.w N33; “anew.
. a ..

Iv

2'".

u
.
:1” s

.u p,
. .
.=£ ' .

n: 59?
ycavm

.. \m. .

 

C. L .i
in”... in v.
1.1.“.
I

 

(-th g
i4
q
to

3

.- nag

This is to certify that the

dissertation entitled

A CLASSIFICATION-BASED ASSESSMENT OF THE
OPTIMAL SPATIAL AND SPECTRAL RESOLUTION
OF COASTAL WETLAND IMAGERY

presented by

Brian L. Becker

has been accepted towards fulfillment
of the requirements for

Ph . D . degree in Geography

 

 

Almak

 

\ Major professor

8\\L0L

 

MS U is an Affirmative Action/Equal Opportunity Institution 0~ 12771

 

LIBRARY
Michigan State
University

 

 

 

PLACE IN RETURN Box to remove this checkout from your record.
To AVOID FINES return on or before date due.
MAY BE RECALLED with earlier due date if requested.

 

DATE DUE DATE DUE DATE DUE

 

 

$212005
9* HM:

 

 

 

 

 

 

 

 

 

 

 

 

 

6/01 eJCIRC/Dateouopss-sz

A CLASSIFICATION-BASED ASSESSMENT OF THE OPTIMAL SPATIAL
AND SPECTRAL RESOLUTION OF COASTAL WETLAND IMAGERY

By

Brian L. Becker

A DISSERTATION

Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of

DOCTOR OF PHILOSOPHY

Department of Geography

2002

ABSTRACT

A CLASSIFICATION-BASED ASSESSMENT OF THE OPTIMAL SPATIAL
AND SPECTRAL RESOLUTION OF COASTAL WETLAND IMAGERY

By

Brian L. Becker

Great Lakes wetlands are increasingly being recognized as vital ecosystem
components that provide valuable functions such as sediment retention,
wildlife habitat, and nutrient removal. Aerial photography has traditionally
provided a cost effective means to inventory and monitor coastal wetlands, but
is limited by its broad spectral sensitivity and non-digital format. Airborne
sensor advancements have now made the acquisition of digital imagery with

high spatial and spectral resolution a reality.

In this investigation, we selected two Lake Huron coastal wetlands, each
from a distinct eco—region, over which, digital, airborne imagery (AISA or
CASI-II) was acquired. The l-meter images contain approximately twenty, 10-
nanometer-wide spectral bands strategically located throughout the visible and
near-infrared. The 4-meter hyperspectral imagery contains 48 contiguous
bands across the visible and short-wavelength near-infrared. Extensive, in-situ
reflectance spectra (SE-590) and sub-meter GPS locations were acquired for the

dominant botanical and substrate classes field-delineated at each location.

Normalized in-situ spectral signatures were subjected to Principal
Components and 2nd Derivative analyses in order to identify the most
botanically explanative image bands. Three image—based investigations were
implemented in order to evaluate the ability of three classification algorithms
(ISODATA, Spectral Angle Mapper and Maximum-Likelihood) to differentiate
botanical regions-of-interest. Two additional investigations were completed in
order to assess classification changes associated with the independent

manipulation of both spatial and spectral resolution.

Of the three algorithms tested, the Maximum-Likelihood classifier best
differentiated (89%) the regions-of-interest in both study sites. Covariance-
based PCA rotation consistently enhanced the performance of the Maximum-
Likelihood classifier. Seven non-overlapping bands (425.4, 514.9, 560.1, 685.5,
731.5, 812.3 and 916.7 nanometers) were identified that represented the best
performing bands with respect to classification performance. A spatial
resolution of 2 meters or less was determined to be the as being most
appropriate in Great Lakes coastal wetland environments. This research
represents the first step in evaluating the effectiveness of applying high-
resolution, narrow-band imagery to the detailed mapping of coastal wetlands

in the Great Lakes region.

to the memory of my grandfathers, for they instilled my appreciation for the
natural world and the game species I passionately pursue. I must also dedicate
this work to my twins, Brianna and Brendin, for unknowingly being my
inspiration and motivation.

iv

ACKNOWLEDGMENTS

I would like to express my appreciation to the three members of my
research committee. Special thanks also goes to my committee chairperson Dr.
David Lusch who has been my supervisor, research partner, and friend.
Without his guidance and support, the completion of this body or work would
have been doubtful. Thanks for statistical help must also go out to Dr. Bruce
Pigozzi. I would also like to thank Robb Macleod at the Midwestern Ducks
Unlimited office, for his support greatly contributed to the success and end-user
applicability of this research. The Great Lakes Fisheries Commission must also

be thanked for their support of my research efforts.

I would like to thank my Parents for their mental and financial support
during my graduate school endeavors. Finally, special thanks must also go out
to my wife Julie, for she supported my Ph.D. endeavors and allowed me to

avoid ”regular” employment just a little longer.

 

TABLE OF CONTENTS

LIST OF TABLES ................................................................................. ix
LIST OF FIGURES ................................................................................ x

LIST OF TERMS ................................................................................. xi
CHAPTER 1: INTRODUCTION .............................................................. 1
1.1 Great Lakes Coastal Wetlands ................................................. 1

1.2 Wetland Remote Sensing ......................................................... 5
1.2.1 Biophysical Relationships ............................................. 6
1.2.2 Aerial Photography of Wetlands ................................... 9

1.2.3 Digital Wetland Imagery ........................................... 12

1.2.4 Satellite Remote Sensing of Wetlands ........................... 13

1.2.5 Airborne Hyperspectral Systems ................................. 14

1.2.6 Hyperspectral Wetland Applications ........................... 20

1.3 Study Sites ......................................................................... 21

1.3.1 Les Cheneaux Island Complex ..................................... 23

1.3.2 Wildfowl Bay Island Complex .................................... 26

1.4 Statement of Problem ........................................................... 28

CHAPTER 2: A Comparison of Image Classzfication Strategies to Optimize Wetland

Botanical Differentiation ......................................................................... 32
2.1 Introduction and Rationale .................................................... 32
2.2 Methods ............................................................................ 35

2.2.1 Airborne Image Acquisition ....................................... 35

2.2.2 Image Geo-rectification .............................................. 37

2.2.3 Imagery Percent Reflectance Transformation ................. 42

2.2.4 Imagery Principal Components Rotation ...................... 46

2.2.5 Accuracy Assessment Regions-of—Interest ..................... 49

2.2.6 Image Classifications ................................................. 52

2.3 Results and Discussion ......................................................... 60
2.3.1 ISODATA - Investigation 1 ......................................... 61

2.3.2 Maximum-Likelihood - Investigation 1 ......................... 64

2.3.3 Spectral Angle Mapper - Investigation 1 ....................... 66

2.3.4 PC-rotated Classifications - Investigation 2 .................... 69

2.3.5 Prentiss Bay Classifications - Investigation 3 .................. 72

2.4 Summary and Conclusions ................................................. 74

vi

CHAPTER 3: Principal components and second derivative analyses of hyperspectral

reflectance data from freshwater coastal wetlands ............................................. 79
3.1 Introduction and Rationale ................................................... 79
3.2 Methods ............................................................................ 83
3.2.1 Spectral Data Pre-Processing ...................................... 83
3.2.2 2“d Derivative Approximations .................................. 90
3.2.3 Principal Components Analysis .................................. 93
3.2.4 Image classifications — Bandset Optimization... .. .... . ........98
3.2.5 Image classifications — Bandwidth Optimization ........... 104
3.2.6 Image classifications - IKONOS Spectral Simulation. ....106
3.3 Results and Discussion ........................................................ 108
3.3.1 Representative Spectral Clusters ................................. 108
3.3.2 2ND derivative Summary .......................................... 112
3.3.3 PCA Loading Curve Inspection ................................. 119
3.3.4 Loading Curve Derivation ........................................ 122
3.3.5 Correlation-based, PCA Factor Interpretation ............... 124
3.3.6 Covariance-based, PCA Dimension Interpretation. . . . .....130
3.3.7 Bandset Optimization - Investigation 4 ....................... 134
3.4 Summary and Conclusions ................................................ 144
CHAPTER 4: A Classification-based Assessment of the Optimal Spatial Resolution
of Coastal Wetland Imagery ................................................................... 149
4.1 Introduction and Rationale ................................................ 149
4.2 Methods ........................................................................ 151
4.2.1 Image Spatial Degradation — Investigation 5 ................ 151
4.3 Results and Discussion ..................................................... 157
4.3.1 Spectral Angle Mapper - Investigation 5 ..................... 157
4.3.2 Mahalanobis Classifications - Investigation 5 ............... 161
4.4 Summary and Conclusions ................................................ 166
CHAPTER 5: CONCLUSIONS AND FUTURE RESEARCH .................... 167
5.1 Classifier Performance and Transferability .............................. 167
5.2 Optimal Spectral Resolution ................................................. 170
5.3 Optimal Spatial Resolution ................................................... 174
5.4 Limitations of the Technology .............................................. 175
5.5 Recommendations for Future Research .................................. 178
APPENDIX A ................................................................................... 180
APPENDIX B .................................................................................... 184
APPENDIX C ................................................................................... 201

vii

APPENDIX D ................................................................................... 214

APPENDIX E ................................................................................... 218

BIBLIOGRAPHY ............................................................................... 225

viii

Table 1.1.
Table 2.1.
Table 2.2.
Table 2.3.
Table 2.4.
Table 2.5.
Table 2.6.
Table 2.7.
Table 3.1.
Table 3.2.
Table 3.3.
Table 3.4.
Table 3.5.
Table 4.1.
Table 4.2.

LIST OF TABLES

Instrument Spectral Sensitivities and Spatial Resolutions ........... 15
Instrument Specifications and Flight Parameters ...................... 38
Instrument Bandsets with FWHM ......................................... 39
Image Rectification Summary ............................................... 42
Image Percent Reflectance Methodology ................................. 47
Prentiss Bay Regions-of-Interest ............................................ 52
Horseshoe Bay, 1-meter Regions-of—Interest ............................. 53
Classification Accuracy Summary .......................................... 63
Collected In-situ Spectral Signatures ....................................... 85
Horseshoe Bay, 4-meter Regions-of-Interest... . . . . . . . . . . . . . . ....101
Bandwidth Manipulation FWHM Parameters ......................... 106
In—situ 2nd Derivative Summary ........................................... 115
Summary of Investigation-4 Classifications ............................ 135
Summary Table, SAM Spatial Degradation ............................. 158

Summary Table, Mahalanobis-Distance Spatial Degradation ...... 162

ix

LIST OF FIGURES

Figure 1.1. Five Characteristic Spectral Signatures ..................................... 8
Figure 1.2. Synthetically Resampled Spectral Signatures ............................. 17
Figure 1.3. Lake Huron Ecoregions ........................................................ 23
Figure 1.4. Prentiss Bay Image Area and Vicinity ...................................... 25
Figure 1.5. Horseshoe Bay Image Area (1 and 4 meter) and Vicinity. . . . .27
Figure 2.1. Horseshoe Bay; Dimensions 1-6 PCA Images ............................ 50
Figure 2.2. Isodata Classification Parameters ............................................ 56
Figure 2.3. Nine Representative Spectra .................................................. 66
Figure 2.4. PC-transformed Spectral Signatures ......................................... 71
Figure 3.1. Representative Spectral Signatures ......................................... 90
Figure 3.2. Derivative Calculation Schematic ............................................ 92
Figure 3.3. 4-meter, Horseshoe Bay Rectangular Image Subset ................... 102
Figure 3.4. Nine Representative Spectra ................................................ 109
Figure 3.5. Representative 2nd Derivative Plots ....................................... 113
Figure 3.6. Frequency of 2“d Derivative Occurrences ............................... 118
Figure 3.7. Key Wavelength Domains - Dimensional PCA Loadings ........... 122
Figure 3.8. Frequency of 2nd-Derivative Occurrences of Loading Curves. .....124
Figure 3.9. Correlation-based, PCA Component Loading Curves ............... 126
Figure 3.10. Covariance-based, PCA Component Loading Curves ............... 131
Figure 4.1. Spatial Manipulation, Image Processing Schematic ................... 155
Figure 5.1. Horseshoe Bay; PC654321 Max-Likelihood Classification 1.76
Figure E-1. Prentiss Bay; False Color Composite ...................................... 219
Figure E-2. Horseshoe Bay; True Color Composite and R018 .................... 220
Figure E-3. Horseshoe Bay; False Color Composite .................................. 221
Figure E-4. Prentiss Bay PCA Image; 4, 3, 2 Color Composite ..................... 222
Figure E-5. Horseshoe Bay PCA Image; 4, 3, 2 Color Composite ................. 223
Figure E-6. Pixel Level Spatial Degradation Patterns ................................. 224

LIST OF TERMS

2nd Derivative Approximation - a slope-based method used to differentiate 15*-
order polynomials that would otherwise not yield higher-order derivatives.

AISA - A hyperspectral sensor deployed by 3DI, Inc. ; Airborne Imaging
Spectrometer for Applications

ARC/INFO - GIS software developed by Environmental Systems Research Inc.
(ESRI)-

ARCVIEW - Desktop GIS software developed by Environmental Systems
Research Inc. (ESRI).

CASI-II - a hyperspectral sensor manufactured and deployed by ITRES Inc. ;
Compact Airborne Spectrometric Imager.

ENVI - The Environment for Visualizing Imagery; image processing software
developed by Research Systems, Inc., a KODAK company.

ERDAS - The Earth Resources Data Analysis System; image processing
software developed by ERDAS, Inc., a Leica Geosystems company.

Hyperspectral Sensors - imaging spectrometers capable of recording spectral
data similar in nature to laboratory spectrometers, making it possible to
construct spectral reflectance curves.

IKONOS - a commercial earth resource satellite remote sensing system
developed by a US. private company designed to provide high spatial
resolution imagery.

Image Classification - the process of establishing a link between an
informational category of interest and a related image spectral class.

Landsat - A series of satellite remote sensing platforms (NASA) carrying
instruments specifically designed to acquire data about earth resources on a

systematic, repetitive, medium resolution, multispectral basis.

Near-infrared Spectrum - that portion of the electromagnetic spectrum
containing wavelengths between 700nm and 1200nm.

xi

NIR Plateau - the longer-wavelength spectral region adjacent to the red edge in
which the NIR reflectance of foliar materials levels out at or near its maximum
reflectance.

Pixels - the basic building blocks of digital images; picture elements.

Principal Components Analysis - a statistical transformation tool used to
generate orthogonal channels or components that are more interpretable and
dimensionally reduced with respect to the non-transformed data.

Red edge - the steep reflectance slope between high pigment absorption in the
red-light zone and high cellular reflectance in the near-infrared zone.

Spectroradiometer - a laboratory or hand-held instrument that acquires light
reflectance data in very narrow, contiguous bands across the visible and
infrared portions of the electromagnetic spectrum.

SPOT - a commercial earth resource satellite remote sensing system (Systeme
Pour l’Observation de la Terre) conceived and designed by the French Centre
National d’Etudes Spatiales (CN ES).

Supervised Classification - classification process in which informational
categories are defined, labeled, and mapped when the spectral characteristics
and / or location of informational classes are known prior to image classification
through a combination of fieldwork, image analysis, and personal experience.

Unsupervised Classification - classification process in which informational
categories are defined, labeled, and mapped with no extensive prior knowledge

of the informational categories found in an image.

Visible Spectrum - that portion of the electromagnetic spectrum containing
wavelengths between 400nm and 700nm.

xii

Chapter 1: INTRODUCTION

1.1 Great Lakes Coastal Wetlands

Estimates indicate that 215 million acres of wetland habitat existed within
the borders of the contiguous United States prior to European settlement (Roe et
al., 1954; p. 501). Approximately 53% of this original acreage has been lost due
to direct and indirect impact by humans (Dahl et al., 1991). Long thought of as
wastelands, wetlands are now recognized as being vital to the health and
stability of our ecosystems. The protection, restoration, and management of our
nation’s remaining wetlands are critical, because they play a vital role in energy
fixation, nutrient assimilation, geo-chemical cycling, sediment stabilization, and
wildlife habitat (Hardisky et al., 1986). Despite current progress in wetland
protection and conservation, there are concerns that these efforts will fall far
short of what is needed to maintain valued wetland functions (Zelazny and
Feierabend, 1988). Freshwater wetlands are of added concern, for they
represented ninety-eight percent of wetland losses that occurred between 1986

and 1997 (Dahl, 2000; p, 10).

In general, coastal wetlands can be described as ecotones, areas of transition
between terrestrial (upland) and open water habitats (Cowardin et al., 1979;

Chadde 1998; p. 2). Inherent to this landscape position is hydrologic

interaction with the physical, chemical, and biological components of the
landscape. The presence of surface or near-surface water is a fundamental
component to all wetlands, that creates the unique physiochemical and
biological conditions that set wetlands apart from both terrestrial and open

water environments (Mitsch and Gosselink, 2000; p. 107).

The Great Lakes contain approximately one fifth of the world’s fresh water
and, along with their connecting waterways, have approximately 15,000
kilometers of shoreline (Prince et al., 1992). There are approximately 1,200 km2
of wetlands along the US. coastline alone, which represents only a small
portion (approximately 30%) of the once expansive coastal marshes found in
this area at the time of European settlement (Cardinale et al., 1998). The most
common type of wetland found along the shorelines of the Great Lakes are
coastal marshes, because the non-woody vegetation that typifies these areas
best tolerates long-term and short-term water level fluctuations (Keddy and
Reznicek, 1986). The remaining coastal marshes have largely been degraded by
a multitude of environmental stressors, the majority of which are directly
linked to the level of residential, industrial, and agricultural development in
this region (Comer et al., 1995). Many Great Lakes coastal marshes are
managed and protected by artificial dikes that serve to protect them from
excessive flooding and wave action, as well as allow active water level

management (Herdendorf, 1987).

Despite the reduced distribution and degraded nature of coastal Great
Lakes wetlands, they are widely recognized as areas of concentrated
biodiversity and productivity (Krieger, 1992). Great Lake wetlands represent,
almost without exception, the most photosynthetically productive areas along
the gradient from upland to open water (Wetzel, 1992). Wetlands are a critical
component of the Great Lakes ecosystem, ultimately serving a wealth of
physical, chemical, and biological functions that contribute to the region's
environmental health. The annual economic value of Michigan’s coastal
wetlands, has been estimated at $51.8 million with respect to fisheries, wildlife,
and recreation (Jaworski and Raphael, 1979). This estimate was undoubtedly
conservative, since it did not incorporate the value of less tangible functions
such as flood mitigation, nutrient assimilation, and sediment stabilization. The
protection and proper management of the remaining coastal wetlands are
directly linked to the economic and cultural prosperity of the region (Raphael,
1987). Effective protection and restoration strategies for Great Lakes wetlands
require knowledge of the nature of wetland health, loss, and degradation

(Detenbeck et al., 1999).

Coastal Great Lakes wetlands are inherently dynamic, for they occupy the
transitional boundary between dry land and open water. Coastal wetlands
experience physical and environmental perturbations at a higher frequency

than that typical of upland and open water habitats as a result of their

3

landscape position and exposure to large-lake processes (Maynard and Wilcox,
1997; p. 17). Coastal zone stressors, such as water-level fluctuations, currents,
storms, ice, sedimentation, and erosion create an unstable and highly variable
environment to which wetland plants have uniquely adapted (Keough et al.,
1999). Wetland botanical diversity is reflective of, and dependent upon, a host
of periodic stressors, especially those associated with water-level fluctuations

(Keddy and Reznicek, 1986).

Hydrology is the single most important environmental determinant for
establishing and maintaining wetland diversity and function, and hydrologic
variability is the norm rather than the exception in most wetlands (Mitsch and
Gosselink, 2000; p. 250). Individual plant species and plant assemblages
display affinities and physiological adaptations for certain water-depth
changes, the end result being a plant community that is continuously in a state
of flux (Maynard and Wilcox, 1986; p. 27). Human and natural stressors
continuously alter and reshape the mosaic of coastal wetland plant species,
resulting in high spatial, temporal, and spatial variability. Different species are
often distributed in rough zones that are determined by and distributed along
coastal hydrologic gradients (Mitsch and Gosselink, 2000; p. 250). It is generally
believed that spatially heterogeneous coastal marshes, those with a highly
interspersed mosaic of plant species and open water areas, support a more

diverse biotic community that homogeneous marshes (Gottgens et al., 1998;

4

Mitsch and Gosselink, 2000; p. 250). The dynamic nature of coastal
environments and the inherent link of spatial heterogeneity to biotic diversity

demands accurate and current wetland characterizations and inventories.

1.2 Wetland Remote Sensing

Since the vegetative community within a coastal wetland is shaped by
environmental stressors, it follows that changes within a plant community
would serve as an indicator of such perturbations. The natural and human
induced stressors influencing Great Lakes coastal wetlands occur at various
scales and can be found at the population, community, and landscape levels
(Maynard and Wilcox, 1986; Pp. 36—40). Remote sensing and landscape ecology
are powerful tools that provide a foundation upon which status and health

assessments of Great Lakes coastal wetlands can be based (Wilcox, 1995).

Proper ground-based wetland characterization and inventory can be
extremely difficult, time consuming, and inaccurate (Seher and Tueller, 1973).
Many scientists have demonstrated that the need for cost-effective methods to
inventory and monitor wetland areas necessitates the use of remote sensing
techniques (Best et al., 1981; Carter and Anderson, 1982; Earnst-Dottavio et al.,
1981 ; Gross and Klemas, 1986: Hardisky et al., 1986; Jensen et al., 1984; Oguma,
1996; and Stewart et al., 1980). Wetland remote sensing falls into two broad

categories: vegetative mapping and the estimation of biophysical wetland

5

parameters (Gross et al., 1989; p. 475). Measurement of biophysical parameters
involves the application of remote sensing data to determine characteristics
such as plant biomass or percent cover. Mapping, the category explored in this
research, primarily involves the delineation of wetland boundaries and the

categorization of plant communities.

1.2.1 Biophysical Relationships

 

Energy across portions of the visible and non-visible electromagnetic
spectrum, either emitted by the sun or generated by mankind, strikes and
interacts with the landscape, including wetlands. The relative proportions of
electromagnetic energy that are absorbed, transmitted or reflected will vary for
different wetland cover types as a result of their structure, condition, or stage of
life (Lillesand and Kiefer, 1994; p. 13). The foundation of wetland remote
sensing is based on distinguishing features or cover-types within a wetland
through the analysis of the electromagnetic energy reflected from substrates

and / or vegetation.

Only a small segment of the electromagnetic spectrum is commonly
utilized for wetland remote sensing, the visible and near-infrared portions of
the spectrum. Energy in the blue (400-500nm), green (500-600nm), red (600-

700nm) and near-infrared (700-1200nm) portions of the spectrum fall within the

range of silicon-detector-based optical remote sensing and are most applicable

6

to wetland remote sensing. In general, the energy within these four broad
wavelength domains is reflected by healthy, green vegetation in a characteristic
pattern or spectral reflectance curve exemplified by the Nuphar advena curve
shown in Figure 1.1. The spectral reflectance curves for three non-vegetative
cover types are also shown in Figure 1.1: gravel, dead biomass, and clear water
over a sand substrate. These non-vegetative signatures do not resemble the
characteristic peak and valley configuration of the living vegetation curves.
Vegetation (dead, dying and healthy), soil, and water are the three basic types
of earth features (Lillesand and Kiefer, 1994; p. 17), all of which play a role in

wetland remote sensing.

Three distinct regions exist within a typical vegetative spectral reflectance
curve or spectral signature, the visible (400-700nm), red-edge (700-775nm), and
NIR plateau (775-1200) (Lillesand and Kiefer, 1994; p. 18). Expression of
reflectance as a percentage allows for a more accurate comparison of values
generated under varied conditions. The relative combination of pigments (e.g.
chlorophyll-a) present in live plant tissues largely dictates foliar reflectance in
the visible range of the electromagnetic spectrum (Jensen, 2000; p. 337).
Although most vegetative reflectance studies focus on the spectral nature of
healthy, living plant tissue, spectral differences are present in green (living),
dying, and necrotic (dead) plant materials (Figure 1.1). Dead biomass

reflectance is significant in coastal wetlands because persistent plant species are

7

 

 

  
  
 

L

message

1

% Reflectance

 

 

H
O

 

 

996 ~
1201 .
8801 -

nanometers

Figure 1.1. Five Characteristic Spectral Signatures

found in most coastal wetlands. Persistent vegetation are those plants whose
dead biomass remains standing at least until the beginning of the following
growing season (Cowardin et al., 1979). The relative proportion of green and
dead / dying biomass varies with many wetland plant species, and the detection
of related spectral differences may lead to better plant community

discrimination (Elvidge and Portigal, 1990).

In general, the NIR portion of the spectrum can be divided into two
regions. The region between 700 to 775nm is commonly termed the red edge or
wall due to the marked change in reflectance by vegetation in this area (Figure
1.1). The NIR plateau is the spectral region adjacent to the red edge in which
the NIR reflectance of foliar materials levels out at or near its peak (Figure 1.1).
As NIR energy enters the mesophyll layer of a leaf, the air spaces and cellular
membranes provide a multitude of surfaces through which N IR light is
refracted. This is evident by the significant rise in reflectance in the infrared
portion of the spectrum near 700nm (Figure 1.1). Intracellular and intercellular
moisture content plays a secondary role in suppressing NIR reflectance of

vegetation, because water absorbs nearly all NIR energy (Jensen, 2000; p. 342).

1.2.2 Aerial Photography of Wetlands

The photographic camera is certainly the oldest and most frequently used
remote sensing instrument (Avery and Berlin, 1986; p. 21). Photography taken
from an aircraft (aerial photography) represents the most common form of
remote sensing applied to wetlands and it is well suited for general mapping

strategies.

Aerial photography was commonly used as early as the 19503 as an upland

land cover evaluation tool for natural vegetation and agricultural crops (Butera,

1983). Wetlands could be differentiated from similarly structured upland
vegetation via infrared photography because hydrophytes tend to reflect higher
levels of N IR energy, which prompted state and federal agencies to increasingly
look towards aerial photography for the purposes of wetland mapping and

inventory (Gross et al., 1989; p. 475).

Film types that are sensitive to NIR electromagnetic energy are generally
preferred for wetland applications because they are more effective in detecting
standing water and determining biomass, as well as differentiating wetland
from upland vegetation (Hardisky et al., 1986). Black-and-white films are
somewhat limited because the human eye is incapable of differentiating the full
range of gray tones found in these panchromatic photographs (Lillesand and
Kiefer, 1994; p. 79). The human eye is much more effective at differentiating the
many hues on color films. Color infrared film (CIR) has emerged as the
preferred wetland photographic media since it takes advantage of both the
sensitivity to NIR and the increased tonal differentiation of color photography
(Hardisky et al., 1986). Increased tonal differentiation is invaluable with respect
to the botanical complexity and physical variation found in many coastal
wetland areas. A host of state and federal agencies utilize CIR photography as

the foundation of their wetland mapping strategies (Carter, 1982).

10

The image resolution of photographic systems is dictated by the physical
size of the photo-reactive silver halide grains (50-100 micrometers), and is
superior to digital remote sensing platforms (Avery and Berlin, 1986; p. 36).
However, the non-digital format of photographic media limits its usefulness
within computerized cartographic and geographic information systems

(Hardisky et al., 1986).

As one would expect, there is less botanical detail in a photograph taken at
20,000 feet (smaller scale) than one taken at 5,000 feet (larger scale). The areal
extent of the wetland to be photographed, the cost of photography, and the
desired botanical detail ultimately determine the desired photographic scale.
Photographs taken at 1:1,000—scale allow extremely accurate vegetation maps to
be produced, but are 10 to 15 times more expensive than the same aerial
coverage at 1210,000-scale (Seher and Tueller, 1973). At scales of approximately
110,000, most large plant communities (25ft2 or 2.3 m2) can be identified and

mapped (Seher and Tueller, 1973).

Aerial photography (1:1,000 to 1215,000 scales) has been found to be a fast,
accurate, and relatively inexpensive method by which wetland boundaries,
botanical changes, and areal extent can be recorded (Mitsch and Gosselink,
2000; p. 107). The two characteristics that limit the applicability of aerial

photography to detailed vegetative mapping strategies are: 1) the limited

11

amount of spectral information (i.e. only three, wide sensitivity ranges)
contained in each image (Jensen et al., 1984), and 2) its non-digital format
(Hardisky et al., 1986). Spectral signatures like those shown in Figure 1.1
cannot be generated, or even approximated, via aerial photographs. If the
assumption is made that more precise spectral reflectance data are needed to
better differentiate wetland plant communities, photographs are not applicable

to such methodologies.

1.2.3 Digital Wetland Imagery

The basic building blocks of digital images are pixels (picture elements),
whose size is determined by both mission altitude and instantaneous field of
View (IFOV). Along-track, digital imaging technologies are based on two-
dimensional photosite arrays (Lillesand and Kiefer, 2000; p. 315) like those
found in today’s hand-held digital cameras. The lst-dimension of such an
array, comprised of a series of end-to-end photosites positioned perpendicular
to the flight line, controls the swath width of the sensor. The 2nd-dimension,
comprised of a series of end-to-end photosites positioned parallel to the flight
line, relates to the spectral resolution of the instrument. Each photosite is a
solid-state sensing device (CCD or CMOS) that responds to electromagnetic
energies striking its surface by generating proportional electronic charges

recorded for each pixel (Lillesand and Kiefer, 2000; p. 112). Arrays with well

12

over 1,000 photosites in the 2nd-dimension are operational today and are able
to simultaneously record very small wavelength regions across much of the
visible and NIR spectrum. Hundreds of 2nd-dimension pixels allow reflectance

curves like those shown in Figure 1.1 to be captured.

Two-dimensional images are generated as the forward motion of the sensor
positions the array so that data can be captured over adjacent portions of the
landscape creating a ribbon-like image. Changing the elevation of a sensor
with respect to a target changes the projected area of each photosite onto the
target (i.e. ground resolution). Airborne platforms can yield a variety of spatial

resolutions, determined by varying mission altitude (Equation 1.1).

Dsp = A* I3 (Eq. 1.1)
DSp = Diameter equal to the spatial resolution (m) or height/width of pixel (m)

A = Mission Altitude (meters)
8 == IFOV (radians)

1.2.4 Satellite Remote Sensingof Wetlands

In the early 19703, the initial launch of the Landsat-1 satellite marked the
practical beginning of spaceborne, earth resource remote sensing (Lillesand and
Kiefer, 2000; p. 373). Over the ensuing thirty years, a variety of systems, with

varied spatial and spectral resolutions, have been developed. Earth resource

13

satellite platforms are in continuous, circular orbits, a vantage point that
provides near-simultaneous digital imagery over large geographic areas.
Satellite platforms continue to provide an overwhelming array of earth resource
data when fitted with remote sensing systems, and the status of these
technologies continues to change as scientific advances yield improved
spacecraft and sensors (Lillesand and Kiefer, 2000; p. 373). Although the rate of
technological change is staggering, all current satellite-based systems display
operational characteristics (Table 1.1) that significantly limit their applicability
to detailed wetland remote sensing strategies. These limitations become
obvious in the context of the two fundamental assumptions of this research: 1)
the spatial heterogeneity of coastal wetlands is best captured by digital imagery
with a spatial resolution at or below 5 meters (ITRES, 2000), and 2) separation of
wetland plant communities is best achieved through the analysis of
hyperspectral reflectance curves. If these assumptions are valid, then aerial
photography and currently available satellite imagery are not applicable to

detailed wetland mapping.

1.2.5 Airborne Hyperspectral Systems

The curves in Figure 1.1 are typical of the data generated by hand-held
Spectroradiometers acquiring data in many, very narrow, contiguous bands
across the visible and infrared portions of the electromagnetic spectrum. Such

Spectroradiometric data have been used to discriminate earth surface features

14

 

Satellite Systems
Landsat(4-5) MSS

 

Landsat TM (4-5)

Landsat-7 ETM+

SPOT

IKONOS

Aerial Photography
B&W Film"

B&W Infrared”
Color Film”

Color Infrared Film“
"‘ = unfiltered

 

Wavelength
500-600nm
600-700nm
700-800nm

800-1100nrn

 

450-520nm
520-600nm
630-690nm
760-900nm
1550-1750nm
1040-1250nm
2080-2350nm

450-520nm
520-600nm
630-690nm
760-900nm
1550-1750nm
1040-1250nrn
2080-2350nm
500-900nrn

500-590nm
610-680nm
790—890nm
510-730nrn

445-516nm
506-595nm
632-698nm
757-853nm
500-900nm

Wavelength

300-720nm
300-960nm
400-700nm
400-900nm

 

Pixel size

79 x 79m
79 x 79m
79 x 79m
79 x 79m

30 x 30m
30 x 30m
30 x 30m
30 x 30m
30 x 30m
30 x 30m
30 x 30m

30 x 30m
30 x 30m
30 x 30m
30 x 30m
30 x 30m
30 x 30m
30 x 30m
15 x 15m

20 x 20m
20 x 20m
20 x 20m
10 x 10m

4x4m
4x4m
4x4m
4x4m
1x1m

Bil—11C!
Band-1
Band-2
Band-3
Band-4

Band-1
Band-2
Band-3
Band—4
Band-5
Band-6
Band-7

Band-1
Band-2
Band-3
Band-4
Band-5
Band-6
Band-7
Mono

Band-1
Band-2
Band-3
Mono

Band-1
Band-2
Band-3
Band-4
Mono

 

Table 1.1.

(Source, Lillesand and Kiefer, 2000)

15

Instrument Spectral Sensitivities and Spatial Resolutions

 

that exhibit diagnostic absorption and reflection patterns over narrow
wavelength regions, patterns that are hidden or blurred in data collected at
coarser spectral resolutions (Lillesand and Kiefer, 2000; p. 363). Hyperspectral
sensors or imaging spectrometers record spectral data similar in nature to
laboratory spectrometers, making it possible to construct spectral reflectance
curves. The characteristic that exemplifies hyperspectral sensors is this inherent
ability to create, or accurately approximate, a continuous spectral reflectance

curve within a spectral domain most appropriate for a target of interest.

The visible and the N IR regions of the spectrum are most applicable to the
spectral differentiation of vegetation. To have utility, a hyperspectral system
would need to capture enough spectral information to effectively approximate
a spectral signature across these regions. Examination of Figure 1.1 reveals
substantial portions of all three vegetation curves in which the slope is
essentially constant, and the data points found between the endpoints defining
these regions of consistency are largely redundant. Therefore, the number of
spectral bands needed to accurately approximate the spectral signature of
vegetation can be reduced without adversely impacting our ability to
differentiate targets based on spectral details. In order to better illustrate this
idea, seven offset, spectral reflectance curves are presented in Figure 1.2. The
original data set (bottom), captured by a hand-held spectroradiometer, is

shown along with four other curves with synthetically reduced spectral

16

 

Resampled Spectral Signatures
"I"'I"'I"'

I I I I I I I I I

ETM+

M /

t

IKONOS _/ CASI-10

AISA-21

 

    

ItrlTlllllllll'IIIIIl

llllllLllllllllllllIlll

 

 

 

 

_ 1 1 1 l 1 1 1 l 1 1 1 I L 1 1 L 1 1 1 l 1 1 1 __..
400 500 600 700 800 900 1000
Wavelength (nm)

Figure 1.2 Synthetically Resampled Spectral Signatures

resolutions (CASI-48, AISA-21, CASI-18, CASI-10). Additionally, the visible
and NIR bands of the IKONOS and LAN DSAT-7 ETM+ instruments are
depicted. Visually, there is little difference between the four curves of highest
Spectral resolution or band number. However, the two curves generated from
satellite sensors poorly emulate the original data set. Much of the data
recorded via the hand-held instrument is redundant, and can be effectively
modeled with a significantly reduced bandset. Reduction in band number is

significant because the constraints of current sensor technologies (i.e.

17

integration time) make it impossible to simultaneously record high spatial and

high spectral detail.

The development of imaging spectrometers brought about increased
interest in utilizing spectral reflectance data of vegetation canopies measured
from the ground to aid in remotely differentiating plant species. The use of
differences in canopy pigmentation and chemistry to classify to the species level
has only recently emerged as a potentially new approach (Zarco—Tejada and
Miller, 1999). Genera- or species-level vegetation differences are most frequent
in the chlorophyll (visible) and water (N IR) absorption regions of the
electromagnetic spectrum (Lillesand and Kiefer, 1994; p. 416). A major
advantage of imaging spectrometry lies in the potential for integrated analysis
comparing calibrated remote sensing imagery with field spectral measurements
(Goetz, 1992; Kruse et al., 1993). A new era of wetland remote sensing began
with the development of hyperspectral sensors capable of potentially exploiting
detailed spectral characteristics as a means to better differentiate vegetation
communities. Hyperspectral sensors provide data with superior spectral
resolution, but spectral resolution alone does not effectively address coastal
wetland remote sensing. In order for remotely sensed imagery to be truly
useful in coastal environments, images must have a spatial resolution of 5
meters or less (ITRES, 2000). Currently, hyperspectral sensors are operational

on several satellite platforms, and many more are expected to become

18

operational over the next decade. Although this is a major step forward in
remote sensing, due to current technological constraints, satellite hyperspectral
data can only be collected at spatial resolutions too coarse for detailed wetland

mapping strategies.

Airborne sensors (photographic and digital) have several characteristics
that make them effective remote sensing systems for wetland studies. Unlike
orbital systems, airborne systems can be tasked when atmospheric and weather
conditions are optimal. In addition, image capture can be timed to correspond
with ecological (e.g. floods), phenological (e.g. flowering), or environmental
(e. g. seiches) conditions. The ability to more easily dictate flight scheduling
allows ground-based efforts, such as the placement of ground control targets, to
be better coordinated with image data collection. Most importantly, airborne
platforms can be flown at relatively low altitudes, a characteristic currently
needed in order to simultaneously collect high spatial and high spectral

resolution data.

Low-altitude, airborne, hyperspectral systems can provide the flexibility,
Spatial resolution, and spectral detail that are ideal for detailed wetland studies.
Airborne hyperspectral sensors, when flown at altitudes appropriate for their
IFOV, can acquire data with spatial resolutions below 1 meter. Such resolutions

are comparable to those of 1:10,000 scale aerial photographs, making airborne

19

hyperspectral images a viable option to aerial photography even without
considering the vastly improved spectral resolutions and digital format
generated by such systems. The combination of hyperspectral sensors and light
aircraft platforms resulted in a powerful tool for coastal wetland studies, one
that provides the spatial, spectral and radiometric resolution needed to

effectively study these environments.

1.2.6 Hyperspectral Wetland Applications

The combination of light aircraft and hyperspectral systems, which capture
superior spectral and spatial resolutions, has been applied to a variety of
remote sensing needs. Coastal applications have included, but are not limited
to: water quality measurements, bathymetric charting, substrate mapping, coral
reef monitoring, oil spill detection, and aquatic / terrestrial vegetation mapping
(ITRES, 2000). Although applications specific to wetlands are not widespread
in the literature, there are several examples of the use of airborne
multi/ hyperspectral sensors in these environments (Borstad Associates, 1995;
Brown and Borstad, 1999; Forsyth et al., 1998; Jupp et al., 1994; Ritter and
Lanzer, 1997; Savastano et al., 1984; Wang et al., 2000; Zacharis et al., 1992).
Common to these investigations is the exploitation of improved spatial or
spectral resolutions to more effectively differentiate wetland vegetation that is
currently unattainable by other means. For example, Wang et a1. (2000)

mapped submergent and emergent macrophytes in Poyang Lake, China to an
20

accuracy estimated at 84 percent utilizing airborne hyperspectral imagery (3.75
meter pixels, 60 bands). This Poyang Lake study demonstrated that multi-
species, botanical associations (e.g. Dense Carex Association, Reed Association)

could be effectively mapped using hyperspectral images.

The wetland studies highlighted above, incorporating hyper / multispectral
airborne sensors, illustrate several pertinent points. First, marine and estuarine
settings represent the vast majority of wetland applications of airborne
hyper/ multispectral strategies currently reported in the literature. Second,
although a limited number of freshwater studies incorporated in-situ
reflectance data to aid in classification, few included detailed analysis of
extensive species-level spectral libraries coupled with very high spatial
resolution imagery (i.e. l-meter pixels). Third, to my knowledge,
multi/ hyperspectral strategies have not been applied to the coastal wetlands of
the Great Lakes. These deficiencies and the current need of many end-users for
improved mapping of Great Lakes coastal resources provided the impetus for

this research.

1.3 Study Sites
Lake Huron is the second largest of the Great Lakes, with an average
surface area of 59,500 km2 and 6,373 km of shoreline (IJC, 1993). In general,

marsh and swamp wetland types are most commonly found along the shores of

21

Lake Huron (Maynard and Wilcox, 1997; p. 52), occupying a variety of
environmental settings ranging from high energy, exposed shorelines to
protected embayments associated with the numerous islands and peninsulas
(Environment Canada, 1994). Complex wetland communities, some of national
as well as global importance (Maynard and Wilcox, 1997; p. 52), are found
along Lake Huron shorelines due to varied disturbance levels, varied hydro-

geomorphology, and the presence of calcareous soils (Smith et al., 1991).

The Lake Huron basin and associated coastal areas, as well as the other four
Great Lakes, have been subdivided into zones or eco-reaches (Figure 1.3) in
which significant concentrations of similar wetland habitats are found (Chow-
Fraser and Albert, 1998). These eco-reaches were delineated via differences in:
climate, geology, hydro-geomorphology, and land use factors, differences that
in combination dictate the development of wetlands with similar characteristics
(Minc, 1997). Two wetland complexes, each from one of the ten distinct eco-
reaches in Lake Huron, were selected as study sites for this investigation. The
first study area is the Les Cheneaux Island complex, falling within the HG3 eco-
reach, found along a stretch of shoreline that has some of the highest wetland
density in northern Lake Huron (Koonce et al., 1999; p. 40). The second study
area is the Wildfowl Bay Island complex, falling within the HG6 eco-reach,

found along the southeastern shoreline of Saginaw Bay, Lake Huron.

22

w...

r‘ *1 :9: 31,275"-

 

f

Figure 1.3. Lake Huron Ecoregions (from Koonce, et. al., 1999)

Les Cheneaux Island Com lex

1.3.1

The Les Cheneaux Islands are found within the most western portion of

located near the Straits of Mackinac, near the town of Hessel,

Lake Huron,

ties along the northern Lake Michigan/ Huron

higan. The marsh comrnuni

Mic

nly found associated with sheltered embayments,

mes are COI'I'IIIIO

shorel

environments providing protection not found along exposed, high-energy

shorelines. An aerial view of the Les Cheneaux Islands reveals a maze of linear

waterways, islands, and peninsulas. In general, the wetlands in this area fall
into two broad types, northern marsh and northern rich fen, albeit with
considerable overlap (Koonce et al., 1999; p. 20). Both typically have vegetative
zonal patterns characteristic of many wetland habitats. Species are generally
distributed along a hydrologic gradient dictated by elevation with wet
meadow/ rich fen communities occupying the highest elevations, transitioning
into shallow then deep emergent zones (Burton et al., 1999). In some instances,
submergent plant species exist within and well beyond the deep emergent
zone. The presence of limestone bedrock at or near the surface is significant
because calcareous wetland soils develop that host plant species (calciphiles)

indicative of such conditions.

The Prentiss Bay wetland (PBW) is located along the eastern boundary of
the island complex, and is situated far back in a deep shoreline incision (Figure
1.4), bounded by Whitefish and Rover points. The embayment formed by
these two peninsulas protects the shallow wetland from most wave energy, but
waves generated by strong southeasterly winds can significantly influence the
wetland. The PBW is relatively unusual in that route M-134 bisects the wetland
parallel to the shoreline, and acts to roughly separate the wet meadow and
emergent zones (Burton et al., 1999). The portion of the wetland north of the
highway is typical of a northern wet meadow dominated by two sedge species,

Carex stricta and Carex Iasiocarpa, although several non-dominant species

24

 

 

 

 

 

Figure 1.4 Prentiss Bay Image Area and Vicinity

indicative of a northern rich fen are interspersed. The shallow and deep
emergent zones are dominated by Typha an gustifolia and Scirpus acu tus,
respectively. Submergent and floating-leaf species are interspersed within the

deep emergent and shallow open water areas of the PBW.

1.3.2 Wildfowl Bay Island Complex

 

The Wildfowl Bay island complex is found along the southeastern shoreline
of Saginaw Bay, Lake Huron. Saginaw Bay is a shallow, gently
slopping glacial lake plain rich in sand and clay. The underlying bedrock is
dolomitic limestone, outcroppings of which form the foundations of many of
the permanent islands in this region. The marsh communities found in
Saginaw Bay are both open and protected lake plain marshes, displaying
botanical characteristics of both northern and southern plant communities

(Koonce et a1, 1999; p. 20).

Horseshoe Bay is a shallow, protected embayment within the Wildfowl Bay
island complex, encircled by Maisou, Middle Grounds, and Heisterman islands
(Figure 1.5). The study site was an elongated rectangle encompassing much of
Horseshoe Bay, portions of the surrounding islands, and a section of the
mainland shoreline near the Geiger Road access site. During the 2000
growing season, due to extremely low water levels in Lake Huron, boat access

to Horseshoe Bay was (at best) restricted to the narrow navigation channels.

26

 

 

 

t”?

1.31)}. I. F E

 

 

inity

Figure 1.5. Horseshoe Bay Image Area (1- and 4-meter) and Vic

27

The shallow embayment wetlands of Saginaw Bay experience dramatic
water-level fluctuations (Maynard and Wilcox, 1997; p. 52). The wetland plant
communities in the study area and elsewhere in Saginaw Bay experience
significant changes in overall vegetative abundance and distribution (Koonce et
al., 1999; p. 20). During the 2000 growing season, the mainland shoreline was
dominated by an expansive Typha angustifolia emergent zone, which was
interspersed with an assemblage of less prevalent emergent genera
(Sagittaria, Alisma, Pontenderia). The submergent community within
Horseshoe Bay and within the shallower portions of the channel separating
the island complex from the mainland was dominated by Vallisneria
americana and Chara sp. A diverse group of plant species was likely present
within Horseshoe Bay proper as a result of the varied topography and
relative protection from disturbance. The higher elevations on the islands
were dominated by upland and wetland graminoides, while the emergent
community was dominated by Scirpus validus, Sagittaria rigida, and Typha
angustifolia. In addition, several floating leaf and less dominant shallow

emergent species were present.

1.4 Statement of Problem

Strategies through which more accurate and detailed wetland maps could
be produced would contribute to continued efforts to protect, monitor, and

restore Great Lakes coastal wetlands. Technological advancements in sensor

28

and computer technologies have led to the development of the next generation
of remote sensors. These hyperspectral systems are capturing imagery at
spatial and spectral resolutions far superior to that available less than a decade

ago.

The increased information content of high-resolution imagery should
ultimately result in more accurate wetland mapping and monitoring
strategies. This is based on the assumption that these new technologies help
to improve the differentiation of coastal wetland plant communities. Thus,
the first step is to explore if indeed botanical differentiation is improved via
these technologies. Few studies have explored the potential benefits and
pitfalls of applying high spatial and spectral resolution imagery to the
characterization and mapping of freshwater coastal wetlands. Of the
studies that have been conducted, even fewer have extensively explored the
role of in-situ reflectance data and classification strategies with respect to the
botanical differentiation of wetland plant communities. An exceptional
opportunity exists to explore the application of high spatial and spectral
resolution imagery in combination with radiometric field measurements in

Great Lakes coastal wetlands.

The objectives of this research are to: 1) identify the best-performing image

Classification tool with respect to field-verified botanical regions of interest; 2)

29

quantitatively examine the properties of in-situ reflectance data to discern the
optimal spectral framework for wetland imagery; and 3) quantitatively discern
the optimal spatial framework for wetland imagery. These objectives will be
addressed by 5 investigations grouped into three independent research

components.

Research Component]:
Identification of the best-performing image classzfication tool with respect to field-

verzfied, botanical regions-of-interest (ROIs).

 

Investigation 1: Identification of the ”best” method of classification
via the direct comparison of the ISODATA, Spectral Angle Mapper,
and Maximum-Likelihood classification results using image-based

(Horseshoe Bay, 1-meter) training data.

Investigation 2: Comparison of the results from [-1 (ISODATA,

 

SAM, and Maximum-Likelihood classifications of raw imagery) to

similarly derived classifications based on Principal Components rotated

imagery.

30

 

Investigation 3: Determination of the transferability of the ”best
approach” identified above (I-1 and I-2) via imagery taken at another

location (Prentiss Bay), with a different sensor (AISA).

Research Component 2:

Quantitatively examine the properties of in—situ reflectance data to discern the

optimal spectral framework for wetland imagery.

Investigation 4: Quantitative examine the properties of in-situ

 

reflectance data to discern the optimal spectral framework for wetland
imagery, to aid in the assessment of SAM and Maximum-Likelihood
classification changes resulting from the manipulation of the spectral

resolution of select imagery.

Research Component 3:

Quantitatively discern the optimal spatial framework for wetland imagery.

Investigation 5: Assessment of SAM and Mahalanobis-Distance
classification changes resulting from the independent manipulation of

the spatial resolution of select images.

Note: Images in this dissertation are presented in color.

31

Chapter 2: A Comparison of Image Classification Strategies
to Optimize Wetland Botanical Differentiation.

2.1 Introduction and Rationale

In the 19705, state and federal agencies increasingly looked towards aerial
photography for the purposes of wetland mapping and inventory (Gross et al.,
1989; p. 475). Film types that are sensitive to NIR electromagnetic energy are
generally preferred for wetland applications because they are more effective in
detecting standing water and determining biomass, as well as differentiating
wetland and upland vegetation (Hardisky et al., 1986). Aerial photography
(1:1000 to 1215,000 scales) has been found to be a fast, accurate, and relatively
inexpensive method by which wetland boundaries, botanical changes, and
aerial extent can be recorded (Mitsch and Gosselink, 2000; p. 107). Two
characteristics limit the applicability of aerial photography to detailed
vegetative mapping strategies: 1) the limited amount of spectral information
contained in each image (Jensen et al., 1986), and 2) its non-digital format

(Hardisky et al., 1986).

The launch of the Landsat 1 satellite (i.e. 19705) marked the practical
beginning of spaceborne, earth resource remote sensing. Over the ensuing
thirty years, a variety of systems, with varied spatial and spectral resolutions,

have been developed. Orbiting satellite-based sensors do provide near-

32

simultaneous coverage of wetland imagery over large geographic areas
recorded in digital format (Gross et al., 1989; p. 476). Satellite platforms
continue to provide an overwhelming array of earth resource data when fitted
with remote sensing systems, and the status of these technologies continues to
change as scientific advances yield improved spacecraft and sensors (Lillesand
and Kiefer, 2000; p. 373). However, regarding their utility for detailed wetland
characterization, current systems are limited by their spatial and / or spectral
resolutions. These limitations become obvious in the context of the two
fundamental assumptions of this research: 1) the spatial heterogeneity of coastal
wetlands is best captured by digital imagery with a spatial resolution of 5
meters or smaller, and 2) differentiation of wetland plant communities is best
achieved through the analysis of hyperspectral (i.e. numerous narrow bands)
reflectance records. Accepting these assumptions means that aerial
photography and currently available satellite imagery are sub-optimal for

detailed wetland mapping.

The development of imaging spectrometers created interest in utilizing
spectral reflectance data of vegetation canopies measured from the ground to
aid in remotely differentiating plant communities. Many next-generation,
satellite-based systems (e.g. COIS, Hyperion, MODIS or Warfighter-l)
incorporate hyperspectral sensors in order to address the spectral limitations of

aerial photography and current satellites systems (ASPRS, 1995; Stoney and
33

Hughes, 1998; Thenkabail et al., 2000). A major advantage of imaging
spectrometry lies in the potential for integrated analysis comparing calibrated
remotely sensed imagery with field spectral measurements (Goetz 1992; Kruse
et al., 1993). The use of systematic differences in canopy pigmentation and
related chemistry to classify to the botanical species level has recently emerged

as a promising new approach.

The development of new-age airborne and satellite-based imaging systems
is making hyperspectral imagery more affordable and easier to obtain.
Although not widespread, the wetland literature contains recent examples of
investigations that utilized hyperspectral systems to map and monitor plant
communities (Savastano et al., 1984; Zacharis et al., 1992; Forsyth et al., 1998;
Borstad Associates, 1995; Jupp et al., 1994; Ritter and Lanzer, 1997; Brown and
Borstad, 1999; Wang et al., 2000). Common to these investigations is the
exploitation of improved spatial or spectral resolution to more effectively

differentiate wetland vegetation.

The remote sensing community is hopeful that the improved resolving
power of these hyperspectral (i.e. spatial, spectral, and radiometric) will lead to
significant improvements in the differentiation of wetland plant communities,

thereby improving coastal wetland mapping and monitoring efforts by

34

improving the accuracy of image classifications. One of the fundamental
questions which remain unanswered is which image classification methodology
is best suited for hyperspectral applications within the context of Great Lakes
coastal wetlands. Answering this question is an important step needed to

advance the application of hyperspectral imagery in coastal environments.

The goal of this component of my research is to perform a comparison of
common classification methodologies in order to evaluate their ability to
capitalize on the high spatial and spectral resolution of hyperspectral imagery.
This goal addresses the first research objective outlined in Chapter 1 - identify
the best-performing image classification tool with respect to the field-verified

botanical regions-of-interest.

2.2 Methods

2.2.1 Airborne Imagg Acquisition

 

Two airborne imaging campaigns provided the digital imagery utilized
throughout most of this research. On August 30, 1999, 3D], Inc.

(www.3dillc.com) captured airborne, digital imagery of the two study sites

 

located in the Les Cheneaux Island complex (Appendix E; Figure E-1). The
sensor utilized by 3DI, Inc. was the Airborne Imaging Spectrometer for

Applications (AISA; www.specim.fi). Single-pass images (360 meters wide)

35

were collected over Prentiss and Mismer Bay, encompassing and roughly
centered on previously established ecological transects (Gathman et al., 1999).
Each of the 21-channel images were approximately 900 meters in overall length,
extending from adjacent uplands, through the wetland proper and ending over

open water (Figures 3.2 and 3.3).

 

On September 11, 2000, ITRES, Inc. (www.itres.com) captured digital
imagery over the Horseshoe Bay study area (Appendix E; Figures E-2 and E3).
The sensor manufactured and deployed by ITRES Inc. was the second
generation, Compact Airborne Spectrometric Imager (CASI-II). Both AISA and
CASI-II are programmable, push broom, imaging spectrometers capable of
being operated in spatial and spectral mode. ITRES, Inc. captured imagery in
both spatial and spectral (hyperspectral) mode. The spatial extent of these
images are large (2 km x 10 km) in comparison to the Les Cheneaux imagery, so
the majority of my image analyses were performed on only the center portion
of the imagery, largely Horseshoe Bay proper (Figure 1.5). This served two
purposes: 1) it reduced the computation time needed to process and classify
these image files, and 2) it eliminated the extraneous portions of the imagery
where little or no ground control or field data were collected. Table 2 .1
provides an overview of the AISA and CASI-II imaging and support systems,
as well as a summary of basic flight parameters. The band center and Full-

Width-Half-Maximum (FWHM) for each spectral band collected are

36

summarized in Table 2 .2. Visual examination of a variety of spectral signatures
from wetland plants provided the means by which band positions were
strategically located throughout the visible and NIR regions of the spectrum for
this 1-meter imagery. The 4-meter CASI-II imagery is truly hyperspectral,

utilizing 48 contiguous, spectral bands across the N IR and visible wavelengths.

2.2.2 Image Geo-rectification

Both sets of imagery were delivered in geo—referenced format as a result of
the on-board positioning technologies (GPS and Inertial Momentum) utilized
by both imaging contractors. Both vendors indicated that the expected
positional accuracy of their 1-meter imagery would be approximately 5-7
meters. At this level of image rectification, the spatial integrity of the in-situ
data (i.e. sub-meter accuracies) would not be preserved. Thus, an in-house geo-

rectification strategy was implemented.

Additional ground control points (GCPs) were identified at each of the
study sites. Visually distinct features, such as pavement markings and culvert
headwalls are ideal GCPs since they can be precisely located on the digital
imagery. In areas where no infrastructure was present (i.e. Horseshoe Bay),

white plywood GCP panels were installed in the wetland prior to the imaging

37

 

 

 

 

 

Sensor AISA CASI—II
Sensitivity Range 400-1000nm 400-1000nm
IFOV 1.0 milliradians 1.3 milliradians
Swath Width 360 pixels 512 pixels
Maximum # channels 288 bands 288 bands
Radiometric Res. 16 bit 12 bit

@ 1 meter pixels 21 bands 18 bands

@ 4.0 meter pixels n/ a 48 bands
Horseshoe Bay (1m-CASI-II)

Date 9/11/ 00
Start time (HBW path 1) 12:21pm
Sun angle (HBW path 1) 52.0 degrees
Start time (HBW path 2) 1:06pm

Sun angle (HBW path 2) 51.3 degrees
Start time (HBW path 3) 1:16pm

Sun angle (HBW path 3) 50.8 degrees
Start time (HBW path 4) 1:26pm

Sun angle (HBW path 4) 50.2 degrees
Horseshoe Bay (4m-CASI-II)

Date 9 / 11 / 00
Start time 12:54pm
Sun angle 51.7 degrees
Les Cheneaux (lm-AISA)

Date 8/30/99
Start time (MBW) 12:18pm
Sun angle (MBW) 48.1 degrees
Start time (PBW) 12:49pm
Sun angle (PBW) 48.2 degrees

 

Table 2.1. Instrument Specifications and Flight Parameters.

38

 

 

 

 

CASI-II FWHM AISA FWHM CASI-II FWHM
48 band in; 21 Band 1_m_ 18 band 1m

414.3 5.7 448.81 n / a 450.5 16
425.4 5.7 508.79 9.73 549.7 6.8
436.5 5.7 549.32 9.73 561 4.9
447.7 5.7 559.17 10.14 572.4 6.8
458.8 5.8 579.45 10.14 584.7 5.8

470 5.8 609.87 10.14 617.9 6.8
481.2 5.8 630.15 10.14 630.2 5.9
492.4 5.8 650.43 10.14 641.6 5.9
503.6 5.8 660.57 10.14 663.6 6.8
514.9 5.8 670.71 10.14 676 5.9
526.2 5.8 680.85 10.14 687.5 5.9
537.5 5.8 690.99 10.14 708.5 5.9
548.8 5.8 701.13 10.14 721 6.9
560.1 5.8 711.97 10.56 767.1 6.9
571.4 5.8 731.33 10.56 800.8 5.9
582.8 5.8 741.89 10.56 813.3 6.9
594.1 5.8 752.45 10.56 880.9 6.9
605.5 5.8 763.01 10.56 927.3 6.9
616.9 5.9 799.97 10.56
628.3 5.9 810.53 10.56
639.7 5.9 821.09 10.56
651.2 5.9
662.6 5.9
674.1 5.9
685.5 5.9

697 5.9
708.5 5.9 ***CASI-II 48 band

720 5.9 ***continued
731.5 5.9 847.1 5.9

743 5.9 858.7 5.9
754.6 5.9 870.3 5.9
766.1 5.9 881.9 5.9
777.7 5.9 893.4 5.9
789.2 5.9 905.1 5.9
800.8 5.9 916.7 5.9
812.3 5.9 928.3 5.9
823.9 5.9 939.9 5.9
835.5 5.9 951.5 5.9

Table 2.2. Instrument Bandsets with FWHM

39

 

 

CASI-II FWHM IKONOS FWHM
10 Band 60cm 4band 4_m_

449.5 15 480.5 71
549.7 6.8 550.5 89
562 5.8 665 66
629.3 6.8 805 96
663.6 6.8
676 5.9
721 6.9
799.8 6.9
812.3 5.9
882.8 6.9

 

 

 

Table 2.2. Continued

overpass. Sub-meter RT-DGPS positions were recorded for all GCP locations.
These GPS data were exported via Trimble Pathfinder Office software into
Arcview shapefiles. The GCPs provided an adequate distribution of control
points to yield sub-pixel image rectification at each study site. The Image-to-
Map Rectification module in the ENVI software provides a dual screen
interactive tool by which images can be geo-rectified. A series of rectification
pairs were created by visually selecting a GCP position on the image that was
then linked to its appropriate location as recorded via the RT-DGPS. Once all of

the viable rectification pairs were selected, an RST (Rotation-Scaling-

40

Translation), nearest neighbor rectification was performed. RST interpolation
utilizes a linear (first-order polynomial), least-squares procedure to model
spatial image corrections (Jensen, 1996. p. 127). This nearest neighbor method
of intensity interpolation was chosen for two reasons: 1) the original input pixel
values were not altered, and 2) the maximum spatial shift associated with any
one input pixel would be one-half pixel, comparable to the sub-pixel accuracies

associated with the in-situ data (Lillesand and Kiefer, 2000. p. 530).

The overall effectiveness of an irnage-to-map rectification can be expressed
as a total root-mean-square error (RMSE), which is the average of all residual
linear distances remaining between image positions and their paired GCT/ GCP
locations after completion of the rectification process. In this research, a total
RMSE at or below the nominal pixel size of the imagery was desired and
achieved (Table 2.3) so that the sub-pixel accuracies associated with the in-situ

data could be effectively compared to the imagery at the pixel level.

2.2.3 Imager; Percent Reflectance Transformation

The baseline image data were recorded as raw radiance values, creating the
need to transform the digital number associated with each pixel into relative
percent reflectance. Generally, calibration surfaces used in such
transformations need to be at least ten times larger than the nominal spatial

resolution (pixel size) of the imagery. Large, homogeneous gravel parking

41

 

Horseshoe Bay (l-meter data)

 

Total RMSE: 0.796261
RST/ Nearest Neighbor Resampling

 

 

 

 

 

 

 

 

 

 

 

 

 

42

GCP# (Image X,Y) (Predict X,Y) Error X,Y) (RMSE)
#1 (1526.50,4610.05) (1527.08,4609.93) (0.58,-0.12) -0.59
#2 (1013.50,4966.45)(1013.28,4966.50) (-0.22,0.05) -0.22
#3 (167145584450) (1670.03,5844.80) (-1.42,0.30) -1.46
#4 (1125.50,6143.50)(1125.86,6143.42) (0.36,-0.08) -0.37
#5 (1862.50,6432.50)(1863.20,6432.35) (0.70,-0.15) -0.72

Horseshoe Bay (4-meter data)

Total RMSE: 0.288911

RST/ Nearest Neighbor Resampling

GCP# (Image Xhfl (Predict X,Y) (Error X,Y) (RMSE)
#1 (414.45,1277.45) (414.12,1277.57) (-0.33,0.12) -0.36
#2 (449.64,1372.64) (450.07,1372.48) (043,016) -0.46
#3 (284.45,1366.45) (284.54,1366.42) (0.09,-0.03) -0.1
#4 (449.45,1586.45) (449.20,1586.54) (-0.25,0.09) -0.26
#5 (498.55,1732.45) (498.61,1732.43) (0.06,-0.02) ~0.06

Prentiss Bav

Total RMSE: 0.931855

RST/ Nearest Neighbor Resampling

GCP# (Image X,Y) (Predict X,Y) (Error X,Y) (RMSE)
#1 685.50,518.50) (684.78,518.32) (-0.72,-0.18) 0.74
#2 680.50,536.50) (680.65,536.39) (0.15,-0.11) 0.19
#3 668.00,553.45) (668.59,554.10) (059,065) —0.88
#4 416.05,593.21) (415.44,593.52) (-0.61,0.31) 0.69
#5 420.00,593.47) (420.37,594.40) (037,093) 1
#6 516.67,628.11) (516.31,627.52) (-0.36,-0.59) 0.69
#7 520.06,628.11) (520.10,628.17) (004,006) 0.07
#8 518.75,606.25) (518.10,605.91) (065,034) -0.74
#9 521.00,606.75) (521.90,606.46) (090,029) 0.95
#10 (519.25,617.50) (519.38,617.08) (0.13,-0.42) 0.44

Table 2.3. Image Rectification Summary.

 

 

 

 

 

#11 (679.50,670.25) (678.92,670.40) (058,015) 0.6
#12 (682.77,672.23) (684.06,672.13) (1.29,-0.10) 1.3
#13 (616.46,716.31) (615.43,715.28) (-1.03,-1.03) 1.46
#14 (623.50,729.00) (624.12,7 29.34) (062,034) 0.7
#15 (551.15,699.92) (550.37,701.10) (-0.78,1.18) 1.42
#16 (377.00,718.00) (377.80,717.07) (0.80,-0.93) 1.22
#17 (398.44,888.22) (397.31,888.08) (-1.13,-0.14) 1.14
#18 (4042288322) (405.21,883.72) (099,050) 1.11
Table 2.3. Continued

areas were present in both the Horseshoe Bay and Prentiss Bay imagery. These
areas served as ground-based calibration surfaces through which image

radiance data were transformed into absolute reflectance.

In order to derive a reflectance coefficient (pa) for the gravel calibration
surface within the Prentiss Bay imagery, near simultaneous, SE-590
measurements were taken of it (ETgravel-pbw) and the in-situ reference panel
(ETmsu) under cloudless skies. The gravel SE-590 radiance measurements
(ETgravel-pbw and ETmsu) were recorded several weeks prior to the imaging
overpass at Prentiss Bay. Although it was my intent, it was not possible to
capture gravel reflectance data during the Prentiss Bay imaging overpass
because 3DI, Inc. flew the mission during an unexpected break in cloud cover.
The moisture level within the gravel region was similar during both the

imaging overpass and the in-situ measurements, allowing the previously

43

collected reflectance data to be utilized. The near-simultaneous SE-590
measurements of the gravel and reference panel radiance allowed pagravel-pbw to

be calculated with Equation 2.1.

pagravel-pbw = ETgravel-pbw/ETmsu * pamsu (for bands 1 through 252) (Eq. 2.1)
ETgravel-pbw = Prentiss Bay gravel radiance

ETmsu = In-situ reference panel radiance

pagmel-pbw = Prentiss Bay gravel reflectance coefficient

pawn = In—situ reference panel reflectance coefficient

A similar procedure was followed in order to derive a reflectance
coefficient (Pagravel-hbw) for the gravel calibration surface within the Horseshoe
Bay imagery (Equation 2.2). Members of the ITRES Inc. field crew captured a
series of gravel (ETgravel-hbw) and Spectralon (Elspec) radiance measurements

using an ASD FieldSpec-Jr spectroradiometer at Horseshoe Bay during the

imaging campaign.

pagravel-hbw = ETgravel-hbw/ ETspec * paspec (for bands 1 through 252) (Eq. 2.2)

ETgravel-hbw = Horseshoe Bay gravel radiance

ETspec = Spectralon reference panel radiance
Pagravel-hbw = Horseshoe Bay gravel reflectance coefficient
paspec = Spectralon reflectance coefficient

In order to create reflectance coefficient values appropriate for each image
bandset, Pagravel-pbw andpagravel-hbw were spectrally resampled into the bandsets
shown in Table 2.2 to yield three synthetic reflectance coefficients, PaAlSA,

44

p°CAsns and PaCASI48. These values, which corresponded well with published
gravel reflectance values (Bowker et al., 1985, Clark et al., 1993), range from
approximately 15 to 40 percent across the wavelength domain of both the AISA

and CASI-II instruments.

The gravel calibration surfaces were among the brightest features (highest
reflectance) throughout the spectral range of the imagery. Two Regions-of-
Interest were delineated that encompassed the areas over which the in-situ
radiance data (ETgravel-hbw and ETgravel-pbw) were captured. The band-specific
mean reflectance (i.e. ETmean-pbw, and ETmean-hbw) and associated standard
deviation (i.e. ETstd-pbw, and ETstd-hbw) were extracted using the ENVI Region of
Interest (ROI) module for the Prentiss and Horseshoe Bay gravel R015. The
band-specific standard deviation values were all at or below five percent.
Generally, calibration surfaces are considered homogeneous, flat reflectors if
their associated standard deviations are less than five percent. The technical
specifications for the well-calibrated Landsat 7, ETM+ imager requires a

radiometric precision of +/ -5 percent.

The division of each band-specific ETmean value by its band-specific
reflectance coefficient (i.e. PaAlSA, PaCASHS or puCASI48) generated radiance values
(E1100) equivalent to those expected from a calibration panel exhibiting 100

percent reflectance, and quantified the total per band irradiance striking the

45

two gravel calibration surfaces. A pseudo-calibration surface was then created
within the ”no-data” region of each image by interactively typing in these band
specific ET 100 values within the ENVI Spatial Pixel Editor. The resulting cluster
of pixels was easily identified against the backdrop of the solid black ”no-data”
region of each image. Based on the assumption that the now quantified
irradiance (E1100) values are constant within any one band and image, percent
reflectance values were calculated for each image pixel by dividing the radiance
value recorded in each (ETpixel) by the appropriate, band-specific, pseudo-
calibration (E1100) value utilizing the ENVI Flat Panel Calibration module. An
example of the series of calculations utilized to transform the raw pixel radiance
values into absolute percent reflectance for the Prentiss Bay imagery is

summarized in Table 2.4.

2.2.4 Imagery Principal Components Rotation

In remote sensing, feature extraction involves the identification of those
statistical characteristics of the remotely sensed data that capture the most
systematic variation (Campbell, 1996; p. 288). Systematic variations, as opposed
to non-systematic variations and / or data noise, ultimately provide the
foundation upon which remote sensing targets are differentiated. A feature-
extraction tool commonly applied throughout the scientific community is
Principal Components Analysis (PCA). PCA has been used to generate

orthogonal channels or components that are more interpretable and

46

 

 

 

 

Rho ET100 ETpixel %r (*100)
Band mmeam pdAISA) ET(mean)/pa(AISA) (Egg (EIpixel/ETIOO)
1 7864 0.9077 8663.52 89 1.03
2 12753 0.9223 13827.46 214 1.55
3 13701 0.9264 14788.92 689 4.66
4 13523 0.9268 14591.66 703 4.82
5 13207 0.9292 14213.24 558 3.92
6 12651 0.9305 13595.55 409 3.01
7 11806 0.9306 12686.71 324 2.55
8 11306 0.9308 12146.91 255 2.10
9 11376 0.9305 1225.32 223 1.82
10 11438 0.9313 1281.23 197 1.60
11 11123 0.9312 11945.10 204 1.70
12 9702 0.9307 10424.68 289 2.78
13 10197 0.9305 10959.21 626 5.71
14 10125 0.9302 10884.50 1158 10.63
15 8861 0.9282 9546.72 2245 23.52
16 9573 0.9275 10321.42 3072 29.77
17 9504 0.9273 10248.72 3428 33.44
18 5833 0.9256 6302.10 2248 35.67
19 8051 0.9236 8716.95 3298 37.83
20 7686 0.927 8330.25 3150 37.82
21 6330 0.925 6861.95 2606 37.98

 

 

Table 2.4. Image Percent Reflectance Methodology

dimensionally reduced with respect to the raw imagery (Jensen, 1996; p. 518;
Singh and Harrison, 1985; and Kauth and Thomas. 1976). PCA has also been
used to address a fundamental problem with multi/ hyperspectral imagery --
extensive inter-band correlation or data redundancy (Lillesand and Kiefer, 2000;
p. 518). The goal of image-based PCA is to compress all of the information
(correlation/ variance) contained in an original image into fewer, independent

(orthogonal) channels or principal components. If the vast majority (>99%) of

47

the meaningful variance were compressed into relatively few, uncorrelated
dimensions, the efficiency of image processing and classification would be
greatly increased through referencing these dimensions instead of the original
data. A detailed description of the statistical framework used to perform PCA
is beyond the scope of this research, but a more complete explanation can be

found in Davis (1986).

The baseline Prentiss and Horseshoe Bay imagery were transformed into
their principal components via ENVI’s Image Transformation module
(Appendix E; Figures E-4 and E-5, respectively). Covariance-based image
transformations are the norm because they focus on major target differences,
resulting in more ”class" separation. During preliminary PCA runs, the
number of dimensions generated was set equal to the number of bands
associated with the imagery. Beginning with the first dimension, each was
systematically displayed in order to visually assess their spatial coherence.
Those dimensions that displayed little spatial coherence (salt-and-pepper
images) were deemed non-meaningful due to the random distribution of the
data (i.e. noise) information contained in that dimension. Based on these
efforts, it was determined that 6 dimensions regularly contained meaningful

(i.e. spatially coherent) data within this research as shown in Figure 2.1.

48

A final PCA rotation was completed on the Prentiss Bay 1-meter, 1-
Horseshoe Bay 1-meter and Horseshoe Bay 4-meter images, yielding a single
output file containing six components (Figure 2.1). The statistics files associated
with each PCA run are presented in Appendix A. Greater than 99% of the
original variance is captured in the first six components of each PC image set,
indicating that the dimensionality of the original imagery had been effectively
reduced into these few, independent data channels. Often, false-color images
generated from combinations of components depict more subtle differences in
color shading and distribution than traditional color—composite images (Jensen,
1996, page 179). Dimensions 2, 3, and 4 did the best job of visually enhancing

the botanical differences found within the Prentiss and Horseshoe Bay imagery.

2.2.5 Accuracy Assessment Regions-of-Interest

 

In-situ reflectance measurements were collected for the predominant
species, species assemblages, and cover types at the Prentiss and Horseshoe Bay
study sites. Care was taken to ensure that each measurement was at the
approximate center of a botanically homogeneous region covering multiple
image pixels. A real-time, differentially corrected GPS (RT-DGPS) receiver was
used to record the approximate center and boundary of each region to sub-
meter accuracies, providing the bulk of the ground-truth botanical data utilized
for classification accuracy assessments. Additional field-verified locations were

delineated within the Horseshoe Bay imagery that were not recorded during

49

Dimension 1 Dimension 2

.4

Dimension 3

       

   

 

Figure 2.1. Horseshoe Bay; Dimensions 1-6 PCA Images

50

initial site visits. The accuracy assessment/ training regions used throughout
Investigations 1 through 3 are shown in Tables 2.5 and 2.6. In addition, the
locations of the training ROIs shown in Table 2.6 are displayed on Figure E-2,

Appendix E.

Each image-based ROI was further divided into an easterly and westerly
region. This division created two adjacent ROIs, each representing the spectral
characteristics of the original ROI from which they were derived, and
independently provided both training (i.e. supervised classification) and
accuracy assessment pixels. Classification results were compared via confusion
matrix parameters (9. g. Kappa Coefficients) that were generated with the ENVI

Post-Classification module with respect to the easterly ground-truthed R015.

The accuracy assessments utilized in this research are based solely on the
botanical ROIs summarized in Table 2.5 and 2.6. Ideally, the entire wetland
would have been ground-truthed allowing a more formal, randomly generated,
accuracy assessment (i.e. 1 % of pixels) to be applied across the entire image.
Such an approach was logistically impossible. However, the ROI-based
accuracy assessments presented in this research remain valid as well as relevant
because they allow multiple classification results to be compared with respect

to their ability to differentiate known botanical communities.

51

 

 

 

 

 

ROI Matrix Code Scientific Name Pixel #
1: Stricta Carex stricta 68
2: stricta50/ lasiocarpaSO C. stricta/Carex lasiocarpa 46
3: lasio/ acutule C. lasiocarpa/Scirpus acu tus 44
4: lasio/ typha50 C. lasiocarpa/Typha 40
5: roadside-veg n/a 48
6: asphalt n/a 50
7: typha80 Typha Angustifolia 62
8: acutu350/ typha50 T. angustifolia/S. Acutus 56
9: acutus40/ soil60 Vallisneria Americana 46
10: Dolomite n/a 44
11: Chara Chara sp. 44
12: Soil70/ typha15/acutus15 T. an gustifolia/S. Acu tus 42
13: Nuphar-12in Nuphar advena 44
14: Potamogeton-15i Potamogeton natans 44
15: 2muck-20cm n/a 56
16: acutus-sparse Scirpus acutus 50
17: vallisneria Vallisneria Americana 94
18: pine Pinus sp. 50
19: shadow n/a 46
20: acutusclumps-dense Scirpus acu tus 100

 

Table 2.5. Prentiss Bay Regions-of-lnterest.

2.2.6 Image Classifications

Image classification, both supervised and unsupervised, is the process of
establishing a link between an informational category of interest and a related
spectral class. Unsupervised classification is based on the fact that most
remotely sensed images are composed of spectral classes that are reasonably
uniform with respect to reflectance (brightness) across one or more spectral

channels, and can therefore be defined, labeled, and mapped (Campbell, 1996;

52

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

3Q! Matrix Code Scientific name Pixel #
1: Chara Chara sp. 96
2: phragmites-clump Phragmites australis 100
3: scival-dense Scirpus validus 108
4: scirpusclump Scirpus validus 100
5: typha-soil2 Typha an gustifolia 100
6: sagrigida60/ 40 Sagittaria rigida 52
7: beach sand n/a 54
8: shallowh20/ sand n/a 80
9: Hetdub Heteranthera dubia 66
10: Sagrigida Sagittaria rigida 84
11: valisneria Vallisneria americana 72
12: deeph20 n/a 94
13: scivalbrown Scirpus validus 62
14: leersia Leersia oryzoides 54
15: muck n/a 54
16: saggram/chara Sagittaria graminea/Chara sp. 60
17: Typhah20 Typha angustifolia 64
18: Nuphar Nuphaadvena 164
19: Eleocharis Eleocharis sp. 88
20: Impatiens I mpatians capensis 200
21: bluejoint Calamogrostis canadensis 232
22: cutgrass-eleocharis L. oryzoides 226
23: Eloecharis-sand Eleocharis sp. 50
24: typha-phrag-scivalmix Typha/Phragmites/Scirpus 90

 

Table 2.6. Horseshoe Bay, l-meter Regions-of-Interest.

53

 

p. 318). The primary advantage of all unsupervised classification strategies is
that no extensive prior knowledge of the informational categories found in an

image is required.

In supervised classification, on the other hand, the spectral characteristics
and / or location of informational classes are known prior to image classification
through a combination of fieldwork, image analysis, and personal experience
(Jensen, 1996; p. 197). The major advantage of supervised classifications is that
the spectral classes can be tailored to the desired informational categories
(Campbell, 1996; p. 328). A major disadvantage of a supervised classification is
that a fixed classification structure is imposed on the data, one that may not
match the natural classes that exist within the n-dimensional data space of the

imagery (Campbell, 1996; p. 328).

2.2.6.1 ISODATA Classification (Investigation 1)

The unsupervised classification tool used in component one of this research
was the ISODATA (Iterative Self-Organizing Data Analysis Technique)
classifier available in the ENVI Classification module (Jain, 1989; Sabins, 1987;
Tou and Gonzalez, 1974). It is an iterative process that begins by initially
establishing a minimum and maximum number of spectral class centroids that
are evenly distributed throughout the n-dimensional (n = number of bands)
data space (RSI, 1999; p. 560). A 48 class ISODATA classification was generated

54

from the Horseshoe Bay, 1-meter imagery using the parameters shown in
Figure 2.2. The 24 regions-of—interest summarized in Table 2.6 were used as

ground-truth data to which the ISODATA results were compared.

In most instances, there were only one or two ISODATA clusters that
captured the vast majority of any particular ground-truth class. When faced
with class overlap, the ISODATA cluster was assigned to the ground-truth class
having the largest proportion of pixels within the cluster. This labeling
prioritization method ensured that the final cluster assignments represented the
best possible classification (i.e. highest possible percent correctly classified).
Upon completion of all class assignments the confusion matrix parameters
referenced in the results section of this chapter were generated. The evaluation
of one or more confusion matrix parameters allowed the performance (ROI
botanical accuracy) of the ISODATA classification of the Horseshoe Bay, 1-
meter image to be assessed and compared to similar Maximum-Likelihood and

SAM classifications.

2.2.6.2 Maximum-Likelihood Classification (Investigation 1)
The first of two supervised classification tools used in this component of the
research was the Maximum-Likelihood classifier available in the ENVI

Classification module. Maximum-Likelihood decision rule values range from 0

55

III ISODTA Par meters
are

    

Figure 2.2. Isodata Classification Parameters

to 1, and represent the probability that a single pixel will be classified into any
one spectral class. The 24 westerly ROIs (Table 2.6) served as the training data,
generating a 25-class classification image (one being unclassified). A
probability threshold of 0.70 was established that caused only those pixels with
probabilities of 70% or greater to be included in any one spectral class. Pixels
with probabilities lower than this threshold for all classes were unclassified. A
probability threshold of 85% is commonly applied to remotely sensed data
within the context of land use and land cover assessment (Anderson et al.,
1976). Due to the general spectral complexity associated with coastal wetlands
and the high precision of the classification categories (species or closely related
assemblages) associated with coastal wetlands, a relaxed threshold of 70%

seemed justified.

56

The easterly ROIs were used for ground-truth data from which a confusion
matrix and associated parameters were generated (i.e. overall accuracy,
producer's accuracy, consumer’s accuracy, kappa coefficient, commission
errors, and omission errors). The evaluation of one or more confusion matrix
parameters allowed the performance (botanical accuracy) of the Maximum-
Likelihood classification of the Horseshoe Bay, l-meter image to be assessed

and compared to similar ISODATA and SAM classifications.

2.2.6.3 Spectral-Angle-Mapper Classification (Investigation 1)

The second supervised classification routine utilized was the ENVI Spectral
Angle Mapper (SAM), which is a physically- based spectral classification tool
that uses spectral angles defined in n-dimensional data space to match pixel
spectra to reference spectral signatures (RSI, 1999; p. 556). It is an automated
method for comparing image spectra to individual spectra or a library of

spectral signatures (Kruse et al., 1993).

A 24-class, SAM classification image was generated using the ROI-based
methodology analogous to that outlined in the previous section. Two
operational parameters were entered that governed the characteristics of the
final SAM classification image. The maximum-value parameter was selected
which causes the classification to include only those pixels with a rule value

(angular sum = a) lower than a specified threshold. Varying (1 across a range of

57

0.1 to 0.4 radians generated a series of preliminary SAM image classifications.
An (I value of .25 was selected for use throughout this research because it
resulted in a limited number of unclassified pixels while maintaining the ability

to delineate similar spectra.

Analogous to the methodologies outlined in the previous section the
easterly ROIs were treated as ground-truth data from which a confusion matrix
and associated parameters were generated. The evaluation of one or more
confusion matrix parameters allowed the performance (botanical accuracy) of
the SAM classification of the Horseshoe Bay, 1-meter image to be assessed and

compared to similar ISODATA and Maximum-Likelihood classifications.

2.2.6.4 PC-rotated Classifications ((Investigation 2)

Figure 2.1 illustrates that the first six principal components extracted from
the hyperspectral imagery contain information linked to the spatial distribution
of the botanical communities, and likely contain information vital to the
differentiation of wetland vegetation. Two Maximum-Likelihood
classifications were generated from the PC-rotated Horseshoe Bay imagery.
The first utilized all 6 PCA dimensions (Figure 2.1) while the second used only
PCA dimensions 2,3,4,5, and 6. Classification results generated without the 1st
dimension (i.e. n-dimensional albedo) addressed whether or not this dimension

is relevant to the differentiation of wetland plant communities. The Maximum-

58

Likelihood classifier was again configured to yield 24 botanical classes based on

the operational parameters outlined in section 2.2.6.2.

Two SAM classifications were also generated from the Horseshoe Bay,
PCA-transformed imagery similar to the classifications outlined in section
2.2.6.3. The first utilized all 6 PCA dimensions (Figure 2.1) while the second
used only PCA dimensions 2,3,4,5, and 6. Confusion matrix parameters from
the baseline (Investigation 1) and PC-rotated (Investigation 2) classifications
were generated with respect to the 24 ground-truth ROIs shown in Table 2.6.
These parameters allowed the relative differences in botanical classification

accuracies between the first two investigations to be directly compared.

2.2.6.5 Prentiss BaLClassifications (Investigation 3)

 

The classification methodologies presented in the previous sections utilized
the Horseshoe Bay, 1-meter CASI-II imagery. Investigation 3 utilized the 1-
meter AISA imagery of Prentiss Bay in order to test whether classifier
performance was consistent across multiple sensor technologies (CASI-II vs.
AISA) and study sites. The regions-of-interest outlined in Table 2.5 were used
in combination with the Prentiss Bay AISA imagery to generate a series of
classifications analogous to those generated from the Horseshoe Bay imagery
(sections 2.2.6.2 through 22.6.4). The ISODATA classification of the Prentiss

Bay imagery significantly under-performed the two supervised classification

59

methodologies and was dropped from further consideration. The 20 westerly
ROIs (Table 25) served as training data, generating a 21-class classification
image (one being unclassified). The easterly ROIs were used as ground-truth
data from which a confusion matrix and associated parameters were generated.
The evaluation of one or more confusion matrix parameters allowed the relative
performance (i.e. botanical accuracy) of analogous Prentiss Bay (i.e. AISA) and

Horseshoe Bay (i.e. CASI-II) classifications to be compared.

2.3 Results and Discussion

Error matrices provided the foundation for comparing the many
classification outcomes generated as part of this research, and provide a
foundation for both descriptive and analytical statistical techniques (Congalton,
1991). Kappa Coefficients and overall-accuracy values best provide a measure
of how well a classification performed regarding the ground-truth data (Table
2.5 and 2.6). Kappa Coefficients are an adjusted accuracy measure in which an
estimate of chance agreement has been subtracted from the baseline overall-

accuracy (Campbell, 1996; p. 390).

When classification outcomes are to be directly compared, it is prudent to
verify if their associated Kappa Coefficients are statistically significantly

different. Each confusion matrix generated for this research was manipulated in

60

a spreadsheet to generate the four independent theta coefficients (Hudson and
Ram, 1987; Congalton et al., 1983), which are intermediaries needed to
calculate the Kappa variance and ultimately a Z-score (Congalton et al., 1983).
Z-scores (equation 2.3) were used to establish whether the Kappa values from
any two classifications were statistically different from each other at or above
the desired confidence limit (i.e. 95%). Unless otherwise indicated within the
classification summary tables, the results of all viable classification parings

were statistically different at or above the 95% confidence limit.

Z = (K1 - K2) / [(sqrt (VK1 + VK2)] Eq. 2.3
Where:
Z = z-score

K = Kappa Coefficient for classification i (Cohen, 1960)
Via = Variance associated with Kappa Coefficient i

The classification results associated with Investigations 1 through 3 are
summarized in Table 2.7, which is partitioned into sections associated with each
investigation. In most instances, there is only a minor difference between an
overall-accuracy value and its paired Kappa Coefficient, indicating that chance
agreement played an extremely small role in the classification results presented

in Table 2.7.

61

2.3.1 ISODATA - Investigation 1

The first confusion matrix shown in Appendix B (i.e. saglm-ISODATA-
48classes) presents the results of the single ISODATA classification outlined in
Chapter 2. The overall accuracy for this optimal ISODATA classification is
62.72%, which is approximately 25 percent lower than the best-performing,
supervised classification outcome. Because the generated matrix parameters
were the result of the prioritization methodology described in section 2.2, their
class-specific meaningfulness is reduced in comparison to the other
classification matrices. For example, one could conclude from the provided
producers-accuracy values that the ISODATA classifier was able to definitively
identify open water areas (i.e. column 12), but was completely unable to
identify class 11, a deep submergent (i.e. Vallisneria americana). Examination of
columns 11 and 12 reveals that 100 percent of the pixels in both the 11th and 12th
ground-truth class fell within a single ISODATA class. Therefore, unless they
were combined into a more heterogeneous class (e. g. water-dominant), either 36
or 47 pixels must represent a classification error. The prioritization scheme
described in section 2.2 mandated that the highest magnitude value (i.e. 47) was
assigned to the overlapping ISODATA class so that the ISODATA matrix
represents the ”best” classification outcome. Further class breakdown or
”cluster-busting” is generally used to increase the interpretability of

overlapping data, and is an iterative process of applying binary masks (0=other,

62

 

 

 

 

 

 

Overall Kappa

Classification Code Accurag Value Inv.
Sagl m_MAXLIKE_70_24classes.cls 86.81 % 0.86 H
Saglm_SAM_25_24classes.cls 76.85% 0.7555 I-1

g1m_ISODATA_48classes.cls 62.72% 0.6023 I—1
Saglm_PCA654321_MAXLIKE_70__24classes.cls 89.79% 0.8921 I-2
Sag1m_PCA65432_MAXLIKE_70_24classes.cls 82.55% 0.8158 I-2
Saglm_PCA654321_SAM_25_24classes.cls 65.87% 0.6409 I-2
Sagl m_PCA65432_SAM_25_24cIasses.cls 60.85 % 0.5878 I-2
prentiss_PCA654321_MAXLIKE_70_20classes.cls 85.50% 0.8441 I-4
“prentiss_SAM_25_20classes.cls 71.75% 0.6966 I-4
**prentiss_PCA65432_MAXLIKE_70_20classes.cls 70.45% 0.6816 1-4
prentiss_PCA654321_SAM_25_20classes.cls 65.99% 0.6344 I-4
prentiss_SAM_PCA65432_25_20classes.cls 56.51 % 0.5311 I-4
prentiss_MAXLIKE_70_20classes.cls 44.42% 0.4046 I-4

 

 

** = lack statistical significance (95%) when paired

Table 2.7. Classification Accuracy Summary.

1=difficult to interpret classes) and re-classifying the original image with
respect to these masks (Jensen, 1996; p. 238). This process is generally repeated
until no newly interpretable classes are identified. Such an approach was
applied to the baseline ISODATA classification in the hopes of improving the
interpretation of several classes. Unfortunately, this process only acted to
partition the ”confused" classes into a greater number of classes, and did little
to improve their delineation. Thus, only the baseline ISODATA results were
presented. The presented ISODATA classification represents the best-

performing, unsupervised classification with respect to the restructured

63

preliminary matrix. Because ISODATA under-performed the other classifiers,

the specifics of its matrix were not further explored.

2.3.2 Maximum-Likelihood - Investigation 1

 

The Maximum-Likelihood classifier assigns each pixel into a spectral class
based on the Euclidean distance between it and all class centroids (means)
within n-dimensional data space (Jensen, 1996; p. 228). Within regions of
overlap, Maximum-Likelihood considers the frequency distribution (assumed
to be normal) related to each spectral class, and performs class assignments
through the examination of relative probabilities. The unaltered confusion
matrix associated with the baseline, Horseshoe Bay, Maximum-likelihood

classification (i.e. sag1m_MAXLIKE_70_24classes.cls) is shown in Appendix B.

The overall classification accuracy is 86.81%, which is nearly 10 percentage
points higher than the second best, Investigation 1 classifier (i.e. SAM). The
highlighted diagonal entries represent the number of ground-truth pixels that
:vere correctly classified. The comparison of these values to the bold-typeface,
column totals illustrates the consistent classification accuracies associated with
the Maximum-Likelihood classifier as applied in this research. The producers-
accuracy values shown indicate that 13 of 24 classes were classified at levels

higher than 95%, indicating that these botanical ROIs were well differentiated.

Examination of Table 2.7 reveals that Maximum-Likelihood is the best
performing classifier throughout Investigation 1, but a few classes displayed
extremely poor classification outcomes. Figure 2.3 presents nine spectral
signatures captured with a hand-held spectroradiometer in order to better
illustrate the weaknesses of Maximum-Likelihood. If the area of Figure 2.3 is
viewed as a decision-space, then the magnitude (i.e. vertical) separation of the
3rd and 4th signature is small although their shapes are different. The
magnitude-based decision rules utilized by Maximum-Likelihood would likely
confuse these two classes despite their different shapes because of their

”magnitude -overlap”.

Class 15 (i.e. muck) within the sag1m_MAXLIKE_70 confusion matrix
further illustrates this idea of magnitude-based, class confusion. This training
class was void of vegetation, and exhibited suppressed reflectance across the
visible and NIR (signature 1 in Figure 2.3). The matrix values associated with
this class reveal that one 1 of 27 pixels was correctly classified, a producers-
accuracy of 3.7%. The 26 remaining pixels fell in six other classes, all of which
were submergent or standing water/ emergent classes. The spectral nature of
these six classes was also suppressed due to the presence of significant amounts
of water and dark substrates (i.e. muck, not sand). Within the context of the
decision space, they are all within the same magnitude-class as the muck

training class, and were subsequently cross-classified. The relatively subtle

65

 

 

 

 

60
a) - , i , L‘L' ,g_1.—.: _
8 I". .
:40-. , , r- '4 ,8. , i i i
.53 I
8 m I __ .-ge-TTTT-ff:~_~4.i6
e ,---.—-— ————— — ,
Q) ’,” u ..7
mm ”___,/’ , 4L - _. m m .LL
°\° / ..... -—- ------ 5
1° “ m * :-:‘-:...‘..;..;a ’2’ ___ W“ “4“
0 —_—_._:\:.::::--:Z.:}"ZZIIVZ’.'ZI;2__
mmmwaneaommansmansso
nanorretets

Figure 2.3. Nine Representative Spectra.

reflectance features that separate these signatures are not large enough within
the context of the total decision space (i.e. 0—60% reflectance, 18 bands) to be

differentiated by the Maximum-Likelihood classifier.

2.3.3 Spectral Angle Mapper — Investigation 1

The confusion matrix associated with the baseline, Horseshoe Bay, SAM
classification (i.e. sag1m_SAM25_24classes.cls) is also shown in Appendix B.
The overall classification accuracy is 76.85%, which is nearly 10 percentage
points lower than the results associated with the Maximum-Likelihood

classification. Again, the highlighted diagonal entries represent the number of

66

ground-truth pixels that were correctly classified. The comparison of these
values to the bold-typeface, colurrm totals illustrates the classification accuracies

associated with the SAM classifier as applied in this research.

The producers accuracy values shown indicate that 6 of 24 classes were
classified at levels higher than 95 %, which is half of that generated from the
Maximum-Likelihood classifier. Overall, 15 of 24 classes displayed accuracies
below those of the Maximum-Likelihood classifier, with only 4 classes
indicating significantly improved classification results. SAM cross-classified
groups with similar greenness without regard to the magnitude-related
differences that the Maximum-Likelihood classifier capitalized on to generate
better results. ”Relative-greenness” errors might be expected within the context
of this research, since many wetland plants exhibit the characteristic peak-and-
valley shape. A shape-based classifier like SAM would likely not perform well
when confronted with training classes containing similar shapes. If this is true,
two patterns should be evident in the baseline SAM confusion matrix: 1) classes
that displayed better (i.e. > 6 pixels) classification results under SAM would be
those that the Maximum-Likelihood classifier confused using magnitude-based
decision rules alone, and 2) those classes that displayed significantly worse (i.e.
> 6 pixels) classification results under SAM should be similarly-shaped classes
that were able to be better separated with magnitude-based, Maximum-

Likelihood decision rules.

67

Four classes (13, 15, 16, and 23) displayed significant classification
improvement (i.e. > 6 pixels) over the Maximum-Likelihood classification. Of
these four, class 23 best illustrates the strength of the SAM classifier as
configured in this research. Class 23 represents a short stature, emergent
species (Eleocharis sp.) distributed over a sandy substrate. This class is very
similar from a biophysical perspective to class 22, which is the same species and
substrate with a grass species (Leeria oryzoides) sparsely intermixed. One would
expect these two signatures to be similar in both shape and magnitude
considering their related cover types. SAM was able to effectively separate
these two classes that Maximum-Likelihood classified poorly. This illustrates
the power of SAM with respect to subtle, relative-greenness reflectance features

that alter the shape (i.e. angular summation) across n-dimensional space.

Class 21 illustrates the inability of the SAM classifier to capitalize on
magnitude differences. This class is one of several graminoid species that
dominated the drier, sandy areas of Middle-Ground Island during the 2000 and
2001 growing seasons. As one might expect, many of these grass species
display relatively green, similarly shaped curves due to their biophysical
similarities. The Maximum-Likelihood classifier was able to separate this class
(Calamagrostis canadensis) from its closely related neighbors with little error.
The SAM classifier, on the other hand, confused this class with other relatively

green signatures, one of which was an emergent, non-graminoid (i.e. Class 10).

68

The magnitude suppression that generally is associated with water-present,
emergent signatures had little bearing on the SAM classification outcome that
only references spectral shape. Thus, despite being completely separated in
magnitude (i.e. y-axis shift), the SAM classifier cross-classified these similarly
shaped classes (i.e. 21 and 10). The overwhelming majority of the errors found
in the baseline SAM matrix are attributable to these types of greenness

classification errors.

Maximum-Likelihood classifications are better suited for coastal wetland
applications due to the magnitude separation (i.e. y-axis) of the predominant
plant communities and genera. The strength of the SAM classifier is found in
its ability to distinguish classes based on subtle shape changes, albeit without
reference to demonstratively important, magnitude differences. A viable
approach might utilize the SAM classifier to help separate only those classes
that Maximum-Likelihood was unable to differentiate, potentially taking

advantage of the relative strengths of both methodologies.

2.3.4 PC-rotated Classifications - Investigation 2

 

The confusion matrices from Investigation 2 are presented immediately
following their associated baseline results in Appendix B. The classification
accuracy associated with the Maximum-Likelihood classifier increased by

approximately 3 percent to an overall-accuracy of 89.79% after PC-rotation of

69

the imagery. PCA acts to redistribute target variance across orthogonal, none-
related dimensions. This redistribution is evident within the producers-
accuracy values shown in the PCA654321, Maximum-Likelihood matrix.

Only 4 classes (4, 6, 16 and 23) displayed improvements in this PCA-based
classification when compared to the analogous Maximum-Likelihood
classification of the baseline imagery. These classes experienced very
significant improvements, but at the expense of 8 other classes whose
producers-accuracies were slightly reduced. This result is consistent with the

more equal distribution of target variance that is characteristic of PCA.

In order to better illustrate the role PCA plays with respect to the
classification results presented in this research, 24 PCA-transformed, spectral
signatures are shown in Figure 2.4. The data were truncated to not include
dimensions 5 and 6 because there was little dimensional variation beyond
dimension 4, which is consistent with the variance percentages presented in
Appendix A. The characteristic peak-and-valley shape of botanical spectra is
no longer present in the PC-rotated signatures. The signatures have been
redistributed across the six PC dimensions, with the first three containing most
of the visible variation (see Appendix A). The fan-like arrangement of
signatures found between the first two dimensions represents a more equitable
distribution of target radiance, and little signature overlap is present.

Reduction in signature overlap would likely improve Maximum-Likelihood

70

0.15 1 \

 

 

c
H
‘ I
,9
vl‘
,//’//
//

Component Loadmgs
/
"

b

a
\\.3:\_.

(“TH
" lit
, ‘
.MI
; 5 M

C
8
\I
it),

 

b
...1

i
\

 

 

,5
5i

 

PCA Dimension
Figure 2.4 PC-transformed Spectral Signatures.

(i.e. magnitude-based) classification results while having little effect on SAM

(i.e. shape-based) results.

The unrotated, Maximum-Likelihood classification resulted in very high
classification accuracies. Thus, one can conclude that the overwhelming
majority of the ground-truth spectral classes were distinguishable based on
reflectance magnitude differences. This being the case, the PC-rotation was not
likely to improve classification accuracies greatly, as evident by the slight

improvement of approximately 3%.

71

The shape or angular complexity (i.e. vectors between successive bands)
has been reduced to the hinge-points associated with the 2nd and 3rd PC-
dimensions. A shape-dependent classification tool like SAM would suffer from
this reduction of angular complexity. The overall-accuracy of the SAM
classification of the PC-rotated imagery was reduced to 65.87%, placing it
among the worst performing classification results found throughout

Investigations 1 and 2.

Maximum-Likelihood and SAM classifications generated from PC-rotated
imagery without the contribution of dimension 1 significantly under-performed
their 6-dimension un-rotated counterparts (i.e. 7% and 5%, respectively). It was
initially hypothesized that the first PC dimension could be removed without
significantly impacting botanical differentiation because PC1 captured the
predominant commonality found in the various cover types. Removal of this
”commonality” dimension, it was surmised, would place added emphasis on
the lower-order dimensions that capture subtle-differences information.
Obviously, this hypothesis must be rejected. PC1 is required in order to

optimize either Maximum-Likelihood or SAM classifications.

2.3.5 Prentiss Bay Classifications — Investigation 3

 

The third tier of results presented in Table 2.7 summarizes the classification

accuracies related to all of the Prentiss Bay classifications. The baseline

72

confusion matrices associated with Investigation 3 are shown in Appendix B.
The ”best” classification methodology identified within Investigations 1 and 2
(Horseshoe Bay) was the Maximum-Likelihood classifier applied to the PC-
rotated imagery, although only a few percentage points separated the overall-
accuracies associated with the baseline (86.8%) and PCA654321 (89.8%)
Maximum-Likelihood classifications. In contrast, the overall accuracy of the
Prentiss Bay Maximum-Likelihood classification was 44.42%. This was the
lowest accuracy level recorded for all of the Prentiss Bay image classification
trials. This surprising lack of performance initially called into question the
”best” classifier identified in Investigations 1 and 2. In light of these results,
three SAM classifications were generated via the Prentiss Bay imagery s (i.e.
baseline, PCA654321, and PCA65432) as a means to cross-validate the findings
of Investigations 1 and 2. The overall-accuracies and Kappa values of the SAM
classifications were very similar to the SAM results from Horseshoe Bay. These
similarities indicate that there is a characteristic inherent to the AISA imagery

that caused Maximum-Likelihood to perform so poorly.

It was established earlier that dimensions generated via PCA are
independent (i.e. orthogonal), as well as less noisy because systematic noise is
sequestered into one or more lower-order dimensions (Lillesand and Kiefer,
2000; p. 520, Jensen, 2000; p. 179). Thus, significant Maximum-Likelihood

improvements should result from PC-rotation in reference to systematically

73

noisy imagery like the baseline, Prentiss Bay data. The overall-accuracy
associated with the Maximum-Likelihood classification of the 6-dimension (i.e.
PCA654321) Prentiss Bay imagery was 85.50%. Although not as high as its
Horseshoe Bay counterpart, this level of accuracy is comparable to the I-1
results. The primary role of the Prentiss Bay image analyses within this
research was to provide a second set of imagery (differing in both sensor type
and location) in order to test the transferability of select methodologies. The
overall PC-rotated accuracies presented in Table 2.7 indicate that the
methodologies applied to the Horseshoe Bay imagery are indeed applicable at
ecologically different locations. Investigation 3 also pointed out that the
characteristics and idiosyncrasies of each sensor can potentially play a
significant role in classification performance. This recognition reinforces the

importance of PCA in wetland image classification.

2.4 Summary and Conclusions

It was established in previous chapters that coastal wetlands could be
botanically complex, exhibiting a rich mosaic of plant species / assemblages that
are distributed along the predominant hydrologic gradient. It is this
characteristic that typically makes wetlands a challenging remote sensing target
due to the relative spectral similarity of the predominantly vegetated landscape.
Botanical training sites (Regions-of-Interest) were delineated that represented

the predominant cover types found in each study site during the 2000 and 2001
74

field campaigns. These ultimately served as the basis for multiple classification

comparisons.

The probability-based Maximum-Likelihood classifier was the best
performing classifier with respect to the differentiation of the field-delineated
ROIs. Approximately 87% of the accuracy assessment pixels (easterly ROIs)
were correctly classified by Maximum-Likelihood, which was nearly ten
percent higher than the comparable SAM classification. It was shown that
probability-based classifiers are magnitude-dependent (y-axis separation), and
their associated errors are typically associated with within-class
misclassifications. Maximum-Likelihood displayed the greatest potential for
differentiating the complex botanical communities found in the Lake Huron
wetlands analyzed in this study. From a practical standpoint, the Maxirnum-
Likelihood classifier would be ideal for delineating major botanical classes
when some degree of within-class error (e. g. confusion of similarly structured

emergent species) can be tolerated.

The baseline SAM classifier under-performed the Maximum-Likelihood
classifier by nearly ten percent. This physically-based classification tool tended
to confuse classes that displayed similar greenness (i.e. shape of the reflectance
signature), and differed from Maximum-Likelihood in that it is insensitive to

magnitude differences. Both the Les Cheneaux and Saginaw Bay imagery was

75

taken as the botanical community neared the end of its growing season, which
likely contributed to the success of the SAM classifier because the plant
community was in various stages of dieback. Imagery taken when the
community was at ”peak green” would likely lead to an increase in greenness

‘

confusion by the SAM classifier.

The strengths of both classification tools could be used in combination to
enhance classification accuracies. Maximum-Likelihood should be used to
parse an image into the predominant classes. Then, complex classes could be
isolated via image masking in order to capitalize on the ability of the SAM
classifier to differentiate similar spectra based on relatively subtle spectral
features that are lost in the magnitude-based approach. This is certainly an area

that needs to be explored further within the context of coastal wetlands.

Covariance-based PCA rotation proved to consistently enhance the
performance of the probability-based classifiers. From a practical standpoint,
this research indicates that PCA rotation should routinely be utilized in
combination with probability- based classification tools. The resulting
dimensionally-reduced signatures are better differentiated by the magnitude-
sensitive classifiers because the image spectra are more equitably distributed
(i.e. more magnitude separation). A more detailed examination of Figure 2.4

revealed that biophysically relevant clusters of spectra do indeed exist,

76

especially between the second and third dimensions. These clusters (e.g. see
dashed spectra) represent similar within-class spectra that proved to be the
underlying weakness of the Maximum-Likelihood classifier. It is these tightly
clustered groups within the rotated spectra that would best be classified via

SAM.

The dimensional reduction that assisted the probability-based classifier (i.e.
Maximum Likelihood) significantly reduced the effectiveness of the SAM
classifier. The angular coefficients that form the foundation of the SAM
classifier are determined by summing the angular complexity of spectra across
all bands. A reduction in bands would inherently reduce the likelihood that
diagnostic differences would be maintained. The inability of the SAM classifier
to utilize PC-rotated imagery creates a logistical concern when the strengths of
both it and a probability-based classifier are to be combined. Ultimately,
rotated and non-rotated imagery would have to be maintained and identically
masked in order to implement a combined (Maximum Likelihood and SAM)

classification strategy.

The Prentiss Bay imagery was incorporated into this research in order to
test the transferability of ”methods” developed with respect to the Horseshoe
Bay, 1-meter imagery. Generally, the outcome of like-configured Prentiss and

Horseshoe Bay classifications were very similar, and trends were common

77

throughout. The PC—rotated Maximum-Likelihood and un-rotated SAM
classifications as outlined in this research were not site dependent, and could
likely provide the basis for basin-wide wetland mapping, especially in

combination.

Although not completely understood, it seems likely that some type of
systematic radiometric noise was inherent to the AISA imagery, as evident by
the poor SAM and Maximum-Likelihood classification results generated via the
raw AISA imagery. PC-rotation is known to sequester such systematic noise
into lower-order dimensions, and this was indeed the case in this research as
evident by the sharp increase in the accuracy of the Maximum-Likelihood
classification (41 percentage points) after PC-rotation of the AISA imagery.
These results suggest that results generated from a probability-based classifier
in conjunction with both raw and rotated imagery could function as a viable

test for subtle, system-level noise.

78

Chapter 3: Principal components and second derivative
analyses of hyperspectral reflectance data from freshwater
coastal wetlands.

3.1 Introduction and Rationale

In the 19705, state and federal agencies increasingly looked towards aerial
photography for the purposes of wetland mapping and inventory (Gross et al.,
1989; p. 475). Film types that are sensitive to NIR electromagnetic energy are
generally preferred for wetland applications because they are more effective in
detecting standing water and determining biomass, as well as differentiating
wetland and upland vegetation (Hardisky et al., 1986). Aerial photography
(1:1000 to 1:15,000 scales) has been found to be a fast, accurate, and relatively
inexpensive method by which wetland boundaries, botanical changes, and
aerial extent can be recorded (Mitsch and Gosselink, 2000; p. 107). Two
characteristics limit the applicability of aerial photography to detailed
vegetative mapping strategies: 1) the limited amount of spectral information
contained in each image (Jensen et al., 1986), and 2) its non-digital format

(Hardisky et al., 1986).

The launch of the Landsat-1 satellite (19705) marked the practical beginning
of spaceborne, earth resource remote sensing. Over the ensuing thirty years, a

variety of systems, with varied spatial and spectral resolutions, have been

79

developed. Orbiting satellite-based sensors can provide near-simultaneous
coverage of wetland imagery over large geographic areas recorded in digital
format (Gross et al, 1989; p. 476). Satellite platforms continue to provide an
overwhelming array of earth resource data when fitted with remote sensing
systems, and the status of these technologies continues to change as scientific
advances yield improved spacecraft and sensors (Lillesand and Kiefer, 2000; p.
373). However, regarding their utility for detailed wetland characterization,
current systems are limited by their spatial and / or spectral resolutions. These
limitations become obvious in the context of the two fundamental assumptions
of this research: 1) the spatial heterogeneity of coastal wetlands is best captured
by digital imagery with a spatial resolution of 5 meters or smaller, and 2)
differentiation of wetland plant communities is best achieved through the
analysis of hyperspectral reflectance records. Accepting these assumptions
means that aerial photography and currently available satellite imagery are

sub-optimal for detailed wetland mapping.

The development of imaging spectrometers created interest in utilizing
spectral reflectance data of vegetation canopies measured from the ground to
aid in remotely differentiating plant communities. In fact, many next-
generation, satellite-based systems (e.g. COIS, Hyperion, MODIS or Warfighter-
1) incorporate hyperspectral sensors in order to address the spectral limitations

of aerial photography and current satellites systems (ASPRS, 1995; Stoney and
80

Hughes, 1998; Thenkabail et al., 2000). A major advantage of imaging
spectrometry lies in the potential for integrated analysis comparing calibrated
remotely sensed imagery with field spectral measurements (Goetz 1992; Kruse
et al, 1993). The use of systematic differences in canopy pigmentation and
related chemistry to classify to the botanical species level has recently emerged

as a promising new approach.

Recent literature has shown that narrow, strategically placed bands may be
crucial in providing additional information with significant improvements over
broad bands in quantifying biophysical characteristics of agricultural crops
(Thenkabail et al., 2000). Although not widespread, the wetland literature
contains recent examples of investigations that utilized hyperspectral systems
to map and monitor plant communities (Savastano, 1984; Zacharis, 1992;
Forsyth, 1998; Borstad Associates, 1995; Jupp, 1994; Ritter and Lanzer, 1997;
Brown and Borstad, 1999; Wang, 2000). Common to these investigations is the
exploitation of improved spatial or spectral resolutions to more effectively
differentiate wetland vegetation, although none explore the independent

contribution of spatial and spectral resolution.

Derivative techniques have long been applied in remote sensing, and have

been found to eliminate background signals, differentiate overlapping signals,

81

and reduce the affects of turbidity in aquatic chlorophyll investigations
(Demetriades-Shah, 1990). Higher-order (i.e. 2nd-order) derivatives become
important within the context of remote sensing because relative reflectance

amplitudes (vertical stretch) increase with derivative order (Fell and Smith,

1982)

In remote sensing, feature extraction involves the identification of those
statistical characteristics of remotely sensed data that capture the most
systematic variation (Campbell, 1996; p. 288). Systematic variations, as
opposed to non-systematic variations (i.e. noise), ultimately provide the
foundation for target differentiation. Within multispectral data, often there is
significant correlation between the information contained in closely adjacent
bands. This is especially true with hyperspectral data sets composed of
narrow, contiguous bands. In principle, feature extraction can significantly
reduce the magnitude of a dataset by isolating essential components (i.e.
information) while discarding redundant or noisy data (Campbell, 1996; p.

288)

A feature extraction tool commonly applied throughout the scientific
community is Principal Components Analysis (PCA). PCA has been effectively

applied in the analysis of multispectral data, resulting in more interpretable

82

data via the reduction of large image data sets into relatively few meaningful
dimensions (Jensen, 1996; p. 172). In essence, the purpose of PCA is to
compress all the information (correlation / variance) contained in an original
data set into fewer, independent (orthogonal) channels - the principal
components (Lillesand and Kiefer, 2000; p. 518). A detailed description of the
statistical framework used to perform PCA is beyond the scope of this research,

although a more complete explanation can be found in (Davis, 1986).

The goal of this component of my research is to perform derivative and
Principal Components Analysis in order to identify those (hopefully few)
recorded bands that are most appropriate for the differentiation of vegetation
types within the coastal wetlands of the Great Lakes. This goal addresses the
second research objective outline in Chapter 1; Quantitatively examine the
properties of in-situ reflectance data to discern the optimal spectral framework

of wetland imagery.

3.2 Methods

3.2.1 Spectral Data Pre-Processing

 

In-situ reflectance measurements were collected for the predominant
species, species assemblages, and cover types at both the Prentiss and

Horseshoe Bay study sites. All spectral data were collected during the 2001

83

growing season (J uly-September). These nadir measurements were taken with
an SE-590 spectroradiometer, outfitted with an 11° field-of-view optic (Spectron
Engineering, Inc. Denver, Colorado), approximately 1 meter above the canopy.
At a sensor height of 1 meter, a circular area with a diameter of 18 centimeters
was sampled. The SE-590 recorded 252 contiguous bands from 365 to 1,125
nanometers, with an approximate bandwidth of 3 nanometers. Prior to the
field campaign, the SE-590 was cross-calibrated with a newly-calibrated
FieldSpec-Pro spectroradiometer from Analytical Spectral Devices

(www.asdi.corn) to ensure the integrity of the spectral signatures being

 

captured.

The SE-590 was pre-programmed to capture eight consecutive
measurements within a few seconds; these were then internally averaged to
yield raw radiance data. Nearly 100 spectra were recorded as summarized in
Table 3.1. Whenever feasible, duplicate or triplicate measurements were made
and averaged in order to characterize additional target variability. The SE-590
was connected via its RS-232 port to the serial port of a laptop computer
running a custom software program. This software package allowed the
reflectance data (raw radiance) from the plant canopy and a reflectance
standard (i.e. reference panel) to be stored in a text-based export file after being

normalized for differences in integration time.

84

 

 

 

 

Sit ate: Code: Description:
Horse / 824-23 HB-l A VG H. dubia
Horse / 824-23 HB-2 H. dubia (surface)
Horse/824-23 HB-3 A VG Chara sp. (10 in)
Horse/824-23 HB—4 A VG S. rigida/Chara sp. (30/70)
Horse / 824-23 HB-5 A VG S. rigida/H. dubia (30/70)
Horse / 824-23 HB-6 A VG S. rigida/Chara sp. 60/40
Horse/824-23 HB-7 A VG S. rigida/Chara sp. (80/20 -15 in water)
Horse / 824-23 HB-8 A VG sand/Chara sp. 8 in (95/5 18 in.)
Horse / 824-23 HB-9 Sand (clouded H20)
Horse/824-23 HB—10 Sand/Chara sp. 8 in (95/5 18 in.)
Horse/824-23 HB—11 A VG V. americana/Sand (70/30 - 12 in.)
Horse / 824-23 HB-12 A VG L oryzoides (90% cover)
Horse / 824-23 HB-13 U. dioica
Horse / 824-23 HB—14 I. capensis (orange blooms)
Horse/824-23 HB—15 A VG S. validus (80% - bloom)
Horse/824-23 HB-16 S. validus/Eleocharis sp. (75fl5 - browning)
Horse / 824-23 HB-17 S. validus/Sand (30/70)
Horse/824-23 HB-18 S. validus/sand (30/70 - 2 in.)
Horse/824-23 HB—19 Chara sp. (3 in.)
Horse/82423 HB-20 A VG Chara sp./S. graminia (80fl0 - 6 in.)
Horse/824-23 HB-21 A VG S. validus (85 % - clumps in 8 in.)
Horse / 824-23 HB—22 T. angustifolia (75% - 4 in. sandy)
Horse / 824-23 HB-23 A VG T. angustifolia/S. rigida/algae (40/40/20)
Horse/907-9 HB-24 A VG P. australis (brown heads)
Horse/ 907-9 HB-25 Bidens sp.
Horse/ 907-9 HB-26 A VG E. maculatum (headed out white)
Horse/907—9 HB-27 T. angusttfolia (50% green, windblown)
Horse / 907-9 HB-28 A VG S. americana (70% cover)
Horse / 907-9 HB-29 A VG S. lahfolia (large leaf)
Horse / 907-9 HB-30 A VG P. amphibium/S. latrfolia (30/70)
Horse/907-9 HB-31 Bidens sp/ P. australis
Horse / 907-9 HB-32 A VG gravel at Ramp
Pre/ 703-25 PB-l A VG T. angustifolia (deadfall)
Pre/ 703-25 PB-2 A VG S. acutus/T. An gustifolia

Table 3.1. Collected In-situ Spectral Signatures.

85

 

 

 

 

Pre/ 703-25 PB-3 A VG S. acu tus (deadfall)

Pre/ 703-25 PB—4 A VG N. advena (mud, no H20)

Pre/ 703-25 PB-5 A VG Chara sp./P. natans

Pre/ 703-25 PB-6 S. acutus (H20)

Pre/ 703-25 PB—7 A VG N. advena (H20)

Pre/ 703-25 PB-8 A VG P. natans (H20)

Pre/ 703-25 PB-9 H20 (stirred)

Pre/ 703-25 PB-10 H20 (clear)

Pre/ 703-25 PB-ll Dolomite

Pre/ 703-25 PB-12 M uck/mud

Pre/ 703-25 PB-13 Blowdown

Pre/ 703-25 PB-14 A VG C. lasiocarpa

Pre/ 703-25 PB—15 A VG C. stricta

Pre/703-25 PB-16 Potentilla sp.(scrub)

Pre/ 703-25 PB-17 C. lasiocarpa/ S. acutus

Pre/ 703-25 PB-18 Salix sp. (shrub)

Pre/ 703-25 PB-19 juncus sp.

Pre/ 703-25 PB-20 S. latifolia (mud)

Pre/ 703-25 PB-21 mud

Pre/ 703-25 PB-22 gravel

Pre/ 703-25 PB-23 Asphaul t

Pre/ 703-25 PB-24 Helianthus sp.

Pre/ 703-25 PB-25 Sci rpus sp.

Pre/827-39 PB-26 C. viriduIa/dry muck soil (80/20)

Pre/827-39 PB-27 C. viridula/dry muck soil (90/10)

Pre/827-39 PB-28 I. canadensis/ dry muck soil (70/30)

Pre/ 827-39 PB-29 ]. canadensis/dolomite rock (80fl0)

Pre/ 827-39 PB-30 C. vi ridula/muck soil (10/90)

Pre/827-39 PB-31 C. viridula/saturated muck soil (5/95)

Pre/ 827-39 PB-32 S. acutus (50/50 green/dead)

Pre/827-39 PB-33 S. acutus (90/10 green/dead - 1 in.)

Pre/ 827-39 PB-34 S. acutus (70/30 green/dead)

Pre/ 827-39 PB-35 C. acu tus (60/40 green/dead)

Pre/827-39 PB-36 T. angustifolia (medium density, sat. soil)

Pre/827-39 PB-37 T. angustifolia/HZO (50/50, 80/20 green/dead)
Table 3.1. Continued.

86

 

 

 

 

Pre/827-39 PB—38 A VG P. natans/muck soil (85/15 no H20)
Pre/ 827—39 PB-39 Nuphar advena/muck soil (80fl0 4 in. H20)
Pre/827—39 PB-40 N. advena/muck soil (60/40, 4in H20)
Pre/827-39 P841 3. acutus (dense, muck soil, 4 in. H20)
Pre/827-39 PB—42 S. acutus (medium, muck, 4 in. H20)
Pre/ 827-39 PB-43 H20 (sandy/muck substrate, 2 in. depth)
Pre/827-39 P344 P. natans/N. flexilis (60/40 - 4 in. H20)
Pre/827-39 PB-45 N. flexilis (95/5 - 8in H20)
Pre/827-39 PB-46 H20 (muck, 8-10 in. depth)
Pre/827-39 PB-47 S. acutus (meduim, 8 in. H20)
Pre/827-39 PB-48 T. angustifolia (sparse, 8 in. H20)
Pre/827-39 PB-49 ' T. angustifolia/S. acutu (sparse, 6 in. H20)
Pre/ 827-39 PB-50 C. lasiocarpa/Cladium mariscoides/S. acutus
Pre/827-39 PB-51 C. lasiocarpa (20 S.acutus stems/mAZ)
Pre/ 827-39 PB—52 A VG C. lasiocarpa (20/80 green/dead)
Pre/ 827-39 PB-53 C. lasiocarpa (10/90 green/dead)
Pre/ 827-39 PB-54 C. stricta (15/85 green/dead)
Pre/ 827-39 PB-55 C. stricta (20/80 green/dead)
Pre/ 827-39 PB-56 C. stricta (30/70 green/dead)
Pre/ 827-39 PB-57 C. stricta (40/60 green/dead)
Pre/ 827-39 PB-58 Alnus sp./C. stricta (90/10)
Pre/ 827-39 PB-59 C. stricta (5/95 green/dead)
Pre/827-39 PB-60 C. lasiocarpa (30 S.acutus stems/mAZ)
Pre/ 827-39 PB-61 C. lasiocarpa (sparse T. an gustifolia)
Pre/ 827-39 PB-62 Solidago rigida
Pre/ 827-39 PB-63 road gravel
Pre/ 827-39 PB-64 Asphaul t

Table 3.1. Continued

87

 

The reference panel used for the in—situ data collection was a 14-inch square
aluminum panel coated with polytetraflouroethylene (PTFE). Reference panel
(Panelmsu) measurements were made before and after each series of
measuremenqts by holding the SE-590 sensor (nadir-looking) 0.5 meters above
the panel to ensure that the 11° field-of—view fell completely upon the surface of
the relatively small reference panel. These panel measurements served as
endpoints for a linear interpolation that provided a reference panel

measurement specific to each botanical measurement and unique solar angle.

The captured radiance values were transformed into values of absolute

reflectance by reference to radiance measurements taken from a calibrated

reference panel (Equation 3.1). A reflectance coefficient (paw...) was generated

Pin-situ = ETin-situ/ ETinterpolated * Pamsu (for bands 1 through 252) (Eq. 3.1)

ETimim = In-situ target radiance
ETinterpolated = Linearly interpolated reference panel radiance
puma = Reference panel reflectance coefficient

pimsim In-situ target percent reflectance

for the in-situ reference panel utilized in this investigation via radiance
measurements (ETmsu) taken of it and a Spectralon panel (ETspec) under
Controlled darkroom lighting. A simple regression (R>.99) established that the
1'8flectance properties of Spectralon and the aluminum panel were indeed

linear. A 2nd-order polynomial was then fit to published Spectralon reflectance

88

coefficients (paspec @ 100nm increments). The resulting polynomial equation
was used in conjunction with SE-590 upwelling radiance (ET) measurements of
both the reference and Spectralon panels to calculate pamsu that is needed to
transform raw radiance signatures into percent reflectance. Because no effort
was made to characterize the bi-directional reflectance of these surfaces, three
steps were taken throughout both the calibration work and the in—situ data
collection in order to minimize its affects: 1) all in-situ data were collected
within 2.5 hours before or after solar noon, 2) only nadir, in-situ measurements
were taken, and 3) sensor / beam geometries were held as constant as practical

during all calibration measurements.

After each signature was transformed into percent reflectance, it was
imported into a spectral library within ENVI (i.e. Environment for Visualizing
Imagery) image processing software. Alphanumeric botanical keys were
created (e.g. HBl; Table 3.1), each corresponding to a specific in-situ
measurement, so that the integrity of the botanical data was more easily
maintained during the migration into ENVI. The transformed data were then
resampled to emulate the spectral characteristics of the hyperspectral imagery
(i.e. 48 bands, approx. 7nm wide) utilized in the second component of this
research. The 96 baseline spectral libraries (Table 3.1, 252 bands) were
resampled within the spectral tools module of ENVI, utilizing the band

center/ FWHM of the 48-hyperspectral bands. These resampled data, examples

89

of which are shown in Figure 3.1, provided the foundation for more detailed

data analyses.

3.2.2 2ml Derivative Approximations

Signatures like those shown in Figure 3.1 are made up of a series of line
segments established by reflectance values recorded at each band center.
Second-derivative values do not exist for line segments, so some type of an
approximation must be used. Second-derivative approximations identify
spectral locations where the slope of the reflectance curve experiences a
relatively abrupt change (i.e. points of inflection). Fundamentally, spectral

regions of abrupt change are inherently more diagnostic than areas with

60
50.
40

30

”/0 Reflectance

20

10

 

0 . .
400 450 500 550 600 650 700 750 800 850 900 950

nanometers

Figure 3.1. Representative Spectral Signatures

90

uniform slopes. A slope-based, numerical method was used to transform the
botanical spectra (Figure 3.1) into 2“d derivative approximations.

A modified version of the slope-based, two-step, derivative-approximation
method presented by Tsai and Philpot (1999) was utilized in this research
(Figure 3.2). The first step determined the slope of a line segment by simply
dividing reflectance value differences by the wavelength interval separating

them (rise / run) as indicated by Equation 3.2.

dlst = (pun- pn)/()\n+1 - An) where n = band number (Eq. 3.2)
dlst = 1st—derivative (line segment slope)

p = absolute (percent) reflectance

A = wavelength (nm)

Second derivatives were calculated similarly, although the equation was
modified so that the wavelength domain across all three data points was
included (Equation 3.3). In addition, a multiplier (0.5) was incorporated in the
denominator to enhance any reflectance amplitude differences (vertical
separation), making the visual differentiation of overlapping spectral signatures
easier. Due to the structure of equations 3.2 and 3.3, the spectral data associated
with the first and last band numbers are lost during the calculation of the 2nd
derivative approximation. This 2nd derivative method was structured so that
the derivative values generated fell exclusively on the center band of the band-

triplet referenced in each derivation (see Figure 3.2).

91

nm= 708.5

 

 

 

 

 

 

 

 

2 R= 6.25 nm= 731.5
R= 5.95
/. 4
3 Concave
1 R= 4.50 0 (anlection
nm= 720 omt
R= 3.52 Convex
nm= 697 - Inflection
Point

lst Derivative (6,25 - 3.25)

(5101*?) —> mmmmm = 0.2608

(Age segment (708.5-697)

lst Derivative (4.50 - 6.25)

(Slope) —> -------- = 0.1521

grate segment (720 - 708.5)

lst Derivative (5.95 - 4.50)

(Slope) —> --------- = 0.1260

Line segment (731.5 - 720)

3—4

2nd Derivative Approx. (4)1521 _ 0 2608)

9101394510139) —> ’ ' =-0.4129

me segmen -

@Point #2 (720 697) 05

2nd Derivative Approx. +

£5.10“? of 51059) _> (0.1260 0.1521) = 0.2781

me segmen
@Point #3 (731.5 - 708.5) 0.5

Figure 3.2. Derivative Calculation Schematic.

92

d2nd = (distnn - d‘lstn)/0.5(}\n+2 - An) where n = band number (Eq. 3.3)

d1st = 15t derivative
d2nd= 2nd derivative approximation
A = wavelength (nm)

The derivative values for each of the 96 spectra (Table 3.1) were placed in a
single spreadsheet to facilitate their analysis. Each paired column of derivative
and band-center values were simultaneously sorted in ascending order with
respect to derivative magnitude. This sorting method ensured that the band
center associated with each derivative value remained linked to that specific
value, thereby preserving the integrity of each data couplet. A second
spreadsheet function was then applied so that the appropriate sign (+/ -) of each
derivative was applied to its linked band center. The 20 largest-magnitude
derivatives (i.e. 10 positive and 10 negative) were extracted from the sorting
spreadsheet and placed into Appendix C. High magnitude derivative values
identify regions of diagnostic spectral change, and they should represent those
bands that are most appropriate for the differentiation of the input spectra (i.e.

wetland vegetation).

3.2.3 Principal Components Analysis

 

The statistical software package SYSTAT (wwwspsssciencecom) was
utilized to transform the 48-channel, in-situ data into their principal

components (i.e. dimensions). SYSTAT’s data reduction module accepts a

93

variety of two dimensional data arrays, the dimensions of which outline the
dependent (rows) and independent (columns) variables of interest. Band
number was established as the independent variable in order to characterize the
explanative power of all 48 bands with respect to the sampled, botanical
community (dependent variable). The information contained in each
resampled spectral library was exported via an ASCII text file into an EXCEL
spreadsheet, with each column relating to the spectral characteristics of one
botanical target and each row referencing a specific band center (wavelength).
This structure was opposite that needed by SYSTAT in order to test the
relationship between the dependent and independent variables outlined above.
Thus, the spreadsheet containing all 96 spectral signatures (Table 3.1) was
transposed in order to yield the appropriate column (band number) and row
(botanical reflectance) data structure. Eight rows of data containing the spectral
information derived from non-vegetative targets such as shallow water and
gravel were eliminated. The remaining 88 rows, signatures taken of areas with
at least some visible vegetation, were subjected to Principal Components

Analysis.

Correlation-based PCA references standardized input variables that have a
mean of zero and a variance of one. Standardization tends to inflate the
contribution of variables whose variance is small, and reduce the influence of

variables whose dimensions are large (Davis, 1986; p. 536). PC analyses based

94

on correlation matrices generate dimensions that emphasize those spectral
features that best capture predominant target similarities (Davis, 1986; p. 536).
Covariance-based PCA, on the other hand, is typically used when the relative
magnitudes of the variables are important because its un-standardized format
enhances magnitude differences and reduces the potential for an insignificant
variable to exert a strong influence on the results (Davis, 1986; p. 536).
Principal components analyses based on covariance matrices generate
dimensions that emphasize those spectral features that best capture
predominant target differences. It was unclear which methodology (covariance
or correlation) was most applicable to band identification within the context of

this research. It was assumed that both provide a different, but meaningful,

perspective.

PCA was applied to the in-situ data in order to identify those bands (i.e.
independent variables) that are most appropriate (i.e. have the highest
component loadings) for capturing the predominant spectral differences within
the 88 reflectance signatures. The 88-signature data array was directly read by
SYST AT and an unrotated PCA output was generated using both the
correlation and covariance methodologies. The primary outputs of interest
were the two SYSTAT output arrays defined by the number of extracted factors

(dimensions) and related component loadings. The number of output factors

95

generated by PCA is typically held equal to the number of substantively
meaningful independent patterns (extracted features) among the variables
tested (Rummel, 1970). Generally, factors (dimensions) are considered
meaningful if their associated eigenvalues are at or above a value of one.
However, in order to remain consistent with the image-based PCA described in
Chapter 2, the first six factors were extracted without examination of their

associated eigenvalues.

The component loadings associated with each individual factor (dimension)
were copied into a spreadsheet along with their related band number. This
copying routine created an intermediate-calculation spreadsheet that contained
a series of columnar loading/ band number data couplets, one couplet for each
of the 12 extracted dimensions (i.e. two sets of six). PCA component loadings
represent a measurement of the relative degree to which each band explains the
relationship between any one factor and the body of dependant variables.
Therefore, a band associated with a relatively high loading explains much of the
correlation/ covariance captured by any one of six dimensions (factors). If a
factor captured that portion of the overall data variance that is inherently
related to the differentiation of botanical targets, then those bands loading
highest on that factor would be well suited for botanical differentiation. These

PCA-derived, band center/ loading couplets provided the foundation for two

96

independent methodologies aimed at identifying key band-centers (i.e.

wavelength domains) highlighted via PCA.

The band center/ loading couplets were further manipulated to yield a
graphical representation of these data. An XY-plot was generated that depicted
component loading values versus wavelength (i.e. band centers). This
methodology allowed the visual inspection of the loading response-curves
associated with the 12 extracted dimensions. The co-plotted, response curves
(i.e. loading graphs) resembled a series of spectral signatures in that they
displayed spectral features (e.g. peaks, valleys, local-minima, local-maxima)
across the visible and N IR wavelengths. Key wavelength features identified
through the visual examination of loading curves (especially those cormnon to
several dimensions) should be most applicable to the differentiation of the

botanical community from which the PCA-based signatures were generated.

The second methodology applied 2nd derivative approximations to the
component loading spectra in order to identify the band centers associated with
the predominant spectral features exhibited by these curves. The band centers
associated with the ten highest-magnitude, 2nd derivative approximations (i.e. 5
largest positive and 5 largest negative) were extracted for each dimension after
the sorting routine described above. This extraction resulted in 120 band centers

(i.e. 12 dimensions x 10 values) that were then sorted in ascending order. This

97

allowed a histogram-like graph of these values to be generated so that the most
botanically relevant (i.e. most frequently occurring) bands could be easily
identified. Key wavelength features identified through the derivation of the
loading curves (especially those common to several dimensions) should be
particularly useful in differentiating the botanical community from which the

signatures were generated.

3.2.4 Image Classifications - Bandset Op_timization

 

Chapter four explored the relative performance of a series of classification
methodologies in order to identify that which was most robust in reference to
the differentiation of vegetative wetland R015. A Maximum-Likelihood
classification generated with respect to PC-rotated imagery proved to be the
best-performing methodology. In some instances, the strengths of the SAM
improved the differentiation of similar spectral classes (i.e. l-meter ROIs). No
attempt was made to spectrally alter the 1-meter imagery in order to improve

or streamline classification results.

The 1-meter AISA and CASI-II imagery referenced throughout chapter 2
were inappropriate for the spectral image manipulations associated with
chapter 3 because their bandsets incorporate a series of non-contiguous bands
(i.e. spectral gaps) across the visible and NIR regions of the spectrum. Unlike

these 1-meter images, the 4-meter, CASI-II imagery (Table 2.1) contained 48
98

contiguous bands across the visible and NIR portions of the spectrum within no
spectral gaps. Band location and width could therefore be altered without fear

of encountering a spectral gap.

Table 3.2 summarizes the 21 ground-truthed, ROIs delineated within the 4-
meter, Horseshoe Bay imagery. Although similar, these ROIs differ in number
and pixel count from those delineated within the l-meter, Horseshoe Bay
imagery due to the method through which they were derived. When
displayed, the geo-rectified 4-meter and l-meter Horseshoe Bay imagery
overlapped each other for that portion of Horseshoe Bay captured in both data
sets. This overlap allowed the centroids of all pixels found in each of the 24, 1-
meter ROIs (Table 2.6) were converted to an ENVI vector (point) layer. This
layer was then displayed on top of the 4-meter imagery, which allowed the 4-
meter pixels that fell within the boundary of the core 1-meter ROIs. In a few
instances, core regions were not available in the 4-meter imagery due to the lack
of image overlap. Whenever possible, these training classes were reestablished
elsewhere in the 4-meter image based on knowledge of the botanical

communities acquired during site visits.

The limited number of pixels identified via this vector-overlay procedure
was insufficient with respect to the two constraints associated with the

Maximum-Likelihood classifier, because it needs at least n + 1 pixels in each

99

training ROI, where n is the number of image bands. The 4-meter Horseshoe
Bay imagery has 48 bands, so at least 49 pixels were needed in each training
class ROI. The second constraint associated with ENVI’s Maximum-Likelihood
classifier is the need for a substantial amount of training class variance. To
address these constraints, a false color composite of the 4-meter imagery was
displayed in ENVI along with the core l-meter pixels identified via the vector-
overlay procedure. ENVI’s interactive histogram module was then activated so
that the relative intensity of the three color-channels (i.e. red = 916.7nm, green =
685.5nm, and blue = 560.1nm) could be manipulated. Each core ROI (Table 2.6)
was expanded by incorporating neighboring pixels that were visually similar as
revealed by color intensity manipulations. This process created 20 larger ROIs
necessary for the Maximum-likelihood classifier while preserving, as best as
possible, the spectral integrity of each training class. An additional class (wave
foam) was added because preliminary, 4-meter classifications indicated that
these pixels remained unclassified. These 21 4—meter ROIs (Table 3.2) provided
the training classes for a series of supervised image classifications generated
with respect to spectrally manipulated 4-meter images. Due to the nature of
PCA, rotated imagery could not be used in this investigation because a
complete 48-band image (i.e. non-rotated) was required. The lack of RT-DGPS
ground-verification of these expanded 4-meter ROIs mandated the
development of an alternative accuracy assessment strategy. A pixel-based,

accuracy assessment procedure was developed with respect to the image subset

100

 

 

 

 

 

ROI Matrix Code Scientific name Pixel #
1 Chara Chara sp. 231
2 scirpus-dense Scirpus validus 124
3 scirpus-clump Scirpus validus 217
4 typha-soi12 Typha angustifolia 53
5 sagrigida60/ 40 Sagittaria rigida 69
6 Beachsand N/a 216
7 shallowh20 N/a 144
8 hetdubia-delta Heteranthera dubia 63
9 sagrigida-delta Sagittaria rigida 92

10 Vallisneria Vallisneria Americana 207
11 deeph20 N/a 318
12 scival-brown Scirpus validus 96
13 saggram/charaROI Sagittaria graminea/Chara sp. 82
14 typha-cut-island Typha angustifolia 116
15 N uphar Nuphar advena 79
16 Eleocharis Eleocharis sp. 97
17 Impatians Impatians btflora 261
18 bluejoint grass Calamogrostis Canadensis 101
19 cutgrass-eleocharis Leersia sp./Eleocharis sp. 93
20 typha-phrag-scivalmix Typha sp./Phragmites australis/Scirpus sp. 101
21 wave foam N/a 60
Table 3.2. Horseshoe Bay, 4-meter Regions-of-Interest.

(250 x 450 pixels) displayed in Figure 3.3. A single Maximum-Likelihood and

SAM classification was generated from the unaltered 4-meter imagery, and all

of the 112,500 classified pixels contained in this rectangular region were treated

as ground-truthed data in contrast to the field-delineated ROIs presented with

respect to Investigations 1-3. These 112,500 evaluation-standard pixels were

assumed to accurately represent the botanical communities found within the

evaluation rectangle. This assumption could be made because its botanical

101

 

 

Figure 3.3. 4-meter, Horseshoe Bay Rectangular Image Subset.

102

validity is functionally irrelevant with respect to the focus of Investigation 4.
The class specific pixel counts established via the 4-meter, baseline (i.e. 48
bands) classifications served as benchmarks so that the relative change
associated with subsequent spectral manipulations could be assessed despite

the lack of ground verification.

Within all of ENVI’s classification modules, bands can be interactively
selected for use during an image classification. Through the use of this
interactive selection tool, a series of independent classifications were generated
by systematically reducing the number of bands referenced during otherwise
similarly configured Maximum-Likelihood and SAM classifications. The first
image manipulation reduced the number of bands to seven, as guided by the
predominant wavelength domains (i.e. occurrence peak) identified via the in-
situ PCA and 2nd derivative methodologies. The results of these efforts
identified the best performing band from each occurrence peak. A series of
classifications were then generated by systematically replacing each of the
optimal bands with their adjacent bands. In some instances, adjacent bands
were weighted equally, or nearly so, with respect to the PCA/ derivative
results. Subsequent classifications were generated to further explore the trade-
off between relative classification change and spectral resolution (i.e. bandset
optimization). The maximum number of bands utilized was 30, which is

approximately 10 more bands than found in the 1-meter AISA and CASI-II

103

imagery. The minimum number of bands was 4, which was selected in order to

emulate the number of bands utilized by the IKONOS sensor (Table 2.2).

Classification performance was assessed through the comparison of relative
change (i.e. classification resiliency) with respect to the Maximum-Likelihood
and SAM evaluation-standard classification images, and confusion matrices
were generated similar to Investigations 1-3. Only three matrix parameters
were referenced throughout Investigation 4, overall accuracy, kappa, and the
producer’s accuracies for each of the 21 classes. These parameters allowed both
the cumulative and class-specific classification resiliency (i.e. relative level-of-
change) associated with each bandset manipulations to be addressed. The
comparison of these values provided the means through which the optimal
bandset associated with my Great Lakes coastal wetland imagery was

identified.

3.2.5 Image Classifications - Bandwidth Optimization

 

A second bandset characteristic somewhat independent of band
number/ position is the spectral width of any one band. Assume that the PC
and 2nd-derivative analyses (referencing 48 bands) correctly identified both
708.5 and 731.5 nm as the two most desirable band locations. If the bandwidths
of these two bands were increased to approximately 25 nanometers, they would

overlap. From a practical standpoint, sensors are not configured to capture

104

bands with significant spectral overlap due to the blurring of the spectral
information specific to each band. If bands were allowed to overlap without
removal, it is equivalent to applying a spectral smoothing (i.e. band averaging)
filter across the bands. As stated in Chapter 1, wide bands tend to mask
relatively subtle spectral features that potentially separate botanical classes that
might only be visible with two or more bands of narrower widths. Thus,
bandwidth potentially plays a critical role in the optimization of a coastal

wetland remote sensing methodology.

A series of synthetically generated text files (Table 3.3) were created in
which the FWHM and band number associated with the optimal bands
identified via the PCA/ derivative methodologies were systematically altered
(e. g. 2x bandwidth). The 12x resampling resulted in bands with a width of
approximately 70 nanometers, which is equivalent to the average bandwidth
associated with the visible-light bands (i.e. bands 1, 2, and 3) of both the
IKONOS and ETM+ sensor. Bands that are located relatively close to each
would inherently be first affected by increases in bandwidth. Overlapping
bands were eliminated according to their 2nd derivative occurrence frequencies.
As one might expect, the potential for such band eliminations increases as the

FWHM envelope around each of the optimal bands expands.

105

A Maximum-Likelihood and SAM classification image was generated with
respect to the bandwidth-manipulated imagery using training ROIs outlined in
Table 3.2. Again, where band overlap occurred, the band with the lower 2“d
derivative frequency was eliminated. These classification results were
compared via three confusion matrix parameters (i.e. kappa coefficients,
overall-resiliency and producers-resiliency) to unaltered (bandwidth increases)
classification referencing the identified optimal bands. Thus, relative
classification change provided insight into the optimal bandwidth for the

classification of my Great Lakes wetland imagery.

 

Band original 2x 3x 4x 6x 8x 12x
Center FWHM FWHM FWHM FWHM FWHM FWHM FWHM
425.4 5.7 11.4 17.1 22.8 34.2 45.6 68.4
514.9 5.8 11.6 17.4 23.2 34.8 46.4 69.6
560.1 5.8 11.6 17.4 23.2 34.8 n/ a N / a
685.5 5.9 11.8 17.7 23.6 35.4 47.2 70.8
731.5 5.9 11.8 17.7 23.6 35.4 n/ a N / a
812.3 5.9 11.8 17.7 23.6 35.4 47.2 70.8
916.7 5.9 11.8 17.7 23.6 35.4 47.2 70.8

 

 

 

Table 3.3. Bandwidth Manipulation FWHM Parameters

3.2.6 Image Classifications - IKONOS Spectral Simulation

The spatial resolution of monochromatic IKONOS imagery, which is the
highest resolution (for multispectral imagery) currently available from a
satellite platform, is 4 meters. Each of the four IKONOS monochromatic bands
encompass a large portion of their respective wavelength domains, blue

106

(710nm), green (890nm), red (660nm) and NIR (960nm). If it is true that
relatively narrow, strategically placed bands are needed to better differentiate
coastal wetland plant communities, then the usefulness of IKONOS or similar

broadband imagery falls under question.

The 48-band, CASI-II imagery was spectrally resampled in reference to an
ASCII text file that contained the band center and FWHM information identical
to the IKONOS sensor (see Table 2.2). This process created a 4-band, pseudo-
IKONOS image that emulated the bandset of the IKONOS sensor (i.e.
bandwidth and band number). A Maximum-Likelihood and SAM
classification was generated from this pseudo-IKONOS image using the 21
training ROIs outlined in Table 3.2. Analogous to previous sections, a
confusion matrix was generated from these two classification results with
respect to their appropriate evaluation standard, and kappa coefficients,
overall-resiliency and producers-resiliency were extracted. These confusion
matrix parameters reflect the additive effect of both bandwidth and band
number manipulations and are directly comparable to the 4-band classification
outlined in section 3.3.3. The evaluation of these confusion matrix parameters
provided insight into the viability of using airborne, IKONOS-like imagery for

the classification of Great Lakes wetland imagery.

107

3.3 Results and Discussion

3.3.1 Representative Spectral Clusters

The 96 spectral signatures referenced in Table 3.1 were simultaneously
displayed in order to visually identify groups or clusters of spectra. The
identification process was conducted without knowledge of the botanical class
each spectral signature represented in order to reduce bias during the selection
process. A representative signature was then pulled from each cluster, and
their appropriate botanical information was examined. This process was not
intended to rigorously group spectra based on a statistical clustering method.
Rather, the intent was to explore the in-situ data to see if easily identifiable
clusters existed, and if these clusters were indeed related to the characteristics

of the botanical community.

Nine spectra were gleaned from the original 96 signatures, and their
associated botanical information was compared (Figure 3.4). The patterns of
these nine spectra encompassed the broad botanical classes making up Great
Lake coastal wetland plant communities (e.g. dense-submergent, emergent,
senescing emergent, substrate dominated). Since this visual inspection
revealed botanically-related clusters, one would expect more advanced spectral
tools and classifiers to be able to similarly, or more effectively, parse the in-situ
and closely related image-ROI spectral data. The spectra shown in Figure 3.4

represent a gradient of botanical signatures ranging from un-vegetated shallow

108

 

 

 

 

 

 

 

 

60
O
50 ‘ ‘* "2.345.“ “ “’
a) v'
U I 8 n
:40 - - - L- - -—
«1 I
'8 o
. —-_~~
gm U T T -_.:__.-a" r¢ '.;-T:;;T-—I--. L6
‘H a I”, ..C‘ .5
m 1” I" ‘.07
mm - - - m— mu
c\° ” ------ ~-5
10 —--.;.—;-_ ----- .’-~ -~ ,.
0" = "'II' / __-.
o .32.. --;-_- ----- 4
' ""7— : ' I t T 7 1 I ---”T

m4505m5506m6507ID7508111850%%0
nanoneters

(1 = un-vegetated shallow water; 2 = sparse submergent vegetation in
shallow water; 3= submergent vegetation in shallow water; 4 =
emergent/submergent vegetation in shallow water; 5 = necrotic vegetation;
6 = gravel; 7, 8 and 9 = high biomass vegetation).

Figure 3.4. Nine Representative Spectra

water (1) to dense, herbaceous vegetation (9). When the IFOV of the SE-590
recorded the radiance (ET) over a target dominated by water, the entire
signature was suppressed (i.e. 1, 2, 3 and 4). Water-dominated targets should
display a minimal amount of reflectance throughout the NIR region since water
absorbs these wavelengths. Signature 2 represents a target that is
predominantly water, but enough plant biomass was present (Najas flexilis) to

increase NIR reflectance above that of a substrate-only, shallow water target (1).

109

Eelgrass/ musk-grass (Vallisneria americana/Chara sp.) beds (3) covered much
of the shallow inundated regions of Horseshoe Bay during the 2000 and 2001
growing seasons. This signature was representative of shallow areas
predominantly covered by submergent vegetation. Indicative of this spectral
class is the double-hump pattern between 700 and 850nm. Note that signature
4 displays a similar shape, although its double-hump is suppressed somewhat
by reflectance in the N IR wavelengths due to the presence of emergent and
submergent biomass. Also note that reflectance in the green region (i.e. 560 nm)
is slightly elevated due to the presence of more green biomass at or near the
surface. An area composed of emergent Sagittaria rigida and submergent
Heteranthera dubia, which covered much of the delta south of Dynamite Cut

(Horseshoe Bay), was the source for signature 4.

Signatures 5 and 6 were displayed with similar line types because they
exhibited relatively flat reflectance curves. Gravel (6) displays relatively
constant reflectance across the visible and N IR wavelengths. Signature 5 is
representative of areas dominated by senescing or necrotic vegetation. Note
that it too displays relatively constant reflectance throughout the visible and
N IR region, although there is still evidence of plant biomass as indicated by the

steep increase in NIR reflectance beginning just beyond 700nm.

110

Signatures 7, 8 and 9 represent landscape facets that are completely
vegetative in nature. Signature 9 exhibits the characteristic peak—and-valley
reflectance pattern of high-biomass green vegetation. This signature (Impatians
capensis) represents one extreme of the botanical gradient, and is typified by the
steepness and length of the red-edge (690-750nm). Signature 8 represents the
dominant, tall-stature emergent vegetation found throughout the Great Lakes
(Typha sp./Scirpus sp.). This signature is strikingly different from 5, (same
genera, mostly non-living, brown biomass), indicative of its level of greenness.
Signature 8 differs from signature 9 in three key areas: 1) flatter reflectance
pattern across the visible domain; 2) lack of a pronounced red absorption
feature (690-700nm); and 3) a more moderate transition into the high reflectance
of the NIR. All of these traits are indicative of a significant spectral contribution

from substrate, necrotic / senescent biomass, or both.

The scope of this research was not to explore every spectral nuance found
in the raw in-situ data. These nine characteristic spectra suggest that digital
spectral methodologies should be able to separate these signatures. In addition,
the presented spectra establish a direct link between spectral reflectance
properties and the biophysical characteristics of wetland plant communities.
The remote sensing community is far from being able to link any one species to
a specific spectral feature, but patterns exist that allow the connection between

biophysical traits and remotely-sensed data.

111

3.3.2 2“d Derivative Summary

 

Figure 3.5 displays the 2nd derivative approximations of the raw signatures
shown in Figure 3.1. The relative magnitudes of these 2nd derivative
approximations (Appendix C) are indicative of the reflectance symmetry (i.e.
parity of 1St derivative absolute values) found at the triplet center. A high
magnitude value indicates distinct reflectance asymmetry at the hinge point
(i.e. a diagnostic change is reflectance). The sign (+/ -) of the 2nd-derivative
approximations found in Appendix C indicates whether the spectral feature is
local reflectance maxima or absorption minima. A positive value indicates that
the band-triplet forms a localized concave or bowl-like spectral feature.
Conversely, a negative 2nd-derivative value indicates a convex spectral feature.
In order to form a concave feature (i.e. positive derivative value), the reflectance
of the center band must be less than the reflectance vector (i.e. trend)
established by its two neighboring bands (i.e. an absorption feature). A convex
band triplet (i.e. negative derivative value) describes a local reflectance peak.

If two band-triplet line segments have identical slopes, the calculated 2“d
derivative approximation would have a value of zero. Thus, 2nd derivative
values near zero are associated with a band triplet reflectance values that are
either generally increasing, steadily decreasing or display very weak absorption
or reflection features. The largest 2“d derivative approximations shown indicate
that there was a relatively large change in reflectance within a band triplet.

The five highest positive and negative 2nd-derivative approximation values

112

 

0.04
0.03 .
0.02 —

0 .72 I

2nd Derivative
Approximation

 

0.01--I---

I- L- LILJL 13‘s,: I

—0.01 AIM-I - ,-
-0.02 Iv _ i
003 -~-4

, _ _
I
I

\
Z/

".3 /J' i f
:«\
N

O
A

4.’

I

I
I
I

‘v
‘(G'

I

 

I I I I

._+. ___ __ -L

Q
<\. .
x \‘ r

1 AVI
I I
_l_

(\.
n‘.

c Q‘
I
I .-
. 4— /)

.f I +_ ____

__V_ _+ __.._ ~74

 

 

-0.04 - I

400 450 500 550 600 650 700 750 800 850 900 950

nanometers

Figure 3.5. Representative 2“d Derivative Plots

generated from the in-situ data are shown in bold typeface in Appendix C.

These 10 rows mark points of inflection that are associated with the

predominant absorption (+) or reflection (-) features for any given spectra. It

should be emphasized that the 2nd-derivative approach outlined in this research

is only applicable to contiguous, symmetric (hyperspectral) reflectance data.

The data contained in Appendix C were further refined to yield a derivative

summary table. The values encountered in each of the bold rows (5 negative, 5

positive) were independently recorded, and redundant values were retained

only with respect to the row in which they first appeared, creating a

magnitude-ranked list of positive and negative values. The 11 wavelengths (i.e.

113

band centers) found within the first column of Table 3.4 represent the largest
magnitude derivatives gleaned from all 88 botanical signatures. The bold
typeface entries in successive columns reveal those band locations that were
introduced upon examination of the next lower magnitude level. Thus, four
new band centers (674.1nm, 766.1nm, 823.9nm and 835.5nm) were introduced
through the compilation of the second tier (1 positive, one negative) of
derivative approximations. This table summarizes the most important band
centers with respect to the 88 in-situ spectra from the perspective of magnitude
only. Eight band clusters or wavelength domains were subsequently identified
via gaps along the wavelength continuum. The 5th cluster was further divided
into its visible and N IR portions because of the marked difference in vegetative
reflectance between these two wavelength domains. The resulting 9 clusters
delimit the wavelength zones that were identified as most significant. It is
likely that these are the spectral regions that are the most useful for

differentiating wetland plant communities.

Several of the wavelength zones delineated in Table 3.4 are associated with
well-known biophysical attributes. Zone one contains absorption features from
a variety of plant pigments: chlorophyll-a at 420nm; chlorophyll-b at 435nm;
beta-carotene at 425nm and 450nm; alpha—carotene at 420nm and 440nm; and
zanthophyll at 425nm and 450nm (Kumar, 2001). Zone two contains bands that

have been related to water turbidity (Shafique et al., 2001). The third zone is

114

 

 

1
685.5
697
708.5
720
731.5
743
754.6
812.3
916.7
928.3
939.7

.L-Z 1-3 1_-4 1:5. Freq-

674.1
685.5
697
708.5
720
731 .5
743
754.6
766.1
812.3
823.9
835.5
916.7
928.3
939.7

 

447.7
514.9
651.2
674.1
685.5
697
708.5
720
731 .5
743
754.6
766.1
812.3
823.9
835.5
847.1
905.1
916.7
928.3
939.7

425.4
436.5
514.9
560.1
651.2
674.1
685.5
697
708.5
720
731.5
743
754.6
766.1
777.7
812.3
823.9
835.5
847.1
905.1
916.7
928.3
939.7

Notes

 

425.4
436.5

Chlorophyll absorption (1)

 

503.6
514.9

Turbidity /
Suspended solids (510mm

(2)

 

548.8
560.1

(3)

Peak green reflectance

 

594.1

(4)

 

651.2

Water absorption (5)

 

674.1
685.5
697

Senescent vegetation (672nm)
Chlorophyll absorption (6)
Senescent vegetation (690nm)

 

708.5
720
731.5
743
754.6
766.1
777.7

Suspended solids (705nm) (7)
Senescent vegetation (705nm)
Water absorption (720/ 731nm)

02 absorption (766nm)

 

800.8
812.3
823.9
835.5
847.1

Water Absorption (8)

 

 

870.3
881.9
893.4
905.1
916.7
928.3
939.7

 

(9)

 

 

Table 3.4.

115

In-situ 2nd Derivative Summary

 

associated with the green reflectance peak from vegetation. The spectral region
comprising zone four is sensitive to yellow-green or yellow reflectance. The
single band comprising zone five (651.2nm) has been related to a weak water
absorption feature and with absorption by monochromatic chlorophyll-a
(Shafique et al., 2001). Reflectance in zone six is dominantly controlled by
chlorophyll absorption. As such, these bands are sensitive to the relative
amount of green vs. brown biomass in each pixel. Carter (1993) found a

significant sensitivity to plant senescence at 672nm and 690nm.

In the first N IR zone (708.5 - 777.7nm), several biophysical controls may
modulate reflectance. Shafique et a1. (2001) found weak relationships between
reflectance at 705nm and concentrations of total suspended sediments and
trichromatic chlorophyll-a. The strongest, short-wavelength water absorption
feature is centered at 722nm. Since green vegetation strongly reflects the NIR
wavelengths, the 720nm band is undoubtedly sensitive to the amount of
standing water within a pixel. The 766.1nm band may be overly influenced by
the strong oxygen absorption feature located between 760 - 765nm. Zone eight
encompasses a spectral region of maximum reflectance by green vegetation,
although a weak water absorption feature occurs near the 812.3nm band. Many
of the submergent vegetation types achieve their peak reflectance in this zone.
Zone eight is also where many of the senesced, large-stature emergent genera

achieve their peak reflectance.

116

Overall reflectance from water is very low throughout the NIR portion of
the spectrum. A distinct, strong, in-vivo water absorption feature is obvious in
vegetation spectra at 960nm. Its affect helps to suppress vegetative reflectance
across zone nine. Dehydration usually accompanies senescence, so reflectance
associated with zone nine is related to the amount of brown biomass in the

pixel.

The results presented in Table 3.4 were based solely on relative magnitude,
without reference to the frequency of occurrence for any single band—center.
Some of the presented band—centers occurred only once among all 88
signatures, and, therefore are not likely to be as botanically important as ones
that occurred frequently. The histogram presented as Figure 3.6 was
constructed from 880 of the 960 derivative values (i.e. 88 vegetative spectra x 10
values) shown in Appendix C; these 880 values did not contain any non-
vegetative spectra. The values were sorted with respect to wavelength in order
to record the number of occurrences associated with each of the band centers.
This frequency analysis highlighted several key band locations or wavelength
clusters. Assuming that these 88 signatures captured the spectral essence of the
typical coastal-wetland plant communities, then the bands highlighted in both
Table 3.4 and Figure 3.6 represent the most powerful bands with respect to

botanical differentiation.

117

100
90 -I

 

 

 

 

 

 

 

 

 

 

 

 

 

10- _
0. ‘.... .
400450500550600650700750800850900950

Occurrences- 2nd Derivatives
8 8 3 8 8 El 8

 

 

 

 

 

 

 

nanometers

Figure 3.6. Frequency of 2nd Derivative Occurrences

It was beyond the scope of this research to systematically examine the raw
spectra and their related derivatives in order to identify unique patterns that
could be linked to the biophysical nature of the sampled plant community.
However, patterns have emerged that link certain derivative signatures with
particular wetland cover types and their associated biophysical characteristics.
Pure water displays a narrow absorption feature centered at 722nm, so it would
be expected that the signatures of targets dominated by standing water (e. g.
submergent species) would exhibit an absorption feature associated with this
wavelength. In contrast, the presence of green biomass causes pronounced
NIR reflectance. Therefore, a strong absorption feature at 720nm (the actual
band center used on the CASI-II instrument) in combination with elevated N IR

118

reflectance (i.e. 731.5nm) could be used to identify submergent vegetation.
Signature 3 [eelgrass/ musk—grass (Vallisneria americana/Chara sp.)] (Figure 3.4)
exhibits a double-hump pattern that is associated with the contrast in
reflectance properties of the 720nm and 731.5nm bands, a pattern highlighted
by derivation. Much more work is needed to explore the species-specific

nuances of the derivative approximations shown in Appendix C.

3.3.3 PCA Loading Curve Inspection

When applied to the in-situ reflectance spectra, Principal Components
Analysis (PCA) generated band-specific component loadings. These
component loadings measured which independent variables (the bands) were
involved with which meaningful dimension and to what degree; they can be
viewed much like correlation coefficients (Rummel, 1970). Since the
meaningful dimensions functionally encapsulated all (>99%) of the spectral
variance of the botanical spectra (dependent variables), the loadings identified
the descriptive power of all 48 bands within the context of the six meaningful

dimensions.

Within the derivative methodologies described in the previous section, each
individual signature was treated as an independent entity, so the results
associated with a particular signature could be evaluated. This is not the case

with component loadings, since they are generated with respect to the

119

combined characteristics (variance) of the dependent variables. Because the
dependent variables (i.e. botanical spectra) are treated jointly within PCA, it is
extremely data dependent. It is therefore impossible to isolate the component
loading associated with a specific spectral signature. Factors scores, which
were also generated, are specific to each individual dependent variable
(spectra). PC factor scores describe the proportionally weighted involvement of
each dependent variable (i.e. botanical spectra) in each meaningful dimension
(Rummel, 1970). Factor scores only explain the relationship between the
orthogonal factors and a specific spectral signature, without reference to
individual independent variables (i.e. bands). Since the goal of my in-situ PC
analyses is to help identify those bands that best differentiate botanical classes
(analogous to derivative methodologies), factor loadings, and not factor scores,

are directly applicable.

Appendix D contains the raw SYSTAT output from both the correlation-
and covariance-based PCA transformations, including dimension specific
statistics (e.g. variance explained) and band-specific component loadings. The
first strategy applied to the baseline PCA data shown in Appendix D was
graphical in nature. Figure 3.7 contains the raw (i.e. unsorted) component
loadings generated with respect to the 88-member in-situ data (i.e. gravel-like
signatures removed). The 12 curves correspond to the 6 extracted dimensions

associated with both the covariance- and correlation-based PCA methods. No

120

attempt was made to emphasize the meaningfulness of either PCA
methodology, and it was assumed that the results generated from both
methodologies were meaningful with respect to the identification of key band
centers. One or more of these signatures mimic the characteristic peak-and-
valley reflectance pattern associated with vegetative spectra. This is strong
evidence that these dimensions are closely linked to the botanical nature of the

input signatures.

Superimposed on Figure 3.7 are a series of bars labeled according to the
band center over which they are centered. The bars were systematically placed
in order to identify the most visually-significant spectral features common
throughout the 12 meaningful dimensions. In some instances, a bar could be
shifted in order to encompass an adjacent spectral feature and band-center. For
example, one could argue that the 708.5nm band could be shifted to the right,
and relabeled as 720 nm. However, the displayed bars identify (by visual
inspection means) the most significant wavelength domains with respect to the
various spectral features exhibited by the component loading curves.

This was not a rigorous mathematical process to identify significant spectral
features. However, the band-centers identified by this simple method
correspond well with those bands having the highest occurrence values shown
in Figure 3.6. At a minimum, the features identified by the visual inspection of

the component loading curves are likely to be botanically relevant bands.

121

1°
01

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

2- Li- L__.. _ W ___.
(D
gels -- --—~ w- —--— _— -
g
3 1 f A J *A *‘ ’4.__+ K ——-- 4
50,5..- i++hw -2 I __
9* \
.2
gas-- -in,_w_m- ---
0)
.§ ,1- _ - -__-_,__ [IMIII I _____
c:
1 5 g 560.1 685.5 / l 823.9 Inle
2 708.5 731.5

 

400 450 500 550 600 650 700 750 800 850 900 950

nanometers

Figure 3.7. Key Wavelength Domains - Dimensional PCA Loadings.
(Note: Covariance PC1 (A) has been broken at 710nm and shifted
downward approximately 13 loading units so it could be displayed)

3.3.4 Loadimg Curve Derivation

A second strategy was applied to the PCA output arrays shown in
Appendix D. This more formal mathematical approach was designed to
address any shortcomings associated with the visual methodology described in
the preceding section. The component loading curves were subjected to the 2“d
derivative approximation methodology outlined in section 3.3.2. Analogous to

the approach used to generate Figure 3.6, the 10 highest-magnitude, derivative

12

values (i.e. 5 + and 5 -) were compiled so that a frequency histogram could be
constructed (Figure 3.8). The overall patterns exhibited are very similar to the
patterns shown in Figure 3.6. Six occurrence-peaks are common to both graphs,
centered at approximately 515nm, 560nm, 685nm, 720nm, 823nm, and 940nm
nanometers. These six band centers wavelength domains are essential with

respect to both the PCA and derivative analyses.

Table 3.2 presented nine distinct wavelength domains that contained
botanically relevant information. Viewing Figures 3.6, 3.7 and 3.8 with respect
to these independent wavelength regions, the optimal band (i.e. most
occurrences) in each region was systematically selected. In addition, the
relative strength of each region can be assessed based on their relative
frequency of occurrence. Applying this strategy, the 685.5nm and 720nm band
centers are the two most important bands within the context of this research.
These are followed in significance by the 607nm and the 731.5nm bands.
Assuming the 88 in-situ signatures utilized to generate Table 3.2 and related 2ml
derivative figures sufficiently captured the spectral essence of a typical Great
Lakes coastal wetland plant community, then the redundancies found in these

data suggest that the identified band-centers are the most botanically relevant.

123

 

 

 

 

 

     

 

 

 

 

 

 

I I I T
10 . I_I. ~1-
I ) ' l
I I I I
'5 3 I , I. I- I
S‘ I I I
'5 ~ : I
D I , I
“U 6 If i . _- r ——1
1 .= as
g as
t: 4" m” T
g i
U I A
u
o 2 ~ , . —
( 1
I i I l I
I 2 I I I l
0' I I I r if I I if I i

 

 

400450500550600650700750800850900950

nanometers

Figure 3.8. Frequency of 2nd-Derivative Occurrences of Loading Curves.

3.3.5 Correlation-based, PCA Factor Interpretation

Kauth and Thomas (1976) derived a modified Principal Components
rotation (T asseled Cap) for Landsat MSS data in which the majority of target
variance is captured in the first two dimensions (Lillesand and Kiefer, 1994; p.
577). These two dimensions have become widely recognized as Brightness and
Greenness. Brightness, which is also called n-dimensional albedo, is the
weighted sum of all bands in reference to the principal variation of soil

reflectance. The second component, greenness, contrasts reflectance in the

124

visible versus N IR wavelength regions, and is strongly correlated with the
amount of green vegetation in the scene. One could argue that the second
dimension should be called non-greenness, since the component loadings (i.e.
transformed eigenvectors) are negative for the NIR bands. Green vegetation
(with its maximum reflectance in the NIR) is inversely related to the second
dimension. Further studies have assigned meaning to a limited number of
additional dimensions (i.e. 3rd = wetness) with respect to 6-band Landsat TM

data (Crist and Cicone, 1984).

As presented in earlier sections, the results of PCA are inherently
dependent on whether a covariance or correlation matrix is used. Covariance-
based PCA, (i.e. non-standardized) is more sensitive to the extremes of the data,
because it both preserves and utilizes the entire magnitude range of the data. In
contrast, correlation-based PCA (i.e. standardized) suppresses patterns within
the magnitude range of the data, and places more emphasis on data similarities.
Because of this fundamental difference, the factors generated via both

methodologies are discussed separately.

Figure 3.9 displays the correlation-based dimensions generated from the 88
spectra referenced throughout this research. These curves are identical to those
presented in Figure 3.7, but the scale was changed to aid in their interpretation.

In order to better interpret the biophysical meaning of the correlation-based

125

1"
N

 

y—l

 

P
00

 

 

 

P
as

 

P
as

O O

 

 

 

 

 

 

 

PCA Component Loadings
G
N

55559555
m¢1hN

 

 

 

 

 

 

 

 

nanometers

Figure 3.9. Correlation-based, PCA Component Loading Curves.

dimensions, it is helpful to visualize a hypothetical reflectance continuum for
the 88 spectral signatures. Three broad categories of cover types are found in
Horseshoe Bay -- water, soils and vegetation, and the original 96 signatures
captured the essence of each. If placed along a continuum of relative
reflectance across all 48 bands, dry, sandy substrates would be the most
reflective surfaces across all bands. Water-dominated targets, on the other
hand, would be the least reflective surfaces across all 48 bands. Eight
signatures were removed from the original 96 because they were non-
vegetative; these contained the extremes of this hypothetical continuum. The

remaining 88 vegetative signatures would fall somewhere in between these two

126

extremes, depending on the level of vegetative biomass, both green and dead,
and its associated percent cover. A sparsely-vegetated, sandy shoreline would

be near the bright end of this continuum.

The characteristic that all of the signatures have in common (i.e. high
correlations among all the variables) is their brightness along this continuum.
Not surprisingly, the lst dimension that captured 68.28% of the total variance,
can be interpreted as n-dirnensional brightness. However, it is n-dirnensional
brightness within the context of a vegetation-dominated data set or landscape.
Note that the highest loads are associated with the green peak (549nm) and the
NIR maxima (710nm). The amount of vegetation present is playing a secondary
role in this dimension. Examination of the factor scores associated with this
dimension further demonstrated that it is related to brightness, albeit seen
through a vegetative filter. The variables that scored lowest were the pure
water, (i.e. very dark pixels). Variables with relatively high biomass, either

terrestrial or emergent/ submergent, displayed the highest scores.

The loading plot associated with the second dimension is reminiscent of an
inverted peak-and-valley reflectance signature typical of vegetation. This
dimension, which captured 29.69% of the total variance, is associated with
greenness. Healthy green vegetation has two reflectance peaks, one centered

near 560 nm (i.e. green) and the other associated with the N IR wavelengths.

127

Note that in Figure 3.8, these wavelengths coincide with the lowest relative
loadings. To illustrate the non-greenness attribute associated with this second
dimension, note that the most ”green vegetation” signature (Impatiens capensis)
of all the in-situ data (Impatiens capensis) produced the largest, negative

component score for component 2.

The component loadings associated with the 3rd dimension, which captured
2.83% of the total variance, are also closely linked to the biophysical
characteristic of the plant communities sampled. The shape of the 3rd-
dimension curve (Figure 3.9) is a mirror- image of one from a very yellow
reflector (i.e. steadily increasing reflectance across the longest of the visible
wavelengths yellow, orange and red). Note that all of the loadings for
components 3-6 are near zero in the NIR. This is because the contributions of
the NIR wavelengths have already been accounted for in the first two
components. Although PC3 captured a relatively small percentage of the
overall data variance, this dimension appears to emphasize target ”non-
yellowness”. In the 3rd dimension, factor scores for blooming goldenrod
(Solidago sp.) and sunflower (Helianthus sp.) were the two highest magnitude,

negative values on this dimension.

The 4th dimension (1.60% of the total variance) appears to represent target

brownness or dead biomass. Note the relatively high contribution of

128

orange / red (i.e. 650-690nm) reflection coupled with a lack of green (550nm) and
a distinct drop in N IR reflectance on the shoulder (700-720nm) of the N IR
region. Brown biomass, especially if moist, would exhibit such a reflectance
pattern. Brown (necrotic) biomass spectra from multiple genera and blooming
common reed (Phragmites australis) were associated with the highest magnitude,
positive factor scores on PC4. Note that the blooms of the common reed were a
deep brown color during spectral sampling, which further substantiates the
notion that dimension 4 is related to spectral brownness. Many of the
submergent cover types, especially those with relatively sparse vegetation,
displayed the highest negative factor scores on PC4. These spectra would have
some brownness due to the spectral contribution of their substrate, but would
certainly have some green and NIR reflectance making their reflectance spectra
distinctly ”not brown”. This dimension (PC4) might be useful in identifying

low-biomass submergent plants.

The loading pattern of PC5 suggests that this dimension is associated with
targets exhibiting a minor green reflectance peak coupled with a sag in N IR
reflectance just in the N IR-shoulder region (700-720nm). Although the sixth
dimension was included in previous discussions, no attempt was made to
interpret its biophysical meaning due to the relatively low amount of target

variance it captured and its near-zero loading pattern.

129

3.3.6 Covariance-based, PCA Dimension Interpretation

 

Figure 3.10 depicts the loading patterns of the covariance-based dimensions
generated from the 88 spectra referenced throughout this research. Again it is
helpful to think of a reflectance continuum to aid in interpreting these patterns.
The continuum described previously with reference to the correlation-based
results was oriented in order to best capture similarities associated with the
variables. A covariance continuum, in contrast, is oriented to enhance the most-
striking differences among the variables. The highest magnitude reflectance
change found throughout the 88 spectra contrasts minimum reflectance in the
short-wavelength visible with maximum reflectance in the NIR. The most
luxuriant vegetation targets would represent one extreme along this covariance
continuum. Non-vegetative, flat reflectors (e.g. open water or muck) represent
the other extreme along this covariance continuum. The NIR portion of the first
covariance dimension was eliminated from Figure 3.10 in order to display the
remaining five dimensions at a scale that would enhance the interpretability of

the patterns.

The loadings associated with the first covariance dimension mimic the
characteristic, peak-and-valley reflectance signature typical of vegetation. This
dimension, which captured 96.87% of the total variance, can only be interpreted
as being sensitive to green foliar biomass. Healthy, green vegetation has two

reflection peaks, one centered near 560 nm (i.e. green peak) and the other

130

2.5

 

Z"
or

H

6

 

 

 

 

.5
01

 

Dimensional PCA Loadings
O
U!

 

 

I
H

 

 

 

 

-1.5~
400450500550600650700750800850900950

 

nanometers

Figure 3.10. Covariance-based, PCA Component Loading Curves

associated with the N IR wavelengths. The overwhelming majority of target
variance was captured by this dimension, indicating that the contrast between
visible and NIR reflectance can be the foundation for differentiating virtually all
of the 88 spectral signatures. The most ”green” signature (Impatiens capensis)
found throughout all of the in-situ data produced the largest, positive factor
score on PClcov. Water covered sand, an extremely flat reflector (i.e. no

visible / N IR contrast), exhibited the highest magnitude negative scores on

PClcov.

131

The covariance-based, 2Ind dimension seems to be indicative of targets with
mildly elevated reflectance in the yellow-green portion of the spectrum coupled
with a minor red-absorption feature and a reflectance peak at the NIR shoulder.
Also striking in this pattern is the rapid decline in reflectance at wavelengths
beyond about 725nm and relatively stable reflectance across the visible
wavelengths. This dimension is associated with floating-leaf vegetation. A
water-dominated, but relatively green, target such as Nuphar advena would
exhibit such a spectral signature. The absorption of the NIR wavelengths
beyond 725nm is due to standing water in the field-of-view. The presence of
interspersed green biomass in the form of floating leaves is responsible for the
subtle, broad yellow-green peak coupled with red-light absorption and the
short-wave NIR reflectance maximum. The highest positive scores related to

this dimension were the water-dominated, floating-leaf communities.

The 3rd covariance dimension, which captured 0.82% of the total variance,
appears to be related to dry, brown biomass. Note the lack of green reflectance
(negative loadings around 550nm) coupled with increasing reflectance across
the orange / red (i.e. 650-690nm) portion of the spectrum and the distinct drop
(to a maximum negative load) in NIR reflectance. Another notable feature of
the loading pattern of PC3C0V is the increasing reflectance in the NIR beyond
about 825nm. This increase suggests that the brown biomass is desiccated and

that little or no water is showing through from below. Brown (necrotic)

132

biomass spectra from multiple genera were associated with the highest positive
factor scores on PC3cov. The submergent and water-dominant emergent cover
types were associated with the highest negative factor scores on this third
covariance dimension. In addition, the most-green signatures (e. g. Impatians
capensis) found throughout the 88 spectra were also negatively associated with
this dimension. The patterns associated with the first three covariance
dimensions suggest a practical use. To differentiate wetland plant genera, one
approach could involve finding differences associated with their greenness,

brownness, and substrate co-dominance.

Generically, the percentage of variance explained by lower-order PC
dimensions (i.e. dimensions 4, 5, and 6), especially covariance-based, is
minimal. The patterns associated with these lower-order dimensions are
inherently difficult to interpret because they must be viewed as being
completely independent of the higher-order patterns. For example, the
covariance-based PClcov dimension is clearly linked to healthy green biomass.
Therefore, all lower-order dimensions (2 through 6) must be interpreted
independently of target ”greenness” because this multispectral attribute has
been sequestered into only the first dimension. Interpretations of the loading
patterns associated with PC4 cov, PC5 cov and PC6 cov were not practical because
of the level of inconsistency associated with their factor scores and the inability

to assign meaning independent of their higher-order counterparts.

133

3.3.7 Bandset Optimization - Investigation 4

 

Three pools of data were generated via the in-situ PCA and 2“(1 deriviatve
methodologies to identify key band locations (Figures 3.6, 3.7, and 3.8; Table
3.4). This information was utilized to systematically reduce the number of
bands referenced during the Maximum-Likelihood and SAM classifications of
the 4-meter, CASI-II imagery. Analogous to earlier sections, the classification
error matrices (Appendix B) were summarized and presented in tabular form
(Table 3.4). The three sections in Table 3.5 reference the three components of
Investigation 4, and were sorted in descending order so that the relative
performance of each classification outcome could be easily compared. The
overall-resiliency values found in Table 3.5 reflect a pixel-by-pixel assessment
of classification change resulting from bandset manipulations. Again, the
calculation of a variance value for each Kappa Coefficient allowed a Z-test
(Congalton et al., 1983) to be conducted between all paired classification
outcomes. A select group of classifications in Investigation 6 were not
statistically different from one another at the 95% confidence limit, and their
classification identifiers were preceded with one or more asterisks (see Table 3.5
footnote). The bulk of the results presented in Table 3.5 references only SAM
classifications. Maximum-Likelihood, which is a probability-based
classification tool, responded poorly to the band reduction strategies. A limited
number of Maximum-Likelihood classifications were generated and their

associated results were only briefly discussed. Thus, bandset reduction

134

 

 

 

 

 

 

 

 

 

Investigation 4A Overall Kappa
Classification Resiliency Statistic
SAM4ikonos 67.74% 0.6522
SAM4band 79.64% 0.7807
SAM7-939 82.49% 0.8111
SAM7—720 83.52% 0.8223
*SAM7-674 85.35% 0.842
*SAM7-743 85.65% 0.8452
**SAM7-548 86.09% 0.8501
***SAM7band 86.27% 0.8519
***SAM7-823 86.26% 0.8519
SAM30band 90.15% 0.8937
SAM24band 94.09% 0.9362
Max7band 59.46% 0.5609
Max24band 73.72% 0.7148
Max30band 78.72% 0.7683
Investigation 4B Overall Kappa
Classification Resiliency statistic
MAX-2x 60.23% 0.5691
MAX-6x 62.58% 0.5950
*MAX-8x 55.84% 0.5223
*MAX-12x 55.76% 0.5216
SAM-2x 86.44% 0.8538
SAM—6x 87.33% 0.8633
SAM—8x 84.87% 0.8364
SAM-12x 83.96% 0.8265
Investigation 4C Overall Kappa
Classification Resiliency statistic
SAM_pseudoIKONOS 71.54% 0.6926
IKONOS_SAM_16classes 77.78% 0.7605

 

 

", *’” = statistically identical
’”‘ = different from SAM7band @ .70 confidence limit

 

 

Table 3.5. Summary of Investigation-4 Classifications

135

strategies are accompanied with an initial reduction in accuracy due to the need

to focus on the SAM classifier.

3.3.7.1 Optimal Band Location - Investigation 4A

 

Seven predominant wavelength domains were identified via all four band-
location approaches (i.e. Table 3.4, Figures 3.6, 3.7 and 3.8), and are distinctly
visible in Figure 3.6. Seven bands were identified that outperformed all tested
band combinations: 425.4, 514.9, 560.1, 685.5, 731.5, 812.3 and 916.7 nanometers,
which became the benchmark (i.e. SAM 7band) against which additional
classifications were compared. Again, relevant confusion matrices are found in
Appendix B. In some instances, the best performing band pulled from each
wavelength domain was not the most frequently occurring band. The overall-
resiliency value of the benchmark, 7-band SAM classification was 86.26%,
indicating that approximately 13% of the pixels were classified differently upon
the removal of 41 of 48 bands. Examination of the SAM7band producers-
resiliency values in Appendix B reveals that eight classes, in three groups (2, 4,
16, 17, 18 and 19; 8; 11), represented the vast majority of classification change.
The first group (highly green) displayed relatively poor classification results
due to cross-classification among themselves. These classes are the most-green
training classes, and one or more bands needed to improve their differentiation
were removed. The second group (class 8), which represents a submergent

species, similarly cross-classified with other submergents. The third group

136

(class 11), which was the deep-water class, cross-classified almost exclusively
with the shallow-water class (i.e. class 7). These findings are consistent with the
more detailed discussion of SAM classification results found in Chapter 2 that

illustrated the fundamental weakness of the SAM classifier: within-class errors.

SAM is a non-probability-based classification tool that separates image
spectra according to a cumulative angular coefficient that is derived from each
spectral data point (i.e. hinge point). Fundamentally, there exists an inverse
relationship between the number of bands and the probability of a key band
associated with a subtle spectral feature needed to differentiate one or more
classes being removed. If only two bands were used, each spectrum would be
reduced to a single, decision-space vector, and all subtle features would be lost.
Inversely, the utilization of the maximum number of bands would ensure all
the recorded, subtle features would contribute during a SAM classification. A
functional classification can be generated with a significantly reduced number
of bands, but the ability to distinguish within-class species or cover types is
reduced. As mentioned in Chapter 1, Wang et a1. (2000) achieved an overall
accuracy of 84% utilizing SAM to classify relatively broad botanical associations
(e.g. reed association, Cm association) using hyperspectral imagery. The use
of broad associations reduces the potential for within-class cross-classification
because the classes have been homogenized. The eight classes that contributed

most to the SAM7band classification change could have been regrouped into

137

several more broad associations in order to improve the overall classification
outcome. However, to reduce classification errors associated with my more

botanically detailed classes, more than these seven optimal bands are needed.
Thus, a trade-off exists between the number of bands and the desired level of

accuracy.

Additional 7-band SAM classifications were shown in Table 3.5 to illustrate
how relative classification change was affected by utilizing alternative bands.
For example, Figure 3.6 indicates that a band centered at 939.9nm would be the
best choice within its occurrence peak. A 7—band classification that included
939.9nm instead of 916.7nm (i.e. SAM7-939) under-performed the SAM7band
classification by approximately 4%. The band centers 812.3nm or 823.9nm are
functionally equivalent, and can replace one another without any change is
classification outcome. Similarly, 548nm can be incorporated without a
significant reduction in classification change. The ability to shift to adjacent
bands without experiencing major classification falloff attests to the level of

information redundancy inherent to the occurrence peaks (Figure 3.6).

Two 4-band SAM classifications were also generated. The first
incorporated, as closely as possible, the band centers (but not the bandwidths)

of the IKONOS sensor (i.e. SAM4ikonos). Over 30 percent of the evaluation-

standard pixels classified differently when only these 4 bands were utilized. In

138

contrast, a second 4-band classification (i.e. SAM4band) that referenced the
best-performing band from the four highest-magnitude, occurrence peaks
shown in Figure 3.6 (i.e. 514.9, 685.5, 731.5 and 916.7nm) exhibited an overall-
resiliency of nearly 80%, approximately 13% higher than SAM4ikonos. Thus,
with respect to classifying the representative botanical communities captured
by the 4-meter Horseshoe Bay imagery, the IKONOS sensor band centers are

not as useful as these four bands.

The final two SAM classifications shown in the first section of Table 3.5
were included to further emphasize the importance of band positioning. At
first glance, one might assume an error occurred during the preparation of
Table 3.5 because it seems logical that the 30-band classification should have
outperformed its 24-band counterpart (i.e. SAM24band). However, as the
results illustrate, this was not true. The bands incorporated into the 30-band
classification (i.e. SAM30band) were the list of bands contained in the last
column of Table 3.4. The 24-band classification more simply incorporated every
other band beginning with 425.4nm. Subtle spectral features that contributed to
the extremely high overall-resiliency of the SAM24band classification are
associated with one or more bands found outside of the nine predominant
wavelength domains. The derivative and PCA methodologies did indeed
identify a subset of bands that are most responsible for the differentiation of the

21 training classes. Spectral features that remained unidentified via these

139

methodologies are needed to achieve very high classification-resiliency. A
hyperspectral bandset, one systematically covering the visible and NIR
wavelengths, is best with respect to the differentiation of the botanical ROIs

referenced in Investigation 4.

The Maximum-Likelihood classifier significantly under-performed the
SAM classifier within Investigation 4, thereby establishing that band reduction
strategies are classifier specific. The nearly 30% difference in classification-
resiliency between the SAM7band (86.26%) and the Max7band (59.46%)
classifications initially questioned the transferability of the seven optimal bands
identified above. However, the seven bands identified above also resulted in
the best-performing Maximum-Likelihood classification. Note that the
resiliency associated with the best-performing Maximum-Likelihood
classification (i.e. Max30band) is largely equivalent to the SAM4band
classification. It seemed surprising that a 4-band and 30-band classification of
the same image would generate comparable results. However, Maxirnum-
Likelihood is a probability-based classification tool unlike SAM, and it proved
to be very sensitive to the removal of bands. In light of these results, the falloff
in classification-resiliency must be associated with the inherent properties of the

classifier itself.

140

Although it is difficult, if not impossible, for a person to visualize beyond
three dimensions, computer-based tools like the Maximum-Likelihood and
Mahalanobis-Distance classifiers are able to accomplish multidimensional
evaluations. Probability-based classifiers establish a distribution of spectral
classes within a decision-space defined by the number of bands (i.e.
dimensions) utilized during the classification. Class distributions are stretched
and pulled apart along the axis of 48, independent dimensions in the case of my
48-band imagery. As the number of bands is reduced, n-dimensional space is
reduced, forcing the class means and their associated probability distributions
into a more narrowly defined space. In light of this, the removal of 41 bands
(i.e. Max7band) would likely have a significant impact on any probability-based

classifier, especially Maximum-Likelihood.

3.3.7.2 Optimal Bandwidth - Investigation 48

The seven optimal bands identified in Investigation 4A were found in
distinct, well separated wavelength domains as evident in Figure 3.6. In fact,
the wavelength interval that separates these 7 bands ranges from 45.2 to
125.4nm. Two pairs of band centers are separated by approximately 46nm:
514.9/560.1nm and 685.5/ 731.5nm. Only these two bands were close enough to
one another to have their FWHM envelopes overlap when bandwidth was
systematically increased to 8 and 12 times the original width. Thus, the first

two SAM resiliency values shown in the second section of Table 3.5 represent

141

classifications where the increase in bandwidth resulted in no overlap, and

subsequently no band elimination occurred (i.e. spectral smoothing).

The SAM-2x and SAM-6x classification results largely disproves the notion
that bandwidth is a detriment to the classification of wetland imagery, as long
as FWHM does not result in band overlap, and thereby band elimination. Note
that both of these classifications exhibited overall-resiliency values equivalent
or higher than the baseline, SAM7band classification. One could argue that
bandwidth does indeed enhance classification resiliency with respect to the
SAM classifier and strategically located bands, because the SAM7band and
SAM-6x classifications are statistically different well above the 95 percent

confidence limit.

The SAM-8x and SAM-12x results represent two 5-band classifications
because bands 560.1 and 731.5nm bands were eliminated due to their FWHM
overlap. Their overall-resiliency and Kappa values are surprisingly similar,
although they are statistically different above a 95 percent confidence limit. The
resiliency values associated with the 4-band and 7-band SAM classifications
generated within Investigation-6A were 79.64% and 86.26%, respectively. As
might be expected, the 5-band, Investigation 48 classifications fell between
these two values. Despite their exaggerated bandwidth, these two

classifications fell inline with their bandwidth-unaltered counterparts, which

142

further dismissed the notion that wide bands (without overlap) are a detriment

to the differentiation of wetland botanical communities.

The Maximum-Likelihood classifications associated with Investigation 4B
need only be briefly discussed for they exhibited the same patterns identified
via the SAM results. The baseline (Max7band) and Max-2x classifications were
functionally identical, while Max-6x exhibited slightly improved, but
statistically different, results. The 8x and 12x, Maximum-Likelihood resiliency
values were reduced as expected when two bands were eliminated. It has
already been discussed that probability-based classifiers like Maximum-

Likelihood are sensitive to the removal of bands.

3.3.7.3 IKONOS Spectral Simulation - Investigation 4C

The SAM4ikonos classification generated in Investigation 4A displayed an
overall-resiliency was 67.74%. This classification only spectrally emulated the
IKONOS sensor, for no band-widening was applied. The SAM_pseudo-
IKONOS classification (53.01 %) summarized in the third section of Table 3.5
both spatially and spectrally emulated the IKONOS sensor. A difference of 4%
separated these two classifications that only differed with respect to relative
bandwidth. It was demonstrated in Investigation 4A that shifting the band
locations with respect to a 4-band classification can result in significant

increases in resiliency (i.e. SAM4band vs. SAM4ikonos). It therefore seems

143

logical that the slight increase (i.e. 4%) in both the SAM and Maximum-
Likelihood resiliency values associated with increased bandwidth (e.g.
SAM4ikonos vs. SAM_pseudIKONOS) can be attributed to the spectral
contribution of more diagnostic bands that fall within the FWHM envelope of

the significantly wider IKONOS—like bands.

3.4 Summary and Conclusions

The 2nd derivative and PCA methodologies presented in this research were
designed to aid in the identification of those bands that are most applicable to
the differentiation of coastal wetland vegetation. This research established nine
distinct wavelength domains (i.e. Table 3.4) with respect to the resampled in-
situ spectra that emulated 48-band hyperspectral imagery taken of the study
sites. The histograms presented in Figures 3.6 and 3.8 helped to further refine
these nine domains by identifying their predominant band center (e. g. 425, 514,
560, 685, 731, 812 and 916nm). The redundancy found in the derivative and
PCA methodologies should not be looked on negatively, for it functions to

cross-validate the results generated via each methodology.

Under the constraints of today’s technology, it is not possible to capture
high spatial resolution imagery with a hyperspectral bandset. If the purpose of
the imagery is to aid in the mapping of large homogeneous cover types, larger

pixels can be recorded in order to allow many more bands to be recorded.

144

Coastal wetlands are often comprised of plant communities that are spatially
complex, which dictates the use of high-resolution (i.e. 1-2 meters) imagery if
the spatial integrity of these communities is deemed relevant. Unfortunately,
small pixels come at the expense of spectral resolution, and only a limited

number of bands would be available if small pixels were recorded.

This research established seven optimal bands (i.e. 425, 514, 560, 685, 731,
812 and 916nm) that generated classification results comparable to those
generated from 48-band, hyperspectral imagery. The reduction of band
number without significant classification change is important because it makes
it practical to utilize small pixels without fear of sacrificing the ability to
differentiate the botanical community. In addition, sensors that record a small
number of bands found within the visible and NIR regions of the spectrum are
relatively inexpensive to build. Certainly the ”best” available spectral
information is recorded by truly hyperspectral sensors, but the use of these
relatively new technologies is prohibited by their high cost (approximately
8570/ kmz) for many potential end-users. The identification of a small number
of wetland-relevant bands should significantly reduce the cost of image
acquisition, which increases the number of potential ”coastal” end-users

considerably.

145

The optimal 7 bands identified in this research show interesting similarity
to those identified in several recent vegetative studies, which might be expected
because, at some level, all plant material have similar spectra. Thenkaboil et al.
(2000) utilized detailed spectroradiometric data in order to recommend 12
bands for use in agricultural crop studies (i.e. biophysical parameters such as
Leaf-Area-Index, Wet-Biomass, and Plant height). They placed maximum
emphasis on three bands that represented maximum reflectance (550 and
920nm) and maximum absorption (682nm) exhibited by four agricultural crops.
In contrast, Carter (1998) and Elvidge and Chen (1995) identified different key
bands to be used with respect to pine vegetative indices, 674 / 780nm and
701 / 820, respectively. The similarities between these and other related studies
suggest that key bands occur that potentially have widespread applicability.
On the other hand, the differences (especially the pine studies) suggest that
optimal band selection is landscape/ target dependent since reflectance spectra
differed even at the species level. These studies did not include imagery or its
classification, which was the focus of this research. The derivative and PCA
methodologies presented in this work show great promise with respect to
identifying optimal bands for image-based wetland analyses. Much more work
is needed in order to test the resiliency of these methodologies and the
applicability of the methodologies put forth in this research across the

overwhelming diversity of coastal wetlands.

146

The replacement of the optimal-7 bands with a neighboring band (except
812 vs. 823nm) caused a statistically significant reduction in classification
resiliency. This suggests that movement of these key bands beyond a few
nanometers in either direction would result in a less botanically robust
classification. One should remember that these band locations were generated
from data collected within botanically similar wetlands despite their location
within different eco-regions. In addition, only late-August/ early-September
imagery was utilized throughout this research. It remains to be seen how static
the band locations remain when these methodologies are applied across the
entire growing season and / or within strikingly different wetland complexes

(e.g. Chesapeake Bay).

It was demonstrated that the probability-based Maximum-Likelihood
classifier is very sensitive to the removal of bands, and should therefore not be
applied when constrained by a small number of bands. In contrast, SAM is
largely insensitive to band removal, as long as key band centers that maintain
the essence of vegetative spectral signatures are maintained. Although SAM is
less sensitive to band removal, it did consistently under-perform Maximum-
Likelihood as determined in Investigations 1 through 3. Somewhat
surprisingly, bandwidth played a minor role with respect to overall
classification resiliency. Classifications generated from width-altered imagery

were equal to, or slightly better than, those generated from unaltered imagery.

147

This conclusion only holds true until the FWHM of adjacent bands began to

overlap (i.e. 8x bandwidth).

The results of this research indicate that the IKONOS band centers are less
than optimal with respect to the differentiation of coastal wetland vegetation
(August/ September imagery). SAM classification results using the four bands
associated with the predominant occurrence peaks shown in Figure 3.6 out-
performed classifications utilizing IKONOS band centers by nearly 10 percent.
This spectral limitation in combination with the 4-meter spatial resolution of
multispectral IKONOS imagery limits its applicability only to broad vegetative

classifications where detailed spatial patterns are not the focus.

148

Chapter 4: A Classification-based Assessment of the
Optimal Spatial Resolution of Coastal Wetland Imagery.

4.1 Introduction and Rationale

In the 19705, state and federal agencies increasingly looked towards aerial
photography for the purposes of wetland mapping and inventory (Gross et al.,
1989; p. 475). Film types that are sensitive to NIR electromagnetic energy are
generally preferred for wetland applications because they are more effective in
detecting standing water and determining biomass, as well as differentiating
wetland and upland vegetation (Hardisky et al., 1986). Aerial photography
(121000 to 1215,000 scales) has been found to be a fast, accurate, and relatively
inexpensive method by which wetland boundaries, botanical changes, and
aerial extent can be recorded (Mitsch and Gosselink, 2000; p. 107). A
characteristic that limits the applicability of aerial photography to detailed

vegetative mapping strategies is its non-digital format (Hardisky et al., 1986).

The launch of the Landsat-1 satellite in 1972 marked the practical beginning
of spaceborne, earth resource remote sensing (Lillesand and Kiefer, 2000; p.
373). Over the ensuing thirty years, a variety of systems, with varied spatial
resolutions, have been developed. Orbiting satellite-based sensors provide
near-simultaneous wetland imagery over large geographic areas recorded in

digital format (Gross et al., 1989; p. 476). Satellite platforms continue to provide

149

an overwhelming array of earth resource data when fitted with remote sensing
systems, and the status of these technologies continues to change as scientific
advances yield improved spacecraft and sensors (Lillesand and Kiefer, 2000; p.
373). However, regarding their utility for detailed wetland characterization, the
current systems are limited by their spatial and / or spectral resolutions. These
limitations become obvious in the context of a fundamental assumption of this
research: the spatial heterogeneity of coastal wetlands is best captured by
digital imagery with a spatial resolution at or below 5 meters. If this
assumption is valid, the majority of currently available satellite imagery is not

applicable to detailed wetland mapping.

The development of new-age airborne and satellite-based imaging systems
is making imagery with high spatial resolution more affordable and easier to
obtain. Although not widespread, the wetland literature contains recent
examples of investigations that utilized hyperspectral systems to map and
monitor plant communities (Borstad Associates, 1995; Brown and Borstad,
1999; Forsyth et al., 1998; Jupp et al., 1994; Ritter and Lanzer, 1997; Savastano et
al., 1984; Wang et al., 2000; Zacharis et al., 1992). Common to these
investigations is the exploitation of the improved spatial or spectral resolutions

to more effectively differentiate wetland vegetation.

150

Remote sensing investigations incorporating airborne hyperspectral
methodologies are ideally suited to address the spatial and spectral complexity
of many wetland environments. The remote sensing community is hopeful that
the improved resolving power of these new-age systems (i.e. spatial, spectral,
and radiometric) will lead to significant improvements in the differentiation of
wetland plant communities, thereby improving coastal mapping and
monitoring efforts by improving the accuracy of image classifications. One
fundamental question that remains unanswered is what spatial resolution is
most appropriate for capturing the botanical complexities of Great Lakes
coastal wetlands. Answering this question is an important step needed to

advance the application of hyperspectral imagery in coastal environments.

The goal of the third component of this research is to generate and classify a
series of spatially degraded, hyperspectral images in order to investigate the
relationship between spatial resolution and botanical differentiation. This goal
addresses the third research objective outlined in Chapter 1; quantitatively

discern the optimal spatial framework of wetland imagery.

4.2 Methods

4.2.1 Image Spatial Degradation — Investigation 5

 

A spatial resolution of 5-meters has been established as a threshold at or

above which detailed wetland remote sensing becomes limited at best

151

(ITRES, 2000). In order to evaluate classification change associated with spatial
degradation (larger pixel size), the baseline Horseshoe Bay, l-meter imagery
was spatially manipulated. This imagery was incrementally resampled by
increasing the functional pixel size by a factor of two, four, and eight, thereby
generating imagery that effectively emulated that recorded at spatial
resolutions of 2, 4 and 8 meters. These spatial resolutions were chosen for three
reasons: 1) they represent a systematic degradation in spatial resolution
through and beyond the established threshold; 2) they represented the
incremental doubling of pixel size; and 3) the 4-meter pixel size matched the
spatial resolution of the satellite-based IKONOS system. The 8-meter
resampling was included to verify the inapplicability of low-resolution (i.e. > 4

meters) imagery to Great Lakes coastal environments.

The EN VI (version 3.2) Image-Resizing module was found to be
inappropriate with respect to spatial (i.e. pixel size) degradation based on the
algorithm it used. When an even-numbered resampling window is desired,
there is no central pixel within the matrix of pixels to be mathematically
combined to yield a larger, resampled pixel. In order to resample a 1-meter
image into 4-meter imagery, sixteen 1-meter pixels would need to be combined
in some manner to yield a single pixel that is four times larger in both the x

(sample) and y (line) direction. Thus, the digital numbers (in my case percent

152

reflectance values) associated with a 4x4 matrix of 1-meter pixels would need to
be averaged to yield a single digital number that would then be assigned to the
larger output pixel. Unfortunately, when faced with an even-numbered input
matrix (2xZ, 4x4, etc.), ENVI assigns the mathematically unchanged value (no
average calculated) found at the upper left input pixel in the window to the
newly created, larger, output pixel. This methodology is fundamentally flawed
because it fails to average all the reflectance values associated with the input

matrix.

Supervised classifiers reference the digital number of each pixel within the
image so that a classification value (training category) can be assigned. Pixels
containing identical digital numbers would be assigned to the same
classification category. Thus, the classification result generated from one, 4-
meter pixel (e.g. DN=6.2) would be functionally and visually the same as that
generated from sixteen, contiguousI-meter pixels (all having the same value,
DN = 6.2) occupying the same space. The process by which the l-meter
Horseshoe Bay imagery was transformed into synthetically coarser resolution
imagery involved a series of image manipulations across three separate
software packages. Only integer values are maintained when an ERDAS (*.lan)
image is exported from EN VI. Subtle reflectance differences can lead to better
differentiation of botanical targets and may be found within the decimal places

of reflectance data. This necessitated multiplying the EN VI reflectance values

153

by 100. The core of the spatial resampling process summarized in Figure 4.1
was the Blockmean function in ArcInfo-GRID. This function partitions the input
grid (1-meter pixels in my case) into blocks (2x2 4x4 and 8x8 in this study),
calculates the mean value for all of the cells (pixels) in the block and writes this
value to each cell in the corresponding blocks in the output grid (ESRI, 2001).
Ultimately, three 1-meter, output images (synthetic-2m, synthetic-4m and

synthetic-8m) were generated.

Because the original spatial resolution was preserved, the 24 R015 that
served as training data for Investigations 1 and 2 (Table 2.6) could also be used
for Investigation 4. Similar to the approach taken in Chapter 3, these ROIS were
used as training sites and both a SAM and Mahalanobis evaluation-standard
was generated through the classification of the spatially unaltered, unrotated

Horseshoe Bay imagery.

After the generation of the two evaluation-standards was completed, three
identically configured SAM and Mahalanobis-Distance classifications were
generated from the three spatially-altered images. Mahalanobis-Distance
replaced Maximum-Likelihood in Investigation 5 because the spatial
resampling methodology reduced training—class variance beyond the allowable
Maximum-Likelihood threshold. No operational parameters were entered for

the Mahalanobis-Distance classifications, which forced all pixels to be classified.

154

 

I ENVI Image Format I

 

 

ENVI Export (integer only)
Multiply by 100

 

 

 

ERDAS (*Jan) Image Format
l-meter Pixels

 

ERDAS Imagine
Import

 

 

 

ERDAS (*.img) Image Format
l-meter Pixels

 

ERDAS Imagine
Save As Grid Stack

 

 

 

 

ARCINFO (*stk) Grid Stack
1-meter Pixels

 

ARCINFO
Open Grid Stack

 

 

 

ARCINFO Grid Format
l-meter Pixels (18 grids)

 

Image Processing

 

 

 

ENVI Image Format
Resampled Pixels

 

T Divide by 100

 

 

 

 

ENVI Image Format
Resampled Pixels

 

ENVI Import
Resampled Pixels

 

 

 

ERDAS (*.img) Image Format
Resampled Pixels

I

 

ARCINFO
Export Imagine Format

 

 

 

 

 

 

 

Blockmean(sag1m, rectangle, 2,2, Data
Blockmean (sagl m, rectangle, 4,4, Data
Blockmean (saglm, rectangle, 8,8, Data

 

 

ARCINFO (*stk) Grid Stack
Resampled Pixels

 

ARCINFO
Recreate Grid Stack

 

 

> ARCINFO Grid Format

Resampled Pixels (18 grids)
Pseudo-2m, 4m, and 8m

 

 

 

Figure 4.1 Spatial Manipulation, Image Processing Schematic

155

Confusion matrices were generated through the comparison of the 6
spatially-altered classification results (i.e. 2xZ-SAM, 4x4-SAM, 8x8-SAM, 2x2-
Mahal, 4x4-Mahal and 8x8-Mahal) with respect to their appropriate image-wide
evaluation standards. Again, the assumption was made that the two
evaluation-standards represent an unverified depiction of the botanical
community of Horseshoe Bay. The botanical accuracy of this assumption is
irrelevant with respect to the focus of this investigation because only relative
change associated with spatial degradation was evaluated. The examination of
the relative change with any output class reflects the classification change (i.e.
class confusion) resulting from the spatial degradation. If the spatial
manipulations had little affect on the classification results, the pixel counts
within each class as summarized by the confusion matrix parameters would
remain relatively constant. If significant changes in these values occurred, then
it was presumed that the spatial resampling caused the spectral characteristics

of a significant number of pixels to be classified differently or to be unclassified.

Fundamentally, as pixel size is increased within a heterogeneous region, the
potential for two or more spectral classes to be included within a single pixel
(mixed pixels) increases. Pixels associated with relatively large, homogeneous
botanical zones (i.e. classes) would be affected the least. Inversely, botanical
zones that have relatively small, areal extents would experience the greatest

change (i.e. reduced producers accuracy) with respect to the evaluation

156

standards. At some level, the discrimination of botanical communities becomes
impossible as the spectral nature of two or more spectrally distinct classes
(pixels) are homogenized into a single, unclassifiable pixel. For example, linear
bulrush (Scirpus validus) zones exist along the edges of dynamite cut, and are
only a few meters in width. If the baseline pixel size is increased to eight
meters, the spectral characteristics of this zone would be homogenized with
neighboring cover types, rendering their identification difficult, if not
impossible. The three classifications generated via both classifiers represent a
systematic progression towards more spatially coarse imagery. Accordingly,
the consecutive examination of their related confusion matrix parameters
provided insight into the spatial resolution needed to differentiate the botanical

communities in my Great Lakes wetland imagery.

4.3 Results and Discussion

4.3.1 Spectral Angle Mapper — Investigation 5

Table 4.1 contains the truncated confusion matrix parameters generated for
the three SAM classification runs associated with Investigation 5. Each
subsequent column represents classification change resulting from the
systematic doubling of spatial resolution. The overall-resiliency and related
Kappa values indicate a systematic increase in classification change (i.e.

decrease in resiliency) resulting from the increase in functional pixel size. The

157

 

 

 

 

 

2x2-SAM 4x4-SAM 8x8-SAM
Overall Resiliency: 77.09% 65.11% 50.31 %
Kappa Coefficient: 0.7536 0.6251 0.4696
Producers Producers Producers
Resiliency Resiliency Resiliency
Image Class (Percent) (Percent) (Percent)
0 Unclassified 11.62 0.86 0.1
1 chara 85.41 67.04 34.23
2 phragmites-clump 57.31 52.06 58.93
3 scival-dense 30.86 15.09 12.78
4 scival-clump 63.75 48.47 24.69
5 typha-soi12 84.09 76.43 69.51
6 sagrigida60/40 72.97 63.55 57.18
7 beachsand 82.7 73.12 61.76
8 shallowh20 64.52 51.81 42.08
9 hetdub 79.04 67.04 61.84
10 sagrigida 59.61 46.2 30.53
11 vallisneria 94.51 87.07 80.55
12 deeph20 93.98 88.85 87.45
13 scivalbrown 73.51 58.94 63.99
14 leersia 66.1 50.97 26.54
15 Muck 46.36 13.9 5.64
16 saggram/chara 87.71 81.84 71.08
17 typhah20 54.92 23.13 4.45
18 nuphar 56.85 34.09 26.18
19 eleocharis 51.17 36.8 29.83
20 impatians 68.18 48.46 32.05
21 bluejoint 56.72 43.84 29.49
22 cutgrass-eleocharis 78.19 64.48 37.01
23 eleocharis-sand 81.22 69.65 16.16
24 typha-phrag-scival 73.25 67.32 62.2

 

Table 4.1.

158

Summary Table, SAM Spatial Degradations

 

Kappa and overall-resiliency values are very similar, indicating that chance

agreement played little role.

The degradation to a 2-meter, functional pixel size caused approximately
23% of all pixels to be classified into a different class compared to the 1-meter
evaluation standard. Spatial degradation to a 4-meter equivalent further
reduced overall-resiliency to approximately 65%, indicating that 35% of all
pixels were classified into a different class. Based on the assumption that the
evaluation standard represents a snapshot of the botanical community, the vast
majority of these errors would represent classification errors. Nearly half of all
the pixels were classified differently as a result of increasing the functional pixel
size to 8 meters. As expected, overall classification outcome is extremely
sensitive to pixel size. The 2-meter results represent a change of 23%, which
could be considered acceptable depending on the nature of the community
being classified. However, few would argue that a 50% change would not

hinder the detailed mapping of coastal wetlands.

A better appreciation for the role of pixel size with respect to wetland
image classification is gained through the examination of the class-specific
producers-resiliency values shown in Table 4.1. The producers-resiliency
values found in the first column (i.e. 2xZ-SAM) represent the percentage of

class—specific pixels that remained consistent with the l-meter, evaluation

159

standard. For example, only 6.02% of the evaluation-standard, deep-water
pixels (i.e. class 12; deeph20) were classified differently (resiliency = 93.98%)

when the functional pixel size was degraded to 2 meters.

Classes 11 and 12 represent cover types that were the least affected by
spatial degradation. These classes (Vallisneria americana and deep-water,
respectively) represent large homogeneous regions. In the interior of such a
region, pixel size can be increased dramatically without fear of encountering a
spectrally different cover type, and only its edges have any potential for
spectral confusion. Due to the spectrally suppressive nature of water, it would
take a considerable amount of a spectrally different cover type (e.g. emergent
vegetation) to alter a classification result. Class 1 (Chara sp.) is also a spatially
extensive class, but it was less resilient than classes 11 and 12. The Chara
evaluation-standard pixels were largely found within the eastern half of
Horseshoe Bay proper. Unlike the deep-water and Vallisneria classes, this
region is significantly interspersed with pockets of deeper water, clumps of
emergent vegetation, and floating leaf plants. Accordingly, the potential for
spectral overlap as pixel size increases is much higher for this class. In the 2-
meter results, only about 15% change occurred, which in among the best-
performing classes. However, when pixel size was increased to 8 meters and
the various cover types mentioned above are encompassed with one pixel,

about 65% of the Chara pixels were classified differently.

160

The very low resiliency value associated with class 7 (beach sand) might at
first seem surprising, considering that this class has been previously described
as being a spectrally distinct, easily-classified cover type. So why did this class
not exhibit resiliency as pixel size increased? The answer lies in its relative
position within the landscape. The majority of the beach sand pixels in the
evaluation standard are associated with narrow, linear sand deposits along the
southern shoreline of Middle Grounds Island. This ribbon of sand is
approximately 25 meters in width, and is interspersed with regions of
vegetation and deadfall. At the 2-meter pixel size, this class displays a
relatively minor change (82.7% resiliency), but its percentage of
misclassification steadily increases with pixel size. The linear nature of the
beach sand zone gives it a relatively low area-to—perimeter ratio. As pixel size
increases, the edge pixels overlap considerably with the surrounding cover

types, resulting in classification change.

4.3.2 Mahalanobis Classifications - Investigation 5

 

Table 4.2 contains the truncated confusion matrix parameters generated for
the three Mahalanobis classification runs associated with Investigation 5.
Again, each subsequent column represents classification change resulting from
the systematic doubling of spatial resolution. The overall-resiliency and related
Kappa values show a systematic increase in classification change resulting from

the increase in functional pixel size. Kappa and overall-resiliency values are

161

 

 

 

 

 

 

2x2-Mahal 4x4-Mahal 8x8-Mahal
Overall Resiliency: 69.92% 42.51% 0.40%
Kappa Coefficient: 0.6771 0.3847 -0.0571
Producers Producers Producers
Resiliency Resiliency Resiliency
Image Class (Percent) (Percent) (Percent)
0 Unclassified 0 0 0
1 chara 50.56 43.68 0
2 phragmites-clump 56.52 48.22 0
3 scival-dense 73.24 52.41 0
4 scival-clump 72.6 29.85 0
5 typha-soi12 58.26 13.47 4.66
6 sagrigida60/40 20.76 14.8 0
7 beachsand 95.76 55.81 0
8 shallowh20 76.46 37.75 0
9 hetdub 77.2 54.5 0
10 sagrigida 72.08 16.97 0
11 vallisneria 85.93 47.3 0
12 deeph20 96.6 86.84 0
13 scivalbrown 84.43 76.98 1.83
14 leersia 68.66 29.13 0
15 muck 35.66 6.15 0
16 saggram/chara 79.83 30.86 0
17 typhah20 64.1 30.9 0
18 nuphar 55.51 20.8 0
19 eleocharis 40.77 9.01 1.73
20 impatians 36.93 23.98 1.87
21 bluejoint 64.64 40.82 0
22 cutgrass-eleocharis 49.86 18.81 0
23 eleocharis-sand 63.2 35.8 0.01
24 typha-phrag-scival 77.41 68.85 0

 

Table 4.2.

162

Summary Table, Mahalanobis-Distance Spatial Degradation.

 

very similar, indicating that chance agreement played little role. The
degradation to a 2-meter, functional pixel size caused approximately 30% of all
pixels to be classified differently compared to the 1-meter evaluation standard.
At the 4-meter pixel size, nearly 58% of the pixels were classified differently.
Assuming that the evaluation standard represents an accurate depiction of the
botanical community, the majority of these pixels would represent classification
errors. Virtually all of the pixels were classified differently as a result of
increasing the functional pixel size to 8 meters. In fact, the negative Kappa
Coefficient for the 8-meter classification indicates that it is less ”accurate" than

one generated randomly.

The overall classification outcome of the Mahalanobis classifier is extremely
sensitive to spatial resolution changes, more so than the SAM classifier. The
probability distribution centered on each training class mean was expanded as
the spectral overlap of all image pixels were homogenized within the spatial
degradation process. The homogenization of image reflectance values increases
the spectral similarity of all classes, and only the most distinct or heterogeneous
classes displayed any resiliency. This spectral smoothing process had a lesser
effect on the physically-based (i.e. non-probability-based) SAM classifier
discussed in the previous section. Because of the extremely poor results
associated with the 8-meter classification, it seems likely that the amount of

training class variance reached a level that rendered the classifier unusable.

163

A better appreciation for the role of pixel size with respect to the
Mahalanobis-Distance classification is gained through the examination of the
class-specific, producers-resiliency values shown in Table 4.2. The producers-
resiliencies found in the first column (i.e. 2xZ-Mahal) represent the percentage
of class-specific pixels that remained consistent with the 1-meter, evaluation
standard. For example, only 3.40% of the evaluation-standard, deep-water
pixels (class 12; deeph20) were classified differently (resiliency = 96.6%) when

the functional pixel size was degraded to 2 meters.

Just as in the analogous SAM examples provided earlier, classes 7, 11 and
12 represent the cover types least affected by spatial degradation. Classes 11
and 12 (Vallisneria americana and deep-water, respectively) represent large
homogeneous regions, and pixel size can be increased dramatically without fear
of encountering a spectrally different cover type. Due to the spectrally
suppressive nature of water, it would take a considerable amount of a
spectrally different cover type (e. g. emergent vegetation) to alter a classification

result.

The best performing class with respect to the 2—meter Mahalanobis-Distance
classification was class 7 (beach sand). Probability-based classifiers perform
well on spectrally distinct classes (i.e. magnitude separation) like class 7, until

the homogenization of reflectance overcomes the probability threshold. This is

164

evident by the significant resiliency falloff in class 7 between the 2- and 4-meter
classifications. Class 24, an inherently mixed class, remained relatively
unaffected by a decrease in spatial resolution from 2 to 4 meters.
Homogenization of this class through spatial degradation should have a
relatively minor affect because it is ”naturally” homogenized (i.e. mixed). The
class-specific results associated with the 8-meter classification only serve as an
indication that the variance threshold of this classifier was exceeded, and little

botanically relevant information can be extracted.

It was previously established (ITRES, 2000) that a spatial resolution of 5
meters or less is appropriate for wetland mapping strategies. Interestingly, this
research indicates that even 2-meter pixels failed to adequately capture (23%
and 30% change) the spatial mosaic of plant species found in Horseshoe Bay.
Less heterogeneous wetland plant communities could probably be classified
adequately with 4-meter imagery, but pixel sizes above this level would not be

recommended.

As described in the previous section, the falloff in classification resiliency is
inherently tied to the spatial configuration of the plant community.
Classification of an image taken of a 500 acre Phragmites marsh would be largely
independent of pixel size (i.e. 1-10 meters) because of the lack of spatial

heterogeneity. As the spatial complexity of the target increases, the spatial

165

resolution required to map it must also increase. Examination of Figure E6 in
Appendix E illustrates the inverse relationship between spatial heterogeneity
and pixel size. A 50x50 pixel portion of the 1-meter, Horseshoe Bay SAM
classification is shown along with its analogous 2x2, 4x4, and 8x8 spatially-
altered counterparts. This series of images illustrates the systematic elimination
of classes and the reduction of class interspersion. The light blue class best
illustrates class elimination because it was completely absent in the 4- and 8-
meter images despite being prevalent in the 1-meter baseline classification. The
yellow class best illustrates the loss in class interspersion, because its rather
complex spatial mosaic found in the baseline classification has been reduced to
non-descript blocks in the coarser imagery. Obviously, the total acreage of this
class has been inflated by the spatial degradation and the small-patch nature of

this class has been totally lost.

4.4 Summary and Conclusions

Care was taken to ensure that the study sites used throughout this research
reflect the potential botanical complexity (both spatial and spectral) of Great
Lake coastal wetlands. Investigation 5 has demonstrated that high spatial
resolution imagery (i.e. 1-2 meters) is paramount to accurately map and

monitor these critical environments.

166

Chapter 5: CONCLUSIONS AND FUTURE RESEARCH

The entirety of this research focused on evaluating the applicability of high
spatial and spectral resolution imagery with respect to the mapping and
monitoring of Great Lakes Coastal wetlands. This focus was explored through
a series of investigations that addressed three underlying themes: 1) classifier
performance and transferability (Investigations 1-3), 2) optimal spectral

resolution (Investigation 4), and 3) optimal spatial resolution (Investigation 5).

5.1 Classifier Performance and Transferability

It was established in previous chapters that coastal wetlands can be
botanically complex, exhibiting a rich mosaic of plant species / assemblages that
are distributed along the predominant hydrologic gradient. It is this
characteristic that typically makes wetlands a challenging remote sensing target
due to the relative spectral similarity of the predominantly vegetated landscape.
Botanical training sites (Regions-of-Interest) were delineated that represented
the predominant cover types found in each study site during the 2000 and 2001
field campaigns. These ultimately served as the basis for multiple classification

comparisons.

167

The probability-based Maximum-Likelihood classifier was the best
performing classifier with respect to the differentiation of the field-delineated
ROIs. Approximately 87% of the accuracy assessment pixels (easterly ROIs)
were correctly classified by Maximum-Likelihood, which was nearly ten
percent higher than the comparable SAM classification. It was shown that
probability-based classifiers are magnitude-dependent (y-axis separation), and
their associated errors are typically derived from within-class misclassifications.
Maximum-Likelihood displayed the greatest potential for differentiating the
complex botanical communities found in the Lake Huron wetlands analyzed in
this study. From a practical standpoint, the Maximum-Likelihood classifier
would be ideal for delineating major botanical classes when some degree of
within-class error (e. g. confusion of similarly structured emergent species) can

be tolerated.

The baseline SAM classifier under-performed the Maximum-Likelihood
classifier by nearly ten percent, but showed promise. This physically-based
classification tool tended to confuse classes that displayed similar greenness
(i.e. shape of the reflectance signature), and differed from Maximum-Likelihood
in that it is insensitive to magnitude differences. Both the Les Cheneaux and
Horseshoe Bay imagery was taken as the botanical community neared the end
of its growing season, which likely contributed to the success of the SAM

classifier because the plant community was in various stages of dieback.

168

Imagery taken when the community was at ”peak green" would likely lead to

an increase in greenness confusion by the SAM classifier.

The strengths of both classification tools could be used in combination to
enhance classification accuracies. Maximum-Likelihood should be used to
parse an image into the predominant classes. Then, complex classes could be
isolated via image masking in order to capitalize on the ability of the SAM
classifier to differentiate similar spectra based on relatively subtle spectral
features that are lost in the magnitude-based approach. This is certainly an area

that needs to be explored further within the context of coastal wetlands.

From a practical standpoint, this research indicates that PCA rotation
should routinely be utilized in combination with probability- based
classification tools. Fundamentally, PCA rotation concentrates target variance
into a small number of independent channels. The resulting dimensionally-
reduced signatures are better differentiated by magnitude-sensitive classifiers
because the image spectra are more equitably distributed (i.e. more class

separation).

The Prentiss Bay imagery was incorporated into this research in order to
test the transferability of ”methods” developed with respect to the Horseshoe

Bay, l-meter imagery. Generally, the outcome of like-configured Prentiss and

169

Horseshoe Bay classifications were very similar, and trends were common
throughout. PC-rotated Maximum-Likelihood and un-rotated SAM
methodologies as outlined in this research were not site dependent, and could
likely provide the basis for basin-wide wetland mapping, especially in

combination.

Although not completely understood, it seems likely that some type of
systematic radiometric noise was inherent to the AISA imagery, as evident by
the poor SAM and Maximum-Likelihood classification results generated from
the raw AISA imagery. PC-rotation is known to sequester such systematic
noise into lower-order dimensions, and this was indeed the case in this research
as evident by the sharp increase in the accuracy of the Maximum-Likelihood
classification (41 percentage points) after PC-rotation of the AISA imagery.
Results generated from a probability-based classifier in conjunction with both
raw and rotated imagery could function as a viable test for subtle, system-level

noise.

5.2 Optimal Spectral Resolution

Under the constraints of today’s technology, it is not possible to capture
high spatial resolution imagery with a hyperspectral bandset. If the purpose of
the imagery is to aid in the mapping of large homogeneous cover types, larger

pixels can be recorded in order to allow many more bands to be recorded.

170

Coastal wetlands are often comprised of plant communities that are spatially
complex, which dictates the use of high-resolution (i.e. 1-2 meters) imagery if
the spatial integrity of these communities is deemed relevant. Unfortunately,
small pixels come at the expense of spectral resolution, and only a limited

number of bands would be available.

The 21nd derivative and PCA methodologies presented in Chapter 3 were
designed to identify those bands that are most applicable to the differentiation
of coastal wetland vegetation. This research established seven optimal bands
(i.e. 425, 514, 560, 685, 731, 812 and 916nm) that generated SAM classification
results comparable to those generated from 48-band, hyperspectral imagery.
The reduction of band number without significant classification change is
important because it makes it practical to utilize small pixels without fear of
sacrificing the ability to differentiate the botanical community. In addition,
sensors that record a small number of bands found within the visible and N IR
regions of the spectrum are relatively inexpensive to build. Certainly the ”best"
available spectral information is recorded by truly hyperspectral sensors, but
the use of these relatively new technologies is prohibited by their high cost
(approximately 8750/ km?) for many potential end-users. The identification of a
small number of wetland-relevant bands should significantly reduce the cost of
image acquisition, which increases the number of potential ”coastal” end-users

considerably. Much more work is needed in order to test the applicability of the

171

derivative and PCA methodologies put forth in this research across the
overwhelming diversity of coastal wetlands. However, consistency across two

eco-reaches of the upper Great Lakes has been demonstrated.

The in-situ- and PC-based derivative methodologies, along with the visual
inspection of the PC-loading curves, formed the foundation for the selection of
the optimal 7 bands. The predominant wavelength domains were identified via
all three approaches indicating that some of these efforts were redundant.
However, a variety of less-significant wavelength domains were identified via
each method. The bulk of the effort associated with these methodologies
involved the preparation of the in-situ signatures and the creation of the
derivative spreadsheet templates. Once these preliminary tasks are completed,
a variety of methodologies can be easily applied (e.g. covariance-PCA,
correlation-PCA, in-situ derivatives and loading derivatives), so it is best to
utilize them all in case one identifies a key spectral band missed by the others.
The associated redundancy should not be looked on negatively because it

cross-validates the results generated via each methodology.

Despite the similarity of the I-4 overall-resiliency values, the replacement of
the optimal-7 bands with a neighboring band (except 812 vs. 823nm) caused a
statistically significant reduction in classification resiliency. This suggests that

movement of these key bands beyond a few nanometers in either direction

172

would result in a less botanically robust classification. One should remember
that these band locations were generated from data collected within botanically
similar wetlands despite their location within different eco-regions. In
addition, only late-August/ early-September imagery was utilized throughout
this research. It remains to be seen how static the band locations remain when
these methodologies are applied across the entire growing season and / or

within strikingly different wetland complexes (e.g. Chesapeake Bay).

It was demonstrated that the probability-based Maximum-Likelihood
classifier is very sensitive to the removal of bands, and should therefore not be
applied when constrained by a small number of bands. In contrast, SAM is
largely insensitive to band removal, as long as key band centers that maintain
the essence of vegetative spectral signatures are maintained. Although SAM is
less sensitive to band removal, it did consistently under-perform Maxirnum-

Likelihood when using many bands.

Surprisingly, bandwidth played a minor role with respect to overall
classification resiliency. Classifications generated from bandwidth-altered
imagery were equal to, or slightly better than, those generated from unaltered
imagery. This conclusion only holds true until the FWHM of adjacent bands

began to overlap and bands were eliminated (i.e. 8x bandwidth).

173

The results of this research indicate that the IKONOS band centers are less
than optimal with respect to the differentiation of coastal wetland vegetation
(August/ September imagery). SAM classification results using the bands
associated with the 4 predominant occurrence peaks out-performed
classifications utilizing IKONOS band centers by nearly 10 percent. This
limitation, in combination with the 4-meter spatial resolution of multispectral
IKONOS imagery, limits its applicability to broad vegetative classifications

where detailed spatial patterns are not the focus.

5.3 Optimal Spatial Resolution

This research established the optimal spatial resolution of coastal wetland
imagery to be at or below 2 meters. The spatial complexities of the studied
coastal wetlands merit the use of such imagery when an accurate depiction of
the botanical community is desired. When more broadly defined vegetative
classes are appropriate, 4-meter imagery can perform adequately. Next-
generation satellite imaging systems are being deployed that will provide
multispectral imagery with improved spatial resolutions (e.g. QuickBird, 2.44-

meter pixels; www.digitalglobe.com). The four, broad bands offered by

 

QuickBird are functionally identical to those of the IKONOS sensor, which were
shown to be less than optimal for coastal wetland studies. Despite its spectral
limitations, classification accuracies generated with QuickBird imagery would

likely increase due to the significant reduction in pixel size. At the time of this

174

writing, only programmable, airborne imaging systems encompass the spatial
and spectral characteristics needed to map and / or monitor Great Lake coastal

wetlands with high classification accuracy.

5.4 Limitations of the Technology

The high accuracy levels at which the methodologies presented in this
research delineated the botanical ROIs was a very positive outcome. The
objectives of this research focused only on the ability to differentiate the
predominant botanical communities found at both study sites, and represented
the first step in the evaluation of hyperspectral technologies in Great Lakes
coastal wetlands. The improved spatial and spectral resolution of
hyperspectral imaging systems show great potential for improving wetland

mapping and monitoring in light of the results of this research.

The accurate differentiation of select cover types does not directly equate to
improved mapping and monitoring, and hyperspectral approaches are not
without limitations. An image-wide, PC-rotated, Maximum-likelihood
classification of the 1-meter Horseshoe Bay CASI-II imagery is presented in
Figure 5.1 in order to illustrate three limitations: 1) class confusion, 2) un-

represented classes and 3) radiometric path differences.

175

 

'3'

\ WI

unclassified

chara
phragmites-clump
scival-dense
scival-clump
typha-soilZ
sagrigida60/40
beachsand
shallowh20
hetdub

sagrigida
vallisneria
deeph20
scival-brown
leersia

muck
saggram/chara
typhah20

nuphar

eleocharis
impatians
bluejoint
cutgrass-eleocharis
eleocharis-sand
typha-phrag-scival-mix

Figure 5.1. Horseshoe Bay; PC654321 Max.-Likelihood Classification.

176

Zone 1 (Figure 5.1) highlights instances where the cattail training class
cross-classified with regions that contained dense stands of other tall-stature
emergents and graminoids in Horseshoe Bay. The hyperspectral techniques
applied throughout this research were not robust enough to separate this
spectrally similar training class from others on an image-wide basis. Zone 2
(also a misclassification) highlights a linear willow thicket (Salix sp.) near
dynamite cut (Horseshoe Bay). This cover class was not visited during field
verification efforts, and therefore was not included in the 24 training classes
utilized for the supervised classifications. The lack of a training class to
represent this cover type means that it must remain unclassified or be
completely cross-classified. Zone 3 highlights the boundary between two
adjacent flight paths that were combined into one image mosaic by ITRES, Inc.
The predominant cover type in this shallowly inundated portion of Horseshoe
Bay was Chara sp. Radiometric differences between these two image paths
caused the Chara class (left path) to be systematically misclassified in one swath
(right) while being correctly classified in the swath containing the region-of-
interest (left). The botanical community in this zone was homogeneous across

both images, indicating that the differences are an artifact of the imagery itself.

The methodologies presented in this research represent the first step in
exploring the role hyperspectral technologies can contribute to coastal Great

Lakes wetland mapping and monitoring. If the results of this work had

177

indicated that few detailed botanical classes could be differentiated, there
would be little hope for hyperspectral technologies to advance coastal mapping
and monitoring. My results indicate there is indeed hope, although limitations
exist that are dictated by the nature of current hyperspectral technologies (e.g.
band number), the spectral complexities of wetland plant communities (e. g.

BRDF, similar greenness), and the methodologies outlined in this research.

5.5 Recommendations for Future Research

Because this research is the ”first step”, the potential for future research is
great. Improved sensor technologies will result in better spatial, spectral and
radiometric resolution. Methodologies similar to those outlined in this research
should be applied to this ”newer” imagery to test if indeed such advancements
will help better define coastal wetland plant communities. A more rigorous
radiometric calibration approach (i.e. atmospheric correction models) should be
applied to my current imagery data and analogous classifications should be
generated in order to evaluate any classification differences. Additional in-situ
spectra should be collected that capture the spectral essence of wetland plant
communities throughout the growing season and under varied environmental
conditions. A more complete spectral library of Great Lakes wetland plant
communities will hopefully lead to better differentiation and improve mapping
and monitoring strategies. In addition, a larger spectral data set could

potentially reveal key spectral domains that were missed in this research that

178

would improve differentiation. A more complete field-verification campaign
should be implemented that would provide a wetland-wide ground-truth
standard. This would allow a more formal accuracy assessment strategy to be
implemented (i.e. randomly generated, image wide), providing the framework
for a statistically defensible mapping of one or more Great Lakes coastal
wetlands. This would also eliminate cover types that were not represented
within the training data used to generate classification results. Ideally, these
mapping efforts would encompass a variety of wetlands, further testing the

applicability of these technologies across the Great Lakes region.

179

Appendix A

180

 

 

Horseshoe Bay l-meter PCA Image

ENVI's Statisitics Output - 6 Dimensions

Band

Omfl©m1¥>wNfl

17
18

Prentiss Bay l-meter PCA Image
EN Vl's Statisitics Output - 6 Dimensions

Band

\OQVO‘U'IhBOOtth—I

Min

Max

-0.693 34.096

0

23

1
O

0
-0.892
-1 .082

Min

-1.191
0
0
0

41.779
42.297
42.859
43.443
44.260
45.287
44.650
44.473
45.160
47.039
49.493
49.816
54.896
56.787
57.009
57.471
60.458

Max

100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100

Mean

2.625

4.448

4.492

4.318

4.172

3.935

3.993

3.808

3.442

3.414

4.091

7.657

9.851
13.549
14.747
15.010
15.405
15.594

Mean

0.831
1.104
1.137
1.131
1.071
1.067
1.048
1.015
0.967
0.978
1.140
1.405
1.655
2.149
2.317
2.424
2.517
2.683
2.705
2.749

m

1.182
1.838
1.867
1.855
1.860
1.880
1.934
1.936
1.921
1.924
2.069
3.484
5.569
10.016
10.481
10.741
12.778
13.565
Total =

54452:

2.020
2.578
2.662
2.722
2.700
2.727
2.736
2.694
2.620
2.645
2.942
3.453
4.054
5.588
6.143
6.461
6.706
7.117
7.154
7.292

Percent Var. Cumm.

Eigenvalues Explained

 

711.311
34.892
5.690
0.829
0.678
0.162
0.119
0.100
0.025
0.019
0.010
0.004
0.003
0.002
0.002
0.001
0.001
0.001

753.848514

383.581
31.387
0.677
0.203
0.143
0.058
0.037
0.032
0.027
0.019
0.016
0.012
0.011
0.009
0.008
0.008
0.008
0.008
0.007
0.006

94.357
4.628
0.755
0.110
0.090
0.022
0.016
0.013
0.003
0.002
0.001
0.001
0.000
0.000
0.000
0.000
0.000
0.000

100

Variance

94.357
98.986
99.741
99.851
99.941
99.962
99.978
99.991
99.994
99.997
99.998
99.999
99.999
99.999
100.000
100.000
100.000
100.000

Percent Var. Cumm.

Eigenvalues Explained

92.150
7.540
0.163
0.049
0.034
0.014
0.009
0.008
0.006
0.004
0.004
0.003
0.003
0.002
0.002
0.002
0.002
0.002
0.002
0.001

Variance

92.150
99.690
99.853
99.902
99.936
99.950
99.959
99.967
99.973
99.978
99.982
99.984
99.987
99.989
99.991
99.993
99.995
99.997
99.999
100.000

 

181

 

 

 

Horseshoe Bay 4-meter PCA Image

ENVI's Statisitics Output - 6 Dimensions

13mg! Min Miles
1 0 0.166
2 0 0.186
3 0 0.208
4 0 0.234
5 0 0.248
6 0 0.250
7 0 0.250
8 0 0.248
9 0 0.246
10 0 0.249
11 0 0.252
12 0 0.257
13 0 0.265
14 0 0.272
15 0 0.277
16 0 0.282
17 0 0.287
18 0 0.288
19 0 0.295
20 0 0.305
21 0 0.300
22 0 0.312
23 0 0.298
24 0 0.307
25 0 0.325
26 0 0.340
27 0 0.330
28 0 0.341
29 0 0.335
30 0 0.350
31 0 0.359
32 0 0.384
33 0 0.378
34 0 0.385
35 0 0.388
36 0 0.390
37 0 0.398
38 0 0.403
39 0 0.402
40 0 0.404
41 0 0.410
42 0 0.412
43 0 0.414

Mean
0.051
0.048
0.047
0.044
0.044
0.044
0.043
0.043
0.044
0.047
0.052
0.056
0.058
0.057
0.055
0.053
0.052
0.050
0.049
0.050
0.048
0.046
0.044
0.043
0.048
0.066
0.089
0.114
0.136
0.150
0.153
0.168
0.166
0.171
0.175
0.177
0.181
0.184
0.184
0.187
0.191
0.193
0.192

Total =

Stdev

0.005
0.005
0.006
0.006
0.007
0.007
0.008
0.008
0.008
0.009
0.011
0.012
0.013
0.013
0.012
0.012
0.013
0.013
0.013
0.014
0.014
0.015
0.015
0.015
0.016
0.018
0.025
0.040
0.059
0.070
0.073
0.080
0.079
0.080
0.080
0.081
0.087
0.092
0.094
0.096
0.100
0.102
0.102

416.256391

100

Percent Var. Cumm.
Eigenvalues Explained

 

0.174
0.003
0.002
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000

97.238
1.593
1.009
0.062
0.031
0.019
0.012
0.010
0.008
0.006
0.002
0.002
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000

Variance

97.238
98.831
99.840
99.901
99.932
99.951
99.964
99.974
99.982
99.988
99.990
99.992
99.993
99.994
99.995
99.996
99.996
99.997
99.997
99.998
99.998
99.999
99.999
100.000
100.000
100.000
100.000
100.000
100.000
100.000
100.000
100.000
100.000
100.000
100.000
100.000
100.000
100.000
100.000
100.000
100.000
100.000
100.000

 

182

 

 

 

 

 

 

44 0 0.419 0.198 0.106 0.000 0.000 100.000
45 0 0.430 0.206 0.112 0.000 0.000 100.000
46 0 0.416 0.192 0.103 0.000 0.000 100.000
47 0 0.459 0.230 0.119 0.000 0.000 100.000
48 0 0.439 0.228 0.118 0.000 0.000 100.000
Total = 0.17854 100 100
Mismer Bay l-meter PCA Image
ENVI's Statisitics Output - 6 Dimensions
Percent Var. Cumm.
13334 _Mi_n _ng M Stdev Eigenvalues Explained Variance

1 -0.377 100 0.707 1.766 332.074 92.792 92.792

2 0 100 0.921 2.259 23.749 6.636 99.428

3 0 100 0.945 2.325 1.499 0.419 99.847

4 0 100 0.919 2.308 0.219 0.061 99.908

5 -0.074 100 0.792 2.072 0.111 0.031 99.939

6 0 100 0.773 2.061 0.044 0.012 99.951

7 -0.196 100 0.744 2.025 0.030 0.008 99.960

8 -0.109 100 0.705 1.972 0.028 0.008 99.968

9 -0.034 100 0.660 1.895 0.022 0.006 99.974
10 -0.609 100 0.664 1.913 0.016 0.005 99.978
11 -0.422 100 0.779 2.174 0.013 0.004 99.982
12 0 100 0.977 2.688 0.010 0.003 99.985
13 -0.192 100 1.184 3.349 0.009 0.003 99.987
14 -0.884 100 1.677 5.102 0.007 0.002 99.989
15 -0.712 100 1.892 5.868 0.007 0.002 99.991
16 -0.753 100 2.020 6.308 0.007 0.002 99.993
17 -1.662 100 2.117 6.577 0.006 0.002 99.995
18 -1.574 100 2.244 6.978 0.006 0.002 99.997
19 -1.759 100 2.263 7.022 0.006 0.002 99.998
20 -3.279 100 2.292 7.103 0.005 0.001 100.000

Total = 357.868 100
Mismer Bay 4—Meter IKONOS PCA Image
ENVI's Statisitics Output - 4 Dimensions
Percent Var. Cumm.
1331;! M15} Mg; M St_dey Eigenvalues Explained Variance

1 276 446 312.795 24.319 16756.310 84.636 84.636

2 242 1631 306.722 39.860 2892.030 14.608 99.244

3 135 1244 198.442 39.426 130.980 0.662 99.905
4 116 1357 300.744 126.742 18.770 0.095 100.000

Total 19798.09 100.000

 

183

 

Appendix B

184

sag1m_ISODATA_48classes.cls

Confusion Matrix:

Horseshoe Bay - l-meter imagery

Original
ISODATA

Ground Truth Class Code

 

O‘IOOOOOOOOOOOOOOONg

QIOOOOOOOOOOOOOQ

17
0
0

QIOOOOO

41

MIOO

24

NIOOOOOOQOOOOOOOOOO

Sign
6
9
23
27
28
15
46
20
22
8
17
32
34
5
11
3
10

clHHMMMGlfifiQBBBBQQ¢¢

185

10
10
10
12
13
13
14
16

47

36

14

45

30

17

 

NNNN
8.0m
omNH
N565
8.3a
cod
no.3
mfiwN
nodm
cod
ommm
2.3
00.05
cod
50.2
00.0
cod
cod
Rdm
CONN
8.2
8.8
8.0
mNAm

3:35“:

muohm

commmmfiou

NN.NN
8st
meme
code
cod
8.8a
RAN
wmdm
cod
codes
wmfih
modN
cod
8.8a
3.5m
N. N
cod
3.3
8.0m
09$
ocdw
Nmma
coda:
N¢.mm
...7—3.8%.
EOE—W
commmmEEO

RE
636
8%
8.8
8.9..
52
3.8
8.6...
$8
52
Sam
was»
8.8
52
SR
852
8.85
8.84
8.8
38
max.»
5.8
«<2
50%
39395.":
5233‘
8.585855

wufin
8.05
5.8
40mm
8.8a
cod
mown
mode.
code.
006
NNNN
nods
8.8a
ood
£sz
afiwm
8.8a
afimw
oodm
8.0m
oon
wv. 5
cod
wméo

3539:

55.5934
mama—65$

_mZuuuwaEn—uafimb vN
938455330 MN
«inseam—0.68559 N
356553 a
2832?: ON
3853? my
Hanna: wa
5239: S
«EnuESwmam mu
355 mm
£8.83 3
5,5513% ma
62885 2
3.85:3, 3
«Emma: 3
page: a
85321.: a
98228: K
$585.53 6
356.8:ng m
9:39.353 v
3:36-138 m
m§_?muzfiw82m N
«.55 n
newsman—ED
8.5155

«Now. scomucwoou «mauv—
.\.~k.$ "58563. :55

 

 

33 mm

33—

OOOOOOOOOOOOOOOOOOOVOOOOHO

RIOOOOOOOOOOOOOOOOOOOOOOODOO

”IOOOOOOOOOOOOOOOOOOOOOOOO®O

RIOOOOOOOOOOOOOOOOOOOOOOOOOO

BIOOOOOOOOOOOOOOOOOOOOOOOOOO

EIOOOOOOOOOOOOOOOOOOOOOOOOOO

BIOO‘OOOOOOOOOOOOOOOOOOOOOOOO

186

sag1m_ISODATA_48classes.cls
Horseshoe Bay - 1-meter imagery

conflnued

13

16

17
18
18
18
18
18
20

12
13
14
18
48

32
0
0
0
0
0
0
0
0
1

8313195118821:

47

{3.8181818151183133

187

32

50 26 27 40 33 42 36 47 31 27 27 30

50

4850 54

Column Total

 

 

 

 

 

 

188

E
03... memoowommooomgogwgam
§OMO$VHHN¢NH 4N COP-4 HQ‘Hr-tm
oflh
Q
~
gfingooonomeorxt-«txtxoot0138001010
3.25 mmmmwmvmvmmNmt-ooo HSSNQ‘
U
_
311: t\ oo
o HAN—4:004:10 cont-r [\N In
é’éwohvaNch-qcoommovmSO§2§N¢
vatmxouoomochovmsorxaocxct-t Q
'- HHHHHHHHHHNNfimN
'U
o
“E
.22.. 5‘
"I.
24":
G5
00“:
0:42:11
so as N [\fibMI-i IQ
fi~8n¢w§§agfioaaaafis
ooooooooooomoooootxm
H H!”
oooooooov-rooomgooom
OOOOOOOOHwt-‘QOHOOOM
Hm L0 H
H
oooooomwoooooooooe
tx-zr 1-1
H
oooooogoooooooooog
H H
ooooooooxoooooooooog
NH
oxaxmhvtxoooooooooooot
Hl—‘H co

 

70_24classes.cls

Horseshoe Bay - l-meter imagery

sag1m_MAXLIKE

 

Ground Truth Class Code

Confusion Matrix

I\IO

\DIO

IOIO

VIC

«no

NIO

HIN

Ima e Class
0 Unclassified

1 chara

32

16
30

13
27

10
10
27

23
31

47
47

36
42 36

27

 

32
33

40
40

27

12
26

50
50

16
33
50

54

45

48 50

42
0

2 phragmites-clump

3 scival-dense

6 sagrigida60/ 40

7 beachsand

8 shallowh20
Column Total

9 hetdub
10 sagrigida_wes

11 vallisneria

4 scival—clump
12 deeph20

5 typha-soi12
24 typha-phrag-scival

13 scivalbrown

14 leersia

16 saggram/ chara

17 typhah20

18 nuphar

19 eleocharis

20 impatians

21 bluejoint

22 cutgrass-eleocharis
23 eleocharis-sand

15 muck

 

NOOOOOOS

w

190

 

8:: 8 8 8:: 8: 8:
8.8 mm.“ 8.8 8.8 828828-283 8 8 8 o o o o
88 8.8 8.8 8.8 ufiméafioma 8 o: o 8: N o o
«8: 88 8.8 8.8 828288898 an «N: o 8 8:: o o
8.8 8.8 8.8: 8.8: «598:3 :N 8: o o o 8: o
8.8 8.8 8.8: 8.8: 28885: 8 8: o o o o 8:
8:: 8.8 8.8 8.8: 2868:... m: 8 o o o o o
8.2: 8.8 8.8 8.8: 8&2 m: 8 o o o o o
8.8 8.~: 2.8 8.8 852g 2.: 8 o o o o o
8.8 8.8 8.8: 8.8 8285883 o: m o o o o o
88 8.8 8.8 8.8 22.6 m: u o o o o o
8.8 8.8 8.8: 8.8 228: v: o: c o o o o
8.8 :28 8.8: 8:28 5:88:88 8: 8 o o o o o
8.8 8.8 8.8: 8.8: 828% N: 8 o o o o o
8.8 8.8 8.8: 8.8: «58%; :: 8 o o o o o
8.8: :88 SE 8.8 «22:8... 8: 8 o o o o o
8.8 88 8.8: 8.8 832: m 8 o o o o o
8.~ 8.8 8.8 8.8: 86,688 m :8 o o o o o
::.:: 8.8 8.8 8.8: 8882.2 2 8 o N o o o
8.8 8.8 8.8: 8:8 8882888 e N: o o o o o
8.8 8.8 8.8 8.8: 88-283 m 8 : o o o o
8.8: 8.8 8.8 8.8 8530-823. :2 8 o o o o o
8.8 $8: 88 8.8 88388 m :8 o o o o o
8.8 8.8 8.8: 8.8 3298:8858 N 8 o o o o o
8.8 8.~: 8.8 8.8 Sufi : 8 o o o o o
7:80.:ng :CoUth: famoumn: Aucmuhwn: fimmfimmm—UGD O m D O H O O
8.82m muohm Hmm§uu< 5333‘ 830 amass :33. m m w law. a
GOmmmmEEOU GOmmmmEO quEn—mCOU muuufimvognm >90.“

88. 38:08.88 8%:
8:28 ".0283 :85

H

g'OOOOOOOOOOOOOOOOOO

SAM_25_24classes.cls

Horses-hoe Bay - l-meter imagery

saglm
Ground Truth Class Code

 

X

Confusion Ma

wlo

BIO

\DIO

[DID

Q‘IO

MIC

NIO

HID

Ima e Class
0 Unclassified

1 chara

10
12
32

28
30

27

27

31
31

47
47

36
36

32
42

33
33

19
15
40

16
27

15
26

38
50

26
14
10
50

14
32

54

30
48 50

27
14

2 phragmites-clump

3 scival—dense

6 sagrigida60/4O
Column Total

4 scival-clump
7 beachsand

5 typha-soilZ
8 shallowh20

9 hetdub
10 sagrigida_wes

11 vallisneria

12 deeph20
24 typha-phrag-scival

13 scivalbrown

14 leersia

16 saggram/chara

17 typhah20

18 nuphar

19 eleocharis

20 impatians

21 bluejoint

22 cutgrass-eleocharis
23 eleocharis-sand

15 muck

 

OOOOHg

N
V

OOOOHON
Q

l\
[\

192

 

8:: 8 8 2: 8:: 8:
2.8 8.8 8.8 8.8 :m>8-8£8-~€b 8 8 m: o o 2 o
8 8 8.8 8 8835688 8 :m c 8 o o o
8.8 8.0 8.8 2.8 284823883 2 2: o o 2: o o
8.2 2.8 8.8 8.8 .5235 :N 8 m o o 8 o

o o 8: 8: 2888:: 8 8: o o o o 8:
8:: 8.8 8.8 8.8 82983 2 8 8 o o o o
8.8 :8 8.8 8.8 5:25 2 8 o o o o o
8.8 8.8 K8 8.8 8:29: 2 8 o o o o o
8 8.8 2.8 8.8 2288888 8: 8 o o o o o
8.8 8.8 8 8.8 8:5 m: 8 o o o o o
8.2 8.8 8 8.8 8.82 2 8 N o o o o
8.8 o 8.8 8: $8.828 2 8 o : : o o
o o 8: 8: 88$:V 2 8 o o o o o
o c 8: 8: 28:83 :: 8 o o o o o
2.8 :28 2.8 2.8 88:88 o: 8 m o o w o
8.8 o 8.8 8: 868:: 8 8 o o o o o
8.8 m8 8 8.8 823088 8 8 o o o o o
8.2 8.8 2.8 8.88 88.888 8 :N o o o o o
8.8 :28 8.8 8.8 888828.“, 8 8 o o o o o
8 8 8.8 8 88-288 m 8 2 o c N o
w 8 8.8 8 8539228 w 8 o o o o o
8.2 8.8 8.8 8.8 888.228. m 8 o o o o o
8 8 8.8 8 8829828828 N 8 o o o w o
8.8 8.8 8.8 8.8 220 : 8 o o o o o
:CmUuma: :CmUumn: aufimUuwn: #75958: Ummfimmw—UGD O O O O O O O

8885 8.82m Nmm38< 595932 man—U owns: ~28. m m MM .HM and.
flammmmfi—EOU GOmmmmEo mquSmCOU 9505665 30.”—

88. 358880 283:
2.8.8 ”8283. :85

EIOOOOOOOOOOOOOOHOOOO
$IOOOOOOOOOOOOOOOHOM

PCA654321_MAXLIKE 7O 24classes.cls

Horseshoe Bay - l-meter imagery

saglm
Ground Truth Class Code

 

O
O

Confusion Matrix

QIO

wIH

[\lo

\OIO

IOIO

Q'IH

MID

NIO

HIN

Ima e Class
0 Unclassified

1 chara

28
32

27
30

14
27

13
27

27
31

47
47

26

36
11
42 36

33
33

36
40

27
27

23
26

50
50

49
50

45

48

0

485054

38

2 phragmites-clump

3 scival-dense

6 sagrigida60/ 40

7 beachsand

8 shallowh20
Column Total

9 hetdub
10 sagrigida_wes

4 scival-clump
11 vallisneria
12 deeph20

5 typha-soil2
24 typha-phrag-scival

13 scivalbrown

14 leersia

16 saggram/chara

17 typhah20

18 nuphar

19 eleocharis

20 impatians

21 bluejoint

22 cutgrass-eleocharis
23 eleocharis-sand

15 muck

OO
NH
moN

Nwd
quN
muOH
mm.mH
Nm.wH

Mn

N
OOLQOOO
ox

Nm.wH

Cwoo

N
mmd
aclmuhlmnd
g
:cmmmmEEoU

34‘
NH
Hm.m
Nu. H

mod
vmw
m.NH

OH
QOO
mem
ONH

O

O
H.wm

O

OH

O
«m. HH

O

N
no.3

v
mwdm

g
a
:ommmmEO

mv. HO
mm
ANNO
OOH
OOH
Nodm
ands
mmdw
HNO
no.3
OONO
OOH
OOH
OOH
ucdw
OOH
OOH
wmdqw
OOH
3mg
mvNO
OOH
OONO
OONO

3589:
5883‘

9883300

 

omfim
mm
OOQO
mafia
OOH
HQOO
3.3
mxw
OO
5m
m HQV
wa
OOH
OOH
QHO
OOH
OO
OOH
3.me
OOH
mo
mmxmw
cm
5.3
g
Nulausuluaw
$333.5

~m>€$ma£9a5mb v"
vagina—302m mu
mtafioflmémmhmaa «N
gowns a
mama—mas: ON
$85080 mH

Sans: OH

853mb .:
Smaflafiwwg OH

x36 mH

2.52 3

550.5338 2

8.58% 2

£6853 2

«Banana 8
@326; m
8:30:55, w
@5383 n.
oioemawcma o
$8-29: m
mfiguAmZum H.
mmcvaaZum m
mfiguémfifiwmfim N
Esau H
Ummmmmfiucb O

wflwlwmalsa

Humw. 583E300 ammax

$2.3 528$. 3:26

mu

:13

OH
UH

§

”833%

OH
mm

33833885;

mm
mm
mw
OH.
HQ
ON

HH

H
H

 

HEPH.
3cm

In
‘3‘

V
N

INOOOOOOOOOOOOOOOCcoooccog

H O
m O
53 O

filNOOOOOOOOOOOOOOOOOOOOO
alt-‘OHOOOOOOOOOOOOOOOOOO

O
O
O

“VHH O

OOH

alOOOOOOOOOOOOOOOOOOOO

MHH OHH OOH 3

OOOOO

O
Q‘

EIMOOOOOOOOOOOOOHOOOO

OOOOOON
m

In
[\

glNOOOOOOOOOOOOOOl-DOO

194

36
47
28
47 31 27 27 30

42

27

 

%
a"; wloO000000800000006000000008
U)
‘D
a:
3:» NgOOOOOOVOOOOOOOOOOOOOOOOOR
u
N
no
'13
3'5 OIOOOOromoooooooooooooooooog
o~ P
2!:
cu
<21
(DIE (pl-DIOOOoogoooooocooooooomooomg
'6
HI
H O
«‘8'. u
g» gwlooooHOOOOcO‘Ov-SQOOOOOOOOOOOOS
1-
<3 .2
Um U
9‘2 _ngOOOQOOOOOOhOOOOOOOOOOOOOOE
I“, “4
cu
fig 5
we _UNIOOROOOOOOOOoooooooooovooos
8m x:
=3fHIOOOOOOOOOOOOOOOOtbOwOOOOOO
e w ..
U
5 cu - '23
1: E a 1:3
'5 ,2. a m H '6 8';
2 (D U 9.. \ Q) (B 05604-0
“5'6 U533) NO 0 3 ": 0)")th
I: «.2 mas—ova ltd < co 'TJ'HE—
° "t: fimfi'afica 6‘: o O ~E H‘A'E—C‘.
.5 U“! a'u'demthngH Emumgfimmm‘:
=’ “8mung-7;,«agoiéoagoacflsxefigguggefiafi
E EEEE>>fi&3%3&58‘§3§$fim8m§38fi%
'~-~ m now" =_. _. _‘
u HDfim883‘33558>n8§8$3cm§nambU
OHNMfi'mwbwdflOv-‘NMVIDVDBQQOHNMQ‘
HHHHHHHHHHNNNNN

195

 

 

5'11!‘ '

L;

 

mm.mm

OH.O
O0.0
ON
mmém
ONOH

OO

mO.Hm

OOO
OmOm
NOOH
OONO

O
O
OvOm
Ow
OH
mO. H
O
ONOO

328.8%

much”.—
commmmEEoU

NN.NO
ON
OH.OH
Ovd
O
VOOH
OOO
mN. HO
BOO
OOH
OO.NO
OOOV
O
O
OO.NO
OH.OH
O
OH.OO
OOOO
ON
Om
HOQO
«N
Omdm

3:8th .

muobm
commmmEO

OH.OO
OOH
NNOO
ONOO
OO
ONHO
HOON
OOH
OOON
O
NO. Hv
OOH
HONN
OH.OO
OO
HN.Ov
ONOO
OOH
NONO
ONO
m.NH
mo
OOOO
ON. Hm

358mm.
Mme—83‘

muufizmaou

ONNO
Nu
vOOO
NOOO
OOH
OOOO
H.Om
ONOH
OOOO

«ONO
HO.Hm
OOH
OOH
OHN
NO.HO
OOH
HO.vH
NOéO
OO
N
OH.OO
ON
NVOO

3:833
NmmSuu<

2335.5
OOOO. scumumwwuou 253—
..»58 ".0382 :55

fi>€$w~£n$amb ON
Ogmgfioflm ON
flannuomfimémfiwusu N
.5333 R
mguamg ON
$858? OH

.8595 OH

omfifib 2
Sanfififlwwam OH

x936 OH
«6.83 «H
550.5338 OH

Raga 2

mtmcflda> HH

«Emu»: 2
£63: O
8:30:97, O
US$583 N.
8333ch e
NOOméamb m
@5314ng v
mmcmuéctum O
QEEuédumEOmEm N
«.85 H
Oflmmmflucb O

3

ON _. —.
NO
O —.
VO —.
N r —.
ON r
NO
5
O
NO
O
VN
O r
OO
NO
O _.
Om
Nv

F?
In
‘- N

r

‘—

v
Ow
VO
O
ON
NV
OO
an H.

38. mm
33—

OOOOONOOOOOOOONOOOOOMFOON
NCOOOOOOOOOOOOOOOOOOOOOQO

 

c(:{'|‘-OOCDOCDCDCDCDOCDOCDO(')CDC>CDOC\ICDC
r-

t")
N

O: O:

O
O

[s
O)

O
O
O
OOH

SIOOVOOOOOOOOOOONOOOOOO

OO_.

00

O

OO_.

SIOOOOOOOOOOOOOOOOOOOO

OOOOOON
m

(D
V

QIWNOOOOOCOFOMOOOOOO
Fir-‘-

196

70_20classes.cls

tiss MAXLIKE

pren

Confusion Matrix

Prentiss Bay I-meter

Ground Truth Class Code

NH

blN

ENC)

IDIH

Q'Iv-c

MIC

NIO

HID

Ima e Class
O Unclassified
1 stricta

20

21

3322

2 stricta/ lasio

3 lasio/ acutus

4 lasio/typha50
5 roadside-veg

6 asphault

25

17
10

20

7 80typha-west

8 acutus/ typha50
9 acutus/ soil60

10 dolomite
11 chara

24
43 41

0
0
0
O
O
0
0
O
O
22
35

0
0
O
0
1
0
10
0
0
9
33

32

32

31 38

29

0 O
0 0
0 O
0 0
0 O
0 0
0 0
0 O
0 0
0 0
3525 25 25

12 soil70/typha15/acutu515
Column Total

13 nuphar-12in

14 potomogeton-lSin
15 muck-20cm

16 acutus-sparse

17 vallisneria

20 acutusclumps-dense

18 pine
19 shadow

197

 

ONH

OH.NOH

NOON

OOOOO

N.O
ON.OO
ONOOH
O
O0.0
O
O
O
OO.mON
358.8%

muobm
commmmEEoU

O
NO
O
O
OO
OOH
OOH
OOH
OOH
OOH
OOH
NOOO
ONOO
OOOO
O
OOH
OOH
OOH
OOH
ON
3
walk—m
:ommmmEO

 

O0.0¢
OOH

NWOO
OOH

OOOOOO

ON.NN
NOON
OOOO
OONO

COCO

OONO
3:855
daunuu<

9885980

OO.NO
g
3%
2333.5

mmcmuémfifiomfiaa ON
3092? OH
8a 2
$8559, NH
omaamméfiaom OH
EoON-v_o=E OH
EmHéBmOo—bouom OH
52-3%: 2
GBBszfibsbsm S
«35 HH
BEE—0O OH
8:8?338 m
omfimbhaaa m
«83-55mNOOO N
2:238 o
Om>¢Eman O
82533 w
955% \ommfl O
2mm— \fioEm N
SUE.“ H

8%38: o
Own—U Owne—

OOOO. scumumwwoou mama!
o\°NO.$ “Nos—38¢ 3985

OOO ON
OHH OO

O

O

O
OOH
NO

1qu
25m

IOOOOOOOOOOOOOOOOOOOO

 

O

N

N
V

IOOOOOOOOOOOOOOOOONOR

O\
H

('0
V

198

tiss_SAM 25 20classes.cls

Prentiss Bay 1-meter

pren
Ground Truth Class Code

 

Confusion Matrix

@IO

wlo

[\IO

\OIO

lOlO

VIC

MIC

NIO

HIO

Ima e Class
0 Unclassified

1 stricta

47
47

11

14
28

22

17
21

21

10
22

21
23

12
16
28

27
31

25
24

19
21

17
22

0
23

28
0

8 acutus/ typha50
Column Total

9 acutus / soi160

10 dolomite

4 lasio/typha50
11 chara

3 lasio/ acutus

5 roadside-veg
6 asphault

7 80typha-west

2 stricta/ lasio
12 soil7O/typha15/acutu515

13 nuphar-12in

14 potomogeton-lSin
15 muck-20cm

16 acutus-sparse

17 vallisneria

20 acutusclumps-dense

18 pine
19 shadow

O
O
O
ONON
O
ONOO
OH.OH
vOOHH
OOOO
OH.OHH
NN.NN
NO
ON
wo. Hv
Om
NHO
OOOO
OO
OvOO
OOO
g
wmolbm
:ommmmEEoU

3
O
O
O

OO
ON
ONNN
O
HN.OO
OOJu
OOOO
NO
OO.N¢
ONH
O
NH.«‘
NO .O
ONNN
NOOO
OO.NH

3589:

38.5
5.3280

OOH
OOH
OOH
OO.NN
OOH
OOOO
OO
H0.0¢
O .NH
OOJEV
O .NO
O. HO
NOOO
O .NO
OOON
OOOO
OOON
HN.OO
ON.OO
OOOO

3:853
HON—33‘

mumfiamcou

OO
OOH
OOH
OOH

3

ON

NN.NN
OOH
ONO

OOOO

Ova
OHO

vH.NO
H.NO
OOH

OOOO

OvOO

NN.NN

OH.OO

OONO

3589:
5832‘

83.695
OOmO. sawmomwoou «gnaw
c\OONHN ”Nu—2393‘ =98>O

88v$m63833m ON

338% OH

2a 3

$8559, NH

mmuamméfiam OH

EoON-v_u=E OH

EOHéBmOoEBom OH

again: 2

aaafimzfibaaa 2

«.85 HH

otEBov OH
8:8)338 a
ommfibkaaa w
umm3-m;m3OO N
3:938 O
O¢>¢Emume O
omafibEas w
.3535 \Emfl O
23:85.“, a
«or: H
OmfimmflocD O

8.48%

OOO

{33328

I"!
H

OH

‘38‘33833393‘33‘33

 

3'5
0
cat-9

OO

N N

(’0
N

200

 

Appendix C

201

 

 

 

 

.........

0.000984
0.001 71 1
0.002017
0.002704
0.003322
0.003469
0.003983
0.004392
0.011892
0.012594

2nd Der.
HB-S

0.000616
0.000718
0.001524
0.00165
0.00195
0.00323
0.004191
0.004466
0.009803
0.009893

.........

2nd Der.
HB—2

-0.01237
—0.00961
-0.00684
-0.00627
-0.00352
-0.00327
-0.00309
—0.00176
-0.00152
-0.00141
0.001482
0.001552
0.001746
0.002255
0.003417
0.003503
0.003624
0.004916
0.010217
0.011146

2nd Der.
HB-6

0.00121
0.001392
0.002025
0.002115
0.002611
0.002815
0.005245

0.00547
0.007181

2nd Der.
HB-3

—0.01592
-0.00704
-0.00683
-0.00635
-0.00539
—0.00279
-0.00249
-0.00226
-0.00112
-0.00111
0.001421
0.001691
0.001796
0.002786
0.0035

0.003837
0.00539

0.007499
0.00762

0.009756

0.001293
0.001795
0.00185
0.002251
0.006691
0.008287
0.00933
0.009607
0.010469

.........

ooooooooo

0.001103
0.001213
0.001464
0.002303
0.002868
0.004007
0.004658
0.005847
0.007218

0.001116
0.001573
0.002298
0.002311
0.003572
0.004878

Band
HB—8

 

202

 

 

 

0.000737
0.001157
0.001257
0.001845
0.002479
0.00292
0.00355
0.005263

2nd Der.
HB-13

-0.01656
-0.01242
-0.01149
-0.00642
-0.00583
-0.00509
-0.00342
~0.00302
-0.00275
-0.00254
0.001516
0.002204
0.002602
0.003332
0.00396
0.005135
0.007379
0.013359
0.013812
0.015648

ooooooooo

0.001264
0.001272
0.001628
0.002376
0.002387
0.00354
0.00498

2nd Der.
FIB-14

-0.03446
-0.02378
-0.01605
~0.01348
-0.01178
-0.01134
-0.0066
-0.00554
-0.00473
-0.00441
0.002161
0.003321
0.004017
0.006681
0.006683
0.013421
0.016018
0.019267
0.023247
0.031895

Band
HB-lO

ooooooooo

0.000763
0.001424
0.001547
0.00266
0.002696
0.004121
0.0(5234

2nd Der.
HB—15

-0.01124

0.001752
0.002136
0.00503
0.006648
0.006743
0.008306
0.009715

2nd Der.
HB-12

-0.02376
-0.01751
-0.01698
-0.00918
-0.00828
-0.00787
-0.00589
-0.0055
-0.00499
-0.00449
0.002693
0.003034
0.00449
0.005521
0.007165
0.008239
0.012789
0.017855
0.020204
0.02338

2nd Der.
HB-16

Band
HB—12

 

203

 

 

 

.........

0.000841
0.001392
0.001 691
0.002974
0.003183
0.008415
0.010558
0.01115
0.011345
0.015149

.........

Band

HB-21

743
731.5
928.3

ooooooooo

0.000837
0.001269
0.00127
0.001574
0.001679
0.00169
0.004026
0.004586
0.00922

0.000987
0.001217
0.003315
0.005278
0.006047
0.006839
0.008864

2nd Der.
M

-0.01662
-0.00836
-0.00665
-0.00648
-0.00612
-0.00289
-0.0028
-0.00274
-0.00263
-0.00253
0.001255
0.001929
0.002269
0.003109
0.003906
0.004845
0.004861
0.004887
0.010683
0.013302

2nd Der.
HB-23

-0.00812
-0.00644
«0.00615
-0.00391
-0.00264
—0.00191
-0.00185
-0.00146
~0.00132
-0.00097

0.001332
0.001626
0.003903
0.004632
0.004715
0.005597
0.006245

Band

0.001829
0.002719
0.00279
0.004666
0.005003
0.005829
0.007683
0.011161

2nd Der.
HB-24

-0.02004
-0.01916
-0.01879
-0.01439
-0.01143
-0.00467
-0.00457
—0.00359
-0.00335
~0.00293
0.001616
0.001873
0.002682
0.002802
0.003505
0.011277
0.014168
0.015448
0.015527
0.021887

Band
HB-24

 

204

 

 

 

 

2nd Der.
FIB-25

-0.02589
-0.02541
-0.01901
-0.01788
~0.00677
-0.00676
—0.00583
-0.00524
-0.00485
-0.00466
0.002898
0.003215
0.003919
0.004794
0.005928
0.013005
0.015914
0.016332
0.022842
0.024662

2nd Der.
HB-29

-0.01676
-0.01645
-0.01559
-0.00866

ooooooooo

0.003236
0.005331
0.006171
0.007493
0.014028
0.018593
0.019981

.........

2nd Der.
HB—26

-0.01778
-0.0157
-0.01189
-0.00992
-0.00815
-0.00326
-0.00323
-0.00309
—0.00255
-0.0025
0.001491
0.001612
0.001852
0.003003
0.005329
0.007693
0.008425
0.014077
0.01478
0.01523

2nd Der.
HB-30

-0.02402
-0.01688
-0.01234
-0.01222
-0.00616
-0.00443
-0.00363
-0.00289
-0.00278
-0.00251
0.001172
0.001739
0.002469
0.002674
0.003375
0.003999
0.009034
0.014666
0.021063
0.025434

939.9
Band

928.3
731.5

2nd Der.
HB—27

-0.02294
-0.01893
—0.01403
-0.01007
-0.00976
-0.00904
-0.006
-0.00531
-0.00311
-0.00288

0.001743
0.002771
0.003602
0.004322
0.011402
0.014222
0.016401
0.018005
0.022899

2nd Der.
HB—31

-0.02642
-0.02225
-0.01613
-0.01405
-0.00457
-0.00439
-0.00401
-0.00386
-0.00371
-0.00339
0.002853
0.003516
0.004531
0.004884
0.005292
0.011189
0.015097
0.01603
0.016523
0.021392

2nd Der.
HB-28

—0.02612
-0.02129
-0.01741
-0.01186
-0.00706
-0.00677
-0.00631
-0.00408
-0.00358
-0.00349

0.003795
0.011126
0.018723
0.01982
0.019825
0.020074

2nd Der.
HB—32

-0.00551
-0.00179
-0.0015
-0.00138
-0.00134
-0.00116
~0.00108
-0.00097
-0.00093
-0.00091

.........

 

 

 

 

 

 

 

0.000913
0.001124
0.001131
0.(X)1696
0.001715
0.(X)1764
0.002394

2nd Der.
PB-S

-0.01101
-0.00845
-0.00627
-0.(X)582
-0.00435
-0.00258
-0.00216
0.00213
-0.00205
-0.00193

0.000736
0.000742
0.000852
0.001116
0.001458
0.001732
0.002852
0.006114
0.014051
0.016412

.........

Band
PB-S

731.5
720
743

928.3

916.7

754.6

708.5

571.4

560.1

639.7

594.1
582.8
492.4
662.6
939.9
503.6
514.9
674.1
697
685.5

0.000834
0.002127
0.00246
0.“)3008
0.003131
0.003192

2nd Der.
PB-6

-0.00348
-0.00227
-0.00214
-0.00182
-0.(X)174
-0.00163
-0.00155
-0.00104
-0.00101
-0.00099

0.000471
0.000674
0.000762
0.000914
0.001007
0.001303
0.001933
0.002451
0.004546
0.004815

Band
PB-6

731.5
743
720

916.7

754.6

812.3

766.1

560.1

905.1

928.3

662.6
503.6
708.5
514.9
835.5
777.7
939.9
674.1
697
685.5

-0.02573
-0.01645
-0.01378
-0.01013
-0.00975
-0.00806
—0.00761
~0.00533
-0.00475
~0.00454

0.002722
0.003557
0.003605
0.003983
0.006879
0.007286
0.008132
0.008389
0.023562
0.034532

Band
PB—3

731.5
743
720

560.1

928.3

754.6

548.8

916.7

537.5

571.4

594.1
526.2
503.6
582.8
708.5
674.1
939.9
514.9
685.5
697

2nd Der.
PB—4

-0.02366
-0.01512
-0.01283
-0.00956
-0.00952
-0.00759
-0.00745
-0.0066
-0.00543
-0.00483
0.002427
0.003275
0.003713
0.004264
0.006571
01116997
0.(X)7802
0.(X)9012
0.021881
0.032092

2nd Der.
PB-8

-0.01453
-0.01258
-0.00675
-0.00433
-0.m426
-0.00371
-0.00299
~0.00299
-0.00272
-0.00245

0.000923
0.001244
0.001362
0.001699
0.001866
0.004135
0.005266
0.006788
0.019428
0.019766

Band

Band
PB—8

731.5
720
743

928.3

708.5

916.7

605.5

560.1

571 .4

639.7

492.4
582.8
628.3
662.6
503.6
514.9
939.9
674.1
697
685.5

 

206

 

 

 

 

000000000

674.1
939.9

0.000999
0.001079
0.001153
0.001157
0.001169
0.001639
0.002209
0.002941
0.006552

2nd Der.
PB-15

-0.01056
—0.00986
-0.00698
-0.00633
-0.00482
-0.00352
—0.00322
-0.00248
-0.00233
-0.00227
0.000924
0.001579
0.001879
0.002291
0.002568
0.004262
0.004336
0.006475
0.013568
0.015389

.........

.........

 

207

 

 

 

 

000000000

0.002724
0.003712

2nd Der.
PB-18

-0.029
-0.02126
-0.01576
-0.01117
-0.00787
-0.00573
-0.00464

-0.0041
-0.00384
-0.00316

000000000

0.002719
0.003714
0.005127
0.006424
0.008042
0.011519
0.013314
0.026453
0.030069

2nd Der.
PB-22

-0.(X)322
-0.00212
-0.00164
-0.00137
-0.00135
-0.00133
-0.00126
-0.00121
-0.00116
-0.00095

0.00181
0.002113
0.004143

IIIIIIIII

0.002203
0.003275
0.004665
0.00509
0.006059
0.009847
0.010249

2nd Der.
PB—23

-0.00169
-0.00149
-0.(X)104
-0.00086

Band
PB-19

.........

2nd Der.
PB-20

0.001797
0.002065
0.003537
0.004442
0.004729
0.005506
0.006414
0.009158
0.015595
0.019587

Band
PB-20

 

208

 

 

 

2nd Der.
PB—25

-0.0115
—0.01001
-0.00624
-0.00462
-0.00302
-0.00295

-0.0029
-0.00262
-0.00238
-0.00196
0.001163
0.00151
0.001744
0.002087
0.003143
0.003577
0.00518
0.005574
0.011682
0.013084

2nd Der.
PB—29

-0.0045
-0.00431
-0.00387
-0.00373
-0.00301
-0.00109
-0.00106
-0.00097
-0.00086
-0.00084

nnnnnnnnn

2nd
PB-26

-0.02419
-0.01351
-0.0112
-0.01007
~0.00788
-0.0071
-0.00545
—0.00489
-0.0043
-0.00391

nnnnnnnnn

0.001274
0.001378
0.001688
0.002983
0.003318
0.004156
0.005838

Band
PB-26

0.004621
0.011139
0.012901
0.013341
0.014272
0.018547

2nd Der.
PB-31

-0.00155
-0.00149
-0.00136
-0.00119
-0.00112
-0.00081
-0.00057
-0.00055
-0.00051
-0.00046

Band
PB-27

aaaaaaaaa

0.001062
0.001473
0.001513
0.002312
0.003184
0.004151
0.007184

0.“)899
0.009623

2nd Der.
PB-32

-0.0062
-0.00443
-0.00378
-0.00266
-0.00239
-0.00152
—0.00104
-0.00101
-0.00081
-0.00067

000000000

ooooooooo

 

209

 

 

 

 

0.000932
0.002471
0.002492
0.002524
0.007284
0.008024
0.008773
0.009357
0.009962

ooooooooo

2nd Der.
PB—34

000000000

0.001464
0.001531
0.001611
0.002864
0.003789
0.005703
0.006793
0.009786
0.019109
0.022103

Band
PB-34

0.001483
0.001552
0.001862
0.003797
0.004352
0.004814
0.006346
0.011663
0.014359
0.018239

0.001324
0.001612
0.0025
0.002566
0.008287
0.010162
0.011265
0.011489
0.012479

2nd Der.
PB—40

-0.01733
-0.0134
-0.01285
-0.0088
-0.00549
-0.0049
-0.00471
-0.00411
-0.00304
-0.00261
0.00187
0.002917
0.003494
0.003957
0.00436
0.00512
0.006879
0.009133
0.011562
0.017627

Band
PB-36

743
731.5
916.7

.........

928.3
731.5

 

~ ...u.-.~.ma_'

 

 

 

210

 

 

 

2nd Der.
PB-41

Band

2nd Der.
PB-42

-0.01004
-0.00798
-0.00476
-0.00425
-0.00279
—0.00216
—0.00127
-0.00087
-0.00087
-0.00074
0.001081
0.001259
0.001438
0.001672
0.002457
0.002503
0.003647
0.003667
0.00516
0.006168

2nd Der.
PB-46

0.003551

Band

0.000799
0.001278
0.001398
0.001438
0.002327
0.002456
0.003096

.........

0.00102
0.001461
0.001827
0.002406
01112754
0.003352
0.004663
0.007192

2nd Der.
PB-48

0.001089
0.001298
0.001498
0.001511
0.001551
0.001667
0.002242
0.002619

 

211

 

 

 

 

ooooooooo

0.001011
0.001164
0.002066
0.003403
0.006333
0.008414
0.010017

Band

ooooooooo

0.001261
0.001364
0.001633
0.001731
0.002204
0.005938
0.00601
0.(X)657

2nd Der.
PB-54

-0.00923
-0.00867
-0.00769
-0.00383
-0.00336
-0.00289
-0.00201
-0.00195
-0.00145
-0.00131

0.001363
0.001378
0.003076
0.003721
0.008001
0.010721
0.012182

ooooooooo

.
.........

0.00163
0.001868
0.002553
0.003143
0.003559
0.009828
0.010366

0.000835
0.001279
0.001767
0.002597
0.003105
0.004554
0.008682
0.011172
0.014414

Band
PB—52

 

212

 

 

 

 

2nd Der.
PB-57

-0.01031

.........

0.001 169
0.001463
0.001669
0.003141
0.004317
0.009131
0.010768
0.013476

2nd Der.
PB-61

—0.00786
-0.00466
~0.00418
-0.00366
-0.00146
-0.00144
-0.00115
-0.00112
-0.00093

ooooooooo

.........

2nd Der.
PB-58

-0.01937
-0.01651
-0.01367
-0.01102
-0.00538
-0.00293
-0.00239
-0.00229
-0.00211
-0.00207
0.001147
0.002113
0.002245
0.002286
0.002576
0.00889
0.009614
0.013233
0.01414
0.015821

2nd Der.
PB-62

-0.0192
-0.01469
-0.01144

-0.0094
-0.00553
-0.00501
-0.00488
-0.00468
-0.00442
-0.00405
0.002069
0.002197
0.00281
0.004376
0.005835
0.006194
0.008782
0.011184
0.017548
0.017552

2nd Der.
PB-59

0.00593
-0.00513
-0.00346
-0.00334
-0.00143
-0.00132
.0.00113
-0.00112
-0.00107
-0.001

0.001304
0.001326
0.002685
0.004021
0.007195
0.007937

2nd Der.
PB-63

-0.00535
-0.00214
-0.00213
-0.002
-0.00157
-0.00151
-0.00145
-0.00141
-0.00113
-0.00102

ooooooooo

0.000273
0.000328
0.000382
0.000387
0.000492
0.000576
0.001385

 

 

 

Appendix D

214

 

m. trust's-3

 

SYSTAT Output
Correlation-based, Principal Components Analysis
CASl-48 Resampled Spectral Library
Merged Data File (88 signatures)

k Dimension Output

 

Band01
Band02
Band03
Band04
Band05
Band06
Band07
Band08
Band09
Band10
Bandll
lBand12
Band13
Lgand14

and15
Band16
Band17
Band18
Band19
Band20
Band21
Band22
Band23
Band24
Band25
Band26
Band27
Band28
Band29
Band30
Band31
Band32
Band33
Band34
Band35
Band36
Band37
Band38

 

1
32.773

Component Loadings

1
0.694
0.721
0.748
0.757
0.759
0.755
0.749
0.748
0.767
0.832
0.921
0.949
0.948
0.942
0.931
0.912
0.898
0.892
0.875
0.854
0.835
0.797
0.758
0.728
0.762
0.921
0.962
0.926
0.877
0.838
0.817
0.803
0.798
0.796
0.794
0.795
0.796

Latent Roots (Eigenvalues)

2
12.816

2
0.585
0.58
0.575
0.581
0.589
0.603
0.616
0.623
0.61
0.533
0.358
0.226
0.177
0.181
0.242
0.326
0.373
0.395
0.429
0.461
0.485
0.531
0.559
0.568
0.527
0.28
-0.064
-0.289
0453
-O.536
057
.059
-0597
-0.6
-0.603
-0.603
-0.602

.3.
1.362

0.39
0.363
0.326
0.291
0.265
0.242
0.219
0.196
0.162
0.111
0.042
-0.014
~0.05
-0.077
-0.11
-0.148
-0.173
-0.188
—0.214
-0.236
—0.247
-0.254
-0.26
-0.262
-0.265
-0.231
-0.113
-0.027
0.025
0.045

0.05
0.056
0.053

0.05
0.048
0.044

0.04

4
0.769

-0.029
-0.002
0.028
0.053
0.067
0.077
0.087
0.09
0.074
0.009
-0.11
-0.198
-0.241
-0.256
-0.24
—0.194
-0.146
-0.101
-0.043
0.014
0.06
0.13
0.21
0.272
0.257
0.057
-0.154
-0.187
-0.128
—0.075
-0.052
—0.034
-0.026
-0.022
-0.019
~0.013
0.008

0.123

-0.076
~0.059
-0.042
-0.027
-0.012
0.005
0.021
0.038
0.053
0.068
0.078
0.08
0.079
0.071
0.057
0.039
0.015
-0.01
-0.019
-0.021
-0.019
-0.001
0.006
0.006
-0.017
-0.115
-0.168
-0.14
—0.077
-0.027
-0.002
0.008
0.009
0.007
0.005
0.003
0.004

0.066

—0.118
-0.084
-0.035
0.005
0.029
0.043
0.053
0.062
0.07
0.067
0.033
-0.014
-0.041
~0.047
-0.038
—0.022
-0.011
—0.005
0.002
0.009
0.01
~0.001
-0.012
-0.028
-0.02
0.029
0.038
0.025
0.017
0.016
0.019
0.023
0.025
0.026
0.028
0.026
0.018

 

215

 

 

 

Band39
Band40
Band41
Band42
Band43
Band44
Band45
IBand46
Band47
Band48

0.799
0.799
0.798
0.799
0.803
0.806
0.808
0.814
0.824
0.827

1

32.773

1

68.277

 

SYSTAT Output
Covariance-based, Principal Components Analysis
CASI-48 Resampled Spectral Library
(Merged Data File (88 signatures)

 

 

-0.597
-0.596
-0.595
-0.592
-0.586
-0.579
—0.573
-0.56

-0.539
-0.526

2

12.816

2

26.699

6 Dimension Output
Latent Roots (Eigenvalues)

1 2

3382.219 72.063

Component Loadings

.1. Z
Band01 0.278 0.935
Band02 0.316 0.974
Band03 0.355 1.012
Band04 0.37 1.056
Band05 0.373 1.085
Band06 0.361 1.107
Band07 0.349 1.13
Band08 0.348 1.156
Band09 0.385 1.19
Band10 0.537 1.24
Bandll 0.905 1.35
Band12 1.289 1.501
Band13 1.518 1.636
Band14 1.569 1.727
Band15 1.41 1.767
Band16 1.174 1.759
Band17 1.06 1.757
Band18 1.019 1.768
Band19

Variance Explained by Components

Percent of Total Variance Explained

 

0.032 0.057 0.013 0.001
0.029 0.07 0.014 -0.002
0.026 0.083 0.015 -0.005
0.022 0.094 0.016 -0.008
0.017 0.103 0.016 -0.013
0.014 0.114 0.016 -0.02
0.01 0.128 0.018 -0.026
0.003 0.137 0.019 —0.033
0 0.15 0.02 -0.05
-0.005 0.168 0.022 -0.057
3. 4 .5. 9
1.362 0.769 0.123 0.066
.3. £1. .5. 9
2.837 1.601 0.255 0.137
3. 4 § 9
28.677 3.744 2.446 1.171
Q 4 § 9
—0.026 0.453 -0.328 0.107
0.007 0.455 -0.312 0.11
0.043 0.451 -0.283 0.118
0.076 0.449 -0.255 0.125
0.098 0.448 -0.228 0.122
0.118 0.446 -0.205 0.113
0.138 0.444 -0.181 0.103
0.151 0.441 -0.157 0.088
0.145 0.423 —0.122 0.068
0.083 0.367 -0.085 0.016
-0.062 0.26 -0.058 -0.097
-0.216 0.14 -0.036 -0.234
-0.315 0.05 -0.009 -0.328
-0.349 -0.01 0.026 -0.351
-0.287 -0.037 0.083 -0.304
-0.164 -0.042 0.143 -0.224
-0.066 -0.055 0.174 -0.153
0.014 -0.067 0.19 -0.09

 

216

 

 

 

Band20
Band21
Band22
Band23
Band24
Band25
Band26
Band27
Band28
Band29
Band30
Band31
Band32
Band33
Band34
Band35
Band36
Band37
Band38
Band39
Band40
Band41
Band42
Band43
Band44
Band45
Band46
Band47
Band48

 

0842
0789
0674
0579
0529
0648
1436
3223
5643
8584
10843
12008
12.553
12.855
12.945
12.987
13.019
13169
13484
13.727
1385
13.973
14.029
13.957
13.815
13686
£3227
1209
11.596

2

33822 72063

2

96.866 2.064

1.727
1.737
1.73
1.688
1.656
1.698
1.894
2.151
2.067
1.389
0.602
0.118
-0.244
-0.369
-0.412
-0.458
—0.45
—O.469
-0.47
—0.481
-0.498
-0.51
-0.482
—0.38
-0.292
-0.225
-0.026
0.246
0.367

3

28.677

3
0.821

0.224
0.309
0.437
0.57
0.678
0.669
0.316
-044
-1.109
-1.366
-1.322
-1.282
-1.194
-1.141
-1.1
-1.063
-0.96
-0.663
-0.223
0.081
0.298
0.505
0.701
0.892
1.097
1.333
1.511
1.677
1.906

Variance Explained by Components

4
3.744

Percent of Total Variance Explained

4
0.107

-0.06
-0.046

0.037
0.054
0.013
-0.234
-0587
-0.743
-0548
-0.19
0.043
0.22
0.231
0.195
0.162
0.116
0.106
0.141
0.132
0.113
0.095
0.062
-0.001
-0.046
-0.075
-0.164
-0.226
-0.282

0.268
0.292
0.318
0.334
0.336
0.335
0.243
-0.056
-0.37
-0.466
-0314
-0.104
0.026
0.112
0.161
0.194
0.213
0.19
0.143
0.143
0.128
0.126
0.121
0.081
0.005
-0.05
-0.143
-0.364
-0.492

1.171

0.034

-0.001
0.022
0.033
0.064
0.093
0.128
0.229
0.322
0.308
0.14
-0.087
-0.198
-0.173
-0.109
-0.037
0.038
0.08
0.106
0.071
0.065
0.078
0.101
0.106
0.098
0.076
0.032
—0.056
~0214
-0.289

 

217

 

 

APPENDIX E

218

 

Figure E-1. Prentiss Bay; False Color Composite.

219

 

Figure E-2. Horseshoe Bay; True Color Composite and ROIs.

220

 

Figure E-3. Horseshoe Bay; False Color Composite.

221

 

Figure E-4. Prentiss Bay, PC-image; 4, 3, 2 Color Composite.

 

Figure E-5. Horseshoe Bay PC-Image; 4, 3, 2 Color Composite.

223

 

l-meter Spatial Resolution

 

 

4—meter Spatial Resolution 8-meter Spatial Resolution

Figure E-6. Pixel Level Spatial Degradation Patterns.

24

BIBLIOGRAPHY

Anderson, I. R., E. E. Hardy, J. T. Roach, and R. E. Witmer. 1976. A land use and
land cover classification system for use with remote sensor data: US. Geological
Survey Professional Paper #964, p. 28.

ASPRS, 1995. Land satellite information in the next decade: Data notebook,
American Society of Photogrammetry and Remote Sensing (ASPRS), Bethesda,
Maryland, USA.

Avery, T. E., and G. L. Berlin. 1986. Fundementals of Remote Sensing and
Airphoto Interpretation. Macmillian Publishers, NY.

 

Best, R. G., M. E. Wehde, and R. L. Linder. 1981. Spectral reflectance of
hydrophytes. Remote Sensing of Environment, Vol. 11: Pp. 27-35.

Borstad Associates Ltd. 1995. Fish habitat mapping: Remote sensing,
verification of imagery and production of digital maps from Baynes Sound and
Comox Harbour. Unpublished report prepared for the Canadian Department of
Fisheries and Oceans. 33 pp.

Bowker, D. E., R. E. Davis, D. L. Myrick, K. Stacy, and W. T. Jones. 1985. Spectral
Reflectances of Natural targets for Use in Remote Sensing Studies. NASA Reference
Publication - 1139.

Brown, L. and G. Borstad. 1999. Mapping wetland breeding habitats in the Frazier
River delta. Backscatter - Association of Marine Remote Sensing Vol. 10(1): Pp.
8—11.

Burton, T. M., D. G. Uzarski, J. P. Gathman, J. A. Genet, B. E. Keas, C. A.
Stricker. 1999. Development of a preliminary invertebrate index of biotic in tegrity for
Lake Huron coastal wetlands. Wetlands Vol. 19(4): Pp. 869-882.

Butera, M. K. 1983. Remote Sensing of Wetlands. IEEE Trans. Geoscience Remote
Sensing. Vol. 21: Pp. 383-392.

Campbell, I. B. 1996. Introduction to remote sensing, 2nd edition. The Guilford
Press, NY.

Cardinale, B. 1., V. ]. Brady, and T. M. Burton. 1998. Changes in the Abundance

and Diversity of Coastal Wetland Fauna from the Open Water/Macrophyte edge
towards shore, Wetlands Ecology and Management Vol. 6(1). Pp. 59-68.

225

Carter, G. A. 1993. Responses of leaf spectral reflectance to plant stress. American
Journal of Botany Vol. 80(3): Pp. 239-243.

Carter, G. A. 1998. Reflectance bands and indices for remote estimation of
photosynthesis and stomatal conductance in pine canopies. Remote Sensing and
Environment. Vol. 63: Pp. 61-72.

Carter, V. P., and R. R. Anderson. 1982. Interpretation of wetlands imagery based
on spectral reflectance characteristics of selected plant species. Proceedings of the
American Society of Photograrnmetry 38th Annual Meeting, Pp. 580-595.

Chadde, SW. 1998. A Great Lakes Wetland FloraJ Pocketflora Press, Calumet,
Michigan.

 

Chow-Fraser, P., and D. A. Albert. 1998. Biodiversity Investment Areas: Coastal
Wetland Ecosystems. State of the Lake Ecosystem Conference-1998. United
States Environmental Protection Agency and Environment Canada.

Clark, R.N., G.A. Swayze, AJ. Gallagher, T.V.V. King, and W.M. Calvin, 1993,
The U. 5. Geological Survey, Digital Spectral Library: Version 1: 0.2 to 3.0 microns,
US. Geological Survey Open File Report 93-592, 1340 pages.

Comer, A. 1., Albert, D.A., Wells, H. A., Hart, B. L., Raab, S. 8, Price, D. L.,
Kashian, D.M., Corner, R. A., and Shuen, D. W. 1995. Michigan’s Native
Landscape, as Interpreted from general Land Ofi‘ice Surveys, 1816-1865. Michigan
Natural Features Inventory Report to Water Division, US. EPA and Wildlife
Division, Michigan Dept. of Nat. Res. Lansing, Michigan.

Congalton, R. G. 1991. A review of Assessing the Accuracy of Classifications of
Remotely Sensed Data. Remote Sensing and Environment. Vol. 37: Pp. 35-46.

Congalton, R. G., 1983. Assessing Landsat Classification Accuracy Using Discrete
Multivariate Analysis Statistical Techniques. Photogrammetric Engineering and
Remote Sensing, Vol. 49(12): Pp. 1671-1678.

Cowardin, L. M., V. Carter, F. C. Golet, and E. T. LaRoe. 1979. Classification of
wetlands and deep-water habitats of the United States. US. Department of the
Interior, Fish and Wildlife Service, Washington DC.

Crist, E. P., and R. C. Cicone, 1984. Application of Tasselled Cap Concept to

Simulated Thematic Mapper Data. Photgrammetric Engineering and Remote
Sensing, Vol. 50: Pp. 343-352.

226

 

Dahl, T. E. 2000. Status and trends of wetlands in the conterminous United States -
1986 to 1997. US. Department of the Interior, Fish and Wildlife Service,
Washington, DC.

Dahl, T.E., C. E. Johnson, and W. E. Prayer. 1991. Status and Trends of wetlands in
the Conterminous United States - Mid 1970’s to mid 1980’s, US. Department of the
Interior, Fish and Wildlife Service, Washington, DC.

Davis, J. C. 1986. Statistics and Data Analysis in Geology. John Wiley 8: Sons,
New York.

Demetriades-Shah, T. H., M. D. Steven, and J. A. Clark. 1990. High resolution
derivative spectra in remote sensing. Remote Sensing and Environment. Vol. 33:
Pp55-64.

Detenbeck, N. E., S. M. Galatowitsch, J. Atkinson, and H. Bell. 1999. Evaluation
perturbations and developing restoration strategies for inland wetlands in the Great
Lakes basin. Wetlands Vol. 19(4): Pp. 789-820.

Earnst-Dottavio, C. L., R. M. Hoffer, and R. P. Morczynski. 1981. Spectral
characteristics of wetland habitats. Photogrammetric Engineering and Remote
Sensing Vol. 47: Pp. 223-227.

Elvidge, C. D., and F. P. Portigal. 1990. Reflectance spectra of green and dry
vegetation derived from 1989 A VIRI S data. Proceedings, fifth Australian remote
sensing conference, Perth Australia: Pp 185-195.

Elvidge, C. D., and Z. Chen. 1995. Comparison of broad-band and narrow—band red
and near infrared vegetation indices. Remote Sensing and Environment. Vol. 54:
Pp. 38—48.

Environment Canada. 1994. Environmental Sensitivity Atlas for Lake Huron’s
Canadian Shoreline. Environment Canada.

ESRI, 2001. ARCINFO users manual, page 158. Redlands, California.Fell, A. F.,
and G. Smith. 1982. Higher derivative methods in ultraviolet-visible and infrared
spectroradiometry. Analytical Procedures. Vol. 54: Pp. 28-32.

Forsyth, F., G. Borstad, W. Horniak, and L. Brown. 1998. Prince Rupert intertidal
habitat inventory project. Unpublished report to Prince Rupert Port Corporation,
the Canadian Department of Fisheries and Oceans, and the City of Prince
Rupert.

227

Gathman, J., and B. Keas. 1999. Les Cheneaux coastal wetland project: A synthesis.
A report submitted to the Michigan Coastal Management Program, Project
Number 99-3099-24 by The Nature Conservancy.

Goetz, A. F. H. 1992. Principles of narrow band spectroscopy in the visible and IR:
instruments and analysis. In Imagingpectrometry: fundamentals and
prospective applications, edited by F. Toselli and J. Bodechtel (Dordrecht:
Kluwer), Pp 21-32.

 

Gottgens, J. F., Swartz, B. P., Kroll, R. W. and Eboch, M. 1998. Long term GIS-
based records of habitat changes in a Lake Erie coastal marsh. Wetlands Ecology and
Management Vol. 6(1): Pp. 5—17.

Gross, M. F., and V. Klemas. 1986. The use of airborne imaging spectrometer (AIS)
data to differentiate marsh vegetation. Remote Sensing of Environment Vol. 19: Pp.
97-103.

Gross, M. F., M. A. Hardisky, and V. Klemas. 1989. Applications to coastal
wetlands vegetation. In Theory and applications of optical remote sensing,
Ghassem Asrar, Editor. John Wiley and Sons, NY.

Hardisky, M. A., Gross, M.F., and Klemas, V. 1986. Remote Sensing of Coastal
Wetlands. Bioscience, Vol. 36(7): Pp.453-460.

Henderndorf, C. E. 1987. Lake Eerie coastal wetlands: an overview. Journal of Great
Lakes Research, Vol. 18(4) Pp. 533-551.

http: / / www.borstad.com/ papers / terranova.html. March 20, 2001. Intertidal
Habitat in Terra Nova National Park.

Hudson W. D., and C. W. Ram, 1987. Correct Formulation of the Kappa
Coeflicient of A greemen t. Photogrammetric Engineering and Remote Sensing,
Vol. 53(4): Pp. 421-422.

Huguenin, R. L0, and J. L. Jones. 1986, Intelligent information extraction from
reflectance spectra: absorption band position. Journal of Geophysical Research. Vol.
91 (B9): Pp. 9585-9598.

International Joint Commission. 1993. Levels Reference Study, Great Lakes St.

Lawrence River Basin, Annex 2. Landuse Management. International Joint
Commission.

228

 

 

 

ITRES Research Limited, 2000. Draft report - Airborne Imaging Systems for
Coastal Applications: An Update. For Joint Airborne Lidar Bathymetry Technical
Center of Expertise, Mobile, Alabama.

Jain, A. K., 1989. Fundementals of Digital Image Processing. Prentice Hall,
Englewood Cliffs, NJ.

Jaworski, E, and N. C. Raphael. 1978. Fish, Wildlife, and Recreational Values of
Michigan’s Coastal Wetlands. Michigan Department of Natural Resources:
Lansing, Michigan.

Jensen, J. R. 1996. Remote Sensing of the Environment : an Earth Resource
Perspective. Prentice Hall. Upper Saddle River, NJ.

 

Jensen, J. R. 1996. Digital Image Processing Prentice Hall. Upper Saddle River,
NJ.

Jensen, J. R., E. J. Christiansen, and R. Sharitz. 1984. Non-tidal wetland mapping
in South Carolina using airborne multispectral scanner data. Remote Sensing of
Environment Vol. 16: Pp. 1-12.

Jensen, J. R., M. E. Hodgson, E. Christensen, M. E. Jr. Halkard, L. R. Tinney, and
R. Sharitz. 1986. Remote Sensing Inland Wetlands: A Multispectral Approach.
Photogrammetric Engineering. Vol. 52(1): Pp. 87-100

Jupp, D. L. B. G. Byrne, J. Anstee, E. McDonald, T. McVicar and D. Parkin.
(1996). Port Phillip Bay sea grass mapping project G2.2. Final report series I, II, III and
IV. CSIRO Division of Water Resources, Consulting Reports 96/.

Kauth, R. J., and G. S. Thomas. 1976. The tasseled cap - A graphical description of
spectral-temporal development of agricultural crops as seen by Landsat. Proceedings,
6th International Symposium on machine Processing of Remotely Sensed Data,
Purdue University West Lafayette, IN. Pp. 41-51.

Keddy, P. A. and A. A. Reznicek. 1986. Great Lakes vegetation dynamics: the role of
fluctuating water levels and buried seed. Journal of Great Lakes Research Vol. 12
Pp.25-36.

Keough, J. R., T. A. Thompson, G. R. Guntenspergen, and D. A. Wilcox, 1999.

Hydrogeomorphic factors and ecosystem responses in coastal wetlands of the Great
Lakes. Wetlands Vol. 19(4): Pp. 821-834.

229

Koonce, J. F., C. K. Minns, and H. A. Morrison. 1999. Biodiversity investment
areas, aquatic ecosystems. State of the Lakes Ecosystem Conference (SOLEC,
1998).

Krieger, K. A., 1992. The Ecology of invertebrates in Great lakes Coastal Wetlands,
Journal of Great Lakes Research Vol. 18(4): Pp. 634-650.

Kruse, F. A., A. B. Lefkoff, J. W. Boardman, K. B. Heidebrecht, A. T. Shapiro, J.
P. Barloon, and A. F. H. Goetz. 1993. The Spectral Image Processing System (SIPS)
- Interactive visualization and analysis of imaging spectrometer data. Remote Sensing
of Environment, Vol. 44: Pp. 145-163.

Kumar, L., K. S. Schmidt, S. Dury and A. K. Skidmore. (in press 2001) Review of
Hyperspectral Remote Sensing and Vegetation Science. In van der Meer, F ., (2001)
(editor). Hyperspectral Remote Sensing (Kluwer Academic Press: Dordrecht)

Lillesand, T.M., and R.W. Kiefer. 1994. Remote Sensing and Image
Interpretation, 3rd Edition. John Wiley and Sons, NY.

 

Lillesand, T.M., and R.W. Kiefer. 2000. Remote Sensing and Image
Interpretation, 4th Edition. John Wiley and Sons, NY.

 

Maynard, L. and D. Wilcox. 1997. Coastal Wetlands. Background Paper, State of
the Lakes Ecosystem Conference (SOLEC, 1996).

Minc, L. D. 1997. Great Lakes Coastal Wetlands: An overview of controlling abiotic
factors, regional distribution, and species composition. A report submitted to the
Michigan Natural Features Inventory, December, 1997.

 

Mitsch, W. J., and J. G. Gosselink. 2000. Wetlands - 4th Edition, John Wiley 8:
Sons, Inc. New York, NY.

Oguma, H., Nakayama, M., N ishio, F., Yamagata, Y. 1996. Studies on metrologies
for monitoring the distribution of wetland vegetation using multispectral data in
Kushiro Mire, Japan, in Proceeding at the Second International Airborne Remote
Sensing Conference and Exhibition, San Francisco, California, Vol. II: Pp.
711-716.

Prince, H. P., P. I. Padding, and R. W. Knapton. 1992. Waterfowl use of the
Lauren tian Great Lakes. Journal of Great Lakes Research, Vol. 18(4): Pp 673699.

Proceedings, 2“d International Symposium on Automated Cartography, Ottawa.

V01. 2: Pp. 883-896.

230

 

 

 

Raphael, N. C. 1987. Prehistoric and historic wetland heritage of the upper great lakes.
Michigan Academician, Vol. 19(3): Pp. 331-365.

Reimold, R. J., J. L. Gallagher, and D. E. Thompson. 1973. Remote Sensing of
Coastal Marsh. Photogrammetric Engineering and Remote Sensing Vol. 39: Pp.
477-488.

Research Systems, Inc., 1999. ENVI User’s Manual, Version 3.2.

Ritter, R. and E. L. Lanzer. 1997. Remote sensing of nearshore vegetation in
Washington State’s Puget Sound. Proceedings of the 1997 Geospatial Conference,
Seattle, WA, Vol. 3: Pp. 527-536. E

Roe, H. B. and Q. C. Ayres. 1954. Engneering for Agricultural Drainage,
McGraw- Hill, New York.

Rummel R. J. 1970. Applied Factor Analysis. Northwestern University Press,
Evanston, IL.

 

Sabins, M. J., 1987. Convergence and consitency of Fuzzy c-means/ISODATA
algorithums. IEEE Transactions Pattern Analysis 8: Machine Intelligence, Vol. 9:
Pp. 661-668.

Savastano, K. J., K. H. Faller, and R. L. Iverson. 1984. Estimating vegetation
coverage in St. Joseph Bay, Florida with an airborne multispectral scanner.
Photogrammetric Engineering and Remote Sensing Vol. 50: Pp. 1159-1170.

Seher, S.J., and Tueller, PT, 1973. Color Aerial Photos for Marshland.
Photogrammetric Engineering. Vol. 39(2): Pp. 489-499.

Shafique, N. A., AC. Bradley, F. Fulk, and S. M. Corrnier, 2001. The Selection of
Narrow Wavebands for Optimizing Water Quality Monitoring on the Great Miami
River, Ohio using Hyperspectral Remote Sensor Data Journal of Spatial Hydrology
Vol. 1(1): Pp. 1-22, (http:/ / www.3patialhydrology.com).

Singh, a., and A. Harrison. 1985. Measuring the fractal dimensions of surfaces.
Proceedings, 2nd International Symposium on Automated Cartography, Ottawa.
Vol. 2: Pp. 883-896.

Smith, P. G. R., V. Glooschenko, V., and D. A. Hagen. 1991. Coastal wetlands of
the three Canadian Great Lakes: inventory, current conservation initiatives and
patterns of variation. Canadian Journal of Fisheries and Aquatic Sciences Vol. 48:
Pp. 1581-1594.

231

Stewart, W. R., V. Carter, and P. D. Brooks. 1980. Inland (non- tidal) wetland
mapping. Photogrammetric Engineering and Remote Sensing Vol. 46: Pp.
617—628.

Stoney, W. E. and J. R. Hughes, 1998. A new space race is one! GIS Wordl, Vol. 11:
Pp. 44-46.

Thenkabail, P.S., R. B. Smith, and E. DePauw. 2000. Hyperspectral Vegetation
Indices and Their Relationships with Agricultural Crop Characteristics. Remote
Sensing and Environment. Vol. 71: Pp. 158-182.

 

Tou, J. T. and R. C. Gonzalez, 1974. Pattern Recognition Principles. Addison-
Wesley Publishing Company, Reading, Massachusetts.

Tsai, F., and W. Philpot. 1998. Derivative Analysis of Hyperspectral data. Remote
Sensing and Environment. Vol. 66: Pp. 41-51.

Wang, J., L. Zhang, and Q. Tong. 1998. The derivative matching for wetland
vegetation identification and classification by hyperspectral data. In Proceedings of
SPIE, Hyperspectral Remote Sensing and Application. Volume: 3502: Pp. 280-

288.

Wetzel, R. G., 1992. Wetlands as Metabolic Gates, Journal of Great Lakes Research,
Vol.18(4) Pp. 529-532.

Wilcox, D. A. 1995. The role of wetlands as nearshore habitat in Lake Huron. In
Munawar, M., T., Edsall, and J. Leach (eds.) The Lake Huron Ecosystem:
ecology, fisheries and management. SPB Academic Publishing, Amsterdam,
The Netherlands. Pp. 223-245.

Zacharias, M, O. Niemann, and G. Borstad. 1992 An assessment and
classification of a multispectral bandset for the remote sensing of intertidal
seaweeds. Canadian Journal of Remote Sensing, Vol. 18(4) Pp. 263-274.

Zarco-Tejada, P. J. and J. R. Miller. 1999. Land cover mappiing at BOREAS using
red edge spectral parameters from CASI imagery. Journal of Geophysical Research.
Vol. 22: p. 27

Zelazny J. and J. S. Feierabend, Editors. 1988. Increasing our Wetland
Resources. National Wildlife Federation, Washington, DC.

232

 

 

 

IIIIIIIIIIIIIIIIIIIIIIIIIIIIII

wuyumwyyywill