surface currents, the method of moments (MoM)
is applied. This method is one the most popular numerical techniques used to analyze the
radiation and scattering from complex structures. In the MoM, first the surface current is
linearly expanded in a set of basis functions with unknown coefficients
ML N L M L N L
j ( k x , k y )  x   I mx n j m ( k x , k y )  y   I my n j n ( k x , k y ) , (19)
m 1 n1 m 1 n 1
where ML and NL control the expansion modes in x and y directions on each interface layer,

respectively, I m n are the complex coefficients in the ς direction (ς = x or y) that need to be
determined, and jm(kx, ky) and jn(kx, ky) are the Fourier transform of the surface current density
components, which are defined only over the electric element. Applying the Galerkin
technique (whereby the test functions are chosen to be identical to the basis ones) the integral
equation is reduced to a system of simultaneous linear equations, which can be compactly
written in matrix form as [V] = [Z][I]. In this notation [V], [Z] and [I] denote, respectively, the
excitation matrix, the impedance matrix and the coefficient matrix. For example, the integral
equation referring to the scattered fields at the interface z = d can be written as
N M L N L  
4 2  E0s x (x , y , d ) J pn (x , y ) dx dy     I mx n   G xx j m j pn dkx dk y 
(0) 
S   0 m 1 n 1 - -
N M L N L   
 I my n   G xy

(0) 
j n j pn dk x dk y , (20)
 0 m1 n1 - -
N ML N L  
4 2  E0s y (x , y , d ) J qn

(x , y ) dx dy     I mx n   G yx j m j qn dkx dk y 
(0) 
S  0 m1 n1 - -
N M L N L   
 I my n   G yy

(0) 
j n j qn dk x dk y , (21)
 0 m  1 n  1 - -
where the left side of (20) and (21) defines the [V] matrix, the double integrals are related to
the [Z] matrix and the variables p and q range equally to the variables m and n, respectively.
(0)
It is important to point out that for notation simplicity the functions G ( k x , k y , dn ) ,

jm(kx, ky) and j pn ( k x , k y ) were written without their respective variables in equations (20)
and (21).
As abovementioned, following the same procedure it is possible to get similar expressions
when magnetic elements are used. Nevertheless, the scattered electric field, the transformed
(0)
surface electric current density and the Green´s function G should be changed by the
scattered magnetic field, the transformed surface magnetic current density and the Green´s
(0)
function R  , respectively.
The double integrations in equation (20) and (21) must be performed numerically, usually in
a very inefficient and time-consuming way. In order to improve the computation efficiency
some mathematical simplifications are employed. These simplifications include the
evaluation of the even and the odd properties of the Green’s functions, the change of the
coordinate system (rectangular to polar: kx = β cos and ky = β sin) and the asymptotic
extraction technique.
2.3 Asymptotic Expressions

In microwave remote sensing applications the major interest consist on the target radiated
power and it’s carried information. Therefore, is necessary the determination of the far
electromagnetic fields scattered by the target. These fields scattered by the multilayer
structure are computed based on asymptotic expressions, which are derived from the
stationary phase method (Collin & Zucker, 1969). The electric far-field, using this method, is
given by
E 0 (r ,  ,  )  
i k0 e  ik0r
2 r

cot ˆ 0 z ( k xe , k ye )  ˆ  0 0 z ( k xe , k ye ) ,  (22)
in a spherical coordinate system, where the intrinsic impedance of free space is represented
by 0, kxe = k0 sin cos and kye = k0 sin sin are the stationary phase points, k0 is the wave
number of the excitation wave, 0z(kxe, kye) and 0z(kxe, kye) are the amplitudes of the
transformed far-field components (in the z direction) in free space and r characterizes the
distance between the receiving antenna and the target. Note that the far magnetic field
H0(r, , ) might be obtained from E0(r, , ). For instance, considering a particular structure
consisting of four layers (N = 2) and having electric and magnetic elements printed on each
interface, the 0z(kxe, kye) and 0z(kxe, kye) are expressed by
4 2
0 z ( k xe , k ye ) 
1

( k xe  0 ) j gx  ( k ye  0 ) j gy  ( k xe 1 ) j1x  ( k ye 1 ) j 1 y 
( k xe  2 ) j 2 x  ( k ye  2 ) j 2 y  ( k ye  3 ) m gx  ( k xe  3 ) m gy 
( k ye  4 ) m1x  ( k xe  4 ) m1 y  ( k ye  5 ) m2 x  ( k xe  5 ) m2 y ,  (23)
2
0 z ( k xe , k ye ) 
4
2
(k ye  0 ) j gx  ( k xe  0 ) j gy  ( k ye 1 ) j 1x  ( k xe 1 ) j1 y 
( k ye  2 ) j 2 x  ( k xe  2 ) j 2 y  ( k xe  3 ) m gx  ( k ye  3 ) m gy 
( k xe  4 ) m1x  ( k ye  4 ) m1 y  ( k xe  5 ) m2 x  ( k ye  5 ) m2 y ,  (24)
with
 0   1 2  1 2  g , (25)
 1   2  1 2 ( 1 g cos  1  i g  1sin 1 ) , (26)
 3   1 2  g  1 2 , (27)
 4   1 2 2 ( g 1 cos  1  i 1 g sin 1 ) , (28)
 0   1  2  g 1 2 , (29)
 1   1  2 2 (  g 1 cos  1  i 1 g sin 1 ) , (30)
 3   1  2 1 2 g , (31)
 4   2 1  2 (  1 g cos  1  i g  1sin 1 ) , (32)
where the terms  1 ,  2 ,  2 ,  5 ,  2 and  5 are given, respectively, by equations (12) to

(17). It is important to observe that all terms of equations (23) and (24) are obtained on
stationary phase points and the transformed surface current densities were written without
their respective variables.
2.4 Simple Target

In order to complete the electromagnetic modeling development, simple target embedded in
a multilayer structure consisting of four layers is considered. The simple targets are
represented by electric and magnetic rectangular dipoles having infinitesimal thickness,
oriented along the x direction and printed on interface z = d2. The dipoles size 2a and 2b
(a >> b) in the x and the y directions, respectively. In this particular situation, the surface
current density along the y direction is neglected since the dipoles width is considered to be
 
very thin. Thus, for either dipole only the Z px22mx 22 matrix needs to be evaluated, being
represented by [Zpm], where the variable  defines the type of dipole, electric ( = e) or
magnetic ( = m). This evaluation involves the Green’s functions G x(02)x (electric dipole) or
R x(02x) (magnetic dipole). After the aforementioned mathematical simplifications (subsection
2.2), the [Zpm] matrix becomes
  
Z    41   A
pm 2
(0) 
x 2 x ( k x , k y , d 2 )d m 2 ( k x , k y )d p 2 ( k x , k y ) dk x dk y 
- -
1 
  /2
 2 0  [ A1 (cos )

2
   A2 (sin ) 2 ] pm  d d 
 0
  / 2
i A3 
  
(cos ) 2 pm  d d  ,

(34)
  0  0
where A x(02x) represents the spectral Green’s function and dm2 means the transformed surface
current density. The spectral Green’s function will be defined by G x(02x) or by R x(02x) in the electric
and magnetic dipole cases, respectively. Similarly, the transformed surface current density will
be described by jm2 or mm2. In each dipole case, the factors A1, A2, A3 and pm are expressed by
4 2  0  2 i A3e
A1e   , (35)
1 
4 3  0  2
A2e  , (36)
2
2
A3e  , (37)
0  2
1   0  5 
A1m    i A3m  , (38)
  2 
 0  5
A2 m  , (39)
1
2
A3m  , (40)
0   2
pm e   2 b 2 J 02 (bk y ) sinc 4 ( k x x / 2 )cos[ k x (x m  x p )] , (41)
with pmm = pme.

Additionally the equation (41) is obtained from the modeling of the surface electric current
density as the summation of piecewise-linear subdomain basis functions (rooftop functions)
taking into account the edge condition. In this equation sinc() = sin()/, J0(.) stands for the
zero-order Bessel function of the first kind and ∆x = 2a/(M+1), where M is the number of
current expansion modes.
From (34) to (41) it is noted that the first double integral of [Zpm] is dependent on the operating
frequency, whereas the second one is not. The former is the well-known Sommerfeld integrals,
which exhibit singularities in the form of branch points and poles; as such, their computation
requires careful attention. The poles (generally complex) correspond to surface and leaky
waves that can be excited in the layers. According to (Marin et al., 1990) the number of surface
wave poles and their locations depends on the thickness of the layers, their electromagnetic
parameters and the wave number. For a multilayer structure and depending on the thickness
of the layers, the Green’s functions might present hundreds of poles, making their integration
a formidable task. Since the number of poles and their locations are not known beforehand
(Newman & Forrai, 1987), the use of a deformed path to compute the integrations seems to be
an efficient way to avoid this problem. Therefore, in this work a parabolic path was chosen to
compute the integrations.
Using the rooftop function as basis function, the excitation matrices [Vpe] and [Vpm], for this
particular structure, are defined only over the S2 region, which is limited by –a and +a in x-
direction and by –b and +b in y-direction being given by
V    F (x , y , d ) D
p 2
 
p 2 x ( x ) D p 2 y ( y ) dx dy 
S2
b
cos( 2 y )

i 1 x p
 2 ( B sini  C cosi ) sinc 2 (1x / 2) e dy , (42)
0 1  ( y / b) 2
where variable  (e or m) defines the type of dipole as defined in equation (34). Therefore,
the function F(x, y, d2) can represent the scattered field E0s x (x , y , d2 ) or H 0s x (x , y , d2 ) and the
complex conjugate of the surface current density D p 2 will stand for J p 2 or M p 2 , with
 = x or y. The constants B, C, 1 and 2 are given by
Be  (Ehi e i  Ehr e  i ) , (43)

Ce   cos  i (Evi e i  Evr e  i ) , (44)
Bm   01 (Evi e i  Evr e  i ) , (45)
Cm   01 cos i (Ehi e i  Ehr e  i ) , (46)
1  k 0 sin i cosi , (47)
2  k0 sin i sini , (48)
with  = k0 d2 cosi, E pi and Epr standing for incident and reflected electric fields having
linear p polarization (horizontal or vertical) and the angles i and i define the direction of
the incident wave. In equations (43) to (48) the subscript v and h stand for vertical and
horizontal linear polarization. The solution of integral equation leads to coefficient matrix
[I], from which it is possible to obtain the surface current density expression over the
scattering element. This density jointly with the spectral Green’s function permits the
evaluation of the electromagnetic fields in spectral domain.
3. Remote Sensing Applications

This section is devoted to some remote sensing applications, which can be derived from the
electromagnetic modeling described in section 2. All results are referred to a simple target
(electric and/or magnetic dipoles) embedded in a four layers structure (N = 2).
3.1 Scattering Matrix

The determination of analytical expressions for the dipoles scattering matrices is the first
application considered in this section. The scattering matrix is the most important parameter
in polarimetric SAR analysis, since it provides complete information about the scattering
mechanism. All polarimetric features that describe the target scattering can be derived from
it. The scattering matrix can be seen as a mathematical characterization of the target
scattering. This matrix linearly relates the electric fields of the wave scattered (Es) by a target
to the ones of the incident wave (Ei). As consequence the scattering matrix can be seen as a
linear operator between Es and Ei, given by
s
S vv S vh  Evi  e  ik0 r
E   EE   e r  
 ik 0 r
s v
S   S  E i , (49)

s
h  hv S hh  Ehi  r
where the subscript v and h stand for vertical and horizontal linear polarization. In order to
determine the scattering matrix of electric and magnetic dipoles embedded in a multilayer
structure, as illustrated in Figure 3, the procedure defined in (Sant’Anna et al., 2007a) will be
followed. This procedure directly relates the elements of the scattering matrix to the
components of the scattered electric field; a standard spherical coordinate system (r, , ) is
chosen to coincide with a coordinate system (k, v, h) defined in terms of horizontal and
vertical linear polarization components, as stated in (Ulaby & Elachi, 1990).
Free Space
(0, 0)
z = d2
Layer 2
ℓ2
(2, 2)
J1 M1
z = d1 z
ℓ1 Layer 1
(1, 1)
z=0
y x
Ground Layer
(g, g)
Fig. 3 – A lateral view of the multilayer structure under analysis.

The incident wave sets up currents on the scattering elements (dipoles), which in turn re-
radiates a scattered wave. The currents induced on the dipoles depend on the incident wave
polarization. Therefore, from the reciprocity theorem, relation (50) between the components
of the transformed far electric field in free space due to each linear polarization
(0hz) ( k xe , k ye )   0 (0vz) ( k xe , k ye ) , (50)
can be obtained, where the superscripts (v, h) are associated to polarization of the incident
wave. Consequently, for an electric planar dipole the induced currents due to horizontal and
vertical linear polarizations are, then, related by
 k ye  ( v )
j 1( hxe) ( k xe , k ye )       j 1xe ( k xe , k ye ) . (51)
 k xe 
This relation leads to the following scattering matrix
 ( cos s )2 /   sin s cos s 

Sed    , (52)
cos s  sin s cos s   (sin s )2 
where the constants  and  are given by
 1 1 i 2  2 k02 0  1
   0 and    cos s j1( vxe) ( k xe , k ye ) .
 2 1  2
Similarly for the magnetic dipole, the relation between the induced currents and its
corresponding scattering matrix are expressed, respectively, by
k  ( v)
m 1( hxe) ( k xe , k ye )      xe  m 1xe ( k xe , k ye ) , (53)
 k ye 
 
  (sin s ) 2 /    sin s coss 

Smd    , (54)
sin s  sin s coss   ( coss )2 
where the constants  and  are represented by
 1 4 i 2 2 k02 0  4
   0 and    cos s m 1( vxe) ( k xe , k ye ) .
24  2
Note that the constants  and , which are directly related to the currents induced on the
dipoles, affect equally all the elements of the scattering matrices. The terms  1 ,  2 ,  1 ,
 4 ,  1 and  4 , given by equations (12), (13), (26), (28), (30) and (32) respectively, are
evaluated at stationary phase points. From equations (52) and (54) it can be also observed
that the scattering matrix expressions are derived for any direction of the scattered wave.
In (Sant’Anna et al., 2008a) it was used two polarimetric features as comparative metrics
between the electric and magnetic scattering matrices. These features included the -angle
derived from the eigenvalue/eigenvector decomposition (Cloude & Pottier, 1997) and a
similarity measurement based on square of correlation coefficient, which was proposed in
(Yang et al., 2001). The results were obtained in (Sant’Anna et al., 2008a) for a structure
composed of layers having the same electromagnetic parameters of free space. They showed
that even though the dipoles present different electromagnetic nature the dominant -angle
for both are equal to 45°. Therefore, based on this feature it was not possible to differentiate
the two dipoles, despite their electromagnetic characteristics. On the other hand, measuring
the correlation coefficient it was shown that it can be used as a discriminating feature
between the two scattering matrices. This statement is promptly verified when the four
layers are assumed to be the free space. In this particular case the constants  and  are
reciprocal real values yielding a correlation coefficient equal to zero, i.e. the scattering
matrices are non-similar.
3.2 Simple Target Characterization

The simple target characterization is the second application considered in this work within
remote sensing context. The characterization is made using several electromagnetic
attributes that are measured at far-field condition. The measurements are done based on a
particular configuration of the four layers structure, in which the dielectric parameters of
layers 1 and 2 are made equals. That is, the structure is then composed by three dielectric
layers (the free-space region, one inner layer and the ground layer) and the dipoles (electric
and magnetic) are printed on interface z = d2, as depicted in Figure 4.
In antenna theory the directivity function (, ) is one of several parameters used to
characterize or define an antenna. According to (Balanis, 1997) this function is defined as the
ratio of the radiation intensity (, ) in a given direction from the antenna to the radiated
intensity average over all direction. The (, ) is expressed by
4  ( ,  )
( ,  )  , (55)
i
where  i   ( ,  ) dΩ is the average radiated power within the solid angle  and the

(, ) function is defined, for a multilayer structure, as
2
1  k0cot   2 2
( ,  )     0 z ( kxe , kye )  02 0 z ( kxe , kye )  , (56)
20  2   
 
with the transformed electric and magnetic longitudinal amplitudes the 0z(kxe, kye) and
0z(kxe, kye) being expressed by equations (23) and (24), respectively.
Considering the structure of Figure 4 and assuming that the size of dipoles are infinitesimal
the function (, ) is promptly computed. Since, in this case, the surface current density
can be considered as constant. The directivity function for a structure characterized by r = 2,
tg = 0, rg = 1 and tgg = 1.010+15 as dielectric parameters values, it was evaluated at
2.25 GHz for printed infinitesimal electric and magnetic dipoles. The respective three-
dimensional graphic of (, ) are illustrated in Figures 5 and 6 for two values of inner layer
thickness, ℓ = 4 mm (0.030) and ℓ = 40 mm (0.30).
Free Space
(0, 0)
J2 M2
z = d2
Layer 2
(1, 0)
ℓ
z = d1 z
Layer 1
(1, 0)
z=0
y x
Ground Layer
(g, 0)
Fig. 4 – A lateral view of the multilayer structure under analysis.

From these figures it can be notice that the radiation patterns are completely different for
both dipoles. This fact it was expected due to the dissimilar electromagnetic feature of the
dipoles. Theoretically, the radiation pattern can be used to differentiate structures which
present these kinds of dipoles. In addition, it can be observed that the shapes as well as the
radiation levels are modified by the variation of the inner layer thickness. The radiation
levels modification may be associated to the modes (guided and leaky) existing within the
structure. Increasing the layer thickness more guided modes might be established inside the
structure, consequently decreasing its scattered power.
(a) (b)
Fig. 5 – Three-dimensional graphic of (, ) for infinitesimal electric dipole: (a) ℓ = 4 mm
and (b) ℓ = 40 mm.
(a) (b)
Fig. 6 – Three-dimensional graphic of (, ) for infinitesimal magnetic dipole: (a) ℓ = 4 mm
and (b) ℓ = 40 mm.
The SCS is other parameter that can be used on the characterization of target scattering
properties. In general, it is function of the operating frequency, wave polarization, incident
and observation angles, target geometry and its dielectric properties. According to
(IEEE, 1993) the target SCS can be expressed by
2
 E s (ks ) 

SCS( k s , k i )  lim 4 r 2 , (57)
r   i 2 
 E (ki ) 

where Ei(ki) is the electric field of incidence plane wave in ki direction and Es(ks) is the
electric field of the scattered wave in ks direction, measured at a distance r of the target.
The multilayer structure (Figure 4) containing dipoles that their length and with are equal to
50 mm and 1 mm, respectively, is used to evaluate the SCS parameter. The dipole length is
oriented along the x-direction while its width is oriented along the y-direction. For SCS
analysis a basis structure characterized by r = 2.33, tg = 1.210-1, ℓ = 50 cm, rg = 5 and
tgg = 2.010-1 is used. The evaluation, conducted at 1.25 GHz, uses a vertically polarized
incident wave at oblique direction defined by i = 40° and i = 60°. The SCS analysis comprises
the influence of the inner layer by varying its thickness (ℓ) and its relative permittivity (r). The
SCS resulting curves are plotted in azimuth (θs fixed at θi and s ranging from −180° to +180°)
and in elevation (θs ranging from −90° to +90° and s fixed at i) directions. It is important to
mention that for each analysis the others structure parameters were kept constant.
The plots of azimuth SCS for the ℓ variation are depicted in Figure 7. It can be observed that
the SCS results for the structure containing the electric dipole are greater than those for the
magnetic one. An inversion of the greatest and the lowest SCS curves, relative to both
structures, is also observed in Figure 7. That is, for instance, the ℓ = 0.20 leads to the greatest
SCS values for the structure having the electric dipole, however conduct to the lowest SCS
values for the structure with the magnetic dipole. From Figure 7b it can be seen that the
greatest RCS level occurs when ℓ = 1.0λ0 and the ℓ values 0.1λ0, 0.2λ0, and 0.5λ0 present the
lowest RCS levels, where the scattering of the dipole seems to be mandatory. The amplitude
of the SCS variation (it is about 4 dB) is similar for both structures.
(a) (b)
Fig. 7 – The azimuth SCS curves for different ℓ values: (a) electric dipole and (b) magnetic dipole.
The effects of the inner layer relative permittivity (r) on the SCS curves are shown for
structure with electric dipole in Figure 8 and for structure that has the magnetic dipole in
Figure 9. In both figures are presented the SCS in azimuth and elevation directions. In
general, similarly to ℓ variation, the SCS values for the structure having the electric dipole
are greater than those for the magnetic dipole. For the structure with the magnetic dipole is
noted that there is not alteration in the curves shape, only in their SCS levels. These SCS
levels increase when the r grows. However, the SCS curves shape as well as their
magnitudes for the structure having the electric dipole are modified by the inner layer r.
From the azimuth SCS curves (Figures 8a and 9a) it can be seen that, for r values greater
than 4, the curves seem to show a kind of complementary. Since, in the i angles where
occur the maximum SCS levels for structure containing the electric dipole are found the
minimum SCS level for the other structure and vice-versa. The SCS is sensitive to
electromagnetic properties of the scattering element since; in general, both dipoles bring
different responses.
(a) (b)
Fig. 8 – The electric dipole SCS curves for different of r values: (a) azimuth and (b) elevation.
(a) (b)
Fig. 9 – The magnetic dipole SCS curves for different of r values: (a) azimuth and (b) elevation.
Several parameters have been proposed to assist the interpretation and the classification of
polarimetric SAR data. For instance, the entropy, the anisotropy and –angle derived from
Cloude-Pottier’s target decomposition theorem (Cloude & Pottier, 1997) has been widely
used for this purpose. The –angle can be also used as characterization parameter since it
provides information about the type of scattering mechanism. In this way, the –angle was
evaluated at 1.25 GHz for the same multilayer structure used on the SCS analysis. In Figure 10
is showed the –angle variation in terms of the inner layer relative permittivity (r).
From the evaluation of –angle it is verified that the average –angle varies completely
different from that expected with –theoretical (around 45°). This fact it is observed for both
structures. It probably occurs due to the existence of cross components in the scattering
matrix of the structures. In contrast, the computed –dominant values are almost constant
and very close to 45° and to 60° for the structure containing the electric dipole and magnetic
one, respectively. Therefore, according to (Cloude & Pottier, 1997), the scattering mechanism
of the structure having the magnetic dipole can not be assumed as a scattering mechanism
produced by an electric dipole. It leads to state that the layers introduce a certain degree of
anisotropy to scattering mechanism of the magnetic dipole since its –dominant value is
characteristic of anisotropic scatter.
(a) (b)
Fig. 10 – The –angle variation in terms of r for structure with: (a) electric and (b) magnetic dipoles.
According to (van Zyl et al., 1987) the polarimetric response is another way to represent the
target SCS. In this graphical representation, the SCS is plotted in function of ellipticity and
orientation angles of the transmitted electromagnetic wave. The polarimetric responses for
the multilayer structures containing the electric and the magnetic dipoles when r = 25 are
presented, respectively, in Figures 11 and 12. The polarimetric response is sensitive to
electromagnetic properties of the scattering element since both dipoles bring different
responses. Besides, it can be observed that the two polarimetric responses seem to be
complementary, because the point where occurs the maximum value of polarimetric
response for one dipole corresponds to the point of minimum value for the other dipole.
This complementary relation can be though as the duality that there exist for both dipoles.
The magnitude of the polarimetric response for electric dipole is greater than for the
magnetic one for this value of r, as can be sustained by Figures 8 and 9.
(a) (b)
Fig. 11 – Polarimetric response of electric dipole: (a) co-polarized and (b) cross polarized.
(a) (b)
Fig. 12 – Polarimetric response of magnetic dipole: (a) co-polarized and (b) cross polarized.
3.3 Polarimetric SAR Images Simulation

The polarimetric SAR data simulation is the third remote sensing application that is
considered in this work, which can be derived from the current electromagnetic modeling.
The methodological approach employed on the polarimetric SAR image simulation is
completely described in (Sant’Anna et al., 2008b and Sant’Anna, 2009). It is based on the
computation of the far electric field scattered by a multilayer structure excited by an elliptically
polarized plane wave, in order to perform the simulation process as realistic as possible.
The multilayer structure used to simulate two sets of polarimetric SAR images is composed
of three isotropic, linear and homogenous layers stacked up in z direction, as depicted in
Figure 4. A rectangular perfect electric dipole, sizing 50  1 mm, is printed on the planar
interface between the free space and the inner layer (interface z = ℓ). This dipole has
infinitesimal thickness and will act as scattering element. A large number of dipoles is
spread over the inner layer in such way that is guaranteed at least 30 elementary scatters
(dipoles) inside of each SAR resolution cell and there is no dominant scatter in the scene.
The dipoles position and local orientation are conveniently (uniformly) distributed over the
inner layer.
The simulated images represent the acquisition of an airborne monostatic SAR flying at
6000 m of height, having a grazing angle of 35º and imaging a terrain where only a small
area of 290  290 m will be consider in the simulation. The simulated images were generated
having 3.0 m of spatial resolution in range and azimuth directions and 2.8 m of pixel spacing
in both directions. The SAR spread point function was represented by a separable two-
dimensional sinc function, whose support region is a square window of 99 pixels.
The first image set, illustrated in Figure 13, was generated for a SAR operating at L-band
(1.25 GHz) and for a multilayer structure having the following characteristics: the ground
layer with relative permittivity rg = 5.0 and loss tangent tang = 2.010-1, the inner layer
characterized by thickness ℓ = 527.63 mm, relative permittivity r = 2.33 and loss tangent
tan = 1.210-4 and both layers having 0 as magnetic permeability. In this set, the images
contain four distinct regions that are different by local orientation of the electric dipoles. The
dipoles are preferentially oriented at 10º, 30º and 20º, respectively, for upper left, upper right
and lower right regions, while for lower left region the dipoles are oriented totally random
(TR). These regions are denoted by A, B, C and D letters, respectively. It is important to
mention that the orientation is referred to the azimuth-axis (az) and the origin of the system
coordinate is located at image center. Note that the VH channel is not presented in Figure 13
because it is equal to HV one, due to reciprocity assumption framework.
A B
D C
rg
az
(a) (b)
(c) (d)
Fig. 13 – L-band amplitude simulated polarimetric SAR images: (a) Phantom, (b) HH, (c)
HV and (d) VV channels.
These simulated SAR images were quantitatively evaluated in (Sant’Anna et al., 2008c). The
amplitude SAR data were investigated by fitting a simple linear regression model to sixteen
samples of 100100 pixels extracted from each region. The polarimetric analysis was based
on the discriminatory capability of the  and  angles derived from the standard Cloude-
Pottier’s decomposition theorem (Cloude & Pottier, 1997) and the orientation angle ()
induced by azimuthal terrain slope (Lee et al., 2000).
Under linear detection and for single look data the ratio of the standard deviation and the
expected value (the coefficient of variation  Cv) over homogeneous area is constant and
equal to [(4-)/]1/2 = 0.5227. This value can be obtained from the moments of the Rayleigh
distribution. As consequence, a simple way of evaluating the simulated data is to check
whether the mean value () and the standard deviation () of the data holds, within
homogeneous area, the linear relationship  = 0.5227*. This analysis was performed by
applying a simple linear regression model to the estimated means and standard deviations
for all sixty four samples in each channel.
The adjusted linear model (Y = b0+b1*X) and expected one (Y = 0.5227*X) are depicted on
Figure 14 in each channel, respectively, by dashed and solid lines. From the adjusted model,
where Y represents the standard deviations and X the mean value, is expected that the
estimated intercept (b0) and slope (b1), should be zero and 0.5227, respectively. It can be
noted that two straight lines are almost coincident, showing that the adjusted model and
theoretical one are approximately equal.
A Student’s t-test for the intercept being equal to zero and the slope being equal to 0.5227 was
performed. These two parameters and their p-values for statistical analysis of the regression
are shown in Table 1. The estimated slope values are around the expected value (0.5227) for all
channels. The estimated intercept values for HH and VV channels are very close to 0, and for
HV channel is a little bit higher than this value. Analyzing the p-values is observed that only
one value lower than 5% is encountered for b1 test in VV channel, that is, for this case there is
no evidence to accept the hypothesis of b1 to be equal to 0.5227 at 95% of significance level.
(a) (b) (c)

Fig. 14 – Linear fit to sample data: (a) HH, (b) HV and (c) VV channels.
p-value*100%
Channel b0 b1
b0 b1
HH 0.0971 0.5070 47.34 94.24
HV 0.2754 0.5030 70.98 75.12
VV -0.0880 0.5245 6.74 4.40
Table 1. Statistics for the linear regression fit.
From the linear fit results it can be concluded that the polarimetric SAR data has one look
and the regions can be considered as homogeneous areas. Additionally, the equivalent
number of looks (ENL) was estimated for the whole image based on all the sixty-four
samples. The estimated ENL for the HH, HV and VV channels are 1.005, 0.994 and 0.998,
respectively, which results in 0.999 as final estimate of ENL computed as mean of those
individual estimates. The estimated values are very close to one, as expected for a single
look data.
Knowing that the -angle provides information about the scattering mechanism and the 
and  angles are related to target orientation angle, it was formed attributes bands from
pointwise estimation of these angles. The ,  and  images and their respective histograms
are shown in Figure 15.
The  image does not show a notably difference among the image regions, leading to a low
discriminatory capability based on only  feature. The -angle mean value of each region is
around 55.96°, value that is greater than 45° (typical value for electric dipole scattering
mechanism), it probably occurs due to the existence of cross components in the scattering
matrix. Comparing the results presented in (Sant’Anna et al., 2007b), where the dipole
scattering was mandatory, it is concluded that the thickness of the inner layer can influences
significantly in the scattering mechanism of the multilayer structure. Therefore, might be
state that the inner layer can introduce a certain degree of anisotropy to scattering
mechanism of the regions.
As abovementioned the difference among the regions comes from the dipole local
orientation, in this sense is expected that  and  angle images present, in general, the best
discriminatory capability. This fact is easily verified by a visual inspection of
Figures 15b and 15c, where the four image regions appear visually distinguishable. The 
and  image histograms exhibit three dominant modes that are related to the values of
regions A, B and C. However, for these regions the variances of  values are smaller than
those of  values, meaning that the former image () presents regions more homogeneous
than the later one ( image), which displays an aspect noisier than  image and this fact
influences on the regions discrimination.
In order to evaluate the discriminatory capability of  and  polarimetric features a simple
threshold classification procedure was employed. The thresholds were defined based on the
mean and one standard deviation of each region (regreg, where reg stands for region).
Over the classified images was applied a mode filter with window of 33 pixels to improve
the classification results. The final classified images that arise from  and the  images are
illustrated in Figure 16. The classification performance was assessed from the matrix
confusion (Tables 2 and 3), overall accuracy and kappa coefficient of agreement and its
sample variance. The confusion matrices were estimated based on all pixels of each region.
(a) (b) (c)

Fig. 15 – Angle images and their respective histograms: (a) , (b)  and (c) .
In general, from Figure 16 can be noticed that all regions are better classified when the 
angles feature is used, this is corroborated by the values of overall accuracy and kappa
coefficient. From a statistical z-test with 95% of significance level is observed that the
classification result of  image is much better than that obtained with  image, showing that
discriminatory power carried by  feature is superior to one gathered by  feature, for the
simulated image used in this study. It is noted that the regions border are better defined in 
image classification, probably due to their smallest class variances.
(a) (b)
Fig. 16 – Classified images arising from: (a)  and (b)  features.
Orig./Class A B C D Orig./Class A B C D
A 96.22 0.02 0.05 3.71 A 97.80 0.17 0.12 1.91
B 0.03 97.34 0.07 2.56 B 0.05 99.76 0.02 0.17
C 0.00 0.15 96.85 3.00 C 0.07 0.95 98.10 0.88
D 8.69 3.66 7.86 79.79 D 1.54 4.39 0.98 93.09
Overall Accuracy = 0.93 Overall Accuracy = 0.97
ˆ  0.90 and ˆ 2  7.46 10 6 ˆ  0.96 and ˆ 2  2.97 10 6
Table 2. Confusion matrix for  image. Table 3. Confusion matrix for  image.
The region D, where dipole local orientation angles are randomly distributed, presents the
greatest confusion in both classification results, being the worst classified region. The region
B (dipole oriented preferentially at 30°) obtained the best region discrimination. The regions
A (dipoles at 10°), B and C (dipoles at 20°) are very well differentiate by the simple
threshold method, using  and  features. The  image classification is characterized by the
largest misclassification of the region D values. The classification results point out that these
two features gather the proper information to distinguish targets whose present differences
in their scatter orientation about the radar line sight.
The variation of the soil moisture content might be symbolized by variations on the relative
permittivity and on the loss tangent of the structure layers. As consequence, simulated
images containing regions that are distinct by these layers dielectric parameters should be
used aiming for the retrieval of soil moisture from SAR data studies. Keeping this idea in
mind the second image set was generated. Images having five different regions were
simulated for a SAR operating at C-band (5.3 GHz) and X-band (9.6 GHz). The dielectric
parameters of the multilayer structure of each image region are presented in Table 4.
Using the HH and VV images were computed two polarimetric attributes derived from the
backscattering coefficient. The former was proposed in (Singh & Dubey, 2007) and is called
by polarization discrimination ratio (PDR). The later was employed in (Tadono et al., 1999)
and is denoted here as polarimetric description square root (PDS). These attributes are,
respectively, given by
0 0
 vv   hh
PDR  0 0
, (58)
 vv   hh
0 0
PDS   vv  hh , (59)
0
where  pq represents the backscattering coefficient for pq polarization.
Region Location r tanr rg tang Dipole Orientation

Upper Left 2.33 1.210-4 5.0 2.010-1 10º
Lower Right 2.33 1.210-4 5.0 2.010-1 30º
Central 2.33 1.210-4 5.0 2.010-1 TR(a)
Lower Left 4.00 1.210-1 8.0 2.010+1 TR
Upper Right 2.33 1.210-4 8.0 2.010+1 TR
(a) TR means totally random.
Table 4. Dielectric characteristic of the structure layers of each image region.
The attribute images are depicted in Figures 17 and 18 for C- and X-bands, respectively. In
these figures are also shown their respective ICM (Correia et al., 1998) classification results.
It is important to mention that it was assumed a bivariate Gaussian distribution to model
the joint density of the PDR and PDS attributes. From the classification results it can be
noted that the five image regions were completely distinguished in both bands. The overall
accuracy reached 99.75% for C-band images and 99.70% for X-band images, showing that
the PDR and PDS attributes properly combined can be used to discriminate image regions
having different dielectric characteristics.
(a) (b) (c)

Fig. 17 – C-band images: (a) PDR, (b) PDS and (c) classified.
(a) (b) (c)

Fig. 18 – X-band images: (a) PDR, (b) PDS and (c) classified.
4. Conclusions
A scattering electromagnetic model for planar multilayer structure is introduced in this
work. The methodological approach used to determine the electromagnetic fields existing in
this type of structure is concisely described. The electromagnetic model potentiality is
demonstrated through some remote sensing applications. These applications include the
establishment of analytical expression for the scattering matrix of electric and magnetic
dipoles, the simple target electromagnetic characterization and the 1-look SAR polarimetric
image simulation.
The characterization of radar scattering from complex targets in terms of a linear combination
of scattering from simpler ones, consists in an effective tool for the analysis of polarimetric
data. From this point of view, knowledge of the scattering matrix from simpler targets is
crucial to extract information gathered by polarimetric SAR data. Besides, the closer to the
exact solution the determination of the scattering matrix is, the more precise the scattering
analysis will be. From polarimetric SAR images and using the knowledge of scattering matrix
it is possible, for instance, to develop mathematical models for natural targets and to derive
several polarimetric target descriptors. It is also possible to improve the classification accuracy
of land cover and use, to identify the scattering mechanisms intrinsic on an ensemble of pixels,
and to built up a sensor calibration processes. In this context, the computation of scattering
matrix plays a central rule in the polarimetric radar remote sensing imagery.
The scattering characterization of dipoles embedded in multilayer structure was carried out
by evaluate the directivity function, the scattering cross section (SCS), the polarimetric
response and the –angle derived from Cloude-Pottier’s target decomposition theorem. The
directivity function revealed that the both infinitesimal dipoles irradiate differently, indicating
their distinct electromagnetic nature. Besides, from the directivity function it could be the
influence of the inner layer thickness on the radiated power. The SCS curves analysis
demonstrated how the inner layer thickness (ℓ) and the relative permittivity (r) of a multilayer
structure can affect the target characterization. The ℓ parameter affects only the SCS levels
while the r influences the SCS shape as well as its level. In general, the SCS shape is dissimilar
for the electric and magnetic dipoles. This statement is also verified from the graphic of
polarimetric response. The estimated average –angle for both dipoles showed great deviation
from the theoretical one. On the other hand, when the dominant –angle is used the estimative
for the electric dipole is in accordance with expected one. However, for the magnetic dipole
the estimated dominant –angle values are constant and greater than 45° (typical value for an
electric dipole), indicating a degree of anisotropy in the scattering mechanism.
The developed image simulation process is not restricted to monostatic airborne sensors, it
can be employed to orbital sensors as well as in bistatic configuration without any
additional complexity. The simulated images can be generated at several operating
frequencies, containing different spatial resolution and pixel spacing among others SAR
parameters. The generated images were evaluated according to several measurements
commonly employed in SAR data analysis. It was checked by using a simple regression linear
model whether the mean value () and the standard deviation () of the data exhibit the linear
relationship  = 0.5227 within homogeneous areas. From regression linear analysis it can be
concluded that this relationship holds for all simulated images. The equivalent number of
looks was estimated, proving that the simulated data have only one look as simulated. It was
also assessed the discriminatory capability of the image regions by applying two classification
approaches, one using a simple threshold procedure and other based on the ICM classifier.
Consequently, from the evaluation results it can be concluded that the simulation process is
working properly, since the results are in accordance with those presented in the literature.
Furthermore this simulation process can be used to improve the understanding of SAR data
properties in different situation, images can be generated to do theoretical as well as
practical studies in several SAR subjects. From these kind of simulated image can be
developed specific tools for digital image processing and to perform studies relative to SAR
data calibration. In the remote sensing context still catches a glimpse the use of the model in
ground penetrating radar (GPR) studies. Finally, from the obtained results it can be
concluded that the presented electromagnetic model is a useful and powerful tool to be
utilized in several remote sensing issues.
5. References
Balanis, C. A. (1997). Antenna theory: analysis and design. New York, John Willey & Sons Inc.
Cloude, S. R. & Pottier ,E. (1996). A review of target decomposition theorems in radar
polarimetry. IEEE Transactions on Geoscience and Remote Sensing, Vol. 34, No. 2, 498-512.
Cloude, S. R. & Pottier E. (1997). An entropy based classification scheme for land
applications of polarimetric SAR, IEEE Transactions on Geoscience and Remote
Sensing, Vol. 35, No. 1, 68–78.
Collin, R. E. & Zucker, F. J. (1969). Antenna Theory: Part 1, New York, McGraw-Hill Book
Company.
Correia, A. H.; Freitas,C. C.; Frery, A. C.; & Sant'Anna, S. J. S. (1998). A user friendly
statistical system for polarimetric SAR image classification. Revista de
Teledetección, Vol. 10, 79–93.
Freeman, A. & Durden, S. L. (1998). A three component scattering model for polarimetric
SAR data. IEEE Transactions Geoscience and Remote Sensing, Vol. 36, No. 3, 963-973.
Institute of Electrical and Electronics Engineers (1993). IEEE Std 1451993: IEEE standard
definitions of terms for antennas. New York.
Israelsson, H.; Ulander, L. M. H.; Martin, T.; Askne, J. I. H. (2000). A coherent scattering
model to determine forest backscattering in the VHF. IEEE Transactions on
Geoscience and Remote Sensing, Vol. 38, No. 1, 238–248.
Lacava, J. C. S.; Proaño De la Torre, A. V. & Cividanes, L. (2002). A dynamic model for
printed apertures in anisotropic stripline structures, IEEE Transactions on Microwave
Theory, Vol. 50, No.1, 22–26.
Lee, J. S.; Schuler, D. L. & Ainsworth, T. L. (2000). Polarimetric SAR data compensation for
terrain azimuth slope variation. IEEE Transactions on Geoscience and Remote Sensing,
Vol. 38, No. 5, 2153-2163.
Lin, C. Y. & Sarabandi, K. (1999). A Monte Carlo coherent scattering model for forest
canopies using fractal-generated trees. IEEE Transactions on Geoscience and Remote
Sensing, Vol. 37, No. 1, 440–451.
Marin, M. A.; Barkeshli, S. & Pathak, P. H. (1990). On the location of proper and improper
surface wave poles for the grounded dielectric slab. IEEE Transactions on Antennas
and Propagation, Vol. 38, No.4, 570–573.
Newman, E. H. & Forrai, D. (1987). Scattering from microstrip patch. IEEE Transactions on
Antennas and Propagation, Vol. AP–35, No.3, 245–251.
Sant’Anna, S. J. S., Lacava, J. C. S. & Fernandes, D. (2007a). Closed form expressions for scattering
matrix of simple targets in multilayer structures, in Proceedings of International Geoscience
and Remote Sensing Symposium, Barcelona, Spain, July 2007, 714–717.
Sant’Anna, S. J. S., Lacava, J. C. S. & Fernandes, D. (2007b). A useful tool to simulate polarimetric
SAR images. Proceedings of Eleventh URSI Commission F Triennial Open Symposium on
Radio Wave Propagation and Remote Sensing. Rio de Janeiro, Rio de Janeiro.
Sant’Anna, S. J. S., Lacava, J. C. S. & Fernandes, D. (2008a). Evaluation of the scattering matrix of
flat dipoles embedded in multilayer structures. PIERS Online, Vol. 4, No. 5, 536–540.
Sant’Anna, S. J. S., Lacava, J. C. S. & Fernandes, D. (2008b). From Maxwell’s equations to
polarimetric SAR images: a simulation approach. Sensor, Vol. 8, No. 11, 73807409.
Sant’Anna, S. J. S., Lacava, J. C. S. & Fernandes, D. (2008c). Analysis of simulated polarimetric
SAR images generated by a multilayer electromagnetic scattering model., Proceedings
of International Geoscience and Remote Sensing Symposium, Boston, July 2008, 1245–1248.
Sant’Anna, S. J. S. (2009). Modelagem do espalhamento eletromagnético de estruturas
multicamadas com aplicações em sensoriamento remoto por micro-ondas, PhD.
Thesis (in Portuguese), ITA, São José dos Campos.
Singh, D. & Dubey, V. (2007). Microwave bistatic polarization measurements for retrieval of
soil moisture using an incident angle approach, Journal of Geophysics and
Engineering, n. 4, 7582.
Tadono, T.; Qong, M.; Wakabayashi, H.; Shimada, M.; Kobayashi, T. & Shi, J. (1999).
Preliminary studies for estimating surface soil moisture and roughness based on a
simultaneous experiment with CRL/NASDA airbone SAR (PI–SAR), Proceedings of
Asian Conference on Remote Sensing. Hong Kong.
Ulaby, F. T. & Elachi, C. (1990). Radar polarimetry for geoscience applications. Norwood, Artech
House.
Yang, J.; Peng, Y. N. & Lin, S. M. (2001). Similarity between two scattering matrices,
Electronics Letters, Vol. 37, No. 3, 193–194.
van Zyl, J. J.; Zebker, H. A. & Elachi, C. (1987). Imaging radar polarization signatures: theory
and observation. Radio Science, Vol. 22, No. 4, 529-543.
Zurk, L. M.; Koistinen, P.; Sarvas, J. & Holmström, L. (2001). Electromagnetic scattering
model for forest remote sensing. Research Reports A38, University of Helsinki,
Helsinki.
On Performance of S-band FMCW Radar for Atmospheric Measurements 645
31
X
On Performance of S-band FMCW Radar

for Atmospheric Measurements
Turker Ince
Izmir University Of Economics
Turkey
1. Introduction
At low microwave frequencies, radars respond to spatial variations in the index of refraction of
the air often characterized by the atmospheric refractive-index structure parameter, Cn2 . While
remote measurement of the atmospheric boundary layer (ABL) is complicated by the low
radar reflectivity of clear-air turbulence structures, frequency-modulated, continuous-wave
(FMCW) radars, having tremendous sensitivity and spatial resolution compared to their
pulsed counterparts, have proven to be a solution to this problem. While Doppler capability
can be added to FMCW radars (Strauch et al., 1976), the unique strength of this technology lies
in its ability to monitor the atmospheric refractive-index structure parameter, Cn2 , with
unparalleled resolution in height and time. Since the first high-resolution atmospheric FMCW
radar, developed by Richter (1969), a number of FMCW radars have been developed for high-
resolution atmospheric probing (Chadwick et al., 1976; Eaton et al., 1995; Hirsch, 1996), which
have provided valuable information about the fine structure and dynamics of clear-air
turbulence. The advent of S-band, FMCW radars has opened a new research field, which was
reviewed by Gossard (1990).
Microwave backscattering from the clear-air ABL is generally a combination of Bragg
scattering from small-scale spatial variations in refractive index caused by turbulence and
Rayleigh scattering from insects, birds, dust, or other airborne particles whose size is much
smaller than the radar wavelength. At S-band frequencies, radars measure clear-air
backscatter from both refractive index turbulence and from insects. Based on the
experimental observations (Wilson et al., 1994; Eaton et al., 1995; Ince et al., 2003; Martin and
Shapiro, 2007), Rayleigh scattering from insects can be the dominant source of clear-air
returns in both the convective boundary layer (CBL) and the nocturnal boundary layer
(NBL). Depending on an application, radar return from airborne insects may be considered
either as source of clutter (such as contamination of weather radar measurements) or as
desirable passive tracers of the clear air (Wilson et al., 1994). Due to its very fine spatial and
temporal resolution, S-band FMCW radars are able to resolve individual insects and
discriminate between discrete particle echoes and distributed clear-air echo.
Despite FMCW radars being excellent tools for turbulence studies in the convective ABL and
in the capping inversion region (Eaton et al., 1995), relatively few atmospheric studies reported
in the literature have used FMCW radar data extensively for both quantitative and qualitative
analysis of clear-air features observed. While S-band FMCW radars have been designed with
the capability to obtain height and time resolutions of 1 m and 1 s, respectively, the degree to
which these resolution limits are obtained in practice depends upon the properties of the
atmospheric echo itself. Range measurement errors due to non-zero Doppler velocities and
finite coherence of the clear-air echo have implications on both spatial resolution and Doppler
estimation of FMCW radars. Although the theoretical background for FMCW radars is well
established, their measurement limitations for atmospheric targets and performance impact for
atmospheric applications has received limited attention in the literature.
The detailed review of the theory of operation of S-band FMCW radar and its application to
atmospheric boundary layer profiling can be found in the literature (Eaton et al., 1995; Ince
et al., 2003). In this study, performance aspects of atmospheric FMCW radars including the
effects of range-Doppler ambiguity, parallax typical of two-antenna radar systems, and
near-field operation are discussed. The data collected during recent field experiments by the
high-resolution S-band FMCW radar, described in Ince et al. (2003), is used to illustrate
system performance. Analysis of FMCW radar signatures of atmospheric targets detectable
at S-band frequencies has been performed by extracting both qualitative morphological
information and quantitative information on the intensity of backscattered power from
clear-air turbulence.
2. Principles
Frequency-modulated, continuous-wave (FMCW) radars may be thought of as a limiting
case of pulse-compression radar where the duty cycle of the transmitted waveform
approaches 100%. They operate by transmitting a long, coded waveform of duration T and
bandwidth B. The improvement factor they gain over pulsed radars of equivalent range
resolution is given by the time-bandwidth product of the waveform BT, which is often
referred to as the compression gain. In FMCW systems, this gain can be very large,
exceeding 60 dB. While several types of frequency coding may be used to yield the
bandwidth B, linear frequency modulation is the simplest and most commonly used method
in atmospheric FMCW radars.
Consider an FMCW radar that transmits constant amplitude linear FM signal of the form
s (t )  exp( jwt  j (a / 2)t 2 ), -T/2 < t < T/2 (1)
where w is the radian frequency and a is the chirp rate in rad/s2. The echo from an
atmospheric target is essentially a delayed, attenuated, and possibly Doppler-shifted replica
of the transmitted signal. The echo is demodulated by mixing it with a portion of the
transmitted signal and low-pass filtering the result. The resulting beat frequency for a point
target at range R0 , moving at radial velocity ur , is given by
2ur 2 R0  4ur 
f r (t )  (  f  ft ), (2)
 c c
where  is the electromagnetic wavelength and f  a / 2  B / T is the chirp rate in

Hz/s. Here, the first term represents the Doppler frequency shift, f D , due to radial motion,
and the second term represents the shift due to the nominal range of the target, which is
exploited primarily by the FMCW radar. The final term represents a defocusing due to the
dilation of the reflected signal’s bandwidth due to the dependence of Doppler frequency to
carrier frequency. For lower atmospheric velocities, the effect of the last term on the
bandwidth of the echo may be safely ignored. Thus, for the case of stationary radar and
targets, the beat frequency is linearly proportional to range.
Signal processing of the echo usually involves matched filtering, which is most commonly
implemented for all ranges simultaneously through spectral analysis via an FFT algorithm.
In this case, the output of the matched filter, which gives the response of the linear FMCW
radar to a moving point target, can be expressed as
sin   f DT  R  R0  / R 
y( R)  ,
  f DT  R  R0  / R  (3)
where f D is the Doppler frequency, and R  c / 2 B  is the expected range resolution for
a transmitted bandwidth B  fT . It is worth noting that this result is equivalent to impulse
response of linear FMCW radar.
For FMCW Doppler radar, Doppler information can be retrieved on a sweep-to-sweep basis
by analyzing the sequence of echoes from a particular range, as discussed by Strauch et al.
(1976). In this case, the sampling frequency is the reciprocal of the sweep period T, the
Nyquist Doppler frequency is 1/2T, and resulting in an unambiguous Doppler velocity
interval of ur   / 4T assuming that range measurement errors due to Doppler can be
ignored. Conventionally, two-dimensional FFT is performed on received signals to obtain
range and Doppler frequency (velocity) spectrum.
The received mean power of a bistatic weather radar from a distributed atmospheric target
of uniform reflectivity  , for circularly symmetric Gaussian shape antenna patterns, is
given by Doviak & Zrnic (1993),
Pt Gt Gr 2 12 R
P R   , (4)
512 ln 2 2 R 2
where Pt is transmitted power, Gt and Gr are transmit and receive antenna gains, 1 is the
one-way 3 dB beamwidth, R is range resolution, and  is radar wavelength. To better
estimate  from the measured echo power at the output of receiver, the radar system must
be calibrated. Based on Tatarskii’s theory of electromagnetic wave propagation in a
turbulent atmosphere (Tatarskii, 1961) and Ottersten’s work in his 1969 landmark paper
2
(Ottersten, 1969), clear-air radars can measure the refractive-index structure parameter ( Cn )
from radar backscattered power. For Bragg scattering from refractive index fluctuations due
to homogeneous, isotropic turbulence,  is commonly related to Cn2 by Ottersten’ well-
known equation:
  0.38Cn21 / 3 . (5)
This equation assumes that the Bragg wavenumber, k B  4 /  lies within the inertial
subrange. Note that Cn2 in (5) is defined as an ensemble average. In practice, such averaging
is performed over space or time assuming ergodicity. Therefore, for S-band radars, Bragg
scatter from clear air depends much less on the wavelength of the radar than that of
Rayleigh scatter ( 4 dependence) by airborne particles.
3. Performance Analysis
3.1 Range and Doppler
FMCW radar performance characteristics, such as range-Doppler coupling, can be seen from
analysis of (3), which is also ambiguity function for linear FMCW signal. The presence of
both range and Doppler terms in the argument of the sinc function illustrates the effect of
target motion on the radar’s ability to locate. Figure 1 shows a typical two-dimensional time
delay (range) – Doppler (velocity) ambiguity surface for a linear FMCW signal with time-
bandwidth product of 1000. As seen from Figure 1, maximum unambiguous Doppler can be
increased (by decreasing sweep period) at the expense of range ambiguity, or alternatively
increasing maximum range results in Doppler ambiguity. By rearranging the terms in the
argument in (3), the apparent range of the target can be expressed as Rapp  R0  f DTR,
where it is easy to see that range mislocation by one resolution cell occurs when f DT  1, or
ur    / 2T which also corresponds to the Nyquist velocity interval for FMCW Doppler
radar. Therefore, targets with unambiguously measured velocities are misregistered by no
more than one half a resolution cell, or misregistration occurs when target velocities are
aliased.
Additionally, the coherence of the atmospheric target during the sweep interval will limit
range resolution of linear FMCW radar. For complex moving targets and for volume
scattering, the coherence time (or the reciprocal of the Doppler spectral width) of the echo
will limit resolution. A distribution of Doppler velocities observed over an integration time,
T, will yield a distribution of apparent ranges. The resulting rms spread in range is
determined by the transformation of the Doppler spectrum to the range domain using the
relationship Rapp  R0  f DTR,
 R   f TR , (6)
where  f is the Doppler spectral width of the echo. From this relation, it is apparent range
resolution and sensitivity are optimized by matching the sweep time to the reciprocal of the
Doppler bandwidth of the echo. In this case, the range spreading is equal to the range
resolution, and the entire Doppler spectrum is confined to one range bin. No improvement
in sensitivity or in resolution is achieved by increasing sweep time beyond this value, as the
resulting echo simply spreads to adjacent range bins.
Fig. 1. The ambiguity surface for a linear FMCW signal with BT = 1000.
To maximize resolution and sensitivity, it is desirable to make both bandwidth and sweep
time as large as possible. The effective value of sweep time is however constrained by the
coherence time of the atmospheric echo. For example, a sampling volume with rms radial
velocity of 1 m/s has a spectral width of 20 Hz at 3 GHz implying a coherence time of
approximately 50 ms.
3.2 Parallax and Near-field Operation

From the weather radar equation (4), the received mean power can be correctly estimated
only if radar targets are in the far field of antenna, that is R  2 D 2 /  , where D is the
diameter of the antenna. Additionally, for bistatic FMCW radar employing two spatially
separated antennas, a correction of the received power is necessary at near ranges to account
for the reduced beam overlap. For Gaussian shaped beams aligned with their axes parallel,
the antenna parallax function, or fractional beam overlap is given by Ince et al. (2003),
d2
C p R   exp( 2 ln 2 ), (7)
12 R 2
where d is the separation distance between transmit and receive antennas, and 1 is the one-
way half-power beamwidth. Note that parallax results in an apparent reflectivity reduction
and parallax error of backscattered power becomes significant within the radiating near-
field of the antennas such that assumption of a far-field Gaussian beam pattern is not valid.
It is also worth noting that the far-field criterion, R f  2 D 2 /  , is based on a conservative
specification of maximum phase error, significant effects on the shape of the main lobe are
not evident until R  R f / 4 (Hansen, 1985).
FMCW atmospheric radars can measure boundary layer targets at minimum distances as
close as 50 m. However, this minimum range lies in the Fresnel region in which antenna
gain and pattern shape vary significantly with distance, and from (4) radar reflectivity is
underestimated for targets in the near field. Based on the universal near-field reflectivity
correction by Sekelsky (2002), 1 dB reduction in reflectivity at R f / 4 is indicated, and at
closer ranges ( R  R f / 4 ) near-field gain reduction becomes much more significant. Since at
closer ranges the beam shape is no longer approximately Gaussian and is strongly
dependent upon the particular antenna design, it is necessary to apply appropriate
reflectivity correction for parallax and near-field gain reduction depending upon the
parameters of radar system. In Figure 2, the relative importance of near field gain reduction
and parallax for the parameters of the S-band FMCW radar system in Ince et al. (2003) is
compared, and in this case parallax has the primary influence at close ranges.
Fig. 2. Reflectivity reduction due to parallax (solid line) and near field (dashed line) effects
for the parameters of the FMCW system in Ince et al. (2003).
4. Experimental Measurements
Data collected by the University of Massachusetts’ high-resolution FMCW radar (Figure 3)
during recent field experiments is used to illustrate system performance and analyze radar
signatures of atmospheric targets detectable at S-band frequencies. Detailed description of
this system and data processing can be found in Ince et al. (2003). The S-band FMCW radar
system implements an internal calibration loop (by injecting into the receiver an attenuated
and delayed sample of the high power amplifier output) to perform calibration of the
atmospheric echo and estimate volume reflectivity.
Fig. 3. The University of Massachusetts S-Band FMCW radar system mounted on flatbed
truck (electronics are contained within the cab).
Figure 4 shows one day continuous radar record of diurnal cycle of the ABL beginning at
14:27 local time (CDT, 19:27 UTC). The radar echo is expressed in terms of the logarithm of
microwave Cn2 obtained using (5), however, this representation is only meaningful for the
clear-air component of the backscatter described by the Bragg scattering mechanism. In
addition to distributed Bragg scatter from clear air, Figure 4 shows strong point echoes,
which are in this study assumed to be entirely due to Rayleigh scatterers. It is worth noting
that the distribution of Rayleigh scatter in reflectivity time-height image appears to reveal
qualitatively additional boundary layer structure not otherwise detectable at S-band.
However, the effect of much stronger Rayleigh echo (due to much stronger wavelength
dependence,  as opposed to 
4 1 / 3
for Bragg scattering) on quantitative backscattered
signal information is adverse. Interestingly, with high spatio-temporal resolution capability
of FMCW radar, by reducing time averaging to below 1 s in Figure 4, it becomes possible to
see undulations of O(10 m) in the radar echo as a consequence of Doppler-induced range
measurement error due to the flapping of bird’s wings. It is estimated that under clear sky
conditions vertical velocities of atmospheric echoes in general will contribute
misregistration of about one range bin at most.
From Figure 4, four-hour time period of convective ABL (between 14:30 to 18:30 local time)
is extracted to perform quantitative analysis of FMCW radar reflectivity. In this case, the
collocated radiosonde (operated by NCAR/ATD) measurements of temperature and
humidity show typical characteristic of convective ABL bounded above by dry air. In the
three panels of Figure 5 respectively reflectivity, Doppler (vertical) velocity, and the
correlation coefficient of successive echoes for one-hour of convective ABL are shown. The
latter two products are the result of the pulse-pair processing (Doviak & Zrnic, 1993)
averaged over 20 pulses (~1 s averaging).
Advances in Geoscience and Remote Sensing
Fig. 4. S-band FMCW radar profiler image showing approximately 24-hour record of the ABL during CASES-99 experiment on
26 Oct beginning at 14:27 CDT.
652
Fig. 5. First hour of convective ABL echo from Figure 4 showing reflectivity, Doppler
velocity, and single-lag correlation coefficient.
The reflectivity image shows that initially significant Rayleigh backscatter is observed both
above and below the capping inversion which peaks near 1000 m altitude at about 17:00
local time. After this time, both the Rayleigh scatter and the distributed Bragg scatter below
the inversion decrease significantly, and the strong echo at the top of the boundary layer
disappears. The velocity image which is derived from the phase of the single-lag covariance
shows structure for some, but not all of the clear-air echo (due to velocity aliasing). The
single-lag covariance panel shows extremely
high correlations for the Rayleigh scattering while the correlation coefficient for the Bragg
scattering is relatively lower on average (due to larger spectral width of clear-air echo).
To discriminate the distributed clear-air echo from particulate scatter, a signal processing
method based on the correlation coefficient of successive echoes can be applied to time-
series of backscattered power. According to this method, magnitude of the sweep-to-sweep
correlation coefficient is compared to an empirically determined threshold (in this case, a
value of 0.8 is used) and high-correlation points (due to Rayleigh scatter) are removed and
filled with estimated data by using least-squares interpolation (without affecting the
statistics of the atmospheric clear-air echo). This processing allows estimation of
atmospheric component of the clear air backscatter. Alternatively, median filtering can be
applied to the two-dimensional radar reflectivity image to differentiate morphologically
differing Bragg scatter (distributed echoes) and Rayleigh scatter (dot echoes). However, in
case of large density of particle echoes, their morphology will be similar to distributed
scatter, hence the filter will not be able to separate two sources of scattering.
Fig. 6. Vertical profiles of estimated mean Cn2 (large data points) obtained from post-
processing over four consecutive 60-min periods. The small dotted profiles show the mean
reflectivity including both Bragg and Rayleigh echo. The dashed lines are normalized
profiles of Cn2 based on the Fairall’s model (1987).
In Figure 6, the computed vertical profiles of mean Cn2 (estimated from postprocessed
reflectivity) for four consecutive 60-min segments are plotted against height. The small
dotted profiles show the corresponding mean reflectivity over the same intervals including
both Bragg and Rayleigh echo. The star (*) profiles indicate estimated clear air component of
the vertical profile over the 60-min interval. The vertical axis of each profile is scaled by the
boundary layer depth, zi , which can simply be obtained from the maximum of the
reflectivity. Given the mean horizontal winds, the vertical profiles are roughly equivalent to
a streamwise spatial average over approximately 10 km. Due to an increased number of
insect echoes in the afternoon boundary layer and insect plumes (tracing sudden vertical
motion from updrafts), Rayleigh scatterers dominate the observed profiles of mean
reflectivity.
In this case, FMCW radar estimates of Cn2 can be used to quantitatively test theoretical
predictions on the ABL. In the free-convection boundary layer, Cn2 is expected to follow a
z 4 / 3 power law, where z is height above ground level (Wyngaard & LeMone, 1980). From
Figure 6, radar estimated Cn2 increases with height inside convective mixed layer (between
2/3
0.2 zi and 0.9 zi the mean reflectivity follows a z profile, a discrepancy of z 2 ) with an
expected sharp turbulence induced peak at the top of the mixed layer (~ zi ). Note in this
case strong ground clutter affects Cn2 measurements below 0.1zi . This observed
discrepancy in Cn2 from its mixed-layer prediction may be due to entrainment effects, which
can be predicted by a model developed by Fairall (1987) based on a top-down and bottom-
up diffusion approach in the entraining, convective boundary layer. In Figure 6, by choosing
appropriate values for Rc and Ac parameters (respectively top-down and bottom-up
components) of the model, the predicted normalized Cn2 profiles is fitted to the vertical
profiles of mean reflectivity from experimental data. Additionally, theoretical models on the
shape of the reflectivity profile in the inversion region have been developed by Wyngaard &
LeMone (1980). The existence of the reflectivity peak is known from in situ measurements
(Wyngaard & LeMone, 1980) and from large-eddy simulation (Muschinski et al., 1999).
In Figure 7, log-variance of estimated Cn2 over four consecutive 60-min periods is compared
against the radiosonde-based potential temperature profiles. It can be seen that much higher
variance occurs across the top of the convective boundary layer (corresponding to the height
of the strongest potential temperature gradient) due to strong refractive index turbulence,
and peak Cn2 variance increases as the potential temperature jump across the inversion
increases.
Fig. 7. The log-variance of estimated Cn2 over four consecutive 60-min periods (left), and
corresponding potential temperature profiles measured by radiosonde (right).
In Figure 8, radar reflectivity time-height image (starting at 0650 UTC in Figure 4) of
nocturnal boundary layer turbulence below and insect layer within a low-level jet in the
NBL is shown. Collocated radiosonde profile of wind speed taken at 0658 UTC is shown in
Figure 9. The FMCW reflectivity shows intensified turbulence and wave activity from
surface extended to heights of about 150 m where sounding shows evidence of strong shear.
Additionally, radiosonde wind measurements shows a low-level jet formed between heights
150 and 550 m with maximum winds of over 16 m/s at approximately 300 m height. In this
case, a layer of insects detected by the radar at heights of the wind speed maximum helps
reveal qualitatively this additional boundary layer structure.
Fig. 8. Nocturnal boundary layer turbulence below and insects within a low-level jet
observed at 0650 UTC.
Fig. 9. Collocated radiosonde profile of wind speed at 0658 UTC.

5. Conclusion
In this study we have discussed and illustrated system performance of FMCW radar
detection of atmospheric targets at S-band frequencies. In particular, the effects of non-zero
Doppler velocities and finite coherence of the atmospheric echo on both spatial resolution of
FMCW radars and on Doppler estimation have been presented. Additionally, we have
considered parallax errors in reflectivity and near-field operation. For the clear-air
atmosphere measurements, the S-band radar’s sensitivity to Bragg and Rayleigh scattering
and the effect of Rayleigh scatterers on vertical profiles of mean radar reflectivity have been
illustrated. Experimental data collected by the high-resolution atmospheric FMCW radar
has been used to perform analysis of S-band radar signatures of atmospheric targets and to
quantitatively test theoretical predictions on the atmospheric boundary layer.
6. References
Chadwick, R.B.; Moran, K.P.; Strauch, R.G.; Morrison, G.E. & Campbell, W.C. (1976).
Microwave radar wind measurements in the clear air. Radio Science, vol. 11, pp.
795–802.
Doviak, R.J. & Zrnic, D.S. (1993). Doppler Radar and Weather Observations. 2d ed. Academic
Press, ISBN-10 : 0122214226.
Eaton, F.D.; McLaughlin, S.A. & Hines, J.R. (1995). A new frequency-modulated continuous
wave radar for studying planetary boundary layer morphology. Radio Science, vol.
30, pp. 75–88.
Fairall, C. W. (1987). A top-down and bottom-up diffusion model of Ct2 and Cq2 in the
entraining convective boundary layer. J. Atmos. Sci., vol. 44, pp. 1009-1017.
Gossard, E.E. (1990). Radar research on the atmospheric boundary layer, In: Radar in
Meteorology, edited by D. Atlas, pp. 477-527, Amer. Meteor. Soc., Boston.
Hansen, R. C. (1985). Aperture theory, In: Microwave Scanning Antennas, vol. 1, edited by R.
C. Hansen, Penninsula, Los Altos, Calif.
Hirsch, L. (1996). Spaced-antenna-drift measurements of the horizontal wind speed using a
FMCW-radar-RASS. Contr. Atmos. Phys., vol. 69, pp. 113–117.
Ince, T. ; Frasier, S.J.; Muschinski, A. & Pazmany, A.L. (2003). An S-band frequency-
modulated continuous-wave boundary layer profiler: Description and initial
results. Radio Science, vol. 38, no. 4, pp. 1072–1084.
Martin, W.J. & Shapiro, A. (2007). Discrimination of bird and insect radar echoes in clear air
using high-resolution radars. J. Atmos. Oceanic Technol., vol. 24, pp. 1215–1230.
Muschinski, A.; Sullivan, P.P.; Wuertz, D.B.; Hill, R.J.; Cohn, S.A.; Lenschow, D.H. &
Doviak, R.J. (1999). First synthesis of wind-profiler signals on the basis of large-
eddy simulation data. Radio Science, vol. 34, pp. 1437-1459.
Ottersten, H. (1969). Atmospheric structure and radar backscattering in clear air. Radio
Science, vol. 12, no. 4, pp. 1179–1193.
Richter, J. (1969). High-resolution tropospheric radar sounding. Radio Science, vol. 4, pp.
1261–1268.
Sekelsky, S. M. (2002). Near-field reflectivity and antenna boresight gain corrections for
millimeter wave atmospheric radars. J. Atmos. Oceanic Technol., vol. 19, pp. 468–477.
Strauch, R.G.; Campbell, W.C.; Chadwick, R.B. & Moran, K.P. (1976). Microwave FM-CW
Doppler radar for boundary layer probing. Geophys. Res. Lett., vol. 3, pp. 193–196.
Tatarskii, V. I. (1961). Wave Propagation in a Turbulent Medium. McGraw-Hill, New York.
Wilson, J.W.; Weckwerth, T.M.; Vivenkanandan, J.; Wakimoto, R.M. & Russell, R.W. (1994).
Boundary layer clear-air radar echoes: Origin of echoes and accuracy of derived
winds. J. Atmos. Oceanic Technol., vol. 11, pp. 1184–1206.
Wyngaard, J.C. & LeMone, M.A. (1980). Behavior of the refractive index structure parameter
in the entraining convective boundary layer. J. Atmos. Sci., vol. 37, pp. 1573-1585.
.
Use of Lidar Data in Floodplain Risk
Management Planning: The Experience of Tabasco 2007 Flood 659
32
X

Management Planning: The Experience
of Tabasco 2007 Flood
Ramos J., Marrufo L. and González F.J.
Instituto de Ingeniería, Universidad Nacional Autónoma de México
Av. Universidad 3000, 04510 Coyoacán, México, D.F.
1. Introduction
Floodplains along rivers are vulnerable and suffer frequent socio-economical losses as
consequence of large rainfall events. The risk of flooding is function of several factors such
as: climate, the material that makes up the banks of the stream, the channel slope and the
topography. The risk of floodplains is increased as a result of the human tendency to use the
river land to cope with community demands. As the river plains are synonymous with
fertile soils with high phreatic levels, their transformation into arable areas or into urban
areas are guaranteed to produce an economical input improving human life. This situation
makes flood less important since the benefits produced outweigh the damage; sometimes
damage could be reduced or mitigated by constructing barriers. The actual problem is when
the flooding generates devastation and the benefits are minimised. This is a situation that
could be present at any moment through the development of the region (Maza and Franco,
1997). This is the case for many of the watersheds in the Gulf of Mexico including the
Grijalva Basin, one of the largest in Mexico, which suffered severe flooding in 2007 affecting
almost 70 % of the Tabasco state. As a consequence around one million residents lost their
homes and the economic impact was calculated in millions of pesos.
The modification of the river landscape is determinant because it alters the natural condition
of the area dredging water bodies, change drainage, building impermeable surfaces and
constructing barriers. These modifications make it difficult to contain the water in the river
channel when prolonged heavy rains take place as is the case in recent times, where
flooding becomes more frequent, extreme, damaging and costly (Werritty, 2006).
A frequent problem is that floodplains lose their capability to preserve the hydraulic
conditions to prevent floods and to function as areas to support ecosystems. In general,
floodplains provide important habitats for plants and animals, and also these areas can
retain water, reduce the flow velocity of a river and minimise erosion reducing the flood
impact downstream (Moss and Monstadt, 2008). However, it is not enough to study the
water movement in the watershed in order to determine the amount of water arriving to the
river channel and to be transported during the flood event. It is also necessary to know the
dynamics of the river and to determine the spatial-temporal distribution of flow as well as
the water level reached in the floodplains (Stabel and Löffler, 2003). The latter is addressed
through the use of flood maps (if they exist) but most of them are theoretical predictions.
Other maps are gathered using remote sensing data obtained by tracking the movement of
the water on the land. However, it is difficult to find a direct relation between ground
elevation and the highest expected river level.
The research reported in this chapter presents a contribution to calculate water levels and
flooding maps using Geographic Information Systems (GIS) to combine existing datasets,
water elevation survey data and hydrological models. There is a major emphasis to use high
accuracy elevation data to model and map a particular flood event (past or future), to assess
the risk of flooding in a particular region (Jones, 2004; Jones et al., 1998). LiDAR (Light
Detection And Ranging) technology has the ability to produce accurate elevation data
throughout the flood plain, in particular flat areas with slopes of less than 50 cm per
kilometre as is the case of the coastal flatplains in Tabasco. LiDAR data allows a better
calibration of hydrological models, to map the flood risk and water levels helping decision
makers to cope with an emergency and to implement strategies after the event (Filin, 2003;
Webster & Stiff, 2008).
2. Gathering topographic data by Digital Elevation Models

As floodplains are neither static nor stable, one needs to consider at least three aspects
related to the land surface characteristic to describe a flood: (a) topography, (b)
geomorphology, particularly when unconsolidated fluvial deposits are the base material,
and (c) hydrology and the extent of recurring floods (DRD, 1991). These three aspects
require the acquisition of different sources of information that in the case of geographic data
is quite difficult to achieve using traditional terrestrial surveys and photogrammetry due to
time consuming and spatial constraints (Lijian et al., 2008). Thus, there is a need to know the
dynamics of the river and to determine the spatial-temporal distribution of flow. The use of
remote sensing techniques has demonstrated to be an efficient tool to map the flood extent
(Horritt & Bates, 2002; Jensen, 2000). These maps have been used to validate 1D or 2D flood
models; however it is not enough to study the water movement in the watershed. It is also
necessary to know the water level reached in the floodplains in order to determine the
amount of water arriving to the river channel and to be transported during the flood event.
In this case, the use of 3D information is a better option since it offers: river depths, river
surface and banks topography, vegetation height, hydraulic structures and identifies
different objects present in the river channel (Mohammadzadeh & Valadan Zoej, 2008). One
alternative to acquire altimetry measures as well as identifying land surfaces features is the
time-of-light information from a LiDAR system mounted on air- or- space based platforms.
2.1 LiDAR data

LiDAR data can acquire high precision and density elevation data of terrain surface since
this technology involves laser altimeters (i.e. Nd:YAG laser or pulsed gallium-arsenide
diode laser), a Global Positioning System (GPS) as a position reference to track the aircraft
trajectory, and an Inertial Measurement Unit (IMU) as the main component of Inertial
Navigation System (INS) to calculate the position, orientation, and velocity of a moving
object without external references (Webster & Dias, 2006a). Thus, knowing the angle and the
absolute position of the instruments it is possible to establish the location (x,y,z) of the laser
pulse destination precisely and accurately obtain a real representation of the ground.
The LiDAR aircraft flew at 500 m to 1 km above the surface, thus the accuracy of the laser
altimeters is around a few centimetres in range (Weitkamp, 2005; Raber et al., 2007). The use
of illumination with visible light from artificial sources has been used for active optical
detection of objects considering relative distance values. These distance values can be
determined measuring the transit time of radiation from the source to the object and back.
This means that when a short pulse of radiation is transmitted in a medium, the radiation
scattered from an object is going to be sensed as a function of the distance of the object at the
surface. Of course the sensed signal is not going to have the same length as the transmitted
pulse, but it will be extended in time (Weitkamp, 2005). This technique is also suitable for
microwaves and sound sources and time-resolving detection systems. LiDAR has sensors
for visible (λ= 400 – 700 nm), ultraviolet (λ= 225 – 400 nm) and infrared (λ= 700 – 1200 nm)
wavelengths. For this reason, LiDAR not only provides elevation data but also provides
information on pressure, temperature, turbulence and wind for gases in the upper, middle
and low atmosphere, and can be used to determine hydrospheric and land use parameters
(oil slicks, phytoplankton, water depth, turbidity, type of land use, health of vegetation,
tree-stem diameters, biomass, and leaf area index, amongst others).
The extraction of topographic parameters using LiDAR gives advantages compared with
other techniques such as photogrammetry since the high density data acquired avoids loss
information. Also, it is not influenced by the shadow angle of the sun and it is not
dependent on the ground control since it can gather a huge quantity of control points.
The high vertical accuracy of LiDAR enables the production of Digital Elevation Models
(DEM), Digital Surface Models (DSM) and a Digital Orthophoto map (DOM) with high
resolution and precision at a low cost compared with traditional techniques (Lijian et al., 2008)
Despite the high accuracy of DEM, it is important to identify the errors associated with the
LiDAR data in order to reduce them (Raber et al. 2007; Webster & Dias, 2006a). The errors
can be associated with several causes, some of them are related to the system (characteristics
of the platform and laser) and conditions (terrain characteristics, vegetation and
atmospheric conditions) (Filin, 2001). Also, Raber (2006) noticed the influence of the land
cover and slope of the surface. To solve this Cobby et al. (2001) proposed a segmentation
method that was tested by Raber et al. (2002) finding that the vertical accuracy is improved.
Other errors are associated with the interpolation method and the data model utilised to
create the DEM (Hodgson & Bresnahan, 2004). In order to reduce the systematic error,
Webster et al. (2006b) suggested to perform an independent validation. In fact LiDAR
processing data includes several steps such as: (a) data collection, (b) data pre-treatment, (c)
filter and classification of point cloud, (d) fusion and application, and (e) ground-object
extraction and modelling.
3. River and floodplain and the 2007 flooding

The present case study is based on the analysis of the Grijalva River Basin flooding in 2007,
affecting almost all of Tabasco state resulting in enormous socio-economical losses.
3.1 Grijalva River Basin: characteristics

The Grijalva River Basin (GRB) is one of the largest basins in Mexico, low-lying in the
Tabasco state at the Villahermosa city, and in the river delta at the Gulf of Mexico. The side
branches of the Grijalva River have high runoff percentages (20%) associated with their
medium permeability with little vegetation or low permeability with dense vegetation. The
Grijalva River is born in the Chiapas state changing its name to Mezclapa after the Malpaso
dam, to recover its name after it crosses Villahermosa city to its end at the Gulf of Mexico.
The Grijalva River is a mature river crossing Tabasco; an alluvial plain that made it to
divagate generating broad plains, wide flows and meanders. Also, it has been divided over
time creating two of the most important branches of the Carrizal River in 1881, and the
Samaria River in 1932. Other important side branches of the Grijalva coming from the
Chiapas Mountain Chain are the Pichucalco, Puyacatengo, Teapa and the La Sierra rivers
(figure 1). The Grijalva River’s total length is approximately 640 km and in its main branch is
the country’s fourth largest hydroelectric plants in the Chiapas state (GET, 2008; INEGI, 2000).
Fig. 1. The Grijalva River and tributaries around Villahermosa city, Tabasco. Landsat TM
image from 28/03/2001
Tabasco is divided into two physiographical regions: Llanura Costera del Golfo Sur (Coastal
plain at the south Gulf) covering almost all the state (94%) with the Sierras de Chiapas and
Guatemala (Chiapas and Guatemala mountain chain) covering the remaining 6%. The main
characteristics of the coastal plains are a gradual sedimentation area refilled with detritus
material and eroded zones with low elevations (less than 60m), the majority related to small
subtracted dominated sand and limestone highland rivers. This made the zone a source of
important lacustrine, palustrine and litoral deposits linked to extended floodplains, abandon
streams and coastal lagoons.
The climate is warm humid subtropical with maritime influence. The mean annual
temperature is between 24 to 28oC with a maximum and minimum mean annual
temperature of 28-30 oC in May and 21-25 oC in January, respectively. The rainy season is
from June to October with a mean annual precipitation in the coast of 1,500 mm with a
gradual increment to the intercontinental zones till 5,000 mm. Two rainfall peaks are
present, one in June with 220 mm, and the other in September with 348 mm; some authors
consider another peak in October with 370 mm. This generates a cloudy atmosphere with a
relative humidity around 80 to 86%. These mean annual values are a result of the extensive
plains that commonly produce air masses from the sea that bring the major contribution to
precipitation (INEGI, 2000).
3.2 Protection along the Grijalva River

The economic activity in Tabasco (mainly petroleum and agropecuary industries) expanded
a rapid growth of most cities, the most important one being Villahermosa, capital of the
state. This city is bordered by the Grijalva River (Carrizal, Mezcalapa and de La Sierra
tributaries) and is quite close to the Gulf of Mexico giving at it an important role. Originally,
Villahermosa was located in highland rivers surrounded by undisturbed rain forest and
water bodies. However, the rapid growth of population changed the physiography of the
city by dredging water bodies, modifying the course of the rivers and removing the
vegetation from the surface to build urban areas (figure 2) (Oropeza, 2004).
Fig. 2. (a) City of Villahermosa in 2003 (Google image 15/03/2003) and (b) proposal for the
growth population for Villahermosa after mid 2007 (PEDUT, 2007)
As Villahermosa was developed along the Carrizal, Mezcalapa and de La Sierra Rivers the
risk of flooding is at its greatest in events of prolonged heavy rainfall. In order to protect the
city from floods the water regime was disrupted constructing technical structures such as
levees, barriers, dykes, roads, etc. In this manner, after the 1955 flooding barriers were
constructed along the Samaria River, and the Malpaso and La Angostura dams were built to
regulate runoff and reduce the risk of flooding in Villahermosa (Maza, 1997). Years later, the
development of the transport communication system in the state and, particularly, in
Villahermosa made it possible to construct levees and barrages, change of the river flow and
dredging of the rivers. Thus there was a dramatic change in the dynamics of the Grijalva
River (GET, 2008) increasing the flood risk since the water levels overpass the channel and
the levees act as retainers of water.
In 1997, Maza studied the hydraulic conditions in the Grijalva-Usumacinta Basin and
attending to the socioeconomic development in the region he proposed several alternatives
to protect against flooding. These actions were reviewed after the 1999 flood. This was an
extreme event with approximately a 20-year return period that affected the Tabasco state
with loss of human lives, houses and an impact on the economy. Some of the proposal
considered the restitution of channels, the use of lagoons and swamps for regulation, and
the improvement of current infrastructure (dams and barriers constructed along the rivers
Carrizal and Grijalva in the Villahermosa city). Some tasks were done, but others were
unfinished and others were unattended. In particular, for La Sierra River it was evident that
hydraulic measures were not taken along the river maintaining its unregulated character.
This was the reason that in the summer and autumn of 2007, several precipitation events
registered along the Gulf of Mexico decreased the free water surface in the rivers,
overtopping them in zones with instable soils or where the barriers failed.
3.3 The greatest flood in Tabasco 2007

In the summer and autumn of 2007, several precipitation events were registered along the
Gulf of Mexico. The majority of this rain was related to 37 tropical depressions and 3 cold
fronts but also it was rain due to the hurricane Dean (August 22nd) and Noel (October 28th).
Exceptionally strong rainfalls took place on three periods: (a) October 10-11 associated with
the cold front No.2 plus a low pressure event below Guatemala, (b) October 22-24 associated
with the cold front No.4 located in Tabasco and Chiapas, and (c) October 28-30 associated
with a cold air mass in the southeast of the Gulf of Mexico plus the cold front No.4 and the
Tropical storm Noel (GET, 2008). After the strong rainfalls in the upper and middle GRB
(specifically in the Ocotepect town in the Chiapas Mountain Chain) with 4000 mm of
accumulated rainfall in 24 hours, the increase of the water levels in the river system and
dams generated by the end of October and the beginning of November, caused severe
flooding, which impacted 70% of the Tabasco state in Mexico, covering all urban and plain
areas. In particular, the Villahermosa capital city of the Tabasco state was covered by water
in around 80% of its territory due to the flooding of the Carrizal and Grijalva rivers that go
through the city, and also by rivers coming from mountain areas, in the border between
Tabasco and Chiapas, such as La Sierra and Pichucalco. As a result, the economic effect was
huge as thousands of residential and non-residential properties were destroyed, and jobs,
crops, animals, and infrastructure were lost. The total volume of overflow was around 1,800
million of m3 for a return period of 100 years (Domínguez et al., 2009).
Different causes have been given for this disaster for example the failure of the protection
system (levees), thousands of inhabitants allocated in the river divagation areas as a result of
the rapid population growth, and modifications in the river flow, amongst others. On one
side, as Tabasco was an alluvial plain with highland rivers, the risk of inundation was
lower. However, most of the low-lying alluvial plains have been filled to allow urban
developments. In consequence, the floodplain area has been increased over a long time. As
floodplains are areas subject to inundation by lateral overflow water from rivers with which
they are associated (Tockner and Stanford, 2002), the effect of its increment is more
significant, particularly the greatest flooding in 2007 with a 100 year return period event
(figure 3).
Fig. 3. Cross section of a hypothetical river flood. Source: DRD, 1991
On the other side, the hydraulic failure related to the levees could not only be explained by
the design (overtopping and foundation), since there were sections that were not
constructed due to financial or political reasons allowing water to pass to the urban areas.
Analysing the information available and using satellite images (Landsat TM and Spot 5) it
was observed that some barriers were constructed into two seasonal lagoons at the south of
Villahermosa that theoretically will be the natural drainage systems to regulate the amount
of water coming from the upper areas. As a consequence of the natural drainage system to
the lagoons together with overtopping the banks of the rivers or the top level of levees, the
water levels of the lagoons increased considerably since the reduction of the water discharge
demanded more time to be drained to the Gulf of Mexico. This magnified the flooding
condition in the city of Villahermosa, which was underwater for weeks (figure 4).
Fig. 4. (a) Grijalva River overflow in Villahermosa city and (b) full coverage flooding in 2007.
As result of the inundated areas that were under water for weeks and in some cases for
months, soils were highly saturated and during the rainy season in 2008, a flood again
occurred with fewer consequences but still with an impact on the economy.
4. Materials and methodology

The project used high-accuracy elevation data from LiDAR to get topographic information
of the flat areas that led to an improvement in the estimation of flood risk, the extent and
depth of the water and to plan efficient protective structures along the Grijalva River. Three
types of data were used: (a) DEM obtained by LiDAR data, (b) field surveyed three
dimension data for unchanged sites, water levels and maximum water levels during the
flooding in 2007 and 2008, and (c) ancillary data to integrate the GIS to provide the
information to map the water levels and flood risk.
4.1 DEM from LiDAR Data

The LiDAR data were acquired for all the Tabasco state between April-May 2008 and these
are being processed by the National Institute of Statistics and Geography (INEGI, 2008).
Figure 5 shows a DEM and DSM generated with 10% of the data processed by the INEGI in
March 2009. The procedure followed by the INEGI (2008) started adjusting the three
dimension cloud points with intensity values according to the National Active Geodetic
Network (RGNA) using the GPS mounted in the aircraft and the ground station. Next, there
is an altimetry adjustment to obtain a correction factor to be applied to the cloud of points.
After this adjustment, a semi-automatic classification and depuration of inconsistencies was
performed in order to obtain the terrain information and to generate the DEM. The DEM
was created from the surface classification giving a regular horizontal spacing of 1m
between each altitude measure. It was provided in binary datasets with a text file with
auxiliary data such as model dimension, type of data, cartographic projection, datum and
metadata. The geodesic elevation was transformed to orthometric elevations using the
Mexican Gravimetric Geoid (GGM05).
Fig. 5. LiDAR data available from the flight between April-May 2008 and the DEM
processed by the INEGI (2008)
The LiDAR data files available for the INEGI were delivered in LAS format (Schukman,
2003) and they were processed using the Lidar Analyst and then the 3D Analyst in ESRI
ArcGis 9.2TM to generate a grid structure where each cell provides a mean elevation in a two
dimensional array. The point data were interpolated to 1m size of the cell from Kriging; in
this case because of the flat topography the re-sampling was done to 0.25 m. After the
interpolation, it was the DEM and DSM were visualised, and the Surface Analyst (Spatial
Analyst) was applied to generate the contour lines. An advantage was the ability to take into
account the DSM and DEM in the project, thus road, bridges, barriers, levees, and dykes
could be identified among other features.
Additionally, the LiDAR DEM made it possible to plot the water movement using vectors
that show the water direction according to the topographic features. Thus, the 25 cm grid
cell permitted identification of the effect of the barriers in the lagoons since the water
instead of flowing downstream was retained and the flood waters were overtopping the top
level of the levees that protect Villahermosa (figure 6).
Flow movement paths in Figure 6 were plotted with the Golden Software Surfer 8.0 for both
DEM and DSM it was very interesting to observe the water movement changes with and
without protective barriers.
Fig. 6. Flow pattern at the southeast of Villahermosa with (DSM) and without (DEM) the
protective barriers against flood. The area in red corresponds to the airport barriers
4.2 Field survey data

The field survey data was obtained during October and November 2008 using differential
GPS. Two differential GPS were used: (a) Pathfinder® ProXH™ Trimble with Zephir
antenna to deliver a decimetre accuracy of 20 cm post-processing (Trimble, 2008), and (b)
SF-2040G integrated GPS/StarFire™ receiver from NavCom with decimetre accuracy (less
than 10 cm after post-processing). The NavCom antenna has a tri-band design with an
integral ground plane to track GPS, SBAS and StarFire™ satellites (NAVCOM, 2008). The
surveyed points were post-processed utilising a first order geodetic control monument
located in the INEGI headquarters at Villahermosa, Tabasco, to adjust the GPS position
according to the registers obtained in the GPS-NAVSTAR system.
The information collected are points that refer to an altitude, water levels in rivers and
lagoons and to the maximum water levels reached in the 2007 and 2008 floods (figure 7).
The 2008 flood was less severe than the 2007 flood but as the soil was still saturated the
inundation also had a significant impact in the region. It was not possible to have direct
measurements of the water levels in 2007, but some marks were still visible on trees in
November 2008. For the 2008 flood the water levels measured with the GPS (NavCom)
showed a nearly consistent height of 3.0 m above sea level compared with the mean 7 m
reached in 2007.
The importance of taking account of the terrain altitude and water level survey points was
to identify possible alternatives to restore the floodplain and recover their natural drainage.
The collected points correspond to different land covers (i.e. pavement, grass, shrub, bare
soil).
Fig. 7. GPS ground control points
Additionally, cross-sections and runoff gauged data were available for five river subbasins:
Tacotalpa-La Sierra-Grijalva, Puyacatengo, Teapa, Pichucalco and Carrizal. This information
was used to model the 2007 flooding in a Bi-dimensional programme to obtain the water
flow in plains (Fuentes et al., 2008). In order to feed the DEM into the hydrologic model, and
the accuracy of the DEM resolution was reduced from 1m to 100 m size grid cell. Although,
Tamiru-Haile & Rientjes (2005) pointed out that this reduction could result in a loss of
terrain information as is illustrated in figure 8.
Fig.8. (a) Grid cell used to model the 2007 flood, and schematic representation of the size
grid cell for (b) 100m and (c) 10m.
5. GIS system
Tamiru-Haile & Rientjes (2005) pointed out that data required for floodplain modelling
needs to be analysed, calibrated and verified. This comparison was done in a GIS that is an
important tool to capture, store, retrieve, analyse and display georeferenced information in
an efficient way. The information in a GIS comes from different databases that are managed
to generate new information, which can be displayed in a map or text file for a specific
purpose. GIS has been demonstrated to be useful to establish the flood extent and the
inundation patterns (DRD, 1991). An important advantage is that the GIS systems can
provide information covering spatial and temporal variations. In this case additional to the
baseline information (precipitation, floodplain limits, stream flow, etc.), other information
related to the flooding effects, DEM and control points were plotted to map not only the
flood extent but also the water levels. It also defines those areas with great inundation risk
and possible alternatives by identifying a natural drainage relieve.
Thus, the LiDAR data were compared using the GPS surveyed points. Figure 9 shows the 25
control points used to compare the altitude terrain and water depth values with the LiDAR
contour lines at the 2007 and 2008 flood (table 1). The control points were located in
pavement roads, bare soil and vegetation (grass, shrubs).
Fig. 9. Control points from GPS NavCom (yellow points) and curve levels obtained from
LiDAR spaced 0.25 m (red lines)
ID Site Y X Cota_NavCom Wlrive r Wllagoon MWL2007 MLW2008

1 Zapotes 2 18.00344737 -92.86025974 8.135 4.38
2 Sandy 18.23583052 -92.66792833 1.512 1.04
3 Porvenir 18.00878829 -92.87709476 4.906 2.673
4 Aztlan pto1 18.07729126 -92.74075388 2.984 1.864 1.624
5 Aztlan pto2 18.09308265 -92.75298726 2.606 1.982 1.602
Grijalva-Corcho 18.09148697 -92.73264987
6 2.636 1.999 1.864
7 Boca de Aztlan 18.09254211 -92.71610197 2.527 1.927
8 Chilapilla River 18.09439884 -92.70301897 2.419 1.908 1.846
9 Juan 18.05470363 -92.78099945 3.16 2.202
10 Grijalva pto7 18.07867476 -92.70327103 2.195 1.908
11 El Cuy 18.06221855 -92.70388467 1.894 1.742 1.732 2.312
12 Brigde La Pigua 2 18.02041803 -92.90815669 5.816 2.816
13 Bridge Grijalva1 17.99634156 -92.91183319 5.216 2.673
15 Bridge Grijalva 3 17.98000875 -92.92200179 5.491 2.689 7.317
Mezcalapa-Pichucalco (Pich) 17.95126038 -92.91668256
16 3.33 2.614
17 Bordo gaviotas 17.97986419 -92.89443397 3.262 2.788
18 Conf sierra-grijal (Grij) 17.96003023 -92.91344147 5.237 3.252
19 Zapotes 1 17.99593865 -92.8457946 3.72 2.349
Potreros Bordo Aeropuerto 17.98458333 -92.83833051
20 2.462 2.19
21 Bordo aeropuerto 17.99895294 -92.83758883 2.581 2.327 3.407
22 Lopez Portillo 18.01384817 -92.80752779 2.526 2.007
23 Lagoon Sn Julian 18.01772961 -92.82322041 1.998 1.926 3.776
24 Sn Julian-Ejido Gonzalez 18.04027172 -92.79818079 3.943 2.205 4.06
25 La piedad 18.04560432 -92.74357392 1.773 1.442 2.182
26 Majagual 18.03774539 -92.74202915 1.704 1.486 2.256
WLriver: water level for river; Wllagoon: Water level for lagoon;
MWL2007: Maximum water level reached in 2007; MWL2008: Maximum water level reached in 2008
Table 1. NavCom GPS control points surveyed in the study area
Generally, to determine the risk of flood a 100 year return period is modelled, in the Tabasco
flooding this theoretical situation was real, thus the flood zone extent was mapped and the
statistical and hydrological model confirmed the magnitude and the extent of the flooding
(Fuentes et al., 2009). The statistical analysis also considered flood events for 1-year, 2-years,
5-year, 10-year, 20-year, 50-year, 100-year, 200-year, 500-year, 1000-year, 2000-year, 5000-
year and 10000-year floodplain (Domínguez et al., 2009). Figure 10B shows the extent of the
2007 flood combining a Spot image from November 6th 2007 (maximum water level reached)
with contour lines every 1m. As one can observe the maximum mean water level was 7m
but in some sites it was even more as dependent upon the topography of the zone
(highlands or low-laying zones). In order to compare and to plot the magnitude of the 2007
flood, the 1999 flood was illustrated where a maximum water level was 4m although in
some places it was of 5 m (Figure 10A).
Fig. 10. Maximum water level reached during the flood events of (A) 1999
6. Results
6.1 Comparison between GPS and LiDAR
An initial comparison was carried out to see if the contour line represents the features at the
borders of the area covered by the DEM. As expected they match very well due to the high
accuracy of the LiDAR DEM.
The next comparison considered the surveyed points obtained with the NavCom GPS (see
Table 1) and the elevations at the same x and y coordinates using LiDAR DEM. Webster &
Dias (2006) proposed a method to valdate LiDAR data against GPS data. They proposed to
consider a radius less than 5 m around the GPS point covering an area with similar
characteristics. Although, the available control points were very few to carry out a rigorous
analysis comparison with the LiDAR DEM. They give a clear idea of the possible causes of
variation. Webster et al. (2006b) noticed that a systematic pattern could be observed due to
the limited precision of the laser ranging system and the motion of the aircraft. In this case
the information related to the aircraft and the acquisition procedure were not available, thus
the points delivered in LAS or text format georeferenced were used to generate the DEM
and DSM directly.
Horizontal values from LiDAR proved to have an accuracy less than 1 cm. However, for
elevation values, one found that for points located close to bridges the root mean square
error (RMSE) was ± 1.75m being the largest error observed. Two very clear examples are
point 1 (Zapotes 2, figure 11a) where GPS indicates a height of 8.135m and the LiDAR DEM
shows a height of 4.5m, and point 15 (Bridge Grijalva 3) where GPS indicates a height of
5.491m and the LiDAR DEM shows a height of 4m. However, using the DSM the RMSE was
reduced to ± 0.55m since vegetation, structural constructions and bridges were present. For
point 1 DSM was 7.5 m and for point 15 it was 5.0m.
For those points located at the river bank (in November it was bare soil but surrounded by
dense canopy, figure 11b) the RMSE was ±0.6m, this was because in some areas sediment
deposits were observed after the flooding mainly related to human intervention to create
grazing zones. The best example is point 8 (Chilapilla River), which through the years has
been converted into grazing land losing completely its floodplain. After the 2007 flooding
people did not agree to dredge the river and opposite to this they refilled the undermined
areas by flood.
Finally, a third type of area showed a RMSE of 0.±0.297m. These areas correspond to flat
areas attached to a lagoon such as San Julian point (figure 11c). In these areas the differences
between height LiDAR and GPS were from 2 to 44 cm with an average of 0.11m and
standard deviation of 0.30m.
Webster & Dias (2006) determined that the radium for validation is a function of the land
cover. Also, they considered that the vertical accuracy of LiDAR needs to be within an
average of 30 cm and for the GPS of 15 cm.
Although the NavCom GPS values were used in the study, the ones obtained using the
PROXH GPS had a similar accuracy with average differences about 15% that the NavCom.
Zapotes 2 Bridge
Chilapilla River
a. b. c.
Fig. 11. GPS points at different land types surveyed on November 2008. (a) Bridge Zapotes 2,
(b) Chilapilla River now a grazing area, and (c) San Julian lagoon, GPS located in an extense
of grass area
6.2 Water levels for flood extent

Analysing the water levels recorded the maximum height observed was at the Carrizal
River bank in Villahermosa with a height of 7.3m and the minimum was observed in the
lagoons area at the northeast of Villahermosa with a height of 2.2m. This confirm the water
level registered at the Gaviotas hydrometric station in the Carrizal River which maximum
value was 7.81m and the average was 5.42m with a flow of 1,772 m3·s-1. Also in the Carrizal
River, the Porvenir station registered a maximum water level of 6.67m and the González
station registered a maximum of 10.20m (average of 8.77m) and a flow of 1,191 m3·s-1. The
Pueblo Nuevo hydrometric station in La Sierra River registered a maximum water level of
8.40m (average of 7.49m) with a flow of 1,092 m3·s-1 (Fuentes et al., 2008). The maximum and
minimum water levels as a function of the topography are significant to determine precisely
the contour line at which the flood was extended during the 2007 flood and also during the
2008 flood. The contour line observed by GIS matched very well with the flood extent
obtained by running the hydrological model (Fuentes et al., 2008). Figure 12 shows the
impact of the water extent over Villahermosa and surrounding areas combining the contour
lines from LiDAR DEM and water levels measured in the field. The progress of the flood
considers a 20 years return period event such as the one in 1999 until a 100 years return
period as the one in 2007. Finally, the figure shows the comparison with the 2D hydrological
model result showing a close match.
The most evident fact was that the levees had the opposite effect to providing a protective
action since they favoured water retains that extended flood along the original floodplain
divagation. In order to avoid a similar flood several actions need to be done to regulate
water flow coming from the mountain’s rivers and to protect Villahermosa and its
surroundings by generating better discharge conditions from the lagoons located at the
south and southeast of Villahermosa.
Villahermosa Villahermosa Villahermosa
Contour Level 5 Contour Level 6 Contour Level 8

20 year return period
100 year return period
Fig. 12. Progress of flood extent for different storm return periods and their comparison with
the 2D hydrological model prediction.
Figure 13 shows the floodplain risk associated with a rainfall with a return period of 100
years in the alluvial plain in Villahermosa and surroundings. As one can observe they are a
low-lying area with slopes less than 50cm, thus the lateral overflow in the river can easily
inundate a vast region characterised by a low-lying land.
Fig. 13. Flood risk associated to the 1999 and 20007 floods showing how vulnerable is the
Villahermosa city
7. Conclusion
Combining existent datasets, hydrological models, results and measured data in the field
together in a GIS provide an effectively tool to analyse a flood event and to map not only
water extent but also water levels to identify highly vulnerable areas. The high resolution
LiDAR data is an important source of information to achieve an accurate representation of
the surface making it possible to map those areas under risk in a 3D geospatial distribution.
As Webster and Dias (2006) proposed an independent validation of the LiDAR data is an
important step to be considered in order to identify errors associated with this dataset.
Unfortunately, a flight report was not available thus an analysis associated to the system
and conditions of the flight was not possible. Nevertheless, a comparative test was carried
out with limited GPS surveyed data points but sufficient to observe some trends. Here the
comparison between LiDAR data and GPS surveyed values offered interesting results since
some differences were mainly associated with the land cover and topography aspects. In
this sense, the less representative data was obtained for those points located in bridges and
levees opposite to the control points attached to lagoons with grass extent lands where
heights were very similar. Although in the first case, the DSM can represent the actual
height accurately, it is important to consider the representiveness of the information
extracted from the LiDAR DEM.
Some authors reported a vertical accuracy for LiDAR DEM of 15 cm, in this case a grid cell
of 25 cm size proves the information necessary to ensure a better distribution of the water
extent and levels reached during the flooding. Also, even if the grid cell resolution was
reduced with a consequent loss of information to apply to the hydrological model, the grid
cell was enough to represent accurately features that traditional DEMs with a 30 m
resolution could not cope with. This provides better and more information related to the
origin, development and effects of the flood.
One important aspect was that a 100 year return period event was plotted allowing the
analysis of a real situation and not a hypothetical one. The 2007 flood in Tabasco showed
that the state and, in particular, the Villahermosa city, is a region subject to flooding as a
result of several human actions that reduced significantly the original floodplain zone. Thus
different actions need to be done at the upper, middle and low Grijalva River Basin in order
to protect effectively the whole area.
Further research considers using a flexible grid cell to extract information every 0.25m
instead of reducing the grid cell resolution. This requires a lot of data and processing time
but the benefits are higher due to the accuracy of the x,y,z LiDAR data. Mapping the flood,
one can identify those elements that failed. The critical one is to recognise the necessity to
recover the floodplains that have been highly degraded through generation of urban or
agricultural areas. By identifying the river status and its lateral connectivity, better
protective systems can be proposed to recover the natural functions and process of the
floodplains.
8. Acknowledgments
This project was founded by the National Water Commission. The LiDAR data was
provided by the sponsor. With thanks to MSc. Víctor Franco, who reviewed the paper for
comments.
9. References
Cobby, D.M., Mason, D.C. & Davenport, J.J. (2001). Image processing of airbone laser
altimetry data for improved river modelling, ISPRS Journal of Photogrammetry and
Remote Sensing, Vol. 56, No2. 2, pages 121-138, ISSN: 0924-2716
Domínguez, R., Carrizosa, E., Arganis, M. & Esquivel, G. (2008). Análisis Estadístico de los
gastos máximos anuales registrados en las cuencas del Bajo Grijalva y de las
descargas de la presa Peñitas, In Plan Hídrico Integral de Tabasco dentro del cual
se Inscribe la Ejecución del Plan de Acción Urgente (PAU) y la Formulación del
Plan de Acción Inmediata (PAI), para la Rehabilitación de la Infraestructura
Dañada por las Lluvias Atípicas de los Días 28 y 29 de Octubre del 2007, Ramos, J.,
& Vélez, L. (Ed.) pp. 1000, Instituto de Ingeniería, UNAM, Report project, México,
D. F.
DRD (1991). Primer on Natural Hazard Management in Integrated Regional Development
Planning, Department of Regional Development and Environment Executive
Secretariat for Economic and Social Affairs Organization of American States with
support of the Office of Foreign Disaster Assistance United State Agency for
International Development, Washington, D.C.
Filin, S. (2003). Recovery of systematic biases in laser altimetry data using natural surfaces,
Photogrammetric Engineering & Remote Sensing, Vol.69, No. 11, November, 2003,
pages 235–1242, ISSN: 0099-1112
Fuentes, O., Domínguez, R., deLuna, F., López, J. & Cruz, J. (2008). Documentación
Hidráulica del evento de 2007 que causó inundaciones en Villahermosa, Tabasco, In
Plan Hídrico Integral de Tabasco dentro del cual se Inscribe la Ejecución del Plan
de Acción Urgente (PAU) y la Formulación del Plan de Acción Inmediata (PAI),
para la Rehabilitación de la Infraestructura Dañada por las Lluvias Atípicas de los

Días 28 y 29 de Octubre del 2007, Ramos, J., & Vélez, L. (Ed.) pp. 1000, Instituto de
Ingeniería, UNAM, Report project, México, D. F.
GET, (2008). Gobierno del Estado de Tabasco, [Web page reviewed on April 2008]
http://www.tabasco.gob.mx/estado/index.php
Horrit, M.S. & Bates, P.D. (2002). Evaluation of 1D and 2D numerical models for predicting
river flood inundation, Journal of Hydrology, Vol. 268, November 2002, Pages 87-99,
ISSN:0022-1694
INEGI (2000). Síntesis de Información Geográfica del Estado de Tabasco, Instituto Nacional de
Estadística y Geografía, ISBN: 9701332733, Aguascalientes, Mexico
Jensen, J.R. (2000). Remote Sensing of the Environment: An Earth Resource Perspective,
Upper Saddle River, New Jersey, Printece Hall Inc., page 544.
Lijian, Z., Zulong, L., Yingcheng, L., Yanli, X., Ming, L., Zhouloei, W., Pei, L. & Xiaolong, L.
(2008). Application and analyses or airborne LiDAR technology in topographic
survey of tidal flat and coastal zone, The International Archives of the Photogrammetry,
Remote Sensing and Spatial Information Sciences, Vol. XXXVII, Part.B3b, Commision
III, page 56ff, July 2008, ISSN:1682-1750, Beijing
Maza, J.A. (1997) Cuenca Grijalva-Usumacinta Estudio de Gran Visión para las Obras de
Protección de la Planicie, elaborado para la Subdirección General de Construcción,
Gerencia Regional Sur, CONAGUA, y Subdirección Técnica, Gerencia de Estudios
de Ingeniería Civil, CFE. Report Project, Instituto de Ingeniería, UNAM, Mexico
D.F.
Maza-Álvarez, J.A. & Franco, V. (1997). Obras de protección para control de inundaciones,
Capítulo 15 del Manual de Ingeniería de Ríos, Vol. 591, pp.185, Instituto de Ingeniería,
UNAM, México, D.F.
Mohammadzadeh, A. & Valadan Zoej, M.J. (2008). A state of art on airborne lidar
application in hydrology and Oceanography: a comprehensive overview,
Proceedings of ISPRS Congress Beijing 2008, The International Archives of the
Photogrammetry, Remote Sensing and Spatial Information Science, Vol. XXXVII,
Part B1 Commission I, pp. 315-320, ISSN 1682-1750, July 2008, Beijing
Moss, T. & Monstadt, J. (2008). Institutional dimensions of floodplain restoration in Europe:
an introduction, In: Restoring flooplains in Europe: policy contexts and project
experiences, T. Moss and J. Monstadt (Ed.), 3-15, IWA Publishing, ISBN 1843390906,
London
NavCom (2008). SF-2040 Pole-Mount StarFire TM/RTK GPS Sensor, NavCom Technology
[Web page consulted on December 2008,
http://www.navcomtech.com/Products/GPS/sf2040g.cfm]
Oropeza, V.M. (2004). Parque Reserva, Península del Carrizal, Bachelor thesis, March 2004,
Universidad de las Américas Puebla, Cholula, Puebla, México
PEDUT (2007). Programa Estatal de Desarrollo Urbano de Tabasco del 2007, Gobierno del
Estado de Tabasco, Mexico
Raber, G.T., Jensen J.R., Hodgson M.E., Tullis, J.A., Davis, B.A. and Berglund J. (2007).
Impact of Lidar nominal post-spacing on DEM accuracy and flood zone
delineation, Photogrammetric Engineering and Remote Sensing, Vol. 73, No. 7, July
2007, pages 793-804, ISNN: 0099-1112
Raber, G.T., Jensen, J.R., Schill, S.R. & Schuckman, K. (2002). Creation of digital terren
maodels using an adaptive lidar vegetation point removal process, Photogrammetric
Engineering and Remote Sensing, Vol. 68, No. 12, pages 1307-1315, ISSN: 0099-1112
Stabbel, E. & Löffler E. (2003). Optimised mapping of flood extent and floofplain structures
by radar EO-methods, Proceedings of the FRINGE 2003 Workshop (ESASP 550), pp.
56.1, CDROOM, Franscati, December 2003, ESA/ESRI, Italy
Tamiru-Haile, A. & Rientjes, T.H.M. (2005). Effects of lidar DEM resolution in flood
modelling: A model sensitivity study for the city of Tegucigalpa, Honduras,
Proceedings of the ISPRS WG III/3, III/4, V/3 Workshop Laser scanning 2005,
September 2005, Enschede, The Netherlands
Tockner, K. and Stanford, J.A. (2002). Riverine flood plains: present state and future trends,
Environmental Conservation, Vol.29, Vol. 3, April 2002, pages 308-330, ISSN: 0376-
8929
Trimble (2008). GPS Pathfinder ProXH Receiver, Trimble Navigation Limited 2008 [Web
page consulted on December 2008, http://www.trimble.com/copyrights.shtml]
Webster, T. & Stiff, D. (2008). The prediction and mapping of coastal flood risk associated
with storm surge events and long-term sea level changes, WIT Transactions on
Information and Communication, Vol. 39, pages 129-138, ISSN 1743-3517
Webster, T.I. & Dias, G.. (2006a). An automated GIS procedure for comparing GPS and
proximal LIDAR evaluations, Computers and Geosciences, Vol.32, May 2006, pages
713-726, ISSN: 0098-3004
Webster, T.L., Forbes, D.L., Mackinnon, E. & Roberts, D. (2006b). Flood-risk mapping for
storm-surge events and sea-level rise using lidar for southeast New Brunswick,
Canadian Journal of Remote Sensing, Vol. 32, No. 2, April 2006, ISSN: 1712-7971
Weitkamp, C., (2005). Lidar: Introduction, In: Laser Remote Sensing, Fujii T. & Fukuchi T.
(Ed.), pages1-36, CRC Press, ISBN:0-8247-4256-7, Florida
Werritty, A. (2006). Sustainable flood management: oxymoron or new paradigm? Area, Vol.
38, No. 1, pp. 16-23, ISSN: 0004-0894
Improving Wetland Characterization
with Multi-Sensor, Multi-Temporal SAR and Optical/Infrared Data Fusion 679
33
X
Improving Wetland Characterization with

Multi-Sensor, Multi-Temporal SAR and
Optical/Infrared Data Fusion
Laura L. Bourgeau-Chavez, Kevin Riordan*, Richard B. Powell,
Nicole Miller** and Mitch Nowels***
Michigan Technological University, Michigan Tech Research Institute,
3600 Green Ct., Ste. 100, Ann Arbor, MI, USA
* Independent Researcher, 85B East Jefferson Rd. Pittsford, NY, USA
** Radiance Technologies, Inc. Bldg 1103, Suite 210 Stennis Space Center MS, USA
*** 8700 SW 174th Street, Miami, FL, USA
1. Introduction
Wetlands provide habitat and food sources for wildlife, protect waterways, act as natural
filtration systems, and serve major ecological roles in the overall health of our local, regional
and global ecosystems. While wetlands serve such vital ecological functions, due to their
limited capacity for adaptation wetland ecosystems are highly vulnerable to change, from
both climatic (IPCC 2008) and anthropogenic sources. In the past, wetlands have been
drained and converted to uplands at an alarming rate and the remaining wetlands are
subject to a number of threats including eutrophication, climate change, invasive plant
species, and the interaction between these stressors. Due to these vulnerabilities, there exists
a need for policy and resource managers to have an operational strategy for monitoring the
extent, composition, and vigor of wetlands at a synoptic scale. For regional areas, such as the
coastal Great Lakes, Boreal Canada, or vast wetland complexes such as the Pantanal,
Mesopotamian marshlands, or the Greater Everglades, cost-effective implementable
methods are necessary. For fine scale studies, cost is generally less of an issue and the
highest resolution data with the highest cost may be justified. In this chapter, we focus on
methods for regional scale mapping and monitoring where the minimum mapping unit of
interest is 5 acres, and thus use moderate resolution satellite imagery (20-30 m). The focus is
on multi-sensor data fusion between SAR single channel and/or multi-channel data and
Optical/IR sensor data. Case studies in the Great Lakes and Boreal peatlands demonstrate
the advantages and widespread utilty of a hybrid SAR-Optical/IR approach.
While many definitions of wetlands exist, both scientific and legal, in essence wetlands are
defined by: (1) the presence of water at, above, or near the ground surface; (2) hydric soils;
and (3) vegetation species adapted to or tolerant of wet soil conditions. Remote sensing can
be used to map and monitor two of the defining wetland features; vegetation type and
surface water/wet soils.
Wetlands have historically been one of the most difficult ecosystems to classify using
remotely sensed data. This difficulty is partially due to the high variability in wetland
morphology. Wetlands can exist in many shapes and sizes, from open wet areas with sparse
vegetative cover to densely forested areas with seasonal flooding. Vegetative cover ranges
from low herbaceous, to shrubby, to forest.
Traditionally, optical data have been used to map wetlands along with other land cover and
land use. However, due to the complexity of wetland ecosystems it would be beneficial to
include a fusion of sensors, operating in different frequencies (thermal, optical, lidar, radar,
infrared) that measure various aspects of wetlands for improved mapping accuracy. Optical
and infrared data are well suited to mapping vegetation ecosystem types and condition.
Complementary to Optical/IR, Synthetic Aperture Radar (SAR) data are capable of
detecting flooding beneath a vegetation canopy, monitoring water levels and soil moisture,
and also for distinguishing other biophysical vegetation characteristics such as level of
biomass.
2. Optical/IR Wetland Mapping Background

Multispectral data that includes near infrared and shortwave infrared bands allow
improved wetland detection and mapping over visible sensors alone. The near-infrared
portion of the electromagnetic spectrum has been used to identify plant and hydrologic
wetland conditions using both color infrared (CIR) aerial photography and satellite remote
sensing systems (Ozesmi and Bauer, 2002). The most broadly used wetlands map in the
United States, the National Wetlands Inventory (NWI), uses aerial CIR photography and
photo interpretation techniques to provide fine scale maps of wetland distribution (Peters,
1994). However, this labor-intensive methodology requires a 10-year repeat interval for new
map production (Wilen and Frayer, 1990).
To effectively monitor changes to wetlands, data collection must be timely (1-5 year
minimum) and cost effective. The National Oceanic and Atmospheric Administration’s
Coastal Change and Analysis Program (C-CAP) uses the Landsat Thematic Mapper (TM)
sensor to provide a more timely and cost effective national system of coastal wetland maps
(Klemas et al., 1993) on a 5 year repeat interval. However, both NWI and NOAA C-CAP
maps offer only broad classes of wetland, such as estuarine emergent or palustrine forested.
Finer classes of actual species or ecosystem types are not mapped.
Since various targets reflect and absorb solar radiation differently, they can often be
distinguished by their spectral reflectance signatures (Jensen, 2007). Spectral reflectance
studies have been useful for determining regions of the electromagnetic spectrum which
provide greatest discrimination between two or more wetland species (Schmidt & Skidmore
2003, Becker et al 2005). However, many studies have concluded that it is difficult to
accurately classify wetland species types based solely on Optical/IR spectral characteristics
(Ozesmi and Bauer, 2002).
3. Satellite SAR Wetland Mapping

SAR data have unique capabilities because the long microwave wavelengths penetrate
vegetation cover and are sensitive to wet soil and flooded conditions that may exist beneath
a canopy. An enhanced signature is often received from a canopy underlain by water due to
a double-bounce effect of the incoming radiation from the smooth water surface and vertical
stems of the canopy. The microwave scattering received by a SAR sensor from a wetland is
dependent upon the wavelength, polarization, and incidence angle at which the energy was
transmitted, the surface roughness, vegetative biomass, dielectric properties of the
vegetation and soils (moisture in the plant canopy and on the ground), and the presence or
absence of a flooded surface. Therefore, the SAR wavelength, polarization and incidence
angle need to be carefully chosen to maximize the scattering to distinguish wetlands from
uplands and to distinguish between wetland ecosystem types. A combination of
wavelengths, polarizations, and/or incidence angles provides the most information about
the various wetlands and thus the greatest capability to effectively map wetland ecosystem
types with SAR.
Current and recently orbiting SAR satellites available commercially are of three different
wavelengths; L-band (~23 cm wavelength); C-band (~5.7 cm wavelength), and X-band (~3.5
cm wavelength). Of these SAR satellites, many are of a fixed incidence angle, but some have
varying incidence angles. To detect flooding beneath a vegetation canopy, steep incidence
angles (<35 degrees) are generally best (Hess et al. 1990). For temperate, sub-tropical and
boreal regions, longer wavelengths such as L-band are more useful for mapping forested
and high biomass herbaceous wetlands than C-band or X-band. C-band or X-band data have
limited ability to map flooding beneath forest canopies. C-band data are most useful in
forests during leaf-off condition and for sparse canopies, and have been used to map extent
of inundation in floodplain swamps of Roanoke (Townsend 2001, 2002, Lang et al. 2008).
In Figure 1, theoretical scattering of a C-band sensor from forested versus herbaceous

landscapes in various dry to flooded conditions is shown. Here it is demonstrated how the
degree of inundation affects the scattering from the herbaceous canopy. In the case of wet
soils and low inundation in an herbaceous canopy, enhanced backscatter is often observed
at C-band, with some double-bounce effects. However, as the water levels increase the
backscatter can first get stronger and then lessen until it reaches a low specular reflection
case (where scattering is away from the sensor) from the water surface in the highly
inundated situation (Kasischke and Bourgeau-Chavez 1997, Kasischke et al. 2003, Bourgeau-
Chavez et al. 2005). Figure 1 also shows the typical scattering from a closed versus open
canopy forest at C-band. With most scattering from the branches and leaves at C-band in the
closed canopy case, and little to no penetration to the ground surface. However, the longer
wavelength L-band generally penetrates a closed canopy forest and has been found to be
best for discriminating flooded from non-flooded forests (Hess et al. 1990, Ramsey 1998,
Bourgeau-Chavez et al. 2001). C-band is best for discriminating emergent wetlands from
agriculture and herbaceous uplands (Bourgeau-Chavez et al. 2001).
SAR polarization is also important, and horizontal send and receive (HH) polarization has
been found to be most useful for detecting wetlands. While, the cross-polarizations
(Horizontal send and Vertical receive HV) are necessary for discrimination of woody versus
herbaceous vegetation types due to their sensitivity to biomass (Ramsey 1998). VV

polarization is also sensitive to soil moisture and flood condition (Bourgeau-Chavez et al.
2001).
Fig. 1. Diagram showing theoretical scattering of C-band energy from forested and
herbaceous ecosystems in dry, wet and flooded conditions, with open and closed forest
canopies.
One primary advantage of using SAR over visible data is the detection of forested wetlands
and the ability for SAR data to be collected irrespective of cloud cover or solar illumination
because it is an active sensor. It is very difficult to detect flooding beneath a forested canopy
with Optical/IR, unless there are large gaps in the canopy. Many researchers have found
significant improvement in distinguishing swamp from other wetland classes and uplands
with SAR (Lang et al. 2008, Grenier et al. 2007, Baghdadi et al. 2001, Hess et al. 1990,
Bourgeau-Chavez et al. 2004).
Several researchers have evaluated the utility of SAR for wetland mapping using single and
multi-date single channel SAR data (Costa et al. 1998, Whitcomb et al. 2009, Arzandeh and
Wang 2002, Rao et al. 1999) and others have evaluated polarimetric SAR (Hess et al. 1990,
1995, Bourgeau-Chavez et al. 2001, Pope et al. 1994, 1997, Crawford et al. 1999, Wang and
Davis 1997, Touzi et al. 2007). See Henderson and Lewis (2008) and Ramsey (1998) for a
more thorough review of past research on wetland ecosystem analysis with SAR. Many
early studies conducted to determine the utility of multi-polarization/multi-frequency data,
focused on NASA’s Shuttle Imaging Radar - C (SIR-C) which was fully polarimetric at L-
and C-bands (L-HH, L-HV, L-VH, L-VV, C-HH, C-HV, C-VH, C-VV and X-VV) and fully
polarimetric AirSAR (P-band (72 cm), L-band, C-band) for wetland classification in tropical
and temperate landscapes (Hess et al. 1995, Pope et al. 1994, 1997, Bourgeau-Chavez et al.
2001). These studies demonstrated the utility of multi-band data and early polarimetric
analyses (e.g. HH-VV phase difference) for mapping forested and herbaceous wetlands
(Pope et al. 1997).
While several researchers have evaluated the use of SAR alone for mapping wetlands, until
more recently, few have evaluated SAR and Optical/IR fusion for wetland mapping
(Lozano-Garcia and Hoffer 1993, Augustein and Warrender, 1998, Toyra et al. 2001, Rio and
Lozano-Garcia 2000, Bourgeau-Chavez et al. 2004, 2008, 2009, Li and Chen 2005, Grenier et al.
2007). Since the SAR and Optical/IR data measure different features of wetlands, it is logical
that a synergistic approach between the two sensor types would increase wetland mapping
accuracy. Further since the presence of standing water causes the SAR energy to interact
differently depending on the dominant vegetation type, it would be advantageous to
combine SAR with optical and infrared data for mapping purposes.
4. SAR and Optical Data Fusion

Of the few broad-scale SAR wetland mapping efforts that have been undertaken, , Canada is
incorporating Radarsat-1 and Radarsat-2 data with Landsat mosaics (100m resolution) and
SPOT data in the development of the Canadian Wetland Inventory (Grenier et al. 2007,
Fournier et al. 2005, pers. comm. Grenier 2009) which will cover the entire country. There is
also an effort underway to use the JERS mosaics (100 m resolution summer and winter
products) of Boreal North America alone to map wetlands across Canada, as has already
been done for Alaska (Whitcomb et al. 2009).
In this chapter, we review case study of multi-sensor, multi-date, SAR-optical/IR fusion

methods and results for mapping wetlands in two main study areas, the Great Lakes and a
Boreal peatland region in Alberta, Canada. The techniques are developed with broad scale
mapping in mind, but are demonstrated on local to regional wetland areas. In all but one of
these case studies, satellite SAR data are used in conjunction with Landsat TM, and in most
cases SAR data from multiple sensors are fused. The exception, is the case studies on
mapping the invasive plant species Phragmites australis on Lake St. Clair, where PALSAR
data are used alone and compared to hyperspectral AVIRIS.
4.1 Case Study: Lakes Ontario and St. Clair Coastal Wetland Mapping
The Great Lakes Coastal Wetlands Consortium (GLCWC) was mandated to develop a
monitoring plan for assessing the health of the coastal wetlands which are vital to the
overall health and maintenance of the Great Lakes ecosystem (Bourgeau-Chavez et al. 2004,
8). The only way to monitor an ecosystem the scale of the Great Lakes basin is through
integrated remote sensing and field observations in a GIS. Landscape indicators in need of
monitoring through remote sensing include wetland type and extent, adjacent land cover,
adjacent land use, intensity of land use, and invasive plant species. The GLCWC sought
implementable, cost-effective yet robust methods with a minimum mapping unit of 5 acres.
To meet these needs a few pilot studies were conducted to demonstrate various monitoring
methods. A hybrid SAR Optical/IR methodology that used satellite sensor data of 30 m
resolution and would allow for detection of areas as small as 1 hectare was found to be the
most promising. This methodology would be cost-effective and data management for the
entire Great Lakes Basin would be reasonable compared to higher resolution, smaller
footprint imagery. And the use of two complementary data types in the mapping was
expected to reduce omission and comission errors.
Two study sites were selected for demonstration of this hybrid SAR-optical data fusion,
coastal areas of Lakes St. Clair and Ontario (Figure 2). Coastal Lake St. Clair was chosen
because it has a diversity of land cover and land use including a large amount of urban and
suburban areas, rural farm fields, and a large amount of wetlands (various species) that
occur at the delta as the river enters the lake. The Lake Ontario study area was chosen
because it is mainly rural, with many agricultural fields, has isolated patches of herbaceous
wetlands, and a relatively large amount of potentially forested wetlands. These two areas
provide different sets of land cover and land use classes and thus an opportunity to test the
hybrid-sensor methodology in diverse settings.
4.1.1 Remote Sensing Data

The data used in this case study were archival SAR and Landsat data from multiple dates,
and multiple sensors (Table 1). For the Landsat images, each containing 7 bands, the blue
and thermal IR bands for each date were removed from the analysis. The blue channel is
generally quite noisy and the thermal band is of coarser resolution than 30 m. The JERS (L-
band) sensor used for this project had horizontal send and receive polarization (L-HH) and
was operational from 1992-8. This sensor has a resolution of 30 m, incidence angle of 35º,
and a footprint of 80 km x 80 km. To complement these data, we also acquired C-band (5.7
cm wavelength) SAR data from the European ERS-1 and Canadian Radarsat-1 (R-1)
satellites. The ERS-1 sensor has vertical send and receive polarization (C-VV) and is
collected at a central incidence angle of 23º. The Radarsat sensor has horizontal send and
receive polarizations (C-HH). It also has pointing capabilities and can be collected in various
modes and incidence angles. The R-1 data used were of standard beam 7 mode, which has
an incidence angle of 47º. Both C-band sensors have 30 m resolution and 100 km x 100 km
footprints.
Fig. 2. Great Lakes pilot study areas for hybrid SAR-Optical/IR mapping for the Great Lakes
Coastal Wetlands Consortium project.
Site Sensor Spring Summer Fall

Lake St. Clair Landsat 23 March 2001 30 August 30 October 2000
TM 2001
Lake St. Clair JERS 28 March 1995 10 August
1998
Lake St. Clair Radarsat-1 3 & 27 October 1998
Lake Ontario Landsat 24 June 1993 12 August 18 December 2002
TM 2002
Lake Ontario JERS 11 April 1993 8 July 1993 17 October 1993
Lake Ontario ERS-1 17 April 1993 7 June 1993 25 October 1993
Table 1. Data used in the Great Lakes Hybrid SAR-Optical/IR wetland and landscape
indicator mapping.
Since this was a demonstration project with minimal funding, data sets chosen were not
optimal but were chosen based on cost and ease of availability (Table 1). At both sites three
dates of Landsat data were used. There were a total of four radar scenes for Lake St. Clair
and 6 radar scenes for Lake Ontario. There are as many as 6 years between the JERS and
Landsat collections for Lake St. Clair and 10 years between data collections for Lake Ontario,
but analysis of the imagery indicated that few major changes in land use have occurred in
these areas during that time period. Also, it was not as much of an issue due to the
methodology; a site would first be checked for vegetation cover in the more recent Landsat
and then checked for flooding in the older SAR. However, it is realized that using data with
such a wide time span will introduce errors. The optimal data set would be from the same
year, with SAR and TM from the same months. However, the datasets met the needs of this
investigation, which was simply to demonstrate a methodology.
4.1.2 Image Interpretation

For the Lake St. Clair site, the imagery and products were evaluated by comparison to the
NWI, land cover/land use maps, field checks (October 2003) and expert field knowledge
(Field ecologist/botanist Dennis Albert of Michigan Natural Features Inventory).
Although the ancillary maps and field work showed many forested wetlands within our
study area, the dates of JERS imagery show all of the forests the same, very bright in the
spring (March 1995) and all are gray in the summer scene (August 1998). It is likely that all
of the forests have a wet ground cover in the spring scene, there may even be wet, melting
snow on the forest floor causing the enhanced signature from all of the forests, and in
August all of the forests are dry with full foliage. However, the differences in backscatter in
the herbaceous vegetation are apparent on these two dates, as well as in the October
Radarsat scenes.
Figure 3 shows a red-green-blue false color composite of the 3 October 1998 R-1 scene, 10
August 1998 JERS scene, and 28 March 1995 JERS scene, respectively. In this composite, the
orange areas were dominated by cattail (Typha spp.), and the green by Phragmites
(Phragmites australis). Phragmites tends to be taller/denser and occurs in less wet locations
than cattail. The red areas of the image are shorter and sparser vegetation, thus they do not
cause enhanced backscatter at L-band, only at C-band. The red areas along the fingers of the
delta are cattail and bulrush (Scirpus spp.) beds and the red area within Dickinson Island is a
flood channel with wild rice (Zizania aquatica), open submergent and emergent vegetation
(Dennis Albert). The dark area in the center of Dickinson island to the west of the kidney
shaped light blue forest area is a wet meadow and appears to be dry in our October 1998 C-
band scene, it has strongest backscatter in the L-band spring scene (blue channel), but not
enhanced backscatter. This combination of R-1 and JERS allows for a good interpretation of
this scene, discerning tall dense herbaceous vegetation from short sparse herbaceous
vegetation, and different hydrological and biophysical properties.
Radarsat Oct 98 JERS Aug 98 JERS Mar 95

Dickinson Island Harsen’s
Island
Upland
Cattail/Scirpus
22.7 km
Lake St.
Clair Wet meadow-
Phragmites
Cattail dominant
©CSA 1998 ©NASDA 1995-8
Fig. 3. Three dates, two sensor false color composites from Radarsat-1 and JERS satellites
over the Lake St. Clair Delta. This Figure clearly defines Typha (cattail), and the invasive
species Phragmites australis from other upland and wetland ecosystems.
At the Lake Ontario test site, we had an ideal seasonal data set with three images each of
JERS and ERS from spring, summer and fall of 1993. Figure 4 presents false color composites
of the two sensor datasets. Both datasets highlight non-forested wetlands (based on NWI) as
green and red shades. In the JERS data, these sites were dark from specular reflection in the
April scene (blue), then some sites were bright in July (green in the composite) while other
sites remained dark in July (red locations in the composite) and all sites were gray in
October (Table 2). In November of 2003, we conducted a field check to determine any
vegetation difference between the red and green areas. The red areas visited were
dominated by mixed grasses. It is likely that in the spring imagery the vegetation is fallen
over or decomposed, with a high water level leading to specular reflection. The water level
must still be high enough to cover much of the grasses and cause specular reflection again in
the July scene, but when the lake water level drops to 74.58 m in October (70 cm drop), this
site has more vegetation exposed and stronger backscatter. In comparison, the green areas
visited in the field contained a mixture of grasses, cattail and shrubs (mainly Cornus
stolonifera). These sites were bright in the summer and gray in the fall in the JERS imagery.
The water level in comparison to the vegetation was likely lower than at the other sites,
causing the enhanced backscatter in July, but with lower soil moisture in the fall the site was
gray. For the same two sites in the ERS, the grass site was dark in the spring but the mixed
shrub/herbaceous site was gray. In the summer the mixed shrub site was bright while the
grass site remained dark. In the fall all sites were gray. While similar patterns emerged for
both sites, the contrast between the non-flooded adjacent forests and the wetlands is
stronger with the JERS, making it easier to map the boundaries of the sites.
Three date ERS false color composite Three date JERS false color composite
17.2 km
©ESA 1993 19.3 km
Fig. 4. Three date ERS false color composite of 25 October, 7 June and 17 April 1993 ERS
imagery over eastern Lake Ontario compared to a 17 October, 8 July and 11 April 1993 JERS
composite.
Spring Summer
Site Sensor Brightness Brightness Fall Brightness
Grass ERS-1 C-VV dark dark gray
Shrub/herbaceous ERS-1 C-VV gray bright gray
Grass JERS L-HH dark dark gray
Shrub/herbaceous JERS L-HH dark bright gray
Table 2. Appearance of “grass” versus “shrub/herbaceous” sites in coastal Lake Ontario in
spring, summer and fall of 1993. Grasses are red areas in JERS composite of Figure 3 and
Shrub/herbaceous are green areas.
There are some forested wetlands within the Lake Ontario scene and they appear to be most
notable in the April scene when the lake water level is the highest (note that coastal
wetlands are hydrologically connected to the Great Lakes), and spring thaw has occurred
and thus flooding is most likely. A comparison was made between assumed flooded forest
and non-flooded forest for each JERS scene/date. The April scene had a 2.3 dB difference
between flooded and non-flooded forest while the July date had only a 0.5 dB difference and
the October date had a 1.7 dB difference. The April scene was then thresholded to values
greater than that of the non-flooded forest. After median filtering the scene with a 5x5
window to remove speckle, the scene was overlaid on a 5, 4, 3 Landsat composite (Figure 5).
The white areas of Figure 5 show the SAR-derived potentially flooded forests. The
backscatter from urban areas is also enhanced and has not been filtered from this scene.
There are also white areas that are likely row plantings of trees. The row structure produces
an enhanced return. The urban areas can be removed by using either the Landsat scene to
mask forest from non-forest or by using the ERS C-band data. The C-band data will have
enhanced backscatter for the urban area but not for the flooded forests. The extent of some
of the enhanced signatures appears to be slightly greater than what is seen in the NWI for
some of the sites (Figure 5) and in other cases it is less.
Forested 73.9 km
Upland
Emergent
Open Water
Fig. 5. Landsat 5, 4, 3 composite with SAR-derived potentially flooded forests (white areas)
overlaid 9 (left). The NWI is presented for comparison (right) with forested wetlands
colored green. Note that the NWI is not the exact area of the Landsat scene.
One technique that has been used somewhat operationally for detecting forested wetlands,
is to map forest cover with Optical/IR and then create either a single-date thresholded
image, as is presented here, or a multi-date SAR image to determine inundation and thus,
forested wetland. Using remote sensing to detect flooding beneath a canopy is limited by the
timing of the data acquisitions. Therefore, only the forests that were inundated on the date
of image acquisition will be mapped as forested wetland. The multi-season data helps
reduce this limitation, but relying on species type and other ancillary data such as hydric
soils, as well as enhanced SAR backscatter to determine wetland status would be more
reliable, hence hybrid approaches are recommended. One useful parameter the SAR can be
used for is to monitor changes in extent of inundation from one date to the next (Bourgeau-
Chavez et al. 2005, Lang et al. 2008, Townsend 2001, Wang et al. 2004).
4.1.3 Mapping Methodology

Several techniques were considered and evaluated for merging the multi-sensor, multi-date
Great Lakes datasets and producing a land cover land use map, but the best method
appeared to be the simplest, which involved creating separate SAR and Landsat land
cover/land use products and then fusing the products in a GIS. This method preserved the
unique characteristics of each sensor, while taking advantage of their complementary nature.
4.1.3.1 Lake St. Clair Study Site

The two-step methodology was first applied to the Lake St. Clair study site where the 15-
band Landsat layer was classified into 12 different land cover features using the maximum
likelihood classification algorithm. The training sites for this supervised classification were
collected from the reference field data and existing maps. The 12 classes created were:
Water, High Density Urban, Low Density Urban, Forest, Emergent Wetland, Wetland
Shrub, Wetland-permanent, Forage Crops, Cropland 1, Cropland 2, Cropland 3, Cropland
4. The second step of the process involved the classification of the 4-band radar image into 9
different classes. These classes were different than that of the Landsat imagery because the
sensors are able to distinguish different phenomena. The classes for the radar classification
were: Forest/Urban, Phragmites, Scirpus/ open submergent and emergent, Cattails, Wet
Meadow, Forage Crops, Cropland1, Cropland2, and Cropland3
The individual classification results were then fused into a single classification. This was
performed by recoding the 12-class Landsat classification into values of 10s (10, 20, 30…120)
and recoding the radar classification into values 1-9. Then the values were added together
on a pixel-by-pixel basis, producing possible values between 11-129. These recombined
classes were then interpreted, by comparison to the reference data, and assigned into a final
11-class file. This final class identification relied on only the Landsat for some classes (such
as water, forest, low density urban), only the radar for others (Phragmites and Scirpus), and
a combination of both for the majority of the classes. The final classes are described below in
Table 3 and the final map is in Figure 6.
Due to limited funding and the archival nature of the dataset, an accuracy assessment was
conducted based on existing maps. The NWI, which is the basemap for the GLCWC, was
first used as a reference. Then IFMAP (Integrated Forest Monitoring, Assessment and
Prescription), which relies on NWI to some extent, was used because it includes upland
classes that NWI does not. IFMAP was produced by the Michigan Department of Natural
Resources. It is mainly based on the analysis of seasonal Landsat imagery (collected between
1997-2001), but is supplemented with selected high resolution images, existing land cover
maps, and large amounts of field work. IFMAP provides a very detailed description of land
cover, but is only available for the Michigan portion of the study area and accuracy could
therefore only be assessed on the U.S. side of the map. NWI and IFMAP have only broad
wetland categories (e.g. palustrine emergent, scrub-shrub, etc).
Class Description
Water Identified through the Landsat, regardless of the radar results
Identified by an urban class in the Landsat imagery and a
High Density Urban
urban_forest (bright) from the radar class
Low Density Urban Identified through the Landsat, regardless of the radar results
Identified through the radar classification and reinforced by
Scirpus being classified as a wetland/vegetative class in the Landsat
imagery
Phragmites being classified as a wetland/vegetative class in the Landsat
imagery
Cattail being classified as a wetland/vegetative class in the Landsat
imagery
Identified as wetlands in the landsat imagery but is different
Wetlands_other
than Scirpus, Phragmites, and cattail
Identified through a Landsat class of shrubland and forest and
Shrubland (shrub wetland)
has a radar classification of wet meadow or shrubland
Identified by the four cropland classes from the Landsat imagery
Cropland
and the three cropland classes from the radar classification.
Identified by landsat imagery (forage, row crop) and radar
Forage Crops/Low Herbaceous
imagery (forage)
Identified through a Landsat imagery forest class (Note: if the
radar were from a time when forested wetlands could be
Forest
identified, this landsat class would be combined with a radar
class)
Table 3. Combined Landsat and radar landcover results for the Lake St. Clair study site.
Using over 3000 randomly selected validation points, comparison to the NWI as reference,
resulted in 94% overall accuracy of our wetlands map. Comparison to IFMAP resulted in
72% overall accuracy when areas of wetland that IFMAP called “open water“were
eliminated. Analysis of the imagery revealed that the timing of data collections and the lake
levels can have a large effect on the boundaries mapped for emergents along the water’s
edge. A SAR comparison of 2 image dates with a change of 19 cm in lake level showed a
huge change in visibility of wetlands on the fingers of the St. Clair river delta (Figure 7). In
this figure, lake water level is 19 cm higher on the first date, causing specular reflection (low
return-dark). A decrease in inundation on the second date reveals the vegetation causing
double bounce scattering (bright return-red). This exemplifies the need for multi-date data
to “see” the wetlands that may be nearly completely inundated by water on a particular
date in both SAR and optical/IR.
In comparison to IFMAP, the hybrid SAR-Optical classification did well with low density
urban, Typha, Phragmites, low herbaceous, Scirpus and cropland (all above 60% user’s
accuracy, Typha above 86%). There were some issues with wet meadow (42.25% user’s
accuracy) where there is confusion in IFMAP with tree species, row crop and herbaceous
upland. Further investigation of this type is needed. The classes were quite different for
IFMAP, but for lowland deciduous, the SAR-Optical map had 81% producer’s accuracy, 65%
and 77% producer’s accuracy for emergent and non-forested wetland and 78% producer’s
accuracy for low density urban.
A visual comparison was also made between our classification of the Canadian side of the
study area using maps produced by Arzandeh and Wang (2002 & 2003) of Walpole Island,
Ontario. In 2002 Arzandeh and Wang used a single Radarsat scene (1997) and Landsat data
(1997), separately, to create two classification maps with eight categories including forest,
urban, swamp, tall grass, water, agriculture, cattail and Phragmites. Their areas of emergent
wetland (cattail and Phragmites) correspond well with the areas that we have mapped as
emergent. However, their maps lack the detail that we gained by combining multiple dates
and two bands of SAR imagery with the Landsat. Their goal was to use texture analysis of a
single date of SAR imagery to improve classification accuracy with a single date of SAR.
Generally, more than one date of imagery is available, but for those cases when only one
date of imagery exists or can be acquired their techniques would be useful.
Figure 6. Landsat and SAR fused land cover classification results for the Lake St. Clair study
area.
Two date false color composite of Radarsat Imagery Lake St. Clair
3 October 1998 27 October 1998
Fig. 7. Two date false color composite of Radarsat imagery. Cyan is the 3 October 1998
image and red is the 27 October 1998 image. Water level dropped by 19 cm between the first
and second date.
4.1.3.2 Lake Ontario Study Site

The same techniques used for Lake St. Clair were applied to the datasets for Lake Ontario,
however the classes were slightly different. First, Landsat imagery was classified into 13
different landcover classes; Water, High Density Urban, Low Density Urban, Deciduous
Forest,Coniferous Forest, Mixed Forest, Row Crops, Low Herbaceous, Bare Soil, Emergent
Wetlands, Shrubland, Fields-Hay, Wetlands-other. The radar imagery was classified into 9
classes; Water, Urban/Flooded Forest, Flooded Shrubland/ Wet Meadow, Emergent
Wetland, Forest1, Forest2, Row Crop 1, Row Crop 2, and Herbaceous Field. These
individual Landsat and SAR classes were then fused into a single product, just as they were
for Lake St. Clair. The 12 final combined classes are described in Table 4. The final map is
presented in Figure 8.
A comparison was made between the SAR-Optical map and the NWI with 5000 random
points Note that the NWI has more generalized classes so the comparison is not straight
forward, and there are over 2 decades between the NWI creation and the SAR-Optical/IR
map, so some differences may be due to changes in the wetland, either succession or loss.
The results in comparison to the NWI showed 94% overall accuracy, with all classes greater
than 89% user’s accuracy, except shrubby wetland for which we only had 13 points, none of
which were correctly classed. The producer’s accuracy in the wetlands was 34% for woody
wetland and 66% for emergent. For the NWI woody wetlands, we labeled 127 out of 208 as
deciduous forest. The problem likely lies in what areas were in fact inundated during the
radar satellite collections. A wet, normal or dry year will provide different wetland extents,
66% agreement for emergent wetlands is quite good considering the likely turnover of some
areas to agriculture and the likely succession of some of the wetlands labeled as emergent in
the 1970s NWI to wetland shrub, as indicated by the field visits and point source field data
from a biocomplexity study conducted by Cornell University (Mark Bain).
Class Description
Water Areas identified as water in the Landsat imagery, regardless of the radar classification
High Density Areas dominated by manmade materials, these areas were identified by high radar
Urban returns that were not forested areas in Landsat
Low Density Urban Areas with a mixture of manmade features and landscape vegetation, these areas were
mainly identified by the Landsat classification
Deciduous Forest Areas of forest that lose their leaves throughout the season, identified through a
combination of the Landsat and radar classifications
Coniferous Forest Areas of evergreen forest, identified through the combination of classifications
Mixed Forest Areas that are a mixture of coniferous forest and deciduous forest, identified through a
combination of classifications
Forested Wetland Areas are forest but have standing water on the ground throughout much of the year,
identified as forest in the Landsat imagery and as urban/flooded forest in the radar
imagery
Emergent Wetland Areas of herbaceous vegetation that are wet at some times of the year, identified
through the combination of radar classification and Landsat classification
Wetland-shrub Areas of short woody vegetation that are wet at some points throughout the year, these
are identified through the combination of sensor classification results.
Crop/pasture Areas of herbaceous vegetation that are not plowed throughout the year, mainly
identified through the Landsat classification
Bare Soil Areas of exposed soil, sand, and/or rock, these areas were mainly identified through
the radar and were confirmed by the Landsat classification
Row Crop Areas that have herbaceous vegetation growing which is plowed at some point during
the season. These areas were identified through a combination of the classification
results.
Table 4. Classes for the combined Landsat and radar classification at the Lake Ontario study
site.
The state of New York did not have a land cover/land use map comparable to IFMAP, and
many errors were found in the National Land Cover Dataset (NLCD). However, field data
collected in the largest coastal wetland complex in the image were available with GPS
locations. The Biocomplexity study of Cornell University allowed for comparison of the
SAR-Optical/IR hybrid map to 55 study points which represented cattail dominated, shrub,
forested, and mixed emergent wetlands. The overall accuracy of this comparison was 89%,
with 91% user’s accuracy for wetland shrub, 89% user’s for emergent wetland, and 67% for
forested wetland. This assessment is quite good considering the likelihood of error in the
geolocation of the study points due to the small plot size in reference to the 30 m pixels.
Some of the points did fall on boundaries of open water/wetland or upland/wetland
causing errors. The producer’s accuracy ranged from 25 to 75% in the reference categories of
specific species types that we did not map. The cattail field points corresponded to the
wetland shrub (dark pink) areas in the Landsat-SAR fused classification. Therefore, the pink
areas of our Landsat-SAR map should be labelled as shrub/high biomass herbaceous
wetlands.
Fig. 8. Landsat and SAR fused land cover classification of the Lake Ontario study area.
4.1.4 Summary
These two study areas demonstrated the utility of a fusion of SAR and Optical/IR data for
mapping landscape indicators of wetland health (wetland extent, adjacent land use
intensity, etc.) in a region surrounded by high intensity urban versus a more rural area. In
one case mapping of forested wetlands was possible and in the other, timing of data
collections did not allow evaluation of forested wetland mapping. However, a simple
approach to the mapping provided desirable results with the archival data, showing the
complementary nature of the two types of sensors. Although the reference data and remote
sensing data were not optimal [there are discrepancies in years of data collection (Landsat
versus SAR (1990s-2001) and years of validation maps (1970s, 2000) and levels of vegetation
classes between the SAR-Optical map and validation maps (fine scale species versus broad
emergent classes)], these case studies demonstrated the added benefits of fusing Optical/IR
data with another complementary sensor such as radar and resulted in a recommendation
by the Great Lakes Coastal Wetlands Consortium for monitoring landscape indicators
(Bourgeau-Chavez et al. 2008). In the next case study, current data are used and ongoing
validation is being conducted for the specific species class levels. This case study is a
continuing investigation and only preliminary results are shown here, however it provides a
better validation of the results through field methods, further exemplifying the utility of
SAR in wetland mapping.
4.2 Case Study: Invasive Phragmites australis Species Mapping on Lake St. Clair
One of the main wetland stressors in the Great Lakes is invasion by the problematic species
Phragmites (Phragmites australis). This species invades native habitat creating dense thickets
and deep detritus that virtually eliminates ecological function. A predicted drop in Great
Lakes water levels due to global climate change is anticipated to increase the spread of the
invasive Phragmites in the Great Lakes coastal zone, and a method to map this species and
its spread would be of great assistance to land managers for control.
Several studies have focused on detecting and mapping invasive species in small
catchments of the Great Lakes with high resolution hyperspectral and/or lidar (e.g. Lopez et
al. 2006, Wilcox et al. 2003). However, such high-resolution mapping of the entire Great
Lakes coastline or comprehensive field studies would be very costly. Others have found 30
m satellite imagery including Landsat, SPOT and Hyperion to be useful for mapping
invasives (Arzandeh and Wang 2003, Pengra et al. 2007), however Landsat has spectral
limitations and Hyperion is no longer operational.
Using a variation of the satellite SAR techniques described in the last section, which
included a delineation of this invasive species using archival multi-date JERS and Radarsat-
I, we are currently evaluating dual polarization PALSAR data, and have plans to
incorporate Radarsat-2 data once it becomes available to distinguish a wider range of
species.

ALOS PALSAR is the follow-on to JERS which showed the greatest potential in previous
studies for mapping Phragmites (Bourgeau-Chavez et al. 2004, 2008). PALSAR was launched
in 2006, and is available in three modes with various imaging parameters. Here we evaluate
the dual-polarization product which has 20 m resolution, two channels (L-HH, L-HV), 70 x
70 km footprint, and is collected at an incidence angle of 34°. Up until recently, most satellite
SAR systems were of a single channel, however with the recent launch of ALOS PALSAR
and Radarsat-2, the utility of multi-channel SAR and polarimetric data are beginning to be
more broadly evaluated for a variety of applications, demonstrating further mapping
capabilities beyond that of single and multi-channel data.
Coincident to this evaluation is an investigation of an airborne hyperspectral NASA AVIRIS

collection from July 2008 over the St. Clair delta. The AVIRIS sensor has 224 bands (400-2500
nm) and was collected with 17 m resolution.
HH/HV PALSAR Band Ratio
14
12
10
8
dB
0
Phrag 1 Phrag Cattail Wet Scirpus- Scirpus
Dickinson Harsens meadow St.Johns beds
Fig. 9. Plot of backscatter from PALSAR (L-HH to L-HV ratio in dB) in Phragmites, Cattail,
wet meadow and Scirpus beds of the Lake St. Clair delta.

Initial observations of an October 2007 dual polarization PALSAR image over Lake St. Clair
revealed that the ratio of the L-HV and L-HH bands shows a 4-5 dB difference between
Phragmites dominated wetlands and other non-forested native wetland types (Figure 9).
The reason for this great divergence is the large difference in vegetation height, density and
biomass of invasive Phragmites versus any other native herbaceous vegetation in the Great
Lakes (Figure 10). Typha generally ranges in height from 1 to 3 m, while Phragmites can
reach heights of more than 3.5 meters. Further, Phragmites forms tall, dense rather
impenetrable stands. It is the sensitivity of L-band SAR to these differences in biomass and
hydrology that allows the distinction between stands dominated by these two species.
Multi-date composites of PALSAR imagery from 2006-8 show the dynamic changes in the
various vegetation cover-types over the growing season (Figure 11). In the top composite of
Figure 11, which is a false color multi-date L-HH representation, the Typha are yellow to
orange, indicating there is a strong return signal in July and October, but low response in
the spring (Table 5). In contrast, most of the Phragmites shows a high response in the spring
and lower in the summer and fall with shades of blue, and purple. In comparison, in the
lower image of the L-HV multi-date false color composite (L-HV is sensitive to biomass),
the Phragmites is cyan, indicating a strong return in the spring, April and May images,
while the Typha is dark in these months, and bright (red) in the fall image.
Phragmites Typha
Fig. 10. Photos of Phragmites: tall scale in winter (left), high density in summer (middle),
and typical Typha (right).
While either L-HH or L-HV multi-date imagery appear to be useful for distinguishing
Phragmites from Typha, there are some stands that would be confused with the L-HV alone
(see arrows with Phragmites label on Figure 11); these are areas of Phragmites that are red in
the lower image and green in the upper image. The different signatures in the PALSAR
imagery from Phragmites-dominated stands is likely due to differences in water levels in the
various seasons. “Phragmites 2“ stands are located in diked areas, and may be wetter and
sparser than “Phragmites 1“ stands (Table 5).
PALSAR Image Date Phragmites 1 Phragmites 2 Typha

Band appearance appearance appearance
L-HH 28 July 2006 dark dark bright
L-HH 09 October 2007 dark bright bright
L-HH 17 April 2008 bright dark dark
L-HV 09 October 2007 dark bright bright
L-HV 26 May 2008 bright dark dark
L-HV 17 April 2008 bright dark dark
Table 5. Appearance of two different Phragmites dominated stands versus Typha stands in
the L-HH and L-HV imagery on the various dates.
4.2.3 Mapping Phragmites with SAR

A simple unsupervised maximum likelihood classification of the four dates of PALSAR data
resulted in four classes of potential Phragmites and two potential classes of Typha spp. Note
that the July 2006 image was from the single channel mode of ALOS PALSAR and thus had
only the L-HH channel, analysis was therefore conducted on an input of 7 channels. Field
observations using a GPS positioning system, in situ photos, and field notes were used to
assess the preliminary “potential Phragmites“ map. Using these field data for validation (29
points), the PALSAR multi-date preliminary map had 92% overall accuracy, with 100%
Lake St.
Clair
Harsens
Island, USA
L-HH 3 Date Composite

28 July 2006
09 Oct. 2007 Phragmites
17 April 2008
Wapole
Island, CA
L-HV 3 Date Composite

09 Oct. 2007
26 May 2008
17 April 2008
Phragmites
Fig. 11. False color composites of PALSAR imagery over Lake St. Clair delta. Top is L-HH
from 28 July 2006 in red, 09 October 2007 in green and 17 April 2008 in blue. Bottom is L-HV
composite with 09 October 2007 image in red, 26 May 2008 in green and 17 April 2008 in blue.
user’s and 80% producer’s accuracy for Typha, and 82% producer’s and 100% user’s
accuracy for Phragmites. Note that the misclassified pixels for Phragmites were small areas
of shrub or Typha within a larger Phragmites dominated area, thus the error is likely due to
resolution (20 m in this case). The utility of the 10 m resolution PALSAR product (although
only L-HH) may resolve this error and is being investigated.
4.2.4 Mapping Phragmites with AVIRIS

A comparison of the Optical/IR spectral signatures of Typha latifolia and Phragmites australis
are shown in Figure 12. These signatures were collected in the field using a
spectroradiometer (FieldSpec 3 JR). The vast differences in these signatures indicate that
separation using Optical/IR remote sensing should be fairly easy, however, using the
spectral angle mapper technique the results were poorer than the SAR methods.
0.30
0.25
0.20
Reflectance
0.15
0.10
0.05
0.00
350 450 550 650 750 850 950 1050 1150 1250 1350 1450 1550 1650 1750 1990 2090 2190 2290 2390 2490
Wavelength (nm)
cattail nadir phragmite nadir
Fig. 12. Spectral signature comparison of Typha latifolia (black) versus Phragmites australis
(red).
Spectral angle mapping is a physically-based spectral classification that uses n-dimensional

(n = number of bands) angles to match pixels to reference spectra. It was run in ENVI, and
determines similarity between the two spectra by calculating the angle and treating them as
vectors in n-dimensional space. Smaller angles represent closer matches to the reference
spectrum. The endmember spectra for Phragmites australis and Typha spp. used by the
classifier were derived from field truth data provided by Michigan Natural Features
Inventory.
Using methods similar to Lopez (2006), the spectral angle mapping of the 224 band (400-
2500 nm), 17 m resolution AVIRIS data using field training data, resulted in 82 % overall
accuracy with 83% user’s and 77% producer’s accuracy for Typha, but only 57% user’s and
67% producer’s accuracy for Phragmites. Note that the validation is based on only 17 points,
because many of the points were used in the training
These are preliminary results and the investigation is ongoing. Methods are being
investigated to determine the best bands from AVIRIS to combine with the L-band SAR for a
data fusion approach. Evaluations will reveal the utility of SAR alone and in combination
with hyperspectral, but the goal is to also determine if any of the bands of existing satellite
Optical/IR systems could be used in lieu of hyperspectral data.
4.3 Case Study: Peatland mapping in Alberta

The last case study is based on the need to understand carbon storage (peat accumulation)
and loss (mainly through fire) in boreal peatlands, which are widely recognized as being
one of the largest terrestrial reservoirs for carbon (C) in the Northern Hemisphere.
Estimating carbon storage and release requires an accurate mapping of peatland type.
Peatlands are defined as wetlands with well developed peat (partially decayed plant matter)
accumulation, generally more than 30 cm deep (Charman 2002). Peatlands actually

represent diverse ecosystem types that differ in hydrology and vegetation, from forested
rain-fed bogs to grass-dominated, saturated, or near-saturated stream-fed fens.
Early research on the ability to map boreal peatlands at a regional scale demonstrated the
utility of merging SAR and Optical/IR data. Early observations included JERS, R-1, and
Landsat imagery. Figure 13 presents Landsat and JERS images in comparison to a detailed
peatland map (based on air photo interpretation and intensive field truth circa 1970-80s)
with open, forested, and wooded categories of bogs, fens and swamps delineated. This
preliminary analysis showed that Landsat would be useful for finding many of the open
fens (see linear features circled in yellow; Figure 13). Whereas wooded fens do not look
different from bogs in the Landsat image, they can be distinguished in the JERS SAR multi-
date imagery (see features circled in pink; Figure 13). Note that the majority of peatlands in
this study region are wooded bogs (salmon color in the peatland map). The linear open fens
are dark in the SAR similar to the open swamps and marshes. Here, C-band Radarsat
should prove useful in distinguishing types of open peatlands. The complementary
information obtained from the spectral reflectance properties of the vegetation combined
with the structural and moisture information from SAR should allow the mapping of both
peatland type (bog, fen, swamp) and level of biomass (open, sparse tree cover, forested).

Two dates of PALSAR imagery were obtained over the local Alberta study area from July
and August of 2007. The dual-polarization PALSAR data included two channels, L-HH and
L-HV. To complement this, a spring and summer data set of Landsat TM imagery were also
obtained from April and August of 2001. Lastly we obtained two R-I images from July 1997
and 2005. These areas are so remote and vegetation growth is slow enough, that a 6-10 year
difference in data collection is not problematic. The only large changes between the years
would be wildfires, but we obtained all data on location and extent of wildfire from the
Canada Forest Service for the time period of study and no fires occurred within the study
area during that time.

Figure 14 shows two false color composites, one from the two dates of L-HV data from
PALSAR and the other from two dates of L-HH data, with the first date as red and the
second as cyan. These composites show how the cyan colored areas help distinguish fen
from bog in the L-HH composite (bottom figure) , since fens are characterized by fluctuating
water levels and flowing water, whereas bogs tend to have water levels that remain below
the moss covered surface with small changes in the short time period of the two image dates
(July and August 2007). The L-HV cross polarized data provide information on the level of
biomass which is essential in discrimination of low herbaceous open fens versus fen
woodlands and open, wooded and forested bogs. This channel also clearly distinguishes the
high biomass upland areas from the lowlands (Figure 14). Further, by using multi-date data,
the seasonal changes in moisture and water levels are useful for discrimination of wetlands
that generally have large changes in moisture/flood conditions (stream-fed fens) versus
those that have smaller changes (rain-fed bogs).
While most boreal peatlands in central Alberta are characterized by low canopy closure,
allowing C-band R-1 to be useful, evaluation of the L-band JERS and PALSAR data
demonstrate the additional definition of forests from ecosystems with low amounts of
aboveground biomass and varying surface wetness conditions. While C-HH R-1 provides
information on low aboveground biomass wetland types, JERS L-HH and PALSAR more
clearly define the differences between forested wetland types, as well as distinguishing high
and low aboveground biomass areas. The R-1 data were included in the mapping methods
for distinguishing swamps since Grenier et al. (2007) found R-1 useful for such purposes.
Fig. 13. Comparison of how different wetland types (bottom left, wetland map based on air
photo and field truth) appear in 3 date JERS L-band imagery (1995-7, top left) versus
Landsat imagery (Sept 88, top right). The pink circled areas are wooded fens and the yellow
circled features are open fens. Most of the area in these scenes is wooded bog.
Open Fen (very low

biomass)
Burned Area
(low biomass )
2 Date PALSAR L-HV Swamps (varying water

July 2007 levels [Red/Cyan])
August 2007
©JAXA 2007
2 Date PALSAR L-HH

July 2007
August 2007
©JAXA 2007
Fen Woodland Bog (no change in
(change in moisture moisture levels)
levels [cyan])
Fig. 14. Two date false color composites of PALSAR L-HV (top) and PALSAR L-HH
(bottom) from July (red) and August (cyan) of 2007.
4.3.3 Mapping Methodology

Unlike the methodology used in the Great Lakes case study, here all data (SAR and
Optical/IR) were fused in an object-based GIS analysis, using Definiens Professional
software. Note, however, that Definiens Professional allows the user to choose which bands
are used for each category being mapped. Thus, only Landsat may be used for one category,
while only SAR is used for a second, and all data are used for a third, etc.
Object-based classification methods involve two phases: 1) spatial objects are formed using a
region-growing segmentation algorithm to merge homogeneous pixels; and then 2) image
classification techniques are applied. The segmentation phase provides additional attributes
describing the spatial context and morphology of features which can inform the
classification beyond spectral values alone. Segmentation can also be reiterated at various
scales to capture the range of features contained in the image. This allows heterogeneous
cover types (i.e., wetlands containing some open water pixels for example mixed with
denser canopy) to be grouped depending on the segmentation scale chosen by the operator,
and can significantly improve map accuracy (Grenier et al. 2007).
We first created segmentation regions defined by the Landsat and SAR. Then using data
from April and August Landsat, two dates of PALSAR L-HV and L-HH and two dates of R-
1 C-HH, we developed a top-down classification approach in Definiens. The methodology
relied on a combination of thresholds and nearest neighbor classifiers in a decision tree.
Using decision rules, we first distinguished land from open water using PALSAR L-HH and
TM band 5 thresholds to map open water, with all non-water pixels being classified as land
(Figure 15). Next the land was divided into burn and non-burn categories using a nearest
neighbor classification of the Landsat data. Note that these burns occurred after the detailed
Airphoto peatland map of Figure 13 was created, which we rely on as reference in our
validation. Non-burned areas were then divided into upland versus wetland, with upland
forest classified using PALSAR L-HV and TM band 3 from August 2001. Next the wetland
classes were mapped. First open fen was mapped using L-HV and Landsat TM. Finally, a
nearest neighbor classification was conducted on the remaining classes: Woodland Bog,
Wooded Fen, and Swamp using April TM 3, 4, 5 and two dates of PALSAR HH, and two
dates of R-1. Figure 15 shows the process, with the final map in Figure 16.
The final peatland map had 80% (Table 5) overall accuracy compared to the air-photo based
map (circa 1970-80s photos), which was created from pre-burn photos (Bourgeau-Chavez et
al., in prep.). The bog had 77 % user’s and 91 % producer’s accuracy, fen had 60% user’s and
producer’s accuracy, upland had 88% user’s and 76% producer’s accuracy. Note that in the
air photo-based reference map, both upland forest and open water were mapped as a
merged class, and this is likely causing some error. Also, the time difference between the
reference 70-80’s air photo map and the SAR-optical map of 2000’s likely resulted in changes
to the landscape, notably the fires that occurred in 1988 and 1998. Also, fens are very
difficult to map on air photos and there may be errors in the reference maps. We did find
errors in some areas mapped as Marsh in the air photo-based map.
4.3.4 Summary
This initial research demonstrates the strong potential of a SAR-Optical/IR approach for
application to large areas for a better understanding of the spatial variation in peatland
types across the boreal landscape. Similar data fusion methods have been (Li and Chen
2005) or are being used (Fournier et al. 2007, Grenier et al. 2007) for mapping peatlands of
Canada. The CWI methods were described earlier (Grenier et al. 2007), but are much coarser
classes. Li and Chen (2005) mapped peatlands of eastern Canada into open versus treed bog,
marsh, swamp and open fen with high accuracy. Their methods involved the use of several
dates of R-1 data (45° incidence), Landsat and a DEM. While we found R-1 to be of limited
use (as did Grenier et al. 2007) in western Canada, it should be noted that eastern Canadian
peatlands are quite different than western peatlands, and the various methods will need to
be assessed for transferability.
We are currently processing imagery to increase the Alberta study area to include a three
scene mosaic of PALSAR from the two dates, using current Landsat from spring, summer,
and fall, and ERS data. ERS data are being used over Radarsat, because of coverage of the
larger area. Field visits are planned to areas in disagreement between the air photo map and
SAR-Optical/IR map for validation and improvement of the mapping approach. Additional
peatland study sites will also be evaluated in eastern Canada, and Alaska, as well as the
Upper Peninsula of Michigan.
Pr oce ss Tree Level 1 Process Tr ee Lev el 2
Orange= land, Blue = water Brown = burn, Red = unburn, blue = water
Process Tr ee Level 3 Pr ocess Tree L evel 4
Brown = burn, Green = upland, Cyan = wetland, blue = water , white = open fen
Figure 15. Process Tree images from the top down approach used for mapping Peatlands in
the Central Alberta study area using Definiens.
10 km
21 km
Figure 16. Hybrid SAR-optical derived map of Peatland types in central Alberta.
Reference AirPhoto Map

user's
Bog Fen Swamp Marsh Upland totals accuracy
Bog 50 1 1 0 13 65 0.77
SAR-Optical Map
Fen 3 6 1 10 0.60
Swamp 1 2 7 1 11 0.64
Marsh 0 1 1 0.00
Water 26 26 1.00
Upland 1 1 1 21 24 0.88
Totals 55 10 9 1 62
producer's 0.91 0.60 0.78 0.00 0.76
Table 5. Accuracy assessment of SAR-Optical Map on vertical axis and Airphoto map as
reference (top). Note that Upland on the airphoto map was labeled “Z” and represented
open water and upland forest.
5. Discussion and conclusions

Many techniques focus on using multispectral data, such as Landsat or Aster, alone or in
combination with ancillary data sets such as soils and topography for wetland mapping.
However, research has shown how SAR and multispectral sensors complement each other
in the classification and monitoring of wetland ecosystems and that SAR represents one of
the most promising sensor types for improving wetland mapping capability (Bourgeau-
Chavez 2004, 2008, Grenier 2007, etc). While multispectral data measure spectral reflectance
and emittance characteristics of various cover types and wetness in open canopied
ecosystems, SAR is sensitive to variations in biomass, structure and soil moisture and flood
condition of landscapes including forests and other closed canopy ecosystems. Forested
wetlands are the most difficult to identify remotely because of the inability of traditional
multispectral sensors to detect moisture beneath the canopy. Radar can not only penetrate a
closed canopy to detect flooding, but since radars are active systems, can acquire data
independently of solar illumination and cloud cover conditions. Thus, data can be collected
during specific conditions relevant to finding seasonally flooded wetlands or seiche-
influenced wetlands. These SAR data can be used not only to detect and define wetlands,
but also to monitor extent of inundation and in some cases level of inundation (Bourgeau-
Chavez et al. 2005, Lang et al. 2008).
The case studies shown here demonstrate the improved mapping capabilities by including
SAR in the traditional Optical/IR methods of mapping. It is important to note that the
timing of the acquisitions can be very important for detection of flooding beneath a canopy,
as was seen in the two JERS dates of imagery over Lake. St. Clair, where neither image date
was able to be used from mapping the flood condition. Further, the importance of multi-
temporal data was demonstrated in the image interpretation for all study sites, but
particularly for the Canadian peatland study in which the subtle changes in backscatter due
to changes in moisture levels allowed the distinction of wooded fens from bogs. Using
knowledge of the phenological changes which occur in one wetland over another, as long as
they are changes that can be detected by a particular sensor, can be key in distinguishing
otherwise similar appearing ecosystems in the mapping process.
While the results shown here focus on amplitude data from SAR, the new era of SAR
sensors have full polarimetric capability and as such, polarimetric decomposition can be
used to understand the type of scattering occurring from a particular ecosystem.
Decomposition variables can be used alone, or as additional bands in the more typical multi-
band classifiers. Variables such as phase difference, which were used in the past have been
further developed to include complex analysis of the full scattering matrix. Decomposition
methods are being developed to use Radsarsat-2 for peatland discrimination using a single
date of imagery (Touzi et al. 2007). These new techniques are in the early stages of
development and have not been tested on a variety of sites yet. But there is great potential
for these new polarimetric satellite sensors, which are just beginning to be explored.
6. Acknowledgements
This research was funded by the Great Lakes Coastal Wetlands Consortium, NASA and the
U.S. Fish and Wildlife Service. Special thanks to Brian Benscoter and Merritt Turetsky for
their assistance with the “truth“ maps of Alberta peatlands.
7. References
Arzandeh, S. and J. Wang. 2002. Texture evaluation of RADARSAT imagery for wetland
mapping. Can. J. Remote Sensing, 28(5):653-666.
Arzandeh, S. and J. Wang. 2003. Monitoring the change of Phragmites distribution using
satellite data. Can. J. Remote Sensing. (29(1):224-35.
Augustein, M. and C. Warrender. 1998. Wetland classification using optical and radar data
and neural network classification. Int. J. Rem. Sens. Vol. 19:1545-1560.
Baghdadi, N., M. Bernier, R. Gauthier, and I. Neeson. 2001. Evaluation of C-band SAR data
for wetlands mapping. Int. J. Rem. Sens. Vol. 22:71-88.
Becker B., Lusch D., & Qi J., (2005) Identifying optimal spectral bands from in-situ
measurements of Great Lakes coastal wetlands using second derivative analysis.
Remote Sens. Environ. 97 238-248.
Bourgeau-Chavez, L.L., M.R. Turetsky, B.W. Benscoter, and E.S. Kasischke. Improving mapping
of boreal peatlands through use of Landsat and satellite SAR imagery, in prep.
Bourgeau-Chavez, L.L. and Richard Powell. 2009. Mapping the Invasive Phragmites with
ALOS PALSAR radar imagery over the Saint Clair River Delta in the Great Lakes.
Society of Wetland Scientists 2009 Conference, Madison, WI. June 21-26, 2009.
Bourgeau-Chavez, L.L., R.D. Lopez, A. Trebitz, T. Hollenhorst, G. E. Host, B. Huberty, R. L.
Gauthier, and J. Hummer. 2008. Chapter 8, Landscape-Based Indicators in Great
Lakes Coastal Wetlands Monitoring Plan, Great Lakes Coastal Wetlands
Consortium, Project of the Great Lakes Commission, funded by the U.S. EPA
GLNPO, March 2008, pp. 143-171. http://www.glc.org/wetlands/final-report.html
Bourgeau-Chavez, L.L., K.B. Smith, Eric S. Kasischke, S. M. Brunzell, E. A.
Romanowicz, and C. J. Richardson. 2005. Remote Monitoring Regional Scale Inundation
Patterns and Hydroperiod in the Greater Everglades Ecosystem. Wetlands, Vol.
25(1):176-191.
Bourgeau-Chavez, L.L., K.Riordan, M. Nowels, and N. Miller. 2004. Final Report to the
Great Lakes Commission: Remotely Monitoring Great Lakes Coastal Wetlands
using a Hybrid Radar and Multi-spectral Sensor Approach. Project no.

WETLANDS2-WPA-06.82pp.
http://www.glc.org/wetlands/pdf/GD-landscapeReport.pdf
Bourgeau-Chavez, L.L., E.S. Kasischke, S.M. Brunzell, J.P. Mudd, K.B. Smith, and A.L. Frick
2001. Analysis of spaceborne SAR data for wetland mapping in Virginia riparian
ecosystems. Int. J. Rem. Sens, Vol. 22(18):3665-3687.
Charman, D. 2002. Peatlands and environmental change. J. Wiley & Sons, London & New
York, 301 p.
Costa, M. , E. Novo, E. Ahern, E. Mitsou, J. Mantovani, M. Ballester, and R. Pietsch. 1998.
The Amazon flood plain through Radar eyes: Lago Grande de Monte Alegre case
study. Can. J. Remote Sensing. Vol. 24: 339-349.
Crawford, M., S. Kumar, M. Ricard, J. Gibeaut, and A. Neuenschwander. 1999. Fusion of
airborne polarimetric and interferometric ASAR for classification of coastal
environments. IEEE Trans. on Geosci. and Rem. Sens., Vol. 37:1306-1315.
Fournier, R. M. Grenier, A. Lavoie, R. Helie.. 2007. Towards a strategy to implement the
Canadian Wetland Inventory using satellite remote sensing. Can. J. Remote
Sensing. Vol. 33: 528-545.
Grenier, M., Demers, A.-M., Labrecque, S., Benoit, M., Fournier, R.A., and Drolet, B. 2007.
An object-based method to map wetland using RADARSAT-1 and Landsat-ETM
images:test case on two sites in Quebec, Canada. Can. J. Remote Sensing, Vol.
33:528-545.
Henderson, F. M., and A.J. Lewis. 2008. Radar detection of wetland ecosystems: a review.
Int. J. Rem. Sens. Vol. 29(20):5809-1835.
Hess, L.L., J.M.Melack, S. Filoso, and Y. Wang. 1995. IEEE Trans. on Geosci. and Rem. Sens.,
33(4)896-904.
Hess, L., J. Melack, and D. Simonett. 1990. Radar detection of flooding beneath the forest
canopy: a review. Int. J. Rem. Sens. Vol. 11: 1313-1325.
IPCC 2008. Climate Change and Water, Intergovernmental Panel on Climate Change Technical
Paper VI, ed. by B. Bates, Z.W. Kundzewicz, S. Wu, and J. Palutikof. June 2008.
Jensen, John R., 2007, Remote Sensing of the Environment: An Earth Resource Perspective,
2nd Ed., Upper Saddle River, NJ: Prentice Hall.
Kasischke, E.S., Smith, K.B., L.L. Bourgeau-Chavez,, E.A. Romanowicz, S. Brunzell, and C.J.
Richardson. 2003. Effects of Seasonal Hydrologic Patterns in South Florida
Wetlands on Radar Backscatter Measured from ERS-2 SAR Imagery. Rem. Sens. of
Env. Vol. 88:423-441.
Kasischke, E.S. and L.L. Bourgeau-Chavez. 1997. Monitoring south Florida wetlands using
ERS-1 SAR imagery. PE&RS, Vol. 63(3): pp. 281-291.
Klemas, V.V., Dobson, J.E., Ferguson, R.L. and Haddad, K.D. 1993. A coastal land cover
classification system for the NOAA Coastwatch
Change Analysis Project. Journal of Coastal Research, Vol. 9(3): 862–872.Li, J. and W. Chen.
2005. A rule-based method for mapping Canada’s wetlands using optical, radar,
and DEM data. Int. J. Rem. Sens. Vol. 26:5051-5069.
Lopez, R.D. D. T. Heggem, D. Sutton, T. Ehli, R. Van Remortel, E. Evanson, and L. Bice. 2006
Using Landscape Metrics to develop indicators of great lakes coastal wetland
condition. US EPA Report EOA/600/X-06/002. www.epa.gov 77 pp.
Lozano-Garcia, D.F., and R.M. Hoffer, 1993, Synergistic effects of combined Landsat-TM and
SIR-B data for forest resources assessment. Int. J. Rem.Sens. Vol. 14(14), 2677.
Ozesmi, S. L., & Bauer, M. E. (2002). Satellite remote sensing of wetlands. Wetlands Ecology
and Management, 10(5), 381–402.
Pengra, B. C. Johnston, and T. Loveland. 2007 Mapping an invasive plant, Phragmites
australis, in coastal wetlands using the EO-1 Hyperion hyperspectral sensor. Rem
Sen Env, 108:74-81.
Peters, D. 1994. Use of aerial photography for mapping wetlands in the United States:
National Wetlands Inventory. Proceedings of the First International Airborne Remote
Sensing Conference and Exhibition held in Strasbourg, France on 11–15 September,
1994 (Ann Arbor: ERIM) pp. 165–173.
Pope, K., E. Reimankova, J. Paris, and R. Woodruff. 1997. Detecting seasonal flooding cycles
in marshes of the Yucatan peninsula with SIR-C polarimetric radar imagery. Rem.
Sens. of Env. , Vol. 59:157-166.
Pope, K., J. Rey-Benayas, and J. Paris. 1994. Radar remote sensing of forest and wetland
ecosystems in the Central American tropics. Rem. Sens. of Env. , Vol. 48:205-219.
Ramsey, E. III. 1998. Radar remote sensing of wetlands. Remote Sensing Change Detection:
Environmental Monitoring Methods and Applications, R. Lunetta and C. Elvidge
(eds) Chapter 13, pp. 211-243 (Chelsea, MI: Ann Arbor Press).
Rao, B., R. Dwivedi, S. Kushwaha, S. Bhattacharya, J. Anand, and S. Dasgupta. 1999.
Monitoring the spatial extent of coastal wetlands using ERS-1 SAR data. Int. J.
Rem. Sens., Vol. 20:2509-2517.
Rio, J. And D. Lozano-Garcia. 2000. Spatial filtering of radar data (RADARSAT) for
wetlands (brackish marshes) classification. Rem. Sens. Environ. Vol. 73:143-151.
Schmidt K. & Skidmore A. (2003) Spectral discrimination of vegetation types in a coastal
wetland. Rem. Sens. Environ. 85 92-108.
Touzi, R., A. Deschamps, and G. Rother. 2007. Wetland characterization using polarimetric
RADARSAT-2 capability. Can. J. Remote Sensing. Vol 33(1):S56-S67.
Townsend, P. 2002. Relationships between forest structure and the detection of flood
inundation in forested wetlands using C-band SAR. Int. J. Rem. Sens., Vol. 22:443-460.
Townsend, P.A. 2001. Mapping seasonal flooding in forested wetlands using multi-temporal
Radarsat SAR. PE&RS, Vol. 67(7):857-864.
Toyra, J. Pietroniro, A. and L. Martz. 2001. Multisensor hydrologica assessment of
freshwater wetland. Rem. Sens. Environ, Vol. 75:162-173.
Wang, Y. 2004. Seasonal change in the extent of inundation on floodplains detected by JERS-
1 Synthetic Aperture Radar data. Int. J. Remote Sensing, 25(13):2497-2508.
Wang, Y. And F. Davis. 1997. Decomposition of polarimetric synthesic aperture radar
backscatter from upland and flooded forests. Int. J. Rem. Sens. , Vol. 18:1319-1332.
Whitcomb, J. M. Moghaddam, K. McDonald, J. Kellndorfer, and E. Podest. 2009. Mapping
vegetated wetlands of Alaska using L-band radar satellite imagery. Can. J. Remote
Sensing, Vol. 35:54-72.
Wilcox, K.L., S. A. Petrie, L.A. Maynard and S.W. Meyer. 2003. Historical Distribution and
Abundance of Phragmites australis at Long Point, Lake Erie, Ontario. Journal of
Great Lakes Research, 29(4):664-680
Wilen, B. and Frayer, W. 1990. Status and trends of U.S. wetlands and deepwater habitats.
Forest Ecology and Management 33/34:181–192.
Remote Sensing Rock Mechanics and Earthquake Thermal Infrared Anomalies 709
34
X
Remote Sensing Rock Mechanics and

Earthquake Thermal Infrared Anomalies
Lixin Wu1,2 and Shanjun Liu2
1Academy of Disaster Reduction & Emergency Management, Beijing Normal University,
Beijing, China
2Institute for Geo-informatics & Digital Mine Research, Northeastern University,
Shenyang, China
1. Introduction
Rock fracturing is the cause of many geo-hazards including tectonic earthquake (EQ), rock
burst, rock sloping and rock pillar failure. Radiation signals such as acoustic emission, radio
frequency emission and electromagnetic (EM) radiation from loaded deforming rock, are
able to provide useful information for monitoring, interpreting and predicting rock
fracturing (Renata, 1977; Brady and Rowell, 1986; Yamada et al., 1989; Martelli et al., 1989).
Based on thermo-elastic theory, thermo-elastic stress analysis (TSA) and stress pattern
analysis by thermal emission (SPATE) were developed for the stress measurement of solid
materials, including homogeneous metal, macromolecular and composite materials,
respectively in 1960’s and 1970’s (Mounatin and Webber, 1978). Luong applied thermovision
to study experimentally the damage processes of concrete and rock (Luong, 1990), but no
reach to the remote sensing on geo-hazards.
In the experiments for investigating the mechanism of satellite thermal infrared (TIR)
anomaly before tectonic EQ (Gorny et al., 1988; Qiang et al., 1990), it was discovered that
there do exist TIR anomaly before rock fracturing (Geng et al., 1992). Later, it was
furthermore discovered that there are obvious TIR features as precursors of rock fracturing,
and that the loaded stress around 0.79  c can be taken as a precaution index for the
stability monitoring of loaded rocks (Wu and Wang, 1998). To explore the laws of infrared
radiation (IRR) variation in the process of rock loading, deforming and fracturing, and to
reveal the possible mechanism of satellite TIR anomaly before EQ, a large amount of IRR
imaging experiments on rock loaded to fracturing were conducted in China (Wu et al., 2000,
2001, 2002, 2003, 2004a, 2004b, 2004c, 2004d, 2006a, 2006b; Deng et al., 2001; Liu et al., 2002).
Hence, a new intersection discipline, Remote Sensing Rock Mechanics (RSRM), which takes
Remote Sensing, Rock Mechanics, Rock Physics and Informatics as its foundations and
serves for remote sensing on geo-hazards, was originated (Geng et al., 1992; Wu et al., 2000).
Based on retrospection to past experiments on RSRM, it was pointed out that there are two
IRR anomalies, being IRR image anomaly and IRR temperature curve anomaly respectively,
can act as rock fracturing precursors. The average IRR temperature (AIRT), being the
integral reflection of surface IRR energy, is applied as a quantitative index to study the
temporal evolution of IRR from loaded rock and to seek for the potential precursors of rock
fracturing. The temporal evolution of AIRT are the comprehensive effect of a series of
physical-mechanical processes inside a loaded rock, such as rock thermo-elastic acting, pore
gas desorbing & escaping, fractures producing & extending, rock frictionating, heat
transferring and environment radiation. The thermo-elastic effect and the frictional thermal
are two of the main mechanisms of increased IRR from loaded rock. RSRM experiments had
revealed the laws of changed IRR from loaded rock and provided scientific interpretations
for the mechanisms of satellite TIR anomaly before tectonic EQs of Ms>5.5.
2. Remote Sensing Rock Mechanics Experiments

2.1 Experiment Methods and Tools
The typical RSRM experiment is comprised of a loader (uni-axial or bi-axial), an infrared
imager and rock samples. As in Figure 1, a bi-axial loader was applied for loading along two
directions, and an infrared imager was applied to detect the surface IRR from loaded rock.
The maximum imaging rate of the imager is 60f/s, and the recording rate was usually set as
1f/s to record the IRR images continuously. Usually, tectonic EQ might be resulted from the
suddenly fracturing of compressively-sheared crust rock, the suddenly breaking of faults at
disjointed zones, the suddenly sliding of compressively-sheared faults or the stability losing
of compressively loaded intersected faults. To simulate the different mechanisms of rock
fracturing and EQ, several typical loading schemes were applied as in Figure 1.
a) b) c) d) e) f) g)
Fig. 1. RSRM experiment schemes to simulate different mechanisms of rock fracturing or
tectonic EQ: a) uni-axially load on a standard cylinder rock sample; b) uni-axially load on a
cylinder rock sample with a central hole; c) compressively-sheared load on a hexahedral
rock sample; d) bi-axially load on three jointed rock samples to frictional sliding; e)
bi-axially load on a damage rock sample with en echelon faults; f) bi-axially load on a
damage rock sample with disjointed faults; and g) bi-axially load on three jointed rock
samples simulating intersected faults.
2.2 Rock Fracturing Precursor: IRR Image Anomaly
2.2.1 Uni-axially loaded rock

Lots of rock samples made from coal, ironstone, sandstone, marble, limestone, granite,
granodiorite, gabbro and gneiss were uni-axially loaded and thermal imaging detected. The
sample size was standard of diameter and length, respectively, 50 and 100mm. It was
discovered that the IRR images of the uni-axially loaded rock have different features for
different fracturing pattern (Wu et al., 2006a). As in Figures 2~4, there are three fracturing
patterns, “X”-shaped, “//”-shaped and “|”-shaped respectively, occurred in our experiments.
The “X”-shaped and “//”-shaped positive IRR abnormal strips foretell the coming of
“X”-shaped shearing fracturing and the coming of “//”-shaped shearing fracturing
respectively, while the “|”-shaped negative IRR abnormal strip foretells the coming of tensile
fracturing.
The “X”-shaped positive IRR abnormal strips generated with loading along the “X”-shaped
shearing zone before peak stress, and got distinguished after peak stress, as in Figure 2. The
rock sample got finally fractured along the “X”-shape shearing zone. The evolution of IRR
abnormal strip had also reflected the fracturing being not symmetrical upper-and lower, in
that the upper part was clear with higher temperature, while the lower part is fuzzy with
lower temperature.
σ=0 MPa σ=96.7 MPa σ=97.2 MPa σ=91.8 MPa σ=0 MPa failure
Fig. 2. The IRR image positive anomaly of “X”-shaped shearing fracturing of an uni-axially
loaded marble sample
The “//”-shaped positive IRR abnormal strips generated with loading along the
“//’-shaped shearing zone at the upper part of sample before peak stress, and got
distinguished after peak stress, as in Fig 3. The evolution of the positive IRR image anomaly
had also reflected the fracturing being not symmetrical in that the upper part of the IRR
anomaly strip was clear with higher temperature, while the lower part was fuzzy excepting
for the final fracturing near the bottom of the sample. Besides, there was strong IRR
anomaly spot at the fracturing center for the intensive accumulation of mechanical energy
and for the intensive generation of frictional thermal at the local central place.
σ=0 MPa σ=119.8 MPa σ=136.5 MPa σ=125.8 MPa σ=0 MPa failure photo
Fig. 3. The IRR image positive anomaly of “//”-shaped shearing-fracturing of an uni-axially
loaded granite sample
The “|”-shaped negative IRR abnormal strip generated with loading along the tensile
fracturing zone of a rock sample before the peak stress, and got distinguished gradually at
the peak stress and after fracturing, as the approximately vertical dark strip in Figure 4. The
same phenomenon for a sandstone sample with a calcite vein was also reported (Wu, et al.,
2000).
26.5
26.4
26.3
26.2
26.1
26.0
25.9
25.8
σ=0 MPa σ=100.3 MPa σ=131.4 MPa σ=0.1 MPa σ=0 MPa failure photo
Fig. 4. The IRR image negative anomaly of “|”-shaped tensile fracturing of an uni-axially
loaded granite sample
2.2.2 Uni-axially loaded rock with a central hole

More than 10 samples with a central hole, modeling the structure stability of loaded rock
tunnels, made from marble and granite were infrared imaging detected. The rock samples
had two kinds of shapes respectively being cylinder with diameter and length, respectively,
50 and 100mm, and regular block with thickness, width and length, respectively, 70, 35 and
100 mm. It was discovered that there were distinguished positive IRR image anomalies
before rock fracturing, and the place of anomaly were exactly the coming fracturing place.
As in Figure 5, the positive IRR image anomalies had reflected the two kinds of fracturing,
respectively being diagonal fracturing (sample 1~5) and fork fracturing (sample 6). The IRR
anomalies, along the fracturing planes and shaped as spots or strips, generated not only on
rock surface but also on the hole’s surface (lateral sample 4 and 5). The temperature
increment is 1~3oC and 4~8oC respectively for marble and granite samples.
sample 1 sample 2 sample 3 sample 4 sample 5 sample 6
Fig. 5. The IRR image positive anomalies of a group of uni-axially loaded marble samples
with a central hole
2.2.3 Compressively sheared rock

More than 20 samples, size 7×7×7×cm3, made from sandstone, marble, limestone, granite
and gneiss, were compressively sheared and infrared imaging detected. Three pairs of steel
platens with shearing angle being 45 , 60 and 70º respectively were applied. The loading rate
was controlled as 2~5kN/s. It was discovered that the IRR temperature of rock surface
changed with loading, and a strip-shaped positive IRR image anomaly generated along the
central shearing plane before fracturing. With loading, the positive abnormal strip got more
and more distinguished and migrated gradually from the upper end to the lower end of the
sample, which foretold that the compressive-shearing fracturing was developing gradually
from the upper end to the lower end of the sample along the central shear plane. Figure 6
shows the typical IRR image series of a compressively sheared limestone sample.
As a special geological phenomenon occurring with the formation of great fault,
penniform-shaped fractures are a group of secondary fractures produced with the formation
of great primary fracture (Nicolas et al., 1977). It happened to occur in our experiments that
there were penniform-shaped fractures produced with a primary fracture in the
compressive loaded rock samples, as in Figure 7. The IRR positive anomaly strips generated
aside the primary IRR strip, passing through the central shearing plane, had reflected the
penniform-shaped fracturing events.
1s 80s 100s 102s

110
104s 106s 107s 110s

Fig. 6. The IRR image positive anomaly of the fracturing of a compressively sheared
limestone sample (time in second)
1s 40s 90s 100s
110s 140s 150s 160s
Fig. 7. The IRR image anomaly of the penniform-shaped fracturing of a compressively

sheared marble sample
2.2.4 Bi-sheared frictional sliding rock blocks

Ten groups of rock samples made from gabbro, granodiorite, limestone and marble were
IRR detected in the process of bi-sheared frictional sliding or viscosity sliding. Each group,
as in Figure 8 and 9, was comprised of three jointed rock blocks whose size respectively be
50×50×100mm3, 50×70×150mm3 and 50×50×100mm3 from left to right, and its friction
area was constant, 50×100mm2. Four contact conditions, symmetrical (yes for rock property
and for its smooth friction surface, as in Figure 8), uncertain symmetrical (yes for rock
property but not for its coarse friction surface, as in Figure 9), unstable asymmetrical (yes for
rock property but not for its staged friction surface) and stable asymmetrical (not for rock
property but yes for its smooth friction surface), were designed and tested respectively (Wu
et al., 2004b).
It is revealed that the evolution of rock surface IRR temperature field is not only correlated
with rock stress, but also correlated with the features of friction surface and rock properties
at both sides. General law lies in that the IRR at the place of stress concentration and strong
friction zone is stronger than that at the place of stress relaxation and weak friction zone. In
condition of friction surface be symmetrical, the IRR image is double butterfly-wings shaped,
as in Figure 8. However, in condition of friction surface be uncertain symmetrical, unstable
asymmetrical or stable asymmetrical, the temporal-spatial evolution of IRR anomaly is
uncertain or unstable, as in Figure 9. The positive IRR anomaly spots, foretelling the
evolution of stress, energy and viscosity-sliding process, may be beads-shaped,
needle-shaped, suspended needle-shaped, strip-shaped, single butterfly-wings shaped or its
evolution in order (Wu et al., 2004b).
1s 71s 151s 311s
Fig. 8. The IRR image positive anomaly of the stick-slipping of symmetrical rock samples
(time in second)
1s 81s 156s 311s
Fig. 9. The IRR image positive anomaly of the stick-slipping of asymmetrical rock samples
(time in second)
2.2.5 Bi-axially loaded rock

Infrared imaging detection on the rupturing of en echelon and collinearly disjointed jointed
faults were done in the process of bi-axial loading. It was revealed that the IRR from loaded
rock surface is correlated with loading stress, which could be divided into five stages as
loading beginning, linear elastic, stress locking, stress unlocking and fracturing(Wu et al.,
2004a). During the stress-unlocking stage, positive IRR anomaly strip generated at the
disjointed zone, as in Figure 10 and 11. The positive IRR anomaly strip around the disjointed
zone has general evolution features as: firstly, the strip gets enhancing; then, gets weakening
(or ‘silence’); and finally, gets enhancing again. The re-enhancing of IRR anomaly strip after
the weakening stage is a meaningful precursor foretelling the place of primary fracturing of
faults or the coming epicenter of an EQ.
1s 227s 236s photo

Fig. 10. The local IRR positive anomaly foretell the fracturing of en echelon disjointed faults
(marble, time in second)
1s 210s 257s photo
Fig. 11. The local IRR positive anomaly foretell the fracturing of collinearly disjointed faults
(marble, time in second)
2.3 Rock Fracturing Precursor: IRR Temperature Anomaly
2.3.1 Quantitative index: AIRT

The evolution of surface IRR from loaded rock is the comprehensive effect of rock
thermo-elastic acting, pore gas desorbing & escaping, fractures producing & extending, rock
frictionating, heat transferring and environment radiation. Being the integral reflection of
surface IRR energy, the average IRR temperature (AIRT) is selected as a quantitative index
to study the evolution of IRR from loaded rock and to seek for rock fracturing precursors
(Wu et al., 2006b).
The infrared imager detects and records the thermal images of loaded rock surface. The
thermogram is comprised of a matrix of color pixels which representing the IRR brightness
temperature of each pixel of the rock surface. For example, the imaging matrix of
TVS-8100MKII infrared imager is 160×120. The IRR temperature of each pixel tends to
fluctuate with time due to the instability of the detector unit and the influence of
environmental radiation, and the IRR temperature of each pixel will not always be the same
for the local difference of rock stress and rock strain. The maximum, minimum and average
value of loaded rock surface IRR temperature, respectively being IRRTmax, IRRTmin and
IRRTave, could be quantitatively obtained from thermogram. The analysis revealed that
IRRTmax and IRRTmin will not change obviously except that IRRTmax might rise suddenly just
before rock fracturing, while IRRTave is to change stably with loading, as in Figure 12. The
physical interpretation lies in that the surface IRRTave is a general reflection of the energy
balance inside the loaded rock.
2 6 .4
IR R T m a x
2 6 .2
2 6 .0
IRRT/¡æ
IRRT/°C
2 5 .8
IR R T a v e
2 5 .6
2 5 .4
IR R T m in
0 50 100 150 200

tim e /s
Fig. 12. The evolution of three indexes of IRR brightness temperature of loaded rock surface
Hence, the IRRTave of rock surface, denominated as AIRT, is selected as a quantitative index
to study the precursors of rock fracturing and geo-hazards. The procedures for AIRT-based
precursor analysis includes: 1) to define a unified boundary of the analyzed region
(resampling region) for all thermogrames; 2) to resample the IRR temperature value from
the data file of each thermograme in time order; 3) to calculate the AIRT of the resampling
region of each thermograme; 4) to draw the AIRT–time curve of the rock sample; 5) to
analyze the evaluation features of the AIRT–time curve and to identify the messages as a
precursor of rock fracturing and hazard; 6) to compare with the qualitative image anomaly
so as to analyze and to confirm the AIRT abnormal precursor.
2.3.2 Influence factors of AIRT curves

1) Loading stages and rock deformation
The stress-strain curve is a basic method for describing rock deformation and for
interpreting rock mechanical behaviors (Hudson and Harrison, 1997). Generally, the
deformation process of loaded rock is divided into four stages respectively being stage-I of
defects compaction, stage-II of linear elastic deformation, stage-III of plastic deformation
and stage-IV of fracturing failure, as in Figure 13. The four characteristic points, E, Y, P and
F are called as elastic-starting point, yield-starting point, peak-stress point and
failure-impending point respectively.
Stage-I: the downward-concave curve section tells that there are some defects such as pores,
fissures and joints inside the rock body, and that the defects is under compaction, which
cause the stress to rise slowly. The more the defects, the severe the curve downward
concave.
Stage-II: the curve section linearly developed tells that the compaction of defects has
finished and the rock is undergoing elastic deformation. The higher the angle of the section
line inclined, the stronger the rock.
Stage-III: the upward-concave curve section tells that there are new fractures developing
inside the rock. The plastic deformation starts, and the new generated fractures together
with the initial defects are possible to cause friction between its two side-faces.
Stage-IV: the curve section turning to drop tells that the fractures are getting wider, longer
and to connect with each other. The rock is losing its strength and stability, and the final
fracturing failure or rock hazard is impending.
For the difference between rock compositions, the details of stress-strain curve of different
rock will be different. As to brittle rock, its stage-II is close to point P and its stage-IV will be
cliff-shaped. Usually, most of the crust rocks are brittle. Five kinds of typical crust rock,
granodiorite, gabbro, gneiss, limestone and marble had been tested in our experiments. The
typical load-displacement curves of the tested rock samples are shown in Figure 14. It tells
that all the tested rocks are brittle.
σ
P 350
F 300
granodiorite
lim estone
Y
250 m arble
gabbro
200
load/kN
gneiss
150
100
E
50
I II III IV ε 0
0.0 0.2 0.4 0.6 0.8 1.0 1.2

0 displacement/mm
Fig. 13. Typical AIRT curve of uni-axially Fig. 14. The load-displacement curves of
loaded rock five tested rocks
2) The influence of loading condition

The experiments discovered that the evolutions of AIRT have different laws in different
loading condition, such as uni-axial loading (constant displacement controlled),
compressively-sheared loading (approximately constant load controlled) and bi-axial
loading (constant displacement controlled).
a) Uni-axial loading
As shown in the left-hand side of Figure 15, the surface facing to the infrared imager is to be
detected, and a rectangle region close to the boundary of the rock sample is defined for data
resampling and analyzing (Liu et al., 2002; Wu et al., 2002). Multiple experiments revealed
that there was slight variation of AIRT at different deformation stage although the AIRT
linearly increased with load and deformation. At stage-I, the AIRT will rise slowly or drop a
little; at stage-II, the AIRT will rise stably; at stage-III, the AIRT will rise quickly than that in
stage-I and stage-II. The right-hand side of Figure 15 shows the comparison of the evolution
of AIRT and the load with rock deformation, which is rock displacement in generally, of a
marble sample.
b) Compressive-shear loading
In condition of compressively shear, the rock sample will always get fracturing along the
shearing plane, which locates near to the central plane of the loaded sample. To minimize
data resampling work and to focus on the key region, a narrow rectangle along the shearing
plane is defined as the resampling and analyzing region (Wu et al., 2004c). Multiple
experiments revealed that the temporal evolution of the AIRT is different with the shearing
angle. Three shearing angles (γ) being 45o, 60o and 70o respectively, are applied. As the
shearing angle changes from 45o to 70o, the temporal evolution changes from monotonic rise,
to drop-to-rise and to monotonic drop in order, as in Figure 16.
25.24
160 load
AIRT 25.20
120
AIRT/℃
P 25.16
load/kN
80
25.12
40
25.08
I II III
0
resampling region (in broken line) 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
displacement/mm
Fig. 15. The typical AIRT curve of uni-axially loaded rock sample (marble)
0 20 40 60 80 100 120 140 160 180 200

28.90
600
load 28.85
500 AIRT
28.80
400
P
load/kN
AIRT/℃
28.75
300
200 28.70
γ
100 28.65
A B C
0 28.60
0 20 40 60 80 100 120 140 160 180 200
time/s
resampling region(in broken line) a) 45°shearing angle
0 20 40 60 80 100
0 20 40 60 80 100 120 140 160 180 250 29.18
250
29.94 load
load AIRT 29.16
200
200 AIRT 29.92
29.14
29.90
150
AIRT/℃
150
AIRT/℃
load/kN
29.88 29.12
load/kN
100 29.86 100

29.10
29.84
50 50 P 29.08
A B 29.82
A B
0
P 29.80 0 29.06
0 20 40 60 80 100 120 140 160 180 0 20 40 60 80 100
time/s time/s
b) 60°shearing angle c) 70°shearing angle
Fig. 16. The typical AIRT curve of compressively-sheared rock samples at different shearing
angles
The mechanism lays in that the ratio of compressive-stress to shear-stress along the shearing
plane decrease with the rise of the shearing angle. The smaller is the shearing angle, the
higher is the compression-shear ratio. In condition of 45o, the load-time curve developed in
three stages, as stage A, B and C in Figure 16a, with load increasing from slow to rapid, and
to slow again. In condition of 60o the load-time curve developed in two stages, as stage A
and B in Figure 16b, with load speed changing from approximate constant to be decrease
slightly. In condition of 70o, the load-time curve developed in two stages, as stage A and B in
Figure 16c, with loading speed changing from slow to rapid.
The compressive action on loaded rock is to cause surface IRR temperature rise, while the
tensile action on loaded rock is to cause surface IRR temperature drop. Actually, both
compressive action and tensile action are to occur along the compressively sheared plane,
and the detected surface IRR is the comprehensive effect of the two actions. It was reached
that (Wu et al., 2004c): 1) in condition of shearing angle being 45o, the surface AIRT will rise
monotonically with loading in that the temperature increment from compressive action and
friction is stronger than the temperature decrement from tensile action in the whole loading
process; 2) in condition of shearing angle being 60o, the surface AIRT will drop
monotonically with loading in that the temperature increment from compressive action and
friction is weaker than the temperature decrement from tensile action before stage-III (point
Y in Figure 13, and point P in Figure 16b); with the friction effect getting strong in stage-III,
the surface AIRT will get to rise in that the temperature from compression and friction get
stronger than the temperature decrement from tensile action; 3) in condition of 70o, the
surface AIRT will drop monotonically with loading in that the temperature increment from
both compression and friction are weaker than the temperature decrement from tensile
action.
c) Biaxial loading
By using of bi-axial loading system and infrared imaging system, the IRR features of two
kinds of disjointed jointed faults, respectively be collinearly and non-collinearly disjointed
faults, were experimentally studied (Wu et al., 2004a). Since all the faults got fractured
finally at the disjointed zone, a circle covering the disjointed region is defined as the
resampling region, as in the left-hand side of Figure 17. It could be known from the
right-hand side of Figure 17 that the IRR from loaded samples is related with load stress,
and the evolution stage could be classified into five stages (I~V) relating with initial
compacting, elastic deforming, stress blocking, stress deblocking and rock fracturing
respectively. From stage-II to stage-IV, the evolution of AIRT has the features of rising to
dropping, and to rising again.
300 2 5 .8 2
250 lo a d 2 5 .8 0
A IR T
200 PA 2 5 .7 8
AIRT/℃
2 5 .7 6
load / kN
150
2 5 .7 4
100
2 5 .7 2
50
2 5 .7 0
0 Ⅰ Ⅱ Ⅲ Ⅳ 1
2 5 .6 8
resampling region(in broken line) 0 50 100 150 200 250
tim e /s
a) collinearly disjointed faults
2 2 .8 0
200
lo a d
A IR T 2 2 .7 6
150
P
A
2 2 .7 2
load/kN
AIRT/℃
100
2 2 .6 8
50
2 2 .6 4
0
Ⅰ Ⅱ Ⅲ Ⅳ Ⅴ 2 2 .6 0
resampling region(in broken line) 0 50 100 150 200 250
tim e /s
b) en echelon disjointed
Fig. 17. The typical AIRT evolution of two kinds of disjointed faults bi-axially loaded
3) The influence of rock characteristics

It was discovered in our experiments that for the most of rock samples in condition of
uni-axial loading, its AIRT approximately rose with loading. But there were a few abnormal
samples made from limestone had shown AIRT features of dropping with loading, as in
Figure 18. The cause lies in that limestone has much more pores than the other rocks.
Usually, there are many gases, such as CH4, CO2, CO and O2 etc., enclosed inside the pores
of rock body (Wang, 2003). With the decrement of pore volume due to the loading
compaction and with the increment of fractures produced inside the limestone sample, the
pore gases will get escaping. The escaping behavior of pore gas needs to absorb thermal
energy from the rock sample. If the heat from compression and friction is lower than that
absorbed by pore gases, the surface AIRT is to drop with loading. If look carefully at the
load-displacement curves in Figure 14, it could be founded that the curve of limestone
concaved downward the most at the compaction stage as compared to that of the other four
kinds of rock, which means that there are more pores inside limestone than the others.
180 26.06 0 50 100 150 200

400 25.74
160 load
AIRT 26.04 load 25.72
140
300 AIRT
120 25.70
26.02
AIRT/�
100
AP
AIRT/℃
load/kN
load/kN
200 25.68
80 26.00
60 25.66
25.98 100
40
25.64
20 25.96
0 0 25.62
0 50 100 150 200
0.0 0.1 0.2 0.3 0.4 0.5 P 0.6
displacement/mm time/s
Fig. 18. AIRT of uni-axially loaded limestone Fig. 19. AIRT short-dropping precursor
sample for uni-axially loaded gabbro sample
2.3.3 The classification of precursors

Analysis to the evolution of AIRT curves discovered that a large amount of rock samples in
condition of uni-axial loading, compressively-sheared loading and bi-axial loading had
presented obvious precursors for rock fracturing. As referring to the general process of
AIRT evolution, the AIRT anomaly precursors for rock fracturing and hazard could be
classified as short-dropping, rapid-rising and dropping-to-rising respectively.
1) Short-dropping precursor
The AIRT curve rises with loading but has a short dropping at loading stage-IV; later, the
AIRT curve will rise again. The bi-axially load on collinearly and non-collinearly disjointed
faults had shown short-dropping precursors as in Figure 17, and the point P was suggested
to be the precursor point of rock fracturing and rock hazard. Figure 19 shows another
typical case of gabbro sample uni-axially loaded. Here, point A is the turning point of AIRT
from rising to short dropping, and point P is another turning point from short dropping to
rising again, which is suggested to be the precursor point of rock fracturing and rock
hazard.
2) Rapid-rising precursor
The AIRT curve rises slowly with loading but turns to rise rapidly before rock fracturing,
and the turning point is exactly the precursor point. Figure 15 and Figure 16a have this kind
of precursor. Figure 20 shows another typical case of marble sample uni-axially loaded.
Here, point P is the turning point of AIRT from rising slowly to rising fast, which is
suggested to be the precursor point of rock fracturing and rock hazard.
3) Dropping-to-rising precursor
The AIRT curve drops slowly with loading but turns to rise just before rock fracturing, and
the turning point is exactly the precursor point. Figure 16b has this kind of precursor. Figure
21 shows another typical case of marble sample uni-axially loaded. Here, point P is the
turning point of AIRT from dropping slowly to rising fast, which is suggested to be the
precursor point of rock fracturing and rock hazard.
0 20 40 60 80 100 120 0 10 20 30 40 50
180 25.26 220 30.00
160 AIRT-time curve 25.24 200 load-time curve 29.98
140 load-time curve 25.22 180 AIRT-time curve
29.96
25.20 160
120
AIRT/℃
25.18 140 29.94
100
AIRT/℃
load/kN
load/kN
25.16 120 29.92
80 P
A
25.14
100
60 29.90
25.12
80
40 25.10 29.88
60
20 25.08 P 29.86
40
0 25.06
0 20 40 60 80 100 120 0 10 20 30 40 50
time/s time/s
Fig. 20. AIRT fast rising precursor for Fig. 21. AIRT dropping-to-rising precursor for
uni-axially loaded marble sample compressively sheared loaded marble sample
2.3.4 The temporal features of precursors

The occurrence moments of the precursors of AIRT of totally 52 tested rock samples are
listed in Table 1(Wu et al, 2006b). Although the loading conditions and the rock samples are
different, the precursor occurrence moment, were very similar as 0.77~0.94σc(σp). Here, σc is
the uni-axial compressive strength, and σp is the peak stress. The precursor occurrence
moment of uni-axially loaded or compressive sheared rock sample is 0.79σc and 0.82σc
respectively. It is worthy to mention that the precursor occurrence moment of bi-axially
loaded collinearly disjointed faults and echelon faults are much different. That for
collinearly disjointed faults was close to the peak stress, 0.87σp, while that for echelon faults
was far away from peak stress, 0.77σp. It provides an important evidence for the complexity
of study on tectonic EQ prediction on shock time, based on satellite infrared remote sensing
and referring to the seismogenic mechanism.
Loading condition Sum of Sum of The ratio: Average of

tested samples with (Sp/St)× precursor
samples, precursors, 100% occurrence
St Sp moment
uni-axial loading 22 9 41% 0.79σc
70° 7 1 14% 0.94σp

compressively-sheared 58%
60° 8 6 74% 0.82σp
loading
45° 11 8 73% 0.77σp
Bi-axial loading for collinearly 2 2 100% 0.87σp
disjointed faults
Bi-axial loading for en echelon faults 2 2 100% 0.77σp
Table 1. Statistic for precursor occurrence in different loading conditions
2.4 Large IRR at Fracturing Centre

As the recording rate of the infrared imager applied was 60f/s, the transient IRR
temperature at the fracturing center could be snapped. In condition of uni-axial loading, the
fracturing center of a brittle rock is usually at the “X”-shaped fracturing center. It is
discovered that the transient IRR temperature at the fracturing center is much higher than
that on rock surface, as in Figure 22, and it is positively related with rock strength and rock
deformation. For some compressively sheared hard rock samples made from gabbro and
gneiss, the transient IRR temperature at the fracturing center is higher than 155oC, which is
the upper limitation of the 2nd temperature range (72~155oC) of the imager applied.
In condition of high angle, 60 and 70º, compressively sheared loading, the fracturing center
is at the center of the fractured shearing zone. Since the ruptured upper block of rock sample
was pushed apart from the steel platen immediately after the abrupt rupturing, usually be
1~2 s after the rupturing, the inside shearing zone got exposed to the imager immediately
and the transient IRR temperature filed was snapped. It was discovered that the IRR
temperature on the inside shearing zone is not only much higher than that of outside rock
surface, but also inhomogeneous distributed, neither even nor centripetal, as in Figure 23. It
means that much more mechanical energy had been converted into frictional thermal and
IRR energy due to the intensive energy accumulation, the sufficient local deformation and
the abrupt frictional sliding at the center of the shearing zone. In other words, the large IRR
temperature at the inside shearing center had reflected the comprehensive effect of local
concentrated energy conversion and frictional thermal.
a) uni-axially
Fig.22. loaded granite
The transient b) compressively
IRR thermograme sheared gabbro-1
of fracturing c) compressively sheared gabbro-2
rock samples
Fig. 22 The transient IRR thermograme of fracturing rock samples
(a) marble-1 (b) marble-2 (c) granite

Fig. 23. The inside IRR isothermal filed on the fracturing zone of compressively sheared rock
samples
Hence, it could be deduced that in condition of great tectonic stress, large deformation
and/or abrupt frictional sliding, large temperature as high as hundreds or thousands of

degree Celsius is possible inside crust rocks. The high temperature could cause partial melt
of crust rock, which provides a scientific explanation for the existence of pseudotachylyte in
some lager faults (Nicolas et al., 1977; Sibson et al, 1980) and for the failure-generated EQ
lights (Martelli et al., 1989). Besides, we can deduce that the continuous shearing
deformation or the abrupt fracturing of highly loaded rock/coal body in a coal mine is
possible to cause local sheared heating of temperature hundreds of degree Celsius, which
might be a potential ignition of local methane (the minimum ignition temperature is 595oC).
3. Remote Sensing Rock Mechanics Model (RSRM-model)

3.1 Thermo-Mechanical Coupling Effect
3.1.1 Thermo-mechanical coupling in a loaded solid

The heat production inside a loaded solid is called as thermo-mechanical coupling effect.
According to the material features and the different deformation stages of a loaded solid, the
thermo-mechanical coupling is classified as thermo-elastic, thermo-plastic and
thermo-viscous respectively for elastic deformation, plastic deformation and viscous
deformation. Generally the rock is a hard brittle solid, its plastic and viscous deformation
could be ignored, and the thermo-elastic effect and the frictional thermal are the two chief
mechanisms of surface IRR from loaded rock. Kelvin coined the thermo-elastic theory in
1853 that the changed physical temperature of a loaded component is correlated to its
changed stress as follows:
T / T   K 0  (1)
Here: T is the absolute temperature of a loaded component (K); T is the changed

temperature (K); K 0 is the thermo-elastic factor (MPa-1); and  is the changed sum of
three principal stresses (   i ,i  1,2,3 , MPa).
As for an isotropic linear elastic solid loaded bi-axially with a free surface, the surface
physical temperature variation is tightly correlated with the sum of two principal stresses
(   i ,i  1,2 ):
T   / C p  [T   ( 1   2 )] (2)
Here: T is the surface absolute temperature of a loaded solid (K); T is the changed
temperature (K);  is the factor of linear expansion (K-1);  is the solid density (Kg· m-3);
Cp is thermal capacity of solid at normal atmosphere (J· Kg-1· K-1); 1  2 are the two
and
principal stresses (MPa). The thermo-elastic factor K is defined as K   /  C p .
For the mechanism of stress measurement with TSA and SPATE, the relationship between
the stress increment and the IRR signal based on equation (2) is as follows (Mounatin and
Webber, 1978):
( 1   2 )  Ath  S
(3)
S   ( 1   2 )  Ath 
1
Here: Ath is a comprehensive factor called as corrective factor, which is a function of solid
surface emissivity, solid surface physical temperature, solid thermo-elastic factor and three
parameters related to the IRR detector, unit in MPa· U-1. S is the increment of
thermo-elastic voltage signal detected (U).
3.1.2 Changed IRR temperature of loaded rock

If the slight change of rock surface emissivity, rock thermo-elastic factor and the physical
parameters of IRR detector during rock loading could be ignored, and if the changed rock
surface physical temperature cannot be ignored due to the thermal exchange and the
frictional thermal, the relationship between Ath and changed rock surface physical
temperature could be expressed as Ath    T 1 . The detected IRR signal S is a direct
representation of surface IRR temperature, i.e., IRRT    S . Hence, the following
equation for changed IRR temperature could be deduced:
IRRT     1  T   ( 1   2 ) (4)
Here: IRRT is the changed IRR temperature (K);  is a constant correction factor
related to rock surface’s emisivity, rock thermo-elastic factor and three parameters of IRR
detector, in unit MPa·K·U-1;  is a transfer factor between detected voltage signal and IRR
temperature (K·U-1).
It means in equation (4) that the changed IRR temperature of rock surface is a direct
reflection of the changed sum of the two principal stresses. If no frictional thermal produced
and the thermal exchange is stable, the IRR temperature of rock surface is to rise with
loading, and the spatial-temporal evolution of surface IRR image will be stable. If there is no
frictional thermal produced but the thermal exchange is unstable, the surface IRR
temperature will be unstable, and the spatial-temporal evolution of surface IRR image will
also be unstable. If there is frictional thermal produced inside and conducted to rock surface,
both the thermal exchange and the surface IRR temperature of rock surface will be unstable,
and the spatial-temporal evolution of surface IRR image will get complicated, as in Figures
2~5, 10, 11.
Especially, in condition of compressive shearing and fictional sliding, there is a large
amount of frictional thermal produced in the friction zone, which is to cause the rise of
physical temperature in friction zone. The rock surface temperature will rise if the thermal
conduction from the friction zone can reach to rock surface. Hence, the IRR image anomaly
will be distinguished and sometimes be large as a combined effect of rock stress and
frictional thermal, as in Figures 6~9.
3.2 Remote Sensing Rock Mechanics model
3.2.1 Energy-Related IRR from Loaded Rock

As a relatively independent closed system comprised of loader head, rock sample and
environmental air, as in Figure 24, the rock deformation, rock fracturing and rock hazard are
all of a complex process of energy input and consumption. If the possible chemical reactions
inside a loaded rock can be ignored, the inputted energy of a loaded rock will include the
mechanical work from loader and the heat input through positive thermal exchange from
loader head and environmental air. The energy consumption by the loaded rock is much more
complex including the energy accumulation in rock and the energy dissipation from rock.
Energy Input Energy consumption
Mechanical work Energy dissipation Energy accumulation
Heat input Kinetic energy Deformation energy
Light radiation Surface energy
Acoustic emission Friction action and heat
Radio frequency emission
Heat output
IR & microwave
Surface physical temperature
Surface IRR temperature

Fig. 24. The IRR mechanism related to the energy accumulation and consumption of a
loaded rock
The energy accumulation in a loaded rock includes the positive elastic-plastic deformation
energy of rock (the positive change of oscillation and rotation energy of mineral molecules),
the surface energy of new produced fractures or fissures, and the friction actions between
mineral molecules, grains, joints, fissures and fractures inside the rock as well as thus
produced frictional thermal. The energy dissipation from loaded rock includes the negative
thermal exchange with the loader head and/or environmental air (i.e., heat output), the
kinetic energy of departed fragments of fractured rock, the light radiation, acoustic emission,
radio frequency emission and IR & microwave radiation.
The thermal exchange and the friction action are to change the heat state of a loaded rock,
and the rock surface physical temperature is a direct index reflecting the heat state of the
loaded rock. Stephen-Boltzmann law states that the IRR strength (radiation flux density) of
any material, at temperature above absolute zero degree, is biquadratic to its surface
physical temperature. Crystal Physics states that the energy jump of molecules oscillation
and/or rotation due to the change of molecules distance, resulting from deformation, is an
important mechanism of electromagnetic radiation. Hence, rock surface IRR is a

comprehensive effect of rock deformation and rock surface thermal state. Rock surface IRR
temperature could be a detective index reflecting rock surface physical temperature and
rock surface deformation field, which implicating the complex physical-mechanical process
inside the loaded rock.
In spite of thermal exchange and plastic deformation, the thermo-elastic effect and the
frictional thermal are two of the main mechanics of changed IRR from loaded brittle rock. In
the stage of elastic deformation, the thermo-elastic effect is the main cause; while in the
stage of plastic deformation or fracturing, the friction-thermal effect plays a great role. At
the moment of rock fracturing or hazard, the fraction-heat effect gets more distinguished.
The frictional thermal effect depends on two factors being frictional force (decided by
normal stress and frictional coefficient) and frictional speed respectively. The larger the
frictional force and the quicker the frictional speed, the more the frictional heat.
3.2.2 RSRM-model based on independent system

The rock sample, load header and environment air could be taken as a closed independent
system, as in Figure 25, and the energy of the loaded rock sample is in a balance state as
follows:
D
G
D
F
T
H
Fig. 25. The energy balance of a loaded rock sample in a closed independent system
M  T  D  F  H  G (5)
Here: M — — The inputted mechanical energy from loader, J; be positive;

T — — The inputted thermal energy from loader and environment air, J; be
positive;
D — — The produced deformation energy of rock sample in elastic and plastic
state, J; be positive;
F — — The sum of consumed rock fracturing energy and formed fracture surface
energy, J; be positive;
H — —The heat energy increment of rock decided by its physical temperature, J; be
positive if temperature rise or be negative if temperature drop;
G — — The energy consumed by the desorbing and escaping of pore gas in rock
samples, J; be positive.
From equation (5) we have:
H  ( M  T )  (D  F  G ) (6)
The change of heat energy will result in the change of physical temperature，surface
radiation energy and IRR temperature of loaded rock samples. Referring to
Stephen-Boltzmann law, the AIRT is a direct index of rock radiation energy, and the
change of AIRT ( AIRT ) must have certain a relationship with the physical temperature
(T) of loaded rock sample as:
E IR  f (AIRT )  T 4 (7)
Here: EIR — — The radiation energy of a loaded rock, J;

AIRT — —The surface average IRR temperature of a loaded rock, K;
 — — The radiation factor of rock sample, 0    1 ;
 — —The constant of Stephen-Boltzmann,   5.6679  10 8 , J×m－2×K－4；
Thermo-elastic effect is the basic mechanism of changed IRR from a loaded rock
sample. Moreover, the desorbing and the escaping of pore gas, the expanding of initial
fissures or joints, the friction between fissures, fractures, joints and grains, the thermal
transfer between the rock, loader header and environment air, and the radiation from the
environment all are to have thermal effect on loaded rock samples. The thermal state is
comprehensively affected by the six factors which are rock stress, pore gas desorbing &
escaping, rock fracturing, heat transferring, rock frictionating and environment radiation
respectively. The equation could be expressed as follows:
AIRT  f [( 1   2 ), G , F1 , H , F2 , E ]

(8)
 f 1 (t )  f 2 (t )  f 3 (t )  f 4 (t )  f 5 (t )  f 6 (t )
Here: AIRT — —The detected change of AIRT , K; be positive if rise, be negative if

drop;
f1 (t ) — —IRR temperature change due to thermo-elastic effect ( ( 1   2 ) , K; be
positive or negative;
f 2 (t ) — —IRR temperature change due to pore gas adsorbing & escaping ( G ), K;
be negative；
f 3 (t ) — —IRR temperature change due to the production of new fractures and the
expansion of initial fissures, joints and new produced fractures ( F1 ), K;
be negative;
f 4 (t ) — —IRR temperature change due to frictional thermal ( F2 ), K; be positive;
f 5 (t ) — —IRR temperature change due to heat transfer ( H ), K; be positive or

negative.
f 6 (t ) — — IRR temperature change due to environment radiation ( E ), K; be
positive.
1) Thermo-elastic effect: f1 (t )
Referring to thermo-elastic theory and equation (4), f1 (t ) could be calculated as:
f1 (t )     1  T  ( 1   2 ) (9)
In condition of uni-axially compressive loading, the load is to cause temperature rise in that
2 is constant zero and the positive  1 will linearly increase with loading. If  1 is
positive, the f1 (t ) will be positive; If  1 is negative, the f1 (t ) will be negative. Hence,
the f1 (t ) will be positive before the compressive stress peak, and will turn to be negative
after the compressive stress peak.
In condition of uni-axially tensile loading, the load is to cause temperature drop in that 2
is constant zero and the negative  1 will linearly increase with loading. If  1 is positive,
the f1 (t ) will be positive; If  1 is negative, the f1 (t ) will be negative. Hence, the
f1 (t ) will be negative before the tensile stress peak, and will turn to be positive after the
tensile stress peak.
In condition of compressively-sheared loading, the rock sample will be compressed by the
normal component of the load. However, as to the central shearing plane, it will suffer not
only compressive stress but also shearing stress. The 1 refers to the positive compressive
stress normal to the shearing plane, while the 2 refers to the shearing stress which is
actually negative tensile stress along the shearing plane. Hence, the AIRT is decided by the
sum of compressive stress and the tensile stress. If  ( 1   2 )  0 , the f 1 (t ) will be
positive; If  ( 1   2 )  0 , the f1 (t ) will be negative. As to Figure 1c, it’s easy to know
that  ( 1   2 ) will be positive if the shearing angle   45 ,  ( 1   2 ) will be zero if
0
the shearing angle   45 , ( 1   2 ) will be negative if the shearing angle   45 .

0 0
2) Pore gas desorbing & escaping effect: f 2 (t )

Any rock has pores of different size more or less inside. Some rock, especially the
sedimentary rock, may has certain gas, such as CH4, CO2, CO and O2, enclosed in the pores
or/and absorbed on the pore surface (Wang, 2003; Yang et al., 1999). Usually, most of the
pores are enclosed and the gas molecules stay inside both in free gassy state and absorbed
state. If the rock is loaded and suffers deformation, the volume of pores will decrease which
results in the escape of gas from the pores. Once the load and deformation cause the
enclosed pores getting fractured, the gassy molecules will escape firstly and the absorbed
molecules will get desorbed to be gassy molecules and escape later. Both desorbing and
escaping actions need to make use of heat energy from the rock, and thus will result in the
AIRT drop of rock surface. Hence, the f 2 (t ) is always negative, and the more the pores
and gas enclosed, the more the negative effect of f 2 (t ) .
3) Fracture effect: f 3 (t )
With loading and deforming, the rock is to get fractured. The new produced fractures together
with the initial fissures and joints will extend both in width and length. The production of new
fractures needs to consume energy, and the extension of fissures, joints and fractures also
needs to consume energy. Hence, the f 3 (t ) is always negative, and the more the fractures
produced and fissures, joints and fractures extended, the more the negative effect of f 3 (t ) .
4) Frictional thermal effect: f 4 (t )
With loading and deforming, the friction action is to occur between rock fissures, rock joints,
rock grains, and new produced fractures. The friction action could be interpreted as: 1) at
the beginning stage of loading, the friction may only be resulted from between rock fissures
and between rock joints; 2) later, rock deformation increases with loading, new fractures are
produced, and the friction between rock grains and between new produced fractures will
join in; 3) finally, at the ending stage of loading, the rock deformation and fractures will be
sufficiently developed, and the frictions between rock grains and between new produced
fractures will be the chief contributors to the frictional thermal. In a word, the f 4 (t ) is
always positive no matter what is the principle friction factor. The more the friction, the
more the positive effect of f 4 (t ) .
5) Heat transfer effect: f 5 (t )
In the process of loading and inside the respectively independent loading system, the heat
exchange is inevitable between the rock sample and the load header, the shearing platen or
cushion-blocks, the surrounded atmosphere, etc. If the current temperature of rock sample is
higher than the others, the heat of rock sample will be transferred out to whose temperature
is lower. If the current temperature of rock sample is lower than the others, the heat of the
others will be transferred into rock sample. Hence, the temperature of rock sample is a
dynamic balance behavior between the heat transferred in and the heat transferred out. If
the heat transferred in is more than that transferred out, the f 5 (t ) will be positive;
otherwise, it will be negative.
6) Environment-radiation effect: f 6 (t )
The IRR detected by infrared imager includes not only the direct radiation from rock surface
itself, but also the reflected radiation from environment. In laboratory, the chief
environment radiations are the scattering sunshine, the moving human bodies and the
illumination lamp. For the uncertain change of scattering sunshine, the movement of human
bodies before the loaded rock sample, and the fluctuation of illumination light, the
environment radiation effect on rock sample will be random. Hence, sometimes f 6 (t ) may
be positive, but sometimes f 6 (t ) be negative. To eliminating the environment-radiation
effect, the human bodies inside the laboratory were not permitted to move during testing
process, the illumination lights were turned off, and the windows as well as its curtains
were closed. Furthermore, some experiments were conducted in the evening so as to avoid
the scattering sunshine completely.
3.3 Experiment Interpretation with RSRM-Model

Due to the comprehensive effects of f1 (t ) ~ f 6 (t ) , the evolution of AIRT will be complex
and will result in different possibility for abnormal AIRT precursors in different rock
loading conditions, which including the loading scheme, rock type, and environment
parameters. The following discussions are based on that f 5 (t ) and f 6 (t ) can be ignored.
3.3.1 Uni-axial loading experiments

For uni-axially loaded rock, f 4 (t ) will take place only after that the rock has sufficiently
deformed and the fractures have sufficiently developed.
At loading stage I and stage II, the rock surface AIRT are decided by f1 (t ) and f 2 (t ) . If
the loaded rock is igneous rock or metamorphic rock, f 2 (t ) is very rare since no gas
absorbed in its pores usually, and the AIRT will rise with loading. If the loaded rock is
sediment rock and with gas closed and absorbed in pores, f 2 (t ) is inevitable, and the
AIRT will rise if f 1 (t )  f 2 (t ) , or be constant if f 1 (t )  f 2 (t ) , or drop if f 1 (t )  f 2 (t ) .
At loading stage III, fractures get sufficiently developed and pores get seriously damaged.
The f 3 (t ) begins to has more and more effect on evolution process of AIRT. AIRT will rise
if f1 (t )  f 2 (t )  f 3 (t ) , or be constant if f1 (t )  f 2 (t )  f 3 (t ) , or drop if
f1 (t )  f 2 (t )  f 3 (t ) .
At loading stage IV, the friction action starts and f 4 (t ) begins to have more and more
effect on the evolution of AIRT. AIRT will rise if [ f1 (t )  f 4 (t )]  f 2 (t )  f 3 (t ) , or be
constant if [ f1 (t )  f 4 (t )]  f 2 (t )  f 3 (t ) , or drop if [ f1 (t )  f 4 (t )]  f 2 (t )  f 3 (t ) .
Since f 2 (t ) is very rare for igneous rock or metamorphic rock, the rise speed of AIRT of
loaded igneous rock or metamorphic rock will get fast at stage IV, and the speed turning
point is suggested to be the precursors point.
3.3.2 Compressive shearing experiments

For compressively sheared rock, not only f1 (t ) but also f 4 (t ) is decided by shearing
angle (  ). If   45 , f1 (t ) will be positive for  ( 1   2 )  0 , and the friction action
0
will be much strong in that  1 , which is normal to the friction plane, is large. If   45 ,
0
f1 (t ) will be zero since ( 1   2 )  0 . If   450 , f1 (t ) will be negative for

( 1   2 )  0 , and the friction action will be much week since  1 is slight.
3.3.3 Biaxial loading experiments

For bi-axially loaded rock, f1 (t ) will always be positive. As to bi-axially loaded en echelon
faults, collinearly and non-collinearly disjointed faults, f 2 (t ) , f 3 (t ) and f 4 (t ) will occur
simultaneously, and the evolution of AIRT will be fluctuated. If

[ f1 (t )  f 4 (t )]  f 2 (t )＋f 3 (t ) , AIRT will rise; if [ f1 (t )  f 4 (t )]＝ f 2 (t )＋f 3 (t ) , AIRT will
keep in the same level, and if [ f1 (t )  f 4 (t )]  f 2 (t )＋f 3 (t ) , AIRT will drop. Usually, the
fact is that [ f1 (t )  f 4 (t )]  f 2 (t )＋f 3 (t ) at the fracturing stage, and there is a short drop
of AIRT, which is called as ‘silence’ before EQ (Ohtake et al., 1981). However, f 4 (t ) will be
an important factor at the later stage of loading for the concentrated formation of fractures
and friction in the disjointed zone, and the final state of AIRT will be rise for
[ f1 (t )  f 4 (t )]  f 2 (t )＋f 3 (t ) .
4. Earthquake Thermal Infrared Anomalies

The prediction of EQ is very difficult, but it's not impossible. A number of signs warning of
EQs, such as foreshock activities, peculiar animal behaviour, increased low frequency
EM-noise, concentrations of radon in water and air, ionosphere and magnetosphere
perturbation, radio frequency emissions, terrestrial gas emanations, EQ clouds, and satellite
TIR anomalies, have been proposed and reported during the past centuries. Satellite TIR
anomaly was firstly reported in 1989, and had been repeatedly verified in the world during
the past 20 years. It is becoming a prospecting space observation technology for seismic
activity monitoring and for EQ predicating.
4.1 General Features of EQ TIR Anomaly

Gorny (1988) firstly reported that there were large area of TIR anomalies in METEO satellite
remote sensing images, spatial resolution being 5 Km and wave length being 10.5~12.5μm,
before many moderate-strong tectonic EQs in the mid-Asia and the east-Mediterranean
region. Tronin (1996) analyzed 10000 about TIR images of channel AVHRR-2 of NOAA in
ten years for the mid-Asia, and reached that there existed average anomaly, 1~5℃, before
the EQs at this region, and that there was obvious statistical relations between the EQ and
TIR anomaly. Qiang (1990), Cui (1999), Liu (1999), Xu (2000), Zhang(2002), Ouzounov(2004),
Arun (2005), Liu (2007), Wu (2008) also reported that there occurred TIR anomalies in
satellite images (NOAA, FY, MODIS) days before more than 100 EQs in Asia (China, India,
Iran, Japan, Kamchatka, Pakistan, Turkey) and Europe (Italy, Greece, Spain). Analysis to all
the reported cases, it was uncovered that satellite TIR anomaly before EQ has the following
features generally (Wu et al., 2009):
1) Temporal features: Satellite TIR anomaly usually appears 1~26 days before shock,
and reaches to its peak 1~2 days before shock, and will disappeared soon after shock.
2) Spatial features: The spatial distribution and geometrical shape of TIR anomaly is
tightly related with tectonic structures such as plate borders and active faults. With
the EQ impending, the TIR anomaly will move to or extend to gradually the
coming epicentre along the structures.
3) Temperature features: the temperature of the TIR anomaly is usually 2～6℃
higher than that of outside or surrounding the TIR anomaly.
4) Magnitude features: there are positive relations somewhat between the TIR
anomaly energy (the anomaly area times the anomaly temperature) and the
magnitude of future shock.
4.2 Anomalies Interpretation with RSRM-Model

The RSRM experimental results are applied to analyze satellite remote sensed TIR anomalies
before several strong EQs in Asia. Referring to the seismogenic mechanisms, the satellite TIR
anomalies are in good accordance with the detected IRR anomalies from rock fracturing and
seismogenic experiments with fault system being simulated with disjointed faults and
intersected faults.
4.2.1 Dongsha Ms5.9 EQ 1992 in Taiwan, China

Dongsha Ms5.9 EQ occurred in Taiwan on Sept 14, 1992. The NOAA satellite images show that
there was TIR anomaly appeared along the regional faults and downfaulted basins before
shock as in Figure 26 (Wu et al., 2006c). There was an isolated and spoon-shaped high
temperature area to the southwest of Taiwan Island 25 days before shock (Aug 19, 1992). The
head of ‘spoon’ locates in the downfaulted basin-Ⅳ and the handle of ‘spoon’ distributes
along fault-6. Satellite TIR image on Aug 22 shows that the high temperature on ‘spoon head’
diminished, but the TIR anomaly on ‘spoon handle’ became wide. Later, the anomaly moves
gradually to the epicenter. TIR image on Sept 9 shows that a large area of high temperature
had appeared around the coming epicenter, and the maximum temperature (in brown color)
appeared south-close to the coming epicenter. It indicated that the TIR anomaly was consistent
with the regional tectonic structures (faults and basins) in spatial position and geometry.
110 112 114 116 118°E 111 113 115 117 119°E
24
epicenter epicenter epicenter
2 2
Ⅰ Ⅱ 2 2
1 1
Ⅳ
Aug 19 1 Aug 22 1 Sept 9
270K 280K 290K 300K 270K 280K 290K 300K
Fig. 26. Satellite TIR anomaly images before Dongsha Ms5.9 EQ (Sept 14, 1992)
4.2.2 Zhangbei Ms6.2 EQ 1998 in China

Zhangbei Ms6.2 EQ occurred in China at 11:50 am, Jan 10, 1998. With the overlay of the
active fault system investigated and deduced, in red lines, on night time NOAA satellite
infrared images, as in Figure 27, it was discovered that (Wu et al., 2007b): 1) 18 days before
shock, Dec 28, 1997, the TIR images on land and sea surface are basically normal, and the
contours of TIR temperature field of Bohai Bay appeared to be along the coastline; 2) 5 days
before shock, Jan 5, 1998, there occurred a positive TIR anomaly strip from Bohai Bay to
Zhangbei passing through Beijing, its temperature was 3°C higher than that of outside, and
the contours of TIR temperature field of Bohai Bay got offset the coastline; 3) 1 day before
shock, Jan 8, 1998, the positive TIR anomaly strip had got wider and its temperature was 7°C
higher than that of outside, and the contours of TIR temperature field of Bohai Bay got in
accordance with the positive TIR strip; 4) 1 day after shock, Jan 11, 1998, the pattern of TIR
images on the surface of land and sea turned to be normal again.

There are several disjointed active faults along the positive TIR strip, including a possible
uncovered great active deep fault going from Bohai Bay to Zhangbei. Besides, to the
southwest of Zhangbei, there are two small active faults pointing to Zhangbei. If extend the
three faults respectively towards Zhangbei, it could be found that Zhangbei is exactly the
intersection point. Hence, the tectonic background around the epicenter is basically
comprised of three intersected active faults as two groups, i.e., the primary fault is the
independent one from Bohai Bay to Zhangbei, and the secondary two act as another group.
The two groups of faults split basically the regional crust into three geo-blocks (geo-block A, B
and C1+C2) as the RSRM experiment model in Figure 28b (rock block A, B and C), and the
secondary two faults act as the two sides of an acute wedge-shaped geo-block (geo-block C2)
as the RSRM experiment model in Figure 28d (rock block C). The geometric features of the
faults and the spatio-temporal features of TIR anomaly were similar. Therefore, the mechanical
mechanism of Zhangbei EQ should be classified to be the failure of an intersected active fault
system, which with an acute wedge-shaped geo-block being its secondary active object.
19pm, Dec 28,1997 19pm, Jan 01,1998
Zhangbei Zhangbei
Beijing Beijing
Bohai Bohai
Bay
19pm, Jan 08, 1998 18pm, Jan 11, 1998
A
Zhangbei Zhangbei
C1 Beijing
Beijing
C2
Bohai Bohai
B Bay Bay
Fig. 27. The TIR images overlaid with detailed active faults before and after Zhangbei Ms6.2
EQ (Jan 10, 1998)
A C
C
B B A
a) 90o intersection b) 75 o intersection c) 60 o intersection d) 40 o intersection

Fig. 28. The IRR anomaly features of simulated tectonic activities due to bi-axially load (Wu
et al., 2007b)
4.2.3 Izmit Ms 7.8 EQ 1999 in Turkey

Izmit Ms 7.8 EQ occurred in Turkey on Aug 17, 1999. In the epicenter zone, there were two en
echelon tectonic faults, and the epicenter is 45km about to the west of the disjointed place of
the two en echelon faults, as in Figure 29. With thermal image of Aug 1&2 being the reference,
the differential thermal images from Aug 6 to Aug 26 were obtained, and it was revealed that
there were NOAA satellite TIR anomaly at the disjointed zone of two en echelon faults since
Aug 6 , 11 days before shock (Tronin, 2000), as in Figure 29. The RSRM experiments on the
simulation of tectonic EQ due to the fracturing of disjointed faults had revealed that there were
IRR anomaly increment and concentrated deformation at the disjointed zone, as in Figure 30.
Obviously, the spatial features of NOAA TIR anomaly before Izmit Ms 7.8 EQ were the same
as that of IRR anomaly before the fracturing of disjointed faults.
epicenter epicenter epicenter
Aug 6 Aug 11 Aug 17
Fig. 29. TIR anomalies 0~11 days before Izmit Ms 7.8 EQ (Aug 17, 1999)
Fig. 30. The IRR anomaly and deformation concentration of a bi-axially loaded granite
sample with disjointed faults
4.2.4 Hengchun Ms7.2 EQ 2006 in Taiwan, China

Hengchun Ms7.2 EQ occurred in Taiwan at 12:26pm, Dec 26, 2006. TIR images from
stationary satellite FY-2 showed that there was TIR anomaly nearby the coming epicenter.
The TIR anomaly appeared to the east of Philippines six days before shock, and moved
gradually toward west to Philippines. Later, the anomaly changed its direction to the north,
and the temperature increased 10oC about one day before shock. Figure 31 shows that the
satellite TIR images around Hengchun on Dec 25, 2006 (Liu et al., 2007). A disjointed
thermal strip (in orange color) appeared on the southwest of Taiwan at 1:00am, and
developed gradually to be an X-shaped thermal anomaly zone at 10:00am. The epicenter

located closely to the cross point of the X-shaped zones. The evolution process of satellite
TIR images was very similar to that of X-shaped thermal IRR anomaly in RSRM experiments
as in Figure 2.
4.2.5 Wenchuan Ms8.0 EQ 2008 in China

Wenchuan Ms8.0 EQ occurred in China at 14:28 am, May 12, 2008. The epicenter locates at
the transferring zone of Tibetan Plateau to Sichuan Basin, which is only 92Km to the NW of
Chengdu, the capital of Sichuan province. Analysis to FY-2C TIR images, as in Figure 32, it
is discovered that there was an high temperature strip with length of 3 000 km appeared on
April 23, 20 days before shock, which started from India Plate and developed to northeast
along the east front of Tibetan and Loess Plateau (Wu et al., 2008). The cause might be the
great accumulation of fictional sliding stress along the east foreland of Tibetan and Loess
Plateau due to the subduction of India Plate to Euro-Asia Plate. The east foreland of Tibetan
and Loess Plateau act as the fictional sliding plane between the west part of China
(including Tibetan and Loess Plateau) and the east part of China (including the North,
Middle and Southwest China). The evolution of satellite TIR images was similar to that of
detected strip-shaped IRR anomaly in the fictional sliding experiments as in Figure 8 and
Figure 9, which shows the evolution of TIR anomaly along the sliding plane.
1:00am 4:00am 10:00am 14:00am

Fig.31. Satellite TIR images one day before Hengchun Ms7.2 EQ (Dec 25, 2006)
14:30Pm, Apr 23 17:30Pm, May 13:30Pm, May

2 12
Fig.32. Satellite TIR images before Wenchuan Ms8.0 EQ (May 12, 2008)
5. Future Researches
5.1 On RSRM Experiment
As a detectable remote sensing signal related with rock stress and physical temperature, IRR
is a meaningful index for studying rock load, rock deformation, rock fracturing and rock
hazard. The temporal evolution of surface IRR from loaded rock is the comprehensive effect
of rock thermo-elastic acting, pore gas desorbing & escaping, fractures producing &
extending, rock frictionating, heat transferring and environment radiation. The IRR image
anomaly referring to the spatial-temporal evolution of IRR from loaded rock is an important
precursor for rock fracturing, and will be meaningful for the predication of geo-hazards
including tectonic EQ. For the practical application of RSRM, deeper research on IRR
imaging detection quantitatively and specially on rock stress and rock hazard for
experimental rock mechanics, rock engineering, tectonic activity and strong EQ is
demanded.
The mechanism of experimental detected IRR anomaly can be theoretically interpreted by
taking the load header, the rock sample and the environment to be a closed independent
system in energy balance state. There are two of the main rock-physics mechanisms,
respectively being thermo-mechanical coupling and frictional thermal due to tectonic stress,
rock fracturing and fictional sliding, for the change IRR from loaded rock samples (Wu, et
al., 2006c). Besides, positive hole (P-hole) activation due to piezoelectricity was suggested to
be another mechanism of IRR from loaded quartz-bearing igneous rock (Freund, 2002), such
as granite, basalt, diorite, and gabbro.
To search for scientific interpretation on the relations among satellite TIR anomaly, rock
stress and experimentally detected IRR anomaly, the EM transferring process from
underground rock body to satellite sensors, through lithosphere, earth surface coversphere
(including soil layers, water bodies and vegetations), the atmosphere and lithosphere,
should be systematically studied. Pulinets pointed out that the incubation of an EQ is to
disturb the ionosphere (Pulinets, 1998), and it was suggested that
lithosphere-atmosphere-ionosphere (LAI) coupling is the mechanism of satellite TIR
anomaly before strong EQ (Molchanov et al., 2004). Nevertheless, the action of earth surface
coversphere on the transferring and the magnifying of EM signals from underground
loaded rock to atmosphere should not be ignored, even if its physical mechanism are not
clear. For the scientific interpretation of satellite TIR anomaly before strong EQ, the
lithosphere-coversphere-ionosphere (LCA) coupling is the key, while for the scientific
interpretation of ionosphere anomaly, the lithosphere-coversphere-atmosphere-ionosphere
(LCAI) coupling should be focused. However, present experiments on LAI, LCA and LCAI
coupling are rather insufficient. Future experiments specially designed to uncover the
mechanisms & laws and to construct the models & quantitative equations of LAI, LCA and
LCAI couplings are expected.
5.2 On EQ Thermal Infrared Anomaly

Although there are uncertain influences from meteorological variation, satellite TIR
anomaly has quite different identification features from that of unseismology-resulted TIR
anomaly. Satellite TIR remote sensing is becoming a prospecting technique for monitoring
tectonic activities and for predicting strong tectonic EQ, which provide a negativism to
that EQ cannot be predicted. Anyway, the practical predication of EQ is not so easy. The
regional tectonic background and the active fault system have extremely important affects
on the incubation of EQ and the TIR anomaly. Especially, the intersected faults,
compressively-sheared faults, and disjointed faults are to control the location and the
routing of the spatial evolution of satellite TIR anomaly, and the brightest spot of TIR
anomaly along the fault, or at the intersection point, or at disjointed zone of faults might
foretell the possible epicenter (Wu et al., 2007b).

First of all, massive observation information including crust stress, land deformation,
atmosphere components, underground water, surface and near-surface temperature,
satellite remote sensing TIR anomaly and EM disturbance in ionosphere should be
integrated together for data fusion and cross checking to analyze comprehensively the
tectonic activity and rock fracturing process. A grid-based distributed database and analysis
tools on TIR remote sensing images, with global and regional tectonic structures being its
background, should be set up to assist the extraction of EQ TIR anomaly in different
temporal and spatial scale. Besides, a quantitative model for tectonic activity analysis and
for EQ magnitude predication based on TIR anomaly should be developed.
The GEOSS under construction is to provide an integral and integrated monitoring on earth
environment, Geo-hazard and global disasters. A generalized remote sensing (GRS) based
on the integration of space-based, aero-based, near-surface based, in-situ based and
underground-based monitoring is forming in the world (Wu and Liu, 2007). The
international broad and sincere cooperation, inside the framework of GEOSS without
discipline exclusion and data privacy, between seismologists, remote sensing scientists,
meteorologists, geophysical scientists, geochemist and spatial information scientists in good
faith is expected. Besides, a powerful spatial information system, especially for EQ early
warning and short-coming prediction based on GEOSS, should be developed. It should has
powerful functions such as massive information classification, smart theme mapping, easy
map-layer overlay, fast features extraction, effective data fusion, intelligent data mine,
powerful knowledge discovery, and easy access and share.
A possible technical procedure based on GRS for satellite TIR anomaly monitoring,
analyzing and early warning of tectonic EQ inside the framework of GEOSS might be that: 1)
the seismological geology background being the foundation of satellite TIR anomaly
analysis; 2) the long-term GPS continuous monitoring and D-InSAR measurement being the
guidance of tectonic stress detection and active fault identification; 3) the underground
water temperature, near surface air temperature, radon & green gas, structural cloud
anomaly dairy monitoring being the forerunner for preliminary identification of coming EQ;
4) the anomaly analysis of satellite TIR, NCEP temperature and ionosphere disturbance
being the dominant for early warning of temporal-spatial-magnitude parameters of EQ; and
5) the additional celestial stress on active faults being a special leading disturbance for
possible tectonic EQ.
6. References
Arun K. & Swapnamita C.(2005). Thermal remote sensing technique in the study of
pre-earthquake thermal anomalies. J. Ind. Geophys. Union, 9(3):197-207
Brady B. & Rowell G. (1986). Laboratory investigation for the electrodynamics of rock
fracture. Nature, 321(29): 488-492
Cui C., Zhang J. & Xiao Q.(1999). Monitoring the thermal IR anomaly of Zhangbei
earthquake precursor by satellite remote sensing technique, Proc. 20th Asia Remote
Sensing Congress, pp.1179-1184, Hongkong
Deng M., Qian J., Yin J., et al.(2001). Research on the application of infrared remote sensing in
the stability monitoring and unstability prediction of large concrete engineering.
Chinese J Rock Mech. Engi., 20(2):147-50.
Freund F.(2002). Charge generation and propagation in igneous rocks. J. Geodynamics, 33(4-5):
543-570.
Geng N., Cui C. & Deng MD, et al.(1992). Remote sensing detection on rock fracturing
experiment and the beginning of Remote Sensing Rock Mechanics. ACTA
Seismologic SINICA, 14(supp) : 645-52.
Gorny V., Salman A., Tronin, A., et al.(1998). The earth outgoing IR radiation as an indicator
of seismic activity, Proc. Acad. Sci. USSR, 30(1): 67-69.
Hudson J. & Harrison J.(1997). Engineering rock mechanics, Elsevier Science Inc., New York,
pp.85-112
Liu D., Peng K., Liu W., et al. (1999). Thermal omens before earthquake. Acta Seismologica
Sinica, 12(6):710-715
Liu S., Wu L., Li J. et al.(2007). Features and mechanism of the satellite thermal infrared
anomaly before Henchun earthquake in Taiwan Region. Science & Technology
Review, 25(6): 32-37
Liu S., Wu L., Wu Y., et al.(2002). Quantitative study on the thermal infrared radiation of
dark mineral rock in condition of uni-axial loading. Chinese J. Rock Mech. & Engi.,
21(11): 1585-89
Luong M.(1990). Infrared thermovision of damage processes in concrete and rock. Engi.
Fracture Mechanics, 35 (1~3): 127-35.
Martelli G., Smith P. & Woodward A.(1989). Light, radiofrequency emission and ionization
effects associated with rock fracture. Geophy. J. Int., 98(2): 397-401.
Molchanov O., Fedorov E., Schekotov A. et al.(2004). Lithosphere-atmosphere-ionosphere
coupling as governing mechanism for preseismic short-term events in atmosphere
and ionosphere. Natural Hazards and Earth System Sciences, 4: 757–767
Mounatin D. & Webber J.(1978). Stress pattern analysis by thermal emission (SPATE), Proc.
Soc. Photo-Opt. Inst. Engi, 164:189.
Nicolas A., Bouchez J., Blaise J., et al.(1977). Geological aspects of deformation in continental
shear zones. Tectonophysics, 42(1):55-73
Ohtake M., Matumoto T. & Latham G.(1981). Evaluation of the forecast of the 1978 Oaxaca
Southern Mexico earthquake based on a precursory seismic quiescence. In:
Earthquake Prediction Maurice Ewing Series American Geophysics Union 4）, pp.53-62
Ouzounov D. & Freund F.(2004). Mid-infrared emission prior to strong earthquakes
analyzed by remote sensing data. Adv. Space Res. 33: 268-273.
Pulinets S. (1998). Seismic activity as a source of the ionospheric variability. Adv. Space Res.,
22(6): 903–907
Qiang Z., Xu X. & Dian C.(1990 ). Thermal infrared anomaly-precursor of impeding
earthquakes. Chinese Science Bulletin, 35(17): 1324-1327
Renata D.(1977). Electromagnetic phenomena associated with earthquakes. Geophsical survey,
3(2):157-174.
Sibson R. et al.(1980). Power dissipation and stress levels on faults in the upper crust. J.
Geophysical research, 85(B11): 6239-6247.
Thomson W.(1853). Trans. R. Soc. Edinburgh, 20: 83-261.
Tronin A.(1996). Satellite thermal survey—a new tool for the studies of seismoactive regions.
J Remote Sensing, 17(8): 1439-1455
Tronin A.(2000). http://www.iki.rssi.ru/earth/ppt/tronin.ppt
Wang X.(2003). A research on the inorganic carbon dioxide from rock by thermal simulation
experiments. Advance in Earth Science, 18(8):515-20
Wu L. & Liu S.(2007a). Generalized remote sensing for solid Earth hazards under condition
of GEOSS. Science & Technology Review, 25(6): 5-11
Wu L. & Wang J.(1998). Infrared radiation features of coal and rocks under loading. Int. J.
Rock Mech. & Min. Sci , 35(7):969-976
Wu L., Cui C., Geng N. et al.(2000). Remote Sensing Rock Mechanics (RSRM) and associated
experimental studies. Int J Rock Mech Min Sci, 37(6): 879–88.
Wu L., Li J. & Liu S.(2009). Infrared anomaly analysis based on reference fields for
earthquake remote sensing. Seismology Geology, 2009(in press)
Wu L., Li J., Xu X., et al.(2007b). Theoretical analysis to impending tectonic earthquake
warning based on satellite infrared anomaly, Proc. 2007 IEEE Int. Geosciences
&Remote Sensing Symposium, pp. 3723-3727
Wu L., Liu S, Shi W. et al.(2003). Experimental study on infrared anomaly of tectonic
earthquake, Proc. SPIE, Remote Sensing For Environmental Monitoring, GIS Applications
& Geology Ⅲ, 5239: 376-387.
Wu L., Liu S. & Wu Y.(2006c). The experiment evidences for tectonic earthquake forecasting
based on anomaly analysis on satellite infrared image, Proc. 2006 IEEE Int.
Geosciences &Remote Sensing Symposium, pp. 2158-216
Wu L., Liu S., Chen Y. et al.(2008). Satellite thermal infrared and cloud abnormities before
Wenchuan earthquake. Science & Technology Review, 26(10):28-29
Wu L., Liu S., Wu Y. et al.(2002). Changes in IR with rock deformation. Int. J. Rock Mech. &
Min. Sci., 39(6): 825-31
Wu L., Liu S., Wu Y., et al.(2004a). Remote Sensing Rock Mechanics (I)：laws of thermal
infrared radiation from disjointed jointed faults fracturing and its meanings for
tectonic earthquake omens. Chinese J Rock Mech. & Engi., 23(1): 24-30
Wu L., Liu S., Wu Y., et al.(2004b). Remote Sensing Rock Mechanics (Ⅱ)：laws of thermal
infrared radiation from bi-sheared faults friction sliding and its meanings for
tectonic earthquake omens. Chinese J Rock Mech. Engi. , 23(2): 192-198
Wu L., Liu S., Wu Y., et al.(2004c). Remote Sensing Rock Mechanics (Ⅳ): laws of thermal
infrared radiation from compressively-sheared fracturing rock and its meanings for
earthquake omens. Chinese J Rock Mech. & Engi., 23(4): 539-544
Wu L., Liu S., Wu Y., et al.(2006a). Precursors for rock fracturing and failure—part I: IRR
image abnormalities. Int J Rock Mech Mining Sci, 43(3): 473-482
Wu L., Liu S., Wu Y., et al.(2006b). Precursors for rock fracturing and failure—part II: IRR
T-curve abnormalities. Int J Rock Mech Mining Sci, 43(3): 483-493
Wu L., Liu S., Xu X., et al.(2004d). Remote Sensing Rock Mechanics (Ⅲ)：laws of thermal
infrared radiation and acoustic emission from friction sliding intersected faults and
its meanings for tectonic earthquake omens. Chinese J Rock Mech. & Engi., 23(3):
401-407
Xu X., Xu X. & Wang Y. (2000). Satellite infrared anomaly before Nantou Ms=7.6 earthquake
in Taiwan, China. ACTA Seismologica SINICA, 22(6): 656-659
Yamada I., Masuda K. & Murakami H.(1989). Electromagnetic and acoustic emission
associated with rock fracture. Phys. Earth Planet. Inter., 57(1-2):157-168.
Zhang Y. & Shen W.(2002). Satellite thermal infrared anomaly before the Xinjiang Qianhai
boder Ms8.1 earthquake. Northwestern Seismological J., 34(1): 1-4

Advances in Geoscience and Remote Sensing PDF

Uploaded by

Copyright:

Available Formats

Advances in Geoscience and Remote Sensing PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Advances in Geoscience and Remote Sensing PDF

Uploaded by

Copyright:

Available Formats

I

First published October 2009

Technical Editor: Goran Bajac

Advances in Geoscience and Remote Sensing,

Remote sensing is the acquisition of information of an object or phenomenon, by the

Dr. Gary Jedlovec

1. Enhancing the Unmixing Algorithm through the Spatial Data Modeling

2. Application of remote sensing to the estimation of sea ice thickness distribution 21

4. Environmental hazards in the El-Temsah Lake, Suez Canal district, Egypt 57

5. Air Pollution Determination Using Remote Sensing Technique 71

6. Some Features of The Volume Component of Radar Backscatter From Thick

7. Theoretical Modeling for Polarimetric Scattering and Information Retrieval

8. Polarimetric responses and scattering mechanisms of tropical forests

9. Discovery of user-generated geographic data using web search engines 207

11. Modeling of wake-vortex detection by a ground-based fiber LIDAR system 249

13. Spectroscopic Microwave Dielectric Model of Moist Soils 279

14. Automated Detection of Clouds in Satellite Imagery 303

15. Remote sensing and the disaster management cycle 317

17. Monitoring Soil Moisture from Spaceborne Passive Microwave Radiometers:

18. Multiwavelength Polarimetric Lidar for Foliage Obscured Man-Made

21. Soil moisture estimation using L-band radiometry 441

22. AggieAir: Towards Low-cost Cooperative Multispectral Remote Sensing Using

25. Dynamical Enhancement Technique for Geophysical Analysis of Remote

27. Application of Multi-Frequency Synthetic Aperture Radar (SAR)

29. Derivative analysis of hyperspectral oceanographic data 597

30. Electromagnetic Scattering Analysis of Simple Targets Embedded in Planar

31. On Performance of S-band FMCW Radar for Atmospheric Measurements 645

33. Improving Wetland Characterization with Multi-Sensor, Multi-Temporal SAR

Enhancing the Unmixing Algorithm through the

3. In-situ Measurements and Remote Sensing Data

Fig. 3. Flow chart of the methodological approach.

4.1 The SMM

Also, j = 1, 2, ..., n ( n = number of components), and i = 1, 2, ..., m (m = number of

spectral bands), with the following constraints: ∑f j = 1, and fj ≥0 for all

4.1.1 Endmember selection

Ris = f Chl rChl + f Dom rDom + f Iss rIss + ε i (2)

4.2 Ordinary Kriging

Water Anisotropy Structure Nugget Sill >range <range Model

4.3 Spatial regression between the SMM and in situ turbidity

Fig. 5. Kernel estimation of the spatial turbidity pattern.

y = Xβ + ε , with ε ~ N (0, σ 2 ) (7)

Where y is an (Nx1) vector of observations on a dependent variable taken at each of N

y = Xβ + u , with u = λWu + ε (9)

Where λ is a scalar spatial error parameter, and u is a spatially autocorrelated disturbance

AIC = −2 * LIK + 2k (10)

4.4 Evaluation of turbidity estimations

5. Results and Discussion

OLS Spatial regression

Turbidity = 99.72 + 221.71Iss + 199.46Chl + 186.53Dom (11)

Application of remote sensing to the estimation

Frequency of sea ice existence

2.1 Submarine sonar profiling

2.2 Moored ice profiling sonar

2.3 Video monitoring system

Fig. 5. Schematic picture of a video monitoring method.

(a) A schematic figure (cited from Shimoda et al. (1997))

2.4 Electromagnetic Induction Sounding

Fig. 7. Schematic picture of EM sounding system.

2.5 Satellite sensors