w - if
at?“ ..

:- OJP . 2

, ‘ . c .
, .. 6.3%.
.0... 7»... J, V , tﬂLx
¢.<\ it... .4 a:
1.3:. taut}. a A .
22%.... E em a},

1

 

 

. . v ‘ a. .

 

inuﬂl 1.31.4‘114: 4”. V . , ‘ . ‘.§\.r|$wrll‘1.lﬂ 1593......1.‘

. niggiyh? ”iyiyg

200}

This is to certify that the
dissertation entitled

‘ HYPERSPECTRAL DATA MODELING FOR WATER
QUALITY STUDIES IN MICHIGAN'S INLAND LAKES

presented by

l

N arumon Wiangwang

has been accepted towards fulﬁllment
of the requirements for the

Ph.D. degree in Geography

W‘

 

 

=7 \ ngor Professor’s Signature
é/Eoﬂé

Date

MSU is an Afﬁrmative Action/Equal Opportunity Institution

 

LIBRARY
Michigan State
University

 

 

 

— —--.-.-o—---c-n-n-c-n-v--._ .—

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 I DATE DUE DATE DUE

 

JUBH 90M

 

[33: . wig

 

SEP 2 9 2010

 

093310

 

 

 

 

 

 

 

 

 

 

 

2/05 p:/ClRC/Date0ue.indd-p‘1

HYPERSPECTRAL DATA MODELING FOR WATER QUALITY STUDIES
IN MICHIGAN'S INLAND LAKES

By

Narumon Wiangwang

A DISSERTATION

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

DOCTOR OF PHILOSOPHY

Department of Geography

2006

ABSTRACT

HYPERSPECTRAL DATA MODELING FOR WATER QUALITY STUDIES
IN MICHIGAN’S INLAND LAKES

By

Narumon Wiangwang

Hyperspectral remote sensing imagery has been used to estimate spatial and
temporal variation of water quality, such as chlorophyll a, transparency, and
suspended solids, primarily for marine and coastal waters. Although
physicochemical properties of marine and inland waters diﬂer, hyperspectral
data and modeling may provide an alternative tool for inland lake assessment.
However, little has been done to identify the most suitable spectral bands for
water quality estimation and there is a lack of quantitative relationship between
water quality and hyperspectral data. The primary objectives of this study are
to identify optimal spectral bands most sensitive to water quality indicators and
to develop improved hyperspectral water quality indicators of inland lakes. The
secondary objective is to determine the most effective ﬁlters for noise removal in

hyperspectral data.

To address these objectives, a ﬁeld campaign was conducted on 42 inland lakes
in Michigan in 2004. Radiometric spectra, Secchi disk depth, dissolved oxygen,
temperature, and light extinction proﬁle data were collected. Water samples
were analyzed for chlorophyll a, suspended solid, total nitrogen, total
phosphorus, non-purgable organic carbon, and phytoplankton species

composition. Spectral radiances were measured with a hand-held spectrometer

(LabSpec® Pro) and with an airborne Imaging Spectrometer for Applications

(AISA) sensor, to correlate the water quality and hyperspectral data.

Principal Component Analysis was used to identify the narrow-wavebands, and
derivative analysis used to determine the region-wavebands. Statistical spectral
water quality indicators were developed to correlate with Secchi depth,
chlorophyll a, total suspended solid, non-purgable organic carbon, diatom
biomass, green algal biomass, and bluegreen algal biomass. These relations
were validated to suggest that high accuracies were achieved for Secchi depth
(R2 0.76 - 0.84), chlorophyll a (R2 0.70 — 0.7 6), and bluegreen algae (R2 0.56 —
0.72). The quantitative relationship between remotely sensed variables and
water quality indicators can be used to extrapolate point—based water quality
measurements to large spatial extents for an improved water quality
assessment. Additionally, the Savitsky Golay ﬁlter was found the best to
remove spectral noises. The innovation of this study is that it developed a
quantitative relationship between hyperspectral data and water quality

variables of inland lakes in Michigan.

Cepyright by
NARUMON WIANGWANG

2006

To my parents, whose love and faith has been unconditional.

ACKNOWLEDGMENTS

Getting to this point has been a long, eventful journey. Although I have a great
feeling of accomplishment upon ﬁnishing my dissertation, there is a stronger
feeling of gratitude for all the people who helped me reach this point and who
helped to make the experience such an intellectually fulﬁlling and enjoyable
one. Mentioning all the people who have helped along the way would ﬁll a
substantial volume. For this and for the miracle of life and all that it brings
with it, I am eternally indebted, ﬁrst of all, to my parents. Beyond that, there
are too many debts to adequately acknowledge, but a few deserve special

mention.

I am grateful that I have two wonderful parents who have raised me to be strong
and to have conﬁdence in myself to try anything and to be conﬁdent in their
unending love. I thank my husband for believing in me. His commitment,

emotional support, dedication, and love made this achievement possible.

While my parents are the reasons I was able to begin this research, Professor
Jiaguo Qi, my advisor, is the reason I was able to complete it. I ﬁrst met Dr. Qi
when I had just begun my Master’s program. Although he was not my major
advisor at the time, I had been working closely with him as he was one of the
research committee members. Since he became my major advisor at the
beginning of my Doctoral program, Dr. Qi has been a constant, buoyant source

of encouragement and inspiration. As a ﬁne academic example, he artfully

vi

combines rigorous standards with gentle support and caring. He always listens
generously to students and respects their ideas, never forcing his own agenda,
but instead working hard to clarify and help them ﬂourish. Dr. Qi has been my
teacher, my director, my employer, my discussant, and always my mentor. I
thank him for the long hours and thoughtful responses. He went beyond the
call of duty, and he did so with complete respect for the ways in which the
personal, political, and professional interact. I wish all doctoral students could
have directors who worked as hard and who made them feel as successful as

Dr. Jiaguo Qi has me.

I wish to express my sincerest appreciation to each member of my committee,
Dr. Joseph Messina, Dr. Robert Jan Stevenson, and Dr. David Lusch. They
have been a pleasure to work with and have always offered advice,
encouragement and support. I would like to express my deepest gratefulness to
Dr. Messina not only for his professional guidance on research but also his
understanding, encouragement and support throughout my time at Michigan
State University. Several times, I wonder how someone, let alone a university
professor, could be incredibly patient with my writing. I am amazed with how
Dr. Messina spent time and effort on revising my thesis, dissertation, and
publication writing through all these years. He too is someone I will always
look up to. I am obliged to Dr. Stevenson for his expertise and his kindness.
He alone brought the entire ecological perspective to the research. I am
thankful for Dr. Lusch’s comments on the research. His detailed insight has
completed this work. Together, these four members of my guidance committee

have tried their best to form pieces of this work into a complete research effort.

vii

Committee members who care about their students and understand their
development as a scholar are invaluable. Superb mentors, they have remained
constant and patient throughout a very long process. The flaws that remain in

this work are mine alone.

My friends and otﬁcemates at Michigan State University made the cold winters
enjoyable and continue to push my thinking further. Thanks especially to
Cuizhen Wang, Yasuyo Makido, Jianjun Ge, Nate Torbick, Meleia Egger,
Meghan Burns, and Eraldo Matricardi and many others for their continuing
understanding and support. They have made my six years away from home
and family so pleasurable. With them, Michigan State University is a ‘home’
away from home. A special thanks goes out to my dearest friends Meghan and
Eraldo for being such warm-hearted individuals. They have always been there
for me. I am blessed to get to know and be their friend. They will forever be in

my thoughts.

I wish to thank The Royal Thai Government, my scholarship sponsor for both
Master’s and Ph.D degrees, for belief in me and for their ﬁnancial support
which helped make all of this possible. The Great Lakes Fisheries Trust
receives my sincere gratitude for supporting the Muskegon River Watershed
Ecological Assessment Project and making it possible for me to conduct
ﬁeldwork, make academic presentations, and work to complete this
dissertation. The funding from this research project, through considerations of
Dr. Qi and Dr. Stevenson, provided a boat for a safe ﬁeld data collection,

valuable hyperspectral data to use in the research, and many more resources

viii

needed in order to complete this research. I am deeply thankful for Dr. Richard
Hauer, University of Montana, for his help providing the airborne hyperspectral
data. He ﬂew his personal aircraft all the way from Montana to acquire the
imagery over the study lakes in Michigan. My gratitude extends to The
Department of Geography for the wonderful academic facility, and The
Graduate School for The Dissertation Completion Fellowship. A million thanks
go to Sharon Ruggles, Graduate Secretary of Department of Geography,
Michigan State University, who always help with regulations and paperwork.
She made my life a lot easier and my experience at Michigan State University

truly delightful.

This research would have been far different without the help of my ﬁeld crews,
Praveen Ummadi, Brian Napoletano, Nathan Torbic, Meghan and Perry. I am
deeply indebted to Meghan and Perry for their dedicated assistance, and always
being there when needed. They have gone beyond duty to help. Dr.
Stevenson’s lab staﬂ and former staff: Vanessa, Lara, Kalina and others, have
provided tremendous help with the ﬁeld and laboratory equipment. Intensive
MATLAB learning could not have gone smoothly without help from Roma —

Rahmatullia Igamberdiev.

Of all people who have parts in my accomplishment, the person who
contributed the most to this research and makes it ever possible is my advisor.

To Dr. Qi, I will forever be thankful.

TABLE OF CONTENTS

LIST OF TABLES .................................................................................. ..XIV
LIST OF FIGURES .............................................................................. . XVIII
CHAPTER 1 .................................................................... . ................. . ...... .. 1
INTRODUCTION .............. . ....... . ........................... 1
1.1 EUTROPHICATION CONSEQUENCES ON WATER QUALITY ...................................... 3
1.2 REMOTE SENSING ALTERNATIVE ...................................................................... 5
1.3 CURRENT PROBLEMS WITH REMOTE SENSING WATER QUALITY ............................ 7
1.4 HYPERSPECTRAL REMOTE SENSING .................................................................. 8
1.4.1 Spectral Data Characteristics .......................................................... 10
1.4.2 Hyperspectral Data Volume and Redundancy ................................. 12
1.5 RESEARCH OBJECTIVES ............................................................................... 12
CHAPTER 2 ............. .............. 14
2.1 SPECTRAL RESPONSE OF WATER BODIES ........................................................ 15
2.2 SPECTRAL CHARACTERISTICS OF CHLOROPHYLL a ............................................ 18
2.3 SPECTRAL CHARACTERISTICS OF COMPLICATING FACTORS (TSS AND DOC) ......... 24
2.3.1 Spectral Characteristics of Suspended Sediments ........................... 28
2.3.2 Spectral Characteristics of Dissolved Organic Carbon ..................... 33
2.4 SPECTRAL CHARACTERISTICS OF ALGAE .......................................................... 35
2.4.1 Spectral Characteristics of Algal Pigments ...................................... 37
2.4.2 Cyanobacteria (Bluegreen Algae) Detection (Potential Toxic Algal
Detection) ........................................................................................ 41
2.5 MOST SENSITIVE SPECTRAL BANDS ................................................................ 44
2.6 SPECTRAL BAND SELECTION METHODS .......................................................... 46
2.6.1 Principle Component Analysis (PCA) ............................................... 46
2.6.2 Spectral Derivative Analysis ............................................................ 49
2.7 SPECTRAL INDICATORS ................................................................................. 51
2.7.1 Spectral Band Ratio Method ............................................................ 51
2.7.2 Statistical Method ........................................................................... 53
CHAPTER 3 .......... ....... . .......................................... 57
3.1 DESCRIPTION OF THE STUDY AREA ................................................................. 58
3.1.1 Climate ........................................................................................... 58
3.1.2 Bedrock Geology ............................................................................. 59

3.1.3 Physiography .................................................................................. 60

3. 1.4 History of Land use / Land cover in Michigan .................................. 60

3. 1.5 Selected Lakes in this Study ........................................................... 63

3.2 REMOTE SENSING DATA ACQUISITION ............................................................ 65
3.2.1 Hand-held Analytical Spectral Devices (ASD) .................................. 65
3.2.2 Airborne Imaging Spectrometer for Applications (AISA) ................... 68

3.3 WATER QUALITY PARAMETER DATA ................................................................ 71
3.3.1 Lab Analytical Methods for Chlorophyll a ........................................ 72
3.3.2 Lab Analytical Methods for Total Suspended Sediment ................... 74
3.3.3 Lab Analytical Methods for Non - Purgable Organic Carbon ............ 74
3.3.4 Lab Analytical Methods for Algal Community Counting ................... 77

3.4 SPECTRAL ANALYTICAL METHODS .................................................................. 80
3.4.1 ASD Data Preprocessing .................................................................. 81
3.4.2 AISA Image Preprocessing ............................................................... 84
3.4.3 Air-Water Interface Correction ......................................................... 89
3.4.4 Noise Reduction .............................................................................. 95
3.4.5 Narrow-Waveband (1 nm) Indicator: Principal Component Analysis 98
3.4.6 Spectral Derivative Analysis .......................................................... 102
3.4.7 Region-Waveband ......................................................................... 108
3.4.8 Area Under Spectral Curve Indicator ............................................. 111
3.4.9 Spectral Curve Height / Width Ratio Indicator ................................ 113
3.4.10 Spectral Band Ratio Technique ................................................... 114
3.4.1 1 Development of New Spectral Indicators for Water Quality .......... 115
CHAPTER 4.... ..... . ......................................... . ...................................... 1 17
4.1 MICHIGAN LAKE WATER QUALITY FROM FIELD OBSERVATION .......................... 117
4.2 OPTIMAL SPECTRAL BANDS ......................................................................... 123
4.3 SPECTRAL INDICATORS ............................................................................... 124
4.3.1 Spectral Indicators for Chlorophyll a ............................................. 124
4.3.2 Spectral Indicators for Secchi Depth ............................................. 128
4.3.3 Spectral Indicators for TSS ............................................................ 133
4.3.4 Spectral Indicators for NPOC ......................................................... 135
4.3.5 Spectral Indicators for Diatom ...................................................... 137
4.3.6 Spectral Indicators for Green Algae ............................................... 140
4.3.7 Spectral Indicators for Bluegreen Algae ......................................... 144

4.4 MODEL VALIDATION ................................................................................... 148
4.4.1 Chlorophyll a Model Validation ..................................................... 148
4.4.1.1 Chlorophyll a ASD Validation ................................................. 149
4.4.1.2 Chlorophyll a AISA Validation ................................................. 151

4.4.2 Secchi Depth Model Validation ...................................................... 153
4.4.2.1 Secchi Depth ASD Validation .................................................. 154
4.4.2.2 Secchi Depth AISA Validation ................................................. 156

4.4.3 TSS Model Validation .................................................................... 158

4.4.3.1 TSS ASD Validation ................................................................ 158
4.4.3.2 TSS AISA Validation ............................................................... 159

4.4.4 NPOC Model Validation ................................................................. 162
4.4.4.1 NPOC ASD Validation ............................................................. 162
4.4.4.2 NPOC AISA Validation ............................................................ 163

4.4.5 Algal Model Validation ................................................................... 165
4.4.5.1 Diatom ASD Validation ........................................................... 166
4.4.5.2 Diatom AISA Validation .......................................................... 168
4.4.5.3 Green Algae ASD Validation .................................................... 170
4.4.5.4 Green Algae AISA Validation ................................................... 172
4.4.5.5 Bluegreen Algae ASD Validation ............................................. 173
4.4.5.6 Bluereen Algae AISA Validation .............................................. 175

4.5 DISCUSSION ON SPECTRAL INDICATORS FOR WATER QUALITY ASSESSMENT ....... 177
4.5.1 Chlorophyll a ................................................................................ 177
4.5.2 Secchi Depth ................................................................................. 179
4.5.3 TSS and NPOC .............................................................................. 181
4.5.4 Algae ............................................................................................. 182
CHAPTER 5 ...... . .................................................................................... 186
CONCLUSIONS AND FUTURE RESEARCH .............................................. 186
5.1 OPTIMAL SPECTRAL BANDS ......................................................................... 187
5.2 SPECTRAL INDICATORS FOR WATER QUALITY ASSESSMENT .............................. 189
5.3 EFFECTIVE FILTERS FOR NOISE REMOVAL ..................................................... 190
5.4 MAJOR FINDINGS ....................................................................................... 19 1
5.5 FUTURE RESEARCH .................................................................................... 193
APPENDIX A ..................................................................................... 196
WATER QUALITY FIELD WORK PROTOCOL ........................................... 196
APPENDIX B ......................................................................................... 202
AISA IMAGERY . .................................................................................... 202
APPENDIX C ......................................................................................... 207
NOISE REMOVAL FILTERS COMPARISON ............................................. 207
APPENDIX D .......................................... . ............................................ .. 213
AREA UNDER THE SPECTRAL CURVE ................................................... 213
APPENDIX E ......................................................................................... 219
MAXIMUM SPECTRAL CURVE HEIGHT ................................................. 2 19

APPENDIX F .......................................................................................... 225

NARROW-WAVEBAND VOLUME REFLECTAN CE ..................................... 225
APPENDIX G ................................................ . ..................................... 23 1
PRACTICAL PROBLEM WITH AISA IMAGE ............................................. 23 1

LIST OF TABLES

Table 2-1 Optical properties of pure water (derived from various sources by
Bukata et al., 1995) ......................................................................... 16

Table 2-2 Characteristics of common major associations of the phytoplankton
in relation to increasing lake fertility (after Hutchinson, 1967 as cited

in Wetzel, 2001) ............................................................................... 36
Table 2-3 Summary of wavebands from literature review ................................ 45
Table 3-1 Spectral conﬁgurations of the 20 band AISA data ............................ 69
Table 3-2 Spectral conﬁgurations of the 30 band AISA data ............................ 70
Table 3-3 Pearson correlation between 20 band AISA and ASD spectra .......... 85
Table 3—4 Pearson correlation between 30 band AISA and ASD spectra .......... 87
Table 3-5 Selected waveband regions for water quality assessment ............... 109
Table 4- 1 Pearson correlation matrix of water quality indicators ................... 122
Table 4-2 Identiﬁed region-wavebands .......................................................... 123
Table 4-3 Identiﬁed narrow-wavebands ......................................................... 124
Table 4-4 Chlorophyll a area models using the optimal wavebands ............... 125

Table 4-5 Chlorophyll a height / width models using the optimal wavebands. 125

Table 4-6 Chlorophyll a narrow-waveband models using the optimal wavebands
..................................................................................................... 127

Table 4-7 Secchi depth area models using the optimal wavebands ................ 128
Table 4-8 Secchi depth a height / width models using the optimal wavebands 129

Table 4-9 SD narrow-waveband models using the optimal wavebands .......... 132

Table 4- 10 TSS area models using the optimal wavebands ........................... 133

Table 4-11 TSS height/width models using the optimal wavebands .............. 133
Table 4-12 TSS narrow-waveband models using the optimal wavebands ...... 134
Table 4-13 N POC area models using the optimal wavebands ........................ 135
Table 4-14 NPOC height / width models using the optimal wavebands ........... 135

Table 4-15 NPOC narrow-waveband models using the optimal wavebands 136
Table 4- 16 Diatom area models using the optimal wavebands ...................... 138
Table 4-17 Diatom height / width models using the optimal wavebands ......... 138
Table 4- 18 Diatom narrow-waveband models using the optimal wavebands . 139
Table 4-19 Green algae area models using the optimal wavebands ............... 141
Table 4-20 Green algae height/width models using the optimal wavebands.. 142

Table 4-2 1 Green algae narrow-waveband models using the optimal wavebands
..................................................................................................... 142

Table 4-22 Bluegreen algae area models using the optimal wavebands ......... 145

Table 4-23 Bluegreen algae height/ width models using the optimal wavebands

..................................................................................................... 146

Table 4-24 Bluegreen algae narrow-waveband models using the optimal
wavebands .................................................................................... 146
Table 4-25 Chlorophyll ASD validation .......................................................... 150

Table 4-26 Pearson correlation matrix of actual and predicted chlorophyll from
ASD ............................................................................................... 150

Table 4-27 Chlorophyll AISA validation ......................................................... 152

Table 4-28 Pearson correlation matrix of actual and predicted Chlorophyll from
AISA .............................................................................................. 152

Table 4-29 Pearson correlation matrix of actual and predicted chlorophyll from
AISA without Mitchell and Muskegon Lakes .................................. 153

Table 4-30 Secchi depth ASD validation ........................................................ 155

Table 4-3 1 Pearson correlation matrix of actual and predicted Secchi depth
from ASD ...................................................................................... 155

Table 4-32 Secchi depth AISA validation ....................................................... 156

Table 4-33 Pearson correlation matrix of actual and predicted Secchi depth
from AISA ...................................................................................... 157

Table 4-34 Pearson correlation matrix of actual and predicted Secchi depth
from AISA without Mitchell and Muskegon Lakes .......................... 157

Table 4-35 TSS ASD validation ..................................................................... 158

Table 4-36 Pearson correlation matrix of actual and predicted TSS from ASD
..................................................................................................... 159

Table 4-37 TSS AISA validation ..................................................................... 160

Table 4-38 Pearson correlation matrix of actual and predicted TSS from AISA
..................................................................................................... 160

Table 4-39 Pearson correlation matrix of actual and predicted TSS from AISA
without Mitchell and Muskegon Lakes .......................................... 16 1

Table 4-40 NPOC algae ASD validation ......................................................... 162

Table 4-41 Pearson correlation matrix of actual and predicted NPOC from ASD
..................................................................................................... 163

Table 4-42 NPOC AISA validation .................................................................. 164

Table 4-43 Pearson correlation matrix of actual and predicted NPOC from AISA

..................................................................................................... 164
Table 4—44 Statistical comparison of modeling and validating data set .......... 165
Table 4-45 Diatom ASD validation ................................................................ 167

Table 4-46 Pearson correlation matrix of actual and predicted diatom from ASD
..................................................................................................... 167

Table 4-47 Diatom AISA validation ............................................................... 168

Table 4-48 Pearson correlation matrix of actual and predicted diatom from AISA
..................................................................................................... 169

Table 4-49 Pearson correlation matrix of actual and predicted diatom from AISA
without Mitchell and Muskegon Lakes .......................................... 170

Table 4-50 Green algae ASD validation ......................................................... 171

Table 4-51 Pearson correlation matrix of actual and predicted green algae from
ASD ............................................................................................... 17 1

Table 4-52 Green algae AISA validation ........................................................ 172

Table 4-53 Pearson correlation matrix of actual and predicted green algae from
AISA .............................................................................................. 173

Table 4-54 Bluegreen algae ASD validation ................................................... 174

Table 4-55 Pearson correlation matrix of actual and predicted bluegreen algae
from ASD ...................................................................................... 174

Table 4-56 Bluegreen algae AISA validation .................................................. 175

Table 4-57 Pearson correlation matrix of actual and predicted bluegreen algae
from AISA ...................................................................................... 176

Table 4-58 Pearson correlation matrix of actual and predicted bluegreen algae

from AISA without Mitchell, Muskegon, and Higgins Lakes ........... 177
Table 4-59 Chlorophyll model performances ................................................. 178
Table 4-60 Secchi depth model performances ............................................... 179
Table 4-61 TSS model performances ............................................................. 181
Table 4-62 NPOC model performances .......................................................... 181
Table 4-63 Diatom models performances ...................................................... 182

Table 4-64 Green algae models performances ............................................... 182
Table 4-65 Bluegreen algae models performances ......................................... 183

Table A— 1 List of selected sample lakes ......................................................... 197

LIST OF FIGURES

Figure 1-1 Hyperspectral “cube” of image data such as recorded by an imaging
spectrometer (after Richard and Jia, 1999) ...................................... 1 1

Figure 2-1 Absorption and scattering of light in pure water (after Jensen, 2000)
....................................................................................................... 17

Figure 2-2 Percent reﬂectance of clear and algae-laden water based on in situ
spectroradiometer measurements (after Han, 1997) ........................ 19

Figure 2-3 Spectra reﬂectance of Frisian waters, measured in situ in August
1995, with Landsat TM bands 1-4 superimposed (after Dekker et al.,
2001) ............................................................................................... 21

Figure 2-4 The lower three curves represent the absorption of water and its
constituents: the dashed curve the absorption of water, the ( ----- )
absorption of 60 mg/l of Chlorophyll, the dotted curve is the DOM
absorption (0. 14 / m @440 nm) and the solid represents the sum of
the water, chlorophyll and DOM absorption. The upper hatched
curve is an observed reﬂectance spectrum over the lake (after Green
1998) ............................................................................................... 22

Figure 2-5 Example reﬂectance from a high chlorophyll content lake (85 mg /1 —
dash line) and one from a low chlorophyll content lake (3 mg/l; TSM
3.5 mg/l — solid line) (after Green, 1998) ......................................... 23

Figure 2-6 Reﬂectance spectra of the water bodies studied in Dall’Olmo and
Gitelson (2005). Some examples are highlighted: curve “Low” (Chl—a =
1 1 mg/l; TSS = 5 mg/l), curve “High” (Chl-a = 89 mg/l; TSS = 21
mg/l), and curve “Moderate” (Chl-a = 24 mg/ 1; TSS = 55 mg/l) ....... 24

Figure 2-7 Absorption spectral properties of optically active constituents in
water (after Dall’Olmo, 2005) ........................................................... 27

Figure 2-8 Percent reﬂectance of algae-laden water with various concentrations
of suspended sediment ranging from 0 — 500 mg/l (after Han, 1997)
....................................................................................................... 30

Figure 2-9 (a) In situ spectral reﬂectance measurements of Clear water and clear
water with various levels of clayey soil suspended sediment
concentrations from 0 -— 1,000 mg/l. (b) In situ spectral reﬂectance
measurements of Clear water and Clear water with various levels of
silty soil suspended sediment concentrations (after Lodhi et al., 1997)

....................................................................................................... 32

Figure 2- 10 Examples of water reﬂectance for diﬂerent chlorophyll a
concentrations in natural waters (after Dall’Olmo, 2005) ................. 35

Figure 2-1 1 Maximum and minimum absorption values in the reﬂectance
spectra of pure algal culture (after Shaﬁque et al., 2001) ................. 39

Figure 2- 12 Two reﬂectance curves from lakes with similar high levels of
Chlorophyll — but one contains phycocyanin, indicating a
cyanobacteria (the phycocyanin absorption peak at 630 nm is
expressed as a trough in reﬂectance), while the other exhibits the
properties of green algae (after Green, 1998) ................................... 40

Figure 2-13 Reﬂectance model for Chlorophyll = 10 mg/ m3 (after Subramaniam
et al., 1999) ..................................................................................... 43

Figure 2- 14 Standard deviation of the data at different wavelengths for Lake
Erken (after Flink et al., 2001) ......................................................... 47

Figure 3-1 Geographic location of the lakes sampled in the study ................... 64

Figure 3-2 The variation of up- and down-welling radiance as measured from
ASD. The 10x of mean of upwelling radiance are plotted to scale the
reﬂectance values up for visual comparison with the downwelling
radiance .......................................................................................... 67

Figure 3-3 Spectral analytical procedure ﬂow diagrams .................................. 80

Figure 3-4 ASD spectral signatures (350 nm — 2,500 nm) of 48 representative
water samples ................................................................................. 82

Figure 3-5 ASD spectral signatures (400 nm - 900 nm) of 48 representative
water samples ................................................................................. 82

Figure 3-6 ASD spectral characteristics of different biophysical dominate in
waters ............................................................................................. 83

Figure 3-7 Comparison between ASD and 20 band AISA spectral signatures of
Higgins Lake. AISA stripe image overlay on Landsat image. Red point
represents the corresponding sample site ........................................ 86

Figure 3-8 Comparison between ASD and 30 band AISA spectral signatures of
Hess Lake. AISA stripe image overlay on Landsat image. Red point

represents the corresponding sample site ........................................ 88

Figure 3-9 Down— and up-welling radiances from water bodies (after Jensen,

2000) ............................................................................................... 90
Figure 3-10 Remote sensing reﬂectance Rf. ..................................................... 94
Figure 3-11 Subsurface volume reﬂectance R(O-) ............................................ 94

Figure 3- 12 Examples of Mean and Savitsky Golay (Sgolay) ﬁlter transformation
on Pickerel Lake .............................................................................. 97

Figure 3-13 Examples of Daubechies (dbl) and Symlet (s8) wavelet de-noising
transformation on Pickerel Lake ...................................................... 97

Figure 3-14 Example of PCA result ............................................................... 102

Figure 3- 15 First derivatives approximation from 4 different ﬁltrations of the
volume reﬂectance spectra ............................................................ 105

Figure 3-16 Second derivatives approximation from 4 different ﬁltrations of the
volume reﬂectance spectra ............................................................ 107

Figure 3- 17 Example of region-waveband selection on Sgolay method .......... 110

Figure 3- 18 Examples of curve ﬁtting on 1at derivative data. Graph (a) and (b)
represent absorption curves, and (c) and ((1) represent reﬂectance

curves ........................................................................................... 1 12
Figure 3- 19 Examples of height and width measurements ............................ 113
Figure 4-1 Bar graphs of SD, CHL, NPOC and TSS ....................................... 118
Figure 4-2 Bar graphs of diatom, green and bluegreen algae ......................... 119

Figure 4-3 Histogram distributions of the raw water quality parameters ....... 120

Figure 4-4 Histogram distributions of LN transformed water quality parameters

..................................................................................................... 12 1
Figure 4-5 Pearson correlation graphs and histogram plots of water quality

indicators ...................................................................................... 122
Figure 4-6 Volume reﬂectance curves ........................................................... 130

Figure 4—7 Examples of volume reﬂectance curves on different SD in B and C
waveband regions .......................................................................... 131

Figure 4-8 Examples of ﬁtting curves on different SD in B and C waveband
regions .......................................................................................... 13 1

Figure 4-9 Comparison between high NPOC lake (Croton dam pond) and other
lakes with similar water condition but lower NPOC ....................... 137

Figure 4- 10 Examples of volume reﬂectance of green algal dominated waters 14 1

Figure 4- 11 Examples of spectra that produce the same height but different
under curve area ........................................................................... 143

Figure 4- 12 Examples of volume reﬂectance of bluegreen algal dominated
waters ........................................................................................... 144

Figure 4- 13 Correlation graphs between actual and predicted CHL from ASD150

Figure 4-14 Correlation graphs between actual and predicted CHL from AISA
..................................................................................................... 152

Figure 4-15 Correlation graphs between actual and predicted CHL from AISA
without Mitchell and Muskegon Lakes .......................................... 153

Figure 4- 16 Correlation graphs between actual and predicted SD from ASD. 155
Figure 4- 17 Correlation graphs between actual and predicted SD from AISA 157

Figure 4- 18 Correlation graphs between actual and predicted SD from AISA
without Mitchell and Muskegon Lakes .......................................... 157

Figure 4- 19 Correlation graphs between actual and predicted TSS biomass from
ASD ............................................................................................... 159

Figure 4-20 Correlation graphs between actual and predicted TSS from AISA
..................................................................................................... 160

Figure 4-21 Correlation graphs between actual and predicted TSS from AISA
without Mitchell and Muskegon Lakes .......................................... 16 1

Figure 4-22 Correlation graphs between actual and predicted NPOC from ASD
..................................................................................................... 163

Figure 4-23 Correlation graphs between actual and predicted NPOC from AISA
..................................................................................................... 164

Figure 4-24 Distribution of modeling and validating data set ........................ 166

Figure 4-25 Correlation graphs between actual and predicted diatom biomass
from ASD ...................................................................................... 167

Figure 4-26 Correlation graphs between actual and predicted diatom from AISA
..................................................................................................... 169

Figure 4-27 Correlation graphs between actual and predicted diatom from AISA
without Mitchell and Muskegon Lakes .......................................... 170

Figure 4-28 Correlation graphs between actual and predicted green algal
biomass from ASD ......................................................................... 171

Figure 4-29 Correlation graphs between actual and predicted green algae from
AISA .............................................................................................. 17 3

Figure 4-30 Correlation graphs between actual and predicted bluegreen algal
biomass from ASD ......................................................................... 174

Figure 4-31 Correlation graphs between actual and predicted green algae from
AISA .............................................................................................. 176

Figure 4-32 Correlation graphs between actual and predicted bluegreen algae
from AISA without Mitchell, Muskegon, and Higgins Lakes ........... 177

Figure A- 1 Trophic conditions of the selected sample lakes ........................... 196

Images in this dissertation are presented in color.

CHAPTER 1

INTRODUCTION

Lakes are valuable water resources. Tens of thousands of inland lakes have
served as crucial resources for drinking water, irrigation, industry,
transportation, recreation, ﬁshing, support of biodiversity, and sheer aesthetic
enjoyment for the 516,000 km2 area of the Upper Midwest region (Carpenter et
al., 1998). These lakes are of immense importance to the region (Lillesand,
2002). Fresh water is a limited resource. Although over 70 percent of the
Earth’s surface is water—covered, only 2.6 percent of it is freshwater. It can be
increased only slightly with tremendous cost, for example, desalinization of
ocean water requires enormous energy, and once the product is obtained
distribution is energetically prohibitive. For many reasons, the usable
freshwater supply is in reality much smaller than the absolute total. Of all the
freshwater, 77 percent is tied up in the polar ice caps and 1 1 percent is stored
in deep groundwater aquifers leaving 12 percent in an active circulation (Brooks
et al., 2003). Water consumption has increased exponentially with demotechnic
growth. And potentially the most serious factor from the growth is the severe

degradation by contaminants to water quality (Wetzel, 2001).

Fresh waters around the world are experiencing accelerating rates of qualitative
and quantitative degradation (Wetzel, 1992). Throughout human history, water
has been used to wash away and dilute pollutants. Pollutant inputs have

increased in recent decades and have degraded water quality of many rivers,

lakes, and coastal oceans. Degradation of these vital water resources can be
measured as the loss of natural ecosystems, their component species, and the
amenities they provide (Postel and Carpenter, 1997; US. EPA, 1996). Water
shortages are increasingly common and likely to become more severe in the
future (Postel, 1997; Postel et al., 1996). In 2001, the United Nations
commemorated a World Day for Water in which it concluded that the demands
for freshwater exceeded supplies by 17 percent and over the next 25 years, two-
thirds of the world population will experience a severe water shortage (Brooks et

al., 2003).

A well documented effect of human activities upon aquatic ecosystems is
eutrophication, a process whereby water bodies, such as lakes, estuaries, or
slow-moving streams receive excess nutrients that stimulate excessive
productivity, simpliﬁcation of biotic communities, and a reduction in the ability
of the metabolism of the organisms to adapt to the nutrient loading. When
eutrophication occurs, excessive inputs often exceed the capacity of the
ecosystems to be balanced, thus the conditions lead to reduced stability of the
ecosystem. In order to effectively maintain the quality of inland waters, it is
necessary to monitor humans’ utilization of these resources independently in
terms of residential, industrial, and agricultural activities (Wetzel, 2001).
Although the fundamental laws of resource utilization may be recognized by
most agencies and industries, they are not being seriously implemented.
Therefore, an effective water quality assessment and monitoring techniques are

needed to maintain a sustainable natural ecosystems.

1.1 Eutrophication Consequences on Water Quality

Eutrophication caused by excessive inputs of phosphorus and nitrogen is the
most common impairment of surface waters in the United States (US. EPA,
1990), with impairment measured as the area of surface water not suitable for
designated uses such as drinking, irrigation, industry, recreation, or ﬁshing.
Eutrophication accounts for approximately 50 percent of impaired lake areas
and 60 percent of the impaired river reaches in the United States (US. EPA,
1996), and is the most widespread pollution problem of US. estuaries (NRC,

1993i

Eutrophication has many negative effects on aquatic ecosystems. Perhaps the
most obvious consequence is the increased growth of algae and aquatic weeds
that interfere with use of the water for ﬁsheries, recreation, industry,
agriculture, and drinking water. As the mass of algae in the water grows, the
water may become murkier and less aesthetically pleasing. Particularly, when
the algae die and decompose, periods of oxygen depletion (hypoxia and anoxia)
occur more frequently (Carpenter et al., 1998). Even living algae create
conditions favorable for some species over others and may cause shifts in the
structure of phytoplankton, zooplankton, and bottom-dwelling (benthic)
communities (Howarth et al., 2000). Eutrophication brings on ecological
changes that decrease the biological diversity — the variety of living organisms -

in the ecosystem (Seehausen et al., 1997).

Explosive growth of nuisance algae are among the most damaging effects of
eutrophication (Anderson and Garrison, 1997). Blooms of harmful algae such
as red and brown tide organisms may become more frequent and extensive.
These algae are harmful to humans and other organisms, sometimes resulting
in human illness via shellﬁsh poisonings (Carpenter et al., 1998). Just as
important, subtle changes in the plankton community and other ecological
factors may trigger reduced growth and recruitment of ﬁsh species and a
lowered ﬁshery production (Howarth et al., 2000). For example, in marine
systems, blooms of Lyngbya majuscule in Moreton Bay, southeast Queensland,
Australia have signiﬁcantly impacted the environment (Relfsema et al., 2001).
In freshwater, blooms of cyanobacteria (bluegreen algae) are a prominent
symptom of eutrophication (Smith, 1998; Kotak et al., 1993; MCComb and
Davis, 1993). These blooms contribute to a wide range of water-related
problems including summer ﬁshkills, foul odors, and unpalatable drinking
water (Kotak et al., 1994; Palmstrom et al., 1988). Water-soluble neuro— and
hepatotoxins, released when cyanobacterial blooms die or are ingested, can kill
livestock and may pose a serious health hazard to humans (Carpenter et al.,

1998).

Various mathematical models have been developed and applied to rivers, lakes,
and estuaries in an effort to monitor, simulate and control eutrophication (e.g.,
Kloiber et al., 2002; Dekker et al., 2001; Flink et al., 2001; Giardino et al.,
2001; Koponen et al., 2001; Pulliainen et al., 2001; Shaﬁque et al., 2001;

Subramaniam et al., 1999). Most water quality models demand comprehensive

water quality sampling programs. In an ideal circumstance, monitoring of
water bodies includes the determination of concentrations of water quality
variables and the processes that generate their spatial distribution and
temporal variation of those variables (Fisher, 1994; Van Stokkom et al., 1993).
However, the conventional measurement of water quality requires in situ
sampling, and expensive and time-consuming laboratory work (Giardino et al.,
2001). Therefore, it is usually based on the determination of concentrations at
one or only a few ﬁxed stations that are assumed to represent the overall
distribution of phytoplankton in a lake, or the spatial interpolation of the
concentration from the stations to obtain continuous ﬁeld maps of the various
water quality parameters (Kallio et al., 2003). Algal blooms are extremely
patchy, both temporally and spatially. Consequently, they often remain
unobserved using the traditional sampling methods based on temporally sparse
sampling at ﬁxed monitoring stations (Harma et al., 2001). Traditional in situ
sampling methods also do not provide the spatial overview that is necessary for
the regional assessment and monitoring of lake water quality (Shaﬁque et al.,
2001). On the other hand, optical indications of water qUality have the
potential of enhancing the abilities of resource managers to monitor water

bodies in a timely and cost-effective manner.

1.2 Remote Sensing Alternative
Remote sensing is deﬁned as acquisition of information about the properties of
electromagnetic waves emitted, reﬂected or diffracted by the sensed objects

without being in direct physical contact. Broad scope remote sensing based

water quality research has been developed to detect environmental indicators
that are useful in assessing, quantifying and monitoring inland water quality.
More fundamentally, the absorption and scattering of light by components of
the lacustrine water column provide basic information about the substances
suspended in the water (Jupp et al., 1994). Although a fairly new method, the
development of spectral indices can be a useful and effective tool for the
diagnosis of water conditions by water resource managers (Shaﬁque et al.,

20011

Remote sensing offers a signiﬁcant source of information, and methods have
been developed for operational large-scale monitoring of water quality (Harma et
al., 2001). For example, remote sensing enables the monitoring of a wide
spatial extent of phytoplankton distribution in the surface water layer more
effectively (Shaﬁque et al., 2001). Reliable spatial coverage and cost-efficient
remote monitoring techniques of lakes and coastal waters are generally growing
in importance as a consequence of increasing symptoms of eutrophication
processes (Giardino et al., 2001; Harma et al., 2001). A number of studies have
shown that applications of remote sensing can meet the demand for the large
sample sizes required of water quality studies conducted on the watershed
scale. Algal blooms have been mapped successfully from remotely sensed data
in a number of different riverine (Lathrop and Lillesand, 1989; Lillesand et al.,
1983), estuarine (Harding et al., 1995; Verdin, 1985), and oceanic (Ruiz—Azuara,
1995) environments around the world. Imagery from satellite and aircraft

remote sensing systems have been used in the assessment of water quality

parameters such as temperature, chlorophyll a, turbidity, and total suspended
solids (Relfsema et al., 2001; Jupp et al., 1994). Estimation of chlorophyll a
distribution in lakes by remote sensing techniques has included the use of
airborne photography (Wrigley and Home, 1974), airborne spectrometry (George
and Malthus, 2001; Ostlund et al., 2001; Heege and Fischer, 2000; Jupp et al.,
1994; Dekker et al., 1992) and satellite sensors (Giardino et al., 2001; Ostlund
et al., 2001; Vos and Rijkeboer, 2000; Dekker and Peters, 1993). Results were

usually reported in the form of concentration maps (Kallio et al., 2003).

1.3 Current Problems with Remote Sensing Water Quality

Data acquisition by remote sensing is fast (e.g., tens of lakes may be acquired
within a day by an airborne sensor or hundreds of lakes by a spacebome
sensor), and large areas can be surveyed over a short period of time. However,
the spectral and spatial conﬁgurations of current aquatic satellite sensors are
not suitable for inland water quality monitoring. In most cases they are not
suitable for phytoplankton monitoring in lakes due to their spectral
conﬁguration and poor spatial resolution (typically from several hundred meters
to 1,000 m). The low spatial resolution of most satellite data can produce
sources of error in empirical approaches used to assess water quality
indicators. For example, a single in situ sample may not be representative of an
entire pixel area. In most cases, a single pixel is greater than several meters in
diameter, and it is rare for a single object or target feature to ﬁll any one pixel.
Thus, the characteristics of any pixel can rarely be considered truly

homogenous (Tsi and Philpot, 1998). In addition, the accuracy of locating the

pixel that corresponds to the in situ sample may be uncertain depending on the

geolocation method used (Giardino et al., 2001).

Spaceborne sensors provide the global coverage of the Earth’s surface
conditions at diﬂerent spatial and temporal resolutions, but the efﬁcacy of all
current spaceborne remote sensing systems for detailed Characterization of
water quality parameters is limited by their spatial and spectral resolutions.
Sensors with high spatial resolution do not have a sufﬁcient number of narrow
spectral bands (e.g., IKONOS, QuiCkBird, Landsat, and ASTER), while narrow-
waveband spaceborne sensors usually have coarse spatial resolution (e. g.,
MODIS). Satellite sensor systems such as Landsat TM and ETM+, and ASTER
currently provide data of sufﬁcient spatial resolution for inland lake
applications. However, their spectral resolution is questionable. It is important
to recognize that these sensors average the spectral information over the entire
width of the spectral band (Dekker et al., 1992). Narrow-waveband sensors
may provide better fundamental information about the biophysical
Characteristics of inland waters. Various ground-based radiometers have been
used with different bandpass ﬁlters to mimic the operational airborne
radiometers and spaceborne scanners. Airborne sensors are generally designed

to serve as a prototype for future spaceborne sensor systems.

1.4 Hyperspectral Remote Sensing
Hyperspectral sensors (0.4 pm - 2.5 pm wavelength range) capture the

hopefully unique spectra (or ‘spectral signature) of an object. These signatures

can be used to identify and quantify the materials of which it is composed (The
Canadian Space Agency, 2003). With this principle, hyperspectral data enable
the identiﬁcation of the Earth’s surface features with greater thematic accuracy.
Airborne imaging spectrometers have been used to assess the trophic status of
lakes and to map the spatial distribution of phytoplankton (Malthus et al.,
1996). The analysis of hyperspectral imagery involves the decomposition of
each reﬂectance pixel into its biophysical constituents. The identity of these
constituents is determined by comparison with ‘library’ spectra of known

materials measured in the ﬁeld or in the laboratory (Richards and Jia, 1999).

The previous generation of spaceborne optical imagers are limited to either
panchromatic or multispectral devices providing only a few spectral bands and
limited resolving power. Numerous effective methods, mostly derived from
multivariate statistics, have been developed and applied successfully for spatial
or spectral analysis of these data (Tsai and Philpot, 1998). Hyperspectral
imagers typically collect data in tens to several hundred contiguous, narrow
bands in the electromagnetic spectrum. The large numbers of bands that are
simultaneously imaged produce vast quantities of information. With higher
resolution, spectrally continuous data, researchers tended to select a subset of
suitable bands to optimize the existing algorithms for multispectral data
analysis or to generate new algorithms based on traditional multispectral
concepts (Penuelas et al., 1994; Martin and Aber, 1993; Chappelle et al., 1992;

Curran et al., 1992).

The new generation of airborne imaging spectrometers, such as the Airborne
Imaging Spectrometer for Applications (AISA) and the Compact Airborne
Spectrographic Imager (CASI), offer considerable advances in terms of
radiometric sensitivity and operational ﬂexibility (George and Malthus, 2001).
The main problems with hyperspectral data are the substantial redundancy of
the information, the difﬁculties in identifying the optical bandwidth and center
wavelength of the bands that maximize the explanation of biophysical
attributes, and the system limitations associated with the storage of the image

data volumes.

1.4.1 Spectral Data Characteristics

Given the enormous number of wavebands recorded, the data produced by the
imaging spectrometers are different from those of multispectral scanners —
leading to the term hyperspectral. The data produced for a given geographical
area can be viewed as a cube, as shown in Figure 1- 1, having three dimensions
that represent spatial location (two dimension) and spectral wavelength (the

third dimension) (Richards and Jia, 1999).

10

 

Figure 1-1 Hyperspectral “cube” of image data such as recorded by an imaging

Spectrometer (after Richard and Jia, 1999)

When displaying remotely sensed data on the display device, only three of the
spectral bands are usually assigned to the red, green and blue color elements of
the device. Careful band selection ensures the most informative display. This
is relatively simple for multispectral data, such as the six 30 m bands from
Landsat, but with hyperspectral data, selecting the three bands to display can
be challenging. Choosing the most appropriate three channels to display is not
straightforward and, in any case, would lead to substantial loss of the spectral
beneﬁts offered by these type of data. However, unless spectral transformations
are employed, a set of three bands comparable to those used with multispectral
imagery are often adopted (near IR, red, green) for simple display of the data

(Richards and Jia, 1999).

1.4.2 Hyperspectral Data Volume and Redundancy

It is obvious that the major differences between multispectral and hyperspectral
data (e.g., Landsat versus AVIRIS) is the number of wavebands (7 versus 224)
and the radiometric quantization (8 versus 10 bits per pixel per band).
Disregarding differences in spatial resolution, the relative data volume per pixel
are 7 x 8 vs. 224 x 10 — i.e., 56 : 2240 bits per pixel. For each pixel, there are
40 times as many bits for AVIRIS as for TM data. Consequently, storage and
transmission of hyperspectral data are issues for consideration (Richards and

Jia, 1999).

Unfortunately, 40 times as much data per pixel does not imply 40 times as
much information can be extracted about the ground cover types being imaged.
Even though additional data often enhance the possibility in discovering that
information, much of it does not add to the potential information content.
Hyperspectral data often contain substantial overlap or redundancy of
information content among the bands of data recorded for a given pixel.
Spectral redundancy means that the information content of one band can be
fully or partly predicted from the other bands in the data (Richards and Jia,

19991

1.5 Research Objectives

The objectives of this study are to:

(1) identify the optimal spectral bands that are most sensitive to water quality
indicators in the various water bodies within Michigan;

(2) develop improved spectral water quality indicators; and

12

(3) determine the most effective ﬁlters for noise removal in hyperspectral remote

sensing data.

Many previous studies relied on the correlation of local in situ measurements of
Chlorophyll a, suspended sediment, with the remote sensing data. These
algorithms are not truly generalizable. They are only good for the particular
location and cannot usually be generalized across space or time. This study
attempts to develop generalizable algorithms that are spatially and temporally
independent. The main purpose is to detect absorption and reﬂectance features
within the spectral data, and then to develop the spectral indicators, such as
under-curve area, curve height / width ratio, or narrow-waveband indices, that
could estimate the water quality parameters (Chlorophyll a, Secchi depth, total

suspended sediment, non-purgable organic carbon, and algae biomass).

l3

CHAPTER 2

LITERATURE REVIEW

Several remote sensing studies have estimated water quality parameters such
as Secchi depth, Chlorophyll a, and bluegreen algae. Some researchers used
multispectral remote sensing data to map the general water quality indicator for
Secchi depth; however, the data do not provide enough spectral resolution for
the detection of algae or accurate assessment of Chlorophyll a. Other
researchers used hyperspectral data to predict suspended sediment, chlorophyll
a and harmful bluegreen algae in the ocean where concentration of
complicating optical factors (e.g., total suspended sediment and dissolved
organic carbon) are very low. Various studies used different analytical
techniques, such as principle component analysis (PCA), derivative
spectroscopy, and regression techniques, to identify optimal narrow spectral
wavebands and develop water quality models. None of the reviewed studies
developed water quality indicators by ﬁtting polynomial curves in the region-
wavebands as presented in this study (Chapter 3). The region-waveband
indicators can be more sensitive to Changes in biophysical variables and less
affected by noise from the atmosphere or the sensor itself than the narrow-
waveband indicators. The literature review presented in this chapter served as
a guideline of what has been done and what can be done to improve water

quality assessment algorithms.

14

2.1 Spectral Response of Water Bodies

When conducting remote sensing investigations on water bodies, it is useful to
understand how pure water selectively absorbs and scatters incident solar.
Bukata et al. (1995) summarized the absorption coefficient a(A), the scattering
coefﬁcient b(A), and the total attenuation coefﬁcient C(A) of pure water molecules
at wavelengths from 250 nm — 800 nm from a number of studies (Table 21).
Several important relationships were observed when the absorption and

scattering data were graphed, as shown in Figure 2-1 (Jensen, 2000).

15

 

Table 2-1 Optical properties of pure water (derived from various sources by

Bukata et al., 1995)

 

 

 

Wavelength Absorption Scattering Total .
(nm) am (m-l) be) (m-l) Attenuatm
C(A) (m1)
250 — ultraviolet 0.190 0.0320 0.2200
300 - ultraviolet 0.040 0.0150 0.0550
320 - ultraviolet 0.020 0.0120 0.0320
350 - ultraviolet 0.012 0.0082 0.0202
400 - violet 0.006 0.0048 0.0108
420 - violet 0.005 0.0040 0.0090
440 - violet 0.004 0.0032 0.0072
460 — dark blue 0.002 0.0027 0.0047
480 - dark blue 0.003 0.0022 0.0052
500 - light blue 0.006 0.0019 0.0079
520 - green 0.014 0.0016 0.0156
540 - green 0.029 0.0014 0.0304
560 — green 0.039 0.0012 0.0402
580 - yellow 0.074 0.00 11 0.0751
600 — orange 0.200 0.0009 0.2009
620 - orange 0.240 0.0008 0.2408
640 - red 0.270 0.0007 0.2707
660 - red 0.310 0.0006 0.3106
680 — red 0.380 0.0005 0.3806
700 - red 0.600 0.0005 0.6005
740 — near-infrared 2.250 0.0004 2.2504
760 - near-infrared 2.560 0.0004 2.5604
800 - near-infrared 2.020 0.0003 2.0203

 

 

 

 

l6

 

Absorption and Scattering Attenuation Coefﬁcient

Figure 2- 1 Absorption and scattering of light in pure water (after Jensen, 2000)

0.10

 

0.09 ~

0.08 ~

0.07 -

0.06 ~

0.05 ~

0.04 -

0.03 ..

0.02 -

0.01 ~

 

0.00

250 300 350 400 450 500 550 600 650 700 750 800
Wavelength (nm)

I

 

I

Absorption

Scattering

viole .

I

T

blue

I

I

green

I

—_

 

red

near-
infrare

Blue wavelength region from approximately 400 nm - 500 nm had the least

amount of absorption and scattering of incident light in the water column, with

the minimum absorption at 460 nm — 480 nm. The wavelengths of violet to

light blue light penetrated further than any other range of light into the water
column because they had the best transmission (Clark et al., 1997). The water
column absorbed incoming irradiance in the green and yellow wavelengths from

520 nm - 580 nm very well with relatively little scattering taking place. Almost

17

 

all of the incident red and infrared (580 nm — 3,000 nm) radiance entering deep
pure water was absorbed with negligible scattering (Figure 2-1). Consequently,
pure water appeared blue to our eyes due to the combined effect of molecular
scattering of violet and blue light (< 520 nm) and signiﬁcant absorption of green
and red light (520 nm — 700 nm) in the same water column. Blue colored
waters are typically found in pure mid-ocean water and deep non-turbid inland

water bodies (Jensen, 2000).

In the natural environment, the spectrum shape characteristics of water from
different lakes differed signiﬁcantly depending on dissolved and suspended
constituents within the water. The trophic state of water strongly inﬂuenced

the spectral signatures (Pulliainen et al., 2001).

2.2 Spectral Characteristics of Chlorophyll a

The spectral reﬂectance characteristics of pure water Changed when Chlorophyll
a was introduced. For example, Figure 2-2 depicted the spectral reﬂectance
characteristics of clear water and the same water laden with algae consisting
primarily of chlorophyll a (Han, 1997). Basically, as chlorophyll a
concentration increased in the water column, the amount of energy reﬂected in
the blue and red wavelengths signiﬁcantly decreased but that in green
wavelength increased. Clear water reﬂected approximately 2 percent between
400 nm and 500 nm and dropped gradually to less than 1 percent at

wavelengths beyond 710 nm (Jensen, 2000). Conversely, the algae-laden water

18

presents four pronounced scattering / absorption features of chlorophyll (Figure
2-2; Jensen, 2000; Han, 1997; Rundquist et al., 1995; Gitelson, 1992):

(1) strong chlorophyll a absorption in blue region between 400 nm and 500 nm;
(2) maximum reﬂectance in green wavelengths around 550 nm (green peak)
caused by relatively lower absorption of green light by algae;

(3) strong chlorophyll a absorption in red wavelengths at approximately 675
nm; and

(4) prominent reﬂectance peak between 690 nm — 700 nm caused by an
interaction of algal-cell scattering and a combined effect of minimum pigment
and water absorption. The height of this peak above the baseline (absorption

trough) has been used to accurately measure chlorophyll amount.

4.0 ‘ '-

3.5 " "

3-0 ' ” clear

2 5 ... water algae-laden
' water

   
  
  
 

2.0 -

1.5

Percent Reﬂectance

1.0-

 

 

0.5 - ~
0.0 t t t i 1'
400 soo one zoo soo 900

Wavelength (nm)

Figure 2-2 Percent reﬂectance of Clear and algae-laden water based on in situ

spectroradiometer measurements (after Han, 1997)

19

Dekker et al. (200 1) illustrated the key spectral features in lake water signature
curves. From their measured reﬂectance spectra in Figure 2-3, it was obvious
that the absorption and scattering of the various constituents created a
distinctive reﬂectance spectrum for each of the water samples. In general, there
was little reﬂectance at shorter wavelengths of 400 nm - 500 nm, due to the
combined effects of absorption by colored dissolved organic matter (CDOM),
inanimate particles (tripton) and phytoplankton pigments. A local maximum in
reﬂectance, caused by a local minimum in the combined absorption eﬂects of
CDOM and tripton absorption (which both exponentially decline with increasing
wavelength) and a low phytoplankton pigment absorption efﬁciency, was found
at approximately 550 nm - 580 nm. The local minimum in reﬂectance at 630
nm is caused by the combined effects of cyanophycocyanin absorption and a
ﬁrst shoulder in the absorption of water was noticeable. As this local minimum
became more pronounced, the relative contribution of cyanobacteria to the total
algal components increased. The local reﬂectance peak at 650 nm is due to a
local minimum in absorption by pigments and an increasingly smaller
contribution from CDOM and tripton absorption. A narrow reﬂectance
minimum is centered at 676 nm, which was the in vivo Chlorophyll a maximum
absorption peak. Beyond 680 nm, reﬂectance increased signiﬁcantly to a
maximum of 14 percent at 706 nm. In the studied lake there was a vast
amount of algae identiﬁable by the large reﬂectance at 706 nm, and the major

algal pigment absorption at wavelengths from 400 nm - 680 nm (Figure 2-3).

20

 

   
 

 

 

 

 
 

 

0,09 . TM1 TM2 TM3 TM4
0.08 r F-

. 0.07 - _a

o

EON " f

o :5

'5 0.05 :

0: ....-—
0.04 4 '—
0.03 ,

0.02 .

 

 

 

 

 

 

wavelength (nm)

Figure 2-3 Spectra reﬂectance of Frisian waters, measured in situ in August

1995, with Landsat TM bands 14 superimposed (after Dekker et al., 2001)

Over the full Spectral range, the shape of the spectral signature for water was
broadly determined by the spectral absorption of dissolved organic matter in the
blue, and by the absorption of chlorophyll a and water itself in the red and
near-infrared (Figure 2-4). Attention had concentrated on the wavelength range
from 600 nm to 740 nm, which included the effect of the interaction of the
water absorption (with a peak absorption near 770 nm; Figure 2-4) and the
chlorophyll a absorption (with a peak near 675 nm). This effect produced a
minimum in absorption and thus a peak in reﬂectance, at about 700 nm. As
chlorophyll a increased the peak size near 700 nm, measured with respect to

adjacent wavelengths, increased in a non-linear fashion with shift to longer

21

wavelengths (i.e., Figure 3-2 showed the peak near 720 nm). Many studies
used red /NIR spectral region in their spectral indices (Kallio et al., 2003; Harma
et al., 2001; Kallio et al., 2001; Gitelson et al., 1993). A notable consequence of
the use of these wavelengths was the negligible effects of DOM in the
Chlorophyll a retrieval compared with techniques using blue and green

wavelengths (Green, 1998; Sathyendranath and Platt, 1989; Tassan, 1988).

 

 

 

 

400 I T — I T E I
3.5 ' , .
3.0 - .
_ Reﬂectance -
3 2.5 . g
E , * 3:
1'5 2.0 ~ . u g.
:0 ‘_ a
0 ”I .0
0‘ 1.5 ~ , a:
1.0 r \
- \ ‘ \
0,5 . ‘— ' \ . E Total absorption 4} ’1’ / Water .
‘ ,u-v.
"me \ ‘ x ‘ " ' “"7. a J '\ Chlorophyll
0'0 —-:;-:;-'—'4'.—'1CT-rrf.::‘_-_ A -------- 1 '——s—1-L
450 500 550 600 650 700 750

Wavelength (nm)

Figure 2-4 The lower three curves represent the absorption of water and its
constituents: the dashed curve the absorption of water, the ( ----- ) absorption of
60 mg/l of chlorophyll, the dotted curve is the DOM absorption (0.14 / m @440
nm) and the solid represents the sum of the water, chlorophyll and DOM
absorption. The upper hatched curve is an observed reﬂectance spectrum over

the lake (after Green 1998)

22

Figure 2-5 presented an example reﬂectance spectrum of a lake with high

chlorophyll a content and one with very little chlorophyll (Green, 1998).

 

 

 

 

6 . . .
5 ~. '1
o ' a ,.
8 4 l- , r ' \“ ,.-, :l l.
0 ~ .- r.
c . .1.
e :
3a 3 _ /"_“"\ ‘ I
a .
E 2 / \
0 / K
8 / \
o ,--f W
a 1 \
\\/’J'\/ .
0
450 500 550 600 650 700 750
Wavelength (nm)

Figure 2-5 Example reﬂectance from a high Chlorophyll content lake (85 mg/ l —
dash line) and one from a low chlorophyll content lake (3 mg /1; TSM 3.5 mg/l —

solid line) (after Green, 1998)

With continued increase in Chlorophyll a content, the reﬂectance peak of water
shifted toward longer wavelengths (Figure 2-6; Gitelson, 1992). Thus, the
indicator of the chlorophyll a concentration of the water column was related to
the shape of the reﬂectance curve in this region, and not simply the peak height

(Green, 1998).

23

 

0.025 . . .

High

   

0.020 -
Moderate ' ‘

  

)

‘7‘ 0.015

(sr

9)
x“ 0.010

 

 

 

500 600 700 800
Wavelength (nm)

Figure 2-6 Reﬂectance spectra of the water bodies studied in Dall’Olmo and
Gitelson (2005). Some examples are highlighted: curve “Low” (Chl—a = 11 mg/l;
TSS = 5 mg/l), curve “High” (Chl—a = 89 mg/l; TSS = 21 mg/l), and curve

“Moderate” (Chl—a = 24 mg/l; TSS = 55 mg/l)

2.3 Spectral Characteristics of Complicating Factors (TSS and DOC)

A general problem concerning remote sensing of all waters was that the
reﬂectance signals were very weak and often also wavelength-speciﬁc (Ostlund
et al., 2001). The radiance leaving water was a function of solar intensity and
angle and the optical properties of the water attenuation, absorption, and
scattering. While solar incoming radiation varied in time, optical properties

varied in relation to the concentrations of optically active constituents, e.g.,

24

phytoplankton pigments, particulate substances, and aquatic humus (Pepe et
al., 2001). The water column contained a mixture of dissolved organics,
inorganic suspended sediments, and chlorophyll a, which masked and
interfered with the spectral identiﬁcation of the Chlorophyll a alone, especially
since the inorganic suspended sediment is a much brighter target than the

chlorophyll.

In the deep ocean environment, dissolved and suspended matter seldom played
important roles and often only one species of algae dominated (Ostlund et al.,
2001). Remote sensing spectral signature of chlorophyll a in the 400 nm — 550
nm region was used to estimate phytoplankton in the oceans (O’Reilly et al.,
1998). Such chlorophyll retrieval algorithms were derived solely from
regression techniques and ignore the speciﬁc absorption and scattering
properties of the water body being remotely sensed. The success of these
algorithms was largely a consequence of the optically simple characteristics of
mid-ocean and many near-coastal waters. Even though the optical properties
of deep off shore waters were primarily a function of phytoplankton
concentration, coastal and inland waters represented a much more complex
optical environment. Water bodies strongly inﬂuenced by land masses
displayed higher orders of optical complexity. This was a result of an increased
number of optically-active components co-existing within the water column, as
well as greater ranges in the variations of the concentrations of these aquatic
components. Due to the optically competitive compositions of coastal,

estuarine, lake, and river water masses, mathematical models developed mainly

25

on marine waters might not adequately describe the same variables of interest
in lakes. The same wavelength region could not be applied in lakes, mainly
because absorption by colored dissolved organic matter hid the spectral
signature of chlorophyll a at these wavelengths. Instead, the interpretation of
chlorophyll a from remote sensing data in lakes was usually based on the 660
nm — 715 nm spectral region (Kallio et al., 2003; Kallio et al., 2001; Gitelson et
al., 1993). Harma et al. (2001) suggested that humic lakes needed to be
separated and interpreted using speciﬁc models developed for these types of
water bodies. Suspended sediment in lakes also had a signiﬁcant effect on
chlorophyll a signature extraction because its reﬂectance was much higher;
therefore, it masked out the chlorophyll a absorption feature. Because of the
same reason, oligotrophic lakes might need to be interpreted separately.
Oligotrophic waters had week signatures that could be interfered easily by other

substances in the water.

Inland water color was related to the types and amounts of these substances in
the water column which interacted with light absorbing or scattering it (Flink et
al., 2001). The optical properties of natural bodies of water were inﬂuenced by
three main components, which could vary independently from each other
(Figure 2-7). These are (IOCCG Report Number 3, 2000):

(1) Phytoplankton - includes phytoplankton and other microscopic free-ﬂoating
organisms found in the illuminated surface layers of water. They were the
living organisms that form the base of the aquatic food web, and were an

important component of the global carbon cycle. The concentration of the main

26

phytoplankton pigment, chlorophyll a, was often taken as an index of
phytoplankton biomass;

(2) Suspended inorganic material - included all inorganic particulate material
that was not included in the phytoplankton component; and

(3) Yellow substances — included the colored, dissolved organic substances, and
also “detrital” particulate material, for example from the degradation of

phytoplankton cells and other organic particles.

 

Dissolved Organic Matter

Absorption
A

  

Pigments

 
 

 

 

 

0 l l l
400 500 600 700 800 900

Wavelength (nm)

Figure 2—7 Absorption spectral properties of optically active constituents in

water (after Dall’Olmo, 2005)

Extracting quantitative information about the constituents of interest from the

remotely sensed data from natural water that contained a mixture of materials

27

was one of the greatest challenges in remote sensing (Goodin et al., 1993). To
begin with, it would be instructive to look at the effect that each of these
constituents had on the spectral reﬂectance characteristics of a water column

(Jensen, 2000).

2.3.1 Spectral Characteristics of Suspended Sediments

Sediment came from a variety of sources, such as agricultural cropland erosion
and urban surface runoff. The particles ranged from ﬁne clay particles (3 pm —
4 pm in diameter), to silt (5 pm - 40 pm), to ﬁne-grain sand (41 pm - 130 um),
and coarse grain sand (131 um - 250 pm). Most of the suspended mineral
sediment was concentrated in the inland and nearshore water bodies (Bukata et
al., 1995). Thus, suspended mineral concentration was usually of no
signiﬁcance to deep ocean remote sensing studies. On the other hand, inland
water bodies might carry a signiﬁcant load of suspended sediment that could
dramatically impact the spectral reﬂectance characteristics of the water bodies

(Jensen, 2000; Nanu, 1993).

For several reasons, it was important to monitor the type, amount, and spatial
distribution of suspended minerals in inland water bodies. First, sediment
affected water quality and its suitability for drinking, recreation, and industrial
purposes. Second, sediment served as a carrier and storage agent of pesticides,
absorbed phosphorus, nitrogen, and organic compounds and could be an
indicator of pollution. Third, photosynthesis by phytoplankton and submerged

aquatic vegetation could be signiﬁcantly impacted as suspended sediments

28

impede the transmission of solar radiation in the water column. These
phytoplankton and aquatic vegetation played a vital role in the food Chain of the

aquatic ecosystem (Jensen, 2000).

Fortunately, remote sensing had been used to monitor the suspended mineral
concentrations in water bodies. The in situ measurements of suspended
mineral concentrations were usually required to derive a quantitative
relationship with the remote sensor data. When collecting samples, the remote
sensor data and the in situ suspended sediment measurements should be
collected on days that have little wind because wind-roughened surface water

created specula reﬂections (Jensen, 2000; Han and Rundquist, 1998).

When both suspended mineral sediment and Chlorophyll were present in the
water body at the same time, a dramatically different spectral response was
produced. For example, Figure 2-8 illustrated the spectral response of water as
red loam sediment concentrations from 0 - 500 mg/l were added to water that
contained algae. For algae laden water, the green peak reﬂectance shifted from

547 nm at 0 mg/l to 596 nm at 500 mg/l (Jensen, 2000).

29

25 "’ Algae-Laden Water with Various

Suspended Sediment Concentrations

2° " 500 mgll

Percent Reﬂectance

 

 

 

400 500 600 700 800 900
Wavelength (nm)

Figure 2-8 Percent reﬂectance of algae-laden water with various concentrations

of suspended sediment ranging from 0 — 500 mg/l (after l-Ian, 1997 )

Figure 2-9 depicted the spectral reﬂectance of clear water and water with
varying suspended sediment concentrations of two different type of soils; clayey
and silty. For a deep clear water, spectral reﬂectance dropped continuously
after approximately 580 nm due to increased absorption in the water column.
Increased in suspended particulates (either inorganic or organic) were related to
increase in overall brightness (Shaﬁque et al., 2001). A water body with

suspended sediment in it would generally appear brighter in imagery than a

30

water body without any suspended sediment. The clayey soil (Figure 2-9b) had
approximately 10 percent lower reﬂectance at all wavelengths than the light-
colored silty soil because it contained more organic matter and was darker in
color (Figure 2-9a). If the suspended particulates were organic in nature, the
reﬂectance data indicated a relative increased at about 705 nm (Shaﬁque et al.,
2001). Reﬂectance increased in the 580 nm — 690 nm region and in the near-
infrared region as more minerals sediments were added to the water bodies.
Thus, the peak reﬂectance Shifted toward longer wavelengths in the visible
region as suspended sediments increased (Jensen, 2000). These results
suggested that: (1) the type of suspended sediments (soil) in waters might be
assessed using the visible wavelength range of 580 nm — 690 nm; and (2) the
amount of suspended minerals in waters where suspended minerals were the
predominant constituent might be estimated using the near-infrared

wavelength range of 714 nm — 880 nm.

31

5.0 -

4-5 T Clayey soll
4.0
3.5
3.0

2.5

 

2.0
1 .5
1 .0

Percent Reﬂectance

 

0.6

l I n I 1 l 1 1 l j

a. 0'0 I l I 1 I I I I I I
400450500660600660700750800860900

Wavelength (nm)

14 -—

v I

12 ._ Sllty soil

10 ..

Percent Reﬂectance
m

 

 

 

 

 

a
6 T” ,1";
4 dbl ’IH
2 clearwater
b. 0 I If + 41 I I I I l I
MMWSGOMSSOTOOTSOSMSSOQW

Wavelength (nm)
Figure 2-9 (a) In situ spectral reﬂectance measurements of Clear water and clear
water with various levels of clayey soil suspended sediment concentrations from
0 - 1,000 mg/ 1. (b) In situ spectral reﬂectance measurements of clear water and
clear water with various levels of silty soil suspended sediment concentrations

(after Lodhi et al., 1997)

32

2.3.2 Spectral Characteristics of Dissolved Organic Carbon

A group of lakes that have high humic concentration made the interpretation of
present multispectral remote sensing data practically impossible. The remotely
measured signal was very low from these lakes due to strong absorption caused
by high concentrations of colored dissolved organic matter (DOM), and low

concentrations of particles causing scattering (Kutser et al., 2001).

The effects of dissolved organic compounds on the absorption of light energy
were marked (Wetzel, 2001). Sunlight penetrates into the water column at
certain photic depth (the vertical distance from the water surface to the 1
percent subsurface irradiance level). Within this depth, phytoplankton
consumed nutrients and converted them into organic matter via
photosynthesis. The process was called primary production. Zooplankton also
consumed the phytoplankton and created organic matter. Bacterioplankton
decomposed these organic material. All the conversion produced dissolved
organic matter (DOM) in the water bodies. The more productive the
phytoplankton, the greater the released of dissolved organic matter. In certain
instances, there might be sufﬁcient dissolved organic matter in the water to
reduce the penetration of light in the water column (Jensen, 2000; Bukata et

al., 1995).

These dissolved humic substances were called yellow substance and could (1)
impact the absorption and scattering of light in the water column, and (2)

change the color of the water (Jensen, 2000). In comparison to pure water, lake

33

water with increasing concentrations of dissolved organic compounds,
particularly humic acids, not only drastically reduced the transmission of light
but shifted absorption selectively. Clear waters had a very high absorption in
red and infrared wavelengths, but a relatively little absorption in UV
wavelengths (300 nm - 400 nm; Figure 2-1). Very low concentrations of
dissolved organic compounds increased UV absorption greatly. Most of the
irradiance in UV, blue, and green wavelengths were essentially absorbed in
much less than a depth of 1 m in lakes highly stained with humic compounds

(Wetzel, 2001).

Phytoplankton was not the only source of dissolved organic matter. For
example, the brownish—yellow color of the water in many rivers in the northern
United States was due to the high concentrations of tannin from the eastern
hemlock (Tsuga Canadensis) and various other species of trees and plants
grown in bogs in these areas. These tannins potentially create problems with

remote sensing of inland water bodies (Jensen, 2000).

Information about the phytoplankton pigments from remotely sensed data of
natural inland water body that was effected by dissolved organic matter (DOM)
was often more difﬁcult to unwind. Figure 2- 10 showed the spectra of natural
waters that were dominated by different concentration of chlorophyll a and

dissolved organic matter.

34

Moderate Chl a High Chl a

“ID 450 500 550 601) 650 700 750 800

Wavelength, nm

   

Wavelength nm

Secchi disk at 50 cm Secchi disk at 25 cm

HthOM

    

.4

we will mm 5511 mm lxm’ roll Tm’l'snll

\‘z’avclelltllh lllll

Secchi disk at 25 cm
Figure 2- 10 Examples of water reﬂectance for different chlorophyll a

concentrations in natural waters (after Dall’Olmo, 2005)

2.4 Spectral Characteristics of Algae

Algae were extremely diverse, and many exhibited a very wide tolerance to
environmental conditions found under natural limnological situations.
Nonetheless, certain Characteristic phytoplanktonic associations occurred
repeatedly in lakes of increasing nutrient enrichment. Some of the commonly
observed major associations were described in Table 2-2 based on the detailed

discussion of Hutchinson (1967) as cited in Wetzel (2001). However, the wide

35

spectrum of intergradations was often observed, and species composition shifts

occurred seasonally from one type to another, especially among more

productive waters. Nevertheless, such Characterizations yielded insight into

regulating environmental factors, thus they were useful from the standpoint of

general correlations between qualitative and quantitative abundance of the

algae and available nutrients.

Table 2-2 Characteristics of common major associations of the phytoplankton

in relation to increasing lake fertility (after Hutchinson, 1967 as cited in Wetzel,

 

 

 

 

 

 

 

 

 

200 1)
General lake Water characteristics Dominant algae Other
trophy commonly
occurring algae
Oligotrophic Slightly acidic; very low Desmids Sphaerocystis,
salinity Staurodesmus, Gloeocystis,
Staurastrum Rhizosolenia,
Tabellaria
Oligotrophic Neutral to slightly Diatoms, Some
alkaline; nutrient-poor especially Asterionella spp.,
lakes Cyclotella and some Melosira
Tabellaria spp., Dinobryon
Oligotrophic Neutral to slightly Chrysophycean Other
alkaline; nutrient-poor algae, especially chrysophyceans,
lakes or more Dinobryon, some (e.g., Synura and
productive lakes at Mallomonas Uroglena); diatom
seasons of nutrient Tabellaria
reduction
Oligotrophic Neutral to slightly Chlorococcal Oligotrophic
alkaline; nutrient-poor Oocystis or diatoms
lakes Chrysophycean
Botryoooccus

 

 

 

36

 

 

Oligotrophic

Neutral to slightly
alkaline; generally
nutrient poor; common

in shallow Arctic lakes

Dinoﬂagellates,
especially some
Peridinium and

Ceratium spp.

Small
Chrysophytes,
cryptophytes,
and diatoms

 

Mesotrophic

or eutrophic

Neutral to slightly
alkaline; annual
dominants or in
eutrophic lakes at

certain seasons

Dinoﬂagellates,
some Pen'dinium

and Ceratium

spp.

Glenodinium and

many other algae

 

 

 

 

 

 

Eutrophic Usually alkaline lakes Diatoms much of Many other
with nutrient year, especially algae, especially
enrichment Asterionella spp., greens and

Fragilaria cyanobacteria

crotonensis, during warmer

Synedra, periods of year;

Stephanodiscus, desmids if

and Melosira dissolved organic

granulata matter is fairly
high

Eutrophic Usually alkaline; Cyanobacteria, Other
nutrient enriched; especially cyanobacteria;
common in warmer Anacystis euglenophytes if
periods of temperate (=Microcystis), organically
takes or perennially in Aphanizomenon, enriched or
enriched tropical lakes Anabaena polluted

 

2.4.1 Spectral Characteristics of Algal Pigments

The amount of chlorophyll a was considered a reasonable representative for the

organic component of optically complex natural waters (Bukata et al., 1995). It

was a good indicator of the quality of lake water as it correlated well with the

total productivity of a lake and; therefore, with the nutrient load and overall

condition of the lake (Koponen et al., 2001).

37

 

Algae contained colored pigments - the chlorophylls, carotenoids, and
biliproteins, which gave them characteristic spectral features. The pigments
were used in the photosynthetic process. All algae and cyanobacteria contained
the photosynthetically active pignent chlorophyll a as it was the primary
photosynthetic pigment of all oxygen-evolving photosynthetic organisms.
Chlorophyll a absorbed light in the blue (near 430 nm) and red (660 nm —- 665
nm) parts of the spectrum, thus giving the substance itself a green color
(Wetzel, 2001). Bluegreen algae, or cyanobacteria had other important
phytoplankton photosynthesizing agents: carotenoids, and phycobilins (primary
phycocyanin in freshwater, and phycoerythrin in marine environments).
Phycocyanin absorbed more toward the yellow and green part of the spectrum

giving the pigment a blue shade.

Fortunately, different genus of phytoplankton appeared as different colors to
sensitive remote sensors because they had different types and concentrations of
pigments (Figure 2- 11). The wavelengths of pigment absorptions could be used
together with nearby wavelengths, which were less affected by the pigment
absorption, to detect the presence of the pigment. Often indices, which were
used for quantiﬁcation, were constructed from reﬂectance at pigment
absorption wavelengths (Flink et al., 2001). Thus, the amount and general type
of phytoplankton might be estimated by recording the optical spectra of the
water body, and information about the health and chemistry of the water body
could be assessed. Changes of optical water condition over time could be

monitored by comparing images taken at different times (Jensen, 2000).

38

0.16

 

0.14 -
0.12 «
0.101

0.08 4

Reﬂectance

0.06 .

 

 

0.04 Green: Chlorophyta
Diatoms: Chrysophyta

 

0.024
I I Blue-green: Cyanophyta

0.00 . . . A . . .
350 400 450 500 550 600 650 700 750
Wavelength (nm)

 

 

 

 

 

 

Figure 2-11 Maximum and minimum absorption values in the reﬂectance

spectra of pure algal culture (after Shaﬁque et al., 2001)

In addition, other pignent peaks might also be apparent in the reﬂectance
curves. The measurement of phycocyanin spectral Characteristic done in a
controlled laboratory environment resulted in a maximum absorption at 630
nm. The absorption feature should also be observable as a trough in the in situ
reﬂectance curve. In Figure 2- 12, the effect of this additional pignent
absorption was demonstrated. The two lakes, with similar chlorophyll
concentrations, differed in that one was predominately by green algae while the
other was predominate by cyanobacteria. The relative size of the chlorophyll a/
water peak near 700 nm was the same; the extra absorption in this region,

most notably at 630 nm, was due to the phycocyanin content. This showed the

39

potential of the remote sensing as “cyanobacteria detectors” which could beneﬁt
environmental monitoring because some of these blue-green algae are toxic

(Green, 1998).

0.08 . 1 . . .

 

0.07 ~ -
' ChlorophylltWater

 
   

0.06
0.05

0.04

4
Phycocyanin

Percentage reﬂectance

     
 

°-°3 DOMtChlorophylls

I

0-02 Chlorophyll a ‘

 

 

0.01 ‘ 1 1 1 ‘
450 500 550 600 650 700 750
Wavelength (nm)

 

Figure 2- 12 Two reﬂectance curves from lakes with similar high levels of
chlorophyll — but one contains phycocyanin, indicating a cyanobacteria (the
phycocyanin absorption peak at 630 nm is expressed as a trough in
reﬂectance), while the other exhibits the properties of green algae (after Green,

1998)

40

2.4.2 Cyanobacteria (Bluegreen Algae) Detection (Potential Toxic Algal
Detection)

Subramaniam et al. (1999) parameterized a standard remote-sensing
reﬂectance model using measured values of Trichodesmium’s inherent optical
properties, namely the spectral dependence of the chlorophyll-speciﬁc optical
absorption cross-sections and the spectral dependence of the chlorophyll-
speciﬁc backseatter cross-sections. Sea truth and data from the Advanced Very
High Resolution Radiometer (AVHRR) were used to map a 300,000 km2

Trichodesmium bloom oﬁ' the Somali Coast in May 1995.

In biological oceanography, changes in optical properties had been used to infer
upper ocean chlorophyll a concentrations, which could, in turn, be related to
primary productivity. It had been very difficult, however, to quantify the
temporal and spatial extent of new production in the oceans, let alone the
contribution of N2 ﬁxation to that ﬂux. The nonheterocystous, colonial
cyanobacterium, Trichodesmium spp., was responsible for most of the N2
ﬁxation in the open oceans (Capone et al., 1997). Hence, a remote-sensing
algorithm capable of distinguishing these organisms from all other
phytoplankton would be of enormous value in constraining estimates of N2

ﬁxation in the world’s oceans.
Together, the optical properties and physiological behavior of Trichodesmium

potentially provided a basis for developing algorithms capable of uniquely

identifying and quantifying their distributions based on remotely sensed

41

information (Subramaniam and Carpenter, 1994). The research analyzed the
backseatter properties, in conjunction with absorption properties, of
Trichodesmium and parameterized a remote sensing reﬂectance model which
derive satellite observations of ocean color in the visible and near-infrared

(Subramaniam et al., 1999).

The backseatter coefficient for Trichodesmium was wavelength dependent.
“Backseatter” peaks centered between 550 nm and 579 nm and at 640 nm were
a consequence of phycoerythrin and allophycocyanin ﬂuorescence, respectively,
rather than true elastic backseatter. Typically, chlorophyll a ﬂuorescence
emission overwhelmed the red absorption band, such that the backseatter
spectrum revealed a peak in this region (Ahn et al., 1992). Remote sensing
reﬂectance spectra of phytoplankton blooms typically contained a “green peak,”
centered around 575 nm. The peak was a consequence of an absorption
minimum in that portion of the spectrum in most phytoplankton taxa, coupled
with a sharp increased in the water absorption spectrum. As a consequence of
both the intrinsically high backscatter and phycoerythrin ﬂuorescence, the
remote-sensing reﬂectance was extremely high around 575 nm for blooms of
Trichodesmium The reﬂectance was approximately ﬁvefold higher than that for
Synechococcus or typical phytoplankton at high densities of Chlorophyll

biomass (10 mg Chl m-3; Figure 2-13).

42

Reﬂectance model for Chl=1o mgim3

 

Seawater

 

-- ----- Trichodesmium colonies
-— Average phytoplankton
-—-- Synechococcus
" -—- Colonies - modeled backseatter

Percent reﬂectance

 

 

 

 

400 450 500 550 600 650 700 750
Wavelength (nm)

Figure 2- 13 Reﬂectance model for chlorophyll = 10 mg/ m3 (after Subramaniam

et al., 1999)

Although remote sensing of phytoplankton in the ocean was primarily based on
water-leaving radiance in the visible, the colonial and buoyancy behaviors of
gas vacuolated cyanobacteria offered opportunities to exploit the red and near-
infrared regions of the spectrum as well. In the speciﬁc case of Trichodesmium,
the virtual absence of solar—induced ﬂuorescence of Chlorophyll a, the optical
brightness resulting from gas vacuoles, and the removal of absorption by water
in the near-infrared permitted the development of simple two-Channel
reﬂectance difference indices from AVHRR data to map distributions of surface

blooms. While this approach was inferior to that utilizing visible color

43

information, AVHRR data could be used to map, retrospectively, the
distribution of surface blooms where there was simultaneous in situ information

identifying the bloom organism.

2.5 Most Sensitive Spectral Bands

Spectral wavebands that were successfully used in the reviewed researches
were summarized in Table 2-3. Several studies have shown spectral bands that
were found link to water quality variables were in common spectral regions.
However, a slight Shift in wavelengths might occur due to the different nature of
research (e.g., controlled laboratory setting versus natural condition) and / or
different condition of water (e.g., high DOC and TSS). When Chlorophyll
increased, the near-infrared peak often shifted to the longer wavelength as well.
The spectral bands that were found to be sigriﬁcant from previous studies were
used as a general guide in the optimal spectral band identiﬁcation process in

later sections.

44

Table 2-3 Summary of wavebands from literature review

 

 

 

 

 

 

 

 

 

 

 

 

 

Parameter Wavelength (nm) Optical Properties Reference
Chlorophyll a 440-450 Distinct absorbance Shaﬁque et al., 2001
in blue
575 “Green peak” from Subramanium et al.,
min algae abs + 1999
sharp increase water
abs
650 Min absorbance Dekker et al., 1992
670-680 Distinct max Kallio et al., 2003
absorbance in red Shaﬁque et al., 2001
Dekker et al., 1992
705 Local peak due to Kallio et al., 2003
max abs at 670 + Shaﬁque et al., 2001
growing water
absorption
SS 705 Type of sediment Shaﬁque et al., 2001
580-690 Amount of sediment
Cyanobacteria 550-579 Phycoerythrin Subramanium et al.,
1999
625 Phycocyanin Jupp et al., 1994
630 Local minimum Dekker et al., 1992
reﬂectance
640 Allophycocyanin Subramanium et al.,

 

 

 

1999

 

45

 

2.6 Spectral Band Selection Methods

Several analytical methods were used to determine wavebands that best
described biOphysical variables. The methods that were used most frequently
in the remote sensing of water studies were principle component analysis and

the spectral derivative analysis.

2.6.1 Principle Component Analysis (PCA)

In order to create an index suitable for water quality variable mapping, the
wavelengths where substance-speciﬁc features existed must be identiﬁed. Such
an index should exhibit high variance, which will reduce the inﬂuence of noise
in subsequent regression analysis. The amount of data collected by
hyperspectral spectrometers is immense and it is, therefore, often necessary to
remove redundancy in the dataset. Principal Component Analysis (PCA) can be
used to determine the inherent dimensionality of the dataset. Flink et al. (2001)
used reﬂectance values derived from CASI bands 35-288. Figure 2- 14 showed
the standard deviation as a function of wavelength. Low variation was

associated with the chlorophyll a absorption wavelengths.

46

x 104 Lake Erken

 

O
.q
«I
ed
a
~¢

U
a
e
l
a
e
0
a
P
e
s
I
0
e
e
I
e
e
I
O
I
e
a
0
.
.
0
e
.
e
e
a
e
O
a
u
a
e
e
o
a
O
u
e
I
I
n
I
I
I
-
a
I
a
e
e
e
a
a
U
-
s
e
.
e
e
u
.
O
a
l

u
e
a
e
a
e
n
u
e
s
A
I
s
o
u
-
e
u
-
e
s
I
e
n
on
e
e
a
I
0
u
o
a
e
I
n
c
o
c
e
.
n
a
a
I
a
.

 

‘
I
I

eeeee

(a)
I
A

Standarddevlatlonol reflectance (mitten)
0|

 

 

600 700 800 900 1000
Wavelength (nm)

 

§N
g.

Figure 2- 14 Standard deviation of the data at different wavelengths for Lake

Erken (after Flink et al., 2001)

A PCA was performed on these correlation matrix, and constructed new
variables called principal components (PCs) as linear combinations of the
original variables. PCA also concentrates the majority of the variance of the
dataset into a few new non-correlated components, thus reducing redundancy.
In this study, more than 96 percent of the total variance of the data was
contained in only three PCs (i.e., any spectra in the lakes could be fairly well
approximated by a weighted sum of only three PCS). Flink et al. (200 1)
reconstructed all 100 lake spectra from their three ﬁrst PCs and then
subtracted them from the corresponding original spectra to derive the

approximation error. Flink et al. (2001) cautioned that PCA should be used

47

with care. The PCs should be identiﬁed with physical phenomena only when
there are obvious connections, as in previous work, e.g., Doerffer et al. (1989),
where they interpreted PCs calculated from TM data as physical variables based
on factor analysis. Their identiﬁed factors were temperature, suspended matter
and aerosol effect. These factors were likely to have been uncorrelated, just as
are the PCS. Other variables such as chlorophyll a and suspended particulate
matter (SPM) (and even suspended inorganic material) were often highly
correlated. One should keep in mind that a PCA gave uncorrelated variables as
its result; therefore, the PCs should be interpreted with care. Measurements of
variables were very seldom uncorrelated in reality, either because the variables
were inherently correlated or because the method used for measuring them

yielded a correlation (Flink et al., 2001).

Many single band algorithms performed well in Lake Erken data, e.g., Band 550
nm gives an R2 of 0.94 between lake spectra and chlorophyll a. However, at the
corresponding wavelength (550 nm) it was difﬁcult to determine if the variations
were caused by chlorophyll a or by some other substance in the water. Thus,
mapping by means of one single band was possible, as long as chlorophyll a
constitutes almost the entire amount of material in the lake. A physically
sound model, however, should include bands at chlorophyll-speciﬁc features
(e.g., a ratio between bands at 708 nm and 680 nm), which is probably the

most widespread way of measuring Chlorophyll a.

48

2.6.2 Spectral Derivative Analysis

Spectroscopic derivatives are obtained by taking the difference between the
reﬂectance of two bands and dividing that value by the difference between the
wavelengths separating the two bands (Philpot, 1991). When the two bands
used in the calculation are adjacent to one another, the result is the ﬁrst
derivative (Shaﬁque et al., 2001). Pepe et al. (2001) developed the chlorophyll a
model using derivative method. The model based on the higher sensitivity of
reﬂectance lst derivative spectra to concentrations of optically active substances
than the radiance reﬂectance spectra themselves. Correlation analyses were
carried out with a ﬁrst derivative at each hand-held spectroradiometer band
pass. Spectral band at 676 nm proved to be the best-correlating wavelength in
most cases, corresponding with a peak in Chlorophyll a absorption (Han and
Rundquist, 1997). On the basis of those results, 676 nm as a sole wavelength
was used to evaluate the applicability of the ﬁrst derivative spectra model to
every lake condition. Considering the results of the ﬁrst-derivative model over
the complete acquired spectrum (380 nm — 780 nm), the result showed that
higher correlation values were dependent on wavelength with respect to
chlorophyll contents and sampling depths. In any case 676 nm proved to be
the most often correlated wavelength. The ﬁrst derivative of reﬂectance at 676
nm was suﬂiciently reliable only when average chlorophyll a contents higher
than 2 ug/l; and when the Cyanophyceae (bluegreen algae) presence less than
20 percent in the phytoplankton biomass. The accuracy of let derivative
method and the near-infrared / red reﬂectance ratio (NIR/red) were tested in the

study. The NIR/ red model was based on the contrast between a local

49

reﬂectance peak feature at approximately 705 nm due to a minimum
absorption by the pignent and the water, and a local reﬂectance minimum
feature at approximately 670 nm due to the absorption maximum of Chlorophyll
a. The results showed that NIR/red model results were less satisfactory than

the ﬁrst derivative one.

Rencz (1999), and Huguenin and Jones (1986) examined a variety of higher-
order derivatives of spectra in an effort to identify the location of individual
absorption regions. While assuming that each absorption was symmetric
around its band center, the method did not require that absorptions have a
speciﬁc Shape. Band centers were identiﬁed where the second derivative of the
spectrum was negative, the fourth derivative was positive, and the ﬁfth
derivative was zero. Like any derivative analysis, this method was highly
sensitive to noise. Therefore, the Huguenin and Jones (1986) approach was
critically dependent on an intelligent smoothing algorithm. Nonetheless, their
approach was capable of resolving overlapping band centers separated by as
little as 0.1 to 1.0 of the full width at half maximum (assuming Gaussian
shaped absorptions). Although derivative technique was sensitive to noise, Tsai
and Philpot (1998) concluded that an algorithm for derivative analysis of
hyperspectal data was a tool that treated hyperspectal data as truly spectrally
continuous data. Moreover, the approach could be used with no need to

assume that the data were generated in highly controlled environments.

50

2.7 Spectral Indicators

In order to assess water quality by using remote sensing data, the relationship
between spectral data and water quality variables such as Secchi depth,
Chlorophyll a, and suspended sediment need to be identiﬁed and quantiﬁed. A
variety of statistical methods such as band rationing and regression techniques
have been used to derive these water quality modeling. Wavebands that were
identiﬁed to be important for water quality indicators were usually set as the
dependant variables while water quality parameters were independent

variables.

2.7.1 Spectral Band Ratio Method

Certain band ratios , could be used successfully for chlorophyll mapping in
inland waters (Koponen et al., 2001; Gitelson et al., 1993). Dekker (1993) and
Gitelson et al. (1993) found that the optimum ratio of spectral radiance or
reﬂectance at two wavelengths (Ax) and (Ay) is achieved where (Ax) was in the
range from approximately 680 nm - 710 nm (corresponding to the chlorophyll a
ﬂuorescence peak and volume scattering from particulate matter) and (Ay) was
at approximately 665 nm - 680 nm (the region of the chlorophyll a absorption

maximum).

 

(Eq 2-1)

“ o l Lu“)
Chla(pg/l) —- a + a [Lay-)1

Pulliainen et al. (2001) employed the wavelength ranges suggested by Gitelson

et al. (1993) in their chlorophyll a retrieval algorithms. The optimum channel

51

ratio was selected empirically using a training data set to determine the highest
correlation with Chlorophyll a concentration. A linear regression model
employing the Channel ratio Iﬂ02nm / Leesmm yielded a maximum value of R2 =
0.94. However, when the predicting waters were affected by various substrates
besides chlorophyll (e.g., humic and high suspended sediment), application of
Eq 2-1 with remote sensing data may encounter some problems especially if
other parameters in addition to chlorophyll a affect the ratio L01.) / L(Ay). This
method might require the data set to be pre-classiﬁed into different sub-groups,
e.g., based on the shape of the measured radiance spectra, in order to increase

the chlorophyll a estimation accuracy.

Koponen et al. (2001) also used the AISA data to measure Chlorophyll a
concentration in lakes. The study found the 702 nm/ 673 nm band ratio
produced the best result. Their results conﬁrmed that an airborne
spectrometer was a useful tool for chlorophyll a monitoring in lakes. The result
corresponded well to the ﬁnding of other authors (e.g., Dekker et al., 1992 or

Gitelson et al., 1993).

George and Malthus (200 1) used an array of wavelength-speciﬁc correlation
coefﬁcients to determine the ‘single band.’ Low coefficient values indicated that
the radiance values at these wavelengths were not inﬂuenced by the presence of
phytoplankton. High positive or negative values indicated that the radiance
values were strongly inﬂuenced by the concentration of phytoplankton. The

strongest correlation between the two variables was recorded in the green and

52

red portions of the spectrum where the ‘r‘ values were positive and reached a
maximum value of 0.86 (P < 0.05). Then, the correlation between all possible
combinations and the measured concentration of chlorophyll a was calculated.
The results suggested that the most effective multi-band algorithm would
contrast the amount of ‘green’ light reﬂected with that absorbed at the ‘blue’
end of the spectrum. The ratio of measurements centered at 550 nm and 440
nm (the ratio identical to the blue /green ratio suggested by Clarke et al., 1970)
was found to perform best for chlorophyll prediction, and the ratio of
measurements at 685 nm and 745 nm (rather similar to the long-waveband
ratio recommended by Dekker, 1993) performed well for waters containing high

concentrations of dissolved organic matter.

2.7.2 Statistical Method

Multivariate statistics have been developed and applied successfully for spatial
or spectral analysis of remote sensing data, usually derived from established
methods in multivariate statistics (Tsi and Philpot, 1998; Richards, 1993; Duda
and Hart, 1973). Giardino et al. (2001) adopted several statistical techniques to
examine the relationship between in situ measured parameters (i.e., Secchi disk
and chlorophyll a) and remote sensing reﬂectance values ﬁ'om the Landsat
Thematic Mapper sensor. These models included linear, exponential and log
transformations. A few previous studies used nonlinear power models (y=aXb)
to address the curvilinear behavior of this relationship (e.g., Cox et al., 1998;
Lathrop, 1992). Although a power model provided a strong correlation,

residuals from it were not normally distributed. In contrast, the semilog

53

equation used by Kloiber et al. (2002) met the model assumptions. A similar
result had been found by Pattiaratchi et al. (1994). Regression models were
used to determine the relationship between the difference between reﬂectance
values difference between TM bands 1 and 3 (TM 1 — TM3), and the ratio
between TM bands 1 and 2 (TMl /TM2). These models allowed the surface
distribution of chlorophyll a concentrations and Secchi disk depths to be
determined with good conﬁdence (the coefﬁcients of determination were 0.99

and 0.85, respectively; Giardino et al., 2001).

Kloiber et al. (2002) took a further step to develop a standard model that used a
consistent equation form for using satellite remote sensing data to estimate key
variables related to lake management issues, such as trophic state condition
and water clarity. Rather than using regression equations where the
independent variables were different for each image, the feasibility of using a
consistent equation form to relate ground observations and satellite data was
examined. A Pearson correlation matrix was developed to examine the relative
strength of correlation between Secchi disk transparency depth (SD) and
various Landsat TM bands and band ratios. Results indicated that regressed
log-transformed SD versus the TM 1 /TM3 ratio plus TM 1 (TM 1/TM3 + TM 1)
provided strong predictive relationships for multiple images over a 25-year
period. However, the eﬂ'ect of increased scattering by suspended particles
impacted much of the visible and near-infrared portion of the spectrum from
about 500 nm to about 850 nm. This scattering effect overwhelmed the subtler

inﬂuence of other features such as the Chlorophyll a minimum. Although a

54

relationship between water clarity and Landsat measured reﬂectance can be
established, this should not be construed to imply that such relationships

could be developed for other water quality variables such as chlorophyll a.

Kloiber et al. (2002) noted the importance of radiometric calibration. The
brightness values of the pixels in a satellite image were affected by sun angle,
atmospheric interference, changes in detector response, and numerous other
factors. If radiometric correction techniques accounted for these factors, then
the coeﬂicients for the models would be more consistent, and one set of

coefﬁcients would apply for different images across time and space.

All relevant publications on techniques that were used with hyperspectral data
were reviewed and discussed, speciﬁcally what had been done, and what
needed to be done in order to improve remote sensing of water quality.
Although some of the spectral bands were identiﬁed in the previous studies,
they were not generalizable due to the disadvantages of empirical methods that
are data dependant. An empirical model was often derived based on
relationship between dependant and independent variables from a speciﬁc set
of data. Spectral indicators based on correlations between local in situ
measurements of water quality and spectral variables at one location may not
represent the relationship between the same variables at different locations.
Spectral indicators developed from wavebands that truly explain optical
properties of the variables such as absorption and reﬂectance features are the

potential solution. However, very few studies have been conducted on deriving

55

the spectral features of water, such as spectral library development, when
compared to those in mineral and vegetation sciences. In addition, the spectral
bands identiﬁed by these derivative techniques were mostly derived from a
specially controlled environment. Therefore, they may not be applicable in the
real natural environment. Water quality studies should be conducted on a
spatial extent large enough to account for the local biophysical conditions, such
as dissolved organic carbon from woods in nearby swamps or suspended
inorganic sediment areas with high slopes, in order for the method to be
generalizable. A study to use spectral bands that were developed based on
optical properties of water quality variables to quantify the relationship between
water quality and spectral data within a natural biophysical condition needed to
be conducted at large scales. This study attempts to develop in the remaining
chapters generalizable algorithms that are spatially and temporally
independent. The objectives of this study are to (1) identify optimal spectral
bands that are most sensitive to water quality indicators in the various water
bodies within Michigan; (2) develop improved spectral water quality indicators;
and (3) determine the most effective ﬁlters for noise removal in hyperspectral

remote sensing data.

56

CHAPTER 3

EXPERIMENTAL DESIGN

Fieldwork was conducted throughout the Lower Peninsula of Michigan. A
majority of the sampled lakes was in the Muskegon River Watershed due to the
extensive amount of ecological research that had been going on within this
watershed. Muskegon River Watershed attracts researchers because it contains
high variation in topography and land use/ land cover types. However, the
experiment was desigied to include lakes in a wide trophic range. The study
site was therefore extended to the entire Lower Peninsula of Michigan. One
purpose of this study was to quantify the relationship between spectral
information and biophysical variables that indicate water quality. Bathymetric
maps were used to predetermine the sample sites within the lakes in order to
diminish the effects of other features, such as lake bottoms and submerge
vegetations. Data were collected when the sky was cloud and haze free to
minimize inconsistency in down-welling radiance and the effect of atmospheric
gases, and when the sun elevation was high above the horizon to reduce the
sun-glint (hotspot) effects. The Characteristics of the study area, spectral and
water quality data acquisition procedures along with analytical methods are

presented in the following sections.

57

3.1 Description of the Study Area
The study site covered almost the entire Lower Peninsula of Michigan.
Therefore, biotic and abiotic components of the Lower Peninsula were described

following (Albert, 1995; Schuette and Skjaerlund, 1994; Veatch, 1941).

3.1.1 Climate

The weather of Michigan was controlled by three major air masses, the
Continental Polar, Maritime Tropical, and the Maritime Polar (Eichenlaub, 1979).
The Continental Polar air mass, forming over land in the Arctic, brought cold,
dry weather in the winter and cool conditions in the summer. The Maritime
Tropical, forming over the waters of the Gulf of Mexico to the south, brought
warm moist winter weather and hot humid summer conditions. The Maritime
Polar air mass originated in the northern Paciﬁc Ocean, although it originally
carried large amounts of moisture, much of this was lost on the western slope
of the Rocky Mountains. The air warmed as it descended from the mountains.
The Maritime Polar air mass brought mild weather with little precipitation to

the Midwest.

The Great Lakes were another major control on climate for Michigan. These
effects increased the amount of storms over and nearby to the lakes during the
winter, but decreased the intensity of storms and increased the stability of air
masses over the lakes during the spring and summer. Areas where elevation
increases rapidly near lakes receive the most lake-effect precipitation. Climate

was responsible for major differences in both soils and vegetation. Along the

58

Great Lakes, the air near the coast warms more slowly in the spring and cools

more slowly in the fall than in the continental climate area.

3.1.2 Bedrock Geology

The continental interior of North America, including all of Michigan, Minnesota,
and Wisconsin, was known as the Central Stable Region or craton, an area that
was relatively stable during the Paleozoic (Dorr and Eschman, 1984). During
the Paleozoic, from Cambrian to Pennsylvanian times, the southern portion of
the craton, including Michigan, Minnesota, and Wisconsin, was intermittently
submerged beneath shallow seas. Marine and near shore sediments, including

limestone, dolomite, evaporites, sandstone, and shale, were deposited over

Precambrian bedrock.

Roughly 31.6 percent of Michigan was comprised of poorly drained soil (Veatch,
1941). The terms “clay soil,” often referred to land underlain by clay at a depth
of a few inches to approximately one foot. This broad group of soil constituted
the greater part of the highly productive and durable agicultural land in the
State of Michigan. An estimated 70 — 75 percent of the original wet or shrub
land underlain by clay had been cleared and drained for some sort of

agicultural use by 1941 (Veatch, 1941).

59

3.1.3 Physiography

Modern physiog'aphy and soils were the result of postglacial erosion and soil
formation processes as the result of glacial deposits during the Wisconsinan
Glaciation of the Pleistocene Epoch. Erosion of bedrock and unconsolidated
materials occurred beneath the advancing glacier. The advancing ice scoured
the bedrock uplands, producing rounded knobs. Rocks and soil materials were
carried in the glacial ice and later redeposited and formed diverse features,
including moraines, drumlins, eskers, kames, and outwash plains. Lakes and
depressions were common in the glacial landscape. Many lakes formed when
large blocks of ice were surrounded by outwash sands as the glacier melted.
Lakes also formed in linear depressions that had been scoured out by the
glacier. Swamps and marshes occur where vegetation colonized shallow
depressions. Michigan’s unique geographical location provided its citizens with
rich freshwater resources including over 11,000 inland lakes. In addition to
ecological value, lakes provided tremendous aesthetic and recreational value for

people in Michigan.

3.1.4 History of Land use / Land cover in Michigan

The present-day vegetation of Michigan resulted from the physical environment,
post-Pleistocene species migration patterns (Albert, 1995), and human
alteration of lands and plant communities. Disturbances such as logging,
agiculture, drainage, ﬁre, and ﬁre exclusion had signiﬁcantly altered plant
cover and composition (Albert, 1995). Located in the Midwestern Corn Belt,

Michigan has an enormous area of agicultural land. Most farmers apply

60

nutrients to maximize their crop yields, but excessive nutrients carried by water
runoff became pollutants downstream. The tile-drained soil of Michigan
enhanced the severity of the nutrient pollution problem by preventing
penetration of nutrients into the ground. As a result, considerable amounts of

fertilizers accumulated into nearby lakes.

During the past 100 — 125 years the natural landscape of Michigan had been
altered by human actions (Veatch, 1953). Humans cannot change the major
elements of the environment, such as the climate, the land forms, the
composition of the Glacial drift covering, or the bedrock; however, they can
make alterations in the cover of vegetation, the fauna, the soil, and the waters.
The effects of man’s activities after 100 years were not geographically uniform
throughout the state, partly because of the differences in regional Climate,
topography, soil and other resources. Those variable impacts resulted in part
from the spatial distribution of the population in Michigan. Historically,
approximately 90 percent of the people in the state lived in the southern half of
the region with only 5 percent in the entire Upper Peninsula (Veatch, 1953).
Out of the approximately 149,734 km2 of the land area of Michigan, 72,843 km2
(48.6 percent) were Classiﬁed by the US. Census (1953) as farmland and as
much as 8,094 km2 as cities, industrial sites, highways, and rural homes
(Veatch, 1953). Agiculture contributed more than 37 billion dollars annually
to the state economy (Schuette and Skjaerlund, 1994). Michigan has many
agicultural advantages, such as an abundance of inland fresh water, fertile

soils and a mild Climate moderated by the Great Lakes. Because of its unique

61

micro-climates, the state of Michigan was ranked second in the nation with
agicultural diversity. Over 100 different food and ﬁber products had been
produced in Michigan. The state has been the lead producer of tart cherries,

blueberries, ﬂowers and edible beans (Schuette and Skjaerlund, 1994).

Land use in Michigan has decreased in agicultural land and pasture. This
decrease in agicultural acreage occurred in concert with increases in urban
and built-up land (National Resources Inventory, 1987). Agiculture was
directly impacted by recent trends in land use patterns. Michigan had not
experienced sig'iiﬁcant increases in population during the last 20-25 years,
although a dramatic shift in the location of residential development had
occurred. As a result, the amount of land used for residential housing had
continued to increase at a rapid rate, placing additional pressures on
agicultural land. When suburbs expanded, they often invaded lands
previously planed in agricultural corps (Schuette and Skjaerlund, 1994). The
impact of increasing residential development was not only in the loss of
farmland, but also impacts existing farm operations. Decreasing farm size led
to increase in crop production by intensiﬁcation of cultivation and improvement
of farm management. Where the natural fertility was not favorable, increased
use of commercial fertilizers was an alternative to help boost production
(Veatch, 1953). This change in farm operations could magnify the non-point

source pollution problem.

62

3.1.5 Selected Lakes in this Study

In situ data were collected from 42 lakes in Michigan from April 24 to August
26, 2004 (Figure 3— 1). Some lakes were revisited in both spring and summer
eresulting in 49 independent in situ measurements of chlorophyll a (CHL), total
suspended sediment (TSS), Secchi disk depth (SD), total nitrogen (TN), total
phosphorus (TP), non-purgable organic carbon (NPOC), light extinction proﬁle,
dissolved oxygen (DO) proﬁle, temperature proﬁle, phytoplankton species
composition, and reﬂectance spectra. In situ upwelling radiance from the water
column was measured with the hand-held spectroradiometer (LabSpec® Pro,
Range 350 nm — 2,500 nm with sampling interval 1.4 nm @ 350 — 1050 nm and
2 nm @ 1,000 nm — 2,500 nm). An intensive ground truth data set was
collected between July 24 — 28, 2004 within days of the airborne hyperspectral

imagery acquisition (July 24 and July 26, 2004), respectively.

63

 

V

Lacs-Chew

I!
‘ Lanae >
O .'
JOdudLale
iGrtLae

kev’

mails

 

 

 

 

Figure 3- 1 Geographic location of the lakes sampled in the study

The ﬁeld measurement lakes were selected to represent a wide range of water
quality characteristics (from hypertrophic waters to oligotrophic waters) based
on the preliminary trophic state data estimated from historic CHL and SD
measurements from three sampling programs: (1) the Michigan Department of

Environmental Quality’s (MDEQ) Lake Water Quality Assessment (LWQA)

64

 

Monitoring Program, (2) the Michigan Cooperative Lake Monitoring Progam
(CLMP), and (3) Professor R. Jan Stevenson’s algal ecology lab, MSU
Department of Zoology. These data were collected in spring and summer
between May - August of 2001 to 2003. Water conditions of the sample lakes
varied from Clear (SD 8.5 m) oligotrophic lakes to turbid (SD 0.5 m)
hypertrophic lakes. Approximately 86 percent of the water depth at the
sampling sites was, at a minimum, twice the Secchi depth. Therefore, the effect
of bottom reﬂections on remote sensing observations was negligible. The rest of
the sample sites were collected at the deepest basins where the bottom was not
visible, although water depth was less than twice the SD. The detailed ﬁeld

procedure is included in Appendix A.

3.2 Remote Sensing Data Acquisition
Two sets of hyperspectral remote sensing data were used in the study. The two
data sets were acquired from a hand-held ASD sensor and an airborne AISA

SCHSOI'.

3.2.1 Hand-held Analytical Spectral Devices (ASD)

Upwelling radiance data (for determination of reﬂectance spectra) were collected
at the same time as water sampling from 42 lakes (48 spectra, 6 revisited).
Recorded radiance represented the vertical ﬂux of energy upward from the
water surface (primarily solar energy backseattered within the water column
and emerging from the water surface). These nadir optical measurements were

collected using the LabSpec FR instrument from Analytical Spectral Devices

65

 

 

(ASD, Inc., 2006), positioned approximately 1 m above the water surface and
approximately 0.7 m from the side of the boat not affected by a shadow. The
radiance samples were taken from the deepest basins of the lakes except for
Higgins Lake because the deepest site was too far for the sampling team to be in
a safe working condition. The samples from Higgins Lake were taken
approximately 3 km offshore where the water was extremely deep (>35 m);

therefore, bottom effect was negligible.

Forty eight representative radiance spectra were recorded consisting of 2, 150
spectral bands ranging from 350 nm to 2,500 nm at l-nm intervals. At least 10
replicate upwelling and downwelling scans were recorded at each sampling
location. These spectra were averaged to produce a representative reﬂectance
sigiature for each lake. Down-welling irradiance was indirectly evaluated by
measuring the reﬂected light from a white, near-lambertain reference
Spectralon plate® (LabSphere, Inc., 2006). This reference panel is made from
sintered polytetraﬂuoroethylene that is a near-perfect Lambertian reﬂector.
Figure 3-2 showed the variation of up- and down-welling radiance of a high

trophic lake (Hess) and a low trophic lake (Higgins).

In summary, spectra collection performed at each sample site comprised (l) the
average of approximately 10 radiance spectra of the reference panel at the
beginning of each measuring session; (2) the up-welling nadir radiance, where

each measurement was an average of approximately 15 spectra collected at

66

 

every sampling site; and (3) another average of approximately 10 spectra of the

reﬂectance panel at the end of the measuring session.

 

Up- and downwelling radiance in lakes

 

‘ — Hess WR — Hess — Higgins VVR —- Higgins .

 

R-Tlunce

 

 

 

400 450 500 550 600 650 700 750 800 850 900
Wavelength (nm)

 

Figure 3-2 The variation of up- and down-welling radiance as measured from
ASD. The 10x of mean of upwelling radiance are plotted to scale the reﬂectance

values up for visual comparison with the downwelling radiance

The fraction of light reﬂected from lake water was very small compared to other
natural surfaces, such as soils and vegetation. Water—leaving radiances from
natural water bodies are commonly less than 10 percent of the total radiance
measured at the sensor. Most of them are even often less than 1 percent
(Gordon, 1987). Typically, in Clear water, the radiance is maximal in the blue (A
= 440 nm), medium in the green (A z 550 nm) and negligible in the near-infrared
(A 2 750 nm). Since the desired water-leaving radiance is only a small part of
the signal recorded by the sensor, accurate radiometric correction is critical

(Gordon, 1987). To extract the true representative reﬂectance of substances in

67

water column, the collected spectra needed to be transformed to reﬂectance
values at diﬁerent wavelengths that: (1) maximally relate to the concentration of
the constituent of interest; and (2) minimize the effects of other optically active

constituents and survey conditions.

3.2.2 Airborne Imaging Spectrometer for Applications (AISA)

Another spectrometer used in the surveys was the Airborne Imaging
Spectrometer for Applications (AISA). AISA is a pushbroom imager with a
charge-coupled device (CCD) array. The two-dimensional array consisted of a
spatial axis of 364 detectors, and a spectral axis of 286 detectors. The
instantaneous ﬁeld of view across the track was one milliradian, which resulted
in 1-m-wide pixels from an altitude of 1,000 m. The Channel combinations of
the two surveys differed slightly: the number of Channels was 20 on July 24
acquisition, and 30 on July 26 acquisition. The integration time in the surveys
was 30 ms for 20 bands imagery acquired on July 24, 2004 and 20 ms for 30
bands acquired on July 26, 2004. During the surveys, the ﬂight altitudes were
approximately 1,200 m. The 20 hand imagery pixel length along the track was
approximately 1.45 m, and across track was approximately 1.09 m (1.09 m x
1.45 m). Ground resolution of the 30 band imagery was approximately 1.06 m

x 1.11 m.

The AISA has ﬂexibility in selecting the sensor’s spatial and spectral resolution

characteristics. It operates at a wavelength range of 430 nm — 900 nm with a

maximum number of 286 Channels (width prog'ammable from 2 nm — 10 nm).

68

The data rate associated with the short integ‘ation times (sampling rates)

required of the sensor in most operational /ﬂight modes, limits the number of

Channels. The selected channels for this study covered the wavelength range

between 434 nm - 900 nm almost continuously with a channel width of 3 nm —

8 nm. Tables 3-1 and 3-2 listed the band-sets selected and indicate the range

of wavelengths covered by each band.

Table 3- 1 Spectral conﬁgurations of the 20 band AISA data

 

 

 

 

 

 

 

 

 

 

Nbr channel Min wvl center wvl Max wvl thm‘ avg.of'fset avg.gain
1 6 434.45 438.43 442.40 7.95 0 2 1.8771
2 17 451.94 455.92 459.89 7.95 0 17.8561
3 22 459.89 463.87 467.84 7.95 0 15.9328
4 47 499.64 503.62 507.59 7.95 0 7.1509
5 59 518.72 522.70 526.67 7.95 0 5.6548
6 94 575.56 579.73 583.91 8.35 0 3.382
7 100 585.58 589.75 593.93 8.35 0 3.2076
8 121 620.65 624.82 629.00 8.35 0 2.9989
9 135 644.03 648.20 652.38 8.35 0 2.9429
10 141 654.05 658.22 662.40 8.35 0 2.8382
1 1 153 674.09 676.60 679.10 5.01 0 3.9341
12 159 684.11 685.78 687.45 3.34 0 5.3284
13 162 689.12 691.62 694.13 5.01 0 3.4042
14 182 723.06 725.65 728.25 5.19 0 3.4821
15 187 731.71 734.31 736.90 5.19 0 3.5483
16 191 738.63 741.23 743.82 5.19 0 3.6101
17 202 757.66 760.25 762.85 5.19 0 4.0223
18 217 783.61 786.21 788.80 5.19 0 5.2317
19 246 833.78 836.38 838.97 5.19 0 5.8467

20 271 877.03 879.62 882.22 5.19 0 7.3689

 

’thm: Full-width, half maximum in nanometers

69

 

Table 3-2 Spectral conﬁgurations of the 30 hand AISA data

 

 

 

 

 

 

 

 

 

 

Nbr channel min.wvl center wvl max.wvl th' avg.offset avg.gain
1 6 434.45 438.43 442.40 7.95 0 32.8156
2 1 1 442.40 446.38 450.35 7 .95 0 30.5965
3 17 451.94 455.92 459.89 7.95 0 26.7842
4 22 459.89 463.87 467 .84 7 .95 0 23.8992
5 34 478.97 482.95 486.92 7.95 0 17.0929
6 47 499.64 503.62 507 .59 7.95 0 10.7264
7 59 518.72 522.70 526.67 7.95 0 8.4822
8 77 547.34 551.32 555.29 7.95 0 6.4266
9 94 57 5.56 579.73 583.91 8.35 0 5.0729

10 100 585.58 589.75 593.93 8.35 0 4.8114
11 121 620.65 624.82 629.00 8.35 0 4.4983
12 135 644.03 648.20 652.38 8.35 0 4.4144
13 141 654.05 658.22 662.40 8.35 0 4.2573
14 153 674.09 676.60 679.10 5.01 0 5.9012
15 159 684.11 685.78 687 .45 3.34 0 7.9926
16 162 689.12 691.62 694.13 5.01 0 5.1063
17 170 702.48 704.99 707.49 5.01 0 5.0089
18 182 723.06 725.65 728.25 5.19 0 5.2232
19 187 731.71 734.31 736.90 5.19 0 5.3224
20 191 738.63 741.23 743.82 5.19 0 5.4151
2 1 202 757.66 760.25 762.85 5.19 0 6.0335
22 207 766.31 768.90 771.50 5.19 0 6.5596
23 217 783.61 786.21 788.80 5.19 0 7.8476
24 222 792.26 794.86 797.45 5.19 0 8.3681
25 232 809.56 812.15 814.75 5.19 0 8.743
26 246 833.78 836.38 838.97 5.19 0 8.77
27 257 852.81 855.40 858.00 5.19 0 9.0624
28 261 859.73 862.33 864.92 5.19 0 9.4207
29 271 87 7 .03 879.62 882.22 5.19 0 11.0533
30 278 889.14 891.74 894.33 5.19 0 12.7728

 

'thm: Full-width, half maximum in nanometers

70

 

The coordinates of the sampling stations were determined in advance. The
AISA sensor was installed aboard an aircraft, which was equipped with a
differential GPS navigation system that made it possible to accurately overﬂy
the sampling stations. The AISA sensor system also incorporates the Fiber
Optic Downwelling Irradiance System (FODIS). FODIS allows for the
concurrent measurement of downwelling and upwelling radiance by the AISA
sensor head. A diffuse collector installed on the top of the plane is connected to
the AISA head via a ﬁber optic cable and collects downwelling irradiance in the
same bandwidth conﬁgrrations as the areas being imaged. The calibration of
the FODIS coupled with the AISA sensor allows for the calculation of apparent
at-platform reﬂectance. The AISA pre-processing software (CaliGeo) provides for
the automatic geometric correction, rectiﬁcation, mosaicing, and calculation of
radiance or apparent at-platform reﬂectance (FODIS ratio). The program uses
the GPS and attitude information to perform the geometric, georeferencing and
mosaicing operations. The AISA data used in this study were radiometrically

and geometrically corrected to apparent reﬂectance at sensor level.

3.3 Water Quality Parameter Data

Field activities at the sites included collection of water samples for laboratory
analysis of CHL, TSS, TN, TP, NPOC, and phytoplankton types, and the on-site
measurement of SD, lake depth, light-extinction proﬁle, DO, and temperature
proﬁle. SD was measured by lowered a Secchi disc (a 20-Centimeter diameter,
black and white disc attached to a calibrated rope with permanent ink marks)

into the water while observing the depth at which it disappeared. The disc was

71

lowered more and then raised until it reappeared. The depth of the water where
the disc vanishes and reappears is the Secchi depth. DO and temperature were
measured using a YSI® 55 Dissolved Oxygen Meter, and the underwater light
was measured using an LI-250 light meter with the LI—193SA underwater

spherical quantum sensor attached to a 30-meter underwater cable.

The water samples for the Chlorophyll a and TSS analyses were collected in
bottles, kept in the dark and ﬁltered at the end of the ﬁeld day through 47 mm
Whatman GF/F ﬁlters (0.7-um pore size membrane). The ﬁlters for pigment
analysis were wrapped in aluminum foil, stored in an ice cooler, transfer and
storage in a freezer, and then analysis was carried out. This occurred within
ﬁve weeks from storage. The ﬁlters for TSS measurement were also wrapped
with aluminum foil and stored in the freezer. Water samples (850 ml) for
phytoplankton analysis were immediately ﬁxed with the iodine solution (M3) on
site. The samples were then concentrated in the laboratory and stored in the
30 ml vials until analysis. Samples for N POC analysis were stored in 30 ml
vials at the sample site, wrapped in aluminum foil, and stored in a refrigerator

until analysis.

3.3.1 Lab Analytical Methods for Chlorophyll a

Samples were removed from the freezer and brought to the dark room, which
only had yellow and orange light. Aluminum wrap were not removed until the
ﬂuorescence light was turned oﬂ‘. Each sample was supplemented with 9 mL

90 percent ETOH and was tightly covered with the cap and the aluminum foil.

72

The samples were sonicated for 15 minutes, covered entirely with aluminum
foil, and then placed in the fridge for 24 hours. Then ﬁlters were removed and
the remaining ETOH solutions were centrifuged for 10 minutes. A text ﬁle (.txt)
was created in the computer connected to the ﬂuorometer to allow ﬁle transfer.
A new Excel spreadsheet was also created and all sample information (sample
site, sample number, date, and ﬁltered volume) were input into the
spreadsheet. Once the samples were centrifuged, they were ready to be
measured. Sample adaptor in the ﬂuorometer was replaced with the solid
standard. The standard was measured by pressing the [*]. The standard
values should be conﬁrmed to match with the calibration values. The solid
standard was then replaced with the sample adaptor. The supernatant of
centrifuged samples were pipetted into 13 mm sample cuvette until
approximately 75 percent full. Blank sample was inserted into the ﬂuorometer
and measured for the ﬂuorescence value. The value was conﬁrmed to be near
zero (~0.05), if not a new blank would be made and used immediately.
Fluorescence of all samples were measured (Rd) and recorded into the
spreadsheet. Samples with high chlorophyll concentration were diluted with
the ratio of 1:5 (1 mL sample with 5 mL 90 percent ETOH). Each sample was
acidiﬁed by adding 8 drops of 0.1 N HCL (0.15 mL for every 5.0 mL of sample)
and stirring thoroughly. Approximately 90 seconds after mixing the acid, the

samples were measured again for the ﬂuorescence values (RC).

73

3.3.2 Lab Analytical Methods for Total Suspended Sediment

TSS determination followed the Filter TSS/ISS method. Approximately eight
TSS samples were taken out from the freezer at a time. Information from the
label (e.g., site, date, volume, ﬁlter number) was recorded. Next, the aluminum
dishes used in the process were labeled and weighed (Dish number. and Dish
Wt). Sample ﬁlters were then individually placed in the dishes, and dried in the
oven for 1 hour at ~ 105°C. The dishes were transferred into the desiccator for 1
hour before being Weighed again for dried weight (Dried Wt). The dish and
sediment were placed in mufﬂe furnace at 500 °C for 15 minutes, and
transferred to the desiccator for 1 hour. The samples were then weighed to

determine loss on ignition (LOI). TSS was calculated using Eq 3- 1 and 3-2).

TSS (mg / L) = (Dried Wt — Dish Wt — Filter Wt) / (Volume Filtered*1,000) (Eq 3- 1)

ISS (mg/ L) = (LOI — Dish Wt — Filter Wt) / (Volume Filtered*l,000) (Eq 3-2)

3.3.3 Lab Analytical Methods for Non — Purgable Organic Carbon

NPOC were analyzed at Hamilton’s laboratory at Kellogg Biological Station,
MSU. The NPOC analytical method was slightly modiﬁed from Hamilton lab
NPOC/ DOC protocol written by Dave Weed. The method was written for the
Shimadzu TOC-chh carbon analyzer with the total nitrogen module (TNM- 1),
and the ASI-V autosampler. The system was set up to automatically add
portions of 2M Hydrochloric acid to each standard and sample, and to sparge

each of them with Chromatogaphic grade or zero air for a predetermined time

74

period. Each sample was sparged just prior to injection onto a combustion
chamber containing platinum catalyst. The vaporized sample then traveled out
to a CO2 detector and to the TNM-l module for chemiluminescence detection of

nitrogen.

(1) Acid Preparation: 2M hydrochloric acid (HCl) was made by adding 41 - 42
mL concentrated HCl to approximately 200 mL of deionized water in a 250 mL

volumetric ﬂask, then diluted to the mark with deionized water.

(2) Standard Preparation: Carbon analysis stock standard was made by drying
potassium hydrogen phthalate (KHP) at 105 - 120°C for 1 hour and cool in
desiccator. The 2.1254 grams KHP was accurately weighed into weigh boat
before transferred to 1 L volumetric ﬂask and diluted to mark with deionized
water (use zero grade water if available for the best results). Nitrogen analysis
stock standard was made by drying potassium nitrate (KNO3) at 105-120°C for
three hours and cool in desiccator. The 7.1282 gn KNO3 was accurately
weighed into weigh boat before transferred into 1L volumetric ﬂask and diluted
to mark with deionized water (use zero grade water if available for the best

results) .

(3) Mixed C /N standards: The following standards were typically made for a

combination NPOC / TN analysis: 5, 10, 25, 50, and 100 parts per million

Carbon and Nitrogen.

75

(4) Autosampler Rack Preparation: The autosampler rack had 93 positions and
used 24 mL vials with corresponding caps and septa. A complete set of
standards (including blanks) were placed at the beginning and end of the rack,
and Check standards (either 5ppm C, N or 10 ppm C, N) were placed among
every 15-20 samples. Approximately 20 mL each of sample or standard were
poured oﬂ into 24 mL vials, then capped with open-hole caps and septa, and
placed into rack. The sample level in the vial was usually even with the top of

the rack.

(5) Instrument Preparation: The two outside rinse water reservoirs were Checked
to make sure they were ﬁlled with deionized water. Front panel was opened to
make sure that the two water traps on the bottom right internal area of the
instrument had water at the correct levels. The front trap had high and low
markings; the back trap should be ﬁlled to the level of the clamp that holds it.
The 2M hydrochloric acid reservoir should be ﬁlled if low. A full HCl reservoir
should contain enough acid to run about ﬁve complete racks (93 samples per
rack), if not more. The main cylinder valve (large silver knob on top of cylinder)
and the regulator valve (black knob) were opened on the zero air compressed
gas cylinder. The instruments were powered on to open gas lines to the
instrument and turn the fans on. The actual instrument settings were
initialized using the TOC software located on the Dell Computer next to the Ion

Chromatograph.

76

(6) Setup using TOC software: The TOC oven must be heated to 720°C and
stabilize prior to running samples / standards. The TN module also required

some stabilization time. This process took at least 20-30 minutes.

(7) Sample Table: The sample table was used to assign methods and rack
positions for each standard and sample. The software was capable of running
the calibration standards, but typically all standards were run as unknowns.

Afterwards, the data was used with MS Excel to create calibration curves.

The principle of Non-purgable organic carbon analysis was to measure samples
that were acidiﬁed (pH less than 2.0) and sparged with CO2 free air. The
sample was injected into a TC (Total Carbon) combustion tube, which had been
ﬁlled with oxidation catalyst and heated to 720°C. The sample was combusted
or decomposed to C02. The combustion product was sent through an IC
reaction vessel, cooled, dried by a dehumidiﬁer and then sent through a
halogen scrubber. The NPOC component was detected for CO2 in a non-
dispersive infrared gas analyzer (NDIR). The peak area count was proportional

to the NPOC concentration of the sample.

3.3.4 Lab Analytical Methods for Algal Community Counting
Phytoplankton community counts used the Soft Algae Counts method. Species
composition and biomass was determined by counting a known volume sample

under a microscope followed the steps:

77

(1) A sample list was created to indicate initial volumes (collected from ﬁeld),
sample ID (what lake and date the sample was from), volume of sample (in lab

before counting).

(2) The initial volume of the sample was measured by comparison with a
marked ‘Standard bottle.’ The standard bottle was made by marking the
outside surface of the empty bottle that was exactly the same as the bottles in
which the original samples were collected with lines that correspond to volumes
precisely measured with a graduated cylinder. The reference 35 mL bottle was

marked every 1 mL.

(3) The outside of the sample bottle that represented lower meniscus level of the
sample was marked with a ﬁne tip permanent marker to precisely measure the
sample volume without transferring the sample into a graduated cylinder of
beaker, which was time consuming and required adding rinse water. The mark

was compared with the closest measure on the ‘Standard bottle’ and the volume
' was measured and recorded. Measurement and recording volumes were done

with all samples.

(4) A pipette was used to transfer 0.1 mL sample into the Palmer-Maloney
Counting Cell. There should be approximately 15-30 natural units in a ﬁeld of
view in order to count a sample in a timely manner without drying of the
subsample in Palmer Maloney counting chamber. If algal density met the

requirement, start counting. If not, the sample needed to be diluted or

78

concentrated further. Algal division, size, and number of colonies were
recorded along with the counting distance on the microscope (for area

calculation) .

(5) Representative biovolume for each algal category was determined. Known
algal genera were recorded when counting. The biovolume of these genera were
derived from Stevenson et al. (1996), Buzzi (2002), and the biovolumes of algal
taxa in samples collected by the USGS National Water Quality Assessment
(NAWQA) Program 2004 (The Academy of Natural Sciences, 2006). Biovolume
for several genera in each category was averaged to obtain the representative
biovolume. These biovolumes were used to multiply cell density for each Class

for each sample.

79

3.4 Spectral Analytical Methods

The spectral analytical process followed the ﬂow diagram showed in Figure 3-3.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Lake Surface Reﬂectance
Correlation with ASD= 0.09-0.97
, 14 Lake AISA spectra (20 bands;17 lakes total)
I 48 Lake ASD Spec” I Correlation with ASD: o.eoo.99 (except Pickerel 0.623) 7
\l hbl 14 Lake AISA spectra (30 bands;19 lakes total) I
34 ASD to develop WQ models \ 7
Air-water surface correction 13 AISA to validate WQ models
Volume Reﬂectance

Filtered Volume Reﬂectance PC A
-Mean (1x3) IBand A:702
-Savitsky-Golay ” -Band 3:435
-Wavelet de-noising Daubechies IBand C:896 (excluded)
-Wavelet de-noising Symmlets

 

 

 

Water Quality Indices
-Secchi
-Chlorophyll a

 

 

 

 

First Derivative -Suspended sediment
IMean (1x3) (excluded) -Green algae
-Savitsky-Golay -Blue-green algae

.. Artifact: 0 values

, _ . -Diatom
ln 1St derlvatlve

-Wavelet Daubechies “
-Wavelet Symmlets

 

 

 

 

 

 

 

 

 

r-v Under Curve Area Index

 

 

 

 

 

 

 

—P Curve HeightIWidth Ratio Index

 

absorption/reﬂectance features extraction

 

 

 

 

 

 

Sgolay Wavelet Db Wavelet S Biophysical characteristics
Az435-475 A2435-485 Az435-475 BLUE abs

8:470-570 8:480-575 8:490-570 GREEN ref

0:565-620 0:550-630 C2560-610 GR/RED edge

0:670-690 0:670-700 0:660-700 RED max chl abs (Lo trop)
E:675-700 E:660-705 RED max chl abs (Hi trop)
Fz690-74O E:690-745 F2690-740 NIR peak (Lo trop)
6:700-740 F2700-745 6700.740 NIR peak (Hi trop)

 

 

Figure 3-3 Spectral analytical procedure ﬂow diag'ams

80

3.4.1 ASD Data Preprocessing

Reﬂectance is the percentage of light reﬂected by a target. It minimizes the
eﬂect of different illumination conditions, thus allows a better quantitatively
measuring of the water color. Radiance measurements from the ﬁeld were
converted to surface reﬂectance by using nearly coincident measurements over
the Spectralon panel (Eq 3-3). These reﬂectance data represent the ratio of
reﬂected energy to incident energy with values ranging from 0.0 to 1.0 (0.0 for
no reﬂectance and 1.0 for 100 percent reﬂectance). The multiple spectra

collected from each site were averaged to determine a mean spectral response

for that lake.

0° in-situ = ET in-situ / ET spec * p° spec (Eq 3'3)
Where:

p0 imam; = In situ target percent reﬂectance

ET in-situ = In situ target radiance
ET spec = Linearly interpolated reference panel radiance

0 s c = Reference anel reﬂectance coefﬁcient
PC

Reﬂectance of forty eight spectra of water bodies are Shown in Figure 3-4. It is
notable that reﬂectance below 400 nm and above 900 nm are dominated by
noise. The relatively small amounts of solar energy outside the 400 nm — 900
nm range result in data with comparatively high levels of noise (Harrington and
Repic, 1995). Further spectral processing and analysis in this research will be

based on the spectral data between 400 nm — 900 nm (Figure 3-5). This range

81

of wavelength has been used in a number of studies on inland waters (Dekker
et al., 2001; Flink et al., 2001; George and Malthus, 2001; Harma et al., 2001;
Koponen et al, 2001; Pepe et al., 2001; Pulliainen et al., 2001; Shaﬁque et al.,

2001; Jensen, 2000; Subramaniam et al., 1999).

Lake surface reﬂectance

Reﬂectance
.0

9

 

 

350 550 750 950 1150 1350 1550 1750 1950
Wavelength

 

Figure 3-4 ASD spectral signatures (350 nm — 2,500 nm) of 48 representative
water samples

 

Lake surface reﬂectance

 

 

       

 

  
  

    
  
    
  
  

-w.~.;ﬁ§f;;- e

.-

0.10
‘L\'
V
‘ 0.08 ‘ ‘
§ 0.06«
8
g
l: 0.04- ‘
'1!
. ‘ H; ' .0"
:2; “M“’ g.» M '\
MW'YTZT—‘tf “‘ “‘- ..:‘*
0.02 :5 ‘._ ’Wh'“mh J\;:VI;'MM

 

400 500 600 700
Wavelength

Figure 3-5 ASD spectral signatures (400 nm — 900 nm) of 48 representative
water samples

82

Optical characteristics of different biophysical conditions in water are shown in
Figure 3-6. Pickerel Lake and Ford Lake exhibit high trophic conditions (CHL =
113.79 mg/L and CHL = 80.54 mg/ L, respectively). The difference between the
two water bodies is that Ford Lake had a much shallower Secchi depth of 0.95
III, while the SD in Pickerel Lake was 2.20 m. Lower SD lakes have higher
reﬂectance. Both reﬂectance curves Show the unique CHL sigrature features
such as high reﬂectance in green wavelengths (500 nm - 600 nm), absorption in
red wavelengths (650 nm - 700 nm), and a secondary peak reﬂectance at near-
infrared wavelengths (near 700 nm). The reﬂectance curves for Brooks Lake
and Tamarack Lake are indicative of waters with moderate concentrations of
suspended sediment (TSS = 64.00 mg/ L and TSS = 51.40 mg/ L). These two

lakes had the highest TSS of all lakes sampled.

 

Optical characteristics of different water condition

 

— Brooks ——Tamarack — - - Pickerel —- - - Ford

 

Reﬂectance

 

 

 

 

 

 

400 500 600 700 800 900
Wavelength

 

 

 

Figure 3-6 ASD spectral Characteristics of different biophysical dominate in

waters

83

3.4.2 AISA Image Preprocessing

Representative reﬂectance values for each ground-truth site were extracted for
each band at locations that coincided with the sampling stations. Mean values
were calculated within 3 x 3 pixels window (approximately 3.2 x 3.3 meters)
around the sampling station. This pixel window size was reasonable enough to
smooth out noise or errors in the spectral data, yet maintain homogeneous
water quality. Pearson correlations were used to investigate the relationship
between the airborne AISA spectra and the hand-held ASD measurements. The
AISA 20-band data set had low, and some negative, correlations with ground-
truth ASD (Table 3-3; Figure 3-7). On the other hand, the 30-band AISA data
set produced very high correlations with the ground measurement ASD (Table
3-4; Figure 3-8). Despite the fact that some of the AISA spectra were correlated
with the ASD spectra of different dates (May and June, due to the lack of ASD
data near the date), and atmospheric effects could play an important role in low
correlations, the sky on July 24 was overcast with patchy clouds that could
have constantly change the incident light. The sky was clear when
measurements were made on July 26. As a consequence, only the 30 hand
data set will be used in further analysis. Figures of 20-band and 30-band AISA

imagery are provided in Appendix B.

84

Table 3-3 Pearson correlation between 20 band AISA and ASD spectra

 

 

 

Lakes AISA Date ASD Date Correlation
Brooks 7 / 24 / 2004 7 / 27 / 2004 0.928
Clear (@Mecosta) 7 / 24 / 2004 7 / 24 / 2004 -0.249
Hess 7 / 24 / 2004 7 / 27 / 2004 0.973
Higgins 7 / 24 / 2004 7 / 27 / 2004 0.776
Houghton 7 / 24 / 2004 7 / 26 / 2004 0.908
Jehnsen 7 / 24 / 2004 5 / 19 / 2004* -0.368
Kimball 7 / 24 / 2004 5 / 26 / 2004* 0.229
Marl 7 / 24 / 2004 7 / 27 / 2004 -0.879
Mecosta 7 / 24 / 2004 7 / 24 / 2004 -0.323
Mitchell 7 / 24 / 2004 7 / 25 / 2004 0.627
Muskegon 7 / 24 / 2004 7 / 28 / 2004 0.424
Roger dam pond 7 / 24 / 2004 7 / 24 / 2004 0.351
Round 7 / 24 / 2004 7 / 24 / 2004 0.020
Silver 7 / 24 / 2004 7 / 25/ 2004 0.801
Tamarack 7 / 24 / 2004 6 / 5 / 2004* 0.090

 

 

 

 

85

* Lakes that ASD was not measured at the same period

 

oﬁm 2958 wécoawoboo 05 ﬂzomoaaou ESQ Com Oman: «smug :o

amigo owes: oQEm «em? .815 2:me do mouse—QB 8.56on <wH< 653 cm 9.8 Qm< nooEon nomEmQEoO mum oasmﬁ

 

 

 

     

 

    

   

 

 

 

 

 

 

 

 

 

 

 

 

 

 

. 595.355
com com com com com oov

. . . . . 88
..... . 8
i l ill'ldﬂll‘dllll i l Foo ”
m
U
no.0 a

$2. . . . 09.. m 1-: ll 8.8

3:835: cunts» 9.3 «:39:
08.0 mwod owed om< 39o 05.0 mood

. . . . 83

l i . 56
w
l l l l l l il l] .88 s
. E88 n um v

886 + x856 u >
I l ii ‘i I . ii- 88
I cod
<m_< teen on new om< 5222. :eaeEBO

 

 

(ﬂ-

3... m— .-
, i—V" - —nc-‘-a‘-.-—.

x

w

331.:—

I, ' I '*4 ‘1 1’
V
~..m~<—-

h.
l.

86

Table 3-4 Pearson correlation between 30 band AISA and ASD spectra

 

 

 

 

 

 

Lakes AISA Date ASD Date Correlation
Arbutus 7 / 26 / 2004 7 / 26 / 2004 0.971
Brooks 7 / 26 /2004 7 / 27 / 2004 0.982
Clear (@Mecosta) 7 / 26 / 2004 7 / 24 / 2004 0.934
Hess 7 / 26 / 2004 7 / 27 / 2004 0.987
Higgins 7 / 26 / 2004 7 / 27 / 2004 0.985
Houghton 7 / 26 / 2004 7 / 26 / 2004 0.914
J ehnsen 7 / 26 / 2004 5 / 19 / 2004* 0.936
Kimball 7 / 26 / 2004 5 / 26 / 2004* 0.591
Marl 7/26/2004 7/27/2004 0.915
Mitchell 7 / 26 / 2004 7 / 25 / 2004 0.832
Muskegon 7 / 26 / 2004 7 / 28 / 2004 0.801
Pickerel 7 / 26 / 2004 5 / 26 / 2004* 0.623
Roger dam pond 7 / 26 / 2004 7 / 24 / 2004 0.892
Sapphire 7 / 26 / 2004 7 / 25 / 2004 0.989
Tamarack 7 / 26 / 2004 6 / 5 / 2004* 0.950

 

87

* Lakes that ASD was not measured at the same period

 

36 29:8 wSccoawoh—oo 05 3:89.58 0509 com .0995 Emacs

:0 >355 owns: 35m «$2 0di mmo: .«o moaumcwum 3.50on <m_< 053 on cam Qm< :ooEon 0080.89.50 m-m aha—ME

 

 

 

 

 

  
 

 

 

 

 

 

 

      

 

 

 

 

 

 

£93.05;
com com con com com oov
. . L r 00.0
w
3.0 m
U
a
00.0
$2. . . . %<J|10
3:83.30 38.50 8.3 one:
.‘ 00¢. .I
02.0 000.0 08.0 03.0 08.0 80.0
_ _ _ 0 8.0
- II I- 8.0
0080 u «m. w
t 00.0 10000.0 n» v
I ‘11 l . l l- 8.0
8.0
<02 0:3 00 can om< 0822. 00022.8

 

 

88

3.4.3 Air-Water Interface Correction
The total radiance, (L1) recorded by the remote sensor is a function of the
electromagnetic energy from the four sources identiﬁed in Figure 3-9 (Eq 3-4)

(Jensen, 2000) .

L1=Lp+Ls+Lv+Lb (E934)
Where:
Lp is an atmospheric noise commonly referred to as path radiance. It is the
portion of the radiance recorded by a remote sensing instrument resulting from
the downwelling solar (Esun) and sky (Esky) radiation that never actually reaches

the water surface.

L3 is sometimes called the free-surface layer or boundary layer. It is the
radiance from the downwelling solar radiance that reaches the air-water
interface but only penetrates it a millimeter or so and is then reﬂected from the
water surface. This reﬂected energy contains spectral information about the
near-surface characteristics of the water body. Unfortunately, if the solar
zenith angle and sensor viewing angle are almost identical, this spectral

component presents only a purely specular reﬂection (sun—glint effect).

L, is the subsurface volumetric radiance. It is the radiance that actually

penetrates the air-water interface, interacts with the water and

organic /inorganic constituents and then exits the water column without

89

encountering the bottom. This radiance provides valuable information about

the characteristics of the water column.

Lb is the portion of the radiance that reﬂected back from the bottom of the
waterbody. If the objective of the study is to obtain information about the
bottom, i.e. bathymetric mapping or coral reef mapping, then this radiance
component may be of signiﬁcant value. However, in water column study,
radiance from the bottom causes diﬁiculty for properly a characterizing the

water column above it.

Downwelling

LP
L
b

 

Figure 3-9 Down- and up-welling radiances from water bodies (after Jensen,

2000)

90

To achieve the goal of identifying the organic and inorganic constituents in the
water column, volumetric radiance (Lu) needs to be extracted from all the other

radiance components being recorded by the sensor (Eq 3-5).
Lu=L:—(Lp+Ls+Lb) (Eq3-5)

The data used in this study were collected 1 m from the water surface with a
hand-held ASD and 1 km above surface on a very clear day with an airborne
AISA. Therefore, atmospheric attenuation (LP) should not have much effect on
these data. Bottom reﬂectance (Lb) was carefully avoided during ﬁeld sampling
(see Appendix B). However, surface sunglint and other surface reﬂection (Ls)
needed to be corrected because it was impossible to avoid in the ﬁeld

measurements.

The analytical process is to convert surface reﬂectance to subsurface volume
reﬂectance, or volume reﬂectance, because it is nearly independent of
atmospheric properties and is almost entirely determined by the optical
properties of the water and its constituents (Bukata et al., 1995). The equation
used by Morel and Gentili (1993; Eq 3-6) was modiﬁed to Eq 3-7, which was
applied to the data to convert the surface reﬂectance (remote sensing reﬂectance)

into subsurface reﬂectance.

Rrs= (I-p)*(1-p')*§(w+stf (Eq 3-6)
1—r*R(0—)*n *Q

 

91

Rrs — Rsurf (Eq 3-7)

R(0—)= ,
(1—p)*(1-p) .. .. _
n2 +r Q (Rrs stjf)

 

 

Where:
Rrs is the remote sensing reﬂectance
R(O-) is the volume reﬂectance
Rsurf is a specular reﬂectance from the surface of the water body

Q is a ratio of downwelling irradiance to upwelling radiance (5 sr"1)

p is an internal Fresnel reﬂectance (0.03)

p' is an air-water Fresnel reﬂection at the interface (0.54)
n is a refractive index of water (1.34)

r is the water-air reﬂection (0.54)

Gege (2001) applied this model for calculating remote sensing reﬂectance in
the WASI 2.0 (Water Colour Simulator) software. Default values indicated in
the brackets of the key of Equation 3-7 were taken from the WASI 2.0 (Gege,

2001)

Upwelling spectra measured above the water can be affected by sun glint
from the water surface (specular reﬂectance, Rwy) due to waves. The
assumption that light absorption by pure water is predominant in NIR (970
nm - 1,000 nm), and the water-leaving radiance in this region is zero, was
suggested to estimate R.m,fin the measured spectra (Ouillon et al., 1997 ). It
should be noted that this assumption may not hold valid in highly turbid

waters (Doxaran et al., 2002). However, the TSS range of the study sites were

92

considered low to moderate; therefore, the waters were assumed not highly
turbid and the assumption of Ouillon et al. (1997) was implemented. Due to
the high noise levels of the collected spectra, the spectral range between 1,250
nm — 1,260 nm were found more stable than other NIR regions (Figure 3-4).
The average reﬂectance from this range was used to remove specular
reﬂectance and wave effects by subtracting it from the whole spectra (Eq 3-7).
Figure 3- 10 and 3-11 show the reﬂectance before and after corrected for air-water
interface. The corrected volume spectra are generally higher than the measured
surface reﬂectance. The volume reﬂectance spectrum that was below zero was
Round Lake. The samples were taken from this lake late in the evening and the
sun elevation angle was very low. There was a fair amount of sun glint reﬂected
from the water surface, which causes a problem when converting from surface

reﬂectance to subsurface volume reﬂectance.

93

 

Surface reﬂectance

 

 
  

Reflectance
O
8

 

 

 

 

Wavelength

Figure 3-10 Remote sensing reﬂectance Rm

 

( Volume reﬂectance

 

0.12 4»

 

Reﬂectance
O
O
m
J:

p
2

v f .
«‘th .. - I
r—Wﬁm~~- '1
0.. 5:... "m

   

 

 

  
  
  

. .0- , .0
$0“ "\\_
5?.» ”a man}:nx‘

  
 

    

 

 

.k, .
7114‘" '~....'.»‘.‘- ..., _ .
0.00 ~ 7 l , ‘ . W'“ v'nﬁ“‘r‘r’r~-i"’ ll?“
-o.04 . . . . . . , . .
400 450 500 550 600 650 700 750 800 350 900

Wavelength

 

 

Figure 3- 1 1 Subsurface volume reﬂectance R(O—)

94

3.4.4 Noise Reduction

The volume reﬂectance were used in the principle component analysis (PCA)
and the derivative analysis in order to determine the optimal spectral bands
that best related to biophysical parameters within the water column. Both PCA
and derivative techniques are based on the variation within the data; therefore,
they are very sensitive to noise in the data. Noise can obscure important
features such as peaks, valleys, or peak widths, or make calculation of signal
features such as slopes, areas, peak widths and so forth difﬁcult. The volume
reﬂectance spectra needed to be ﬁltered to produce the data that are
uncontaminated by noise signal from the atmosphere or the sensor itself. The
ﬁlter should maintain the sharpest absorption /reﬂectance features in the
original signal. Mean and Savitsky Golay ﬁlter, and Daubechies and symlet
wavelet de-noising transformations were used (Schmidth and Skidmore, 2004).
Several conﬁgurations of these methods were applied to the volumetric
reﬂectance (Appendix C). The methods were compared based on the criteria
that the selected model must smooth out high frequency noise while
maintaining the smallest features that could be associated with biophysical
attributes (absorption troughs and reﬂectance peaks). If the higher wavelet
order and level did not provide more optical information than the lower ones,

the simpler models were selected.

In this process, the volume spectra were ﬁltered with: Mean 1x3, Savitsky Golay

(Sgolay), Daubechies de-noising wavelet, and Symlet de-noising wavelet. Mean

ﬁlter is the simplest method. Savitzky Golay seeks to preserve shapes of

95

reﬂectance peaks. The Savitzky Golay ﬁlter method essentially performs a local
polynomial regression to determine the smoothed value for each data point.
This method is superior to adjacent averaging because it tends to preserve
features such as peak height and width, which are usually 'washed out' by

adjacent averaging (Figure 3-12).

The general wavelet de-noising procedure involves three principle steps:
decomposition, threshold selection, and reconstruction. First, a wavelet is
selected and a level at which the wavelet decomposition will be computed. In
this study, two wavelets were selected as the best wavelet options, Daubechies
order 1 level 3 (dbl) and Symlet order 8 level 3 (s8). The coefficients of all
component of a third-level decomposition (that is the third-level approximation
and the ﬁrst three levels of detail) are returned concatenated into one vector, C.
Vector L gives the lengths of each component. Second, select a threshold such
as global thresholding (gbl) and apply it to the detail coefﬁcients. Third,
compute wavelet reconstruction using the original approximation coefﬁcients

and the modiﬁed detail coefﬁcients (Figure 3-13; The MathWorks Inc., 2006).

96

 

Original

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

0,0 . 7 1 1
400 soo 600 700 800 900
Wavelength
01 Mean 1x3 [01 savmy Golaylsaolav)
O
2 (
G
“a
a K
0.0 . . 0 . 1 0.0 1 . . .
400 500 600 700 300 900 400 500 600 700 800 900
Wavelength Wavelength

 

 

 

 

Figure 3- 12 Examples of Mean and Savitsky Golay (Sgolay) ﬁlter transformation

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

on Pickerel Lake
lnal
0.1 Orig
400 500 600 700 800 900
Wavelength

o1 Daubechleﬂ Level3 (db1) o1 Symlet. Level3 (so)
8 8
i i
s '5

K

0.0 - . 0 . o.o . . .
400 500 600 700 800 90 £00 500 600 100 800
Wovebm 1 Wavelet“:

 

 

 

Figure 3- 13 Examples of Daubechies (dbl) and Symlet (s8) wavelet de-noising

transformation on Pickerel Lake

97

 

 

3.4.5 Narrow-Waveband (1 nm) Indicator: Principal Component Analysis

A number of approaches have been developed to maximize the spectral variance
of remote sensing data to assist in the discrimination of different surface
materials. The most widely used of these techniques is principal components
analysis (PCA) (e.g., Becker et al., 2005; Lillesand and Kiefer, 2000; Jensen,
1996; Davis, 1986). PCA has been effectively used with multispectral data
through the compression of the information into fewer, independent (orthogonal
channels) components. The principal components transformation assists
removing the inherent redundancy through a decorrelation process followed by
the discarding of low variance components (Richards and Jia, 1999). It is
sometimes used as a tool for feature reduction in multispectral data analysis.
When employing the method with hyperspectral data, the high computational

load is the main concern (Richards and Jia, 1999).

PCA requires no a priori information and is based on the entire spectral domain
of the data set (Rencz, 1999). PCA is generally based on the correlation matrix
between spectral channels. Conceptually, the process deﬁnes a new set of
orthogonal uncorrelated coordinate system or vector space that maximizes the
amount of variance that each axis describes. In doing so, it produces a space
in which the data have the most variance along its ﬁrst axis, the next largest
variance along a second orthogonal axis, and so on. In general, the later
principal components would be expected to show little variance and thus can be
ignored. Eigenvalues explain the amount of the spectral variability contained in

each component, and the eigenvectors explain the relative proportion or

98

contribution of each band to that component (Rencz, 1999). The technique is
only of value however if the spectral structure of the data is distributed

substantially along the ﬁrst few axes (Richards and Jia, 1999).

Generally, high data variance is usually needed for separating different features
in the data. Thus, higher order components with low variance can be discarded
without signiﬁcant information loss. The ﬁrst few high variance features
(principal components) are usually selected. Increasing use of high spectral
resolution remote sensing requires larger image storage. Bandwidth
compression has become more important. Another advantage of PCA is that the
original spectral or image data can be reconstructed from the reduced
representation (using an inverse principal components transform) although
with loss of information. Since the reconstructed data may differ, the
information loss is sometimes referred to as distortion from the original

(Richards and Jia, 1999).

PCA is thus an effective tool for summarizing the common components of
spectral variability of the hyperspectral data into a small number of variables.
The separation of materials in a remote sensing scene by PCA is driven by the
spectral variability of the substances and is thus linked to spectral properties
(Rencz, 1999). However, it is important to recognize that some features selected
in this way may be misleading. For example, original noisy bands will lead to
some principal components with high variance but low separability (Richards

and Jia, 1999).

99

Fourteen out of 48 water spectra were separated for validation. Thirty four
ﬁltered spectra were used to develop the spectral indicators for water quality.
These 34 ﬁltered volume reﬂectance at wavelength between 400 nm — 900 nm
were used in the PCA analysis. The spectral data were arranged in the format
that each row represented spectral data of the lake and each column referenced
a speciﬁc waveband. SPSS® software package was used to perform PCA and

transform the 500 independent spectral bands into their principal components.

Band number was the independent variable. PCA was performed with the
minimum eigenvalue: 1.0. Correlation-based and covariance-based PCA were
performed to compare the emciency of the two methods. Both PCA methods
used standardized input variables that have a mean of zero and a variance of
one. This standardization tends to inﬂate the contribution of variables whose
variance is small, and reduce the inﬂuence of variables whose variations are
large (Davis, 1986). PC analyses based on correlation matrices generate
dimensions that emphasize the features that best capture the predominant
similarities among the variables (Davis, 1986). PCA based on covariance
matrices generate dimensions that emphasize the features that best capture

predominant target differences. Both methods were investigated and compared.

Unrotated PCs from both methods produced a very high percent of total
variance explained in the 1m component, which made meaningful interpretation
difficult. Rotation of the factors to right angles (uncorrelated) with the ﬁrst

removes the effects of variables highly loaded on the ﬁrst factor and enables an

100

assessment of patterns independent of them (Rummel, 1970). PCA was
performed again with the verimax rotation. Rotated correlation-based and
rotated covariance-based PCA produced almost the same results with the same
ranges of high component loadings, but different loading values. The results
consistently showed that there were three signiﬁcant components to extract
from the spectra of 34 lakes. The percent variance explained by the 4th
component dropped dramatically from the ﬁrst three. Although both PCA
methods produced consistent results, principle components from correlation-
based PCA provided more conﬁdence in locating the spectral bands because of
the more deﬁned peak. Therefore, three principle components, which explain
the whole data set while signiﬁcantly reduce redundancy in dataset from 500
bands to 3 principle bands, were produced from correlation-based PCA. The
three principle components (Figure 3- 14) consisted of PCI: band 700 nm (NIR
peak), PC2: band 435 nm and 455 nm (blue absorption) for different ﬁlter
methods, and PC3: band 896 nm which was suspected to detect the
atmospheric effect because water absorbed most incident light after 800 nm.
These bands were selected repeatedly as the principle components regardless of
different ﬁlter and different PCA methods. Each component explained 36
percent, 37 percent, and 26 percent of the variance of the data. The ﬁrst three
components together explained more than 99 percent of the variance of the
data. However, only PCI and PC2 will be used in the water quality indicator
model development because they contain information about biophysical

characteristics in waters.

101

 

 

 

O 20 Filtered Volume Reﬂectance
PC1
0.15
8
C
€0.10
“5
n:
0.05
0.00
400450500550600650700750800850900
Wavelength

 

 

Figure 3- 14 Example of PCA result

The ﬁrst principle waveband was located precisely at the near-infrared
reﬂectance peak of high trophic lake spectra. The second principle wavebands,

especially band 435 nm, were located close to blue absorption region.

3.4.6 Spectral Derivative Analysis

Although the massive amount of hyperspectral data (the hundreds of spectral
bands) may cause difﬁculties in traditional image processing and data handling
techniques, it provides information to allow analyses based on a knowledge of
spectroscopic principles. Data used in spectroscopy are usually collected under
controlled laboratory conditions with full control of the intensity and spectral

distribution of the illumination as well as viewing geometry (Tsai and Philpot,

102

1998). However, having a well-deﬁned spectrum and knowledge of spectral
features means that a scientiﬁc approach to interpretation of the real world
features can be carried out (Richards and Jia, 1999). Absorption features (seen
as localized clips) are often observed in the reﬂectance spectra of speciﬁc
substances. These features provide characteristics of the substrates of interest
when sumcient spectral resolution is obtained. Characterization, and thus
automatic detection of such absorption features, is of particular interest in
hyperspectral image recognition. Absorption features can be characterized by
their locations (bands), relative depths and widths (full width at half the
maximum depth). A complete spectrum is divided into several regions and

absorption features (Richards and Jia, 1999).

Derivative spectroscopy is particularly promising tool among the techniques
developed in spectroscopy for use with remote sensing data (Tsai and Philpot,
1998). Spectral derivative techniques have long been applied in spectral
studies, and have been found to eliminate background signals, differentiate
overlapping signatures, and reduce the affects of turbidity in aquatic
chlorophyll a investigations (Demetriades-Shah, 1990). In hyperspectral
analyses, derivatives are found to be very sensitive to noise. Derivatives of the
second order or higher should be relatively insensitive to variations in
illumination. Nonetheless, ﬁrst and second order derivatives have been the
most commonly used by scientists (Tsai and Philpot, 1998; Fell and Smith,

1982; Butler and Hopkins, 1970).

103

After random noise was minimized from the 34 spectral measurements, the
derivatives were calculated from volume reﬂectance for each of the four ﬁlter
methods: Mean, Sgolay, dbl and s8 (Figure 3-15). Spectroscopic ﬁrst
derivatives were obtained by taking the difference between the reﬂectance of two
bands and dividing that value by the difference between the wavelengths
separating the two bands. The ﬁrst step is to determine the slope of a line
segment by simply dividing reﬂectance value diﬂ'erences by the wavelength

interval separating them (rise / run) as indicated by Eq 3-8.

d1“ = (9m: - on) / (A... -2...) (Eq 3-8)
where:
n = band number

dlst = lst-derivative (line segment slope)

percent reﬂectance

'0
II

A = wavelength (nm)

104

950on ooSBoouoc 0833 05 mo anew—aha «Genomic V Eek.“ Goumﬁunoamn 3332.020 09E 2 -m PSME

 

59.2285
com

con oom

 

 

 

o>z~>teu amp “023.: an

 

 

 

55:.cl.e>s>>

can con coo

 

 

 

4253.8 03 023.: 5000

 

3493/4191) Isl.

engenyep ISI-

 

 

 

 

59.2020;
000 000 005 000 000 000
_ . . . . . 20.0-
_
_ _ .800-
_ . _ .
. . v ‘ _ . __ .800- m
. ._ ‘ l. _ l. l _ . ... . w.
; , 7.. . _ +080 u.
. l .. . _ A
_ . , 0......
m . _ 000.0 m.
. $0.0
05300330900030. \ - 0.00
59.335;
80 80 8x 000 08 80
_ . _ . _ 20.9
036.
. 08.9 m
J, . .r , _ . .. ”I
, . 1 . 000.0 u.
_ . __ _ _._. W
25.0 9
m
. Sod
-14.-) tidlrillll 1-)) 20.0
52.0.0 .2 028E 58: L

 

 

 

 

 

 

 

105

Spectral absorption /reﬂectance features were revealed in the 1m derivative
products, especially in Savitsky Golay ﬁlter method. These optical
characteristics are related to the various constituents in the water column. For
example, green reﬂectance of CHL between 560 nm — 620 nm and maximum
red absorption of CHL near 680 nm. The derivative transformation reveals
waveband regions that are most appropriate for the differentiation of the input
spectra. The mean ﬁlter method did not provide useful wavebands; it was thus

omitted from further consideration.

The 2nd derivative approximations are calculated using Eq 3-9 to investigate
whether more pattern or information can be extracted (Figure 3-16). An
expansion multiplier (0.5) was incorporated in the denominator as this was

found to enhance the differences in reﬂectance amplitude (Becker et al., 2005).

(12nd = (dIStrHl - dIStn) / 0.5 (Ann — An) (Eq 3-9)
where:

n = band number

dlst = 1st derivative

d2nd = 2M derivative approximation

A = wavelength (nm)

106

«.50on 80882.30 0.829» 2.3 Co wcovaba unouogp v Sod Gonnamxoaam 83803.net vacuum Sim ocsma

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

59.20:; 5a:e.o>a>> .
000 000 00k 000 00... 09. 000 000 02 000 000 000
. . . . . . . . . _ 0000.
r 800.
z . n . . . . I
W. l .. _ _ _ 80° N
w. ._ . . . . . _ . 08.0 P
m. l. .. a m.
m _ l _ _ _ m
u 1 ._ . _ 80.0 m.
.
m 08.0
000.0 000.0
0 m 0 a> w c 20
- ||| (‘1 ozmzcou new “conﬁrm“ v|L r >3 t .0 u N u a: 3.0 ill! W
1
l - ﬁncﬂg‘ 593.235
000 08 02 80 08 00.. 000 000 02 000 08 000
. . . . 000.0- _ . . . T _ 000.0-
80.0- 03.0-
_. 0
, «000. z 080.
. w w.
7.. ,_...___..__I ._ __ . _ . D. __ ., i2... _. l . D.
J .. .__, _ ; M0809 : , 0809
,7____ , _. . . m. . w
_. . __ . _.. l w _ A.
w - ~80 m. 08.0 m.
A
a
_ - 30.0 03.0
ll 053.30 EN 028.: 5000 000 0 l 0202...... 0.." 022.0 :8: 000 0

 

 

 

Second derivatives did not provide much information compared to the ﬁrst
derivative because no additional clear absorption or reﬂectance features could

be extracted. Therefore, only the 18‘ derivatives will be used in further analyses.

3.4.7 Region-Waveband

Region—wavebands were determined based on the observed features in
derivative curves and the reviewed literature (Table 3-5, Figure 3- 17). Previous
studies indicate single spectral bands that were successfully used to explain
water quality variables. None of the studies developed water quality indicators
based on a region-waveband. In this study, region-wavebands were indicated
using narrow band locations ﬁ'om the reviewed researches as a guideline. The
ﬁrst-order derivatives from ﬁltered volume reﬂectance spectra were used as a
primary source in the waveband determination. Spectral features were
obviously showed in Figure 3-17. These features indicate waveband regions
that are most sensitive to changes in biophysical variables. Reviewed
researches provided the information to associate each spectral region to its
optical properties (Table 3-5). These wavebands were used to develop spectral
indices for water quality in the next step. Lakes that have high and low trophic
conditions have different optical properties. To insure the ability to differentiate
the same biophysical characteristic but different lake trophic condition, some of

the region-wavebands were overlapped.

108

Table 3-5 Selected waveband regions for water quality assessment

 

 

 

Sgolay dbl s8 Biophysical characteristics

A:435-475 A:435-485 A:435-475 BLUE absorption

B:470-570 B:480-575 82490-570 GREEN reﬂectance

C:565-620 C:550-630 C:560-610 GR/ RED edge

D:670-690 D:670-700 D:660-700 RED max chl a absorption for
Low trophic lake

E:675-700 E:660-705 RED max chl a absorption for
High trophic lake

F:690-740 F:690-7 45 F:690-740 NIR peak for Low trophic lake

G:700-740 G:700-745 G:700-740 NIR peak for High trophic lake

 

 

 

 

109

 

 

, Sgolay Filtered Volume Reﬂectance

A B C DEFG

.0
N
o

_O
.3
U1

 

Reﬂectance
O
0‘

 

l 400 450 500 550 600 650 700 750
1 Wavelength

 

 

 

 

. Sgolay ﬁltered 1st derivative
I A a c o E FG

 

1st derivative

 

 

 

 

 

 

 

 

 

 

 

Wavelength

 

Figure 3— 17 Example of region-waveband selection on Sgolay method

110

3.4.8 Area Under Spectral Curve Indicator

First, derivative curves for each ﬁlter method are separated into region-
wavebands (Table 3-5). Figure 3-17 displays spectral and derivative curves with
continuous lines, but the data are in fact the series of discrete points. To
extract the area under the curve, a 2“(1 order polynomial function was ﬁtted to
these data points for each lake (Figure 3-18). To calculate the area that
represents relative inherent constituents (e.g., smaller area for lower CHL
concentration and larger area for higher CHL concentration), the derivative data
need to be adjusted such that all the data points fall above zero (for absorption
curves) or below zero (for reﬂectance curves). For the absorption curves, the
minimum value within the region was added to all data within the region. For
reﬂectance curves, the maximum value was added to the rest of the data within

the same region. Appendix D shows the area index for the three ﬁlters.

Daubechies wavelet created a staircase artifact in the ﬁltered signals, which
produce a large number of zero values in the 1‘t derivative product. Fitting a
polynomial function to data that has numerous zero values weighs the curve
down near zero. Dbl curve ﬁtting was adjusted to eliminating all the zero

values from the function in order to minimize the artifact of this ﬁlter method.

111

mezzo conﬂuence Someone: .3

US. .0. .050 .3350 Cow—98m“... «Concave 3. one 8. £90.00 dump o>ﬂw>tov a. Go madam 0.5.5 .o moagm w. -m 0.3%....

 

.0 50:30.5! . 59.2255 «wood. .0
0.0“. . 9.2: . 3.0.. . _ 080 0- 0.0“. . 9.2... . 8.3 ..

 

 

00000
e co

 

 

 

 

 

 

 

 

 

 

 

 

m | I
m
M.
u
M.
0 ill
A .
. ... o.
0:. o
o: mm» can ow» on» m2. 0:. no» can m5 mom 80 non mum 08
E: 02.62. 8.on also 9.3... E: 30.63 .5on ciao acne...
.0 505.02.; 505.055 .a
cum on... can 06 cm: one m3 one am: o9. one one m3. 9... m9

 

 

 

 

 

 

O c m . s O o
0. n. .0... 9. 0 _ .0
E: 273‘ aloum ciao gar...

 

 

 

Eon. . 9.2: . 8.3 L 88.0
E: 2.90: 3.on 02:0 0:3...

 

 

 

 

 

 

 

 

112

3.4.9 Spectral Curve Height] Width Ratio Indicator
Maximum height of the same polynomial ﬁtting function measured from y = O
was determined for each lake and each ﬁlter (Appendix E). The width used in

calculation of height / width ratio is basically the width of the region-waveband

 

 

 

 

 

 

 

 
  

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

(Figure 3-19).
Fitting curve Sgolay 435-475 nm
0.0006 .
[ . Lotus - chks . Ford I
O 0 a ‘ . : 0 . . . . ° °
A
0.0004 -, --—7—.— —~q~ +~ _._°_._ 39:; ° ° ° .
O o . " " a a a n ‘ O ‘ A a
a n ‘ A A A A
t D Iﬂ\b ‘ 4)
3.! 0.0002 ~ .. . _ e , ‘
. . " ' l»
‘ Height '1 \
0.0000 = . r a u I . r 1‘
435 440 445 450 455 460 465 470 475
a_ Wavelength
Fitting curve Sgolay 565-620 nm
565 570 575 580 585 590 595 600 605 610 615 620
0.0000 ‘ . I I I ﬂ IT I ﬂ I I r
0.. a " ..O 000
-0.0005 « , ”"-. .. _______ j .
e. be __ ___V‘ I4 .. ..
O O
00
-0.0010 «————~-—— 4r~--~——w -.-.--. - .
O O
O O
O O
O
is -0.0015 .. 4 — ,_____ , V../ —— ~-~ - -— ~—
F O. .0
-00020 P °°“°"‘. ML.)
r0 Lotus - Hicks a Ford]
-0.0025 I
b. Wavelength

 

 

 

Figure 3- 19 Examples of height and width measurements

113

3.4.10 Spectral Band Ratio Technique

Band ratios have been most used extensively in the analyses of broadband
sensors such as Landsat Multispectral Scanner (MSS), Thematic Mapper (TM),
and Systeme Pour I’Observation de la Terre (SPOT). Ratios can also be used for
the analysis of data from two major classes of absorption: visible to near-
infrared absorptions such as the spectral characteristics of chlorophyll a and
other photosynthetic pigments. Ratio indices, created by dividing one spectral
channel by another, are widely used in biophysical investigations. The method
has been used to enhance the discrimination of surface spectral characteristics.
Band rationing can be very eﬁ'ective spectral analysis tools when applied with a
well-formulated rationale and spectroscopic basis. The major advantages of the
ratio indexes are that effects of bi—directional reﬂectance are removed and the
relative color properties of substances are enhanced (Rencz, 1999). Common to
spectroscopic tools, the data must be corrected for the system contribution
(e.g., dark current) or environmental signal (e.g., atmospheric path radiance)
that is additive to the measured radiance. These additive noises contribute

signiﬁcantly to errors in the results (Rencz, 1999).

Spectral features that are strong in biophysical characteristics (e.g., the green
peak and red edge in chlorophyll spectra) can be used with band ratios to
rapidly detect areas with these properties and therefore can be effective for
environmental monitoring (Rencz, 1999). PCA and spectral derivative
techniques were used to determine the optimal wavebands and region-

waveband that are sensitive to changes in biophysical attributes in waters. The

114

selected bands will be used both as the single bands alone and the ratio bands.
Single-band algorithms have an advantage of simplicity but multi-band

algorithm has an advantage in that they can be applied in situations where the
target is not evenly illuminated. Therefore, multi-band variables can be used in

a wide range of circumstance (George and Malthus, 2001).

Ratios were calculated for both area and height / width region-waveband (e. g.,
A/B, A/C, A/D, ..., G/D, G/E, GF). The narrow-wavebands were originally
selected by PCA. It was observed that these bands were closest to the lower
edge wavelengths of the region-wavebands (e.g., 435 nm, 470 nm, 480 nm, 490
nm, 550 nm, 560 nm, 565 nm, 660 nm, 670 nm, 675 nm, 690 nm, and 700
nm). Therefore, narrow-wavebands (l-nm) from the lower edges of region
wavebands were selected for the narrow-waveband indicators in addition

(Appendix F). Ratios were calculated for all narrow-wavebands.

3.4.1 1 Development of New Spectral Indicators for Water Quality

The relationship between the spectral indicators (e. g., area, height/width, and
narrow-waveband) and the water quality indicators (e.g., SD, CHL, TSS, NPOC,
diatom, green and bluegreen algae) measured at the same sites was determined.
Multiple regression methods were used to test the correlation of water quality
observations and spectral data both with the region—waveband, and narrow-
waveband and their multi-band ratio. The water quality parameters and
corresponding ASD spectral indicators from 34 lake data set were used to

develop the models and determine the empirical coefﬁcients for SD, CHL, TSS,

115

NPOC, diatom, green, and bluegreen algae algorithms. Stepwise regression was
ﬁrst performed on all variables to investigate all possible combinations of
spectral bands that potentially best explain biophysical characteristics in the
water. These wavebands were then used in the determination for the ﬁnal
models based on their level of signiﬁcance (P s 0.05, two-tailed analysis),
correlation coeﬂicient, distribution of residuals, tolerance value, and optical
properties of the bands. The models will be validated by the 12 remaining ASD
data to determine and compare the accuracy of the model to ASD sensor. The
best performance models for each water quality variable will be applied to the
AISA data set from the same 12 lakes to investigate an ability to expand the

algorithm to an airborne hyperspectral data.

116

CHAPTER 4

RESULTS AND DISCUSSION

4.1 Michigan Lake Water Quality from Field Observation

Initial observations of water quality indicators include bar graphs, histograms
and Pearson correlation matrices of the parameters (Figure 4-1 to 4-5; Table 4-
1). The histograms (Figure 4-1 and 4-2) reveal that both Secchi depth (SD) and
chlorophyll a (CHL) concentration varied widely among the sampled lakes, but
some of the other biophysical parameters such as total suspended solids (TSS),
non-purgable organic carbon (NPOC), diatom, green and bluegreen algae had
only a few samples with high values, and the rest of the data values were low.
This caused some of the variables to be non-normally distributed (Figure 4-3).
Both logarithm base 10 (LOG) and natural logarithm (LN) transformations of
these data were compared. Both transformations generated more normally
distributed data, but the LN produced a range of values that were more
appropriate to use in the regression models (Figure 4-4). Table 4-1 and Figure
4-5 demonstrated that Secchi depth (SD) had a relatively strong negative
correlation with chlorophyll a (CHL) and total suspended solids (TSS), meaning
that CHL and TSS are the major constituents that change clarity in the sample
waters. The relationship between SD and CHL was stronger. SD and TSS had
a relatively strong correlation with CHL and green algae, suggesting that most

of the TSS was caused by algae rather than inorganic substances (e.g., silt).

117

 

 

Chlorophyll a concentration (mg/L)

0
0

    

0 8
uogenueeuoo

O
n

 

TSS concentration (mg/L)

8

9
uonanueouoo

 

Lakes

 

 

 

Secchi Depth (m)

 

Lakes

 

 

NPOC concentration (mg/L)

0
N
uonenueouoo

O

O

 

Lakes

 

 

118

Figure 4- 1 Bar graphs of SD, CHL, NPOC and TSS

 

Green algae biovolume (um3)

3.0E +07

 

 

Lakes

0E+07
Lakes

01.0907
0 0E+00
Q

N
uopanueauo

 

 

 

Dlatom biovolume (um3)

 

r v v v v

Bluegreen algae biovolume (pm3)

_-!—_----l.
v v v v v v

 

«"9

¢©
2 06407
156+07
OE+07
5 06406
0 0E+00

$9

nomad-ewes

 

 

Q
Lakes

 

 

EOE+07

4.0E+07 +

52 05+07 (
1.0907 .-
o oe+oo

Q

 

 

119

Figure 4-2 Bar graphs of diatom, green and bluegreen algae

3803.0: 02 m2 300038 00.. mm. 30003.00 02 m0.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

5 5 4 3. 2 1. D 7. 3 5. 4. 3 2 1. O. 6 5 4 3
. . . . 2 1. 0.
0 0 0 0 0 O 3% INIIEITWLIIUW 0 0 O 0 0 0 00
ﬂ . H w - w
I l W N
.. a
_H w. m R
C I I m P I w m
N U
_ m
I w 3803.00 00. m0.
. _ .0. . s 2 o a 6. 4. 2 0.
$000
H . 0 0 0 0 0 0 0 0 _ 0 .
0 o O 0
E30 .500 .0000
I 1
300030: :0. mm: 3003.0: um. m2 3803.0: 02 mm. ..
S. 4. 3. 2. 1. 0 8. 7. 6. 5. a” 3. 2. 1. 0. 6 S 4 3 2 1 0
0 Lu 0 D 0 0 O O D 0 p 0 O 0 0 0 D. 0 0 0 .w
1 a a a a 8 a J a a a q a 7 J1 q q . a n — h n n u
. 7 r m m m m w w m m o
I 6 «:30
1 $ 0
I l
I 5 I 2 N
H I m E
. 4 m a n m
E T
. 3 s . w o
. m
I —1 2 I — m h
I. _ 1 I . 0 —
0 h. m. 5 00 . . _ 0 0 — r p
:3
. o .0000 ESQ

30

DIATOM

Figure 4-3 Histogram distributions of the raw water quality parameters
120

20

10

 

32633 no. on:

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

3 2 1. D
0 0 0 3
I — ‘
.. T ..
.I 2
- . _
5 O S 00
.500
$3903.03 UR m2
3. 2 1. 0.
O 0 0 3
, 3
I ﬂ 2
l 0
. . . 1
6 2 8 4 0 .
.Soo

LN(CHL)

LN(SECCH|)

 

383:5 no. m2

3
0

2
o

LN(NPOC)

 

 

 

 

 

 

 

 

 

 

- b b .
o 5 0 5 o
2 4| 1
E30
3363203 Ba mmq
5. 4. 3. 2 1” D
0 O O O o 3
a d d d d
1 k
n n b n
5 0 s o 5 0
2 2 1 1
3:00

3
LN(TSS)

333.8 8. mm-

6543210.
3

 

 

 

 

 

 

 

 

 

 

 

 

 

 

o. a a o. a a o
- - - - - - 2
I — 5
4|
I o
n n 5
0 o o 0
3 2 4|
E300
$8836: uoq m2
3. 2. 1. 0.
I 7
gm 1
I 6
1 1
r — — 5
I F ‘
1
I 3
1
b n 2
5 O 5 01
4| 1
2.500

 

 

 

 

 

 

 

 

M
E
F.
R
G
E
U
L
m.\
m 3838: Sq we.
4. 3. 2 1. 0.
0 0 0 0 3
A d a a
_
_
.. J
. p F
m w m s o
m E300
E
E
R
w
m

20

1S
LN(DIATOH)

10

Figure 4-4 Histogram distributions of LN transformed water quality parameters

121

Table 4- 1 Pearson correlation matrix of water quality indicators

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

8D CHL T88 NPOC DIA‘I‘OM GREEN BLUEGREEN
8D 1.000
CHL -0.546 1.000
T88 -O.463 0.377 1.000
NPOC -0.151 0.050 0.357 1.000
DIATOM -0.l35 0.042 0.305 0.720 1.000
GREEN -0.388 0.393 0.370 0.554 0.653 1.000
BLUEGREEN -0.343 0.268 0.284 -0.069 -0.125 0.137 1.000
f
O
0
Lu
(0
gru-
n
.1
I
0
O)
O)
.—
O
O
2
I
O
'2
5
1 .—
0 a
2
g n
O
9 11.. ..
Z all n u
LIJ
E
s l) g: )
D B
a .- =£~ ° - I‘
SECCHI CHL TSS NPOC DIATOM GREEN BLUEOREEN

Figure 4-5 Pearson correlation graphs and histogram plots of water quality

indicators

122

4.2 Optimal Spectral Bands

Two types of optimal spectral wavebands were identiﬁed, the region-wavebands
and the narrow—wavebands (Tables 4-2 and 4-3). Region-wavebands were
determined using the spectral derivative spectroscopy method. Important
absorption and reﬂectance features that associated with optical water quality
variables were extracted from the ﬁrst derivative curves. Narrow-wavebands
were determined using Principle Component Analysis together with the results
from the region-waveband determination. PCA identiﬁed three principle
wavebands that explained over 90 percent of variation in the 500 wavebands.
One of the waveband was suspected to cause by atmospheric noise and
therefore excluded from the ﬁnal optimal waveband set. The wavebands at the
lower edge of the region-wavebands were added to the narrow-waveband set
because these wavebands represented the maximum peak or trough in water

spectral signatures.

Table 4-2 Identiﬁed region-wavebands

 

 

 

Sgolay dbl s8 Biophysical characteristics

A:435-475 A:435-485 A:435-475 BLUE absorption

Bz470-570 8:480-575 Bz490-570 GREEN reﬂectance

C:565-620 C:550-630 C:560-610 GR / RED edge

D:670-690 D:670-700 D2660-700 RED max chl a absorption for
Low trophic waters

E:675-700 E:660-705 RED max chl a absorption for
High trophic waters

F:690-740 Fz690-745 Fz690-740 NIR peak for Low trophic waters

G:700-740 G:700-745 Gz700-74O NIR peak for High trophic waters

 

 

 

 

123

 

Table 4-3 Identiﬁed narrow-wavebands

 

 

Narrow-band Wavelength Biophysical characteristics
A 435 BLUE absorption
B 455 BLUE absorption from PCA
C 470 Beginning of GREEN reﬂectance
D 565 GREEN and RED edge
E 670 RED max chl a absorption for

Low trophic waters
F 675 RED max chl a absorption for
High trophic waters

690 NIR peak for Low trophic waters

 

 

 

H 700 NIR peak for High trophic waters

 

4.3 Spectral Indicators

Several optimum spectral indicators were developed to relate water quality
variables with the spectral information of the same site. These indicators may
be used to assess water quality parameters where adequate remote sensing

data is available without the needs for in situ water samples.

4.3.1 Spectral Indicators for Chlorophyll a

Spectral wavebands were determined using PCA and derivative methods
(detailed description in Section 3.4.5 and 3.4.7) and are summarized in Tables 4-2
and 4-3. The selected wavebands from the three ﬁlter methods were regressed

against LN(CHL). The ﬁnal models were compared in Tables 4-4 to 4-5.

124

 

Table 4-4 Chlorophyll a area models using the optimal wavebands

 

 

 

 

 

 

 

 

 

Filter Band Coeﬂ'icient Tolerance P(2-tail) R2 Adjusted R2
Sgolay AG -1.687 0.472 0.000 0.787 0.757
DG 1.509 0.175 0.028
FE -0.331 0.708 0.003
GD 1.132 0.188 0.005
CONSTANT 3.234 0.003
dbl AC 7.263 0.874 0.000 0.710 0.680
AF -3.081 0.935 0.000
DF 2.166 0.882 0.010
CONSTANT 2.160 0.064
s8 AG -3.268 0.912 0.000 0.765 0.749
CE -1.304 0.912 0.005
CONSTANT 7.352 0.000

 

Table 4-5 Chlorophyll a height / width models using the optimal wavebands

 

 

 

 

 

 

 

 

 

Filter Band Coefficient Tolerance 112-tail) R2 Adjusted R2

Sgolay HWAG -1.930 0.800 0.000 0.7 18 0.699
HWBD 5.181 0.800 0.013
CONSTANT 5.891 0.000

dbl HWCF -4.891 0.592 0.000 0.599 0.572
HWFB -1.881 0.592 0.006
CONSTANT 10.042 0.000

s8 HWAG -2.244 0.787 0.000 0.709 0.690
HWCD —1.570 0.7 87 0.005
CONSTANT 6.139 0.000

 

125

 

 

 

 

 

 

The regression models selected 2 — 4 variables out of 49 variables (7 single—
bands and 42 ratio-bands). The wavebands that were selected most repeatedly
are A (blue absorption) and G (NIR peak for high trophic). These two
wavebands link directly to the optical characteristics of CHL as reviewed in
Section 2.2. This regression result proves that the narrower region-waveband G
(700 nm — 740 nm, which represents the waveband region of higher trophic
lakes) correlated better to the CHL than the broader region-waveband F (690
nm - 740 nm, which represents lower trophic lakes). Bands D and E also
appeared in the models, but they do not contribute as large an inﬂuence as
bands A and G do (consider variable coefﬁcients, e.g., larger positive or negative
coefﬁcient value suggested a larger inﬂuence of the variable in the model). Both
band D and E represent the regions of maximum CHL absorption in the red
wavelengths. They could be used to differentiate CHL in water because higher
CHL should produce a deeper absorption feature (valley) in this region. It is
notable that the shorter wavelength band D (670 nm - 690 nm) was selected
more frequently than the longer wavelength band E (675 nm — 700 nm) in the

region-waveband indices.

126

Table 4-6 Chlorophyll a narrow-waveband models using the optimal wavebands

 

 

 

 

 

 

 

 

 

 

Filter Band Coefficient Tolerance P(2-tail) R2 Adjusted R2
Sgolay A -38.605 0.264 0.000 0.785 0.754
E 42.536 0.125 0.004
EH -5.108 0.367 0.000
GA -1.549 0.102 0.025
CONSTANT 9.801 0.000
dbl A -18.323 0.649 0.011 0.709 0.679
H 19.859 0.469 0.004
F0 4566 0.513 0.003
CONSTANT 7.414 0.000
s8 A -24.190 0.731 0.001 0.723 0.694
E 17.116 0.488 0.017
EG -3.598 0.605 0.000
CONSTANT 7 .239 0.000

 

 

 

The narrow-wavebands that were selected most frequently are A (435 nm, blue
absorption), E (670 nm, maximum red absorption), and G (690 nm, NIR peak).
Band A and G consistently performed well in the regression models for both
narrow-waveband and region-waveband indicators. Band E was expected to
appear to improve the prediction power of the model. It is noticed that between
band G (690 nm) and H (700 nm), which both represent NIR peak of lower and
higher trophic waters, band G was selected more often and had a larger

inﬂuence in the models.

127

4.3.2 Spectral Indicators for Secchi Depth
Region-wavebands (area and height / width) indicators and narrow-wavebands

reﬂectance indicators from three ﬁlter methods were regressed against LN (SD).

The ﬁnal models were compared in Tables 4-7 to 4-9.

Table 4-7 Secchi depth area models using the optimal wavebands

 

 

 

 

 

 

 

 

 

Filter Band Coefficient Tolerance P(2-tail) R2 Adjusted R2
Sgolay AB -5.560 0.473 0.004 0.736 0.700
AG 0.792 0.352 0.001
BF -2.183 0.575 0.029
EC -0.937 0.307 0.009
CONSTANT 3.955 0.012
dbl BD -0.252 0.636 0.006 0.638 0.615
DC -8.735 0.636 0.000
CONS 4.195 0.000
SS C —36.227 0.166 0.000 0.718 0.679
BC -0.858 0.167 0.000
BE -1.418 0.453 0.001
FD 1.69 0.378 0.000
CONSTANT 2.888 0.000

 

128

 

 

 

Table 4-8 Secchi depth a height / width models using the optimal wavebands

 

 

 

 

 

 

 

 

Filter Band Coefﬁcient Tolerance P(2-tail) R2 Adjusted R2
Sgolay HWC -47.187 0.356‘ 0.000 0.781 0.759
HWBA —1.742 0.426 0.000
HWGC -0.597 0.762 0.000
CONSTANT 4.559 0.000
dbl HWBC -7.035 0.108 0.000 0.596 0.570
HWCB -1.912 0.108 0.027
CONSTANT 9.084 0.001
SS HWC -23.473 0.249 0.049 0.741 0.716
HWAG 0.909 0.537 0.001
HWBC -1.297 0.182 0.002
CONSTANT 1.465 0.078

 

 

 

 

 

The wavebands that were selected most repeatedly are band B (green
reﬂectance) and C (green and red edge). These two wavebands together covered
the entire green reﬂectance peak, which is the range that has the highest
variation in reﬂectance (Figure 4-6). Savitsky Golay models produced the
higher R2 for both area and height / width methods, suggesting that 70 percent
of the SD can be explained by the spectral indicators. Band A (blue absorption)
also appeared in Sgolay and s8 models. This band can measure clarity of the
water because it could penetrate deeper than other band regions (Figure 4-6).
SD mostly related to overall reﬂectance of all particles suspended in the water
column. Unlike CHL, SD does not need detail absorption / reﬂectance features

to detect the differences among lakes.

129

 

‘ Sgolay Filtered Volume Reﬂectance

 

‘

,, .-.,, - -E
r" J 4_,v:_h“\‘\\‘7:

Reﬂectance
.0
8

 

 

 

   

 

 

Wavelength

 

Figure 4-6 Volume reﬂectance curves

When looking at the volume reﬂectance within the waveband B and C regions, it
was observed that the spectral signatures did not arrange from low SD to high
SD (Figure 4-7). However, the 1at derivative curves shows a clear pattern of SD
arranged from low to high (Figure 4—8). Derivative processes in this algorithm

improve the ability to relate remote sensing spectra to water quality parameters.

130

 

Sgolay Volume Reﬂectance

 

 

 

 

 

 

   

 

 

 

 

 

0.10
.‘i. ’H'.-'-“'.
. ’ "\
0.08 ~ 7 - -/ ——i ‘7 ~—
g ,.‘ / \‘0‘
“r3 ,/
> 0.06 '- Jr ———~ k * —-§
': ...--’
o -‘ ‘2 ....\¢"°.“W k u
C: ,w—m‘“” ‘_“
13 0.04 ._,..._.--_w-— ' -
F
— - -Poly. (Hicks 1 m) 7 __ ,
0'02 ——Poly. (Chipewa 2 m)
— - -Poly. (Cubs m)
—Pol . Kli eer
0.00 y ( n9 ) r .
470 520 570 620

 

Wavelength

 

 

Figure 4-7 Examples of volume reﬂectance curves on different SD in B and C

waveband regions

 

Sgolay 1st Derivative

 

 
 
   

 

 

 

 

 

 

 

0.0010
0 0.0005 "’
>
a
E
o .
0 0.0000
“
a
F
‘0'0005 ‘ — - -Poly. (Hicks 1 m)
-—-—-Poly. (Chipewa 2 m)
_ . .Poly. (Cub5m) . ‘
-o.oo1o —‘—'P°'Y-(K""9°'8m) , . _.
470 520 570 620
Wavelength

 

 

Figure 4-8 Examples of ﬁtting curves on diﬂ'erent SD in B and C waveband

regions

131

Table 4-9 SD narrow-waveband models using the optimal wavebands

 

 

 

 

 

 

 

 

 

Filter Band Coefficient Tolerance P(2-tail) R2 Adjusted R2
Sgolay A -7. 1088 0.737 0.020 0.796 0.768
DA 1.1003 0.1 15 0.001
DC - 1.8842 0.102 0.000
GC -1.0169 0.583 0.000
CONSTANT 2.871 0.000
dbl E -11.875 0.289 0.000 0.877 0.855
BA 6.929 0.386 0.000
ED -0.791 0.558 0.000
GD - 1.36 0.291 0.000
GH -2.212 0.258 0.004
CONSTANT -1.1 15 0.304
s8 D -10.0724 0.387 0.000 0.865 0.841
AB -8.5365 0.313 0.000
BD 0.8514 0.439 0.017
DE -0.2994 0.192 0.01 1
EC —1.7623 0.159 0.000
CONSTANT 11.172 0.000

 

 

 

 

The R2 are very high for all narrow-waveband SD models. Similar to the region
wavebands models, the narrow-waveband models also have band A (435 nm).
Narrow-waveband models tend to select band G (690 nm, NIR peak for lower

trophic waters) over band H (700 nm, NIR peak for higher trophic waters).

132

4.3.3 Spectral Indicators for TSS

Region-wavebands (area and height / width) indices and narrow-wavebands

volume reﬂectance indicators from three ﬁlter methods were regressed against

LN(TSS). The ﬁnal models for TSS were compared in Tables 4-10 to 4-12.

Table 4-10 TSS area models using the optimal wavebands

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Filter Band Coefficient Tolerance P(2-tail) R3 Adjusted R2

Sgolay DA 0.420 0.878 0.019 0.351 0.286
ED 1.088 0.983 0.017
GE 0.295 0.870 0.047
CONSTANT 0.405 0.498

dbl DC 3.171 1.000 0.001 0.286 0.264
CONSTANT 1 .748 0.000

88 EC 0.179 0.996 0.037 0.283 0.237
GF 2.344 0.996 0.015
CONSTANT 1.375 0.000

Table 4- ll TSS height/width models using the optimal wavebands

Filter Band Coefficient Tolerance P(2-tail) R2 Adjusted R2

Sgolay HWGC 0.213 1.000 0.004 0.231 0.207
CONSTANT 2.273 0.000

dbl HWEC 0.599 1.000 0.009 0.197 0.172
CONSTANT 1.262 0.010

s8 HWB 39.605 0.977 0.022 0.267 0.220
HWCA -0.874 0.977 0.012
CONSTANT 2.288 0.000

 

 

133

 

 

 

 

 

 

Table 4- 12 TSS narrow-waveband models using the optimal wavebands

 

 

 

 

 

 

 

 

 

 

Filter Band Coefficient Tolerance P(2-tail) R2 Adjusted R2

Sgolay FH - 1.224 1.000 0.000 0.346 0.326
CONSTANT 3.876 0.000

dbl FH -1.073 1.000 0.001 0.318 0.296
CONSTANT 3.737 0.000

s8 E 8.3417 0.344 0.034 0.532 0.468
BC 3.1053 0.444 0.002
DF 0.2785 0.213 0.015
EC 1.034 0.264 0.001
CONSTANT - 1.9603 0.102

 

 

 

TSS models have low R2 overall. Wavelet s8 produces better models in some
cases but the reported R2 are not much higher than Sgolay method. Several
waveband regions were selected in the models including band D, E and C.
These are the band range (565 nm - 700 nm) that varied the most when

suspended sediments were added to the water (Figure 2-9).

The narrow-waveband models selected band F (675 nm) and H (700 nm). These
are the maximum CHL absorption band and NIR peak band, which have been
used extensively in vegetation indices. The combination of these two bands
should indicate CHL. It is possible that these wavebands were selected because
most of the TSS in the data set may be dominated by algal biomass, which also
refers to CHL. TSS and CHL concentration distributions are actually look alike
(Figure 4- 1). However, correlation coefﬁcients of TSS models are not as high as

CHL. This may be because TSS lab analysis used the simple mass

134

measurement, which could encounter high error when TSS concentration is

low. On the other hand, CHL analysis uses equipment that is very sensitive to

CHL concentration.

4.3.4 Spectral Indicators for NPOC

The ﬁnal models for NPOC were compared in Tables 4— 13 to 4-15.

Table 4-13 N POC area models using the optimal wavebands

 

Filter Band

Coefficient Tolerance P(2-tail)

R2 Adjusted R3

 

 

 

 

 

 

 

 

Sgolay EC* 0.485 1.000 0.052 0. 116 0.087
CONSTANT 1.882 0.000

dbl EA* 0.262 1.000 0.076 0.098 0.069
CONSTANT 1 .514 0.001

88 EC* 0. 194 1.000 0.067 0.104 0.075
CONSTANT 1 .843 0.000

 

*Insigniﬁcant variable

Table 4-14 NPOC height / width models using the optimal wavebands

 

 

 

 

 

 

 

 

 

Filter Band Coefficient Tolerance 'P(2-tail) R2 Adjusted R2
Sgolay HWEC 0.100 1.000 0.037 0.133 0.105
CONSTANT 1.902 0.000
db 1 HWEC 0.602 1.000 0.020 0.164 0. 137
CONSTANT 0.945 0.082
s8 HWAC* 0.177 1.000 0.053 0.115 0.087
CONSTANT 1 .848 0.000

 

*Insigniﬁcant variable

135

 

 

 

 

 

 

Table 4- 15 NPOC narrow-waveband models using the optimal wavebands

 

 

 

 

 

 

 

 

 

 

Filter Band Coefficient Tolerance P(2-tail) R2 Adjusted R2|
Sgolay CH* -0.0872 1.000 0.12 1 0.07 6 0.046
CONSTANT 2.4076 0.000
dbl DF" -0.1206 1.000 0.124 0.075 0.045
CONSTANT 2.4426 0.000
s8 CE* -0. 1243 1.000 0.103 0.084 0.054
CONSTANT 2.4441 0.000

 

 

 

*Insigniﬁcant variable

Most of the NPOC models are not valid and one that is valid has a very low R2
because there are very few high NPOC lakes in the data set. Although the
model was performed on LN(NPOC), which helped transformed the highly
skewed data to a normally distributed data set, a wide range of NPOC data was
lacking in the regression model development. Another reason could be that
NPOC absorb incident light so much that the measured spectral signature of
the two high NPOC lakes was so low and no spectral feature could be extracted
from them. Figure 4-9 compared spectral signature of the highest NPOC Croton

dam pond with the two other similar water condition lakes but lower NPOC.

It was almost consistent that band E and C have been selected very often, both
in signiﬁcant and insigniﬁcant models. NPOC absorb light signiﬁcantly in
shorter wavelengths; therefore, the useful wavebands (if there is one) should be

located in longer wavelengths, such as band D to G (Figure 2-7 and 2- 10).

136

 

 

0.12

 

Croton (NPOC=36mg/L; Cl-I.=27.19mg/L; SD=1.70m)
— - - Kent (NPOC=9.47mg/L; CHL=31.59mg/L; SD=1.2m)
- - - -Kimbal| (NPOC=7.67mg/L; Cl-I.=36.22mg/L; SD=1.75m)

’.

 

 

 

-----

 

 

 

o.oa~——--———-— - — __

 

Volume Reﬂectance

0.04 As ,. ' ~f

 

 

 

 

0.00

400 500 600 700 800 900
Wavelength

 

 

 

Figure 4-9 Comparison between high NPOC lake (Croton dam pond) and other

lakes with similar water condition but lower NPOC

4.3.5 Spectral Indicators for Diatom

Region-wavebands (area and height / width) indices and narrow-wavebands
volume reﬂectance indices from three ﬁlter methods were regressed against
LN(DIATOM), LN(GREEN) and LN(BLUEGREEN). The ﬁnal models for diatom

were compared in Tables 4-16 to 4- 18.

137

Table 4- 16 Diatom area models using the optimal wavebands

 

 

 

 

 

 

 

 

 

 

Filter Band Coefficient Tolerance P(2-tail) R2 Adjusted Ra
Sgolay AD 2.670 0.322 0.048 0.280 0.208
DC 10.127 0.186 0.003
GC -3.522 0.345 0.004
CONSTANT 5.638 0.030
dbl BD -0.67 0.904 0.069 0.289 0.190
CA* -4.16 0.125 0.087
FA 3.852 0.148 0.020
GC -17 .065 0.336 0.022
CONSTANT 20.061 0.000
s8 BA 3555 0.633 0.021 0.160 0.106
CG* -1.828 0.633 0.113
CONSTANT 19.298 0.000

 

 

 

*Insigniﬁcant variable

Table 4-17 Diatom height / width models using the optimal wavebands

 

 

 

 

 

 

 

 

 

 

Filter Band Coefficient Tolerance P(2-tail) R2 Adjusted R2

Sgolay HWEC 0.752 0.617 0.005 0.264 0.216
HWGA -1.729 0.617 0.005
CONSTANT 1 1.663 0.000

dbl HWD -26.578 0.230 0.01 1 0.203 0.151
HWED -12.835 0.230 0.011
CONSTANT 27.574 0.000

s8 HWAC* 0.733 1.000 0.085 0.090 0.062
CONSTANT 10.126 0.000

 

 

 

*Insigniﬁcant variable

138

Table 4— l8 Diatom narrow-waveband models using the optimal wavebands

 

 

 

 

 

 

 

 

 

 

Filter Band Coefficient Tolerance P(2-tail) R2 Adjusted R2

Sgolay AD 4.718 1.000 0.004 0.230 0.206
CONSTANT 8.474 0.000

dbl FG —20.348 0.177 0.006 0.248 0.200
HF - 12.028 0. 1 77 0.003
CONSTANT 43.596 0.000

s8 AC 7.654 1.000 0.006 0.212 0.187
CONSTANT 5.059 0.034

 

 

 

There were only a few diatom dominated lakes in the data set. Almost all of
these few lakes that had high diatom biomass also had high green algae
biomass. Therefore, the diatom dominated signatures to use in the model
development was lacking. The R2 of all models were low; however, Sgolay
produced better models than other methods in every case. Signiﬁcance of the
models was determined based on Pvalue (P S 0.5; two-tailed test). When P
value of a variable was higher than 0.5, the model was considered invalid and
the variable that was most insigniﬁcant was eliminated from the regression
analyses. The regression model was performed again based on the remaining
variables. In some cases, other variables became signiﬁcant after the ﬁrst
variable was removed and the model became signiﬁcant. Otherwise, the
variable that had the next highest insigniﬁcant Pvalue was removed and the
regression model was performed again until all remaining variables produced a

signiﬁcant valid model.

139

The region wavebands that were selected repeatedly were D (maximum
chlorophyll absorption in red) and C (green and red edge). Waveband C (565
nm - 620 nm) could represent the general CHL curve, which is the major
pigment of all algae. According to Figure 2-11, The diatom spectral signature
has the prominent trough around 670 nm - 690 nm, which is the region
waveband D. The height/width model did not have band D, but it had band E
which was very similar (67 5 nm - 700 nm). Other bands that were selected,

band A and G, were the bands that were found to be the indicators of CHL.

The narrow-waveband models selected waveband in different region. According
to Figure 2-11, band D (565 nm) selected by Sgolay method was the distinctive
but very small valley appeared in diatom curve. This trough would be difﬁcult
to identify the region waveband indicators. Wavelet dbl method did selected
band F (67 5 nm), which is the exact same prominent trough that showed in the
region waveband models. However, Sgolay produced a slightly better result,

therefore, it was used in validation process.

4.3.6 Spectral Indicators for Green Algae
Examples of green algal dominated lakes are shown in Figure 4-10, and the

final models for green algae were compared in Tables 4- 19 to 4-21.

140

 

 

 

 

 

 

 

 

 

 

 

 

Green Algae Signatures
0.12
Belleville (Green=16.805,432: Bluegreen=1.254.403; Diatom=274.486)
— — Muskegon (Green=7.543.489; Bluegreen=276.293; Diatom=143.672)

g - . . -Jehnsen (Gneen=4.203.119; Bluegreen=0; Diatom=308.218)
g 0.09 +--— _ v + ‘ —
5
a:
g 0.06
>
3
:13
C
5' 0.03
.3:

O.m l T I l 7 T

400 450 500 550 600 650 700 750
Wavelength

 

 

 

Figure 4- 10 Examples of volume reﬂectance of green algal dominated waters

Table 4-19 Green algae area models using the optimal wavebands

 

 

 

 

 

 

 

 

 

Filter Band Coefficient Tolerance P(2-tail) R2 Adjusted R2

Sgolay AD 1.642 0.190 0.047 0.328 0.285
EA 2.266 0.190 0.002
CONSTANT 9.847 0.000

dbl DA 2.734 1 .000 0.008 0.203 0.178
CONSTANT 13.071 0.000

s8 8* -39.786 0.171 0.081 0.231 0.154
AB -6.883 0.167 0.019
GB 7.548 0.423 0.016
CONSTANT 19.238 0.000

 

*Insigniﬁcant variable

141

 

 

 

Table 4-20 Green algae height / width models using the optimal wavebands

 

 

 

 

 

 

 

 

 

 

Filter Band Coefficient Tolerance P(2-tail) R2 Adjusted Ra
Sgolay HWEA 0.375 1.000 0.006 0.2 17 0. 192
CONSTANT 13.343 0.000
dbl HWAF* -1.267 1.000 0.064 0.103 0.075
CONSTANT 16.098 0.000
s8 HWAF -4.772 1.000 0.013 0.177 0.151
CONSTANT 1 7 .84 0.000

 

 

 

*Insigniﬁcant variable

Table 4-21 Green algae narrow-waveband models using the optimal wavebands

 

 

 

 

 

 

 

 

 

Filter Band Coefiicient Tolerance P(2-tail) R2 Adjusted R2
Sgolay A -28.2 0.384 0.022 0.372 0.309
D 34.403 0.210 0.001
AB 14.341 0.406 0.000
CONSTANT -0.01 0.998
dbl BA -9.489 0.738 0.002 0.421 0.363
BC -78.591 0.740 0.006
GF 4.002 0.985 0.002
CONSTANT 99.102 0.001
s8 A -28.015 0.377 0.021 0.429 0.350
H 56.429 0.142 0.001
AC 3.854 0.758 0.007
EA -2.113 0.156 0.013
CONSTANT 12.596 0.000

 

142

 

 

 

Area and narrow-waveband models tend to perform better than the

height /width model for green algae. All the models produce relatively low R2;
therefore, they could not be compared with a high conﬁdence but it is observed
that height / width may not be as sensitive indicator to green algae as area does.
Due to the 2nd order polynomial ﬁtting method, lakes with different algal
biomass may produce ﬁtting curves with different area but the same height.

Figure 4- 1 1 demonstrated how height may be less sensitive to changes in

biophysical parameters than area.

 

 

 

 

 

 

 

 

 

 

 

 

o Spectrum1 a Spectrum2
Poly. (Spectrum1) — - -Poly. (Spectrum2)
0.0009
3 0.0006 ~
‘6
c
2
e
a
3 0.0003 «
1
0 l l r T -
480 500 520 540 560
1st Derivative

 

 

Figure 4-11 Examples of spectra that produce the same height but different

under curve area

Although Sgolay did not produce the best models in every case, it was selected
as the method to validate because it produced better results overall. Different

ﬁlters to use for the same water quality parameter cause a difﬁculty in

implementation.

143

The region-waveband that was selected most repeatedly was band A (blue
absorption). Either or both band D and E (the maximum CHL absorption
bands) were also appeared in every model. These wavebands are the indicators
of chlorophyll pigment, which is the major component in green algae.

The narrow-waveband models selected band A (435 nm) and B (455 nm) most
often. Both region-waveband and narrow-waveband model selected band A,

which is the blue absorption. This band may be a good indicator of green algae.

4.3.7 Spectral Indicators for Bluegreen Algae
Examples of bluegreen algal dominated lakes is showed in Figure 4-12, and the

ﬁnal models for bluegreen algae were compared in Tables 4-22 to 4-24.

 

 

 

 

 

 

 

 

 

 

Bluegreen Algae Signatures
0.15
Paw Paw (Bluegreen=14.049.906; Green=2.001.618; Diatom=160.862)
— — Pickerel (Bluegreen=12,999.799; Green=381.352: Diatom=23.920)
g - - - .Hicks (Bluegreen=9.445,346; Gmen=4.630,117; Diatom=3.124)
i
“5 0.10
a:
o
E
2
o
>
3.:
z 0.05 .
t:
E
i
0.00 I l I T I l
400 450 500 550 600 650 700 750
Wavelength

 

 

 

Figure 4-12 Examples of volume reﬂectance of bluegreen algal dominated

waters

144

The ﬁrst observation from the reﬂectance curve alone shows that the slight
bumps in blue wavelengths around 435 nm — 47 5 nm are noticeable. This blue
reﬂectance pattern did not appear clearly on the green algae curves (Figure 4-

10) except for Belleville Lake, which also has a high bluegreen algal biomass.

Table 4-22 Bluegreen algae area models using the optimal wavebands

 

 

 

 

 

 

 

 

 

Filter Band Coefficient Tolerance P(2-tail) R2 Adjusted R2
Sgolay AG -3.123 0.646 0.000 0.634 0.580
CA 1.977 0.756 0.005
DB -66.849 0.788 0.000
DE -7.727 0.811 0.044
CONSTANT 28.569 0.000
Dbl DB -22.962 0.644 0.013 0.200 0.144
DC* 12.158 0.644 0.066
CONSTANT 16.569
s8 D -192.032 0.618 0.001 0.370 0.326
BA 5.106 0.618 0.001
CONSTANT 18.946 0.000

 

*Insigniﬁcant variable

145

 

 

 

 

Table 4-23 Bluegreen algae height / width models using the optimal wavebands

 

 

 

 

 

 

 

 

 

Filter Band Coefficient Tolerance P(2-tail) R2 Adjusted R2
Sgolay HWF 133.395 0.542 0.001 0.436 0.353
HWAF -35.544 0.212 0.001
HWBF 69.307 0.495 0.001
HWDA -1.191 0.239 0.004
CONSTANT 1.266 0.834
Dbl HWBE 20.307 1.000 0.008 0.213 0.187
CONSTANT 6.387 0.032
s8 HWDB -3.233 1.000 0.000 0.373 0.352
CONSTANT 23.298 0.000

 

Table 4-24 Bluegreen algae narrow-waveband models using the optimal

 

 

 

 

 

 

 

 

wavebands
Filter Band Coefficient Tolerance P(2-tail) R2 Adjusted R2

Sgolay BE -2.107 0.405 0.000 0.364 0.320
HC -2.7 76 0.405 0.021
CONSTANT 19.926 0.000

Dbl BF -2.328 0.383 0.000 0.569 0.505
GH - 1 5.242 0.204 0.006
HE -7.000 0.22 1 0.009
HF -7.160 0.209 0.015
CONSTANT 44.931 0.000

s8 BE —1.429 0.961 0.000 0.477 0.441
EF 10.362 0.961 0.002
CONSTANT 5.577 0.088

 

 

146

 

 

 

 

 

 

 

Again, area and narrow-waveband models perform better than the height/width
model for bluegreen algae (discussed in Section4.3.6; Figure 4-11). The models
produce fairly good R2, approximately 50 - 60 percent of the bluegreen algal
biomass could be predicted by the area and the narrow-waveband spectral
indicator models. There were not as many toxicity lakes that were dominated
with bluegreen algae comparing to green algae. Nevertheless, the models select
spectral regions that could detect bluegreen algae. Giving a potential toxic of
this type of algae and the importance to assess and monitor for it in the real
ecosystem environment, 50 percent chance of detecting it with remote sensing
technology is useful. This hyperspectral remote sensing algorithm does not
require as much time as the traditional way of taking samples back to the lab
analysis. It allows a frequent monitoring by a hand-held sensor, or even a
spatial overview with an airborne or a satellite sensor. Hyperspectral satellite
sensors suitable for inland water monitoring are currently unavailable.
However, research on locating waveband regions and spectral band width such

as this study would help facilitate the future sensor conﬁgurations.

The region wavebands selected most frequently were band A (435 nm - 475 nm,
blue absorption) and D (670 nm — 690 nm, maximum red absorption). Band A
is a good indicator of blue green algae. It separates bluegreen algae from other
algae type because bluegreen algae reﬂect light in blue wavelength regions,
whereas other algae absorb light in these regions (Figure 4-10 and 4- 12).
Waveband D is an indicator of chlorophyll maximum absorption. It is expected

to present in the models because most of the lakes that were dominated by

147

 

bluegreen algae had relatively high trophic. Waters that have high chlorophyll
would have a clear prominent absorption feature in band D. The narrow-
waveband models selected band B (455 nm) and E (670 nm), which are in the
ranges of band A and D in the region-waveband indicators. The narrow-
waveband indices show that wavelength 455 nm, which is close to the center of
the slight bump in blue wavelengths, is a good indicator of bluegreen algae.
Band H appears in the models as it is an indicator of higher CHL spectral

signature.

4.4 Model Validation
The spectral indicators for each water quality variable were validated with 14

samples from both ASD and AISA data set.

4.4.1 Chlorophyll a Model Validation

AISA sensor was ﬂown on 14 lakes on July 26, 2004. Of all the 14 lakes, ASD
data are available for 13 of these lakes. Most of the measurements were made
within 3 days of the AISA ﬂy-over. One of these 13 lakes, Marl Lake, was
extremely shallow. The deepest basin of the lake was only 1.5 meters and it
was impossible to launch the boat to collect the sample. ASD and water
samples were collected at the clock, which did not fall within the path of the
AISA image. The ASD reading was impacted largely by the bottom effect.
Spectral signature of this site did not purely represent inherent constituents

within the water column. Therefore, it was appropriate to exclude the lake from

148

the validation data set. Consequently, the validation data set contained 12 ASD

and 12 AISA spectra of the same lakes.

4.4.1.1 Chlorophyll a ASD Validation

The best spectral indicator for CHL was from Sgolay method. This method
consistently performed better than the other approaches, and was therefore
selected to use in the validation process. Although the area indicator tended to
produce better results than the height / width index and narrow-waveband index
in general (in terms of correlation coefﬁcient and distribution of residual), all

three indices were validated.

The ASD radiance spectra from 12 lakes were converted to volume reﬂectance
and ﬁltered with Savitsky Golay method. The ﬁltered volume reﬂectances were
then calculated for the 1st derivative. The derivative products were separated
into narrow-wavebands and region-wavebands according to Table 4-2 and 4—3.
Second order polynomial functions were ﬁtted to the derivative data of region-
waveband, and area under the curve, height and width of the function were
calculated (detail described in Section 3.4.7 to 3.4.9). These waveband
variables were used in the CHL models (Tables 4-4 to 4-6) to calculate the
predicted CHL concentration for each lake (Table 4-25). The predicted CHL was
then correlated with the real CHL concentration measured from the sampled

water of the same site (Table 4-26; Figure 4-13).

149

Table 4-25 Chlorophyll ASD validation

 

 

 

 

Lake Actual LNchl Predicted LNchl
Area Height] Width Narrow Band
HIGGO727 0.122 0.857 0.096 1.465
CLME0724 2.270 2.505 2.718 2.482
SAPP0725 2.315 2.787 2.800 2.725
ARBU07 26 2.371 1.922 1.882 1.903
MITC0725 2.612 2.919 3.228 2.907
HOUG0726 2.851 2.577 2.641 2.633
ROGE0724 3.001 2.967 3.365 3.179
MUSK0728 3.334 3.626 3.928 3.377
JEHN0519 3.359 2.664 2.890 2.781
HESSO727 3.515 4.901 3.336 4.215
BR000727 3.928 4.199 3.514 3.713
TAMA0605 4.456 3.133 3.304 2.866

 

 

 

 

 

 

 

Table 4-26 Pearson correlation matrix of actual and predicted chlorophyll from
ASD

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Predicted CHL Area Method Height/Width Method Narrow Band Method
Actual CHL 0.782 0.865 0.749
ASD PREDICTED CHL

_l

I

O

_l

< .

D

B

<

AREA HEIGHTNVIDTH NARROW

Figure 4-13 Correlation graphs between actual and predicted CHL from ASD

Correlations between actual and predicted CHL were quite strong for all

methods. In a complex real natural ecosystem, many organic and inorganic

150

constituents coexist at the same time. These substances have different optical
properties that could interfere or mask the signal of CHL in water. Therefore,
the correlation coeﬂicients of 0.75 - 0.87 were considered to be high especially
for the studies that base on the natural environment rather than a controlled
laboratory experimental condition. The CHL models selected spectral bands
that are strongly related to the optical properties of CHL. These spectral bands
were able to differentiate and provide a very good indication of CHL

concentration of various trophic states in Michigan’s inland lake waters.

4.4.1.2 Chlorophyll a AISA Validation

AISA spectra were extracted from the 30 bands AISA imagery using the 3x3
area of interest (AOI). The average reﬂectance values within the AOI were used
as the representative spectra of the lakes. These spectra were converted to
volume reﬂectance. The AISA volume reﬂectances did not need to be ﬁltered
because the high frequency noise was not present. The average bandwidth was
5 nm — 8 nm. Filtering these spectra could result in losing useful

absorption / reﬂectance features. The ﬁltered volume reﬂectance was calculated
for the 18‘ derivative. The derivative was then separated into narrow- and
region-wavebands according to the Sgolay spectral indicator model. Polynomial
functions 2'“d order were ﬁtted to the derivative data for each band regions, and
under curve area, height and width of the function were calculated (detail
described in Section 3.4.7 to 3.4.9). These spectral indicators were used to
predict CHL concentration for each lake (Table 4-27). The predicted CHL was
then correlated with the real CHL concentration measured from the sampled

water of the same site (Table 4-28; Figure 4-14).

151

Table 4-27 Chlorophyll AISA validation

 

 

 

 

Lake Actual LNchl Predicted LNchl
Area Height/Width Narrow Band

HIGG0727 0.122 2.978 0.071 0.905
CLME0724 2.270 2.560 1.493 1.821
SAPP0725 2.315 2.371 2.688 2.438
ARBU0726 2.371 3.201 -0. 181 1 .652
TAMA0726 2.543 2.289 2.852 2.170
MITC0725 2.612 2.461 0.806 2.603
HOUG0726 2.851 2.319 2.917 2.090
JEHN0726 2.866 2.714 2.016 1.849
ROGE07 24 3.001 1.467 4.617 2.657
MUSK0728 3.334 1.969 -0.293 1.204
HESSO727 3.515 1.713 3.689 3.962
BR0007 27 3.928 2.053 3.019 3.591

 

 

 

 

 

 

 

Table 4-28 Pearson correlation matrix of actual and predicted chlorophyll from

AISA

Predicted CHL Area Method Height/Width Method Narrow Band Method
Actual CHL -0.646 0.463 0.668

 

 

 

 

 

 

 

 

AISA PREDICTED CHL

O
O
O
O O
‘ e
O 0 ° .

HEIGHT/WIDTH NARROW

 

   

ACTUAL CHL

 

 

 

 

 

AREA

Figure 4-14 Correlation graphs between actual and predicted CHL from AISA

Correlation between the actual and predicted CHL from AISA data was lower

than expected. Two of the 12 AISA spectra were not reliable. Mitchell Lake

152

image could not be successfully processed in an image export. However, part of
the image that was successfully exported, although far from the in situ sample
location, was used. The reﬂectance value was extracted from that part of the
image. Muskegon Lake contained a large amount of bad scan lines (Appendix
G). It was almost impossible to extract reﬂectance in 3x3 pixel window that was
not affected by the defective scan lines. These two lakes were excluded from the
data in the correlation process. The result shows a signiﬁcant improvement on

height / width and narrow-waveband indicators (Table 4-29; Figure 4- 15).

 

Table 4-29 Pearson correlation matrix of actual and predicted chlorophyll from
AISA without Mitchell and Muskegon Lakes '

 

Predicted CHL Area Method Height/Width Method Narrow Band Method

 

 

 

 

 

 

Actual CHL -0.627 0.685 0.836

 

AISA PREDICTED CHL

 

 

 

 

 

 

ACTUAL CHL

 

AREA HEIGHTNVIDTH NARROW

Figure 4- 15 Correlation graphs between actual and predicted CHL from AISA

without Mitchell and Muskegon Lakes

4.4.2 Secchi Depth Model Validation

Validation is tested on the ASD and the AISA data from the same 12 lakes.

Spectral radiance from both data sets was transformed to surface reﬂectance.

153

 

Air-water interface correction and was applied to the reﬂectance. Three de-
noising ﬁlters were performed on the ASD data, and the 1*!t derivative was
calculated for each lake. The derivative products were then separated into
regions determined in the previous section (section 3.4.7). Under curve area,
height and width rationing were calculated. These indices, as well as the
narrow-waveband reﬂectance values were used in the multivariate regression
models. Since Sgolay produced the better results in the most biophysical
models, it was used to validate the efﬁciency of the models. Although Sgolay
did not produce the best SD model in the narrow-waveband model, the R2 was
not low. For an implementation purpose, it is more practical for the decision
makers to use the models that were calculated from one ﬁlter method to predict
water quality variables rather than applying different ﬁlters to the data for

different variables.

4.4.2.1 Secchi Depth ASD Validation

The ASD radiance spectra from 12 lakes were converted to volume reﬂectance
and ﬁltered with Savitsky Golay method. The ﬁltered volume reﬂectance was
then calculated for the 1st derivative. The derivative products were separated
into narrow-wavebands and region-wavebands according to Tables 4-2 and 4-3
(Sgolay indicator). Polynomial functions 2nd orders were ﬁtted to the derivative
data of region-waveband, and under curve area, height and width of the
function were calculated (detail described in Section 3.4.7 to 3.4.9). These
waveband variables were used in the Secchi models to calculate the predicted

SD for each lake (Table 4-30). The predicted SD was then correlated with the

154

 

real SD concentration measured from the sampled water of the same site (Table

4-31; Figure 4- 16).

Table 4-30 Secchi depth ASD validation

 

Predicted LNsd

 

 

 

Lake Actual LNsd
Area Helm] Width Narrow Band
HESSO727 -0.598 0.203 0.283 1.663
BR000727 0.000 0.475 0.534 1 .010
ROGE0724 0.182 0.396 0.619 1.728
SAPP07 25 0.336 0.902 0.855 2.457
MUSK07 28 0.405 0.363 -0.067 1.302
HOUG0726 0.470 0.805 1.127 1.771
TAMA0605 0.531 0.865 0.623 1.169
MITC0725 0.642 0.574 0.589 1.543
JEHN0519 0.693 0.795 0.880 2.122
CLME07 24 1.065 0.995 1.075 2.1 14
ARBU0726 1.163 1.368 1.183 2.429
HIGG0727 2.054 2.208 1.920 3.258

 

 

 

 

 

 

Table 4-3 1 Pearson correlation matrix of actual and predicted Secchi depth

 

 

from ASD
Predicted SD Area Method Height/Width Method Narrow Band Method
Actual SD 0.918 0.810 0.718

 

 

 

 

 

 

ACTUAL SD

ASD PREDICTED SD

 

 

 

 

 

 

 

AREA

HIGHTIWIDTH

NARROW

Figure 4- 16 Correlation graphs between actual and predicted SD from ASD

155

Correlations between actual and predicted SD were very strong for all methods.
The outlier point was Higgins Lake, which was the clearest lake in the data set
(SD = 7.8 m; CHL = 1.13 mg/ L). Comparing within the Sgolay models, region-
waveband models are more accurate than the narrow-waveband model. The
selected bands were able to differentiate and provide a very good indication of

SD of various lakes in Michigan.

4.4.2.2 Secchi Depth AISA Validation

AISA spectra were extracted from the 30 bands AISA imagery and processed
according to Section 4.4.1.2. Sgolay SD models were applied to the AISA data
set to predict SD for each lakes (Table 4-32). The predicted SD was then
correlated with the actual reading SD from the same sample site (Table 4-33;

Figure 4- 17).

Table 4-32 Secchi depth AISA validation

 

 

 

 

 

Lake Actual LNsd Predicted LNsd
Area Height] Width Narrow Band
HESS07 27 -0.598 3.459 -2.900 2.822
BROOO7 27 0.000 3.949 - 1.588 3.720
ROGE0724 0.182 6.216 -1.769 2.489
SAPP0725 0.336 2.474 1.131 3.330
MUSK07 28 0.405 -0.242 -0.433 1.132
HOUG0726 0.470 2.099 0.893 2.997
TAMA0605 0.588 2.246 0.323 2.303
MITC0725 0.642 1.942 -0.154 6.967
JEHNOS 19 0.916 1 .991 1.352 5.997
CLME07 24 1.065 2.423 1.410 4.213
ARBU07 26 1.163 -0.990 2.236 5.810
HIGG07 27 2.054 - 10.069 2.238 5.843

 

 

 

 

 

156

 

Table 4-33 Pearson correlation matrix of actual and predicted Secchi depth

 

 

from AISA
Predicted SD Area Method Height/Width Method 1Narrow Band Method
Actual SD -0.798 0.869 0.581

 

 

 

 

 

 

AISA PREDICTED SD

 

 

ACTUAL SD

 

 

 

 

 

 

AREA HEIGHT/WIDTH NARROW
Figure 4-17 Correlation graphs between actual and predicted SD from AISA

Mitchell Lake and Muskegon Lake spectra were not reliable (explained in
Section 4.4.1.2). These lakes were removed and the result shows an

improvement on the narrow-waveband model (Table 4-34; Figure 4- 18).

Table 4-34 Pearson correlation matrix of actual and predicted Secchi depth
from AISA without Mitchell and Muskegon Lakes

Predicted SD Area Method Height/Width Method Narrow Band Method
Actual SD -0.820 0.872 0.735

 

 

 

 

 

 

 

 

AISA PREDICTED SD

 

ACTUAL SD

 

 

 

 

 

 

AREA HEIGHTNVIDTH NARROW

Figure 4- 18 Correlation graphs between actual and predicted SD from AISA
without Mitchell and Muskegon Lakes

157

4.4.3 TSS Model Validation

The ASD and the AISA radiance spectra from 12 lakes were converted to volume
reﬂectance and ﬁltered with Savitsky Golay method. The ﬁltered volume
reﬂectance was then calculated for the 1&3t derivative. The derivative products
were separated into narrow-wavebands and region-wavebands according to
Tables 4-2 and 4-3 (Sgolay indicator). Polynomial functions 2nd orders were
ﬁtted to the derivative data of region-waveband, and under-curve area, height
and width of the function were calculated (detail described in Section 3.4.7 to
3.4.9). These waveband variables were used in the TSS models to calculate the

predicted TSS for each lake (Table 4-35).

4.4.3.1 TSS ASD Validation
The predicted TSS values were correlated with the actual TSS concentration

measured from the sampled water of the same site (Table 4-36; Figure 4-19).

Table 4-35 TSS ASD validation

 

 

 

 

Predicted LNtss
Lake Actual LNtss
Area Height] Width Narrow Band
CLME0724 2.128 2.424 2.448 2.425
HIGG0727 2.197 2.313 2.348 2.338
MUSK0728 2.262 2.856 2.747 2.802
JEHN0519 2.303 2.479 2.494 2.515
ROGE0724 2.542 2.510 2.735 2.616
ARBU0726 2.580 2.410 2.375 2.498
SAPP0725 2.879 2.559 2.462 2.591
MITC07 25 2.944 2.569 2.681 2.570
HESSO7 27 3.401 2.92 1 2.993 2.682
HOUG0726 3.440 2.488 2.512 2.511
TAMA0605 3.940 2.597 2.548 2.526
BR000727 4.159 2.760 2.818 2.676

 

 

 

 

 

158

 

Table 4-36 Pearson correlation matrix of actual and predicted TSS from ASD
Predicted TSS Area Method Height/Width Method Narrow Band Method
Actual TSS 0.453 0.442 0.294

 

 

 

 

 

 

 

 

ASD PREDICTED TSS

 

ACTUAL TSS

 

 

 

 

 

 

AREA HEIGHTNVIDTH NARROW

Figure 4-19 Correlation graphs between actual and predicted TSS biomass from
ASD

TSS concentration for each lake was predicted using the TSS spectral indicator
model developed in section 4.3.3 applied to spectral data from a handheld ASD
spectroradiometer. The models tend to underpredict TSS values when
concentration is high and slightly overpredict values when concentration is low.
This may be because the data set used to develop the models was mostly from
low TSS lakes. Therefore, the models tend to underpredict high TSS values
resulting in relatively low correlations between actual and predicted TSS
showed in the correlation graph point pattern. Comparing among all methods,
the area model appears to have a stronger correlation between predicted and

actual TSS.

4.4.3.2 TSS AISA Validation
AISA spectra were extracted from the 30 bands AISA imagery and processed
according to Section4.4.1.2. Sgolay TSS models were applied to the AISA data

set to predict TSS for each lakes (Table 4-37). The predicted TSS was then

159

correlated with the actual reading TSS from the same sample site (Table 4-38;

Figure 4-20).

Table 4-37 TSS AISA validation

 

 

 

 

 

 

Lake Actual LNtss Predicted LNtss
Area Height/Width Narrow Band
CLME0724 2.128 2.847 3.168 2.377
HIGG0727 2.197 2.677 2.866 2.251
MUSK0728 2.262 3.812 3.923 2.417
ROGE0724 2.542 3.212 4.139 2.526
ARBU0726 2.580 2.628 2.787 2.411
TAMA0727 2.703 2.735 3.565 2.528
SAPP0725 2.879 2.509 3.192 2.424
MITC0725 2.944 2.599 3.796 2.526
JEHN0726 3.158 2.614 3.150 2.262
HESSO727 3.401 1.844 4.626 2.652
HOUG0726 3.440 2.673 3.330 2.433
BROOO7 27 4.159 2.237 4.207 2.655

 

 

 

 

 

Table 4-38 Pearson correlation matrix of actual and predicted TSS from AISA

 

Predicted TSS Area Method Height/Width Method Narrow Band Method

 

 

 

 

 

 

 

 

Actual TSS -0.640 0.456 0.580
AISA PREDICTED TSS
(D
(D
'—
_.l
<
D
I.—
0
< e

 

 

 

 

   

 

AREA HEIGHTNVIDTH NARROW
Figure 4-20 Correlation graphs between actual and predicted TSS from AISA

160

Two of the 12 AISA spectra were not reliable (discussed in Section 4.4.12).
Therefore, they were removed from the validation. The new prediction improved

only slightly (Table 4-39; Figure 4-21).

Table 4-39 Pearson correlation matrix of actual and predicted TSS from AISA
without Mitchell and Muskegon Lakes

Predicted TSS Area Method Height/Width Method Narrow Band Method
Actual TSS -0.631 0.562 0.590

 

 

 

 

 

 

 

 

AISA PREDICTED TSS

 

 

 

 

 

ACTUAL TSS

   

 

AREA HEIGHT/WIDTH NARROW

Figure 4-21 Correlation graphs between actual and predicted TSS from AISA
without Mitchell and Muskegon Lakes

Although the uncertain data were removed from the validation, the results
remain almost the same. Accuracy of AISA data was much lower than the ASD
because of the AISA conﬁguration. Spectral band location and band width of
the sensor used in this study did not support the use of area model. Narrow-

waveband method produced the better result in this case.

161

4.4.4 NPOC Model Validation
Although Sgolay did not produce the best models in every case, it was selected
as the method to validate because it produced better results for overall

biophysical parameters.

4.4.4.1 NPOC ASD Validation

The predicted N POC data (Table 4-40) were correlated with the real N POC

concentration measured from the sampled water of the same site (Table 4-41;

Figure 4-22).

Table 4-40 NPOC algae ASD validation

 

 

 

 

Lake Actual Predicted Lanoc
LN npoc Area Height/ Width Narrow Band

ARBU0726 1.550 2.075 2.063 2.146
BR000727 2.086 2.220 2.222 2.285
CLME0724 2.496 2.166 2.167 2.147
HESSO727 2.447 2.210 2.189 2.281
HIGG07 27 1.227 2.128 2.105 1.953
HOUG0726 2.259 2.229 2.223 2.236
JEHN0519 1.926 2.259 2.240 2.206
MITC0725 2.47 9 2.341 2.330 2.285
MUSK0728 2.257 2.516 2.520 2.276
ROGE0724 2.044 2.398 2.360 2.289
SAPP0725 2.123 2.248 2.218 2.128
TAMA0605 2.539 2.366 2.408 2.247

 

 

 

 

 

162

 

Table 4-41 Pearson correlation matrix of actual and predicted NPOC from ASD

 

 

 

 

 

 

 

 

 

 

 

 

 

Predicted NPOC Area Method Height/Width Method Narrow Band Method
Actual NPOC 0.491 0.539 0.712
ASD PREDICTED NPOC

O

O

(L

2

.1

<

D

'—

0

<

 

 

AREA

HEIGHTNVIDTH NARROW

Figure 4-22 Correlation graphs between actual and predicted NPOC from ASD

The correlation graphs indicate no correlation between the actual and the

predicted NPOC. Similar to TSS, NPOC data set mostly contain low NPOC

lakes. A wide range of data was lacking in the model development process.

Therefore, a model that captures speciﬁc optical features of NPOC was not

successfully produced in this study.

4.4.4.2 NPOC AISA Validation

The predicted NPOC concentration were calculated and correlated with the

actual concentration values (Tables 4-42 and 4-43; Figure 4-23).

163

 

Table 4-42 NPOC AISA validation

 

 

 

Predicted Lanoc
Lake Actual Lanoc
Area Height/Width Narrow Band
J EHN0726 1.227 1.256 2.637 2.043
MITC0725 1.550 1.379 2.510 2.136
SAPP0725 2.044 0.689 3.445 2.252
HOUG0726 2.086 0.905 3. 1 16 2.266
ARBU0726 2.123 1.240 2.755 2.155
MUSK0728 2.257 0.866 3.959 2.263
CLME0724 2.259 1.030 2.871 2.214
BROOO7 27 2.447 1.074 3.044 2.276
HESSO727 2.479 0.624 3.345 2.230
ROGE07 24 2.496 1.054 2.92 1 2.1 57

 

 

 

 

 

 

 

 

Table 4-43 Pearson correlation matrix of actual and predicted NPOC from AISA
Predicted NPOC Area Method Height/Width Method Narrow Band Method
Actual NPOC -0.569 0.499 0.718

 

 

 

 

 

 

 

 

AISA PREDICTED NPOC

/-‘. .

AREA HEIGHTNVIDTH NARROW

 

     

 

 

 

 

ACTUAL NPOC

 

Figure 4-23 Correlation graphs between actual and predicted NPOC from AISA

The correlations between actual and predicted NPOC are very low. Considering
correlation graph in Figure 4-23, there is no relationship between the predicted
and actual NPOC for region-waveband models. However, narrow-waveband

model appears to have a better correlation. This reﬂects the same explanation

164

 

with the TSS models that AISA conﬁguration used in this study may not

support region-waveband TSS models prediction. Therefore, no further test has

been done on excluding Mitchell and Muskegon lake data.

4.4.5 Algal Model Validation

The spectra of 12 lakes from ASD and AISA were used in the validation. The

data set used to develop the models and the data set used for validation were

compared (Table 4-44, Figure 4-24). Green algae was the only parameter that

has approximately the same distribution between the modeling and validating

data sets. Most of the validating data set for diatom and bluegreen algae had

lower algal biomass.

Table 4-44 Statistical comparison of modeling and validating data set

 

 

 

 

 

 

 

 

 

 

Diatom Green Bluegreen
Data set

Model Validate Model Validate Model Validate
Min 1259 30145 381352 868721 0 0
Max 471 1 0080 9496427 27080095 24628622 19053542 1 1 150160
Mean 1 807403 957288 4363092 5196300 4960194 3057239
Median 96388 2609 l 0 2001618 3065090 3343379 1 516637
Std. Dev. 7954953 2468981 6390964 6060236 5578742 398220 1

 

165

 

 

Model Data Setot Diatom Biomass Validate Data Set of Diatom Biomass
50907 5.0907

 

 

4.0907 - n- “—777 , .1___ 4,0907 . ,_-_ . .__

83.0907 kﬁ~ -1 ~ _- 3,0907. ————— —-
.9

 

 

 

 

 

 

 

2.0E+07 i , , ,7 ll 2,054.07 4... . .. -- .. __ —.__
1.0E007 — ———» — - ——- — « 1_oe+o7 . - .m _ . .
0.0900 Illa- 0,0900 iluww-__n_
Lake Lake
Model Data Set of Green Algal Biomass Validate Data Set of Green Algal Biomass
3.0E+07 3.0E+07

 

2.SE+07 255.07 «M h 1mm- _

2.0E+07 ~ - —
g 1.590% ~—
19 1.0907 <

5.0900 «

0.0900 - -

2.0907 - W A
Lalte

A7. , _‘(

7 i _ﬁ4

Biomass

1.SE+07

1.0E*07

 

5.0E+06

 

 

 

0.0E+00

 

 

 

Validate Data Set of Bluegreen Algal Biomass

 

 

 

 

“5‘07 2.907
2.E+07 — —- , ~ - - >—+ 2.5.07 T~ﬁmnl ..__ __ __.- _,-.E._ .‘_1
g ”M7 £19074
D

5.5+06 5.Et06 4*

 

 

 

 

 

 

 

 

0.5+m 0.E*00 .

 

 

 

 

   

 

 

Figure 4-24 Distribution of modeling and validating data set

4.4.5.1 Diatom ASD Validation

The ASD radiance spectra from 12 lakes were converted to volume reﬂectance
and ﬁltered with Savitsky Golay method. The ﬁltered volume reﬂectance was
then calculated for the Ian derivative. The derivative products were separated
into narrow-wavebands and region-wavebands according to Tables 4-2 and 4-3
(Sgolay indicator). Polynomial functions 2"(1 orders were ﬁtted to the derivative
data of region-waveband, and under curve area, height and width of the

function were calculated (detail described in section 3.4.7 to 3.4.9). These

166

waveband variables were used in the diatom biomass models to calculate the

predicted diatom biomass for each lake (Table 4-45). The predicted diatom

biomass was then correlated with the real diatom biomass concentration

measured from the sampled water of the same site (Table 4-46; Figure 4-25).

Table 4-45 Diatom ASD validation

 

 

 

Predicted LNdiatom
Lake LNdiatom
Area Height/ Width Narrow Band
JEHN0519 7.539 9.174 10.222 11.787
MITC0725 10.314 9.439 9.852 1 1.622
SAPP07 25 11.046 11.841 1 1.926 12.364
HOUG0726 11.934 11.263 1 1.592 12.922
ARBU0726 12.116 10.850 11.361 11.474
MUSK0728 12.161 11.994 11.821 11.974
CLME0724 12.554 12.265 12.415 11. 142
BR000727 12.582 11.632 11.968 12.050
HESSO727 12.639 12.134 11.943 12.667
ROGE0724 13.560 13.634 12.271 1 1.911
HIGG0727 13.579 12.616 1 1.960 12.201
TAMA0605 16.066 13.6 11 12.879 13.632

 

 

 

 

 

 

 

Table 4-46 Pearson correlation matrix of actual and predicted diatom from ASD

 

 

 

 

 

 

 

 

 

 

 

 

Predicted Diatom Area Method Height/Width Method Narrow Band Method
Actual Diatom 0.887 0.841 0.498
ASD PREDICTED DIATOM BIOMASS

E

.9

<_r_ .

O

_l

<

I)

'6

<

 

 

AREA

HEIGHTNVIDTH NARROW

Figure 4-25 Correlation graphs between actual and predicted diatom biomass

from ASD

167

 

 

Correlations between actual and predicted diatom were quite strong for both
region waveband methods. Given a complexity in natural waters where many
algal types coexist, the region waveband models could predict nearly 80 percent
of diatom biomass (R2 = 0.787). The selected waveband regions are sensitive to
changes in diatom biomass. Despite the fact that lakes that were used for
validation have very low diatom biomass, and the data that used to create the
models have very few diatom dominated lakes to begin with, wavebands used in

the model were able to capture diatom optical characteristics.

4.4.5.2 Diatom AISA Validation

AISA spectra were extracted from the 30 bands AISA imagery and processed
according to Section 4.4.1.2. Sgolay SD models were applied to the AISA data
set to predict diatom biomass for each lake (Table 4-47). The predicted diatom
biomass was then correlated with the actual reading diatom biomass from the

same sample site (Table 4-48; Figure 4-26).

Table 4-47 Diatom AISA validation

 

 

 

 

 

 

 

 

 

Lake LNdiatom Predicted LNdiatorn
Area Height] Width Narrow Band

JEHN0726 7.539 -11.573 19.376 12.007
MITC0725 10.314 -11.819 18.185 11.994
SAPP0725 11.046 -2.418 16.887 13.249
HOUG0726 11.364 -5.579 19.955 1 1.564
ARBU07 26 1 1.934 -3.237 18.499 10.936
MUSK0728 12.116 1.900 15.652 9.777
CLME0724 12.161 -11.614 21.813 11.396'
BR000727 12.554 0.088 16.598 8.934
HESSO727 12.582 -5.122 17.681 11.539
ROGE0724 13.560 0.285 26.576 1 1.177
HIGG0727 13.579 ~13.855 19.632 12.519

 

168

 

 

Table 4-48 Pearson correlation matrix of actual and predicted diatom from AISA
Predicted Diatom Area Method Height/Width Method Narrow Band Method
Actual Diatom 0.346 0.245 -0.280

 

 

 

 

 

 

 

 

AISA PREDICTED DIATOM BIOMASS

e O
0 o
e e e
e
O
0
e
e

AREA HEIGHT/WIDTH NARROW

 

 

 

 

 

 

ACTUAL DIATOM

 

Figure 4-26 Correlation graphs between actual and predicted diatom from AISA

Two of the 12 AISA spectra were not reliable (discussed in Section 4.4.1.2).
Therefore, they were removed from the validation. The result shows no
improvement at all (Table 4-49; Figure 4-27). The reasons could be that the
model is not sensitive to changes in diatom biomass, or the AISA spectral
setting used in this study does not support the estimation of diatom biomass.
Since ASD validation shows a very good result, the model should not be the
reason that causes such a low predictability with the AISA data. AISA data set
proved to produce high accuracy when applied with CHL (chlorophyll a) and SD
(Secchi depth) models. However, the average bandwidth of 5 nm — 8 nm (mostly
8 nm for the selected bands), may not provide the detail features needed for
biomass estimation of the algae. For example, band A (435 - 47 5 nm) has ASD
input of 40 data point to the polynomial ﬁtting curve, whereas, AISA only has 5
data point input into the model. In this case, AISA waveband actually ranges

from 434.45 nm to 486.92 nm, which is not exactly same range as selected in

169

the models. The worst case is band D (670 nm — 690 nm), which only have 3
data point input from AISA with spectral range between 674.09 nm to 694.13
nm. Table 4-49 showed Pearson correlation matrix of actual and predicted

diatom from AISA without Mitchell and Muskegon Lakes.

Table 4-49 Pearson correlation matrix of actual and predicted diatom from AISA
without Mitchell and Muskegon Lakes

Predicted Diatom Area Method Height/Width Method Narrow Band Method
Actual Diatom 0.315 -0.245 -0.283

 

 

 

 

 

 

 

 

AISA PREDICTED DIATOM BIOMASS

/ .7. \
g/

AREA HEIGHTNVIDTH NARROW

 

\

   
 

 

 

 

 

 

ACTUAL DIATOM

Figure 4-27 Correlation graphs between actual and predicted diatom from AISA

without Mitchell and Muskegon Lakes

4.4.5.3 Green Algae ASD Validation

The spectral indicator for green algae derived from Sgolay method consistently
performed better than the other approaches. It was therefore selected as the
best method and was used in the validation process. The predicted green algal
biomass was calculated and correlated with the real green algal biomass
measured from the sampled water of the same site (Tables 4-50 and 4-51;

Figure 4-28).

170

 

Table 4-50 Green algae ASD validation

 

 

 

 

 

 

 

 

Predicted LNgreen
Lake LNgreen
Area Height] Width Narrow Band
JEHN0519 13.039 14.149 14.384 19.987
MITC0725 13.675 14.512 14.824 17.338
SAPP0725 14.164 14.723 14.073 24.155
HOUG0726 14.662 14.345 14.308 18.697
ARBU0726 14.781 14.901 14.974 16.542
MUSK0728 14.993 15.467 15.545 14.168
CLME0724 15.215 14.335 14.448 16.954
BR000727 15.251 14.565 14.944 17.821
HESSO727 15.254 14.346 14.789 19.185
ROGE0724 15.676 14.271 14.785 17.950
HIGG0727 15.887 16.358 15.780 14.922
TAMA0605 17.019 14.425 14.651 19.728

 

 

, _

Table 4-51 Pearson correlation matrix of actual and predicted green algae from

ASD

 

Predicted green Area Method Height/Width Method Narrow Band Method

 

Actual green 0.249 0.360 -0.243

 

 

 

 

 

 

ASD PREDICTED GREEN ALGAL BIOMASS

O 0

 

 

 

 

 

 

ACTUAL GREEN

 

AREA HEIGHTNVIDTH NARROW

Figure 4-28 Correlation graphs between actual and predicted green algal

biomass from ASD

171

 

There was no correlation between the actual and the predicted green algae. It
could be because green algae do not have a unique or distinctive absorption /
reﬂectance features other than the appearance of normal chlorophyll curve.
The model that captures speciﬁc optical features that could detect green algae

is not successfully produced in this study.

4.4.5.4 Green Algae AISA Validation
The predicted green algal biomass were calculated and correlated with the

actual green algal biomass (Tables 4-52 and 4-53; Figure 4-29).

Table 4-52 Green algae AISA validation

 

 

 

 

 

 

Lake LNgreen Predicted LNgreen
Area Height] Width Narrow-Band '
J EHN07 26 13.039 15.003 13.754 16.155
MITC0725 13.675 15.421 13.650 18.967
SAPP0725 14.044 15.757 13.894 12.954
HOUG0726 14.164 14.504 13.806 18.624
ARBU0726 14.662 15.336 13.682 15.429
MUSK0728 14.781 15.339 13.854 19.032
CLME0724 15.2 1 5 15.240 13.66 1 15.982
BR000727 15.254 15.9 11 13.776 16.495
HESSO727 15.67 6 20.985 14.730 18.881
ROGE07 24 15.887 16.937 13.664 10.815
HIGG0727 17 .019 15.692 13.880 19.035

 

 

 

172

 

Table 4-53 Pearson correlation matrix of actual and predicted green algae from

 

 

 

 

 

 

 

 

 

AISA
Predicted Green Area Method Height/Width Method Narrow Band Method
Actual Green 0.405 0.273 0.017
AISA PREDICTED GREEN ALGAL BIOMASS
6
DJ
0:
(D
..l
<
D
5
<

 

 

 

 

 

 

 

AREA

HEIGHTNVIDTH NARROW

Figure 4-29 Correlation graphs between actual and predicted green algae from

AISA

The correlations between actual and predicted green algae are very low.

Considering correlation graph in Figure 4-29, there is no relationship between

the predicted and actual green algae. Therefore, no further test has been done

on excluding Mitchell and Muskegon lake data.

4.4.5.5 Bluegreen Algae ASD Validation

The spectral indicator for bluegreen algae derived from Sgolay method was used

in the validation process. The predicted bluegreen algae was correlated with

the real bluegreen algal biomass measured from the sampled water of the same

site (Tables 4-54 and 4-55; Figure 4-30).

173

 

 

 

Table 4-54 Bluegreen algae ASD validation

 

 

 

 

Predicted LNbluegreen
Lake LNbluegreen
Area Height / Width Narrow Band
JEHN0519 6.982 8.353 11.979 5.160
MITC0725 13.478 15.184 15.740 13.203
SAPP0725 13.501 14.828 14.910 10.995
HOUG0726 13.940 16.128 16.173 13.681
ARBU0726 14.368 14.713 14.659 13.586
MUSK07 28 14.379 12.830 1 1.462 10.587
CLME0724 14.899 13.211 14.373 13.176
BR000727 15.664 14.988 14.997 1 1.384
HESSO727 15.992 16.150 15.292 11.387
ROGE0724 16.227 15.337 15.298 13.275

 

 

 

 

 

 

Table 4-55 Pearson correlation matrix of actual and predicted bluegreen algae

from ASD

 

Predicted Bluegreen Area Method

Height] Width Method lNarrow Band Method

 

 

 

Actual Bluegreen

0.849

 

0.536

 

0.811

 

ASD PREDICTED BLUEGREEN ALGAL BIOMASS

 

 

 

 

 

 

 

ACT BLUEGREEN

AREA

HEIGHTNVIDTH NARROW

Figure 4-30 Correlation graphs between actual and predicted bluegreen algal

biomass from ASD

174

 

 

Although the correlation coefﬁcients are high, there was no strong correlation
between the actual and the predicted green algae considering the correlation
graph. It could be because the validate data set does not have high bluegreen
biomass lakes that have the strong optical features mentioned in Section 4.4.5

(Figure 4-24).

4.4.5.6 Bluereen Algae AISA Validation
Statistical correlation was performed on the predicted bluegreen algal biomass
were calculated and correlated with the actual biomass (Tables 4-56 and 4-57;

Figure 4-31).

Table 4-56 Bluegreen algae AISA validation

 

 

 

 

 

 

 

 

Lake LNbluegreen Predicted LNbluegreen
Area Height [Width Narrow Band

J EHN07 26 6.982 - 104.087 - 128.282 9.732
MITC0725 13.478 - 11.129 -37.148 12.828
SAPP0725 13.501 - 17.873 -45.334 10.719
HOUG07 26 13.940 -6.085 - 15.480 13.159
ARBU0726 14.075 -23.37 2 -51.623 12.200
MUSK07 28 14.368 —29. 146 - 103.32 1 12.438
CLME07 24 14.379 - 1.465 - 129.525 11.945
BR000727 14.899 -23.879 -4.993 12.069
HESSO7 27 15.664 -22.247 -70.829 11.858
ROGE0724 15.992 -30.9 18 -121.843 12.744
HIGG0727 16.227 -26.086 -46.916 12.434

 

175

 

Table 4-57 Pearson correlation matrix of actual and predicted bluegreen algae
from AISA

 

Predicted Bluegreen Area Method Height] Width Method Narrow Band Method
Actual Bluegreen 0.793 0.308 0.746

 

 

 

 

 

 

 

AISA PREDICTED BLUEGREEN ALGAL BIOMASS

AREA HEIGHTNVIDTH NARROW

 

   

 

 

 

 

 

ACT BLUEGREEN

Figure 4-31 Correlation graphs between actual and predicted green algae from

AISA

Although correlation coefﬁcients are high, correlation graph also showed no
relationship between AISA predicted and actual bluegreen algae. The Higgins
lake data that were outliers was an exceptionally deep and clear lake. An
experiment had been made to exclude Mitchell, Muskegon, and Higgins lakes to
investigate if the model is able to predict bluegreen algal biomass at all since it
was expected to predict a half of the actual data according to the model R2
(Tables 4-22 to 4-24). Correlation graph shows some degree of correlation in
the region waveband methods, but the correlation appears to be negative
suggesting that the model could not be used effectively with this AISA data

(Table 4-58; Figure 4-32).

176

 

 

 

Table 4-58 Pearson correlation matrix of actual and predicted bluegreen algae

 

from AISA without Mitchell, Muskegon, and Higgins Lakes
Predicted Bluegreen lirea Method Height/Width Method Narrow Band Method
Actual Bluegreen -0.751 -0.516 0.164

 

 

 

 

 

 

 

 

AISA PREDICTED BLUEGREEN ALGAL BIOMASS

D

AREA HEIGHTNVIDTH NARROW

 

 

 

 

 

 

ACT BLUEGREEN

 

Figure 4-32 Correlation graphs between actual and predicted bluegreen algae

from AISA without Mitchell, Muskegon, and Higgins Lakes

4.5 Discussion on Spectral Indicators for Water Quality Assessment
Spectral indicator models developed from Savitsky Golay ﬁlter performed best
among the four ﬁlters. Performances of the models developed from Savitsky
Golay ﬁlter were summarized in this section. The dependant variables (spectral
wavebands) were discussed in terms of their linkages to the water quality

variables.

4.5. 1 Chlorophyll a

The best CHL spectral indicator models and their accuracies were summarized

in Table 4-59.

177

 

Table 4-59 Chlorophyll model performances

 

 

Index Model Model Validation Validation
R2 ASD R2 AISA R2
Area LN(CHL) = -1.687(A/ G) + 1.509(D/ G) 0.757 0.612 0.393

- 0.331(F/E) + 1.132(G/D)

 

 

+ 3.234
Height/ LN(CHL) = -1.930(A/G) - 5.181(B/D) 0.699 0.748 0.469
Width + 5.891
Narrow LN(CHL) = -38.605(A) + 42.536(E) - 0.754 0.561 0.699

5.108(E/H) — 1.549(G/A) +

 

 

 

 

 

 

 

9.801

 

The blue absorption band A (435 nm or 435 nm — 475 nm) and the ratio of it
with the NIR peak band G (690 nm or 700 nm — 740 nm) were the key
wavebands for chlorophyll assessment. Blue wavelengths not only contains the
most information of water column as they penetrate deeper than other
wavelengths, but they also represent an absorption region of CHL. NIR
wavebands are no doubt the indicator of CHL. NIR wavelengths have been used
in the inland vegetation indicators and blue wavelengths have been used in

water quality indicators.

The ratio between maximum CHL absorption in red wavelengths and the NIR
peak were selected repeatedly (region-waveband: D / G, F/E, G / D and narrow-
waveband: E / H). The RED / NIR ratio has also been used extensively in

vegetation studies. These wavebands are truly indicators of CHL.

178

 

 

4.5.2 Secchi Depth
The best SD spectral indicator models and their accuracies were summarized in

Table 4-60.

Table 4-60 Secchi depth model performances

 

 

Index Model Model Validation Validation
R2 ASD R2 AISA R2
Area LN(SD) = - 5.560(A/B) + 0.792(A/G) 0.700 0.843 0.672

— 2.183(B/F) - 0.937(E/C)

 

 

+ 3.955

Height/ LN(SD) = - 47.187(C) — 1.742(B/A) 0.759 0.656 0.760
Width — 0.597(G/C) + 4.559

Narrow LN(SD) = -7.1088(A) + 1.1003(D/A) 0.768 0.516 0.540

— 1.8842(D/C) - 1.0169(G/C)
+ 2.871

 

 

 

 

 

 

Shorter spectral wavelengths (blue waveband) between 435 nm — 475 nm (band
A) explained clarity in water. Blue wavelength region approximately 400 nm -
500 nm has the least amount of absorption and scattering of incident light in
the water column (Figure 2-6). These wavebands penetrate into the water
column deeper than other wavelengths. Reﬂectance from them contains
information of constituents in the water column. Therefore, they can be used to
differentiate clear and turbid waters. Longer wavebands (> 520 nm) get
absorbed by water very quickly; therefore, they could not detect and separate

clear water from turbid water. The longer wavebands that appeared in the

179

 

models were mostly the indicators of chlorophyll. For examples, N IR peak for
band G, maximum CHL absorption for band D and E. Since turbidity in the
studied lakes was dominated by algal biomass, these selected bands could

improve the predictability of the SD model.

When incident light penetrates into the water column, it interacts with inherent
substances that are dissolved and suspended. Clarity measured by a Secchi
disk can results from almost everything optical objects in the water, such as,
algal biomass, inorganic suspended solid, and dissolved humic substances.
After the incident light interacts with substances in water, it reﬂects back and
gets measured by the remote sensor. The higher reﬂectance usually associates
with the higher substances. Except for the case that water contains high
concentration of dissolved organic carbon or other absorbing agents that low
reﬂectance would indicate high concentration of constituent in water. Secchi
depth (SD) is the measurement of visibility link to clarity of the water. It is
almost a direct optical measurement. Low SD reading usually means turbid
water. Turbid water has a lot of substances in the water for the incident light
to interact with and reﬂect back. Therefore the spectral reﬂectance of this lake
is expected to be high. Section 4.3.2 and Figure 4— 11 showed that derivative
product of the volume reﬂectance is higher in more turbid water (lower SD

value) and lower in clearer water (higher SD value).

180

 

4
, a.

 

4.5.3 T88 and NPOC

The best TSS and NPOC spectral indicator models and their accuracies were

summarized in Tables 4-61 and 4-62.

Table 4-61 TSS model performances

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Index Model Model Validation Validation
R2 ASD R2 AISA R3
Area LN(TSS) = 0.420(D/A) + 1.088(E/ D) 0.286 0.205 0.398
+ 0.295(G/E) + 0.405
Height/ LN(TSS) = 0.213(G/C) + 2.273 0.207 0.195 0.316
Width
Narrow LN(TSS) = -1.224(F/ H) + 3.876 0.326 0.086 0.348
Table 4-62 NPOC model performances
Index Model Model Validation Validation
R3 ASD R3 AISA R2
Area No Valid Model
Height/ LN(NPOC) = 0.100(E/ C) + 1.902 0.105 0.291 0.249
Width
Narrow No Valid Model

 

 

 

 

 

 

Majority of the lakes used in this study did not have high TSS or NPOC (Figure

4-1 and 4-3). Although the data were LN transformed to avoid the skewness

problem (Figure 4-4), the accuracy of the TSS models were low and only

height/width indicator produced a valid NPOC model. Although humic

181

 

 

 

substances (NPOC) produces distinctively low reﬂectance curve by signiﬁcantly

absorb incident light (Figure 4— 12), one or two low reﬂectance signatures were

not sufﬁcient to produce a quantitative regression model.

4.5.4 Algae

The best algal spectral indicator models and their accuracies were summarized

in Tables 4-63 to 4-65.

Table 4-63 Diatom models performances

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Index Model Model Validation Validation
R2 ASD R2 AISA R2
Area LN(Diatom) = 2.670(A/ D) + 0.208 0.787 0.099
10.127(D/C) — 3.522(G/C)
+ 5.638
Height/ LN(Diatom) = 0.752(E/C) - 1.729(G/A) 0.216 0.707 0.060
Width + 11.663
Narrow LN(Diatom) = 4.718(A/ D) + 8.474 0.206 0.248 0.080
Table 4-64 Green algae models performances
Index Model Model Validation Validation
Ra ASD R2 AISA R2
Area LN(Green) = 1.642(A/D) + 2.266(E/A) 0.285 0.062 0.164
+ 9.847
Height/ LN(Green) = 0.375(E/A) + 13.343 0.192 0.130 0.075
Width
Narrow LN(Green) = -28.200(A) + 34.403(D) 0.309 0.059 0.000
+ 14.341(A/B) — 0.010

 

 

 

 

 

 

182

 

 

 

Table 4-65 Bluegreen algae models performances

 

 

Index Model Model Validation Validation
R2 ASD R2 AISA R2
Area LN(Bluegreen) = - 3.123(A/ G) 0.580 0.721 0.564

+ 1.977(C/A) — 66.849(D/B)
— 7.727(D/E) + 28.569

 

Height/ LN(Bluegreen) = 133.395(F) 0.353 0.287 0.266
Width — 35.544(A/F) + 69.307(B/F)
- 1.191(D/A) + 1.266

 

Narrow LN(Bluegreen) = - 2.107(B / E) 0.320 0.658 0.027

 

 

 

 

 

— 2.776(H/C) + 19.926

 

 

Comparing all algae indicators, bluegreen model reported high R2 but the
correlation plot showed no correlation (Figure 4-30 to 4-32). In fact, only ASD
diatom correlation graph showed a relatively strong correlation (Figure 4-25).
None of the other algal validation demonstrated a relationship between the
predicted and the actual algal biomass. AISA hyperspectral data in this study
could not be effectively used with the algal indicators at all. The band width
and location setting for the data used in this study may not be appropriate to
the algal division assessment, although it worked better in the chlorophyll

models.

There could be potential algal classiﬁcation error associated with the algal

biomass data, such as misclassiﬁcation of Dinobryon (a Chrysophyte) to the

green algae class. There was also some uncertainty associated in the

183

 

calculation of biovolume because no one reference source contained all of the
standard biovolume for the algae found in studied samples. The standard
biovolume for each species were determined from three different sources
(Section 3.3.4) and some of them were different among the sources. In addition,
not all of the species cell biovolumes could be found in the reference sources,
and the exact number of cells within a unit was not recorded in the counting
process. Therefore, the accuracy of biovolume determination was limited by
insufﬁcient data. Algal density was counted by unit (colonies), not by cell.
Approximate number of cells per unit was determined before multiply the
density with cell biovolume. The estimated biovolume of known species were
used to determine a constant biovolume factor to represent each algal category.
These factors were then multiplied with algal density for each lake. The fmal
biovolume of algae in each lake represents a relative biomass of algae in each
algal division. Figure 4-24 showed the range of algal biomass in the data sets.
Diatoms dominated only one lake, green algae dominated a few lakes, and
bluegreen algae (especially microcystis) appeared in many lakes and had an
almost continuous range. This may explain why the bluegreen indicator models
report higher R2 than other algal models. Bluegreen algae have absorption and
reﬂectance features that are different from other algal type (discussed in Section
4.3.7), and wavebands where these features occur were selected by the
regression model. Therefore, with larger and more complete data, there is a

potential to assess bluegreen algae.

184

 

 

Comparing model accuracies of all variables, area region-waveband models
performed better than narrow-waveband indicators. Region-waveband
indicators take into account the wider range of wavelength that could be
responsible to the predicting variables. The results showed the generalizability
of the area region-waveband models that could capture changes in the signal of

responding biophysical parameters in the water.

185

 

 

 

CHAPTER 5

CONCLUSIONS AND FUTURE RESEARCH

Hyperspectral remote sensing provides a valuable tool in water quality
assessment. It has been successfully used to estimate spatial and temporal
variation of water quality parameters primarily for marine waters, which has
less optical complicating factors such as suspended solids and dissolved
organic carbon than inland waters. Most of the previous hyperspectral studies
on inland waters were performed over small spatial extents (e.g., one or a few
lakes) or in controlled environments (e.g., laboratory spectral reading of several
chlorophyll and suspended sediment conditions). This study was conducted in
a natural environment of a large spatial extent (statewide scale) where several
biophysical variables coexist. A total of 48 individual measurements of
radiometric spectra, chlorophyll a, Secchi disc depth, suspended solid, non-
purgable organic carbon, and phytoplankton species composition including
diatom, green, and bluegreen algae, were collected during spring and summer
of 2004. Hyperspectral sensors used in the study included a hand-held
spectrometer (LabSpec® Pro, Range 350 nm — 2,500 nm, 1 nm interval with
sampling interval 1.4 nm @ 350 nm - 1,050 nm and 2 nm @ 1,000 nm — 2,500
nm) and an Airborne Imaging Spectrometer for Applications (AISA, Range 434
nm - 900 nm, 3 nm — 8 nm bandwidth). The objectives of the study were (1) to
identify optimal spectral bands that are most sensitive to water quality
indicators in the various water bodies within Michigan; (2) to develop improved

spectral water quality indicators; and (3) to compare the ﬁlter methods for noise

186

 

 

 

removal in hyperspectral remote sensing data. The purpose of the study was to
develop generalizable algorithms that were spatially and temporally

independent. Each of the objectives is discussed in the following sections.

5.1 Optimal Spectral Bands

The most sensitive spectral bands that can be eﬂ'ectively used in inferring water
quality information were identiﬁed in this study. Two sets of wavebands were
located including narrow-wavebands and region-wavebands. The most
sensitive narrow-wavebands (1 nm bandwidth) were at 435 nm, 455 nm, 470
nm, 565 nm, 670 nm, 675 nm, 690 nm, and 700 nm. Biophysical
characteristics associated with these spectral bands were indicated in Table 4-
3. The wavebands were identiﬁed using principle component analysis (PCA)
and spectral derivative methods based on absorption and reﬂectance properties
of the biophysical variables of interest. Region—wavebands were identiﬁed using
a 1*“ derivative technique to detect spectral regions that were sensitive to
changes in water quality variables. These spectral regions were 435 nm — 475
nm, 470 nm — 570 nm, 565 nm - 620 nm, 670 nm - 690 nm, 675 nm — 700
nm, 690 nm — 740 nm, and 700 nm — 740 nm. Table 4-2 explained biophysical
characteristics associated with these wavebands. Different waveband regions

were identiﬁed for different trophic waters.

Two spectral regions were detected for each of the near-infrared (NIR) and red

regions to represent high and low trophic waters. Because the NIR peak shifted

to the longer wavelengths when chlorophyll a increased, it was reasonable to

187

 

 

detect two spectral regions and compare which one was more sensitive to water
quality parameters. The results clearly showed that the higher trophic
waveband (waveband G) was selected in statistical process to use in the water
quality indicator models more frequently for the NIR reﬂectance peak region,
but the lower trophic waveband (waveband D) was used in the models more
ﬁ'equently for the red absorption region. For the NIR case, the selected
waveband G had a narrower bandwidth (Figure 3-17), which automatically
avoided an effect from the red absorption region (immediately next to it 3n the
lower wavelength side). For the red spectral region case, waveband D covered

the absorption feature better for most lakes.

Validation for each of the variables showed that 8—nm bandwidth AISA data
provided similar accuracy as the l-nm-bandwidth ASD data in the narrow-
waveband models. However, when uSed in area and height/width region-
waveband models, the 8-nm-bandwidth AISA produced lower accuracy than 1-
nm-bandwidth ASD. The wide bandwidth of AISA particularly affected an
implementation of water quality parameters in the region-Waveband method,
especially the height / width indicator method. It lowered the‘ sensitivity of the
spectral band regions when nOt enough datawere available for the polynomial
ﬁtting process. For example, waveband D (670 nm — 690 nm) has 20 data
values from the 1-nm-bandwidth ASD to input in the polynomial curve ﬁtting,
but only 3 data values from the 8-nm-bandwidth AISA. Results of this study
suggest that the bandwidth for hyperspectral data to be effectively used in

water quality assessment should be! approximately 3 nm, or not more than 4

188

 

 

 

nm, and the waveband locations should be concentrated on wavelengths
between 435 nm - 740 nm (with an exclusion of wavelength between 620 nm —

670 nm).

5.2 Spectral Indicators for Water Quality Assessment

Quantitative relationships between water quality and remote sensing
parameters with polynomial ﬁtting were derived for seven water quality
variables including Secchi depth (SD), Chlorophyll (CHL), Total Suspended Solid
(TSS), N on-Purgable Organic Carbon (N POC), diatom biomass, green algal
biomass, and bluegreen algal biomass. Three spectral indicators - area,

height / width, and narrow-wavebands — were developed and compared. Area-
under—spectral-curve indicators were found to be the best indicators in this
study in terms of goodness—of-ﬁt reported by the regression models. Narrow-
waveband indicators had the second best accuracy but the method is easier to
use because these indicators used the ﬁltered volume spectral reﬂectance from
the lakes without processing through derivative calculation and polynomial
ﬁtting. ‘Height/width indicatorswere not as sensitive to changes in biophysical

variables as the other two indicators.

Among the seven spectral indicator models for water quality, Secchi depth,
chlorophyll a, and bluegreen algal biomass models showed the most promising
results in terms of using remote sensing to map water quality. High accuracies
were reported in terms of goodness-of-ﬁt when validated by hand-held and

airborne spectrometer. The results showed goodness-of-ﬁt ranked from the

189

 

 

 

Secchi depth (R2 0.76-0.84), chlorophyll a (R2 0.70-0.76), and bluegreen algae
(R2 056-072) models. Spectral indicators for chlorophyll a used spectral
regions from both low and high trophic types of water. They are therefore
independent of trophic states. The models can be used to assess chlorophyll a
condition in any waters regardless of trophic condition. The spectral indicators
for TSS (R2 021-040), NPOC (R2 011-025), diatom (R2 0.10-0.79) and green
algae (R2 013-031) did not produce accurate results due to limited availability
of a wide range of these data. The data collection focused on a wide range of
trophic state, which reﬂected in a continuous range of chlorophyll a and Secchi
depth. Historical data for TSS, NPOC, and algal compositions in lakes were not
available before the ﬁeld data collection was conducted. Concentrations of
these variables were low in most of the sampled lakes; therefore, the models
were developed based on data that were bias toward the lower concentrations

and tended to underestimate the actual values.

5.3 Effective Filters for Noise Removal

A spectral ﬁltering techniques need to be applied to hyperspectral data in order
to remove noise generated from the atmosphere or within the sensor itself.
Three hyperspectral denoising ﬁlters — Savitsky Golay, wavelet Daubechies and
wavelet Symlet - were compared for their ability to remove noise in the data
while maintain sharp spectral absorption /reﬂectance features. Savitsky Golay
proved to be the best method. It smoothed noise within the data while
preserving sharp peaks and troughs because the ﬁlter calculated a local

polynomial for every determined wavelength range (Figure 3-12). Symlet

190

 

 

 

wavelet was the second best method. It generated de-noised spectral signatures
that seemed smooth but some of the high frequency noise still remained in the
signature (Figure 3-13). The remaining noise showed in the 1%t derivative
product. Unlike Savitsky Golay, Symlet tended to ﬂatten out the curves, which
reduced the slope of the curve. As a result, derivatives from Symlet ﬁlter
appeared ﬂat and the wavelength range of absorption and reﬂectance features
were more difficult to determine (Figure 3- 15). Daubechies wavelet performed
poorly in this study. It created a staircase artifact in the de-noised spectra,
which generated a tremendous amount of zero values in the derivative product
(Figure 3312 and 3- 15). The process required elimination of zero data, which
reduce the spectral resolution of continuous l-nm bandwidth to an inconsistent
bandwidth depending on the location of zero values. When amounts of data
were eliminated, polynomial ﬁtting was based on fewer amounts of data than

other ﬁlter methods.

5.4 Major Findings

The important contributions of this study include the development of an
algorithm to answer remote sensing questions of which spectral wavebands are
effective in inferring water quality information, and what ﬁlter method can be
effective in removingnoise and preserving important absorption and reﬂectance
features. The study also answers the water quality monitoring question, “can
remote sensing can be used effectively to assess water quality variables in
natural ecosystem?” Quantitative relationships between remotely sensed

spectra‘and three water quality indicators - Secchi depth, chlorophyll a

191

 

 

concentration, and bluegreen algal biomass — were developed with high
accuracies in this study. Through this study, it was demonstrated that
hyperspectral remote sensing could be used to infer water quality variables.
Validation methods in the study showed that results from a hand-held
hyperspectral instrument can be extended to airborne scale suggesting that
hyperspectral remote sensing has a promising potential to be used to map
water quality condition and facilitate improved water quality monitoring in large

spatial extents at low cost.

The study also demonstrated that ﬁeld measurements are unpredictably
variable due to cloud condition, time of measurement, sun elevation angle,
relative sun-target-sensor angle, and shadow casting from the boat. These
potential sources of error could be avoided by taking multiple measurements at
one site so that the spectra could be averaged and the eﬁect of water surface
glare minimized. Reference reﬂectance from the white reﬂectance panel must
always be recorded immediately before and after the water reﬂectance
measurements are taken in order to ensure the correction for inconsistent
incident light. Water surface spectra should be taken as far from the boat as
possible to avoid shadow effect. Measurements must be done consistently at
the nadir angle regardless of waves on the water surface to keep the
measurement of all lakes on the same standard. More spectra may be

necessary if the water surface is rough.

192

 

 

 

Based on this study, a limited number of spectral bands were identiﬁed for
water quality analysis. Based on the number of wavebands selected, it is
suggested that hyperspectral data are not needed if multispectral data are
designed to have narrow bandwidth position in the sensitive spectral
wavelengths. The spectral wavebands identiﬁed in this study suggest the
conﬁguration of prototypes for the future satellite sensors, which provide useful

spectral information of hundreds of water bodies at one time at reasonable cost.

5.5 Future Research
( 1) A new algal cell counting method needs to be employed to produce a better
accurate in situ data. The algal samples should be counted to 300 units as did

in this study but number of cells in a unit should also be recorded.

(2) Future research should include samples from more humic and high-
suspended sediment lakes. The sampled lakes in this study were determined
based on their trophic states because information on TSS or DOC of majority of
lakes in Michigan was not obtainable. A more complete set of these data can
improve the determination of spectral indicator models in the statistical

process.

(3) Narrower and more frequent wavebands from an airborne hyperspectral
sensor should be tested. Although the ASD proved useful according to the
results and conclusion of this study and could save time and cost for water

sampling and laboratory analysis, an airborne sensors such as AISA provides

193

 

 

additional spatial information such as mapping of water quality variables on a
larger area. ASD measurements involve traveling to the lake, launching the
boat, traveling to the sample sites, and then moving to the next lake. A
maximum of seven lakes could be sampled in a working day by ASD method,
but tens of lakes could be sampled by an airborne sensor in a shorter period of

time.

(4) An analytical method similar to this study should be explored, deriving
waveband regions and under-spectral-curve area directly from the spectral
reﬂectance of the water instead of the derivative product of it. The derivative
generally separated the reﬂectance curve in half. The method worked well in
this study because it responded to changes in biophysical variables very well.
However, by visual observation, the entire absorption / reﬂectance curves from
spectral reﬂectance seemed to respond with water quality as well. Therefore,

they deserve further study.

194

 

 

 

APPENDICES

195

APPENDIX A

WATER QUALITY FIELD WORK PROTOCOL

1. Total 46 Michigan inland lakes, over 50 percent of which were in the
Muskegon River Watershed, were selected based on their trophic states to have
the widest range of trophic condition possible (Figure A-l; Table A-l). Trophic
State Index (TSI) values were calculated from historical chlorophyll a (CHL)
measurements from three sampling programs: (1) the Michigan Department of
Environmental Quality’s (MDEQ) Lake Water Quality Assessment (LWQA)
Monitoring Program, (2) the Michigan Cooperative Lake Monitoring Program
(CLMP), and (3) Professor R. Jan Stevenson’s algal ecology lab, MSU
Department of Zoology. These data were collected in spring and summer

between May - August of 2001 to 2003.

 

 

 

 

 

LakeTSKCHL)
80
O
_ k ,1, , , _ , v.1 L_--.___._______._______ ;_’._
W 60 .0.
8 ...QO..
3 ...eoooo’
3 404——~---——--v 400—04~~—-~ ,___‘_~_1__r_ —— ~h ﬂ
0
a .ee
2 O
p.
20» 4 f -1 --.,-,_-i_--1 1 a 1 1 1 1 1,11, *4 few—~49
0 II T w TTTTT
o 9 9 e} e‘ e‘ '0 \\ ° ‘\ 0‘ ’b\\ 6‘ ‘J
0‘ 0‘°~O“‘°.e§v°.‘°o°’°<°°4~°o°
6°03 V3,0 '0 0&6 +\ Q- ‘9‘05 0‘90 (96‘ 3° 2‘ ‘0 Q? Q5, 9'.
40’ T
0‘
Lake

 

 

 

Figure A-l Trophic conditions of the selected sample lakes

196

Table A-1 List of selected sample lakes

 

 

 

TSI from T8! from
Lake CHL Lake (continued) CHL
Glen (Big) 31.00 Muskegon 53.40
Higgins 33.28 Hardy Dam Pond 54.09
Chemung 35.00 Houghton 54.50
Sapphire 35.00 Fremont 54.60
Arbutus 37.00 Mitchell 55.98
Long Lake (Grand Traverse) 37.00 Randall 56.92
Silver 40.00 Croton dam pond 57.88
Clear (St. Joseph) 40.40 Pickerel 60.12
Diamond 40.50 Hess 62.00
Mecosta 40.98 Brooks 62.74
Klinger 4 l .50 Bear (Kalkaska) 70.61
Eagle 43.00 Belleville (Washtenaw) unknown
Marl Lake 44.94 Cass (Oakland) unknown
Big (Osceola) 45.50 Chipewa unknown
Rogers Darn Pond 46.00 Cub unknown
Tamarack 46.10 Ford (Washtenaw) unknown
Little Whiteﬁsh 46.48 Hicks unknown
Clear (Mecosta) 46.66 Kent (Oakland) unknown
Horsehead Lake 46.77 Loon (Oakland) unknown
Round (Mecosta) 48.50 Maceday (Oakland) unknown
Kimball 50.50 Oakland (Oakland) unknown
Pontiac (Oakland) 52.33 Orchard (Oakland) unknown
Jehnsen 53.37 Paw Paw unknown

 

 

 

 

197

 

 

2. Field work had been done in 13 weeks in two periods Apr 24 — Jun 12 and
Jul 17 - Aug 28 to collect:

2.1 GPS point coordinates

2.2 Spectral signatures using the ASD

2.3 Lake depth

2.4 Secchi-depth (SD)

2.5 Light extinction proﬁle using LI-250 light meter

2.6 Dissolved Oxygen (DO)

2.7 Temperature

2.8 Chlorophyll a (Chl-a) I
2.9 Total suspended solid (TSS)
2.10 Phytoplankton

2.11 Nutrients (TP and TN)

2.12 Dissolved organic carbon (DOC)

3. Procedure:

3.1 Launch the boat.

3.2 Turn on ASD and GPS.

3.3 Drive the boat to the deep basin of the lake. Lower the anchor. Record the

GPS coordinate.

 

198

 

3.4 Measure the lake depth by lowering the Secchi disc to the bottom of the

lake. Record the depth. Pull the Secchi disc up until it is visible. Record the

 

Secchi-depth (SD).

3.5 Measure the light, DO, and Temperature. All the sensors are tied together
with the lowering frame. Be sure to lower the lowering rope and not the sensor
cables. Begin measurement ﬁom the surface of water (where water just covers
the sensors). For lakes with shallower SD (< 4 m) record light, DO, and Temp
every 0.5 m. For lakes with deeper SD (> 4 111) record every 1.0 m. Lower the

sensors until the depth where light is lower 1 percent of the light at surface (at

the beginning). Turn off the meters.

3.6 Water samples are collected for Chl-a, TSS, nutrients, algae, and color
analysis. Samples for Chl-a, TSS, nutrients, and algae analysis will be taken
using three methods: 1) Photic depth, 2) Secchi depth, and 3) Epilimnion depth.
Samples for TN, TP, and DOC are taken using the photic depth method only.

- Photic depth can be calculated by multiplying the Secchi depth by 2.5.
The depth is then divided by four to indicate the actual depth at which the
water was collected each time. The mixture of the water from these four depths
represents the column of the photic layer. After four liters of water is collected

and drained into a bucket, it is transferred into one acid-washed 250ml-bottle

 

for nutrients, two 30ml-glass vials for DOC, one 1,000rnl-bottle for algae, one

1,000ml-bottle for Chl a, and one 1,000ml-bottle for TSS analysis.

199

- Secchi depth method collects water from the Secchi depth column.
The depth is divided into four depths. Approximately four liters of water are
collected from four depths and mixed in a bucket. The sample is transferred
into one acid-washed 250ml-bottle for nutrients, one 1,000ml-bottle for algae,
and one 1,000ml-bottle for Chl-a and one 1,000ml-bottle for TSS analysis.

- The epilimnion layer does not appear in all lakes at the time of
sampling. For lakes that have drastic changes in temperature proﬁles, one
more set of samples will be collected for Chl—a, TSS, nutrients, and algae
analysis. The epilimnion depth is divided by four to indicate the actual depths
that samples are taken from.

- The chemical (M3) is added to the algae bottles to preserve algae cells.
The rest of the samples are placed in a cooler. Samples for Chl-a and TSS are
ﬁltered in situ before covered with aluminum foil and kept in a freezer until
analysis. Samples for DOC analysis are covered with aluminum foil on site and

kept in a fridge until analysis. Samples for nutrients are kept in a freezer.

3.7 To take the spectral signatures

- Connect laptop with the ASD (after the ASD has warmed up for 15
minutes). Turn on the computer and navigate to FR B&W (an icon on the
screen).

- Select ‘Spectrum save’ and change the path to the workspace folder to
record spectral signatures. Change the starting number to 1. Change the view

angle to match the angle on the sensor.

200

 

- Take out the white reﬂectance panel and hold the sensor above it. Be
careful not to have any shadow on the panel. Click on ‘OPI" and let the sensor
calibrate. Once ﬁnished, continue to hold the sensor above the white
reﬂectance panel and click on ‘WR’. Perform the WR calibration. Save the
white reﬂectance by pressing spacebar approximately 10 times.

- Hold the sensor approximately 1.5 meters above the water surface and
save the water signature (approximately 10- 15 signatures). Take another 10
signatures of white reﬂectance.

- Leave the laptop on if the distant to the next site is short. Move to the
next site immediately or click on ‘quit’ and turn off the computer to save the
laptop battery. Be sure to name the ﬁle differently when turning back on so it
will not overwrite the existing signatures.

- To ensure enough samples in a bad weather situation, the ASD light
source may be used if necessary when there is thick cloud cover.

- Move to two or three more sites in the lake to measure for the GPS
coordinates, spectral signatures, and the Secchi depth. No water sample is
taken at these sites. Turn off both laptop and ASD after ﬁnish the last site on
the lake to save the batteries.

In summary, for each lake 3-4 sites are measured for spectral signature
and SD, but only one deepest basin site is measured for Chl-a, TSS, algae,
nutrients, and DOC. Depending on stratiﬁcation of the lakes, two or three sets
of Chl-a, TSS, nutrients, and algae samples are taken from each lake at the

same site but at different depth.

201

 

 

APPENDIX B
AISA IMAGERY

AISA 20 bands image

 

Long Lake (Grand Traverse) Silver Lake
:2 Q! "‘-

  

  
  

 
  
   

...5.
Lake

.-ef’.’
arl

5.».

   

‘u'lb 9:3 is 3"».

Lake Mitchell .. Mecosta and Round Lakes

 

e . if" . _
Kimball Lake Fremont Lake

203

     

.4

r

 
   
 

 

,. "J
. ir-—s'a‘.-.r"

Muskegon Lake - ... Houghton Lake

 
   

'4‘.

Rimball Lake I Brooks and Hess Lakes

204

  

 

 

Roger Dam Pond Clear Lake (Mecosta)

 

. “ . ‘7 u . .. u L
orsehead Lake Jehnsen Lake

205

I at“! 4‘; 5x.
Sapphire Lake

      

”at!

. Arbutus Lake Long Lake (Grand Traverse)

ﬂ

1

   

206

 

59353.5
8o 89 coo

 

o>=a>teo «or 2:3 Sc oucﬁoecem eE:.e>

 

ll .. l ~86 _ .--

595.653
8» can can con 8..

 

5,
9

0
940341199 “I

    

. .- .. --- .1);- . l. 8.0
225.. 5.. 8:83:02 eE:_o>

eeueroeuei SuﬂllOA

 

 

 

59.0.05;

com 89 com con 8* com

 

 

053:2... new «.33 En coca—coco: eE:_o>

59.0.05;
can con 08 com o3

 

0
940341199 131

 

 

 

 

:-. - .-.1.---. . l.l-.l :1. on...
«.23 En conﬂuecem oE=.e>

eouepeuer eumloA

 

 

 

ZOmEA—SOO mMHBA—h A4>OSHN Hmuoz

O nanny—Hm:

207

 

 

59.12925

 

 

com con 000 com oov

 

”WEN-WP Isl

 

 

 

 

 

 

¢>=u>t¢u “up 223 Sn 3:800:3— cE=_o>

 

£322.25
can con coo com

 

£26.. 5.. oocsgccm aE:_o>

cad

ofo

scum oumloA

 

208

 

59.2033
8m 8m 2:. So can

 

 

ooo

59.2033
can So 8w.

 

 

 

 

 

 

A
. L w
u . n
w m
m. m
P.
m m
U
9
0
- . -. - 1 1. - ..1 . 1 «Ed 7 1- .1 11 .1-1 1 .1 .1 - and
0333.3 «3 226... «an 8:800:3— aE=_o> 205.. «at 35302.3. 25...;
- mil—34235 1 59.22.25
com can con 08 com 8v 25 com can 8» 8m 8..
. . _ . _ . «8.9- *- . . 1 . mod-
A
a _ 83. M
r _ V0001 Iv Wﬂmﬁ.wm/ﬁqll-l-,. 1.-1- M
_ E 7,. , i w . /!,,1- . m
0 ., . _ . _ i. .1 w
_ 1.. > . , _ _ . _ . m. .
__ ‘ ‘ .. 80.0 I a
. . i. f, . , . . 0 .. 4 , W. N
_ < z, .. . _ m . w
, .0 V8.0 9
u
m
‘ 8o... .
1 1 1 11 1 «:8 m .
1 .80

 

3:52.00 «3 «.33 «at 3:83:3— 2.5...

>

 

 

 

N—M>0.J NAU- OOCSOf—MWK ﬁrm—“1:1;-

 

 

 

 

 

 

209

 

593.033.

000 can com com oov

 

 

 

 

 

 

595.353
08 con 08 com

 

 

gyro-4 «at 8:83.31 vEr.o>

  

. ood

 

 

mum emnloA

 

210

 

59.0.0225

8m con 80 com 8*

 

3.3330 «3 226.. 3.. 3:80.53. oE=.o>

 

 

 

 

 

 

 

1593.32» .

com can 000 com oov

1

 

 

 

c>=a>too «or 203.. an... 3.58:3. oE:_o>

593.033

 

 

 

 

 

 

 

 

 

 

 

So 8» 2: 8» 8m 2:

_ . . . no...
A
W
n

. w
9
W
m
_- 1 1.1.- 1.1 1 1 1. 11 1 and
233 new oocﬂoacoz oE:.o>
522.55
com com 2: a8 8m 2:.

T _ . _ _ 8.0
A
0|

. n
m
9
W
U
0
. O
1 1. 1. 1.. 1 .93

«.25.. 3.. 3:58:01 oE:.o>

 

 

 

211

 

new

5223!:
:8 8‘. 8o 8... 8v

. 1 . .11. 1.- .1153...
a>za>too «up 230.. ma: 3:83.31 cE:.o>

 

 

ooa com com coo oom oov

595.052.

 

 

 

 

 

 

 

 

51:035....
8» 85 coo 8m 8..

1

 

 

 

 

 

 

 

“92111199 I“

 

 

     

              

r..¥|’

. ...ulﬂ mhnhl1...umlbls I!

I 1.2.?

rt
,

1‘.

com com 02. com com 00?

 

.1 11 .1- 11. -1...

«.03.. man 3:83.31 35:5

0
".
0

In
".
O

o
N.
o

 

 

 

212

APPENDIX D
AREA UNDER THE SPECTRAL CURVE

Area under the curve need in area index agolay altered model data set

 

 

 

 

Band A B C D E l" G
Lake 435-475 470-570 565-620 670-690 675-700 690-740 700-740

BEAR0819 0.0201 0.1189 0.0464 0.0189 0.0236 0.1153 0.0107
BELL0826 0.0169 0.1615 0.0386 0.0326 0.0401 0.1413 0.0408

8160820 0.0174 0.1378 0.0358 0.0178 0.0225 0.1205 0.0153
BR000526 0.0132 0.1425 0.0305 0.0198 0.0275 0. 1162 0.0135
CASS0810 0.0142 0.1347 0.0303 0.0176 0.0229 0.1190 0.0144
CHEM0826 0.0143 0.1333 0.0304 0.0186 0.0221 0.1201 0.0146
CHIP0820 0.0124 0.1403 0.0384 0.0183 0.0234 0.1208 0.0160
CLSJO603 0.0 188 0.1405 0.0452 0.0187 0.0243 0.1208 0.0164
CRO’I‘0605 0.0109 0.1245 0.0189 0.0195 0.0245 0.1217 0.0173
CUBO819 0.0200 0.1348 0.0470 0.0181 0.0221 0.1185 0.0130
DIAM0824 0.0246 0.1403 0.0740 0.0168 0.0214 0.1264 0.0206
EAGL0824 0.0157 0.1361 0.0390 0.0166 0.0208 0.1189 0.0137
FORD0826 0.0280 0.1910 0.0533 0.0384 0.0649 0.1531 0.0675
GLEN0819 0.0347 0.1004 0.0899 0.0154 0.0196 0.1226 0.0167
HESSOS26 0.0363 0.1815 0.0582 0.0225 0.0402 0.1549 0.0586
HICK0820 0.0227 0.1734 0.0398 0.0293 0.0432 0.1509 0.0536
HIGGO608 0.0213 0.1159 0.0435 0.0173 0.0233 0.1145 0.0106
KENT0826 0.0 150 0.1458 0.0332 0.0220 0.0259 0.1299 0.0245
KIMBOS26 0.0272 0.1665 0.0195 0.0144 0.0171 0.1573 0.0491
KLIN0603 0.0203 0.1251 0.0476 0.0172 0.0228 0.1172 0.0126
LOON0810 0.0158 0.1375 0.0314 0.0182 0.0238 0.1176 0.0132
LOTU0807 0.0301 0.1569 0.0716 0.0147 0.0175 0.1313 0.0238
MACE0807 0.0278 0.1495 0.0654 0.0168 0.0212 0.1218 0.0162
MEC00608 0.0141 0.1295 0.0287 0.0179 0.0231 0.1205 0.0158
MECOO724 0.0224 0.1551 0.0385 0.0205 0.0216 0.1290 0.0210
AWR10728 0.0140 0.1361 0.0254 0.0213 0.0284 0.1277 0.0248
OAKL0807 0.0160 0.1340 0.0298 0.0170 0.0219 0.1179 0.0131
ORCH0810 0.0280 0.1478 0.0685 0.0160 0.0191 0.1259 0.0189
PAWP0824 0.0244 0.1710 0.0594 0.0228 0.0296 0.1442 0.0403
PICK0526 0.0255 0.1729 0.0483 0.0261 0.0408 0.1471 0.0502
PONT0807 0.0274 0.1721 0.0458 0.0198 0.0252 0.1425 0.0371
RAND0603 0.0130 0.1294 0.0278 0.0198 0.0247 0.1188 0.0144
ROGE0519 0.0147 0.1344 0.0153 0.0193 0.0245 0.1285 0.0243
SILV0725 0.0264 0.1465 0.0582 0.0187 0.0230 0.1228 0.0174

 

 

 

 

 

 

 

 

213

 

Area under the curve need in area index agolay ﬁltered validation data set

 

 

 

 

Band A B C D E l" G

Lake 435-475 470-570 565-620 670-690 675-700 690-740 700-740
ARBU0726 0.0193 0.1334 0.0545 0.0163 0.0217 0.1199 0.0150
BROOO7 27 0.0297 0.1754 0.0388 0.0217 0.0270 0.1697 0.0637
CLME0724 0.0167 0.1403 0.0381 0.0185 0.0223 0.1243 0.0185
HESSO727 0.0427 0.1982 0.0415 0.0207 0.0281 0.1846 0.0797
HIGG0727 0.0229 0.1095 0.0464 0.0174 0.0236 0.1127 0.0090
HOUGO726 0.0165 0.1432 0.0309 0.0176 0.0221 0.1254 0.0200
JEHN0519 0.0160 0.1413 0.0302 0.0189 0.0235 0.1231 0.0180
MITC0725 0.0164 0.1412 0.0244 0.0183 0.0231 0.1303 0.0251
MUSK0728 0.012 1 0.1305 0.0231 0.0233 0.0302 0.1263 0.0237
ROGE0724 0.0187 0.1456 0.0224 0.0201 0.0239 0.1348 0.0291
SAPP0725 0.0152 0.1360 0.0332 0.0189 0.0251 0.1200 0.0161
TAMA0605 0.0118 0.1273 0.0242 0.0199 0.0242 0.1184 0.0133

 

 

 

 

 

 

 

 

214

 

Area under the curve need in area index dbl ﬁltered model data aet

 

 

 

 

Band A B C D l" 0
Lake 435-485 485-575 550-630 675-700 690-745 700-745

BEAR08 19 0.1893 0.3664 0.4808 0.1100 0.4328 0.0978
BELL0826 0.1781 0.5562 0.4315 0.1552 0.4967 0. 1710
BIG0820 0.1638 0.4882 0.4465 0.0892 0.4571 0. 1097
BR000526 0.1295 0.4497 0.3803 0.1180 0.4223 0.0894
CASSOSIO 0.139 1 0.455 1 0.3962 0.1056 0.439 1 0.1018
CHEM0826 0.1322 0.4784 0.4374 0.0897 0.4602 0.1124
CHIP0820 0.1404 0.5042 0.4579 0.0903 0.463 1 0. 1219
CLSJ0603 0.1820 0.5054 0.4556 0.1030 0.4824 0.1367
CROT0605 0.1151 0.4034 0.3644 0.1067 0.4807 0.1404
CUB0819 0.179 1 0.4693 0.5153 0.0937 0.4744 0.1175
DIAM0824 0.2078 0.4756 0.5168 0.0947 0.4714 0.1279
EAGL0824 0.1566 0.4786 0.4323 0.0924 0.4547 0.1 149
FORDO826 0.2328 0.5876 0.4819 0.1885 0.5192 0.2120
GLEN0819 0.2436 0.2622 0.5618 0.0926 0.4766 0.1261
HESS0526 0.2350 0.5584 0.4794 0.1504 0.5019 0.1854
HICK0820 0.2133 0.6205 0.4299 0.1769 0.5304 0.2173
HIG60608 0.2024 0.3394 0.4702 0.0899 0.4442 0.1023
KENT0826 0.1569 0.5148 0.4120 0.1110 0.4833 0.1306
KIMBOS26 0.2041 0.5413 0.3436 0.0846 0.4951 0.1530
KLIN0603 0. 1850 0.3899 0.4859 0.0905 0.4438 0.1097
LOON0810 0.1573 0.4954 0.4072 0.1022 0.4594 0.1 171
LOTU0807 0.2245 0.5154 0.5119 0.0747 0.4828 0.1313
MACE0807 0.1904 0.4753 0.4469 0.0924 0.445 1 0.1014
MEC00608 0.1370 0.4419 0.4037 0.1077 0.4737 0. 1368
MEC007 24 0.1970 0.6308 0.4667 0.1049 0.4998 0.1455
AWR10728 0.1447 0.4895 0.4010 0.1242 0.4759 0.1400
OAKL0807 0.1606 0.4713 0.4266 0.0790 0.4795 0.1226
ORCH0810 0.2727 0.4788 0.5624 0.0860 0.4804 0.1217
PAWP0824 0.2064 0.5942 0.4947 0.1307 0.5355 0.1866
P1CK0526 0.2194 0.5948 0.4734 0.1708 0.5244 0.2187
PONT0807 0.2397 0.5915 0.4345 0.1130 0.5210 0.1692
RAND0603 0.1335 0.4357 0.4143 0.1005 0.4853 0.1340
ROGE0519 0.1428 0.4442 0.3335 0.1123 0.4328 0.0981
SILV0725 0.2386 0.5051 0.5399 0.0989 0.4879 0.1319

 

 

 

 

 

 

 

215

 

Area under the curve naed in area index dbl filtered validation data set

 

 

 

 

Band A B C D F G
Wavelength 435-485 485-575 550-630 675-700 690-745 700-745
ARBUO726 0.1674 0.4489 0.4517 0.0916 0.4496 0.1126
BR000727 0.2673 0.7211 0.4611 0.1501 0.6208 0.2688
CLME0724 0.1566 0.5044 0.4420 0.1070 0.4673 0.1263
HESSO727 0.3171 0.8309 0.4566 0.1496 0.6180 0.2723
HIGG0727 0.1983 0.3219 0.4685 0.0995 0.4264 0.0920
HOUG0726 0.1496 0.5164 0.3987 0.1046 0.4661 0.1231
JEHN0519 0.1534 0.4568 0.3756 0.1122 0.4353 0.0997
MITC0725 0.1615 0.5078 0.3796 0.0951 0.4743 0.1324
MUSK0728 0.1234 0.4334 0.3888 0.1298 0.4523 0.1234
ROGE0724 0.1815 0.4970 0.3642 0.1151 0.4692 0.1282
SAPP0725 0.1499 0.4728 0.4066 0.1112 0.4512 0.1111
TAMA0605 0.1065 0.4322 0.4067 0.1035 0.4733 0.1163

 

 

 

 

 

 

 

216

 

Area under the curve used in area index a8 ﬁltered model data set

 

 

 

 

Band A B C D E l" 0
Lake 435-475 470-570 565-620 670-690 675-700 690-740 700-740

BEAR0819 0.0608 0.080 1 0.0506 0.0697 0.0781 0.1429 0.0403
BELLO826 0.0572 0.1215 0.0380 0.0850 0.0916 0.1695 0.07 12
BIG0820 0.06 16 0.0994 0.0375 0.0664 0.0739 0. 1480 0.0454
BR000526 0.0613 0.1053 0.0322 0.0702 0.0778 0.1440 0.0440
CASSO810 0.0607 0.0970 0.0323 0.0672 0.0749 0. 1467 0.0446
CHEM0826 0.0532 0.0964 0.0312 0.0682 0.0755 0. 1485 0.0452
CHIP0820 0.0569 0.1021 0.0391 0.0669 0.0746 0.1486 0.0458
CLSJ0603 0.0584 0.1014 0.0471 0.0679 0.0756 0. 1486 0.0463
CRO’I‘0605 0.0573 0.0880 0.0223 0.0700 0.0777 0.1493 0.0474
CUBOS 19 0.0593 0.0952 0.0493 0.0670 0.0747 0. 1456 0.0419
DIAM0824 0.061 1 0.1005 0.0763 0.0634 0.070 1 0. 1543 0.0508
EAGL0824 0.0556 0.0975 0.0386 0.0664 0.0747 0.1474 0.0438
FOR00826 0.0629 0.1476 0.0508 0.0985 0.1080 0. 1803 0.1001
GLEN0819 0.0769 0.0592 0.0942 0.0631 0.0703 0.1503 0.0467
HESSOS26 0.0755 0.1364 0.0563 0.0704 0.0767 0.1832 0.09 17
H1CK0820 0.0614 0.1322 0.0384 0.0805 0.0866 0.1790 0.0845
HIGGO608 0.0647 0.0767 0.0471 0.0681 0.0767 0.1423 0.0404
KENT0826 0.0589 0.1067 0.0344 0.0700 0.0762 0. 1581 0.0547
KIMBOS26 0.0737 0.1216 0.0214 0.0582 0.0628 0.1852 0.0795
KLIN0603 0.0647 0.0848 0.0506 0.067 2 0.0751 0.1447 0.0429
LOON08 10 0.0607 0.0989 0.0330 0.0681 0.0758 0.1455 0.0432
LOTU0807 0.0686 0.1 1 19 0.0727 0.0598 0.0659 0. 1596 0.0539
MACE0807 0.0684 0. 1060 0.0673 0.0655 0.0730 0.1495 0.0462
MEC00608 0.0598 0.0923 0.031 1 0.0673 0.0748 0. 1482 0.0460
MECOO724 0.0654 0.1150 0.0405 0.0651 0.0709 0.1561 0.0497
AWR10728 0.0576 0.0981 0.0278 0.0716 0.0788 0.1556 0.0553
OAKL0807 0.0584 0.0955 0.0317 0.0678 0.0756 0. 1443 0.0422
ORCH0810 0.0656 0.1047 0.0702 0.0631 0.0691 0.1536 0.0483
PAWP0824 0.0582 0.1306 0.0578 0.0680 0.0731 0.1722 0.0702
PICK0526 0.0624 0.1308 0.0489 0.0780 0.0852 0. 1745 0.0807
PONT0807 0.0664 0.1284 0.0466 0.0649 0.0699 0.1708 0.0675
RAND0603 0.0551 0.0916 0.0300 0.0704 0.0782 0.1469 0.0443
ROGEOS 19 0.0609 0.0962 0.0184 0.0705 0.0779 0.1563 0.0546
SILV07 25 0.0628 0.1053 0.0604 0.0664 0.0736 0.1501 0.0467

 

 

 

 

 

 

 

 

217

 

Area under the curve need in area index 88 ﬁltered validation data set

 

 

 

 

Band A B C D E l" G
Lake 435-475 470-570 565-620 670-690 675-700 690-740 700-740
ARBU0726 0.0606 0.0949 0.0568 0.0653 0.0728 0. 1477 0.0452
BR000727 0.0608 0. 1307 0.0388 0.0635 0.0664 0. 1967 0.0927
CLME0724 0.0573 0.1018 0.0401 0.0670 0.0742 0.1519 0.0485
HESSO727 0.0723 0.1501 0.0405 0.0630 0.0659 0.2134 0.1109
HIGG0727 0.0670 0.0710 0.0509 0.0689 0.0776 0.1404 0.0394
HOUG0726 0.0604 0.1049 0.0324 0.0652 0.0724 0. 153 1 0.0498
JEHN0519 0.0622 0.1025 0.0319 0.0681 0.0752 0.1509 0.048 1
MITC0725 0.0619 0.1022 0.0263 0.0667 0.0734 0.1581 0.0554
MUSK0728 0.0575 0.0934 0.0262 0.0748 0.0823 0.1541 0.0542
ROGE0724 0.0627 0.1065 0.0243 0.0680 0.0741 0.1626 0.0590
SAPP0725 0.0594 0.0981 0.0351 0.0685 0.0760 0. 1478 0.0462
TAMA0605 0.0536 0.0905 0.0266 0.0697 0.0771 0.1460 0.0431

 

 

 

 

 

 

 

 

218

 

APPENDIX E
MAXIMUM SPECTRAL CURVE HEIGHT

Maximum height need in height] width index egolay ﬁltered model data aet

 

 

 

 

Band A B C D E F G
Lake 435-47 5 470-570 565-620 670-690 675-700 690-7 40 700-740

BEAR0819 0.0005 0.0013 0.0011 0.0011 0.0012 0.0024 0.0004
BELLO826 0.0005 0.0019 0.0009 0.0018 0.0019 0.0034 0.0014
BIG0820 0.0004 0.0014 0.0009 0.0009 0.0011 0.0025 0.0005
BR000526 0.0003 0.0015 0.0007 0.0011 0.0012 0.0025 0.0005
CASS0810 0.0004 0.0014 0.0007 0.0009 0.0010 0.0024 0.0004
CHEM0826 0.0004 0.0014 0.0007 0.0011 0.0011 0.0025 0.0005
CHIP0820 0.0004 0.0015 0.0010 0.0010 0.0012 0.0025 0.0005
CLSJ0603 0.0005 0.0015 0.0011 0.0010 0.0011 0.0025 0.0005
CROT0605 0.0004 0.0013 0.0004 0.0010 0.0012 0.0025 0.0005
CU80819 0.0005 0.0015 0.0011 0.0010 0.0011 0.0025 0.0005
DIAM0824 0.0007 0.0017 0.0018 0.0009 0.0010 0.0026 0.0006
EAGL0824 0.0005 0.0014 0.0009 0.0011 0.0010 0.0025 0.0004
FORD0826 0.0008 0.0023 0.0012 0.0033 0.0032 0.0044 0.0026
GLEN0819 0.0009 0.0015 0.0023 0.0009 0.0009 0.0025 0.0005
HESS0526 0.0010 0.0022 0.0013 0.0020 0.0020 0.0042 0.0022
HICK0820 0.0006 0.0020 0.0009 0.0020 0.0021 0.0037 0.0017
HIGGO608 0.0006 0.0014 0.0010 0.0009 0.0010 0.0023 0.0003
KENT0826 0.0004 0.0016 0.0008 0.0012 0.0013 0.0027 0.0007
KIMBOS26 0.0007 0.0017 0.0005 0.0008 0.0009 0.0035 0.0015
KLIN0603 0.0005 0.0014 0.0011 0.0009 0.0010 0.0024 0.0003
LOON0810 0.0004 0.0014 0.0007 0.0009 0.0011 0.0024 0.0005
LOTU0807 0.0008 0.0018 0.0019 0.0008 0.0009 0.0028 0.0008
MACE0807 0.0007 0.0017 0.0017 0.0009 0.0010 0.0025 0.0005
MEC00608 0.0004 0.0013 0.0006 0.0009 0.0010 0.0024 0.0004
MECOO724 0.0007 0.0016 0.0010 0.0011 0.0012 0.0028 0.0009
AWRIO728 0.0004 0.0014 0.0005 0.0012 0.0013 0.0028 0.0008
OAKL0807 0.0004 0.0014 0.0007 0.0010 0.0011 0.0024 0.0003
ORCH0810 0.0008 0.0017 0.0018 0.0009 0.0011 0.0026 0.0006
PAWP0824 0.0007 0.0020 0.0015 0.0014 0.0014 0.0033 0.0013
PICK0526 0.0007 0.0020 0.0011 0.0020 0.0020 0.0037 0.0017
PONT0807 0.0007 0.0019 0.0012 0.0011 0.0012 0.0032 0.0011
RAND0603 0.0004 0.0013 0.0006 0.0011 0.0012 0.0024 0.0004
ROGE0519 0.0004 0.0014 0.0003 0.0010 0.0011 0.0027 0.0007
SILV0725 0.0007 0.0017 0.0015 0.0010 0.0011 0.0025 0.0005

 

 

 

 

 

 

 

 

219

 

Maximum height need in height] width index agolay ﬁltered validation data set

 

 

 

 

Band A B C D E F G
Lake 435-475 470-570 565-620 670-690 67 5-700 690-740 700-740
ARBU0726 0.0005 0.0015 0.0013 0.0008 0.0010 0.0025 0.0005
BR000727 0.0008 0.0019 0.0010 0.0013 0.0014 0.0039 0.0018
CLME0724 0.0005 0.0015 0.0009 0.0010 0.0011 0.0025 0.0006
HESSO727 0.0012 0.0022 0.0011 0.0013 0.0014 0.0046 0.0027
HIGG0727 0.0006 0.0013 0.0011 0.0009 0.0010 0.0023 0.0003
HOUG0726 0.0005 0.0015 0.0007 0.0009 0.001 1 0.0026 0.0006
JEHN0519 0.0004 0.0015 0.0007 0.0010 0.0011 0.0026 0.0006
MITC0725 0.0004 0.0015 0.0005 0.0010 0.0011 0.0028 0.0008
MUSK0728 0.0003 0.0014 0.0005 0.0012 0.0013 0.0028 0.0008
ROGE0724 0.0005 0.0015 0.0006 0.0011 0.0012 0.0029 0.0009
SAPP0725 0.0004 0.0014 0.0007 0.0010 0.0011 0.0025 0.0005
TAMA0605 0.0003 0.0013 0.0005 0.0010 0.0012 0.0025 0.0005

 

 

 

 

 

 

 

 

 

220

Maximum height used in height/width index dhl ﬁltered model data set

 

 

 

 

Band A B C D I" G
Lake 435-485 485-5 75 550-630 67 5-700 690-7 45 700-745

BEAR08 19 0.0038 0.0048 0.0068 0.0067 0.0081 0.0032
BELL0826 0.0047 0.0071 0.0065 0.0069 0.0 108 0.0050
BIG0820 0.0036 0.0059 0.0066 0.0063 0.0089 0.0029
BR000526 0.0027 0.0052 0.005 1 0.0050 0.0080 0.0023
CASSOB 10 0.0030 0.0053 0.0054 0.0046 0.0082 0.0025
CHEM0826 0.0032 0.0056 0.0066 0.0073 0.0090 0.0030
CHIP0820 0.0050 0.0065 0.0069 0.0053 0.0089 0.0034
CLSJ0603 0.0041 0.0069 0.0066 0.0069 0.0092 0.0035
CROT0605 0.0029 0.0047 0.0046 0.0078 0.0094 0.0040
CUBO819 0.0037 0.0061 0.0077 0.0095 0.0102 0.0032
DIAM0824 0.0045 0.0064 0.0073 0.0048 0.0088 0.0034
EAGL0824 0.0038 0.0060 0.0063 0.0118 0.0087 0.0031
FORD0826 0.0053 0.0076 0.0076 0.0089 0.01 19 0.0065
GLEN0819 0.0050 0.0037 0.0078 0.0061 0.0093 0.0035
HESS0526 0.0049 0.0072 0.007 1 0.0074 0.0108 0.0050
HICK0820 0.0050 0.0079 0.0070 0.0078 0.01 13 0.0057
HIGGO608 0.0045 0.0044 0.0062 0.0044 0.0081 0.0025
KEN'I‘0826 0.0042 0.0060 0.0062 0.0055 0.0097 0.0036
KIMBOS26 0.0042 0.0064 0.0052 0.0041 0.0095 0.0038
KLIN0603 0.0040 0.0051 0.0065 0.0048 0.0084 0.0029
LOON0810 0.0035 0.0060 0.0060 0.0064 0.0088 0.0031
LOTU0807 0.0051 0.0067 0.0076 0.0050 0.0094 0.0038
MACE0807 0.0040 0.0059 0.0064 0.0056 0.0087 0.0025
MECOO608 0.0029 0.0052 0.0055 0.0055 0.0090 0.0034
MECOO724 0.0050 0.0079 0.0084 0.0094 0.0 107 0.0071
AWRIO728 0.0032 0.0060 0.0057 0.0057 0.0093 0.0035
OAKL0807 0.0035 0.0053 0.0061 0.0045 0.0 101 0.0029
ORCH0810 0.0061 0.0069 0.0086 0.0089 0.0102 0.0029
PAWP0824 0.0045 0.0078 0.0083 0.0061 0.0105 0.0049
PICK0526 0.0045 0.0077 0.0077 0.0077 0.01 16 0.0057
PONT0807 0.0061 0.0073 0.0068 0.0054 0.0100 0.0043
RAN D0603 0.0032 0.0051 0.0058 0.0065 0.0093 0.0034
ROGE0519 0.0030 0.0050 0.0043 0.0052 0.0080 0.0022
SILV0725 0.0058 0.0067 0.008 1 0.0091 0.0108 0.0051

 

 

 

 

 

 

 

221

 

Maximum height used in height/width index dbl ﬁltered validation data set

 

 

 

 

Band A B C D l" G
Lake 435-485 485-575 550-630 675-700 690-745 700-745
ARBU07 26 0.0036 0.0055 0.0062 0.0040 0.0084 0.0026
BR0007 27 0.0067 0.0101 0.0080 0.0075 0.0 124 0.0067
CLME0724 0.0033 0.0061 0.0070 0.0078 0.0088 0.0032
HESSO727 0.0078 0.0113 0.0086 0.0071 0.0138 0.0082
HIGG07 27 0.0043 0.0040 0.0064 0.0061 0.0081 0.0025
HOUGO7 26 0.0032 0.0062 0.0060 0.007 3 0.0090 0.0032
JEHN0519 0.0032 0.0051 0.0051 0.0057 0.0081 0.0024
MITC07 25 0.0041 0.0059 0.0056 0.0049 0.0091 0.0041
MUSK07 28 0.0030 0.0052 0.005 1 0.0056 0.0089 0.0032
ROGE0724 0.0042 0.0057 0.0053 0.007 8 0.0089 0.0032
SAPP0725 0.0031 0.0057 0.0056 0.0055 0.0084 0.0027
TAMA0605 0.0029 0.0052 0.0057 0.0092 0.0101 0.0028

 

 

 

 

 

 

 

222

 

Maximum height need in height/width index :8 ﬁltered model data set

 

 

 

 

Band A B C D E F G
Lake 435-47 5 470-570 565-620 670-690 67 5-700 690-740 700-740

BEAR0819 0.0017 0.0011 0.0013 0.0019 0.0019 0.0031 0.0010
BELL0826 0.0016 0.0017 0.0009 0.0027 0.0026 0.0040 0.0022

3160820 0.0017 0.0013 0.0010 0.0018 0.0018 0.0031 0.0012
BR000526 0.0016 0.0014 0.0008 0.0019 0.0019 0.0030 0.0013
CASSOSIO 0.0016 0.0012 0.0008 0.0018 0.0017 0.0030 0.0012
CHEM0826 0.0017 0.0012 0.0008 0.0018 0.0018 0.0031 0.0013
CHIP0820 0.0016 0.0014 0.0011 0.0018 0.0018 0.0031 0.0013
CLSJO603 0.0016 0.0014 0.0012 0.0018 0.0018 0.0030 0.0012
CRO’I‘0605 0.0016 0.0011 0.0005 0.0018 0.0018 0.0030 0.0012
CU80819 0.0018 0.0013 0.0013 0.0018 0.0018 0.0031 0.0011
DIAM0824 0.0018 0.0015 0.0021 0.0017 0.0017 0.0032 0.0014
EAGL0824 0.0015 0.0013 0.0011 0.0018 0.0019 0.0032 0.0011
FORD0826 0.0018 0.0023 0.0013 0.0038 0.0036 0.0052 0.0035
GLEN0819 0.0021 0.0009 0.0026 0.0017 0.0016 0.0031 0.0013
HESS0526 0.0021 0.0021 0.0014 0.0025 0.0025 0.0049 0.0030
HICK0820 0.0018 0.0019 0.0010 0.0027 0.0027 0.0044 0.0026
HIGG0608 0.0017 0.0010 0.0012 0.0018 0.0018 0.0030 0.0010
KENT0826 0.0017 0.0014 0.0009 0.0020 0.0020 0.0033 0.0015
KIM80526 0.0019 0.0016 0.0006 0.0016 0.0016 0.0041 0.0023
KLIN0603 0.0017 0.0012 0.0013 0.0017 0.0017 0.0029 0.0011
LOON0810 0.0016 0.0013 0.0008 0.0018 0.0018 0.0030 0.0012
LOTU0807 0.0019 0.0016 0.0021 0.0016 0.0016 0.0033 0.0015
MACE0807 0.0019 0.0015 0.0019 0.0017 0.0017 0.0031 0.0012
MECOO608 0.0015 0.0012 0.0008 0.0018 0.0018 0.0030 0.0012
MECOO724 0.0019 0.0015 0.0011 0.0020 0.0019 0.0035 0.0016
AWR10728 0.0015 0.0013 0.0006 0.0020 0.0020 0.0034 0.0015
OAKL0807 0.0015 0.0012 0.0008 0.0019 0.0019 0.0030 0.0011
ORCH0810 0.0018 0.0015 0.0019 0.0017 0.0017 0.0033 0.0013
PAWP0824 0.0017 0.0019 0.0016 0.0022 0.0022 0.0039 0.0021
PICK0526 0.0018 0.0019 0.0013 0.0025 0.0025 0.0043 0.0026
PONT0807 0.0018 0.0018 0.0013 0.0020 0.0019 0.0038 0.0019
RAND0603 0.0015 0.0012 0.0007 0.0019 0.0019 0.0030 0.0012
ROGE0519 0.0016 0.0012 0.0004 0.0019 0.0018 0.0033 0.0015
SILV0725 0.0022 0.0015 0.0017 0.0018 0.0018 0.0031 0.0013

 

 

 

 

 

 

 

 

223

 

Maximum height uaed in heigfht/width index a8 ﬁltered validation data set

 

 

 

 

Band A B C D E F G

Lake 435-475 470-570 565-620 670-690 675-700 690-740 700-740
ARBUO726 0.0017 0.0013 0.0015 0.0017 0.0017 0.0030 0.0012
BR000727 0.0017 0.0018 0.0011 0.0021 0.0021 0.0045 0.0027
CLME0724 0.0015 0.0013 0.0011 0.0018 0.0018 0.0031 0.0013
HESSO727 0.0023 0.002 1 0.0012 0.0021 0.0021 0.0052 0.0035
HIGG0727 0.0017 0.0010 0.0013 0.0018 0.0018 0.0028 0.0010
HOUGO726 0.0021 0.0014 0.0008 0.0018 0.0017 0.0032 0.0014
JEHNOS 19 0.0016 0.0013 0.0008 0.0018 0.0018 0.0031 0.0013
MITC0725 0.0016 0.0013 0.0006 0.0018 0.0018 0.0033 0.0015
MUSK07 28 0.00 16 0.0012 0.0006 0.0021 0.002 1 0.0033 0.0016
ROGE0724 0.0017 0.0014 0.0006 0.0019 0.0019 0.0035 0.0017
SAPP0725 0.0016 0.0013 0.0009 0.0018 0.0018 0.0031 0.0012
TAMA0605 0.0016 0.0012 0.0006 0.0019 0.0019 0.0031 0.0012

 

 

 

 

 

 

 

 

224

 

APPENDIX F

NARROW-WAVEBAND VOLUME REFLECTAN CE

Volume reﬂectance need in narrow-band index agrolay ﬁltered model data aet

 

 

 

 

 

 

 

 

 

 

 

 

Band A B C D E F G H
Lake 435 455 470 565 670 675 690 700

BEAR0819 0.04 10 0.0438 0.0462 0.0419 0.0094 0.0099 0.0096 0.0076
BELL0826 0.0464 0.0476 0.0486 0.0845 0.0426 0.0442 0.0560 0.0586
BIG0820 0.0309 0.0325 0.0338 0.0457 0.0207 0.0206 0.0197 0.0172
BR000526 0.0108 0.0102 0.0099 0.0262 0.0087 0.0080 0.0094 0.0095
CASS0810 0.0335 0.0330 0.0333 0.0420 0.0238 0.0235 0.0225 0.0205
CHEM0826 0.0358 0.0352 0.0357 0.0429 0.0249 0.0256 0.0247 0.02 18
CH1P0820 0.0328 0.0315 0.0307 0.0449 0.0165 0.0162 0.0159 0.0137
CLSJ0603 0.0334 0.0353 0.0373 0.0530 0.0172 0.0170 0.0170 0.0153
CROT0605 0.0293 0.0267 0.0260 0.0246 0.0210 0.0211 0.0216 0.0198
CUBOB 19 0.0346 0.0372 0.0396 0.0503 0.0149 0.0152 0.0142 0.0113
DIAM0824 0.0794 0.0850 0.0889 0. 1067 0.0357 0.0350 0.0336 0.0305
EAGL0824 0.0506 0.0500 0.0516 0.0623 0.0342 0.0346 0.0320 0.0295
FORD0826 0.0429 0.0499 0.0552 0.1213 0.0627 0.0603 0.0807 0.0987
GLEN0819 0.1216 0.1326 0.1401 0.1203 0.0363 0.0359 0.0329 0.0295
HESS0526 0.0859 0.0979 0.1060 0.1629 0.0954 0.0913 0.0979 0. 1052
HICK0820 0.0355 0.0398 0.0431 0.0906 0.0577 0.0562 0.0673 0.0733
HIGGO608 0.0490 0.0520 0.0551 0.0474 0.0170 0.0168 0.0154 0.0141
KEN'I‘0826 0.0365 0.0364 0.0368 0.0565 0.0314 0.0318 0.0345 0.0318
KIMB0526 0.0445 0.0506 0.0556 0.0948 0.0839 0.0827 0.0794 0.0740
KLIN0603 0.0421 0.0447 0.0473 0.0486 0.0122 0.0119 0.0106 0.0086
LOON0810 0.0133 0.0135 0.0145 0.0261 0.0074 0.0073 0.0068 0.0052
LOTU0807 0.0642 0.0714 0.0780 0.1108 0.0418 0.0412 0.0377 0.0329
MACE0807 0.0364 0.0429 0.0482 0.0740 0.0173 0.017 1 0.0152 0.0123
MECOO608 0.0584 0.0581 0.0583 0.062 1 0.0447 0.0445 0.0437 0.0416
MECOO724 0.0318 0.0363 0.0388 0.0674 0.0373 0.0379 0.0390 0.0338
AWRIO728 0.0304 0.0300 0.0300 0.0402 0.0279 0.0275 0.0301 0.0300
OAKL0807 0.0265 0.0273 0.0278 0.0360 0.0199 0.0199 0.0180 0.0158
ORCH0810 0.0574 0.0633 0.0694 0.0935 0.0312 0.0308 0.0284 0.0241
PAWP0824 0.0576 0.0633 0.0671 0.1132 0.0525 0.0514 0.0560 0.0551
PICK0526 0.0346 0.0407 0.0448 0.093 1 0.0506 0.0487 0.0571 0.0634
PONT0807 0.0431 0.0491 0.0545 0.1007 0.0596 0.0585 0.0603 0.0579
RAND0603 0.0242 0.0226 0.0228 0.0262 0.0112 0.0116 0.0121 0.0104
ROGE0519 0.0278 0.0276 0.0281 0.0361 0.0398 0.0400 0.0402 0.0386
SILV0725 0.0415 0.0473 0.0523 0.0746 0.0252 0.0251 0.0250 0.0222

 

225

 

Volume reﬂectance need in narrow-band index agony ﬁltered validation data set

 

 

 

 

 

 

 

 

 

 

 

 

Band A B C D E F G H
Lake 435 455 470 565 670 675 690 700
ARBU0726 0.0653 0.0676 0.0698 0.0794 0.0328 0.0323 0.0302 0.0279
BR000727 0.0863 0.0929 0.0999 0.1491 0.1156 0.1140 0.1181 0.1153
CLME0724 0.0315 0.0316 0.0335 0.0485 0.0196 0.0198 0.0193 0.0162
HESSO727 0.0694 0.0818 0.0947 0.1663 0. 1263 0.1240 0.1278 0.1261
HIGG0727 0.0492 0.0534 0.0570 0.0436 0.0109 0.0 107 0.0094 0.0082
HOUG0726 0.0316 0.0316 0.0337 0.0507 0.0300 0.0296 0.0287 0.0259
JEHNOS 19 0.0222 0.0227 0.0236 0.0386 0.0196 0.0197 0.0197 0.0 173
MITC0725 0.0288 0.0295 0.0305 0.0455 0.0351 0.0348 0.0344 0.0320
MUSK0728 0.0372 0.0357 0.0352 0.0398 0.0285 0.0289 0.0328 0.0331
ROGE0724 0.0326 0.0349 0.0366 0.0552 0.043 1 0.0433 0.0443 0.0415
SAPP0725 0.0256 0.0259 0.0264 0.0369 0.0145 0.0142 0.0144 0.0133
TAMA0605 0.0270 0.0252 0.0246 0.0257 0.0156 0.0158 0.0167 0.0140

 

226

 

Volume reﬂectance used in narrow-band index dbl ﬁltered model data aet

 

 

 

 

Band A B C D E l" G H
Lake 435 450 455 480 550 670 690 700

BEAROB 19 0.0409 0.0434 0.0437 0.0471 0.0458 0.0094 0.0092 0.0073
BELL0826 0.0461 0.0473 0.0473 0.0489 0.08 13 0.0424 0.0569 0.0587
BIG0820 0.0312 0.0324 0.0324 0.0343 0.0445 0.0203 0.0191 0.0172
BR000526 0.0108 0.0103 0.0103 0.0097 0.0235 0.0086 0.0098 0.0099
CASSOB 10 0.0335 0.0330 0.0330 0.0334 0.0404 0.0239 0.0222 0.0204
CHEM0826 0.0358 0.0353 0.0353 0.0351 0.0402 0.0253 0.0237 0.0219
CHIP0820 0.0323 0.0318 0.0318 0.0305 0.0429 0.0 164 0.0157 0.0129
CLSJ0603 0.0331 0.0353 0.0353 0.0379 0.0526 0.0 174 0.0168 0.0150
CROT0605 0.0286 0.0273 0.0273 0.0261 0.0245 0.0209 0.0212 0.0196
CUBO819 0.0345 0.0366 0.0366 0.0403 0.0508 0.0150 0.0136 0.0110
DIAM0824 0.0795 0.0841 0.0844 0.0899 0.1099 0.0361 0.0333 0.0309
EAGLO824 0.0504 0.0500 0.0500 0.0519 0.0613 0.0351 0.0304 0.0295
FORD0826 0.0428 0.0485 0.0494 0.0570 0.1 157 0.0627 0.080 1 0.1003
GLEN0819 0.12 17 0.1308 0.1321 0.1427 0.1283 0.0369 0.0319 0.0298
HESSOS26 0.0861 0.0958 0.0974 0.1089 0.1607 0.0948 0.0974 0.1057
HICK0820 0.0351 0.0391 0.0391 0.0441 0.0844 0.0586 0.0687 0.0737
HIGGO608 0.0493 0.0515 0.0515 0.0560 0.0508 0.0173 0.0155 0.0136
KENT0826 0.0362 0.0365 0.0365 0.0371 0.0537 0.0314 0.0344 0.0322
KIMBOS26 0.0446 0.0494 0.0502 0.0583 0.0890 0.0841 0.0790 0.0747
KLIN0603 0.0424 0.0443 0.0444 0.0484 0.0514 0.0122 0.0103 0.0086
LOON0810 0.0 133 0.0 134 0.0134 0.0149 0.0246 0.0079 0.0066 0.0049
LOT00807 0.0643 0.0702 0.0704 0.0809 0.1 102 0.0422 0.0373 0.0332
MACE0807 0.0362 0.0414 0.0425 0.0503 0.0732 0.0175 0.0150 0.0123
MECOO608 0.0581 0.0580 0.0580 0.0582 0.0615 0.0449 0.0435 0.0420
MECOO7 24 0.0326 0.0374 0.0374 0.0404 0.0624 0.0363 0.0388 0.0334
AWR10728 0.0298 0.0297 0.0297 0.0302 0.0391 0.0279 0.0305 0.0300
OAKLO807 0.0263 0.0274 0.0274 0.0282 0.0346 0.0202 0.0182 0.0156
ORCH0810 0.0575 0.0623 0.0623 0.0712 0.0946 0.0314 0.0279 0.0234
PAWP0824 0.0574 0.0625 0.0627 0.0680 0.1080 0.0528 0.0576 0.0549
P1CK0526 0.0347 0.0400 0.0400 0.0462 0.0889 0.0519 0.0589 0.0639
PONT0807 0.0430 0.0484 0.0485 0.0568 0.0954 0.0599 0.0610 0.0583
RANDO603 0.0239 0.0225 0.0225 0.0227 0.0250 0.0110 0.0120 0.0099
ROGEOS 19 0.0278 0.0276 0.0276 0.0283 0.0344 0.0397 0.0402 0.0387
SILV0725 0.0425 0.0468 0.0468 0.0541 0.0761 0.0254 0.0244 0.0220

 

 

 

 

 

 

 

 

 

227

 

Volume reﬂectance need in narrow-band index dbl ﬁltered validation data aet

 

 

 

 

 

 

 

 

 

 

 

 

Band A B C D E F G H
Lake 435 450 455 480 550 670 690 700
ARBU0726 0.065 1 0.0670 0.0671 0.0704 0.0805 0.0330 0.0298 0.0280
BR000727 0.0869 0.0916 0.0916 0.1028 0.1435 0. 1 162 0.1190 0.1166
CLME0724 0.0313 0.0315 0.0315 0.0342 0.0463 0.0201 0.0188 0.0162
HESSO727 0.0720 0.0798 0.0798 0.0994 ' 0.1579 0. 1276 0.1290 0.1260
HIGG0727 0.0492 0.0529 0.0530 0.0581 0.0482 0.0110 0.0087 0.0084
HOUG0726 0.0315 0.03 15 0.0315 0.0338 0.0475 0.0302 0.0283 0.0259
JEHN0519 0.0222 0.0226 0.0226 0.0240 0.0360 0.0196 0.0198 0.0175
MITC0725 0.0287 0.0295 0.0297 0.0309 0.0427 0.0353 0.0344 0.0317
MUSK0728 0.0372 0.0357 0.0357 0.0349 0.0400 0.0284 0.0330 0.0335
ROGE0724 0.0325 0.0346 0.0349 0.037 1 0.0509 0.0433 0.0440 0.0420
SAPP0725 0.0255 0.0257 0.0257 0.0266 0.0360 0.0146 0.0145 0.0133
TAMA0605 0.0272 0.0250 0.0250 0.0245 0.0251 0.0153 0.0163 0.0138

 

228

 

Volume reﬂectance uaed in narrow-band index a8 ﬁltered model data set

 

 

 

 

Band A B C D E F G H
Lake 435 450 490 560 660 690 698 700

BEAR0819 0.0410 0.0428 0.0480 0.0436 0.0093 0.0099 0.0075 0.007 1
BELL0826 0.0462 0.0470 0.051 1 0.0844 0.0461 0.0567 0.0590 0.0588
BIG0820 0.0311 0.032 1 0.0348 0.0454 0.0224 0.0198 0.0177 0.0 171
BR000526 0.0108 0.0103 0.0098 0.0257 0.0115 0.0096 0.0101 0.0098
CASS0810 0.0336 0.0329 0.0338 0.0417 0.0251 0.0227 0.0210 0.0205
CHEM0826 0.0360 0.0348 0.0353 0.0424 0.0252 0.0251 0.0222 0.0216
CHIP0820 0.0325 0.0319 0.0317 0.0441 0.0181 0.0162 0.0141 0.0133
CLSJ0603 0.0333 0.0345 0.0392 0.0532 0.0193 0.0175 0.0 155 0.0152
CRO’I‘0605 0.0292 0.0274 0.0252 0.0246 0.0217 0.0218 0.0200 0.0 198
CUBOS 19 0.0345 0.0373 0.0424 0.0509 0.0158 0.0149 0.0116 0.0 1 10
DIAM0824 0.0791 0.0833 0.09 17 0.1085 0.0389 0.0340 0.0313 0.0305
EAGL0824 0.0507 0.0488 0.0531 0.0620 0.0347 0.0329 0.0288 0.0290
FORD0826 0.0427 0.0478 0.06 13 0.1212 0.0754 0.0799 0.0986 0.1010
GLEN0819 0.1213 0.1300 0.1450 0.1235 0.0381 0.0331 0.0301 0.0295
HESS0526 0.0856 0.0947 0.1135 0.1633 0.1079 0.0966 0.1068 0.1063
HICK0820 0.0355 0.0384 0.047 1 0.0897 0.0661 0.0680 0.0744 0.0743
HIGGO608 0.0490 0.0509 0.0574 0.0489 0.0174 0.0159 0.0141 0.0138
KENT0826 0.0363 0.0365 0.0386 0.0559 0.0338 0.0352 0.0328 0.0319
KIMBOS26 0.0445 0.0490 0.0632 0.0930 0.0875 0.0797 0.0755 0.0741
KLIN0603 0.0422 0.0439 0.0507 0.0497 0.0132 0.0106 0.0091 0.0087
LOON0810 0.0 134 0.0130 0.0 160 0.0254 0.0087 0.0073 0.0054 0.0049
LOTU0807 0.0640 0.0693 0.0858 0.1 112 0.0445 0.0383 0.0336 0.0325
MACE0807 0.0362 0.04 10 0.0547 0.0743 0.0185 0.0156 0.0 127 0.0122
MECOO608 0.0583 0.0580 0.0583 0.0620 0.0463 0.0440 0.0421 0.0417
MEC007 24 0.0326 0.0354 0.04 19 0.0665 0.0405 0.0401 0.0346 0.0336
AWR107 28 0.0301 0.0301 0.0308 0.0399 0.0306 0.0303 0.0306 0.0303
OAKL0807 0.0263 0.0270 0.0293 0.0356 0.0198 0.0181 0.0 164 0.0160
ORCH0810 0.0571 0.061 1 0.0754 0.0942 0.0332 0.0294 0.0245 0.0239
PAWP0824 0.0574 0.0616 0.0701 0.1129 0.0593 0.0569 0.0566 0.0553
PICK0526 0.0345 0.0391 0.0497 0.0930 0.0585 0.0575 0.0639 0.0643
PONT0807 0.0430 0.047 5 0.0609 0.0998 0.0651 0.0611 0.0594 0.0581
RAND0603 0.0244 0.0228 0.0233 0.0259 0.0115 0.0127 0.0 106 0.0 101
ROGEOS 19 0.027 9 0.0276 0.0291 0.0356 0.0398 0.0403 0.0391 0.0386
SILV07 25 0.0420 0.0454 0.0564 0.0749 0.0273 0.0253 0.0226 0.0218

 

 

 

 

 

 

 

 

 

229

 

Volume reﬂectance uaed in narrow-band index 38 ﬁltered validation data set

 

 

 

 

 

 

 

 

 

 

 

Band A B C D E F G H

Lake 435 450 490 560 660 690 698 700
ARBU0726 0.0652 0.0668 0.0713 0.0803 0.0344 0.0304 0.0284 0.0279
BR000727 0.0868 0.0910 0.1074 0.1480 0.1241 0.1191 0.1176 0.1157
CLMEO724 0.0317 0.0310 0.0351 0.0479 0.0213 0.0197 0.0167 0.0163
HESSO727 0.0714 0.0776 0.1054 0.1644 0.1353 0.1284 0.1284 0.1265
HIGGO727 0.0490 0.0520 0.0586 0.0455 0.0112 0.0093 0.0083 0.0083
HOUG0726 0.0319 0.0310 0.0347 0.0499 0.0324 0.0291 0.0263 0.0257
JEHNOS 19 0.0222 0.0224 0.0251 0.0380 0.0210 0.0200 0.0179 0.0173
MITC0725 0.0287 0.0294 0.0324 0.0447 0.0372 0.0347 0.0327 0.0320
MUSK0728 0.0371 0.0359 0.0348 0.0401 0.0306 0.0331 0.0335 0.0333
ROGE0724 0.0325 0.0343 0.0383 0.0540 0.0453 0.0449 0.0422 0.0415
SAPP0725 0.0254 0.0257 0.0271 0.0368 0.0166 0.0147 0.0137 0.0133
TAMA0605 0.0270 0.0251 0.0242 0.0253 0.0165 0.0173 0.0147 0.0142

 

230

 

 

APPENDIX G
PRACTICAL PROBLEM WITH AISA IMAGE

2......

...,“.1. W .. .1- a}...

pm.

 

231

REFERENCES

Ahn, Y. H., Bricaud, A., and A. Morel, 1992, Light Backscattering Efﬁciency and
Related Properties of Some Phytoplankters, Deep-Sea Research, 39: 1835-1855.

Albert, D. A., 1995, Regional Landscape Ecosystems of Michigan, Minnesota,
and Wisconsin: A Working Map and Classiﬁcation, General Technical Report NC-
1 78, St. Paul, MN: US. Department of Agriculture, Forest Service, North
Central Forest Experiment Station, Jamestown, ND: Northern Prairie Wildlife
Research Center Online,

http: / / www.npwrc.usgs. gov / resource / habitat / rlandscp / rlandscp.htm

(Version 03JU N 1998).

Anderson, D. M. and D. J. Garrison, editors, 1997, The Ecology and
Oceanography of Harmful Algal Blooms, Limnology and Oceanography,
42: 1009- 1305.

ASD Inc, 2006, Analytical Spectral Devices Inc., http: / /www.asdi.com.

Asrar, G., 1989, Theory and Applications of Optical Remote Sensing, New York,
Wiley Series in Remote Sensing, John Wiley 85 Sons, Inc.

Becker, B. L., Lusch, D. P., and J. Qi, 2005, Identifying Optimal Spectral Bands
from In Situ Measurements of Great Lakes Coastal Wetlands using Second-
derivative Analysis, Remote Sensing of Environment, 97:238-248.

Bukata, R. P., Jerome, J. H., Kondratyev, K. Y., and D. V. Pozdnyakov, 1995,
Optical Properties and Remote Sensing of Inland and Coastal Waters, New York:
CRC Press, 362 pp.

Butler, W. L. and D. W. Hopkins, 1970, Higher Derivative Analysis of Complex
Absorption Spectra, Photochemistry and Photobiology, 12:439-450.

Buzzi, F., 2002, Phytoplankton Assemblages in Two Sub-basins of Lake Como,
Journal of Limnology, 61(1): 1 17-128.

Capone, D. G., Zehr, J. P., Paerl, H. W., Bergman, B., and E. J. Carpenter,

1997, Trichodesmium, a Globally Signiﬁcant Marine Cyanobacterium, Science,
276:1221-1229.

232

Carpenter, S. R., Caraco, N. F., Correll, D. L., Howarth, R. W., Sharrley, A. N.,
and V. H. Smith, 1998, Nonpoint Pollution of Surface Waters with Phosphorus
and Nitrogen, Ecological Applications, 8(3)559-568.

Chappelle, E. W., Kim, M. S., and J. E. I. McMurtrey, 1992, Ratio Analysis of
Reﬂectance Spectra (RARS): An Algorithm for the Remote Estimation of the
Concentrations of Chlorophyll a, Chlorophyll b, and Carotenoids in Soybean
Leaves, Remote Sensing of Environment, 39:239-247.

Clark, C. D., Ripley, H. T., Green, E. R, Edwards, A. J., and P. J. Mumby, 1997,
Mapping and Measurement of Trophical Coastal Environments with

Hyperspectral and High Spatial Resolution Data, International Journal of Remote
Sensing, 18(2):237-242.

Clarke, G. K., Ewing, G. C., and C. J. Lorenzen, 1970, Spectra of Backseattered
Light from the Sea Obtained from Aircraft as a Measure of Chlorophyll
Concentration, Science, 167:1 1 19- 1 121.

Cox, R. M., Forsythe, R. D., Vaughan, G. E., and L. L. Olmsted, 1998, Assessing
Water Quality in the Catawba River Reservoirs using Landsat Thematic Mapper
Satellite Data, Lake and Reservoir Management, 14:405-416.

Curran, P. J ., Dungan, J. L., Macler, B. A., Plummer, S. E., and D. L. Peterson,
1992, Reﬂectance Spectroscopy of Fresh Shole Leaves for the Estimation of
Chemical Concentration, Remote Sensing of Environment, 39:153-166.

Dall’Olmo, G., 2005, Advanced Algorithms for Remote Estimation of Water
Quality, http: / / calmaps.un1.edu/ snrscoq/ 1, PPT-ﬁle.

Dall’Olmo, G. and A. A. Gitelson, 2005, Effect of Bio-optical Parameter
Variability on the Remote Estimation of Chlorophyll-a Concentration in Turbid
Productive Waters: Experimental Results, Applied Optics, 44(3):412-422.

Davis, J. C., 1986, Statistics and Data Analysis in Geology, John Wiley 85 Sons,
New York.

Dekker, A. G., 1993, Detection of Optical Water Quality Parameters for
Eutrophic Waters by High Resolution Remote Sensing, PhD Thesis, Vrije
Universiteit, Amsterdam.

Dekker, A. G. and S. W. M. Peters, 1993, A TM Study of Eutrophic Lakes in the
Netherlands, International Journal of Remote Sensing, 14:799-82 1.

233

Dekker, A. G., Malthus, T. J., Wijnen, M. W., and E. Seyhan, 1992, The Eﬂ'ect
of Spectral Bandwidth and Positioning on the Spectral Signature Analysis of
Inland Waters, Remote Sensing of Environment, 41:2 1 1-225.

Dekker, A. G., Vos, R. J ., and S. W. M. Peters, 2001, Comparison of Remote
Sensing Data, Model Results and In Situ Data for Total Suspended Matter (TSM)
in the Southern Frisian Lakes, The Science of the Total Environment, 268: 197—

2 14.

Demetriades-Shah, T. H., Steven, M. D., and J. A. Clark, 1990, High Resolution
Derivatives Spectra in Remote Sensing, Remote Sensing of Environment, 33:55-
64.

Doerffer, R., Fisher, J ., Stossel, M., Brockmann, C., and H. Grassl, 1989,
Analysis of Thematic Mapper Data for Studying the Suspended Matter
Distribution in the Coastal Area of the German Bight (North Sea), Remote
Sensing of Environment, 28:6 1-73.

Dorr, J .A., Jr. and D. F. Eschman, 1984, Geology of Michigan, University of
Michigan Press, Ann Arbor, MI: 470 pp.

Doxaran, D., Froidefond, J .M., Lavender, S., and P. Castaing, 2002, Spectral
Signature of Highly Turbid Waters: Application with SPOT data to Quantify

Suspended Particulate Matter Concentrations, Remote Sensing of Environment,
81 : 1 49- 16 1.

Duda, R. O. and P. E. Hart, 1973, Pattern Classiﬁcation and Scene Analysis,
Wiley, New York.

Eichenlaub, V.L., 1979, Weather and the Climate of The Great Lakes Region,
The University of Notre Dame Press, Notre Dame, IN, 333 pp.

Fell, A. F., and G. Smith, 1982, Higher Derivative Methods in Ultraviolet, Visible
and Infrared Spectrophotometry, The Analytical Proceedings, 19:28-32.

Fisher, S. G., 1994, Pattern, Process and Scale in Freshwater Ecosystems:
Some Unifying Thoughts, In Aquatic Ecology — Scale, Pattern and Process, edited
by P. S. Giller, A. G. Hildrew and D. G. Raffaelli (London: Blackwell Sciences
Ltd), 575-592.

Flink, P., Lindell, T., and C. Ostlund, 2001, Statistical Analysis of Hyperspectral

Data from Two Swedish Lakes, The Science of the Total Environment, 268:155-
169.

234

Fraser, R. S., 1998, Multispectral Remote Sensing of Turbidity Among Nebraska
Sand Hills Lakes, International Journal of Remote Sensing, 19:301 1-3016.

Gege, P., 2001, A Software Tool for Simulation and Analysis of Optical In-situ
Spectra, 4th Berlin Workshop on Ocean Remote Sensing, Berlin, Germany.

George, D. G. and T. J. Malthus, 2001, Using a Compact Airborne
Spectrographic Imager to Monitor Phytoplankton Biomass in a Series of Lakes
in North Wales, The Science of the Total Environment, 268:215-226.

Giardino, C., Pepe, M., Brivio, P. A., Ghezzi, P., and E. Zilioli, 2001, Detecting
Chlorophyll, Secchi Disk Depth and Surface Temperature in a Sub-Alpine Lake
using Landsat Imagery, The Science of the Total Environment, 268:19-29.

Gitelson, A. A., 1992, The Peak Near 700 nm on Radiance Spectra of Algae and
Water: Relationships of its Magnitude and Position with Chlorophyll
Concentration, International Journal of Remote Sensing, 13:3367-3373.

Gitelson, A., Garbuzov, G., Szilagyi, F., Mittenzwey, K. H., Kamieli, K., and A.
Kaiser, 1993, Quantitative Remote Sensing Methods for Real-time Monitoring of
Inland Waters Quality, International Journal of Remote Sensing, 14: 1269- 1295.

Goodin, D. G., L. Han, R. N. Fraser, D. C. Rundquist and W. A. Stebbins, 1993,
Analysis of Suspended Solids in Water Using Remotely Sensed High Resolution
Derivative Spectra, Photogrammetric Engineering and Remote Sensing,
59(4):505-510.

Gordon, H. R., 1987, Calibration Requirement and Methodologr for Remote

Sensors Viewing the Oceans in the Visible, Remote Sensing of Environment,
22: 103- 126.

Gordon, H. R., Brown, 0. B., and M. M. Jacobs, 1975, Computed Relationships
between the Inherent and Apparent Optical Properties of a Flat Homogenous
Ocean, Applied Optics, 14:417-427.

Green, S., 1998, The Effect of Chlorophyll Concentration on Airborne
Hyperspectral Reﬂectance, www.ucd.ie / app-phys/ stuart/ MODEL.HTM.

Hamilton, M. K., Davis, C. O., Rhea, W. J., Pilorz, S. H., and K. J. Cardel, 1993,

Estimation Chlorophyll Content and Bathymetry of Lake Tahoe using AVIRIS
Data, Remote Sensing of Environment, 44(2/3):217-230.

235

Han, L., 1997, Spectral Reﬂectance with Varying Suspended Sediment
Concentrations in Clear and Algae-Laden Waters, Photogrametric Engineering
and Remote Sensing, 63(6):701-705.

Han, L. and D. C. Rundquist, 1997, Comparison of NIR/ RED Ratio and First
Derivative of Reﬂectance in Estimating Algal-Chlorophyll Concentration: A Case
Study in a Turbid Reservoir, Remote Sensing of Environment, 62:253-261.

Han, L. and D. C. Rundquist, 1998, The Impact of a Wind-roughened Water
Surface on Remote Measurements of Turbidity, International Journal of Remote
Sensing, 19(1):195-201.

Han, L., Rundquist, D. C., Liu, L. L., Fraser, R. N., and J. F. Schalles, 1994,
The Spectral Response of Algal Chlorophyll in Water with Varying Levels of
Suspended Sediment, International Journal of Remote Sensing, 15:3707-3718.

Harding, L. W., Itsweire, E. C., and W. E. Esalas, 1995, Algorithm Development
for Recovering Chlorophyll Concentrations in the Chesapeake Bay using

Aircraft Remote Sensing, 1989-91, Photogrammetric Engineering and Remote
Sensing, 61(2): 177-185.

Harma, P., Vepsalainen, J ., Hannonen, T., Pyhalahti, T., Kamari, J ., Kallio, K.,
Eloheimo, K., and S. Koponen, 2001, Detection of Water Quality using
Simulated Satellite Data and Semi-empirical Algorithms in Finland, The Science
of the Total Environment, 268:107-121.

Harrington, J. and R. L. Repic, 1995, Hyperspectral and Video Remote Sensing
of Oklahoma Lakes, Papers and Proceedings of Applied Geography Conferences,
18:79-86.

Heege, T. and J. Fischer, 2000, Sun Glitter Correction in Remote Sensing
Imaging Spectrometry, The XV Ocean Optics Conference, 1 6 — 20 October 2000
(Falls Church, VA: Ofﬁce of Naval Research).

Howarth, R., Anderson, D., Cloern, J ., Elfring, C., Hopkinson, C. Lapointe, B.,
Malone, R., Marcus, N., McGlathery, K., Sharpley, A., and D. Walker, 2000,
Nutrient Pollution of Coastal Rivers, Bays, and Seas, Issues in Ecology, Issue 7,
www.esa.org/ science / Issues.

Huguenin, R. L. and J. L. Jones, 1986, Intelligent Information Extraction from

Reﬂectance Spectra: Absorption Band Positions, Journal of Geophysical
Research, 91:9585-9598.

236

IOCCG, 1998, Minimum Requirements for an Operational Ocean-colour Sensor
for the Open Ocean, Reports of the International Ocean-Colour Coordinating
Group, Number 1, Dartmouth, Canada.

Jensen, J. R., 1996, Remote Sensing of the Environment: An Earth Resource
Perspective, Prentice Hall, Upper Saddle River, NJ.

Jensen, J. R., 2000. Remote sensing of the environment: An Earth resource
perspective, Prentice Hall series in geographic information science, 544 pp.

Jupp, D. L. B., Kirk, J. T. O., and G. P. Harris, 1994, Detection, Identiﬁcation
and Mapping of Cyanobacteria — Using Remote Sensing to Measure the Optical
Quality of Turbid Inland Waters, Australian Journal of Marine and Freshwater
Research, 45:801-828.

Kallio, K., Koponen, S., and J. Pulliainen, 2003, Feasibility of Airborne Imag'ng
Spectrometry for Lake Monitoring — A Case Study of Spatial Chlorophyll a
Distribution in Two Meso-Eutrophic Lakes, International Journal of Remote
Sensing, 24( 19) :377 1-3790.

Kallio, K., Kutser, T., Hannonen, T., Koponen, S., Pulliainen, J ., Vepsalainen,
J ., and T. Pyhalahti, 2001, Retrieval of Water Quality from Airborne Imaging
Spectrometry of Various Lake Types in Different Seasons, The Science of the
Total Environment, 268:59-77.

Kloiber, S. M., Brezonik, P. L., Olmanson, L. G., and M. E. Bauer, 2002, A
Procedure for Regional Lake Water Clarity Assessment using Landsat
Multispectral Data, Remote Sensing of the Environment, 82:38-47.

Koponen, S. Pulliainen, J ., Servomaa, H., Zhang, Y., Hallikainen, M., Kallio, K.,
Vepsalainen, J ., Pyhalahti, T., and T. Hannonen, 2001, Analysis on the
Feasibility of Multi-source Remote Sensing Observations for Chl a Monitoring in
Finnish Lakes, The Science of the Total Environment, 268:95-106.

Kotak, B. G., Keneﬁck, S. L., Fritz, D. L., Rousseaux, C. G., Prepas, E. E., and
S. E. Hrudey, 1993, Occurrence and Toxicological Evaluation of Cyanobacteria]
Toxins in Alberta Lakes and Farm Dugouts, Water Research, 27:495-506.

Kotak, B. G., Prepas, E. E., and S. E. Hrudey, 1994, Blue Green Algal Toxins in
Drinking Water Supplies: Research in Alberta, Lake Line, 14:37-40.

237

Kutser, T., Herlevi, A., Kallio, K., and K. Arst, 2001, A Hyperspectral Model for
Interpretation of Passive Optical Remote Sensing Data from Turbid Lakes, The
Science of the Total Environment, 268:47-58.

LabSphere, Inc. , 2006 ,
http: / / www. globalspec.com / Supplier / Proﬁle / Labsphe re.

Lathrop, R. G., 1992, Landsat Thematic Mapper Monitoring of Turbid Inland
Water Quality, Photogrammetric Engineering and Remote Sensing, 46:286-280.

Lathrop, R. G. and TM. Lillesand, 1989, Monitoring Water Quality and River
Plume Transport in Green Bay, Lake Michigan with SPOT- 1 imagery,
Photogrammetric Engineering and Remote Sensing, 55(3):349-354.

Lillesand, T. M., 2002, Combining Satellite Remote Sensing and Volunteer
Secchi Disk Measurement for Lake Transparency Monitoring, Proceedings,
National Monitoring Conference.

Lillesand, T. M. and R. Q. Kiefer, 2000, Remote Sensing and Image
Interpretation, 4th Edition, John Wiley and Sons, NY.

Lillesand, T. M., Johnson, W. L., Deuell, R. L., Lindstrom, O. M., and D. E.
Meisner, 1983, Use of Landsat Data to Predict the Trophic State of Minnesota
Lakes, Photogrammetric Engineering and Remote Sensing, 49(2):219-229.

Lodhi, M. S., D. C. Rundquist, L. Han and M. S. Juzila, 1997, The Potential for
Remote Sensing of Loess Soils Suspended in Surface Waters, Journal of the
American Water Resources Association, 33(1): 1 1 1-127.

Malthus, T. J ., Place, C. J ., Bennet, S., and S. North, 1996, An Evaluation of
the Airborne Thematic Mapper Sensor for Monitoring Inland Water Quality,
Proceeding of the 22'"71 Annual Conference of the Remote Sensing Society,
September 1 996, University of Durham, 317-342.

Martin, M. E. and J. D. Aber, 1993, Measurements of Canopy Chemistry with
1992 AVIRIS Data at Blackhawk Island and Harvard Forest, in The Forth
Annual JPL Airborne Geoscience Workshop, Washington D. C.

MCComb, A. J. and J. A. Davis, 1993, Eutrophic Waters of Southwestern
Australia, Fertilizer Research, 36: 104- 114.

Morel, A. and B. Gentili, 1993, Diﬂ‘use Reﬂectance of Oceanic Waters: 11.
Bidirectional Aspects, Applied Optics 32:6864-6879.

238

Nanu, L., 1993, Effect of Suspended Sediment Depth Distribution on Coastal
Water Spectral Reﬂectance: Theoretical Simulation, International Journal of
Remote Sensing, 14(2):225-239.

National Resources Inventory, 1987, Michigan Data, USDA Soil Conservation
Service, East Lansing, Michigan.

NRC (National Research Council), 1993, Managing Wastewater in Coastal Urban
Areas, National Academy Press, Washington, D. C., USA.

O’Reilly, J. E., Maritorena, S., Mitchell, B. G., Siegel, D. A., Carder, K. L.,
Garver, S. A., Kahru, M., and C. R. McClain, 1998, Ocean Color Chlorophyll
Algorithms for SeaWiFS, Journal of Geophysical Research, 103:24937-24953.

Ostlund, C., Flink, P., Strombeck, N ., Piersson, D., and T. Lindell, 2001,
Mapping of the Water Quality of Lake Erken, Sweden, from Imaging
Spectrometry and Landsat Thematic Mapper, The Science of the Total
Environment, 268: 139-154.

Ouillon, S., Forget, P., Froidefond, J .M., and J. J. Naudin, 1997, Estimating
Suspended Matter Concentrations from SPOT Data and from Field

Measurements in the Rhone River Plume, Marine Technology Society Journal,
31(2):]5-20.

Palmstrom, N. S., Carlson, R. E., and G. D. Cooke, 1988, Potential Links
between Eutrophication and Formation of Carcinogens in Drinking Water, Lake
and Reservoir Management, 4: 1-15.

Pattiaratchi, C., Lavery, P., Wyllie, A., and P. Hick, 1994, Estimates of Water
Quality in Coastal Waters using Multi-date Landsat Thematic Mapper Data,
International Journal of Remote Sensing, 15: 157 1— 1584.

Penuelas, J., Gamon, J. A., Fredeen, A. L., Merino, J ., and C. B. Field, 1994,
Reﬂectance Indices Associated with Physiological Changes in Nitrogen- and
Water-limited Sunﬂower Leaves, Remote Sensing of Environment, 48: 135-146.

Pepe, M., Giardino, C., Borsani, G., Cardoso, A. C., Chiaudani, G., Premazzi, G.,
Rodari, E., and E. Zilioli, 2001, Relationship Between Apparent Optical
Properties and Photosynthetic Pigments in the Sub-alpine Lake Iseo, The
Science of the Total Environment, 268:31-45.

239

Philpot, W. D., 199 1, The Derivative Ratio Algorithm: Avoiding Atmospheric
Eﬁ'ects in Remote Sensing, BEE Transactions on Geoscience and Remote
Sensing, 29(3):350 357.

Postel, S. L., 1997, Last Oasis. Second edition, Island Press, Washington, D. C.,
USA.

Postel, S. L. and S. R. Carpenter, 1997, Freshwater Ecosystem Services, G. C.
Daily, editor. Nature’s Services. Island Press, Washington, D. C., USA, pg 195-
2 14.

Postel, S. L., Daily, G. C., and P. R. Ehrlich, 1996, Human Appropriation of
Renewable Fresh Water, Science, 271:785-788.

Pulliainen, J ., Kallio, K., Eloheimo, K., Koponen, S., Servomaa, H., Hannonen,
T., Tauriainen, S., and M. Hallikainen, 2001, A Semi-Operative Approach to
Lake Water Quality Retrieval from Remote Sensing Data, The Science of the
Total Environment, 268:7 9-93.

Rencz, A. N., 1999, Remote Sensing for the Earth Sciences (Manual of Remote
Sensing — Third Edition), John Wiley 85 Sons, Inc, 707 pp.

Richards, J. A., 1993, Remote Sensing Digital Image Analysis (2nd Edition),
Springer Verlag, New York.

Richards, J. A. and X Jia, 1999, Remote Sensing Digital Image Analysis: An
Introduction, Springer, 363 pp.

Ritchie, J. C. and C. M. Cooper, 1991, An Algorithm for Estimation Surface
Suspended Sediment Concentrations with Landsat MSS Digital Data, Water
Resources Bulletin, 27(3):373-379.

Roelfsema, C., Dennison, B., Phinn, S., Dekker, A., and Brando, V., 2001,
Remote Sensing of a Cyanobacteria] Bloom (Lyngbya majuscule) in Moreton
Bay, Australia, BEE.

Ruiz-Azuara, P., 1995, Multitemporal Analysis of “Simultaneous” Landsat
Imagery (MSS and TM) for Monitoring Primary Production in a Small Tropical

Coastal Lagoon, Photogrammetric Engineering and Remote Sensing, 61(2): 187-
198.

Rummel, R. J, 1970, Applied Factor Analysis, Northwestern University Press,
Evanston, IL.

240

Rundquist, D. C., Han, L., Schalles, J. F. and J. S. Peake, 1996, Remote
Measurement of Algal Chlorophyll in Surface Waters: The Case for the First

Derivative of Reﬂectance near 690 nm, Photogrammetric Engineering and Remote
Sensing, 62(2):195-200.

Rundquist, D. C., Schalles, J. F. and J. S. Peake, 1995, The Response of
Volume Reﬂectance to Manipulated Algal Concentrations Above Bright and
Dark Bottoms at Various Depths in an Experimental Pool, Geocarto
International, 10:5-14.

Sathyendranath, S. and T. Platt, 1989, Remote Sensing of Ocean Chlorophyll:
Consequence of Nonuniform Pigment Proﬁle, Applied Optics, 28:490-495.

Schmidt, K. S. and A. K. Skidmore, 2004, Smoothing Vegetation Spectra with
Wavelets, International Journal of Remote Sensing, 25(6): 1 167-1184.

Schuette, B. and D. M. Skjaerlund, 1994, Michigan Farmland and Agriculture
Development Task Force: Policy Recommendations and Options for the Future
Growth of Michigan Agriculture: A Report to Governor John Dagler/ Michigan
Farmland and Agriculture Development Task Force, The Task Force, 1 994,
Lansing, Michigan.

Seehausen, 0., van Alphen, J. J. M., and F. Witte, 1997, Cichlid Fish Diversity
Threatened by Eutrophication that Curbs Sexual Selection, Science, 277: 1808-
181 1.

Shaﬁque, N. A., Bradley, A. C., Fulk, F., and S. M. Cormier, 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, 1(1):1-22.

Slater, P. N., 1980, Remote Sensing: Optics and Optical Systems. Reading Mass:
Addison-Wesley.

Smith, V. H., 1998, Cultural Eutrophication of Inland, Estuarine, and Coastal
Waters. In M. L. Pace and PM. Groffman, editors, Successes, Limitations, and
Frontiers in Ecosystem Science, Springer-Verlag, New York, New York, USA.

Stevenson, R. J ., Bothwell, M. L., and Lowe, R. L., 1996, Algal Ecology:
Freshwater Benthic Ecosystems, Academic Press, San Diego, pg 718.

241

Subramaniam, A. and E. J. Carpenter, 1994, An Empirically Derived Protocol
for the Detection of Blooms of the Marine Cyanobacterium Trichodesmium using
CZCS Imagery, International Journal of Remote Sensing, 15:1559-1569.

Subramaniam, A., Carpenter, E. J ., Karentz, D., and P. G. Falkowski, 1999,
Optical Properties of the Marine Diazotrophic Cyanobacteria Trichodesmium
spp. 1. Absorption and Spectral Photosynthetic Characteristics, Limnology and
Oceanography, 44:608-617.

Tassan, S., 1988, The Effect of Dissolved Yellow Substance on the Quantitative
Retrieval of Chlorophyll and Total Suspended Sediment Concentrations from
Remote Measurements of Water Colour, International Journal of Remote
Sensing, 9(4):787-797.

The Academy of Natural Sciences, 2006, Biovolumes of Algal Taxa in Samples
Collected by the USGS NAWQA program,
http: / /diatom.acnatsci.org/ nawqa/ 200 1biovol.asp.

The Canadian Space Agency, 2003, Hyperspectral Remote Sensing: Applications
for Marine, Coastal and Inland Waters,
www.space.gc.ca/ asc/ eng/ satellites / hyper_marine.asp.

The MathWorks, Inc., 2006, Documentation for MathWorks Products, R2006a,
www.mathworks.com/ access / helpdesk/ help / helpdeskhtrnl.

Tsai, F. and Philpot, W., 1998, Derivative Analysis of Hyperspectral Data,
Remote Sensing of Environment, 66:41-51.

US. EPA (US. Environmental Protection Agency), 1990, National Water Quality
Inventory, 1988 Report to Congress, Oﬁ'rce of Water (4503F), US. Government
Printing Ofﬁce, Washington, D. C., USA.

US. EPA (US. Environmental Protection Agency), 1996, Environmental
Indicators of water quality in the United States. EPA 841-R-96-002. USEPA,
Ofﬁce of Water (4503F), US. Government Printing Ofﬁce, Washington, D. C.,
USA.

Van Stokkom, H. T. C., Stokman, G. N. M., and J. W. Hovenier, 1993,

Quantitative Use of Passive Optical Remote Sensing Over Coastal and Inland
Water Bodies, International Journal of Remote Sensing, 14:541-563.

242

Veatch, J. O., 1941, Agricultural Land Classiﬁcation and Land Types of
Michigan, Michigan State University, Agricultural Experiment Station, Special
Bulletin: no. 231 (First revision).

Veatch, J. O., 1953, Soil and Land of Michigan, Memoir (Michigan State
University, Agricultural Experiment Station, no. 7.

Verdin, J. P., 1985, Monitoring Water Quality Conditions in a Large Western
Reservoir with Landsat Imagery, Photogrammetric Engineering and Remote
Sensing, 51(3):343-353.

Vos, R. and M. Rijkeboer, 2000, Monitoring of Algal Blooms in Case 2 Waters
with SesWIFS, The XV Ocean Optics Conference, 1 6 — 20 October 2000 (Falls
Church, VA: Ofﬁce of Naval Research).

Wetzel, R. G., 1992, Clean Water: A Fading Resource, Hydrobiologia,
243/244:21-30.

Wetzel, R. G., 2001, Limnology: Lake and River Ecosystems (3rd Edition),
Academic Press, San Diego, 1006 pp.

Wrigley, R. C. and A. J. Home, 1974, Remote Sensing and Lake Eutrophication,
Nature,250:2 17-2 15.

243

.‘

 

 

111111111111111111111111111111111111