Harken, James

MAPPING WETLANDS AND POTENTIAL WETLAND RESTORATION AREAS
IN BLACK HAWK COUNTY, IOWA USING
OBJECT-ORIENTED CLASSIFICATION
AND A GIS-BASED MODEL
An Abstract of a Thesis
Submitted
In Partial Fulfillment
of the Requirements for the Degree
Master of Arts
James Harken
University of Northern Iowa
July 2004
ABSTRACT
Wetlands are transitional lands between terrestrial and aquatic systems that
provide many benefits, including: floodwater retention, non-point pollution treatment,
wildlife habitat, and soil-erosion control. Wetlands in Iowa have decreased over 95% in
the last 200 years. Therefore, there is a need to map and monitor these resources, as well
as to determine potential sites for wetland restoration. In Black Hawk County, wetland
maps are outdated, and ground surveys have proved to be too time-consuming and
expensive. Traditional pixel-based automated classifiers of remotely-sensed imagery
have also proven to be inaccurate in classifying wetlands because of spectral confusion.
This study tests multispectral data, hybrid data, hyperspectral data, a seasonal matrix, and
a new object-oriented classifier. These are tested against traditional multispectral, pixel-
based (ISODATA and Maximum-Likelihood) classifiers both to see if wetland
classification accuracies from remotely-sensed imagery can be increased and to produce
an updated wetlands map for Black Hawk County. A hyperspectral image of Eddyville,
Iowa is tested to evaluate how well wetlands are classified when a hyperspectral image is
used with an object-oriented classifier and a hyperspectral pixel-based (Spectral Angle
Mapper or SAM) classifier. A GIS-based wetland restoration model is developed to
identify potential wetland restoration sites in Black Hawk County.
This study shows that the object-oriented classifier is more accurate in identifying
wetlands and overall land-cover than pixel-based ones (ISO DATA, Maximum-
Likelihood, SAM) in both multispectral, hybrid-multispectral, and hyperspectral imagery.
The summer/fall seasonal matrix produced unacceptable accuracies. Wetlands in Black

Hawk County decreased by 1500 acres (plus or minus an error margin of 375 acres) from
1983 to 2003. The restoration model identified 2,971 acres in Black Hawk County as
being highly suitable, 34,307 acres as being moderately suitable, and 121,271 acres as
having low suitability for wetland restoration. The results are available at http://gisrl-
9 .geog. uni. edu/wetland.
Limitations of the study include file size when using the object-oriented classifier,
image availability for the seasonal matrix, and the number of variables employed in the
GIS-based restoration model. The future direction of the study lies in obtaining
hyperspectral data for Black Hawk County, more current Landsat multispectral imagery
for the seasonal matrix, and testing of more non-parametric classifiers, such as the CART
algorithm.
MAPPING WETLANDS AND POTENTIAL WETLAND RESTORATION AREAS
IN BLACK HA WK COUNTY, IOWA USING
OBJECT-ORIENTED CLASSIFICATION
AND A GIS-BASED MODEL
A Thesis
Submitted
In Partial Fulfillment
of the Requirements for the Degree
Master of Arts
James Harken
University of Northern Iowa
July 2004
11
This study by: James Harken
Entitled: Mapping Wetlands and Potential Wetland Restoration Areas in Black

Hawk County, Iowa using Object-Oriented Classification and a GIS-
Based Model
has been approved as meeting the thesis requirement for the
Degree of Master of Arts
('
.-,of-----4
Dr. Ramanathan Sugumaran, Chair, Thesis Committee
'//z-1/JJ.J> c.,
Date Dr. Tim Strauss, Thesis Committee Member
'-//iJ /(J '(

Date
ll1
ACKNOWLEDGEMENTS
I would like to express my sincere appreciation to the Iowa Space Grant
Consortium for providing funding for the initial part of my research, all of the Geography
faculty at the University of Northern Iowa, and especially my committee members of Dr.
Sugumaran, Dr. May, and Dr. Strauss. I would also like to thank Steve Finegan of the
Black Hawk County Conservation Board, Barb Berquam and Kim Veeder from the Black
Hawk County GIS office, and Pete Kollasch from the Iowa Department of Natural
Resources, for their feedback and advice, especially during the beginning phases of the
research.
iv
TABLE OF CONTENTS
PAGE
LIST OF TABLES ...... ......... ...... .. ........ ...... .... .............. .............. ................................ vi
LIST OF FIGURES .......................... ......... ....................... ............ .............................. Vll
CHAPTER I. INTRODUCTION................................................................................ 1
The Importance and Status of Wetlands.......................................................... 3
Goal and Objectives......................................................................................... 7
Research Questions.......................................................................................... 7
CHAPTER IL LITERATURE REVIEW ..... .... .. .. .. .. ... .. .. ... .. .. ...... .. .... .. .... .. .... .. ... ..... .. 9
Traditional M ultispectral Classification of Wetlands.................................... 10
Hyperspectral Classification of Wetlands........................................................ 13
Object-Oriented Classification of Wetlands.................................................... 16
Wetland Restoration Models Using G IS.......................................................... 18
CHAPTER III. METHODS........................................................................................ 23
Study Area 23
Data Used 23
Black Hawk County Multispectral Images .......................................... . 23
Eddyville Hyperspectral Image ........................................................... .. 25
Image Analysis and Classification ................................................................. .. 27
Object-Oriented Classification ............................................................ . 28
Multispectral CIR Image Classification .............................................. . 30
Multispectral Landsat ETM Image Classification .............................. . 31

V
PAGE
Hyperspectral CASI Image Classification........................................... 33
GIS-Based Restoration Model......................................................................... 35
CHAPTER IV. RESULTS AND DISCUSSION......................................................... 39
Multispectral CIR Classification...................................................................... 39
Multi spectral Landsat ETM Classification .. ..... .. .. ..... .. .. .. .. .. .. ... .... .. ..... ...... .. ... . 44
Hyperspectral CASI Classification.................................................................. 48
Overall Classifier Comparison......................................................................... 53
GIS-Based Restoration Model Results............................................................ 55
Wetlands Change in Black Hawk County....................................................... 59
Web-Based Data Dissemination...................................................................... 60
CHAPTER V. CONCLUSION................................................................................... 63
Recommendations for Future Research........................................................... 66
REFERENCES ............................................................................................................ 67
APPENDIX: MAPS..................................................................................................... 75
Vl
LIST OF TABLES
TABLE PAGE
1 Error Matrix CIR Supervised Classification..................................................... 42
2 Accuracy Percentages CIR Supervised Classification...................................... 42
3 Error Matrix CIR Unsupervised Classification................................................. 42
4 Accuracy Percentages CIR Unsupervised Classification ................................. 43
5 Error Matrix CIR Object-Oriented Classification (GPS ground truth points).. 43
6 Accuracy Percentages CIR Object-Oriented Classification .................. ........... 43
7 Error Matrix Landsat ETM 15 m Supervised Classification............................ 45
8 Accuracy Percentages Landsat ETM 15 m Supervised Classification .. .. .... .... . 46
9 Error Matrix Landsat ETM 15 m Unsupervised Classification........................ 46
10 Accuracy Percentages Landsat ETM 15 m Unsupervised Classification......... 46
11 Error Matrix Landsat ETM 15 m Object-Oriented Classification.................... 47
12 Accuracy Percentages Landsat ETM 15 m Object-Oriented Classification..... 47
13 Error Matrix CASI SAM Classification (Ground Validation Pixels)............... 52
14 Accuracy Percentages CASI SAM Classification ...... .. .................. .................. 52
15 Error Matrix CASI Object-Oriented Classification (Ground Validation

Objects)............................................................................................................. 53
16 Accuracy Percentages CASI Object-Oriented Classification........................... 53
17 Overall Classification Accuracy for Different Classifiers................................ 55
18 Results from GIS-Bascd Restoration Model..................................................... 59

vii
LIST OF FIGURES
FIGURE PAGE
1 Ecology of Wetlands Systems.......................................................................... 2
2 Iowa Surface Area Covered by Wetlands......................................................... 5
3 Black Hawk County, Iowa................................................................................ 24
4 Hyperspectral Image 50 Acre Study Area........................................................ 26
5 Multispectral Imagery Processing Flowchart .... ................... ............................ 27
6 Hyperspectral Image Processing Flowchart...................................................... 28
7 Example of a Fuzzy Function........................................................................... 30
8 GIS-Based Wetland Restoration Model for Black Hawk County.................... 37
9 Research Mapping Results................................................................................ 39
10 Results from Multispectral One Meter CIR Image Classification.................... 40
11 Results from Multispectral 15 Meter Landsat ETM Image Classification....... 44
12 Results from CASI Hyperspectral Image Classification .......... .. ... .. ... ...... ..... ... 51
13 Wetlands Change in Black Hawk County .. ... ..... .. ... .. ...... .. .. .... ... ... .. ...... ........ .. . 59
14 Screen Shot of Homepage................................................................................. 61
15 Screen Shot of Web-Based Data Interface........................................................ 62

CHAPTER 1
INTRODUCTION
Wetlands are an important ecosystem. Cowardin, Carter, Golet, and LaRoe
(I 979) provide the official federal definition of wetlands: "Wetlands are lands transitional
between terrestrial and aquatic systems where the water table is usually at or near the
surface or the land is covered by shallow water" (p.1 ). Other definitions include
"Wetlands are a mix of characteristics from terrestrial or upland areas and the
characteristics of aquatic or water environments" (Lyon, 1993, p. 7), " ... places where
plants and animals live amid standing water or saturated soils, also called swamps,
sloughs, marshes, bogs, fens, seeps, oxbows, shallow ponds, or wet meadows" (Cohen,
2001, p. 1), and the US Army Corp of Engineers Wetlands Delineation manual: "Those
areas that are inundated or saturated by surface or ground water at a frequency and
duration sufficient to support and that under normal circumstances do support, a
prevalence of vegetation typically adapted for life in saturated soil conditions"
(Environmental Laboratory, 1987, p. 9).
There are several wetland classifications available in the literature. One of the
most important is defined by the U.S. Fish and Wildlife Service which classifies wetlands
into five types: Palustrine (non-tidal freshwater habitats and open water less than 20
acres), Estuarine (deep water tidal habitats), Marine, Lacustrine (open water greater than
20 acres), and Riverine, defined as freshwater rivers and streams; (Dahl, 2000). All of
these wetland categories must have one or more of the following three attributes: (a) at
least periodically, the land supports predominately hydrophytes; (b) the substrate is
2
predominantly undrained and hydric (soil that has developed anaerobic conditions); and
(c) the substrate is nonsoil and is saturated with water or covered by shallow water at
some time during the growing season of each year. A pictorial representation of two
typical freshwater inland wetlands is given in Figure 1.
Terrestrial Wetland Deepwater

system aquatrc system
Hydrology
• Dry----.i<lf- Intermittently to ~14---Permanently flooded •
permanently flooded
Terrestrial Wetland
system
I-- l __ l -
Hydrology
- -- Ground water -- -
• Dry ---t->1<1-- Intermittently to ~
permanently flooded
Figure 1. Ecology of Wetland Systems.

(Source: Mitsch & Gosselink, 2000)
3
The Importance and Status of Wetlands
Wetlands compromise only 3 to 6% of the earth's land surface area, but they
provide human populations with a host of goods and services, including food storage,
water quality maintenance, agricultural production, fisheries, and recreation (Acreman &
Hollis, 1996). They are critical to flood protection, and act like sponges to soak up water
and release it slowly. Although most wetlands store an average of approximately three
feet of water, a single acre of wetland can store up to five feet, or 1.66 million gallons of
flood water (Sierra Club, 2000). Wetlands are also believed to play a significant role in
global climate change by acting as a source of atmospheric greenhouse gases such as
methane and a sink for both carbon (wetlands contain 15-22% of the world's soil carbon
pool), nitrogen, and methane (Takeuchi, Tamura, & Yasuoka, 2003; Trettin, Song,
Jurgensen, & Li, 2001). Global biodiversity is also enhanced by wetlands because they
are vital for the survival of a disproportionately large number of threatened and
endangered species (Mitch & Gosselink, 2000). Wetlands have become a popular way
for treating contaminated surface and wastewaters, and are particularly suited for treating
non-point pollution, such as agricultural and urban runoff (Dierberg, DeBusk, Jackson,
Chimeny, & Pietro, 2002). They can also lessen soil erosion, and moderate stream
temperature (critical for certain species survival like trout, Budlong, 2002). Lastly,
wetlands have been found to preserve archeological remains (Chapman & Cheetham,
2002).
Despite these proven advantages, wetland conversion to other land uses has been
a problem historically and continues to the present day. However, the last few decades
4
have witnessed an enormous rise in awareness of the importance of wetlands. Nationally,
at the time of European settlement, the continental United States contained an estimated
221 million acres (89 .5 million hectares) of wetlands, or 9% of the total surface area.
Over time, wetlands have been drained, dredged, filled, leveled, and flooded to the extent
that less than half of the original acreage remains (Dahl, 1990; Whittecar & Daniels,
1999).
Within the state of Iowa, wetlands were viewed as a hindrance to land
development and agriculture. In less than 150 years, these rich resources were drained,
filled, or otherwise altered, drastically changing the face of Iowa's land. Similar
percentages are given concerning the amount of wetland losses in Iowa. One study
places the loss at 95% (Arbuckle & Pease, 1999) and another 90%-95% (Cohen, 2001).
In a mandated report to Congress by the U.S. Fish and Wildlife Service, only two other
states showed higher wetlands losses than Iowa: California and Ohio (Dahl, 1990).
According to the Iowa Department of Land Stewardship [IDALS] ( 1998), the amount of
wetlands six years ago covered only 1.2% of Iowa's surface area, compared to 11% two-
hundred years ago (Figure 2).

5
1.20%
01803
@J 1997
11%
Figure 2. Iowa Surface Area Covered By Wetlands.
This is a loss of approximately 3.5 million acres, or an area approximately the size
of the state of Connecticut. The reduction of wetlands in Iowa has also contributed to the
fact that Black Hawk, Hamilton, Johnson, Linn, Story, and Tama counties were
designated federal (flood) disaster areas five times from 1989-1998 (Sierra Club, 2000)
and all of Iowa's 99 counties were designated federal (flood) disaster areas at least once
during that time.
The loss of these critical resources (wetlands) in Iowa with some 92% of the land
being used for agriculture (Dung, 2003), and their documented value, shows an urgent
need to monitor these resources, measure their changes, and provide a method for
identifying potential wetland restoration areas. Traditionally, wetlands are delineated
using ground surveys. However, these surveys are difficult and time-consuming (Lyon,
1993; Yasuoka et al., 1995). Geospatial technologies, such as remote sensing,
Geographical Information Systems (GIS), and Global Positioning Systems (GPS) can
provide an alternative and possibly better solution to mitigate the before-mentioned
problems (Goldberg, 1998). In addition, remote sensing data can be used for the
6
following: (a) to detennine the extent of wetlands, (b) to identify the type of wetland
resource, (c) to characterize the general wetland land cover type, ( d) to identify
submergent and emergent wetlands, and (e) to supply details about the resource (Lyon &
McCarthy, 1995). Geographical Infonnation Systems and GPS can be used effectively
for natural resource management, conservation, and restoration (Konecny, 2003). This
includes inventorying and updating wetlands (Houhoulis & Michener, 2000). The need
to update the last wetlands survey undertaken for Black Hawk County (completed by the
National Wetland Inventory and Iowa Department of Natural Resources and based upon
aerial photos taken in 1983 and 1984), is the justification for this research.
According to the Iowa Department of Natural Resources (2004), "Wetlands are
one of the easiest and most quickly restored elements of natural landscape, and they can
provide nearly instantaneous wildlife habitat. The Wildlife Bureau offers technical
assistance to landowners interested in restoring wetlands on their properties." The Iowa
Natural Resource Conservation Service (NRCS) administers the Wetland Reserve
Program aimed at returning fonner wetland areas that have been cropped. The
Emergency Wetland Reserve Program also works to place pennanent easements on land
that has a flood history, returning it to wetland conditions. Wetland detennination and
mitigation assistance is provided for United States Department of Agriculture wetland
compliance programs.
7
Goal and Objectives
The main research goal is to map and identify potential wetland restoration areas
in Black Hawk County, Iowa using remote sensing and GIS technologies. To achieve
this goal, the following four objectives are presented:
1. Extract up-to-date and accurate wetland areas from multispectral and
hyperspectral images;
2. Evaluate different image classifiers, specifically object-oriented, Maximum-
Likelihood, ISODATA, and Spectral Angle Mapper (SAM);
3. Analyze different GIS wetland restoration models from the literature and
create such a model for use in Black Hawk County;
4. Disseminate the final results through the Internet via Arc Internet Map Server
(ArcIMS).
Research Questions
On the basis of the goal and objectives of the study, the research questions are as
follows:
1. How well does the object-oriented classifier perform in comparison to
traditional ones, such as Maximum-Likelihood and ISODATA, for the delineation of
wetlands using multispectral imagery in Black Hawk County?;
2. Can data fusion, specifically between Landsat Enhanced Thematic Mapper
(ETM) multispectral and ETM panchromatic images, improve wetland classification?
3. Is the object-oriented classifier more accurate than SAM in high resolution
hyperspectral image classification of wetlands?

8
4. What role does summer and fall seasonality play in wetland classification
when using remote sensing data?
5. What are the most important factors in a GIS-based wetland restoration model
for Black Hawk County?

9
CHAPTER2
LITERATURE REVIEW
The value of wetlands and their contributions to a healthy ecosystem have been
gaining increasing recognition over the past few decades as have the spectral and spatial
resolution ofremote-sensing satellites since Landsat was first launched in 1972. Along
with the increased power of geographical information systems, mapping and monitoring
wetlands and other ecosystems with remotely-sensed imagery is proving to be an
indispensable tool for understanding these valuable resources and keeping wetland
inventories current. In Iowa 92% of the land is in agricultural use (the highest in the
nation) and Iowa ranks in the top three states in wetland losses (Dahl, 1990). Therefore,
there is a need to update decades-old inventories and research potential wetland
restoration areas.
The literature compiled for this study is broken down into classification
techniques of wetlands from remotely-sensed imagery and GIS-based wetland restoration
models. The review is based on the most current studies published. Although the
literature gives an excellent, solid foundation in multispectral assessment of wetlands,
recent software introductions (Benz, Hofmann, Willhauck, Lingenfelder, & Heynen,
2004) and availability of hyperspectral data give the chance to further research methods
of wetland classification. Lastly, wetland restoration models are reviewed to develop a
model that can be applied to the landforms of the study area.

10
Traditional Multispectral Classification of Wetlands
Traditionally, Landsat Multispectral Scanner (MSS), Landsat Thematic Mapper
(TM), and the French Systeme Pour !'Observation de la Terre (SPOT) satellite systems
have been used to study wetlands (Lunetta & Balogh, 1999; Shaikh, Green, & Cross,
2001; Shepard, Wilkinson, & Thompson, 2000; Toyra, Pietroniro, & Martz, 2000). Other
studies have included the moderate-resolution remote-sensing platforms of the National
Oceanic and Atmospheric Administration's Advanced Very High Resolution Radiometer
(AVHRR), the Indian Remote Sensing Program (IRS), the Japanese Earth Resources
Satellite (JERS-1), the European Remote-Sensing Satellite (ERS-1), the Shuttle Imaging
Radar (SIR-C), and lastly, the Canadian Radio Detection and Ranging Satellite,
RADARSAT (Alsdorf, Smith, & Melack, 2001; Bourgeau-Chavez et al., 2001; Chopra,
Verma, & Sharma, 2001). Some of the earliest work included visual interpretation of
aerial photographs. Unsupervised classification or clustering is the most commonly used
classification to map wetlands and maximum-likelihood is the most common supervised
method (Ozemi, 2000). To aid in the low wetland accuracy percentages that usually
accompany these classification methods (30 - 60% accuracies), multi-temporal and
ancillary data are often used along with various models to improve classification
accuracies. Ancillary data provide a practical solution to help solve the problem of
distinguishing among spectral similarities in wetlands, agricultural fields, and forests
(Houhoulis & Michener, 2000).
However, there are limitations in delineating wetlands using traditional, optical,
multispectral techniques. One limitation on the use of optical data for wetland mapping
11
is their inability to penetrate vegetation canopies, and thus their inability to remotely
sense flooding beneath a closed canopy (Bourgeau-Chavez et al., 2001). There has been
some research done on wetlands using radar data (Alsdorf et al., 2001; Bourgeau-Chavez
et al., 2001; Rio & Lozano-Garcia, 2000) as well as LIDAR (MacKinnon, 2001), but the
majority has been concentrated on Landsat TM, MSS, SPOT, and airborne Color Infrared
(CIR) photos.
Previous studies for classification of wetlands using rule-based classifiers have
shown mixed results. Jenssen and Middelkoop (1992) showed improvements of six to
twenty percent accuracies for crop cover classification of Landsat TM images over the
maximum-likelihood classifier. Halid (1997) had a decrease in accuracy of land cover
changes using a knowledge-based classifier compared to a maximum-likelihood one
(78% overall accuracy decreasing to 44%). However, he noted that rule-based
classification had the advantages of being quicker and requiring less field work. Sader,
Ahl, and Liou (1995) reported overall accuracies of 80% and 82% for wetlands in Maine
(in Orono and Acadia National Parks, respectively). Wetland producers accuracy in
Acadia National Park was determined to be 77% and wetlands users accuracy 62%. In
Orono National Park, wetlands producers accuracy was determined to be 66% and
wetland users accuracy 82%. Forested wetlands were emphasized in the study. Ozemi
(2000) noted that rule-based classifiers generally provide more accurate classification
results than the traditional maximum-likelihood method, but not always. In addition, she
noted that classification accuracies were much greater using two dates of imagery for
Landsat TM (leaf-on and leaf-off). This indicates that seasonal comparison of images for
12
wetland classification is probably needed. Hodgson, Jensen, Mackey, and Coulter (1987)
also indicated wetlands could be better defined on imagery acquired in spring when the
water table was high.
Houhoulis and Michener (2000) created a rule-based method of wetland change
detection using National Wetlands Inventory and SPOT data for a study area in the Flint
River Basin in south-west Georgia. Their utilization of the modulus to reduce data
volume and provide spectral variability was added to the attribute table of the wetland
polygons along with majority land-cover attributes to determine the change criteria
(within one standard deviation). They also used a custom Arc Macro Language (AML)
script to determine thresholds and provided an accuracy assessment of over 10% of the
12,000 wetland polygons used in the study. The overall accuracy of the study was an
impressive 96%, with 90% accuracy for changed wetlands and 8% of the wetlands
showing a conversion to other land uses. The reasoning behind the study was that the
National Wetland Inventory (NWI) coverages were two decades old and needed to be
updated. Limitations of the study include the fact that since only previously surveyed
wetlands were monitored for change, the accuracies were artificially high compared to
wetlands delineation from scratch. Also, no ground-trothing was performed, accuracy
was limited to the 20-meter resolution of the SPOT data, and no allowance was made for
created or mitigated wetlands that could have been created within the past twenty years.
Other work has been done using multi-sensor assessment (Toyra et al., 2001) and neural
networks (Han, Cheng, & Meng, 2003; Ozemi, 2000).

13
Hyperspectral Classification of Wetlands
Hyperspectral classification of wetlands is relatively new and the literature not yet
fully developed. Recently, only a few researchers have reported the use of hyperspectral
images for wetland mapping. Relevant studies include Anderson, Garono, and Robinson
(2003), Bakker and Schmidt (2002), Carter, Wells, and Lewis (2004), Juan, Jordan, and
Tan (2000), and Schmidt and Skidmore (2002). This dearth of studies exists perhaps
because hyperspectral imagery requires more complex software and more powerful
computers for processing than multispectral imagery. It is also more expensive, but
according to the following research, has yielded more accurate results than traditional
multi spectral imagery classification. The following sections provide a brief background
and summarize the available literature.
Studies using pixel classifiers, such as SAM, Minimum Noise Fraction, and
Matched Filter, in conjunction with hyperspectral imagery include Marcus, Legleiter,
Aspinall, Boardman, and Crabtree (2003), Salem and Kafatos (2001), and Underwood,
Ustin, and DiPietro (2003).
The Spectral Angle Mapper (SAM) algorithm is a physically-based spectral
classifier that uses an n-dimensional angle to match pixels to reference spectra (ENVI,
2002). The mathematical formula for SAM is as follows:
a= cos · 1 LXY (1)

✓L(X) 2 L(Y) 2
Where in Equation 1;
a = angle formed between reference spectrum and image spectrum

X = image spectrum
Y = reference spectrum
14
The advantage of the Spectral Angle Mapper technique over traditional
Maximum-Likelihood and ISODATA techniques is that the illumination differences
across landscapes (e.g., different aspects) do not create false differences between pixels
of the same composition (ERDAS, 2002). For a detailed description of the SAM
technique see Salem and Kafatos (2001) and ENVI (2002).
Carter, Wells, and Lewis (2004) evaluated the potential ofITD VNIR IOE (a type
of sensor) hyperspectral imagery to detect invasive wetland plant species in northern
Mobile Bay, Alabama, in September of 2003. Ground resolution was one meter, and the
wavelengths captured were in the 400 to 1000 nm range. They were successful in
detecting Chinese tallow tree (Tridica sebifera), and water hyacinth (Eichhornia
crassipes), as well as mapping native wetland plants. The researchers continue to
evaluate different algorithms for use in coastal wetlands.
Schmidt and Skidmore (2003) studied 27 salt marsh vegetation types in a coastal
Dutch wetland and concluded that statistical variation of wetland vegetation reflectance
spectra is possible in the visible to short-wave range. They used a three step analysis to
test difference between type classes, used continuum removal as a normalization
technique in the visible range (although it failed in the infrared range), and measured the
distance of the vegetation types in spectral space using the Bhattacharyya and Jeffries-
Matusita distance measures. S-Plus software was used to process the 579 bands between
400 and 2500 nm with a gap between 1820 to 1940 nm for atmospheric water absorption.
A GER spectrometer was used to measure the in situ reflectance on 132 vegetation plots.
The bands found to be the most useful for discriminating wetland vegetation types were
15
between 740-1820 nm in the shortwave infrared and between 400 to 700 nm in the visible
spectrum. Six wavelength bands were then selected out of the above mentioned bands
based on their higher frequency of statistically different median reflectance and their
more-or-less spacing across the whole spectrum. Those bands are: 404, 628, 771, 1398,
1803, and 2183 nm. This study provides a foundation for other researchers wishing to
test those specific bands for their own wetland study areas.
Bakker and Schmidt (2003) concentrate on edge filtering for hyperspectral images
in agriculture and salt marsh test areas. They conclude that hyperspectral edge filters can
assist in image interpretations. Lastly, Juan et al. (2000) flew a hyperspectral mission
over Fort Drum Marsh in Florida using an unspecified hyperspectral sensor that collected
64 wavebands in the 399.2 to 920.5 nm range. They were successful in delineating the
wetland species from the airborne hyperspectral imagery, but did not release what
wavebands were most sensitive for different plant species.
Anderson, Garono, and Robinson (2003) used Compact Airborne Spectrographic
Imager (CASI) Imagery along with Landsat 7 ETM+ images to map wetlands along the
Columbia River. They originally wanted to map the entire area with CASI, but ran into
time and budget issues. Their configuration for the CASI imagery was 19 bands from
459.3 nm to 819.8 nm, and 1.5 meter spatial resolution. They masked out the urban areas
and used National Wetland Inventory maps along with ground truthing to create the
classification. They were able to determine over 80 different classes with the CASI
imagery, 20 of which were purely spectrally determined. They also used ERDAS
Imagine software and the ISODATA unsupervised classification algorithm, where 6-7
16
major habitat types were identified and then continuously cut from the originally 19-band
mosaic until all spectral classes fit their criteria of narrowness. Their accuracy
assessment has still not been completed.
Lastly, Underwood et al. (2003) mapped iceplant (Carpobrotus edulis)
successfully in a southern California coastal habitat using the Minimum Noise Fraction
algorithm and band-ratio indices. Salem and Kafatos (2001) used the SAM algorithm
along with hyperspectral imagery to detect oil spills in Chesapeake Bay, and concluded
such a method minimized the limitations of conventional remote-sensing techniques (i.e.,
multispectral and aerial photographs). Marcus et al. (2003) evaluated one meter, 128
band hyperspectral imagery for mapping in-stream habitats, depths, and woody debris in
Yellowstone National Park. They concluded that clear water was necessary to measure
depth, and that tree canopy cover was also a problem. They accomplished high overall
accuracies ranging from 69 to 99%. One method (classifier) not seen in the hyperspectral
and wetlands literature is the object-oriented one, discussed in the next section.
Object-Oriented Classification of Wetlands
Object-oriented classification is relatively new to the field ofremote sensing and
most of the studies completed have taken advantage of high-resolution imagery
(IKON OS, QuickBird, etc.) for land-cover classification. Of particular interest to many
researchers is urban area classification due to the functions associated with eCognition
software. However, other research has focused on natural resource and wetland
classification, as shown by many studies (Antunes, Lingnau, & Da Silva, 2003; Civco,
17
Hurd, Wilson, Song, & Zhang, 2002; Gomes & Marcal, 2003; Ivits & Koch, 2002; Kaya,
Pultz, Mbogo, Beier, & Mushinzimana, 2002; van der Sande, de Jong, & de Roo, 2004).
van der Sande et al. (2004) divided one meter, four-band IKONOS-2 imagery into
different land cover segments with an overall accuracy of 74%, and then used that
thematic map as an input for a flood-simulation model. They were able to then
successfully estimate flood damage for local land-use planners and insurance companies.
Gomes and Marcal (2003) used 9-band 15-meter Advanced Spaceborne Thermal
Emission and Reflection Radiometer (ASTER) imagery to revise a 1995 land-cover data
set for the Vale do Sousa region in northwest Portugal. Their overall accuracy was
71.5%; forested areas, which were emphasized in the study, had an average accuracy of
46.3%.
Antunes et al. (2003) segmented a 4 band IKONOS image to identify riparian
areas that could not be spectrally differentiated in the northern part of the state of Parana
in southern Brazil. They needed to map declining wetland areas for resource
management because of increased agricultural activities. Their accuracies were 75.4%
for riparian vegetation and 78.6% for swamp vegetation. They also ran a Bayesian
Maximum Likelihood classifier for the same areas and came up with 56.0% for riparian
areas and 45.3% for swamp vegetation. Although they showed promising results, there
was a disappointing lack of detail in their exact pre-processing and methodology steps.
Civco et al. (2002) compared knowledge-based and object-oriented techniques
(among others) for land cover change detection in the Stony Brook Millstone River
watershed in New Jersey using Landsat ETM+ data. They concluded that no single
18
method was superior for their study data and area. However, they admitted "The image
segmentation and object-oriented classification holds much promise." and "The image
segmentation and object-oriented classification and change detection appeared to have
produced better overall results, especially in terms not only detecting and characterizing
the nature of change, but also in minimizing the salt-and-pepper effect caused by isolated
and non-contiguous pixels" (p. 8). Ivits and Koch (2002) used six European test sites and
IRS panchromatic and Landsat ETM imagery along with object-oriented classification to
develop a preliminary landscape habitat ecological analysis.
Kaya et al. (2002) acquired RADARSAT-1 data to map wetlands and other land
cover types in coastal Kenya to assess malarial risk. Their object-oriented approach
resulted in 85.5% overall accuracy and 65.3% accuracy for wetlands. They tested
Synthetic Aperture Radar (SAR) data because of lack of availability of multi-spectral
cloud free images for coastal, tropical regions. Some problems they encountered with the
data included backscattering returns being classified as wetlands, as well as certain forest
types (mangrove) also being wrongly classified as wetlands. Harken and Sugumaran
(2004) found that an object-oriented classifier had a high degree of accuracy in
classifying freshwater wetlands using 60 cm CASI hyperspectral imagery in a study area
in Eddyville, Iowa.
Wetland Restoration Models Using GIS
As Hey and Philippi (1999) note, wetlands can be restored to provide functions
that have been lost. They also note that wetland restorations are most effective when they
currently occupy less than 10% of the area to be restored. There are no standard models
19
for restoring wetlands as there are for determining and mapping wetlands (which in itself
is a complex, time-consuming procedure). However, there have been five studies
completed where remotely-sensed I GIS-based wetland restoration models have been
created and implemented. Berman, Rudnicky, Berquist, and Hershner (2002) worked in
Virginia and the Lower Mississippi River Conservation Committee (2001) completed a
study in the Mississippi Alluvial Valley in Missouri. Sader et al. (1995) worked in
Maine, Braster and Radish (1996) in western Iowa, and Riverlink (2000) in North
Carolina.
Other useful studies include wetland hydrological modeling (Brown, Johnston, &
Cahow, 2003; Loiselle, Bracchini, Bonechi, & Rossi, 2001; Tsihrintzis, John, &
Tremblay, 1998; Whittecar & Daniels, 1999) as well as wetland nutrient modeling (Wang
& Mitsch, 2000), wetland soil carbon modeling (Trettin, Song, Jurgensen, & Li, 2001),
wetland habitat modeling (Wakeley, 1988) and wetland buffer modeling (Budlong,
2002). The end product of the Lower Mississippi River Conservation Committee (2001)
model is a raster map, where each 30-meter cell has a arbitrary weighted value of 7 to 75,
which is to be interpreted as an indicator of relative probability of a given grid cell to
deliver water quality benefits ifrestored. The model's purpose was to prioritize areas for
forested wetland areas on private land next to the Mississippi River in south-east
Missouri. They used ARC/INFO and Arc View Grid Analyst software, as well as State
Soil Geographic Database (STATSGO) soil coverages, a Digital Elevation Model
(DEM), and geomorphology coverages. Hydrology (flooding, topography) was given

20
73.33% of the total model weight, and reforestation (soils) 26.67%. The reasoning
behind the weighting was not given.
Riverlink (2000) developed three disparate mountain wetland restoration models;
one for a general need assessment that identified 94 watersheds, a second to identify
high-probability wetland restoration areas that identified 140,000 acres of land, and a
third to identify large parcels of land, 25 acres or more, that identified 4 77 potential sites,
and 78 high potential sites. Hydrologic units were determined to have too coarse a
resolution for the study needs so small management units were created in Arc View based
on a flow accumulation of 5,000 grid cells or approximately 1,148 acres. Grid cells in
each layer of the model (wetlands, building starts, agriculture, roads, elevation, sewer,
and conservation/natural resource areas) were ranked on their presence or absence, their
linear distance from each other, and what percent of the grid cell they covered. The cells
were then scored and regrouped into three natural break categories of restoration
potential, high need, medium need, and low need. They also used another natural break
(Jenks) regrouping based on final parcel size; i.e., their need was to develop wetland
restoration areas in the largest tracts possible.
Budlong (2002) used three factors in determining potential riparian habitat buffers
in the Whitewater River Watershed in south-eastern Minnesota. They were: proximity
of row crops to streams and rivers, slope, and proximity of feedlots to rivers and streams
of the watershed. It should be noted that in most of the restoration models reviewed,
proximity to a hydrological feature (usually a river or stream) and slope were always used
as model factors. Hydric soils were also found to be important in ranking potential
21
wetland restoration areas, and these areas were always preferred to be agricultural.
Budlong's ranking system divided subwatersheds into high, moderate-high, moderate, or
low restoration potential. To achieve this goal of ranking, percentages were used for
land-cover types within the 50-meter stream buffer (>65% row crop area meant high
potential, etc.), mean slope value within 300 meters of all hydrological features, and total
areas of feedlots within the 50-meter stream buffers. The final equation was: (x = row
crop land-cover%, y = slope, and z = feedlot areas) RESTORATION POTENTIAL= (x
* 0.65) + (y * 0.25) + (z * 0.10). One of the most important conclusions from this study
was that riparian stream buffers should be adjacent to the headwater streams of a
watershed for maximum ecological effect.
Berman et al. (2002) used ARC/INFO software, a land-cover layer derived from
30-meter Landsat United States Geological Survey (USGS) imagery, a digital Soil
Survey Geographic Database (SSURGO) layer, a hydrology layer, a National Wetlands
Inventory (NWI) layer, and conservation-area layers. They based their analysis on
wetland functions. Polygons were ranked as good, high, or excellent according to water
quality, flood control, sediment control, erosion control, and wildlife habitat. Landscape
position and surrounding land-cover was also used to assign rankings. Agricultural areas
were again favored in the ranking schema.
Lastly, Braster and Radish (1996) wanted to identify current land uses and offer
alternatives to land managers of floodplain areas. They wanted to do this by using GIS,
strengthening relationships with local organizations and landowners, and providing
informed development strategies. Chi-square values were developed for the variables of
22
depth of water table, number ofNWI wetlands, presence ofNWI wetlands, position of
soil mapping units, com-suitability ratings, presence of hydric and non-hydric soils, and
proximity to levees. Lo git modeling was used and weights and ratings for the variables
were based on the chi-square for the first model (Model 1) and field experience for the
second model (Model 2). The formula for the composite score was= (W 1 * R + (W *
1
)
2
R 2) + ... + (WA7 * R"7) where W 1 is the weight for the variable mapped data layer 1 and
R I was the rating assigned to the category on data layer 1. Weightings were based on a
GIS map arithmetic approach, after Anderson (1992). After applying the models to the
312 selected study sites frequency statistics were generated. Both models showed a high
improvement over chance (83.0 and 82.1 % respectively) in predicting high-probability
locations of wetland restorations.
All of these studies in both wetland classification and restoration methods have
been important in the fields of wetland delineation and restoration research. Their
limitations include not bringing together updated wetland classification and restoration
models and unacceptable accuracies. This study will attempt to address some of those
issues through the use of a new classifier (object-oriented), up-to-date data sets, and a
unique site context (the Iowan Surface and Southern Iowa Drift Plain landforms, more
specifically Black Hawk County and Eddyville) in which to apply the methods and
potential wetland area restoration model. Also, in the literature the majority of
hyperspectral mapping of wetlands has been concentrated in coastal and estuarine areas,
and not in freshwater inland areas.

23
CHAPTER3
METHODS
Study Area
Two study areas were used in this research: Black Hawk County, located in the
northeastern part of the state and encompassing 567 square miles, and 50 acres near
Eddyville, in south-east Iowa. The imagery used included 30-m Landsat ETM+ for
Black Hawk County and 60 cm CASI for Eddyville, discussed further in the next section.
The two study areas were chosen for the following reasons: (a) Black Hawk County
because imagery was available at no cost, and local experts could critique the
methodology along the way; and (b) Eddyville because it is the only portion of the state
with a hyperspectral dataset where wetlands are present.
Data Used
Black Hawk County Multispectral Images
Black Hawk County, Iowa is the fourth most populous county in the state and is
located at 42.491N Latitude and 92.367W Longitude. The multispectral imagery used for
classification is as follows: An April 2002, I-meter resolution Color Infrared Photo
mosaic (Figure 3, left), obtained from the Iowa Geographic Image Server, a September
2000, 30-m Landsat ETM+, and a July 1999, 30-m Landsat ETM+ obtained from the
University of Northern Iowa's STORM Project (Figure 3, right). Two hybrid data sets
were created by pan-sharpening the Landsat images with their 15-m panchromatic band,
Principal Components Analysis and a Matrix of the two seasonal Landsat images. The
choice of the data sets was based on their no cost availability and their temporal
24
applicability (all three within the last four years). This is pertinent because one of the
project goals was to create an updated wetlands map for the county using the most up-to-
date imagery available. As stated previously, the current wetlands map, created by the
Iowa Department of Natural Resources (IDNR) and the NWI, is based on aerial
photograph interpretation and field surveys done almost twenty years ago. The vector
ancillary data used for cross checking the multispectral imagery, classification accuracy
assessments, and as direct inputs into the restoration model (wetland areas, hydrology,
soils, and conservation areas) were obtained from various sources, including the USGS,
NRCS, IDNR, NWI, and Iowa Geographic Map Server. Additionally, data were acquired
from the Black Hawk County GIS office. The software used with the multispectral
imagery was ERDAS Imagine 8.6, eCognition 3.0, and ArcGIS 8.2.
Figure 3. Black Hawk County, Iowa. One meter CIR image (left)
and 15 m Landsat ETM image (right).
25
Eddyville Hyperspectral Image
Eddyville, Iowa is a small town located in the south-central part of the state along
the Des Moines River at 41.160N Latitude and 92.631W Longitude. The hyperspectral
image used for classification was flown with the CASI sensor in July of 200 I for a Iowa
Department of Transportation & National Consortium on Remote Sensing in
Transportation - Environmental Group project (Iowa Department of Transportation &
National Consortium on Remote Sensing in Transportation - Environmental Group
[NCRST-E], 2002). The 2001 image is a mosaic of seven flight lines and has a spatial
resolution of 60 cm with 25 contiguous spectral bands, each of which is approximately
0.018 micrometers with a range of 350 to approximately 2500 nanometers (Figure 2). In
addition, a I-meter Color Infrared Image from the IDNR, SSURGO maps and NWI data
were used for training and accuracy assessment. The software used to process and
classify the hyperspectral image was ENVI 3.6 and eCognition 3.0.
The 60-centimeter 2001 Eddyville image encompasses approximately 969 acres
and contains unique ecological habitats. The Iowa Department of Transportation
discovered this when they planned a highway bypass northeast of the city and citizens
informed the IDOT of the protected species and habitats (NCRST-E, 2002). However,
only a SO-acre test portion of the study area was classified in this research (See Figure 4).
Wetland vascular plant species in the area include such species as: Festuca rubra L. (red
fescue), Pycnanthemum tenuifolium (Slender mountain mint), Polygonum persicaria
(Spotted ladysthumb), Conyza sp., Phalaris arundinacea (Reed canarygrass), Galium
aparine (Goose-grass), Utica dioica (nettles), and Marus alba (White mulberry). All of
26
the before mentioned species occur on the 1996 National List of Vascular Plant Species
that Occur in Wetlands, published by the U.S. Fish and Wildlife Service (United States
Fish and Wildlife Service, 1996).
Figure 4. Hyperspectral Image 50 Acre Study Area.

27
Image Analysis and Classification
The following sections provide an overview of object-oriented classification and
the processing behind the multispectral and hyperspectral images. Figures 5 and 6
demonstrate the overall flow of the multispectral and hyperspectral image analysis
process.
Multispectral Data
30 m Landsat ETM+ I mCIR
Data Fusion
Seasonal Matrix Pan-sharpened

July+ Sept.
Training Areas Identified
Object-Oriented !SODATA Unsupervised Maximum Likelihood

Classifier Classifier Classifier
Accuracy Assessment
Final Output Final Output Final Output

Seasonal Matrix Landsat CIR
Figure 5. Multispectral Imagery Processing Flowchart.

28
CASI Data
Pre-Processing (Minimum Noise

Fraction Algorithm)
Band Selection
Training Areas Identified

(RO!s)
Spectral Angle Mapper Multiresolution

(SAM) Classifier Segmentation
00 Classification
Output Rule Images

Training Areas
Retrained
Accuracy Assessment
Final Output Final Output

SAM 00
Figure 6. Hyperspectral Image Processing Flowchart.
Object-Oriented Classification
In contrast to traditional image-processing methods, the basic processing units of
object-oriented analysis are image objects or segments, and not single pixels (Baatz &
Schape, 2001 ). The reasoning behind this is the expected result of many image-analysis
tasks is the extraction of real-world objects. Representation of image information is
based on the networking of these image objects, which must be explicitly worked out in
contrast to implicit neighbor objects on the pixel scale. Scale is an important

29
consideration in object-oriented analysis because it determines the occurrence or
nonoccurrence of a certain object class, i.e., a house or a subdivision, or a field or an
ecosystem. This is achieved by a strict hierarchical structure that allows relations
between objects and their sub-objects and super-objects. Single pixel objects represent
the smallest possible processing scale. Other information used in object-oriented analysis
includes tone, shape, texture, context, and information from other object layers.
The method of segmentation of the image objects is important, as there are an
almost infinite number of solutions. They can be roughly grouped into two categories:
knowledge-driven (top down) and data-driven (bottom up). Examples of data-driven
segmentation include unsupervised spectral classification, region-growing algorithms
from seed pixels, and texture-segmentation algorithms. According to the eCognition
User Guide (Baatz & Schape, 2001), image segmentation in the eCognition software is
essentially a heuristic optimization procedure which locally minimizes the average
heterogeneity of image objects for a given resolution over the whole scene. The
parameters that must be set for image segmentation in eCognition include: (a) aliases, (b)
layer weights, (c) image-object level, (d) scale parameter, (e) segmentation mode, (f)
composition of homogeneity criterion, and (g) type of neighborhood.
Classification is based on fuzzy systems which use a degree of probability to
express an object's assignment to a class. Please refer to Figure 7 for a graphical example
of a fuzzy function. The membership value lies between 1.0 and 0.0, where 1.0 expresses
full membership/probability and 0.0 expresses absolute non-membership/improbability.

30
In eCognition, the software used for this project, supervised classification was
used to create training areas to classify the data, and ultimately, to see how well wetlands
were classified.
µ = 0.5
µ = 0.0
Figure 7. Example of a Fuzzy Function. A crisp set M (rectangle) and the fuzzy sets
A and C (triangles) over a feature range X.
Multispectral CIR Image Classification
The unsupervised classification of the CIR I-meter (4.6 GB file size) was
completed using ERDAS Imagine's ISODATA algorithm with the following parameters:
120 classes with a convergence threshold of .95 and 30 maximum iterations. To identify
separable clusters in the histogram, 120 classes were selected. Classes were then
identified by visual interpretation based on the original false-color image and recoded
(merged) into 6 general classes based on the Anderson, Hardy, Roach, and Witmer
(1976) USGS classification system: Wetland (includes Woody and non-Woody
Wetlands), Mixed Forest, Artificial Surface, Fallow/Bare Soil, Mixed Grasses (includes
Mixed and Herbaceous Grasses), and Open Water. The supervised classification of the
CIR was completed with ERDAS Imagine's Maximum-Likelihood Classifier, using a

31
created signature file of polygon AO I's by visual interpretation and NWI ancillary data
and grouped into the same six general classes used in the unsupervised classification:
Wetland, Mixed Forest, Artificial Surface, Fallow/Bare Soil, Mixed Grasses, and Open
Water. The accuracy assessment for the CIR was performed by generating 300 random
stratified points or 50 points per class. The points were then visually interpreted on an
unclassified 2002 CIR image.
For the object-oriented classification of the 1-m CIR image, a 2,881-acre subset
was classified, due to file size restraints within the software. This specific limitation is
discussed in greater detail in Chapters 4 and 5. Layer 3 of the CIR image was given a
slightly higher weighting (1.0 versus 0.8 for layers 1 and 2) based on its proven
vegetation sensitivity characteristics (Lillesand & Kiefer, 2000). For the accuracy
assessment of the subset image, GPS ground trothing was performed for 35 areas during
June of 2003.
Multispectral Landsat ETM Image Classification
The unsupervised classification of the Landsat ETM+ 15-m image of Black Hawk
County was done using the same parameters as the other unsupervised classification
(CIR) to insure statistically comparable results. The ISODATA algorithm was used to
separate the image into 120 classes with a convergence of 0.95 and 20 maximum
iterations. Classes were identified by visual interpretation, histogram separability, and
ancillary data and then grouped into six classes. A Row Crop class was substituted for
Fallow/Bare Soil for this classification because the image was captured in September and
much more planted vegetation was present than in the April CIR. The other five classes
32
remained the same. The supervised classification of the Landsat image was completed
using the ERDAS' Maximum-Likelihood Classifier, using a created signature file of
polygon AO I's based on visual interpretation and NWI ancillary data and grouped into
the same six classes as the unsupervised Landsat classifications. The accuracy assessment
was performed by generating 240-300 random stratified points or 40-50 points per class
depending on pixels-per-class availability and visually interpreted using both an aerial
photo and an unclassified TM image.
For the object-oriented classifier in eCognition, Layers 3, 4 and 5 on the Landsat
Image as well as layer 3 (SWIR) on the CIR were given slightly higher weightings (1.0
compared to 0.8) during the initial segmentation based on their proven vegetation
sensitivity characteristics. Each data set in eCognition was classified according to an
average of 185 objects per class and 50 samples per class were tested (6 classes) for a
total of 300 random sample points for the accuracy assessment (except for the Landsat
ETM 30-m where lack of objects kept the points down to 20 per class). Objects generally
ranged from 5-15 pixels in size for the CIR and 94-95 pixels for the Landsat image.
A seasonal matrix (Landsat summer/fall images) was created in ERDAS
Imagine's interpreter function under GIS analysis to address the third research question
and followed the same classification and accuracy assessment procedures as mentioned
above. Figure 5 shows the overall multispectral processing flow for both the CIR and
Landsat ETM+ data.

33
Hyperspectral Image Classification
The Spectral Angle Mapping Wizard was run in ENVI against the subset
Eddyville hyperspectral image. The wizard is composed of 10 steps, however only 4 of
the most pertinent steps are shown in the flowchart in Figure 6. The first step in the
image analyses was to select suitable bands and to reduce noise. A Minimum Noise
Fraction (MNF) algorithm was run to determine the inherent dimensionality of image
data, segregate noise, and to reduce the computational requirements for subsequent
processing (Boardman & Kruse, 1994). The MNF is a linear transformation that consists
of two separate principle component analysis (PCA) rotations, separating noise from
signal and compressing spectral information to a few bands (Green, Berman, Switzer, &
Craig, 1988). Based on the MNF output graph of eigenvectors and by visually inspecting
the new bands, 16 of the 25 bands were selected as inputs for the classification. The next
step in the flow chart is the identification of training areas or RO I's (Regions oflnterest),
or supplying spectral endmembers as stated in the SAM wizard. Regions of Interest were
selected from ground control points and augmented with visual interpretation. For the
entire 969 acre image, 82 ground truth points were available, 41 of which were used to
develop training areas and 41 of which were used to develop ROI's for accuracy
assessment purposes. Training and accuracy areas were also grown from seed pixels and
manually delineated into polygons based on visual interpretation of a I-meter Color
Infrared Image along with corresponding digital SSURGO soil maps and National
Wetlands Inventory data. The Spectral Angle Classifier was then run using a maximum
angle of O.10. Output rule images were also generated to see if any of the classes were
34
poorly identified. A comparison between the classified image and validation areas was
generated using standard post-classification techniques, resulting in a confusion matrix
(see Table 14). For more information on post-classification techniques using ENVI and
hyperspectral images, refer to Underwood, Ustin and DiPietro (2003).
Nine classes were chosen to represent all wetland types present as well as
additional aggregated land cover types. Wetland types were based on ground truth
assessments and other land cover types were based on the Anderson Level classification
(Anderson et al., 1976). The classes include three classes of wetland: Open Water,
Aquatic Vegetation, and Flooded Forest, and the other land cover classes include:
Floodplain Crop, Upland Crop, Artificial Surface, Herbaceous Cover, Shadow, and
Mixed Forest. These are shown in the final output maps (Figure 10).
For the object-oriented classifier, the same procedures were followed for steps
one through four of the flowchart shown in Figure 5. The same 16 bands used for the
SAM classifier were exported to an ERDAS Imagine format (.img), and then subset into
a smaller file size (because of eCognition's file size limitation), and lastly imported into
eCognition. The hyperspectral image was segmented using the following parameters:
33-pixel average object-size (derived from a segmentation parameter of 10 pixels), equal
weighting given to each of the 16 bands or layers, and standard nearest-neighbor
relationship for the class hierarchy. For more information on object-oriented
classification see Benz (2001), Baatz and Schape (1999), and Darwish, Leukert, and
Reinhardt (2003).
35
A standard accuracy assessment was also run in eCognition, using ground truth
points and visually interpreted areas, resulting in a confusion matrix output analogous to
ENVI's: overall accuracy, wetland users accuracy, wetland producers accuracy, and the
kappa statistic.
GIS-Based Restoration Model
A restoration model for Black Hawk County was developed to identify areas that
cannot be defined as current legal wetlands, but due to their nature of soil properties,
distance to surface hydrological features, and elevation, were most likely wetlands in the
past. It was also created to reveal what areas that would provide the most benefit for the
least cost and time when planning conservation within the county.
The model shown in Figure 8 is based on Berman et al. (2002), Braster and
Hadish (1996), Budlong (2002), Cowardin et al. (1979), Lower Mississippi River
Conservation Committee (2001), Riverlink (2000), Sader et al. (1995), and the US Army
Corp of Engineers Wetland Delineation Manual (Environmental Laboratory, 1987).
These authors found hydric soil, low slope, and distance to hydrological features and
existing wetlands as the most important variables. Those variables were weighted in this
model accordingly except for low slope, as the entire county study area is flat enough to
disregard that as a factor. Other variables included in the literature but not deemed
applicable to the study area of Black Hawk County were defining hydrological basins,
distance to levees, and amount of forest cover. Since this study is focused on wetland
restoration sites and not current delineation and mapping of wetlands, criteria one of the
36
federal wetland definition (hydrophytic vegetation) was not used. Significant land cover
changes (mostly conversion to agriculture) in the study area during the last 150 years
have resulted in destruction of most hydrophytic vegetation.
Four data layers were incorporated into the model shown in Figure 8: (a) a
SSURGO soil type layer from the NRCS, (b) a hydrology layer (rivers and streams) from
the IDNR, (c) an existing wetlands layer from the NWI, and (d) a shapefile of Black
Hawk County Conservation Areas obtained from the Black Hawk County GIS office. All
shapefiles were converted into coverages in ArcGIS 8.2 to build topology and converted
to the same projection, Universal Transverse Mercator Zone 15 North, North American
Datum 27. The four coverages were then converted into raster (grid) format in ArcGIS
Spatial Analyst to facilitate weighting and ensure uniform cell size between layers. A 30-
meter cell size was chosen as this was the lowest resolution of confidence according to
the accuracy assessments included in the layers' metadata. Also, a 1992 30-meter USGS
Landcover map was used to initially mask out areas unsuitable for restoration (urban,
bare rock and sand, open water, existing wetlands) and a 1996 county roads map from the
IDNR was used to mask out a 30-m buffer along roadsides (see Berman et al., 2002).
37
Area of Interest Land cover Mask & Roads buffer

Mask
Soils Index Hydrology Index Existing Areas Index
Hydric Soil Proximity to County Conservation

Surface Water Areas
Existing
Wetlands
Ranked Potential Wetland Restoration Areas
Figure 8. GIS-Based Wetland Restoration Model for Black Hawk County.
Weighting was accomplished by evaluating which criteria were the most
important for the study area of Black Hawk County. The Existing Areas Index was
evaluated to be the most important, with areas adjacent or contained in an existing county
conservation area ranked as a one (on a scale of one through four, one being the highest),
and the adjacent to existing wetlands factor ranked as a three. The reasoning behind this
is that wetlands have a greater chance of being restored and are easier to manage if they
are to be located in land already owned or adjacent to county conservation land (S.
Finegan, personal communication, May, 2003). The Soils Index was ranked as the next
in importance, because to meet the federal definition of a wetland, the wetland must
contain hydric soil (Cowardin et al., 1979). Therefore poorly-drained, hydric soil was
given a ranking of two. Lastly, proximity next to a surface hydrological feature was
ranked at four, because of the importance given this variable in previous studies
38
(Budlong, 2002). All cells in each layer were reclassified using Spatial Analyst's
Reclassify function by adding a column in the attribute table (RANK) and providing a
score. The final equation, adapted from Budlong (2002), was: (x = existing area index
total, y = soil index total, z = hydrology index total) RESTORATION POTENTIAL= [(x
* 0.85) + (y * 0.65) + (z * 0.40)].

39
CHAPTER4
RESULTS AND DISCUSSION
The classification results can be thought of as a binary tree, as shown in Figure 8.
Multispectral classification results for Black Hawk County include three classifiers for
the CIR imagery, and three classifiers for the Landsat ETM imagery. Hyperspectral
classification results for Eddyville include two classifiers for the CASI imagery.
Wetlands Mapping Products
Multispectral Hyperspectral
CIR Landsat ETM CASI
ISODATA
Figure 9. Research mapping results.
Multispectral CIR Image Classification
Figure 9 depicts the CIR image classification with three classifiers: ISODATA,
supervised Maximum-Likelihood, and Object-Oriented. In Figure 8, light/white areas
represent wetlands, black, water, and gray tones other land-cover classes. For full-color
results of the CIR classification, see Appendix: Maps 1 through 3.

40
CIR - IS0DATA classified CIR-ML CIR - 00 image

image classified image classified image
Figure 10. Results from Multispectral One Meter CIR Image Classification.
The object-oriented method gave the highest overall accuracy when classifying
the CIR (73.2%) while the Maximum-Likelihood classifier gave the highest wetland
producers accuracy (75.0%) while the object-oriented classifier gave the best wetland
users accuracy (50%; see Table 13). The object-oriented classifier in general performed
worse than expected, and in contrast to previous studies in which high-resolution imagery
(four meter IKONOS) had been used to identify wetlands (Antunes et al., 2003). Two
possible reasons include incorrect scale parameters used in the segmentation step and
poor spectral resolution. The eCognition (object-oriented) software consistently
performed better the more layers present there were to segment (see Table 6, 12, and 17).
However, the one-meter CIR was very useful as an ancillary data source and wetlands
could be manually (visually) delineated. Thus, CIR imagery is a cost-effective solution
to agencies seeking to define wetlands from remotely-sensed imagery. Also, the entire
county was not classified with the object-oriented classifier and the CIR image, because
of file limitations in the eCognition software, a problem that also surfaced with the
Eddyville hyperspectral image. For more information, refer to Chapter 5, Conclusion.

41
The ISODATA unsupervised algorithm differentiated artificial surfaces, tree
canopies and wetland areas better than the Maximum-Likelihood supervised method; it
was also slightly more accurate overall. Strangely, the supervised method identified the
open-water class more effectively.
Table 1 provides the error matrix for the CIR Maximum-Likelihood supervised
classification, and Table 2 shows the various accuracy percentages for different types of
land-cover classes. Similarly, Table 3 displays the error (or confusion) matrix for the
CIR unsupervised ISODATA algorithm classification, and Table 4 lists the class
accuracy percentages for the before mentioned method. Lastly, Table 5 represents the
confusion matrix for the object-oriented classification method, again using the 2002
Color Infrared one meter aerial photo. Table 6 illustrates the class accuracy percentages
for the CIR object-oriented method. One random point for the CIR supervised
classification accuracy assessment had to be discarded since it fell out ofrange of the
image; similarly 16 points had to be discarded for the unsupervised classification
accuracy assessment.
42
Table 1
Error Matrix CIR Supervised Classification
Class Artificial Open Fallow/Bare Mixed Mixed

Wetland Total
Surface Water Soil Grasses Forest
Artificial
II 3 34 2 0 0 50
Surface
Open Water 49 0 0 0 0 50
Fallow/Bare
0 48 0 0 50
Soil
Mixed
0 0 45 3 50
Grasses
Wetland 4 0 7 12 18 9 50
Mixed
43 0 2 2 49
Forest
Total 19 53 132 59 24 12 299
Table 2
Accuracy Percentages CIR Supervised Classification
Class Producers Users

Kappa
Accurac~ Accurac~
Artificial Surface 57.89 22.00 0.1671
Open Water 92.45 98.00 0.9757
Fallow/Bare Soil 36.36 96.00 0.9284
Mixed Grasses 76.27 90.00 0.8754
Wetland 75.00 36.00 0.3041
Mixed Forest 16.67 4.08 0.0007
Note. Overall Accuracy 57.86%, Overall Kappa 0.4941.
Table 3
Error Matrix CIR Unsupervised Classification

Wetland Total
Artificial
36 12 0 0 50
Surface
Open Water 7 12 25 0 2 4 50
Fallow/Bare
0 48 0 0 50
Soil
Mixed
0 0 0 50 0 0 50
Grasses
Wetland 2 0 30 2 II 5 50
Mixed
0 5 16 6 22 50
Forest
Total 47 13 120 69 20 3I 300
43
Table 4
Accuracy Percentages CIR Unsupervised Classification

Kappa
Accurac:i::: Accurac:i:::
Artificial Surface 76.70% 72.00% 0.6457
Open Water 92.31% 24.00% 0.2056
Fallow/Bare Soil 40.00% 96.00% 0.9333
Mixed Grasses 72.46% 100.00% 1.0000
Wetland 55.00% 22.00% 0.1643
Mixed Forest 70.97% 44.00% 0.3755
Note. Overall Accuracy: 59.67%, Overall Kappa: 0.5160.
Table 5
Error Matrix CIR Object Oriented Classification (GPS ground truth points)

Wetland Total
Artificial 0
4 0 0 0 0 4
Surface
Open Water 5 0 0 0 0 6
Fallow/Bare
0 0 4 0 0 0 4
Soil
Mixed
0 0 5 2 0 8
Grasses
Wetland 0 0 0 0 6 4 10
Mixed
0 0 0 0 2 3
Forest
Total 5 5 5 5 10 5 35
Table 6
Accuracy Percentages CIR Object Oriented Classification

Kappa
Accurac;r Accurac:i:::
Open Water 83.3 100.0 1.00
Fallow/Bare Soil 100.0 80.0 0.778
Mixed Grasses 62.5 100.0 1.00
Wetland 68.75 50.0 0.5 I 7
Mixed Forest 33.3 20.0 0.135

44
Multispectral Landsat ETM Image Classification
The object-oriented classifier outperformed the pixel-based methods (ISODATA
& ML) when classifying the September 2000 Landsat imagery. Overall accuracy was
higher in both the 30 m (73.9%) and 15 m (90.7%) images (Table 13). However, wetland
identification accuracy was only better than the pixel-based methods when spatial
resolution was increased (73.7% producers accuracy, 66.7% users accuracy).
Segmentation parameters were taken from previous studies (Antunes et al., 2003; Fisher,
Gustafson, & Redmond, 2002; Gomes & Marcal, 2003; Meinel, Neubert, & Reder, 2001;
Schiewe, 2001) who also used multispectral satellite imagery and reported generally
similar accuracies for different land cover types using the object-oriented classifier. In
Figure 11, the results of the Landsat ETM image classification are shown side by side for
comparison. For full color maps of the classifications refer to Appendix: Maps.
ETM- IS0DATA classified ETM-ML ETM - 00 image

image classified image classified image
Figure 11. Results from Multispectral 15 Meter Landsat ETM Image Classification.
45
In Figure 11, light areas represent wetlands, black, water, and gray tones other
land cover types. Table 7 illustrates the confusion matrix for the pan-sharpened Landsat
and Maximum-Likelihood supervised classification. Table 8 then gives the accuracy
percentages for the before mentioned method according to class. Similarly, Table 9
represents the error matrix for the ISO DATA unsupervised classification of the Landsat
image in ERDAS Imagine and Table 10 explains the class producer and users accuracy
for the ISODATA method. Lastly, Table 10 shows the object-oriented classification
error matrix for the classified Landsat ETM+ and Table 11 the object-oriented class
producer and user accuracies.
Table 7
Error Matrix Landsat ETM 15 m Supervised Classification
Class Artificial Open Mixed Mixed

Row Crop Wetland Total
Surface Water Grasses Forest
Artificial
51 3 42 30 3 9 138
Surface
Open Water 0 15 0 0 0 0 15
Row Crop 0 0 39 0 0 0 39
Mixed
0 0 0 60 0 0 60
Grasses
Wetland 0 3 0 6 21 3 33
Mixed
0 0 0 3 6 6 15
Forest
Total 51 21 81 99 30 18 300
46
Table 8
Accuracy Percentages Landsat ETM I 5 m Supervised Classification

Kappa
Accurac:r Acrnraci'.
Open Water 71.43 100.00 1.0000
Row Crop 48.15 100.00 1.0000
Mixed Grasses 60.61 100.00 1.0000
Wetland 70.00 63.64 0.5960
Mixed Forest 33.33 40.00 0.3617
Table 9
Error Matrix Landsat ETM I 5 m Unsupervised Classification

Artificial
38 0 0 2 0 0 40
Surface
Open Water 0 38 0 2 0 0 40
Row Crop 0 0 38 0 2 0 40
Mixed
0 0 0 40 0 0 40
Grasses
Wetland 0 0 6 12 16 6 40
Mixed
0 0 8 2 10 20 40
Forest
Total 38 38 52 58 28 26 240
Table 10
Accuracy Percentages Landsat ETM I 5 m Unsupervised Classification

Kappa
Accuraci'. Accuracl'.
Open Water 100.00 95.00 0.9406
Row Crop 73.08 95.00 0.9362
Mixed Grasses 68.97 100.00 1.000
Wetland 57.14 40.00 0.3208
Mixed Forest 76.92 50.00 0.4393

47
Table 11
Error Matrix Landsat ETM 15 m Object Oriented Classification

Artificial
19 0 0 0 0 0 19
Surface
Open Water 0 20 0 0 0 0 20
Row Crop 0 0 23 0 0 0 23
Mixed
0 0 33 2 0 36
Grasses
Wetland 2 0 0 14 2 19
Mixed
0 0 0 0 5 18 23
Forest
Total 20 22 24 33 21 20 140
Table 12
Accuracy Percentages Landsat ETM 15 m Object Oriented Classification

Kappa
Accuracz: Accuracz:
Open Water 100.0 90.9 0.894
Row Crop 100.0 95.8 0.95
Mixed Grasses 91.7 100.0 1.0
Wetland 73.7 66.7 0.614
Mixed Forest 78.3 90.0 0.837
For the Landsat imagery, the ISODATA classifier performed as well as the
object-oriented one for classifying land-cover types other than Wetland and Mixed Forest
types. Overall accuracy for the unsupervised classifier was superior to the supervised
classifier in both the Landsat and CIR imagery, suggesting that a "cluster-busting"
method of determining land-cover classes is more accurate than traditional Maximum-
Likelihood classification. Also, the above results (Tables 7 through 12) and the overall
Landsat results (in Table 17) are based on classification of a partial September 2000
Landsat ETM+ image. A full scene for Black Hawk County was available for July of
48
1999, and was used for the seasonal matrix and used for the final Black Hawk County
wetlands map (see Appendix A: Map 7), but not for classifier comparison due to the
large amount of flooding present on the Landsat ETM+ July 1999 image.
Hyperspectral CASI Classification
Figure 11 and Tables 13 through 16 show the classified outputs and accuracy
assessment for the object-oriented and SAM classifiers for the CASI image. Average
wetland producers accuracy for the Spectral Angle Mapper classifier was 79.3%,
somewhat higher than in other comparable studies, such as Garono, Schooler, and
Robinson (2003). They achieved 74% accuracy with the ERDAS ISODATA
unsupervised algorithm to map tidal wetlands along the lower Columbia River with CASI
imagery. The greatest confusion between wetland classes for the SAM classifier was
between the flooded forest and mixed (upland forest) categories, and also with the
emergent (herbaceous) land cover class. This has also been found in many other studies,
due to the inability of the wavelengths to penetrate the vegetation canopies. Possible
solutions to this problem include RADAR (Bourgeau-Chavez et al., 2001) and LIDAR
(MacKinnon, 2001) to penetrate dense vegetation canopies. Confusion between
emergent herbaceous vegetative cover and wetland classes is also fairly well documented
(Ozemi, 2000). A workable solution to this problem is the extraction of individual plant
species from the hyperspectral imagery, which was not completed in this study due to
time constraints.
In the object-oriented classification, average wetland producers accuracy for the
object-oriented classifier was 97.6%. Object-oriented classifiers have been shown to

49
increase accuracy of wetland classification in multispectral imagery (Antunes et al.,
2003). The results of this classification appear to be valid also for hyperspectral imagery.
Wetland users accuracy for the object-oriented classifier (86.7%) was lower than
producers accuracy, mirroring the SAM classifier. Confusion between classes was
mainly limited to forested wetland and forested upland, which was also a problem with
the SAM classifier. The accuracy assessment is based on the 50-acre study area. The
comparison between these two classifiers revealed some interesting results. The object-
oriented classifier produced better overall accuracy (92.3% vs. 68.2%) and better wetland
class accuracy (97.6% vs. 79.3%) than the SAM classifier. Wetland Users Accuracy was
lower than Producers Accuracy in both classifiers, suggesting that these two methods are
more suited to detecting wetlands than for managing them from a users standpoint.
The last wetlands survey completed for Mahaska, Wapello, and Monroe counties
was conducted by the Iowa Department of Natural Resources in 1996 and is based upon
National High Altitude Program Color Infrared Photographs taken in 1983 and 1984.
The total wetland acreage for the entire hyperspectral image area according to that
information is 53 acres. For the 50-acre study area that was classified in this research,
2.6 acres of wetlands were identified from the last wetlands survey. Comparison of the
1996 wetlands survey against the 2002 1-m CIR aerial photo clearly shows that many
wetland areas have been developed or farmed or have shifted, necessitating an updated
survey. The SAM classifier for the 50-acre study area identified 4.1 wetland acres, 0.2 of
which were Open Water, 0.6 acres of Aquatic Vegetation, and approximately 3.3 acres of
Flooded Forest. The object-oriented classifier for this study identified 3.9 acres, 0.3 of
50
which were Open Water, 0.6 acres of Aquatic Vegetation, and approximately 3 acres of
Flooded Forest.
There are also issues with these two classifiers for wetland classification. Known
sources of bias include the fact that pixel-based classifiers such as SAM in hyperspectral
imagery tend to perform best when extracting individual spectra of individual plant
species (ERDAS, 2002; ENVI, 2002) and this study grouped different species of wetland
vegetation into generic land-cover classes, a fact that might have favored the object-
oriented classifier, which inherently classifies such object-based primitives. Other studies
have shown higher accuracies using hyperspectral imagery and pixel-based methods
(such as MNF, ration indices, etc.) to extract individual plant species spectra (Garono et
al., 2003; Underwood, 2003). File size in eCognition is also another limitation. The
version of eCognition that was used, 3.0, was unable to segment and classify files larger
than 100 MB, which in this study represented 15-20 acres of the total 969 acre image.
Therefore, the CASI image had to be divided into 60 different tiles. Definiens Imaging,
the parent of eCognition software, has told the researchers that this file size limitation
will be corrected in the release of eCognition 4.0. In Figure 12, the results of the
hyperspectral classification are shown.

51
0 40 80 160 Meters 0 40 80 160 Meters
Figure 12. Results from CASI Hyperspectral Image Classification. The left-hand image
was classified using the eCognition Object-Oriented Classifier and the right-hand image,
the ENVI Spectral Angle Mapper classifier. In the left-hand (object-oriented
classification) image, the dark tones represent open water, saturated soil, and upland
(non-wetland) forest. The lighter areas are aquatic vegetation and herbaceous cover. The
gray-tone classes represent dry bare soil and flooded (wetland) forest. On the right-hand
side image (Spectral Angle Mapper classification), dark tones represent open water,
saturated soil, and aquatic vegetation. Lighter areas represent upland and flooded forests,
and gray-tones herbaceous cover and dry bare soil.
52
Table 13
Confusion Matrix for SAM Classification (Ground Validation Pixels)
Open Aquatic Flooded Artificial Upland Flooded Herbaceous Mixed

Class Shadow Total
Water Vegetation Forest Surface Cro2 Cro2 Cover Forest
Unclassified 13 0 32 22 0 5 3 17 6 98
Open Water 230 0 0 192 0 0 0 0 0 422
Aquatic
0 318 0 0 0 0 0 0 0 318
Vegetation
Flooded 191 0 496
0 0 0 0 38 86 3
Forest
Artificial
0 0 0 496 0 0 0 0 319
Surface
Upland Crop 0 0 0 35 516 62 0 0 0 613
Flooded 0 2 0
0 0 0 0 0 0 280
Crop
Herbaceous 0
0 0 84 0 3 141 49 4 561
Cover
Mixed
0 0 125 0 199 188 0 421 12 387
Forest
Shadow 0 0 II 0 0 13 0 181 71 276
Total 243 318 443 745 715 274 182 754 96 3770
Table 14
Accuracy Percentages CASI SAM Classification
Class Prod. acc.(%) User acc.(%)

Open Water 94.65 54.5
Aquatic Vegetation 100.00 100.00
Flooded Forest 43.12 59.87
Artificial Surface 66.58 100.00
Upland Crop 72.17 84.18
Flooded Crop 68.61 45.58
Herbaceous Cover 77.47 50.36
Mixed Forest 55.84 75.04
Shadow 73.96 25.72
Wetland Avg. (3 classes) 79.3 71.5

53
Table 15
Confusion Matrix for Object-Oriented Classification (Ground Validation Objects)

Open Aquatic Flooded Artificial Upland Flooded Herbaceous Mixed
Class Shadow Total
Water Vegetation Forest Surface Croe Croe Cover Forest
Unclassified 0 0 0 0 0 0 0 0 0 0
Open Water 5 0 0 0 0 0 0 0 0 5
Aquatic 0 0
0 19 0 0 0 0 0 19
Vegetation
Flooded 0
0 0 13 0 0 0 0 14
Forest
Artificial
0 0 0 10 0 0 0 0 0 10
Surface
Upland 0 0 20
0 0 0 0 20 0 0
Crop
Flooded
0 0 0 0 0 20 0 0 2 22
Crop
Herbaceous 0 0 0 0 19 0 0 19
0 0
Cover
Mixed
0 7 0 0 0 0 20 0 28
Forest
Shadow 0 0 0 0 0 0 0 0 6 6
Total 5 20 20 10 20 20 20 20 8 143
Table 16
Accuracy Percentage Classifications CASI Object-Oriented Classification
Class Prod. acc. (%) User acc. (%)

Open Water 100.0 100.0
Aquatic Vegetation 100.0 95.0
Flooded Forest 92.9 65.0
Artificial Surface 100.0 100.0
Upland Crop 100.0 100.0
Fl oode<l Crop 90.9 100.0
Herbaceous Cover 100.0 95.0
Mixed Forest 71.4 100.0
Shadow 100.0 75.0
Wetland Avg. (3 classes) 97.6 86.7
Overall Classifier Comparison
The overall classification accuracy for different classifiers is provided in Table
17. Accuracy comparisons between the classifiers were completed using the same areas;
for example, county-wide stratified random points for the CIR and Landsat images, and
the same 50-acre subset for the hyperspectral Eddyville image. Accuracies increased
54
(both for the wetland class and the overall average) when spatial resolution of the Landsat
imagery was sharpened with the panchromatic band; also, the unsupervised ISO DATA
algorithm performed better for overall accuracy than the supervised Maximum-
Likelihood classifier. This is consistent with the results of other studies (Ozemi, 2000).
The object-oriented classifier increased overall accuracy with the Landsat imagery over
the traditional pixel-based classifiers, but did not increase wetland-identification accuracy
until the spatial resolution was increased (see Table 13). The CIR imagery in general
performed poorly with all automated classifiers, suggesting that even though the spatial
resolution was very sharp, either more bands such as Landsat 4 & 5, (see Chen 2002), are
needed to detect vegetation, or seasonality played a role because the imagery was flown
in late April/early May of 2002 before the growing period of many wetland vascular
plants in the northeastern part of the state. The most accurate results came from the
hyperspectral object-oriented approach and the pan-sharpened Landsat object-oriented
approach.
The seasonal matrix of the pan-sharpened Landsat images produced lower
accuracies than anticipated, especially for identifying wetland areas. It did, however,
increase accuracies for row-crop cover and herbaceous cover. This may be due to the
large amount of flooding present in the July 1999 Landsat image. Landsat imagery
remains a valid choice for large-scale wetlands mapping projects, especially with the
added capability of the panchromatic band.

55
Table 17
Overall Classification Accuracy for Different Classifiers
HYPER- HYPER-
CIR CIR LSAT LSAT LSAT LSAT SPECTRAL SPECTRAL
CIR lm lm 15 m 15 m 15 m 15 m LSAT CASI 60 CASI 60
lm ISO- Max- ISO- Max- Pan Seasonal 30m cm cm
0-0 DATA Like DATA Like 0-0 Matrix 0-0 0-0 S.A.M.
Overall
Accuracy 73.2 59.7 57.9 79.17 64.0 90.7 67.3 73.9 91.7 68.2
Wetland
Producers 68.8 55.0 75.0 57.14 70.0 73.7 77.78 58.8 94.6 79.3
Accuracy
Wetland
Users 50.0 22.0 36.0 40.0 63.6 66.7 46.67 50.0 86.7 71.5
Accuracy
Kaeea 0.701 0.52 0.49 0.75 0.552 0.888 0.622 0.667 0.904 0.637
Note. 0-0: Object-Oriented Classifier, ISODATA: Unsupervised Classification, 120 Initial Classes
Max-Like: Maximum Likelihood Algorithm, S.A.M.: Spectral Angle Mapper Algorithm.
GIS-Based Restoration Model Results
The results of the wetland restoration model are as follows. Black Hawk County,
Iowa encompasses an area of 567 square miles, or 362,880 acres. From that initial
acreage, 56,729 acres were masked out as unsuitable based on the USGS Landcover
raster layer of the following classes: (a) Low Intensity Residential, (b) High Intensity
Residential, (c) Commercial/Industrial/Transportation, (d) Bare Rock/Sand/Clay, (e)
Quarries/Strip Mines/Gravel Pits, (f) Urban/Recreation Grasses, (g) Open Water, (h)
Woody Wetlands, and (i) Emergent Herbaceous Wetlands. The accuracy assessment for
that 1992 data set was made publicly available on March 17, 2004 (United States
56
Geological Survey, 2004). The overall accuracy for Black Hawk County was only 53%,
but for the classes masked out the accuracy was 60.4%. Urban areas were chosen as
unsuitable restoration areas as well as bare areas and existing wetlands. Landcover
classes left as suitable for restoration included: (a) Deciduous Forest, (b) Mixed Forest,
(c) Grasslands/Herbaceous, (d) Pasture/Hay, (e) Row Crops, and (f) Small Grains.
From the remaining 306,151 acres, 93 .4 acres in 10 parcels were also masked out
as they are recognized as wetland parcels by the county assessor's office. Non-urban
county roads along with a 30-meter buffer totaling 24,724 acres were also masked out to
eliminate right-of-way areas owned by the Iowa Department of Transportation (see
Berman et al., 2002). That left 281,334 acres or 440 square miles for wetland restoration
consideration. Based on the flow chart in Figure 8 on page 36, cells classified as having
soil that was hydric with poor drainage were given a score of 4 in the SSURGO soil data
layer. Cells that did not meet this criterion were eliminated from consideration, as
according to the federal definition of a wetland (see page one) a wetland must contain
hydric soil, hydrophytic vegetation, or be in an area where the water table saturates a
non-soil substrate or covers the area with shallow water periodically. As hydrophytic
vegetation would not necessarily be present in a disturbed land-cover area (such as an
agriculturally based one) and water table depth information was not available, the hydric
soil criterion was chosen as the ranking factor or Step 2 in the flowchart. Eliminating
cells that were not hydric resulted in narrowing the suitability area down to 121,271
acres. The next step was determining if the potential wetland restoration area was within
a buffer of 20 meters for a hydrological feature (stream) or 50 meters of the Cedar River,
57
as wetland areas are proven floodwater storage areas (Sierra Club, 2000) and wetlands
restored adjacent to hydrological features moderate stream temperature and reduce
erosion (Budlong, 2002). These cells were given a score of 1, and totaled 17,491 acres.
Step 4 included identifying cells that were adjacent to existing wetlands. For the wetland
areas, a 1996 Iowa Department of Natural Resources wetlands layer was used (based on
1983/84 aerial photos) instead of the updated Black Hawk County wetlands map
produced earlier in the research as the wetland producer and user accuracies were lower
(73.7 and 66.7% respectively) in the updated wetlands map. However, the model was run
using the updated wetlands map and produced results less than one standard deviation
from the mean as compared with using the older wetlands data, indicating no dramatic
shifts in wetland areas or total acreage, a fact also shown by the Black Hawk County
Wetlands timeline in Figure 13 on page 57. Cells that were adjacent to existing wetlands
were given a score of 2 (see also Hey & Philippi, 1999).
Lastly, cells that were adjacent to or contained within county conservation areas
were given a score of 3. The reasoning behind this is that wetlands have a greater chance
of being restored and are easier to manage if they are to be located in land already owned
or adjacent to county conservation land (S. Finegan, personal communication, May,
2003). The final equation from the methodology section was: (x = existing area index
total, y = soil index total, z = hydrology index total) RESTORATION POTENTIAL= [(x
* 0.85) + (y * 0.65) + (z* 0.40)]. Cell scores were computed using ArcGIS Spatial
Analyst raster calculator according to the above equation (for example, if a cell had a
58
perfect ranking, meeting all desirable criteria, [((3+2)*0.85) + (4*0.65) + (3*0.40)], it
would have a score of 8.05.
Three categories were then defined using a Jenks Natural Breaks method: Cells
with scores of 3.9 through 8.05 were ranked as most suitable areas, cells with scores of
2.7 through 3.8 were given a ranking of medium suitability, and cells with scores of 2.6
were given a ranking of low suitability. Cells with scores lower than 2.6 were deemed as
unsuitable or unclassified. For a map of these areas, refer to Map 9 in Appendix A.
There were not too many surprises in the results of the model, as all areas were close to
surface hydrological features and generally were in areas where wetlands were present
historically. The topography of the county does not vary greatly, and some areas have
been known to county conservation officials for some time as highly-suitable areas for
wetland restoration, such as the Crane Creek watershed and areas in the southeastern part
of the county along the Cedar River. What this study contributes, through the use of GIS,
is to demonstrate where restoration of wetlands could and should take place if county or
state resources become available. Table 18 displays a numerical summary of the
restoration model. Figure 13 references a timeline for wetland changes in Black Hawk
County.
59
Table 18
Results from GIS-Based Restoration Model
Cell TYPe Cell Total Area (acres)

Black Hawk County 362,880
Highest Suitability 2,971
Medium Suitability 34,307
Lowest Suitability 121,271
Unsuitable 204,331
USGS Landcover Mask 56,729
Wetland Parcel Mask 93.4
County Roads Buffer Mask 24,724
Hydric Poorly Drained Soil 121,271
Adjacent to Hydrology 17,491
Adjacent to Wetlands 30,590
Adjacent or Contained in County
7,170
Conservation Area
Wetlands Change in Black Hawk County
50000
~ 40000
~
< 30000
§ 20000
~ 10000
0 +-------=
1832 1983 1992 2004
Source: IDNR Source: NW! Source: USGS Source: Harken
Figure 13. Black Hawk County Wetlands Timeline.
As shown by the above timeline, Figure 13, wetlands have decreased in the past
twenty years in Black Hawk County. This is possibly due to natural variations in the
hydrological cycle, agricultural practices, or image bias. While the acreage amount of
decrease is not great, it still shows a need for restoration planning and implementation.
As Hey and Philippi (1999) note, wetlands can be restored to provide ecological benefits
60
(wildlife habitat, non-point pollution treatment, flood water storage) that have been lost.
They also noted that wetland restorations are most effective when the wetlands occupy
less than 10% of the area to be restored, as is the case in Black Hawk County where
wetlands currently account for only 5% of the county's surface area.
Web-Based Data Dissemination
The Black Hawk County wetland project homepage is available to the general
public at http://gisrl-9.geog.uni.edu/wetland/ (Figure 14). The homepage explains the
goals and objectives and also methodologies and protocols developed in this project. It
also provides a summary of results, links to other wetland sites, and a comprehensive list
of references. Also available on the website is a technical report published for the Iowa
Space Grant Consortium in January 2003.
Figure 15 shows a screen shot of the ArcIMS viewer. An ArcIMS-based web
page was created so stakeholders in the project as well as the general public could access
the results and use them for their own needs. According to ESRI (2003), ArcIMS is
software specifically designed to serve geographic data on the Internet, and to develop
Web pages that communicate with maps. Potential uses of the website include: (a)
landowners identifying parcels of land that would be highly suitable to restore wetlands,
(b) corroborating evidence for local government officials for conservation planning, and
(c) general information on wetlands in Black Hawk County for the public. Users of the
website can select different layers to display, as well as use built-in functions such as a
measuring, querying, and buffering tools.

61
fl< f.dl i,,. '""!'" .tooo u,;,

o,... · e l.'!,l Lt (Q;P"""' -fy,-•.. "5'- 0 9.1:;., l:il • ., ~
Wetl&ndt ~ tl'lwiU~I la~d# l,et'~ttn ~ I and •quatic•y,wnt -~~ pr,:mde i:re.ny gwd1 fU\d semces,i.tlcluding f!'.ood ~l'l!tentian,
".t'•t«(j'J.aUtytMJntQW'lc:""".UdW. habJtat,and flX1 erotionCQ!lfrot. WettaA4' irl.low• t,4w dec:N>Ue4owr'$1$~inth• hm:2':IO yea:,. Th• lou ot
thr.-1111 rt:tn..i~ and bm~ftt• ,h,wr, & n«-dtomapan.d l'cWIJ:mtW'«li1.nd11:a k,w~r.,~flywht.ffeumnt. ~lti.nd ma.pt art:Ndl'lff.d,u tJthni".AH
ill Bt.dr.Haw"k ~ur.ty. Tradlttotial w«W\d ground ,arvq-, aN:tootiMI c:on,uming llild e:,;pa,.t!W: though,~d l"W'de 1e1111ng ln'lagerycan
r,cn,14tt1MII' ud m.t:~~vt tol1rticu\a., Thi.I rtudy~ tht hl~,l~~iltty0f iui~g hlg:h moluuon aJrl>offst Colet fo1ta.red (cm)1n-.gtry,rru.i1t1 ..~ I
hyt,rill dirta,an4 high moiulwti hypa-spodnil Jm11.guyla1tapand u~litt.lt ""i!!:lrw.4&1- it1 W111ra. l!\addi:lton,thll projl!d ~ treiWotta!.
cl.u1tfie!'.I! "'Jth objtct:-o.t1e11ted (OO)ctau~6atian t!!chnlque:,to lmptt)vt ~«la.nd mappit13 accuracy We1:lia.nd1 1n E:lite:tHll"f1c.County 1howcd il
11ight d«.'ft<IIIH of ,wghly l!XXI 111cn:t (+/•an U'Nt'Ull.rgin of37S acm: )from.l983-2QOO:, ht gen'"'l,the COclaa,lfi!!rptirlanred bttterthan L,ditional
pixel-1,:,oHd l«hmquu (Mai<iMun\ t.Jr.i!Uhood fl!,ld tSOOATA)fortheCIR and l&nduJt. 1n-.ap:ryand bfflt-.rth4n l:ht$pectfl;1 Antle MApper($AM)
ltthoJqutbtl1.1C,\Slin•r•fU:1•tly,a 'Wcb-1,,Jtdtool WU dew,lope:d Uling:ArclM$to diFFemirtattthe do'lta devdOf"!diromthu: m&d)'to
,t.tbholdets. '
• Jou.nwl "11klt So.'bnduJcm:

0 ClaHlficatlanoflo'fa Wetl.uidt'J.tlng:•Aut-oh'l#Jiypei,~~t:'. ACCA'lf'M'9nOf~Mgle:Mappe.r("'.JAM)Cla,1"1eraadaA
O\,jttt-Orl.mt11d (OO)Ap~ ~Cana.di&nJouma.I oJ'Rcrmtt $ttntmg (•ubm.'ttcd)
~ b;iplemen.tlttJ a GlS~aud Wd:land P•ton.tloll Model rot Bl•clt Hawk O:,,,i,nt.y,JA. (under~re.tion)
• Cont«nuct Pru1nt1lion1;
o 2003lGJCCotlht~na,AJ:¥1/9,Iowa;~S~q,t,1Wcfo11t4Dy~fnB1adrHaa1kC~1A
o 2003 ArcU•en ConI~ce,Omaha,Nt!brad:a:
Ut:tntArcQSWM~tff«r.tti4JW«1,rJ~wnSCftti181odtHaa.,1f:C~l,1,
o 2004. ESFJhrtmnttlru1111t U,erCookrua,San Oi.ego,Ca.lik;mia: ll.rdM.S flfdW.tr411d R~icnmBrct1-H1tVkC'1U.~1J\{fort~Qdog)
• Put.nerddp, cn«tt-4
o But~m,BlftckHa.,u-lr.CountyGISoffi.c:e
o ~n.Finesan,.BW:kJ-Uwl::Coo.LntyCotwffli"i\tlonBc».rd
O ht.411 Y~l1ud1, 1owa Oep11utvmil cl N"tural Re,'aun;e,
¢,JQwa Ptp•rtil'lt'otolTiu1portation Wetland• Utut
Figure 14. Screen Shot of Homepage

62
I Black Hawk County Wl!tlands Projl!ct Microsoft tntMnl!t lxplorM ~@~I

f'l!t !,_ Yle!II P'-..te, Tools Help
_O ._...~ _Q· ~_[~l_~ 1-e..... <fl,-.. II?- ~t~-~ ~ :__D -5
Figure 15. Screen Shot of Web-Based Data Interface

63
CHAPTERS
CONCLUSION
In this study multispectral (CIR; ETM) and hyperspectral (CASI) images were
tested for wetland classification using different classifiers (Maximum-Likelihood,
ISODATA, Object-Oriented and Spectral Angle Mapper). The object-oriented classifier
produced superior results over traditional pixel-based classifiers (ISODATA; Maximum-
Likelihood) in multispectral imagery (73.7% vs. 57.14%; 70.0%) when mapping
wetlands but only when spectral resolution was increased (i.e., 15-m 6-band Landsat
imagery achieved superior results over 1-m 3-band Color Infrared Aerial photography)
and the spatial resolution of the Landsat imagery was increased (Pan-sharpened from 30-
m to 15-m).
The results for Eddyville also clearly showed that hyperspectral images enabled
more accurate wetland mapping than multispectral datasets when using the object-
oriented classifier (94.6%) and the SAM classifier (79.3%). The object-oriented
classifier in this case also performed better than the pixel-based (SAM) classifier. A
seasonal comparison of Landsat imagery to identify wetlands did not produce accurate
results, perhaps due to extensive flooding present in the summer (July) imagery.
The answers to the research questions from Chapter 1, Page 7, are as follows:
1. The Object-Oriented classifier was superior and more accurate in comparison
to the pixel-based Maximum-Likelihood and ISODATA for the delineation of wetlands
using multispectral imagery in Black Hawk County.

64
2. Data fusion between a Landsat ETM multispectral and ETM panchromatic
band increased the accuracy of wetland classification in Black Hawk County.
3. The Object-Oriented Classifier was more accurate than the SAM classifier for
identifying wetlands in the Eddyville hyperspectral (CASI) image.
4. Seasonality may play a significant role in classifying wetlands from remotely-
sensed imagery; however this study did not yield anticipated higher wetland detection
accuracies.
5. The most important variables for a GIS-based wetland restoration model in
Black Hawk County were: (a) hydric soil; (b) proximity to surface hydrological features;
(c) proximity to existing wetlands; and (d) proximity to existing conservation areas
owned by the county.
The results of the GIS-based model used in this study for wetland restoration in
Black Hawk County identified far more acres than initially believed were suitable for
such purposes (56% of county land area deemed unsuitable, 33% low suitability, 10%
medium suitability, 1% highly suitable). Two highly-suitable identified areas had already
been previously targeted by conservation officials for wetland easements or restorations
should funding become available.
Known sources of error include the fact that any wetland identified through
remotely-sensed imagery must be field-checked by a qualified ecologist or biologist in
order to qualify for legal status or protection. Wetlands in Black Hawk County showed a
slight decrease of roughly 1500 acres(+/- an error margin of 375 acres) from 1983-2003.
A web site with an ArcIMS viewer was created in order to disseminate information to the
65
stakeholders involved in the study, as well as the general public. Information available
on the website include: a summary of findings, maps of classification results, wetland
links, a comprehensive bibliography, as well as a technical report published for the Iowa
Space Grant Consortium. Available on the ArcIMS site is the ability to arrange different
layers, as well as measurement, buffer, and query tools.
In conclusion, it is the findings of this research that wetland classification from
multispectral imagery in the study area can be accurately completed if spatial resolution
is increased by data fusion, but not at the cost of spectral resolution. A non-parametric
object-oriented classifier can also identify freshwater inland wetlands for the study areas
of Black Hawk County and Eddyville more accurately than traditional pixel-based
(ISODATA, Maximum-Likelihood) ones. Hyperspectral imagery is preferable to
multispectral imagery in identifying freshwater inland wetlands because of increased
spectral resolution. The object-oriented classifier also identifies wetlands more
accurately using hyperspectral imagery, but has limitations for large file sizes.
Limitations of the research include: (a) image availability for the seasonal matrix,
(b) classifying only a subset of the multispectral and hyperspectral imagery with the
object-oriented classifier, and (c) a limited number of variables used in the GIS-based
restoration model.
66
Recommendations for Future Research
The future direction of this study lies in testing more non-parametric classifying
methods, such as a CART (classification and regression tree) algorithm since other
studies have shown it to be more accurate than a purely spectral based classifier
(Sugumaran, Pavuluri, & Zerr, 2003). In addition, greater long-term seasonality will also
be addressed, as other studies have stressed the importance of multi-seasonal variation in
detecting wetlands via remote-sensing imagery (Ozemi, 2000; Houhoulis & Michener,
2000). Future efforts for the restoration model include adding more variables, such as
land ownership, as well as field testing of high potential sites for evidence ofhydrophytic
vegetation and confirmation of hydric soils.

67
REFERENCES
Acreman, M. C., & Hollis, G. E. (1996). Water management and wetlands in Sub-
Saharan Africa. Gland, Switzerland: IUCN.
Alsdorf, D. E., Smith, L. C, & Melack, J.M. (2001). Amazon floodplain water level
changes measured with interferometric SIR-C radar. IEEE Transactions on
Geoscience and Remote Sensing, 39(2), 423-431.
Anderson, B. D., Garono, R., & Robinson, R. (2003). CASI and Landsat: Developing a
spatially linked, hierarchical habitat cover dataset along the lower Columbia
River, USA. Retrieved December 9, 2003, from
http://www.waterobserver.org/event-2003-06/pdf/paper-04-June-05-
BAnderson.pdf
Anderson, J. R., Hardy, E., Roach, J. T., & Witmer, R. E. (1976). A land use and land
cover classification system for use with remote sensor data (United States
Department of the Interior). Washington DC: U.S. Government Printing Office.
Anderson, P. F. (1992). Primer for Geodesy: GIS teaching software for landscape
planning. Ames: Department of Landscape Architecture, Department of
Agronomy, Land Use Analysis Laboratory, Iowa State University.
Antunes, A. Z. B., Lingnau, C., & Da Silva, J. C. (2003). Object-oriented analysis and
semantic network for high resolution image classification. Retrieved January 14,
2003, from http://www.definiens-imaging.com/documents/publications/felipe.pdf
Arbuckle, K., & Pease, J. L. (1999). Managing Iowa habitats: Restoring Iowa wetlands.
[Iowa State University Extension, Pm-1351h]. Ames: Iowa State Press.
Baatz, M., & Schape, A. (1999). Multiresolution segmentation: An optimization approach

for high quality multi-scale image segmentation. In the AGIT Symposium
Angewandte Geographische Informationsverarbeitung XI [Geographic
Information Processing Vol. 11]. Salzburg: Definiens Imaging.
Baatz, M., & Schape, A. (2001). eCognition User's Manual. [Computer software and
manual]. Munich, Germany: Definiens Imaging.
Bakker, W. H., & Schmidt, K. S. (2002). Hyperspectral edge filtering for measuring
homogeneity of surface cover types. Photogrammetry & Remote Sensing, 56,
246-256.
Benz, U. (200 I). Definiens Imaging GmbH: Object-oriented classification and feature
detection. IEEE Geoscience and Remote Sensing Society Newsletter, 8, 16-20.
68
Benz, U., Hofmann, P., Willhauck, G., Lingenfelder, I., & Heynen, M. (2004). Multi-
resolution, object-oriented fuzzy analysis ofremote sensing data for GIS-ready
information. ISPRS Journal of Photogrammetry & Remote Sensing, 58, 239-258.
Berman, M. R., Rudnicky, T., Berquist, H., & Hershner, C. (2002). Protocols for
implementation of a GIS-based model for the selection ofpotential wetlands
restoration sites in southeastern Virginia. Gloucester Point, VA: Center for
Coastal Resources Management, Virginia Institute of Marine Science, College of
William and Mary.
Boardman, J. W., & Kruse, F. A. (1994). Automated spectral analysis: A geological

example using AVIRIS data, northern Grapevine Mountains, Nevada. In
Proceedings, Tenth Thematic Conference, Geologic Remote Sensing (pp. 407-
418). San Antonio, TX: Environmental Research Institute of Michigan.
Bourgeau-Chavez, L. L., Kasischke, E. S., Brunzell, S. M., Mudd, J.P., Smith, K. B., &
Frick, A. L. (2001 ). Analysis of space-borne SAR data for wetland mapping in
Virginia riparian ecosystems. International Journal of Remote Sensing, 22(18),
3665-3687.
Braster, M., & Radish, G. (1996). Floodplain land use and resource management
alternatives for western Iowa. Ames: Iowa State University Press.
Brown, T. N, Johnston, C. A., & Cahow, K. R. (2003). Lateral flow routing into a
wetland: Field and model perspectives. Geomorphology, 53, 11-23.
Budlong, R. C. (2002). The use of spatial data in creating a riparian buffer suitability
model: Whitewater River watershed, Minnesota. Winona, MN: Dept. of Resource
Analysis, St. Mary's University of Minnesota, U.S. Fish and Wildlife Service.
Carter, G., Wells, T., & Lewis, D. (2004, March) Hyperspectral remote sensing of
invasive wetland plants in northern Mobile Bay. Paper presented at the annual
meeting of the American Association of Geographers, Philadelphia, PA.
Chapman, H.P., & Cheetham, J. L. (2002). Monitoring and modeling saturation as a

proxy indicator for in-situ preservation in wetlands: A GIS-based approach.
Journal ofArchaeological Science, 29, 277-289.
Chen, J. H., Chun, E. K., Tan, C.H., & Shih, S. F. (2002). Use of spectral information for
wetland evapotranspiration assessment. Agricultural Water Management, 55,
239-248.
69
Chopra, R., Verma, V. K., & Sharma, P. K. (2001). Mapping, monitoring and
conservation of Harike wetland ecosystem, Punjab, India, through remote sensing.
International Journal ofRemote Sensing, 22(1 ), 89-98.
Civco, D. L., Hurd, J. D., Wilson, E. H., Song, M., & Zhang, Z. (2002, April). A
comparison of land use and land cover change detection methods. Paper
presented at the meeting of the American Society for Photogrammetry and
Remote Sensing, Denver, CO.
Cohen, D. (2001). Iowa wetlands. [Iowa Association of Naturalists, Iowa Biological

Community Series, IAN-204]. Ames: Iowa State University Press.
Cowardin, L. M., Carter, V., Golet, F. C., & LaRoe, E.T. (1979). Classification of
wetlands and deepwater habitats of the United States. (U.S. Fish and Wildlife
Service). Washington, DC: U.S. Government Printing Office.
Dahl, T. E. (1990). Wetlands losses in the United States 1780's to 1980's. (U.S. Dept of
the Interior, Fish and Wildlife Service). Washington, DC: United States
Government Printing Office.
Dahl, T. E. (2000). Status and trends of wetlands in the conterminous United States 1986
to 1997. (U.S. Dept of the Interior, Fish and Wildlife Service). Washington, DC:
United States Government Printing Office.
Darwish, A., Leukert, K., & Reinhardt, W. (2003, June). Image segmentation for the
purpose of object-based classification. Paper presented at the meeting of the
International Geoscience and Remote Sensing Symposium, Paris, France.
Dierberg, F. E., DeBusk, T. A., Jackson, S. D., Chimeny, M. J., & Pietro, K. (2002).
Submerged aquatic vegetation-based treatment wetlands for removing phosphorus
from agricultural runoff: Response to hydraulic and nutrient loading. Water
Research, 36, 1409-1422.
Dung, J.E. (2003). Agricultural land evaluation in Black Hawk County using geospatial
technologies. Unpublished master's thesis, University of Northern Iowa, Cedar
Falls.
ENVI. (2002). 3.6 User's Manual [Computer Software and Manual]. Rochester, NY:
Kodak Research Systems.
Environmental Laboratory. (1987). Corps of engineers wetlands delineation manual.

[Technical Report Y-87-1]. Vicksburg, MS: U.S. Army Engineering Corp.
70
ERDAS. (2002). ERDAS Imagine v. 8.6 Spectral Analysis User's Guide [Computer
Software and Manual]. Heerbrugg, Switzerland: Leica Geosystems.
ESRI. (2003, May). Arc/MS 4 architecture and functionality (White Paper J-8900).
Redlands, CA: Environmental Research Systems, Inc.
Fisher, C., Gustafson, W., & Redmond, R. (2002). Mapping sagebrush/grasslands from
Landsat TM-7 imagery: A comparison of methods (U.S. Department of the
Interior, Bureau of Land Management, Montana/Dakota State Office).
Washington, DC: U.S. Government Printing Office.
Garono, R., Schooler, S., & Robinson, R. (2003). Using CASI imagery to map the extent
ofpurple loosestrife (Lythrum salicaria) in tidally influenced wetlands of the
Columbia River estuary. Retrieved December 9, 2003, from
http://www.waterobserver.org/
Goldberg, J. (1998). Remote sensing of wetlands: Procedures and considerations.

Retrieved January 24, 2003, from
http://www.vims.edu/rmap/cers/tutorial/rsceol.htm
Gomes, A., & Marca!, A. (2003, May). Land cover revision through object based
supervised classification ofASTER data. Paper presented at the meeting of the
American Society for Photogrammetry and Remote Sensing, Anchorage, AK.
Green, A. A., Berman, M., Switzer, P., & Craig, M. D. (1988). A transformation for
ordering multispectral data in terms of image quality with implications for noise
removal. IEEE Transactions on Geoscience and Remote Sensing, 26(1), 65-74.
Halid, M. (1997). Land use-cover change detection using knowledge-based approaches:

Remote sensing and GIS. Retrieved May 2, 2003, from
http://www.gisdevelopment.net/aars/acrs/1997/ts8/ts8001 pf.htm
Han, M., Cheng, L., & Meng, H. (2003). Application of four-layer neural network on
information extraction. Neural Networks, 16, 547-553.
Harken, J., & Sugumaran, R. (2004) Classification ofIowa wetlands using an airborne
hyperspectral image: A comparison of Spectral Angle Mapper (SAM) classifier
and an Object-Oriented (00) approach. Manuscript submitted for publication.
Hey, D. L., & Philippi, N. S. (1999). A case for wetland restoration. New York: Wiley &
Sons.
71
Hodgson, M. E., Jensen, J. R., Mackey, H. E., & Coulter, M. C. (1987). Remote sensing
of wetland habitat: A wood stork example. Photogrammetric Engineering &
Remote Sensing, 53, 1075-1080.
Houhoulis, P. F., & Michener, W. K. (2000). Detecting wetland change: A rule-based

approach using NWI and SPOT-XS data. Photogrammetric Engineering &
Remote Sensing, 66(2), 205-211.
Iowa Department of Agriculture and Land Stewardship. (1998). Iowa wetlands and
riparian areas conservation plan. Retrieved April 30, 2003, from
http://www. ag. iastate. edu/centers/iawetlands/NWlhome.html
Iowa Department of Natural Resources. (2004). Restoring wetlands. Retrieved April 18,
2004, from http://www.iowadnr.com/wildlife/files/restorwet.html
Iowa Department of Transportation & National Consortium on Remote Sensing in

Transportation - Environmental Group. (2002). NCRST-E remote sensing
mission to Eddyville, Iowa in conjunction with the Iowa Department of
Transportation. Retrieved December 2, 2003, from
http://www.ncrste.msstate.edu/publications/posters/ncrste_ eddyville_mission_200
l.pdf
Ivits, E., & Koch, B. (2002). Object-oriented remote sensing tools for biodiversity
assessment: A European approach. In Proceedings of the 22nd EARSeL
Symposium. Rotterdam, Netherlands: Millpress Science Publishers.
Jenssen, L. L., & Middelkoop, H. (1992). Knowledge-base crop classification of Landsat

Thematic Mapper image. International Journal ofRemote Sensing, 13(15), 2827-
2837.
Juan, C., Jordan, J. D., & Tan, C. H. (2000). Application of airborne hyperspectral
imaging in wetland delineation. Retrieved October 13, 2003, from
http://www.gisdevelopment.net/aars/acrs/2000/ts 16/hype0006pf.htm
Kaya, S., Pultz, T. J., Mbogo, C. M., Beier, J. C., & Mushinzimana, E. (2002). The use of
radar remote sensing/or identifying environmental factors associated with
malaria risk in coastal Kenya. Retrieved June 19, 2003, from
http://www.pcigeomatics.com/tech-papers/igarss02_kaya_paper.pdf
Konecny, G. (2003). Geoinformation: Remote sensing, photogrammetry and Geographic

Information Systems. New York: Taylor & Francis.
Lillesand, T. M .. , & Kiefer, R. W. (2000). Remote sensing and image interpretation. New
York: John Wiley & Sons.
72
Loiselle, S., Bracchini, L., Bonechi, C., & Rossi, C. (2001). Modelling energy fluxes in
remote wetland ecosystems with the help ofremote sensing. Ecological
Modelling, 145, 243-261.
Lower Mississippi River Conservation Committee. (2001). Spatial model to prioritize

wetland forest restoration on private lands in the Mississippi alluvial valley to
enhance water quality. Retrieved March 19, 2003, from
http://www.lmrcc.org/newmodel. pdf
Lunetta, R., & Balogh, M. (1999). Application of multi-temporal Landsat 5 TM imagery

for wetland identification. Photogrammetric Engineering and Remote Sensing, 65
(11), 1303 -1310.
Lyon, J. G. (1993). Practical handbook for wetland identification and delineation. Ann
Arbor, MI: Lewis Publishers.
Lyon, J. G., & McCarthy, J. (Eds.). (1995). Wetland and environmental applications of
GIS. Boca Raton, FL: Lewis Publishers.
MacKinnon, F. (2001). Wetland application ofLIDAR data: Analysis of vegetation types

and heights in wetlands. Nova Scotia: Applied Geomatics Research Group,
Centre of Geographical Sciences.
Marcus, W. A., Legleiter, C. J., Aspinall, R. J., Boardman, J. W., & Crabtree, R. L.
(2003). High spatial resolution, hyperspectral (HSRH) mapping of in-stream
habitats, depths, and woody debris mountain streams. Geomorphology, 55(1-4),
363-380.
Meinel, G., Neubert, M., & Reder, J. (2001). The potential use of very high resolution
satellite data for urban areas: First experiences with IKONOS data, their
classification and application in urban planning and environmental monitoring.
Remote Sensing of Urban Areas, 35, 196-205.
Mitsch, W. J., & Gosselink, J. G. (2000). Wetlands: Characteristics and boundaries.

New York: Wiley & Sons.
Ozemi, S. L. (2000). Satellite remote sensing of wetlands and a comparison of

classification techniques. Unpublished doctoral dissertation, University of
Minnesota, Minneapolis.
Rio, J. N. R., & Lozano-Garcia, D. F. (2000). Spatial filtering of radar data

(RADARSAT) for wetlands (brackish marshes) classification. Remote Sensing of
the Environment, 73, 143-151
73
Riverlink. (2000). GIS management models to assist wetland restoration efforts in

Buncombe and Haywood counties, North Carolina. Retrieved February 22, 2003,
from http://www.riverlink.org
Sader, S. A., Ahl, D., & Liou, W. S. (1995). Accuracy of Landsat-TM and GIS rule-based
methods for forest wetland classification in Maine. Remote Sensing of the
Environment, 53, 133-144.
Salem, F., & Kafatos, M. (2001). Hyperspectral image analysis for oil spill mitigation. In
Proceedings of the Asian Conference on Remote Sensing: Vol. 1. (pp. 748-753).
Singapore: National University of Singapore, Center for Remote Imaging.
Schiewe, J., Tufte, L., & Ehlers, M. (2001). Potential and problems of multi-scale
segmentation methods in remote sensing. Retrieved February 22, 2003, from
http://www.definiens-imaging.com/documents/publications/GIS200106034. pdf
Schmidt, K. S., & Skidmore, A. K. (2003). Spectral discrimination of vegetation types in

a coastal wetland. Remote Sensing of the Environment, 85, 92-108.
Shaikh, M., Green, D., & Cross, H. (2001). A remote sensing approach to determine
environmental flows for wetlands of the Lower Darling River, New South Wales,
Australia. International Journal ofRemote Sensing, 22(9), 1737-1751.
Shepherd, I., Wilkinson, G., & Thompson, J. (1999). Monitoring surface water storage in
the north Kent marshes using Landsat TM images. International Journal of
Remote Sensing, 21(9), 1843-1865.
Sierra Club (2000). Report: Wetlands restoration in waiting. Retrieved October 13,
2003, from
http://www.sierraclub.org/wetlands/reports/wetland_restoraation/iowa.asp
Sugumaran, R., Pavuluri, M. K., & Zerr, D. (2003). The use of high-resolution imagery
for identification of urban climax forest species using traditional and rule-based
classification approach. IEEE Transactions on Geoscience and Remote Sensing,
41(9), 1933-1939.
Takeuchi, R., Tamura, M., & Yasuoka, Y. (2003). Estimation of methane emission from
a west Siberian wetland by scaling technique between NOAA A VHRR and SPOT
HRV. Remote Sensing of the Environment, 85, 21-29.
Toyra, J., Pietroniro, A., & Martz, L. W. (2001). Multi-sensor hydrologic assessment of a
freshwater wetland. Remote Sensing of the Environment, 75, 162-173.
74
Trettin, C. C., Song, B., Jurgensen, M. F., & Li, C. (2001). Existing soil carbon models
do not apply to forested wetlands (USDA Forest Service General Technical
Report SRS-46). Washington, DC: U.S. Government Printing Office.
Tsihrintzis, V. A., John, D. L., & Tremblay, P. J. (1998). Hydrodynamic modeling of

wetlands for flood detention. Water Resource Management, 12, 251-269.
Underwood, E., Ustin, S., & DiPietro, D. (2003). Mapping non-native plants using
hyperspectral imagery. Remote Sensing of the Environment, 86, 150-161.
United States Fish and Wildlife Service. (1996). National list ofplant species that occur
in wetlands. Retrieved May 2, 2003, from http://www.nwi.fws.gov/bha
United States Geological Survey. (2004). Accuracy assessment of 1992 national land
cover data, Region 7. Retrieved March 21, 2004, from
http://landcover.usgs.gov/accuracy/pdf/region7.pdf
van der Sande, C., de Jong, S., & de Roo, A. (2004): A segmentation and classification
approach ofIKONOS-2 imagery for land cover mapping to assist flood risk and
flood damage assessment. ISPRS Journal of Photogrammetry & Remote Sensing,
58, 217-229.
Wakeley, J. S. (1988). A method to create simplified versions of existing habitat

suitability (HIS) models. Environmental Management, 12(1), 79-83.
Wang, N., & Mitsch, W. J. (2000). A detailed ecosystem model of phosphorus dynamics
in created riparian wetlands. Ecological Modelling, 126, 101-130.
Whittecar, G. R., & Daniels, W. L. (1999). Use of hydrogeomorphic concepts to design

created wetlands in southeastern Virginia. Geomorphology, 31, 355-371.
Yasouka, Y., Yamagata, Y., Tamura, M., Sugita, M., Pornprasertchai, J., Polngam, S., et
al. (1995). Monitoring of wetland vegetation distribution and change by using
microwave sensor data. Retrieved November 29, 2003, from
http://www.gisdevelopment.net/aars/acrs/1995/ts 1/ts 1001 pf.htm
75
APPENDIX
MAPS
76
2002 CIR Unsupervised Classification
0 8,000 16 ,000 32 ,000 Meter s
Class_Names - Mixed_Forest
- Fallaw/Bare Soil - Mi xed_Grasse s

Artificial_Surface - Open_Water
D Wetland
Map l: 2002 CIR Unsupervised Classification , Bl ack Hawk County

77
2002 CIR Supervised Classification
0 8,000 16,000 32,000 Meters
Clas s_Names ~ Mi xed_Forest

~ Fallaw/Bare Soil ~ Mixed_Grasses
Artificial_Surface ~ Open_Water
Wetland
Map 2: 2002 CIR Supervised Classification, Black Hawk County

78
2002 CIR Object-Oriented Classification
Land Cover type 1111 Herbaceous Wetlands ~t--·.........---·......

ut ___:-+-
.·1---..-.....-...---40.2 Miles
Ill Open Water 1111 Urban

~ Bare Rock/Sand 1111 Fallaw/Bare Fie lds
Ill Woody Wetland 1111 Mixed Grasses s
1111 Mixed Forest
Map 3: 2002 CIR Object-Oriented Classification, Black Hawk County

(Subset image, near Gilbertville)
79
2000 Landsat Unsupervised Classification
0 6,250 12,500 25,000 Meters
Class_Names 1111 Open_Water

1111 Mixed_Forest 1111 Row_Crop
1111 Mixed_Grasses C=:J Artificial_Surface
C=:J Wetland
Map 4: 2000 Landsat Unsupervised Classification, Black Hawk County

80
2000 Landsat Supervised Classification
0 7.000 14.000 28.000 Meters
Class_Names . . Row_Crop
. . Mixed_Forest Artificial Surface
. . Mixed_Grasses . . Open_Water
s
CJ Wetland
Map 5: 2000 Landsat Supervised Classification, Black Hawk County

81
2000 Landsat Object-Oriented Classification
0 6,500 13,000 26,000 Meters
Class_Names Ill Artificial_Surface N
Ill Mixed_Grasses Ill Row_Crop

Ill Mixed_Forest c=J Wetland s
Ill Open_Water
Map 6: 2000 Landsat Object-Oriented Classification, Black Hawk County

82
1999 Landsat Object-Oriented Classification
0 3.5 7 14 Miles
Class_Names C:J Wetlands

1111 Row_Crop/Mixed_Grasses C:=J Artificial_Surface s
1111 Open_Water 1111 Mixed_Forest
Map 7: 1999 Landsat Object-Oriented Classification, Black Hawk County

83
Object-Oriented Classifier Spectral Angle Mapper Classifier
0 40 BO 160 Meter s
-~,-'
LEGEND - Upland Crop
1111 Open Water (Wetland) - Flooded Fore st (Wetland)
- Floodplain Crop - Shadow

University of Northern Iow a
- Mixed Fore st - Aqua ti c Vegetat ion (Wetland) Dept. of Geography
January 200 4
c=J Artificial Su rfa ce - Herbaceo us
Map 8: 2001 CASI Object-Oriented Classification (left), Eddyville, lA

2001 CASI Spectral Angle Mapper Classification (right), Eddyville, [A
(16 acre subset of 50 acres)
84
Black Hawk County Roads
D 9,000 18,000 36,000 Meters
- Bla ck Hawk Co unty Polygon

Black Hawk County Road s
•-
Map 9: Bl ac k Hawk County Roads Layer Used in GIS-Based Model

85
Black Hawk County

Conservation Areas
0 9,000 18,000 36,000 Meters
- Black Hawk Coun ty Polygon

[==:J Black Hawk Coun ty Conservation Areas
·•·
Map 10: Black Hawk County Conservation Areas Layer Used in GIS-Based Model
86
Soil Survey Geographic

Database (SSU RGO)
Black Hawk County, IA
0 8,750 17,500 35,000 Meters
- SSU RGO BH C •·
Map 11: SSURGO Soil Coverage Layer Used in GIS-Based Model

87
Existing Wetland Areas 1983/84

0 8,750 17,500 35,000 Meters
- Black Hawk County Polygon

Exi sting Wetland Areas
-• -
Map 12: NWI Wetl and Areas Layer Used in GIS -Based Restoration Model
88
Urban Areas and County Roads Mask
-~~e':i~•···········-·=·1
.a•i ....:.:• !1•••••••••• 1111•111111
1111!" ~, . . . . . . . . . . . . . . . . . .~.
U••·····•• 1111!111
~' 11•• "•• n••••• ••• n
~-~'lli• ••···•-•.1•~1
!:i!'(]II -
iilll111• 1iil
••r~11:, ~,-~..c•••••11••••••1.1
tnJ~ .•_,
--IQ
--- ~
Ill~a11,. -
~ ~'
-, ~~-• ·-····
•lwai~---·-1 ..,...11.11•--·-·
-~• -Ila;,... ,, .
.,• - •• 111•• 1 "Ifill - •• l! •• -
1111.• •-»• ;)111·~...
:1••~!-•11••···-·~•·-··1
.iiiiiiii.::.:a:'. • -•.,....,.•••-
,,.-1-.. ·-----·--
'- ••------··
• -•1r~•· ,~. .• ,...~
r-1-• '!"
1•11111
I
•lf,1·-111111·~~1l =-••··
1• - • ll• ••'-°111.~Wlil •--
• a1-• II
1 • ••=1~1111111
11
~• . ,a-•-
••::••··••11••··••t1"_.,..1,_••·
•• I . . . . . . . . :.
•••••••••11•••• • • a...,.,....:rl~
• • ••••• •• ll• ••••• B1j~lL..a~
0 9,000 18,000 36 ,000 Met ers

r - - - + - --+-------+----+-----+--+-------1-~
- Bla ck Hawk Coun ty Polygon ·• ·
Map 13 : Landcover and Roads Buffer Mask Layer Used in GIS-Based Restoration
Model
89
Low Suitability Potential

Wetland Restoration Areas
D 9,000 18,000 36 ,000 Meter s
- Black Hawk Co unty Polyg on

·•
Map 14: Low Suitability Potential Wetland Restoration Areas identified by GIS-Based
Model
90
Medium Suitability Potential

0 9,000 18,000 36,000 Meters
- Black Hawk Co unty Polygon

·• -
Map 15: Medium Suitability Potential Wetland Restoration Areas identified by GIS-
Based Model
91
High Suitability Potential

0 9,000 18,000 36,000 Meters
- Black Hawk County Polygon

·•· •
Map 16: High Suitability Potential Wetland Restoration Areas identified by GIS-Based
Model

Harken, James

Uploaded by

Copyright:

Available Formats

Harken, James

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

Harken, James

Uploaded by

Copyright:

Available Formats

MAPPING WETLANDS AND POTENTIAL WETLAND RESTORATION AREAS

IN BLACK HAWK COUNTY, IOWA USING

AND A GIS-BASED MODEL

of the Requirements for the Degree

University of Northern Iowa

provide many benefits, including: floodwater retention, non-point pollution treatment,

expensive. Traditional pixel-based automated classifiers of remotely-sensed imagery

have also proven to be inaccurate in classifying wetlands because of spectral confusion.

based (ISODATA and Maximum-Likelihood) classifiers both to see if wetland

classification accuracies from remotely-sensed imagery can be increased and to produce

used with an object-oriented classifier and a hyperspectral pixel-based (Spectral Angle

Mapper or SAM) classifier. A GIS-based wetland restoration model is developed to

identify potential wetland restoration sites in Black Hawk County.

Likelihood, SAM) in both multispectral, hybrid-multispectral, and hyperspectral imagery.

The summer/fall seasonal matrix produced unacceptable accuracies. Wetlands in Black

9 .geog. uni. edu/wetland.

IN BLACK HA WK COUNTY, IOWA USING

AND A GIS-BASED MODEL

of the Requirements for the Degree

University of Northern Iowa

This study by: James Harken

Entitled: Mapping Wetlands and Potential Wetland Restoration Areas in Black

has been approved as meeting the thesis requirement for the

Degree of Master of Arts

'-//iJ /(J '(

I would like to express my sincere appreciation to the Iowa Space Grant

LIST OF FIGURES .......................... ......... ....................... ............ .............................. Vll

The Importance and Status of Wetlands.......................................................... 3

Goal and Objectives......................................................................................... 7

Traditional M ultispectral Classification of Wetlands.................................... 10

Hyperspectral Classification of Wetlands........................................................ 13

Object-Oriented Classification of Wetlands.................................................... 16

Wetland Restoration Models Using G IS.......................................................... 18

CHAPTER III. METHODS........................................................................................ 23

Black Hawk County Multispectral Images .......................................... . 23

Eddyville Hyperspectral Image ........................................................... .. 25

Image Analysis and Classification ................................................................. .. 27

Object-Oriented Classification ............................................................ . 28

Multispectral CIR Image Classification .............................................. . 30

Multispectral Landsat ETM Image Classification .............................. . 31

Hyperspectral CASI Image Classification........................................... 33

GIS-Based Restoration Model......................................................................... 35

CHAPTER IV. RESULTS AND DISCUSSION......................................................... 39

Multispectral CIR Classification...................................................................... 39

Hyperspectral CASI Classification.................................................................. 48

Overall Classifier Comparison......................................................................... 53

GIS-Based Restoration Model Results............................................................ 55

Wetlands Change in Black Hawk County....................................................... 59

Web-Based Data Dissemination...................................................................... 60

Recommendations for Future Research........................................................... 66

1 Error Matrix CIR Supervised Classification..................................................... 42

2 Accuracy Percentages CIR Supervised Classification...................................... 42

3 Error Matrix CIR Unsupervised Classification................................................. 42

4 Accuracy Percentages CIR Unsupervised Classification ................................. 43

5 Error Matrix CIR Object-Oriented Classification (GPS ground truth points).. 43

6 Accuracy Percentages CIR Object-Oriented Classification .................. ........... 43

7 Error Matrix Landsat ETM 15 m Supervised Classification............................ 45

8 Accuracy Percentages Landsat ETM 15 m Supervised Classification .. .. .... .... . 46

9 Error Matrix Landsat ETM 15 m Unsupervised Classification........................ 46