The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLI-B8, 2016

XXIII ISPRS Congress, 12–19 July 2016, Prague, Czech Republic

CLASSIFICATION OF LISS IV IMAGERY USING DECISION TREE METHODS

Amit Kumar Verma a, *,P.K Garg b, K.S Hari Prasad b, V.K Dadhwal c
a
Research Scholar, Geomatics Engineering Group, IIT Roorkee, Roorkee-247667, India - amit768@gmail.com
b
Professor, Civil Engineering Department, IIT Roorkee, Roorkee-247667, India - (gargpfce, suryafce)@iitr.ernet.in
c
Director, National Remote Sensing Centre, ISRO, Hyderabad-500042, India - dadhwalvk@hotmail.com

Commission VIII, WG VIII/8

KEY WORDS: Crop, Vegetation Indices, Texture, Decision Tree, Classification, LISS IV

ABSTRACT:

Image classification is a compulsory step in any remote sensing research . Classification uses the spectral information represented by the
digital numbers in one or more spectral bands and attempts to classify each individual pixel based on this spectral information. Crop
classification is the main concern of remote sensing applications for developing sustainable agriculture system. Vegetation indices
computed from satellite images gives a good indication of the presence of vegetation. It is an indicator that describes the greenness,
density and health of vegetation. Texture is also an important characteristics which is used to identifying objects or region of interest
is an image. This paper illustrate the use of decision tree method to classify the land in to crop land and non-crop land and to classify
different crops. In this paper we evaluate the possibility of crop classification using an integrated approach methods based on texture
property with different vegetation indices for single date LISS IV sensor 5.8 meter high spatial resolution data. Eleven vegetation
indices (NDVI, DVI, GEMI, GNDVI, MSAVI2, NDWI, NG, NR, NNIR, OSAVI and VI green) has been generated using green, red
and NIR band and then image is classified using decision tree method. The other approach is used integration of texture feature (mean,
variance, kurtosis and skewness) with these vegetation indices. A comparison has been done between these two methods. The results
indicate that inclusion of textural feature with vegetation indices can be effectively implemented to produce classified maps with 8.33%
higher accuracy for Indian satellite IRS-P6, LISS IV sensor images.

1. INTRODUCTION traditional hard classification techniques are parametric in nature

and they expect data to follow a Gaussian distribution, they have
Classification of satellite imagery plays an important role in been found to be performing poor results. In order to overcome
many application of remote sensing. Classification is a method this problem, non-parametric classification techniques such as
by which labels or class identifiers are attached to the pixels artificial neural network (ANN) and Decision tree classification
making up remotely sensed image on the basis of their spectral (DT) are used. The non-parametric property means that non
characteristics. These characteristics are generally measurements homogenous, non-normal and noisy data sets can be handled, as
of their spectral response in different wavebands. They also well as non-linear relations between features and classes, missing
include other attributes (e.g. Vegetation indices and Texture). values, and both numeric and categorical inputs (Quinlan, 1993).
Decision tree technique includes a set of binary rules that define
Spectral vegetation indices in remote sensing have been widely meaningful classes to be associated to individual pixels. Different
used for the assessment and analysis of the biomass, water, plant decision tree software are available to generate binary rules. The
and crops (Jackson and Huete, 1991). Vegetation indices (VI) software takes training set and supplementary data to define
enhances the spectral information and increases the separability effective rules. In this study decision tree approach is used for
of the classes of interest therefore it influences the quality of the land cover studies using LISS IV sensor data.
information derived from the remotely sensed data.

Texture is also one of the important characteristics used in 2. THE STUDY AREA
identifying objects or region in an image. It is an innate property
of virtually all surfaces which includes the pattern of different The selected area for this study is village Foloda which is located
crops in a field. Texture contains important information about the in Muzaffarnagar District, India, Measuring approximately 8 km2
structural arrangement of surfaces and their relationship to which lies between 29°36'22.70"N - 29°38'41.11"N Latitude
surrounding environment. In pixel-based approach, each pixel is and 77°47'50.26"E - 77°50'38.21"E Longitude. The ground truth
classified individually, without considering contextual information of the study area, including field wise information of
information. Several studies have explored the potential for using various crops and non-crop were collected using Trimble JUNO
these texture statistics derived from satellite imagery as input Global Positing System (GPS).The main crop growing in this
features for land cover classification (Haralick et al, 1973, Harris, region are sorghum, paddy, wheat and sugarcane. The study area
1980, Shih et al, 1983). is shown in Figure 1.

Many algorithms have been developed and tested to classify

satellite images. There are two approaches namely supervised
and unsupervised classification, known as hard classifiers. The

* Corresponding author

This contribution has been peer-reviewed.

doi:10.5194/isprsarchives-XLI-B8-1061-2016 1061
 The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLI-B8, 2016
XXIII ISPRS Congress, 12–19 July 2016, Prague, Czech Republic

4. METHODOLOGY

4.1 Data Pre-processing

Satellite based multispectral imagery contains various

quantitative information related to surface and atmosphere. The
procedure of retrieving surface reflectance or removing
atmospheric contamination from satellite measured radiance is
called atmospheric correction. To extract the accurate
information about surface we need to correct atmospheric
influence. In this study atmospheric correction is done by
SACRS2 (Scheme for Atmospheric Correction of Resoursat-2
Sensor), which is developed by Space Application Centre (SAC),
Ahmedabad, India. SACRS2 is based on the parameterization of
the equations describing radiative transfer model in the
atmosphere. It is a GUI based atmospheric correction model
developed using signal simulations by the radiative transfer
model 6SV-code (Vermote et al., 2006). The GUI of SACRS2
and flow diagram of atmospheric correction is shown in Figure 3
Figure 1. The study area and Figure 4 respectively.

3. SATELLITE DATA USED

Crop classification during Kharif season using satellite data is

generally hampered due to the cloud cover problem. Because of
the cloudy days, completely cloud free data set of Resourcesat- 2
(IRS-P6) was not available from July to October. Indian satellite
IRS-P6, LISS-IV sensor data (Path 96, Row 50) of September 15,
2013 imagery has been taken for this work. The LISS-IV sensor
has three band in different region of EMR (b2: Green, b3: Red,
b4: NIR). The satellite image of LISS IV sensor is shown in
Figure 2 and the details of LISS-IV data is shown in Table 1.

Figure 3. Atmospheric correction SACRS2 model

LISS IV Sensor
DN Raw Data

Input Stack
(Green, Red and NIR)

Lmin, Lmax , Julian Date,

Figure 2. False colour composite image
SACRS2 MODEL Viewing Geometry,
For Atmospheric Correction Atmospheric Inputs
(AOT, WV, O3 etc)
Spectral Swath Spatial Radiometric
Bands (nm) Width Resolution Resolution

b2 : 520-590
TOA Reflectance Surface Reflectance
b3 : 620-680 70/23 Km 5.8 meter 10 bit
b4: 770-860
Figure 4. Flow diagram of atmospheric correction

Table 1. LISS-IV sensor data specification

This contribution has been peer-reviewed.

doi:10.5194/isprsarchives-XLI-B8-1061-2016 1062
 The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLI-B8, 2016
XXIII ISPRS Congress, 12–19 July 2016, Prague, Czech Republic

4.2 Generation of Various Vegetation Indices

Various vegetation indices have been developed by linear

combination of red, green and near-infrared spectral bands
(Basso et al., 2004). Vegetation indices are more sensitive than
the individual bands to vegetation parameters (Baret and Guyot,
1991). The eleven vegetation indices Normalized Green (NG),
Normalized Red (NR), Normalized Near Infrared (NNIR),
Vegetation Index Green (VI green), Difference Vegetation Index
(DVI), Normalized Difference Vegetation Index (NDVI), Green
Normalized Difference Vegetation Index (GNDVI), Normalized
Difference Water Index (NDWI) Optimized Soil Adjusted
Vegetation Index (OSAVI), Modified Soil Adjusted Vegetation
Spectral Index (MSAVI2) and Global Environmental Monitoring
Index (GEMI) are generated using reflectance of green, red and
NIR bands of LISS IV sensor. ENVI 5.1 band math function is
used for formulation of vegetation indices. The different
formulae of vegetation indices are shown by equations (1)-(11)
and generated images is shown in Figure 5.
 green
NG  (1)
 green  red  nir
red
NR  (2)
 green  red  nir
nir
NNIR  (3)
 green  red  nir
 green  red
VIgreen  (4)
 green  red Figure 5. Generated vegetation indices images
DVI   nir   red (5)
4.3 Generation of Various Texture Images
  red
NDVI  nir (6)
nir  red Image is a function f(x,y) of two space variables x and y,
x=0,1,…..,N-1 and y=0,1,……, M-1. The function f(x,y) can take
nir   green discrete values i=0,1,…..,G-1, where G is the total number of
GNDVI  (7)
nir   green intensity levels in the image. The intensity level histogram is a
function showing the number of pixels in the whole image, which
 green  nir have this intensity:
NDWI  (8)
 green  nir
N 1 M 1

OSAVI 
nir  red
nir  red  L
(1  L) (Where L= 0.16) (9) h(i)    ( f ( x, y), i)
x 0 y 0
(12)

MSAVI2 
Dividing the values h (i) by total number of pixels in the image
(2* nir  1  (2* nir  1) 2  8*( nir  red ) (10) one obtain the approximate probability density of the intensity
levels
2
   0.125  p(i )  h(i ) / NM , i  0,1,......G  1 (13)
GEMI   (1  0.25 )   red  (11)
 1  red ) 
Different useful image parameters can be worked out from the
[2*(  2nir   2red )  1.5nir  0.5red ]
Where   histogram to quantitatively describe the first-order statistical
nir  red  0.5 properties of the image. Most often the so-called central moments
(Papoulis 1965) are derived from it to characterize the texture
[Eq. 1: (Sripada et al., 2006); Eq. 2: (Sripada et al., 2006); Eq. 3: (Levine 1985, Pratt 1991), as defined by Equations (14)-(17)
below and the generated texture images is shown in Figure 6.
(Sripada et al., 2006); Eq. 4: (Gitelson et al., 2002); Eq. 5:
(Tucker, 1979); Eq. 6: (Rouse et al., 1974); Eq. 7: (Buschmann
G 1

 ip(i)
and Nagel, 1993); Eq. 8: (McFeeters, 1996); Eq. 9: (Rondeaux et
al., 1996); Eq. 10: (Qi et al., 1994); Eq. 11: (Pinty and Verstraete, Mean(  )  (14)
1992)] i 0

This contribution has been peer-reviewed.

doi:10.5194/isprsarchives-XLI-B8-1061-2016 1063
 The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLI-B8, 2016
XXIII ISPRS Congress, 12–19 July 2016, Prague, Czech Republic

G 1 4.4.1 Decision tree classification based on vegetation indices:

Variance( 2 )  
i 0
(i   )2 p(i) (15) The steps used in this classification are given below:

G 1
Node 1: if VIgreen < 0.0971 then node 2 else if
Skewness( 3 )   3
 (i  )
i 0
3
p(i) (16)
VIgreen >= 0.0971 then node 3 else class fallow
Node 2: if DVI < 0.0303 then node 4 else if
DVI >=0.0303 then node 5 else class fallow
G 1 Node 3: if NNIR < 0.642 then node 6 else if
Kurtosis( 4 )   4  (i )
i 0
4
p(i)  3 (17) NNIR >= 0.642 then node 7 else class orchard
Node 4: class = settlement
Node 5: if NR < 0.297 then node 8 else if
NR >= 0.297 then node 9 else fallow
Node 6: class = water
Node 7: if GEMI < 0.302 then node 10 else if
GEMI >= 0.302 then node 11 else orchard
Node 8: if MSAVI2 < 0.0597 then node 12 else if
MSAVI2 >= 0.0597 then node 13 else sorghum
Node 9: class = fallow
Node 10: class = poplar tree
Node 11: if NDVI < 0.586 then node 14 else if
NDVI >= 0.586 then node 15 else orchard
Node 12: if DVI < 0.0386 then node 16 else if
DVI >= 0.0386 then node 17 else sugarcane
Node 13: class = sorghum
Node 14: if OSAVI< 0.198 then node 18 else if
OSAVI >= 0.198 then node 19 else sugarcane
Node 15: if GEMI<0.3042 then node 20 else if
GEMI >= 0.3042 then node 21 else orchard
Node 16: class = sorghum
Node 17: class = fallow
Node 18: class = orchard
Node 19: if DVI < 0.0439 then node 22 else if
Figure 6. Generated texture images
DVI >= 0.0439 then node 23 else sugarcane
Node 20: if NG < 0.252 then node 24 else if
NG >= 0.252 then node 25 else orchard
4.4 Decision Tree Classification Node 21: if NDWI < -0.416 then node 26 else if
NDWI >= - 0.416 then node 27 else other crops
The decision tree is an approach where pixels are classified based Node 22: if DVI < 0.0422 then node 28 else if
on a sequence of binary decisions (Safavian and Landgrebe, DVI >= 0.0422 then node 29 else sugarcane
1991). According to decision tree, the first conditional statement Node 23: if GNDVI < 0.377 then node 30 else if
leads to the second, the second to the third and so on. Decision GNDVI >= 0.377 then node 31 else sugarcane
tree is an inductive learning algorithms which generates Node 24: class = poplar tree
classification tree using the training samples. MATLAB 15a was Node 25: class = orchard
used to build decision trees. In this study training samples are Node 26: class = orchard
selected based on Google Earth and GPS field observations. The Node 27: class = paddy
characteristics of training sample ROIs is summarize in Table 2. Node 28: class = paddy
Node 29: class = sugarcane
Node 30: class = orchard
Number of ROIs Number of pixels
Node 31: if NDWI < -0.383 then node 32 else if
Water 11 203 NDWI >= - 0.383 then node 33 else sugarcane
Node 32: class = sugarcane
Fallow 16 320 Node 33: if DVI < 0.044 then node 34 else if
DVI >= 0.044 then node 35 else sugarcane
Settlement 9 272 Node 34: class = sugarcane
Node 35: class = paddy
Poplar Tree 12 237
4.4.2 Decision tree classification based on vegetation indices
Orchard 9 210 and textural features: The steps used in this classification are
given below:
Sugarcane 21 421

Paddy 16 252 Node 1: if GEMI < 0.274 then node 2 else if

GEMI >= 0.274 then node 3 else 1
Sorghum 7 203 Node 2: if NR < 0.272 then node 4 else if
NR >= 0.272 then node 5 else 5
Table 2. Number of ROIs and Pixels in each class Node 3: if VIgreen < 0.0841 then node 6 else if
VIgreen >=0.0841 then node 7 else 1
Node 4: class = water

This contribution has been peer-reviewed.

doi:10.5194/isprsarchives-XLI-B8-1061-2016 1064
 The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLI-B8, 2016
XXIII ISPRS Congress, 12–19 July 2016, Prague, Czech Republic

Node 5: class = settlement Orchard 83.54 86.71 89.54 88.91

Node 6: if NDVI < 0.264 then node 8 else if
NDVI >= 0.264 then node 9 else 1 Sugarcane 82.64 78.18 85.17 88.46
Node 7: if GEMI < 0.294 then node 10 else if
GEMI >= 0.294 then node 11 else 7 Paddy 71.25 81.71 81.44 79.83
Node 8: class = fallow
Node 9: if kurtosis < 1.595 then node 12 else if Sorghum 77.94 75.17 79.18 82.31
Kurtosis >=1.595 then node 13 else 6
Node 10: if mean < 0.0022 then node 14 else if OA= 81.09 OA= 89.42
Mean >= 0.0022 then node 15 else 4 K=0.79 K= 0.87
Node 11: if NDVI < 0.490 then node 16 else if
NDVI >= 0.490 then node 17 else 3 Table 3. Classification accuracy
Node 12: class = sorghum
Node 13: class = fallow
Node 14: if NDVI < 0.461 then node 18 else if 5. RESULTS AND CONCLUSIONS
NDVI >= 0.461 then node 19 else 4
Node 15: class = orchard The final classified images are shown in Figure 7 and Figure 8.
Node 16: if skewness < 0.0011080 then node 20 else if
Skewness >= 0.0011080 then node 21 else 7
Node 17: if DVI < 0.046901 then node 22 else if
DVI >= 0.046901 then node 23 else 2
Node 18: class = orchard
Node 19: class = poplar tree
Node 20: if skewness < -0.0001811 then node 24 else if
Skewness = -0.0001811 then node 25 else 7
Node 21: class = paddy
Node 22: if GNDVI < 0.401 then node 26 else if
GNDVI >= 0.401 then node 27 else 2
Node 23: if GDVI < 0.0420 then node 28 else if
GDVI >= 0.0420 then node 29 else 3
Node 24: class = paddy
Node 25: if variance < 5.59059e-06 then node 30 else if
Variance >= 5.59059e-06 then node 31 else 7
Node 26: class = orchard
Node 27: class = poplar tree
Node 28: class = paddy
Node 29: class = orchard
Node 30: class = sugarcane
Node 31: class = orchard

4.5 Accuracy Assessment Figure 7. Classified image using decision tree (VI)

Accuracy assessment is used to compare the classification results

with reference data, which is assumed to be true for determining
the classification results. Many methods are used to analyse the
accuracy of remotely sensed data (Congalton and Green, 1999,
Koukoulas and Blackburn, 2001). In this work, confusion matrix
or error matrix method is used (Foody, 2002). Reference data has
been taken during the field visit on September 18-21, 2013. Total
650 pixels have been selected for various classes to determine the
accuracy. The accuracy assessment has been done using ERDAS
IMAGINE software. The producer’s accuracy (PA), user’s
accuracy (UA), overall accuracy (OA) and kappa coefficient (K)
values are given in Table 3.

DT
Class Name DT ( VI)
(VI + Texture)

PA UA PA UA

Water 91.31 84.54 96.21 92.91

Fallow 87.64 81.28 95.82 92.48

Settlement 92.18 89.71 97.71 96.23

Figure 8. Classified image using decision tree (VI & Texture)
Poplar Tree 82.17 84.54 91.64 89.32

This contribution has been peer-reviewed.

doi:10.5194/isprsarchives-XLI-B8-1061-2016 1065
 The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLI-B8, 2016
XXIII ISPRS Congress, 12–19 July 2016, Prague, Czech Republic

Indian satellite IRS-P6 LISS IV sensor imagery has been Levine, M., 1985. Vision in Man and Machine, McGraw-Hill.
classified using decision tree method. The first decision tree was
constructed based on only vegetation indices and the second one McFeeters, S. K., 1996. The use of normalized difference water
was constructed using vegetation indices with textural features. index (NDWI) in the delineation of open water features.
The final image was classified into eight major classes (water, International Journal of Remote Sensing (17), pp. 1425–1432.
fallow, settlement, poplar tree, orchard, sugarcane, paddy and
sorghum). The overall accuracy and kappa coefficient is found to Papoulis, A., 1965. Probability, Random Variables and
be 81.08 % and 0.79 for decision tree using vegetation indices Stochastic Processes, McGraw-Hill.
method. Inclusion of textural feature with vegetation indices
decision tree, overall accuracy and kappa coefficient is 89.42 % Pinty, B. and Verstraete, M. M., 1992. GEMI: a non-linear index
and 0.87 respectively. The results indicates that LISS IV imagery to monitor global vegetation from satellites. Vegetatio (101), pp.
can be effectively implemented to produce classified maps with 15–20.
higher accuracy.
Pratt, W., 1991. Digital Image Processing, Wiley.

Rondeaux, G., Steven, M. and Baret, F., 1996. Optimization of

ACKNOWLEDGEMENTS soil-adjusted vegetation indices. Remote Sensing of Environment
(55), pp. 95–107.
The authors are thankful to Indian Institute of Technology
Roorkee for providing software’s for the study (ENVI 5.1, Rouse, J. W., Haas, R. H., Schell, J. A. and Deering, D. W., 1974.
ARCGIS 10.2.1, MATLAB 2015a and JUNO GPS) and Monitoring vegetation systems in the Great Plains with ERTS
the farmers for providing necessary information and NASA Goddard space Flight Centre 3d ERTS-1 Symposium, pp.
support during the field visit. 309-317.

Safavian, R. S. and Landgrebe, D., 1991. A survey of decision

tree classifier methodology. IEEE Transactions on Systems Man
REFERENCES and Cybernetics (21), pp. 660–674.
Baret F. and Guyot G., 1991. Potentials and limits of vegetation Shih, E. H. H., and Schowengerdt, R. A., 1983. Classification of
indices for LAI and APAR assessment. Remote Sensing of Arid Geomorphic Surfaces using Landsat Spectral and Textural
Environment (35), pp.161–173. Feature, Photogrammetric Engineering and Remote Sensing
49(3), pp. 337-347.
Basso, B., Cammarano D and De Vita, P., 2004. Remotely
sensed vegetation indices: Theory and applications for crop Sripada, R. P., Heiniger, R. W., White, J. G., Meijer, A. D., 2006.
management. Italian Journal of Agrometeorology (53), pp. 36– Aerial colour infrared photography for determining early in-
53. season nitrogen requirements in corn. Agronomy Journal (98),
pp. 968–977.
Buschmann, C., Nagel E., 1993. In vivo spectroscopy and
internal optics of leaves as basis for remote sensing of vegetation. Tucker, C. J., 1979. Red and photographic infrared linear
International Journal of Remote Sensing (14), pp. 711–722. combinations for monitoring vegetation. Remote Sensing of
Environment (8), pp. 127–150.
Congalton, R. G., and Green, K., 1999. Assessing the accuracy
of remotely sensed data: principles and practices, Lewis Vermote, E., Tanré, D., Deuzé, J. L., Herman M., Morcrette J. J.,
Publishers, Boca Raton. and Kotchenova, S. Y., 2006. 6S User Guide on Second
Simulation of a Satellite Signal in the Solar Spectrum -Vector
Foody, G. M., 2002. Status of land cover classification accuracy (6SV), Version 3.
assessment, Remote Sensing of Environment (80), pp. 185-
201.

Gitelson, A., Kaufman Y. J., Stark, R., and Rundquist, D., 2002.
Novel algorithms for remote estimation of vegetation fraction.
Remote Sensing of Environment (80), pp. 76–87.

Haralick, R.M., K.Shanmugam and I.Dinstein., 1973. Textural

Features for Image Classification, IEEE Transactions on Systems
Man and Cybernetics, Vol. SMC-3: pp. 610- 621.

Harris, R., 1980. Spectral and Spatial Image Processing for

Remote Sensing, International Journal of Remote Sensing 1(4),
pp.361-375.

Jackson, R. D., and Huete, A. R., 1991. Interpreting vegetation

indices, Preventive Veterinary Medicine 11(3–4), pp. 185-200.

Koukoulas, S., and Blackburn, G. A., 2001. Introducing new

indices for accuracy evaluation of classified images representing
semi-natural woodland environments, Photogrammetric
Engineering and Remote Sensing, (67), pp. 499-510.

This contribution has been peer-reviewed.

doi:10.5194/isprsarchives-XLI-B8-1061-2016 1066