...i.
(it;
.53.,1H7xlx
Immu‘flqng‘UJI.
I 35.; T
\r! it"
-. Av ll
. Aha. ...
ctmmkzh.b.t.?#fld
.Ilr...t~fl. £5ng
xl ul , 0.
Jun
:1: S
2.. . as:
iihufiz

«33:13-, .9! ,

iloIIivnl. .
11...!!! 1...}!!!
15.71.251.421)!

. . Ital!!!“- .174"

x if}!!!

.2. . {cyclic

i flaunt... 9‘4...”

"Ida/$4 GNU}!!!

if;
:11}

3 ‘ lxlllflnr
3"”: I!!!

. i.‘ g
ilkwl «(llul’lcuflltilng ‘
fi.€1..6§ ,

.Ill...

 

$3.3”? Jami» ‘ . . ..
gt"... 5.. ..o‘. 10". 1?]..hI-IE. “YEW-.4. MLu ‘ , .
v. 5-: , n : A

 

This is to certify that the

dissertation entitled

The Use of Remote Sensing in Freshwater Wetland
and Lake Studies

presented by

Stacy A. C. Nelson

has been accepted towards fulfillment
of the requirements for

Ph.D. degreein Fish. & Wildl.

 

 

92a £3;

Major professor

Date q//7/0 2'

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

 

 

 

. LIBRARY i
Michigan State
University

 

 

 

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

 

DATE DUE DATE DUE DATE DUE

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

610.1 c:/ClRC/DateDue.p65-p. 15

THE USE OF REMOTE SENSING IN FRESHWATER WETLAND AND LAKE
STUDIES

By

Stacy Arnold Charles Nelson

A DISSERTATION

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

DOCTOR OF PHILOSOPHY

Department of Fisheries and Wildlife

2002

ABSTRACT

THE USE OF REMOTE SENSING IN FRESHWATER WETLAND AND LAKE
STUDIES

By

Stacy Arnold Charles Nelson

Wetlands and freshwater lakes are currently under threat from anthropogenic pressures
that are occurring across large spatial scales. We need approaches to study these changes
that can accommodate these large spatial scales. For example, the Barataria Basin,
Louisiana, is an extensive wetland and coastal estuary system of great economic and
intrinsic value that has experienced high rates of coastal land loss. However, little
information exists on whether freshwater wetlands in the upper basin have changed. My
objectives in chapter 1 were to: 1) quantify land cover change in the upper basin over
twenty years from 1972- 1992; and to 2) determine land cover transition rates among land
cover types. Using 80-m resolution Landsat MSS data, I found significant changes in
land cover occurred within the upper Barataria Basin over the study period with a 21%
increase in total wetland area. If present trends in total wetland area continue, bottomland
hardwood forests may be eliminated by the year 2025. In Chapter 2, I examined whether
Landsat-7 ETM+ could be used to measure water clarity across a large range of lakes. My
objectives were to: 1) develop a model to estimate lake water clarity from Landsat data
calibrated on 93 lakes in Michigan, and to 2) examine how the distribution of water
clarity across the 93 lakes influences the model. My regression model resulted in a much
lower r2 value than previously published studies. So, to examine the role of the

distribution of calibration lake water clarity data to the model, I simulated a calibration

dataset with a different water clarity distribution by sub-sampling the original dataset.
The subsampled dataset had a much higher percentage of lakes with shallow water
clarity. The regression model for the subsampled dataset resulted in a much higher r2
value. My results show that the use of Landsat to measure water clarity is sensitive to the
distribution of water clarity used in the calibrated dataset. In Chapter 3, I examine the use
of Landsat-5 to monitor macrophytes in inland lakes. My objectives were to: 1) determine
if different aquatic plant cover types could be detected using Landsat-5; and to 2)
determine if I could improve predictions of macrophytes in lakes by including lake
characteristics in the models. Using logistic regression models, I found statistically
significant relationships between most macrophyte measures in 12 calibration lakes and
Landsat—5 spectral values. However, using another lake to validate the model resulted in
poor model validation. Thus, it is still unclear exactly how useful Landsat data is for
monitoring aquatic plants. Remote sensing provides an effecting monitoring tool for
large area studies in historical wetland change. I demonstrate one application of this in
chapter 1. In aquatic studies, remote sensing can play a vital role in reducing the cost,
labor, and time required to monitor regional inland lake water clarity and assess the
distribution of macrophytes. However, caution should be used in assessing water quality
based solely on remotely sensed data until our methods have been refined or better

sensors are developed for aquatic applications.

This body of work is dedicated to my parents, and God from whom all blessing flow.

ACKNOWLEDGEMENTS

I would like to acknowledge the many people whose assistance has allowed me to reach
this level: Patricia Soranno, David Lusch, David Skole, Jack Liu, Jiaguo Qi, Walter
Chowmentowski, Chris Barber, Danielle Gwinn, Jay Samek, Sam Batzli, Jonathan
Chipman, Ralph Bednarz, Kevin Hoffman, John Whitlock, Tamara Brunger-Lipsey,
Kristy Rogers, Sarah Wills, N arumon Wiangwang, Marc Linderman, Russel Minnerick,
Stephen Aichele, Orlando “Ace” Sarnelle, Ken Frank, Jim Bence, Emily Szali, Daniel
Hayes, William Taylor, Ken Poff, James Jay, Jane Thompson, Jim Brown, Sarah Cline,
Carla Dombroski, Julie Traver, Terie Snyder, Jill Cruth, Deana Haner, Lesley Knoll,
Andrew “Drew” Kramer, Sherry L. Martin, Konrad Kulacki, Michael Belligan, Sarah
Walsh, Alan Wilson, and Kendra Spence Cheruvelil. If I’ve forgotten anyone, please

know that you help was truly a blessing. Thank you all so very much.

PREFACE

This dissertation examines applications of remote sensing to freshwater studies,
especially focusing on regional applications across large geographic areas. Iexamine
aspects of both the land surrounding freshwater ecosystems as well as features of the
water itself. In particular, I have used remote sensing to assess regional changes in
wetland area, regional lake water clarity, and to measure freshwater macrophyte
distributions across multiple lakes. The dissertation is divided into three separate, but
related chapters, preceded by an introduction that provides the overall background for
this research. Each of these chapters stands alone as a separate manuscript that has been
or will be submitted for peer-reviewed publication. Chapter 1 has been published already
in the journal, Environmental Management (see citation below). Chapter 2 has been
submitted for publication to the journal, Remote Sensing of the Environment. Chapter 3
is in preparation for submission to the journal, Aquatic Botany. Below are the citations

for each of the chapters including the co-authors who have contributed to each chapter.

Chapter 1:
Nelson, S.A.C., P.A. Soranno, and J. Qi. May 2002. Land cover change in the
upper Barataria Basin Estuary, Louisiana from 1972-1992: Increases in wetland

area. Environmental Management 29(5): 716-727.

vi

Chapter 2:
Nelson, S.A.C., P.A. Soranno, KS. Cheruvelil, S.A. Batzli, and D.L. Skole.
Assessing regional lake water clarity using Landsat and the role of inter-lake

variability. Submitted to Remote Sensing of Environment September 2002.

Chapter 3:
Nelson. S.A.C., KS. Cheruvelil, and RA. Soranno. Remote sensing of freshwater
macrophytes and the influence of lake characteristics. To be submitted to Aquatic

Botany, October 2002.

vii

TABLE OF CONTENTS

List of Tables .......................................................................................... x
List of Figures ....................................................................................... xiii
Introduction ............................................................................................ 1

Chapter 1: Land Cover Change in the Upper Barataria Basin Estuary,

Louisiana, from 1972-1992: Increases in Wetland Area ........................................ 11
Introduction ................................................................................. l 1
Study Area .................................................................................. 14
Materials and Methods .................................................................... 15

Satellite Image Data ............................................................... 15
Supervised image classification ................................................. 16
Accuracy assessment ............................................................. 17
Land change analysis and transition matrices ................................. 18
Results ....................................................................................... 18
Discussion and Conclusions ............................................................... 20
Acknowledgements .......................................................................... 25
Literature Cited ............................................................................. 25

Chapter 2: Assessing regional lake water clarity using Landsat and the

role of inter-lake variability ........................................................................ 29
Introduction ................................................................................. 29
Materials and Methods .................................................................... 31

Study Area ......................................................................... 31
Field Observations ................................................................ 32
Satellite Image Data ............................................................... 32
Statistical Analysis ................................................................ 33
SDT Distributions ................................................................. 34
Results ....................................................................................... 34
Discussion and Conclusions ............................................................... 35
Acknowledgements .......................................................................... 39
Literature Cited .............................................................................. 40

Chapter 3: Remote sensing of freshwater macrophytes and the influence

of lake characteristics ............................................................................... 44
Introduction .................................................................................. 44
Materials and Methods ..................................................................... 48

Study Area .......................................................................... 48
Macrophyte Sampling ............................................................ 48
Lake Characteristics ............................................................... 49
Satellite Imagery ................................................................... 51

viii

Statistical Analysis ................................................................. 51

Logit Model Validation ........................................................... 54

Results ........................................................................................ 55
Discussion and Conclusions ............................................................... 58
Acknowledgements .......................................................................... 62
Literature Cited ............................................................................. 63
Conclusions .......................................................................................... 69
Major conclusions and recommendations from this research ........................ 70
Logistical issues involved in this research .............................................. 73

Future use of remote sensing in freshwater studies .................................... 79
Literature Cited ............................................................................. 81
Appendices ............................................................................................ 84
Appendix 1: Chapter 1 Tables and Figures ...................................................... 85
Appendix 2: Chapter 2 Tables and Figures ..................................................... 103
Appendix 3: Chapter 3 Tables and Figures ..................................................... l 15

LIST OF TABLES

Chapter 1:

Table 1.1A. Error matrix for comparison between the 1992 classification to the higher
resolution Landsat TM (30m) reference scene from 20 October 1995. The
number of correctly classified pixels for each land cover are on the diagonal in
bold.

Table 1.1B. Classification accuracy for comparisons between the 1992 classification to
the higher resolution Landsat TM (30m resolution) reference scene. The
reference scene totals column refers to the number of pixels in each land cover
identified within the Landsat TM reference scene. The classified totals are the
number of pixels in each land cover from our 1992 classification scene. The
number correct is the number of pixels in our classification that match the Landsat
TM scene. Accuracy (%) is the percent correctly classified for each land cover.
The overall classification accuracy produced from this assessment was 91% with
an overall Kappa statistic of 0.88.

Table 1.2. The total area of land cover change between each time step (ha) and the rate
of land cover change (ha/yr) between each time step for each land cover category.
Positive numbers are gains in land cover area and negative numbers are losses in
land cover.

Table 1.3A. The total area (ha) of land converted from one land cover to the next
between 1972 and 1985. The areas of no change are in parentheses. The total

column represents the land use areas in the earlier time step.

Table 1.3B. The total area (ha) of land converted from one land cover to the next
between 1972 and 1992, and between 1985 and 1992. The areas of no change are
in parentheses. The total column represents the land use areas in the earlier time
step.

Table 1.4A. The area (percent) of land converted from one land cover to the next
between 1972 and 1985. For example, the percent of agricultural land present in
1972 that was converted to urban land in 1985 is 35%. Areas of no change are in
parentheses. The total column represents the land use areas in the earlier time
step.

Table 1.4B. The area (percent) of land converted from one land cover to the next
between 1972 and 1992, and between 1985 and 1992. 1972 and 1985 to 1992.
For example, the percent of agricultural land present in 1972 that was converted
to urban land in 1992 is 40%. The areas of no change are within parentheses. The
total column represents the land use areas in the earlier time step.

Table 1.5A. The area (percent) of land converted from one land cover to the next between
1972 and 1985. Changes here show the total percentage of each land cover in the
basin as a whole that changed into other categories. For example, 9% of the total
basin area was converted from agriculture to urban land from 1972 to 1985. The
total column represents the land cover areas in the earlier time step. The areas of
no change are in parentheses.

Table 1.5B. The area (percent) of land converted from one land cover to the next between
1972 and 1985, and between 1985 and 1992. Changes here show the total

percentage of each land cover in the basin as a whole that changed into other

xi

categories. For example, 10% of the total basin area was converted from
agriculture to urban land from 1972 to 1992. The total column represents the land

cover areas in the earlier time step. The areas of no change are in parentheses.

Chapter 2:
Table 2.1. Landsat remote sensing studies of Secchi depth using multiple inland lakes.

Table 2.2. Information for the 93 study lakes used to collect field observations.

Chapter 3:

Table 3.1a. Physical characteristics and image information for study lakes.

Table 3. lb. Limnological characteristics of study lakes.

Table 3.2. Results from the logit regressions for the three models.

Table 3.3. T-test p values from testing the difference between the means of the beta
coefficients (log transformed) for the 12 individual lakes used in the August and
September logit models.

Table 3.4. Regression results of the model coefficients estimated on individual lake
datasets versus selected lake variables. Values in bold are significant at the 0.10
level. See table la. for lake characteristic units.

Table 3.5. Regression results of Landsat TM2 and TM4 versus selected pelagic lake
characteristics using the combined dataset of all lakes (N = 12). See table la for
lake characteristic units.

Table 3.6. Results of the model validation using Winfield Lake. The total number of

littoral zone sites in Winfield Lake is 137.

xii

LIST OF FIGURES

Chapter 1:

Figure 1.1. Zones within the entire Barataria Basin. LA that include the upper swamp
forest/freshwater marsh zone (this study), the brackish marsh zone and the
saltwater marsh zone, which represent the upper, middle and coastal regions
respectively. Three levels of elevations are shown in white, gray, black. Hashed
areas are inland lakes.

Figure 1.2. Classified land cover from Landsat MSS images for each time step.

Figure 1.3. Percent cover of each land cover class in the upper Barataria Basin in each
time step.

Figure 1.4. The area (ha) of land cover occurring within human developed land cover
classes only (A), and within natural land cover classes only (B) in each time step.

Figure 1.5. Changes in total wetland area (ha) and wetland types from in each time step.

Chapter 2:

Figure 2.1. The study area within Path 21 of the Landsat Worldwide Reference System-2
(WRS) in Michigan, USA. The three individual scene rows within Path 21 used
in this study are designated as R29, R30, and R31. All images are from the same
acquisition date, 28 August 2001.

Figure 2.2. Regression of Landsat-7 ETM+ data (ratio of two bands, ETMl and ETM3)
and field measured Secchi depth transparency (SDT; natural log transformed) for

the 93 study lakes. The regression model is ln(SDT) = 1.830 (ETMl/ETM3) -

xiii

2.976. The rmse residual value (standard error of estimate) has the same units as
the predicted ETM values and represents the standard deviation. The dashed lines
represent the 95% confidence interval.

Figure 2.3. Histogram of the percent of lakes in each 0.5 m Secchi depth category for the
47 lakes in Kloiber et a1. (2000), and for the 93 lakes in our study (complete
dataset). Vertical dashed lines represent the limits of the three trophic state
categories (Forsberg & Ryding, 1980).

Figure 2.4. Histogram of the percent of lakes in each 0.5 m Secchi depth category for the
47 lakes in Kloiber et a1. (2000), and for the 17 lakes in our study (subsampled
dataset). Vertical dashed lines represent the limits of the three trophic state
categories (Forsberg & Ryding, 1980).

Figure 2.5. Regression of Landsat-7 ETM+ data (ratio of two bands, ETMl and ETM3)
and field measured SDT (natural log transformed) for our subsampled dataset.
The regression model is ln(SDT) = 2.630 (ETMl/ETM3) - 5.009. The rmse
residual value (standard error of estimate) has the same units as the predicted
ETM values and represents the standard deviation. The dashed lines represent the
95% confidence interval.

Figure 2.6. Regression lines of Landsat-7 ETM+ data (ratio of two bands, ETM] and
ETM3) and field measured SDT (natural log transformed) for our complete and
subsampled datasets for slope comparison. The difference between the two slopes
was not statistically significant (ANCOVA, p=().247). The dashed lines represent

the 95% confidence intervals around each regression.

xiv

Figure 2.7. Scatter plots of the residuals for the complete and subsampled datasets
against the predicted value. The dashed lines represent the 95% confidence
intervals.

Figure 2.8. Scatter plots of the residuals for the complete and subsampled datasets
against observed Secchi depth. The dashed lines represent the 95% confidence
intervals.

Figure 2.9. Histogram of the percent of lakes in each 0.5 m Secchi depth category from a

statewide lake database (n=675) and for our 93 study lakes (complete dataset).

Chapter 3:

Figure 3.1. The study area within Path 21 of the Landsat Worldwide Reference S ystem-2
(WRS) in Michigan, USA. The three individual scene rows within Path 21 used
in this study are designated as R29, R30, and R31. Images R29 and R30 are the
acquisition date 05 September 2001. Image R31 is from the acquisition date 04
August 2001.

Figure 3.2. Landsat TM 30 m raster image of a sample lake with 40 m grid of our sample
points overlain.

Figure 3.3. The distribution of different plant categories in each lake presented as the
total number of sites that had plants present in each category divided by the area
(ha) of each lake's littoral zone. * Winfield Lake was not used to develop any
models and was used only to validate the model.

Figure 3.4. Plant cover in the entire lake (lake percent plant cover) and in the littoral

zone (littoral percent plant cover) for each of the study lakes. The lake percent

XV

plant cover is the total number of sites where plants were found divided by the
total number of sample points in the lake. The littoral percent plant cover is the
total number of sites where plants were found divided by the total number of
sample points in the littoral zone. The littoral zone is defined as g 4.5 m lake
depth. * Winfield Lake was not used to develop any models and was used only to
validate the model.

Figure 3.5. The distribution of plant categories and plant cover levels within the littoral
zone of the lakes comprising our three logit model datasets.

Figure 3.6. The distribution of plant categories and the plant cover levels within the
littoral zone of Winfield lake, the validation lake.

Figure 3.7. The average percent concordance for individual lake models averaged by

image date.

xvi

INTRODUCTION

Environmental Applications of Remote Sensing

Historical investigations of our natural ecosystems predominantly have focused
on local or small scale studies because researchers in the past were limited by phenomena
they could either physically examine in the field, or the area which could be sampled and
returned to the laboratory during the study period (Hardisky et a1. 1986). Today,
increased understanding of ecosystem interactions and functions have led to much
broader scale research approaches as the need to investigate the ecosystem as a whole
becomes more apparent (Sullivan 1994; Brouha 1994). In particular, as the local,
regional, and global land surface is continually transformed as a direct result of earth’s
growing human population and competition for resources, there is a continual threat to

the sustainability of sensitive natural areas (Brouha 1994).

In fact, human’s ability to alter the natural environment has been noted one of the
greatest concerns for global environmental change (Brown et al. 1992, Southwick 1996;
Wilke and Finn, 1996). Today, the global population is over 6 billion and the human
population continues to grow at a rate of 90 million/year; this type of population growth
can have severe consequences on the natural environment (Brown et al. 1992). For
example, it is now estimated that we have transformed approximately 1/3 to 1/2 of earth’s
ice-free surface, thus making human-caused land use/land cover change one of the most
important components of global change (Wilke and Finn, 1996). In the U.S., one of the
most prominent areas of land cover change involves the extent of coastal and inland

wetlands loss that has occurred since European settlement. According to the US. Fish

and Wildlife Service and the National Wildlife Federation, out of the > 81 million ha of
wetlands that were present in the continental United States in the early 1700's, only 35 -

40 million ha remained as of the mid 1980’s (Dahl et al. 1991).

Human’s effects on the natural environment have also led to an additional need
for understanding the effects of cumulative small scale land use transformations that
occur across large regions. Western and Pearl (1989) note that the speed and scale with
which we are modifying the Earth limits our ability to use traditional field methods over
large areas. Green et al. (1994) suggested that remote sensing provides a mechanism
versatile enough to keep up with the rapid and extensive changes made to the Earth’s
surface as the world’s population grows. Because remote sensing offers a tremendous
potential as a rapid, large-area, synoptic source for data acquisition, it has become an
important part of many ecological studies (Lillesand and Kiefer 1994).

Since the 1970’s, technological advances and better sensor configurations have
further refined the use of remote sensing to include monitoring and research of
agriculture, forestry and range resources, land use mapping and planning, geology and
environmental applications, and coastal and water investigations. In particular, remote
sensing has been identified as an effective means of detecting and monitoring land use
and land cover change in the terrestrial environment (Green et al. 1994). However,
satellite remote sensing of inland lakes has been less studied as a result of the difficulties
inherent in interpreting reflectance values of water (Verbyla 1995). Clear water provides
little spectral reflectance because longer wavelengths are absorbed and the reflected
shorter wavelengths, which sensors rely on for surface feature detection, are subject to

atmospheric scattering. Although sensors such as the Landsat Multispectral Scanner

(M88) and Thematic Mapper (TM; and now ETM+) were primarily designed for land
studies, and may not be completely applicable to aquatic studies, recent improvements
now provide better spatial and spectral resolutions than previously available, which may
improve their usefulness to aquatic studies, such as for small inland lakes (Kloiber et al.
2000).

Despite these limitations, remote sensing still has the potential to provide a
valuable tool for monitoring freshwater systems. Several studies have successfully used
remote sensing to measure lake characteristics such as chlorophyll and Secchi disk
transparency, suspended sediments and dissolved organic matter (Lathrop and Lillesand
1986, Jensen et al. 1993, Narumalani et. al. 1997, Lillesand et al. 1983, Khorram and
Cheshire 1985, Dekker and Peters 1993, Kliober et al. 2000). Other studies have used
remote sensing to map aquatic plant abundance in large areas of homogenous cover or for
species detection in smaller areas they have used high resolution hand or aircraft sensors
(Ackleson and Klemas 1987, Armstrong 1993, Penuelas et. al. 1993, Lehmann et al.
1997). While remote sensing has been used in large-scale land change studies for many
years, few studies have used Landsat data to measure water quality parameters of
multiple lakes, over large regions, or covering multiple images. Additionally, current
research on detecting aquatic macrophytes on a regional scale is still lacking.

One additional challenge to regional remote sensing is combining images from
multiple dates or regions of varying atmospheric influences (i.e. haze, cloud cover, etc.).
Song et al. (2001) found that atmosphere corrections are necessary when using two or
more satellite images collected on different dates, such as imagery used in land use/cover

change studies. Multi-temporal image to image differences are a result of atmospheric

absorption and light scatter, which can be highly variable from one period of time to the
next (Moore 1980). This difference can have a pronounced effect on the interpretation of
imagery for trend analyses when using multiple images that are collected on different
days, months, or years. For example, for aquatic studies, Brivio et al. (2001) found that
without the appropriate corrections for atmospheric scattering and transmittance effects,
successful remote sensing of lake water quality parameters could not be accurately
achieved from four Landsat TM images. Brivio et al. (2001) were able to develop two
image-based rectification models which were used to correct lake atmospheric effects in
multi-seasoned Landsat images of Lake Lseo and Lake Garda, Italy. However, accurate
and simple procedures to correct for these effects are still an area of future research
(Rahman et al. 2001, Song et al. 2001).

These atmospheric limitations may be reduced with many recent technological
advances. A few of the newest sensors include IKONOS, the EOS (Earth Observing
System) Terra sensors, and the Landsat ETM+. These sensors will ultimately employ
more advanced data calibrations than possible with some of the older sensors. However,
the advantage of the older Landsat platforms is that they provide consistent historical data
not yet available from new sensors and the temporal and spatial resolutions provided by
Landsat M88 and TM are still more effective in monitoring landscape changes. For
example, since the early 1980's, remote sensing techniques have become commonly used
to detect wetland change (Ricketts 1992). However, few studies have taken advantage of
the wealth of historical data available from the satellite platforms that have been

operating across several decades, and that are becoming widely available and inexpensive

(Munyati 2000). Images obtained using the Landsat Multispectral Scanner (MSS)
sensor presently represent our longest, continuous satellite remotely sensed historic
record.

The availability of newer data will, however, allow for the integration of cross-
platform data in developing band and image sharpening algorithms which will serve to
fill in the gaps present in the older system data acquisitions (clouds, haze, etc.) and
enhance the ability to develop multi-image calibrations for cross referencing.
Furthermore, three of the main Terra sensors (MODIS - Moderate Resolution Imaging
Spectrometer, ASTER - Advanced Spaceborne Thermal Emission and Reflectance
Radiometer, and MISR — Multi-angle Imaging Spectroradiometer) have been identified as
being particularly useful in the remote sensing of large area land cover and land use
changes and vegetation dynamics studies (NASA 1998, Jensen 2000). However, aquatic
remote sensors have yet to see the production of sensors specifically designed to monitor
freshwater environments.

The primary focus of my research involves the application of remote sensing to
studies of freshwater ecosystems. This body of research is important because there is an
increasing need to assess large-scale changes in aquatic ecosystems as a result of
anthropogenic impacts, just as has been done for terrestrial ecosystems. I have examined
a wide range of applications in this research, focusing on two important ecosystems
(lakes and wetlands) from two different states (Louisiana and Michigan, USA).

Large area wetlands and inland lakes within a large region represent two areas

where the use of large-scale monitoring techniques would prove very beneficial, as

research and management agencies are limited in their resources to develop rapid,
regional assessments from ground based inventories.

My first chapter investigates changes in a large area wetland system (628,600 ha)
in southeastern Louisiana, the Barataria Basin. This area has experienced well
documented wetland loss along the coastal margin for many years. However, less is
known about wetland changes occurring in freshwater portion of the upper basin. In this
study, I investigate historical changes within this area using Landsat MSS over three time
periods; 1972, 1985, and 1992.

My second chapter investigates the application of data from the Landsat-7 ETM+
sensor to effectively predict Secchi Disk Transparency across a large region (94, 350
kmz) within the lower peninsula of Michigan, USA. In this study, I use linear regression
models to explore the statistical relationship between Landsat data and the wide
distribution of Secchi values (collected in situ) found throughout a large area, such as the
lower peninsula of Michigan.

My third chapter investigates the application of data from the Landsat TM sensor
to detect aquatic macrophyte distributions in inland lakes across a large region (70,000
kmz) within the lower peninsula of Michigan, USA. Like the second chapter, this study
also uses statistical models to determine the relationship between Landsat data and a wide
range of in situ data across a large area. In this study, macrophytes from randomly
chosen lakes are used to develop logistic regression models which indicate the probability
of correctly classifying the various levels of plant cover based on the Landsat data and

the depth at each individual sample point of plant cover.

Theses research efforts represent three areas of research where an assessment
tool such as the one provided by remote sensing is valuable, both from and ecological and
an operational point of view. Furthermore. this collective research presents important
information which may ultimately provide the impetus for further detailed research in my
study sites and in the area of regional remote sensing for wetland and inland lakes. We
clearly need more research on the application of current and future sensors to wetland and
freshwater studies, and I hope my dissertation significantly contributes to this important

area.

Literature Cited
Ackleson S.G., Klemas V., 1987. Remote-sensing of submerged aquatic vegetation in
lower Chesapeake bay: a comparison of Landsat MSS to TM imagery. Remote
Sensing of Environment. 22, 235-248.
Armstrong, R., 1993. Remote sensing of submerged vegetation canopies for biomass
estimation. International Journal of Remote Sensing. 14, 621—627.
Brouha, P. 1994. The NMFS habitat protection program: lost in the bureaucracy.
Fisheries, vol. 18(10), 1-4.
Brown, L.R., C. Flavin, and H. Kane. 1992. Vital Signs. W.W. Norton Publishing. New
York, New York. Pp 437.
Brivio, P.A., C. Giardino, E. Zilioli. 2001. Validation of satellite data for quality
assurance in lake monitoring applications. Science of the Total Environment.

26823-18.

Dahl, T.E., Johnson, C.E., Frayer, W.E. 1991. Wetlands status and trends in the
conterminous United States, mid-1970's to mid- 1980's. US. Department of the
Interior, Fish and Wildlife Service, Washington. DC, 1-15 pp.

Dekker, A.G., and Peters S.W.M., 1993. The use of the Thematic Mapper for the
analysis of eutrophic lakes: a case study in the Netherlands. International Journal
of Remote Sensing. 14, 799-821.

Green, K., D. Kempka, and L. Lackey. 1994. Using remote sensing to detect and monitor
land-cover and land-use change. Photogrammetric Engineering and Remote
Sensing. 60(3): 331-337.

Hardisky, M.A., M.F. Gross, V. Klemass. 1986. Remote sensing of coastal wetlands.
Bioscience. 36(7): 599-606.

Jensen, J.R., Narumanlani, S., Weatherbee, O.. Mackey. H.E., Jr., 1993. Measurement of
seasonal and yearly cattail and waterlily changes using multidate SPOT
panchromatic data. Photogrammetric Engineering and Remote Sensing. 52, 31-
36.

Jensen, JR. 2000. Remote sensing of the environment an earth resource perspective.
Prentis-Hall, Inc. Upper Saddle River, NJ. Pp. 544.

Khorram, S., and Cheshire H. M., 1985. Remote sensing of water quality in the Neuse
River Estuary. North Carolina. Photogrammetric Engineering and Remote
Sensing. 51, 329-341.

Kloiber, S.M., T.H. Anderle, P.L. Brezonik, L. Olmanson, M.E. Bauer, D.A. Brown.

2000. Trophic state assessment of lakes in the Twin Cities (Minnesota, USA)

region by satellite imagery. Archive Hydrobiologie Special Issues Advances in
Limnology. 55:137-151.

Lathrop, R. G., and Lillesand, T. M., 1986. Utility of Thematic Mapper data to assess
water quality. Photogrammetric Engineering and Remote Sensing, 52, 671—680.

Lehmann, A., and Lachavanne, J .B., 1997. Geographic information systems and remote
sensing in aquatic botany. Aquatic Botany. 58. 195-207.

Lillesand, T. M., Johnson, W.L., Deuell, R. L., Linstrom, O. M., and Meisner, D. E.,
1983. Use of Landsat data to predict the trophic state of Minnesota lakes.
Photogrammetric Engineering and Remote Sensing. 49, 219 - 229.

Lillesand, T.M. and R.W. Kiefer. 1994. Remote Sensing and Image Interpretation. John
Wiley and Sons, Inc. New York, New York. Pp 750.

Moore, GK. 1980. Satellite remote sensing of water turbidity. Hydrobiological Sciences.
25:407-421.

Munyati, C. 2000. Wetland change detection on the Kafue Flats, Zambia, by
classification of a multitemporal remote sensing image dataset. International
Journal of Remote Sensing 21(9): 1787-1806.

Narumalani, S., Jensen, J.R., Althausen, J.D., Burkhalter, S.G., Mackey, H..,E Jr., 1997.
Aquatic macrOphyte modeling using GIS and logistic multiple regression.
Photogrammetric Engineering and Remote Sensing. 63, 41 -49.

NASA. 1998. NASA ’s Earth Observing System — EOS AM-l . Greenbelt: NASA Goddard

Space Flight Center. Pp. 43.

Penuelas, J., Gamon, J.A., Griffen, KL, and Field, CB, 1993. Assessing coummunity
type, plant biomass, pigment composition, and photosyntheic efficiency of aquatic
vegetation on spectral reflectance. Remote Sensing of Environment. 46, 1 10-1 18.

Rahman, H. 2001. Influence of atmospheric correction on the estimation of biophysical
parameters of crop canopy using satellite remote sensing. International Journal of
Remote Sensing. 22(7): 1245-1268.

Ricketts, Peter J. 1992. Current approaches in geographic, information systems for
coastal management. Marine Pollution Bulletin 25(1-4):82-87.

Song, C., C.E. Woodcock, K.C. Seto, M.P. Lenney, and SA. Macomber. 2001.
Classification and change attention using Landsat TM data: when and how to
correct atmospheric effects? Remote Sensing of Environment. 75:230-244.

Southwick, CH. 1996. Global ecology in human perspective. Oxford university press.
New York, N.Y.. Pp. 392.

Sullivan, J .K. 1994. Habitat status and trends in the Delaware Estuary. Coastal
Management, vol. 22, 49-79.

Verbyla, D.L., 1995. Satellite Remote Sensing ofNatural Resources. CRC Lewis
Publishers, New York, NY. Pp. 198.

Western, D. and M. Pearl. 1989. Conservation biologvfor the next century. Oxford
university press. New York, N.Y.. Pp. 423.

Wilkie, BS. and J.T. Finn. 1996. Remote sensing imagery for natural resources

monitoring: a guidefor/irst-time users. Columbia University Press. New York, New

York. Pp. 296.

10

Chapter 1: Land Cover Change in the Upper Barataria Basin Estuary, Louisiana,

from 1972-1992: Increases in Wetland Area

Introduction

The extent of coastal and inland wetlands in the United States has declined
markedly since European settlement. According to the US. Fish and Wildlife Service
and the National Wildlife Federation, more than 81 million ha of wetlands were present
in the continental United States in the early 1700's, but only 35 - 40 million ha remained
as of the mid 1980’s (Dahl and others 1991). Wetland losses have been attributed to a
number of factors such as agricultural land conversions, urbanization, and erosive natural
processes. The draining and filling of wetlands for agriculture accounted for as much as
54% of total wetland losses between the mid-1950‘s and the 1970's, while urban land
conversions accounted for only 5% (Dahl and others 1991). Because of recent legislation
protecting wetlands and changes in how wetlands are valued, it is unclear whether these

land use transitions have continued from the 1970’s to today.

Natural changes in coastal wetlands result from wind and wave pressures,
episodic storm events, land subsidence, variations in hydrological cycles leading to
flooding or drying out, and eustatic sea level rise along coastal margins (Mitsch and
Gosselink 1986). Currently, coastal Louisiana, which contains approximately 27% of the
coastal wetlands found within the continental United States, is experiencing estimated
loss rates ranging from 6,000 to 12,450 ha/yr (Visser and others 1999). These estimates
are the highest rates of wetland loss in the United States and have been attributed to high

land subsidence rates and sea level rise compounded by the artificial levying of the

11

Mississippi River Deltaic Plain and other factors (e. g. canal dredging and erosion)
(Scaife and others 1983. Conner and others 1987, Visser and others 1999). Although
coastal estuaries, which are experiencing such high loss rates, have been well studied,
there has been less documentation of the changes occurring in the freshwater portions of
the estuary systems, which are subject to additional pressure from agriculture and urban

development.

In particular, within the Barataria Basin, Louisiana, a large coastal estuary, much
attention has focussed on the high coastal wetland loss and increases in open water in the
lower coastal basin that are a result of anthropogenic modifications and natural processes
(Turner 1997, Sklar and Browder 1998, Martin and others 2000). However, little
information exists on how other components of the estuary have changed, especially the
freshwater wetlands in the upper basin. Some have suggested that increased
urbanization, industrialization and agricultural practices will lead to increasingly
eutrophic conditions within the Basin’s upper freshwater zones (Stow and others 1985).
Because the upper basin is tightly linked hydrologically to the lower basin, large changes
in the upper basin may strongly impact the lower basin.

Understanding changes in the Barataria is especially important because an
estimated 97% of all commercially valuable Louisiana Gulf of Mexico fisheries species
require some or part of their life cycle on the productivity of the Barataria and adjacent
coastal estuarine basins. This estimate makes up approximately 20% of the US.
commercial seafood harvest, approximately one half billion pounds of fish and shellfish
per year. However, this region has only recently been officially designated as a

nationally valuable resource, being included in the U. S. Environmental Protection

Agency’s (EPA) National Estuary Program in 1990 as part of the Barataria-Terrebonne
estuarine system.

Since the early 1980's, remote sensing techniques have become commonly used to
detect wetland change. Remotely sensed images and geographic information systems
(GIS) are being used to address critical coastal resource management problems, providing
researchers with the ability to make rapid decisions at large spatial scales using recent
data (Ricketts 1992). However, fewer studies have taken advantage of the wealth of
historical data available from the few satellite platforms that have been operating across
several decades, and that are becoming widely available and inexpensive (Munyati 2000).
For example, the images obtained using the Landsat Multispectral Scanner (MSS) sensor
presently represents my longest, continuous, satellite remotely sensed historic record.
This sensor spans a time period where known large changes have occurred in land cover
(1972—present). Although, few studies have used the MSS sensor’s 20-plus year, archival
data source for historic wetland change determination. several researchers have used
Landsat MSS data to identify wetland change across smaller time periods with a fairly
high degree of accuracy (Jensen and others 1986, Haaek 1996, Munyati 2000). The
coarse resolution of MSS (80m) provides an additional advantage in that it allows

coverage of a large spatial area.

My objectives were to quantify land cover change in the upper Barataria basin
from 1972—1992, to determine land cover transition rates among land cover types, and to
evaluate the accuracy of MSS to detect wetland and other land cover change. I obtained
Landsat MSS data from the North American Landscape Characterization (NALC) data

archive and classified land cover in each scene using a supervised image classification

approach. I use the term land cover to refer to both human (agricultural and urban) and
natural (wetland and forest) land covers. I found large changes in both human and
natural land covers across the study period that are different from known changes in the

lower, coastal regions of the Barataria basin.

Study Area

The Barataria Basin (Figure 1.1) makes up the eastern-most region of the
Barataria-Terrebonne estuary system. This 628,600 ha area is located in southern
Louisiana south of New Orleans and extends into the Gulf of Mexico at its widest
portion. Natural and artificial levees, completed in the 1940’s, hydrologically isolate the
Barataria from sources of riverine sediment input (Conner and others 1987, Miller and
others 1995). The Barataria can be divided into three zones based on a salinity gradient.
The upper freshwater portion, or upper Barataria Basin. consists of swamp
forest/freshwater marshes (0-2 ppt), the middle low salinity portion (2-10 ppt) is made up
of brackish marsh, and the lower portion (6-22 ppt) is the saltwater coastal marsh zone
(Conner and others 1987, Miller and others 1995). For this study, I focus on the upper
swamp forest/freshwater marsh zone only. which makes up 27% (170.370 ha) of the total
basin area. There are two open water bodies within the upper basin: Lac des Allemands
and Lake Boeuf, which did not change across the study period so were removed from the
land cover change analysis.

Agriculture and urban land covers are found in higher elevation areas along the
levee system of the upper Barataria, whereas the freshwater marsh, and swamp forest are

in the lowest elevation areas (Figure 1.1). Bottomland hardwood forest exists in less

frequently flooded lower elevation areas and along the lower perimeter of the levee
system. The bottomland hardwood forest land cover is comprised largely of American
elm (Ulmus americana), sweetgum (Liquidambar styraciflua), sugarberry (Celtis
laevigata) and swamp red maple (Acer rubrum var. drummondii). Swamp forest is
comprised of baldcypress (Taxodium distichum) and water tupelo (Nyssa aquatica).
Freshwater marsh is comprised largely of a floating marsh known locally as “flotant”.
Flotant consists of vegetative mats of detritus, algae, and plant roots that support
maidencane (Panicum hemitomon), spikerush (Eleocharis sp.) and bulltongue (Sagittaria
falcata).

The majority of freshwater input into the upper Barataria is from rainfall (Conner
and others 1987). Average annual rainfall measured at the New Orleans, Louisiana
Weather Data Center from 1972 to 1992 was 171 cm/year (+/- 36 cm/yr S.D.). Annual
rainfall for each of the three time periods used in this study was 163, 170 and 208 cm/yr
for 1972, 1985 and 1992 respectively. During my study period, annual rainfall showed

no significant trends through time (R2=0.009. P=0.68).

Materials and Methods
Satellite Image Data

I acquired four. 80m resolution, Landsat MSS scenes covering the upper Barataria
Basin from the North American Landscape Characterization (NALC) data archive (Earth
Resources Observation Systems Data Center Distributed Active Archive Center Sioux
Falls, SD) for three time periods: 1 October 1972, 10 October 1972, 31 August 1985, and

5 October 1992. Because the 1972 data were sensed from the older Landsat (1, 2, and 3)

 

s

.si,i

Hi.-

KR,

platforms with a larger swath width, these scenes were slightly out of phase with the
1985 and 1992 scenes and required two scenes to cover the extent of the upper Barataria
Basin. All scenes were geometrically rectified to a digitally scanned, 1:100,000 scale
USGS quadrangle topographic map covering the upper Barataria Basin. Each scene was
coregistered with a corresponding digital elevation model (DEM) and transformed to
Universal Transverse Mercator (UTM). A nearest neighbor algorithm was used to
perform the resampling procedure and the image-to-map registrations which yielded a
root-mean-square error of 0.95 pixels for all data. The two 1972 scenes were mosaicked
to create a single 1972 image.
Supervised Image Classification

To ensure consistency among classifications of the three time steps, I performed
two pre-processing procedures on each scene. First, each scene was normalized to the
1992 scene using contrast stretching and histogram matching procedures to reduce
atmospheric differences across all three scenes. Second. I developed a ratio of bands 1
and 2 to correct distortions inherent in each scene. The band ratioing procedure reduced
the scene distortion by dividing brightness values of pixels in one band by the brightness
values of the corresponding pixels in another band.

I then used a supervised classification procedure to classify the 1992 scene using
spectral training samples, seed pixels, and a maximum likelihood algorithm. I identified
six major land cover types as described by Conner and others (1987): agriculture, urban,
bottomland hardwood forest, swamp forest, freshwater marsh, and open water.
Histograms of all bands were developed for each land cover class to determine average

spectral signatures. Spectral signature separability was computed to determine the

statistical distance between signatures. The final spectral signatures for the 1992
classification, resulting from a band combination of 4. 2, and my band ratio, were then
used to classify the 1972. 1985, and 1992 scenes. I then applied a 3 x 3 focal majority
filter model to all three classifications, which is a post classification smoothing operation
to reduce the number of misclassified pixels. Finally, I manually recoded < 10% of the
pixels in all three classified images that were obviously misclassified when compared to
the raw scenes. For example, some obviously urban and agricultural pixels were
misclassified as freshwater marsh land cover. Freshwater marsh should not occur in the
higher elevation areas. This obvious misclassification represented the majority of my
manual pixel reclassifications.

Accuracy Assessment

To estimate the classification accuracy, I developed two error contingency
matrices from the 1992 classification using two different reference scenes. and calculated
the probability of misclassifications through estimates of omission and commission
errors. For the first assessment, 1 used a 28 September 1995 Landsat Thematic Mapper
(TM) scene (30m resolution) that covered 100% of the study area, and yielded an overall
classification accuracy of 91% using 256 randomly generated points (Table 1.1). For the
second assessment, I used high resolution (3m) photographs from a 3-band digital aircraft
camera from 20 October 1995 that covered approximately 43% of the study site. The
MSS data was subset to match the spatial extent of the aircraft photographs. One
hundred ten points were randomly generated to represent 43% of the total area. The
overall classification accuracy of this assessment was 87%. Because both assessments

yielded similar results, I present data for only the Landsat TM scene since it covered the

entire study area. No reference imagery was available for the 1970’s and 1980’s
classifications. However, given the high accuracy of the 1990’s classification, I assumed
that the identical techniques used in the development of all three classifications would
have produced comparable accuracy. Field validation would have provided optimal
assessment of my classification accuracy. However, due to the large size of my study
area and the fact that my study used historical data, ground-truthing was not possible
(Jensen and others 1995).

Land Change Analysis and Transition Matrices

Land cover area and percentages for the entire upper basin were calculated for
each time step for individual cover types, and for two broader categories: developed land
(urban and agriculture) and natural land cover (bottomland hardwood forest, swamp
forest, and freshwater marsh). I developed transition matrices to identify shifts between
individual land cover types in successive time steps and overall changes occurring
throughout the entire study period. To develop these matrices I calculated: (1) the area of
land in ha converted from each land cover into each of the remaining categories, (2) the
proportional area of land (%) converted from each land cover into other categories, and
(3) the total proportion (%) of each individual land category converted into other

categories.

Results
In all time steps, agriculture and urban land cover dominated the upland areas
around the perimeter of the basin and natural land classes dominated the lower elevation

areas (Figure 1.2). However, distinct changes in land cover occurred across the time

18

steps (Figure 1.3). Agricultural land, bottomland hardwood forest and freshwater marsh
all decreased through time; whereas urban land and swamp forest increased. Because
open water remained unchanged throughout the study period. I removed this category
from all further analysis.

Across the broader categories, there were small changes in total human dominated
land cover and in total natural land cover (Figure 1.4). However, there were large
changes within each category. In 1972, agriculture land cover made up 75% of the
human developed land, but only approximately 50% by 1992. Urban land cover made up
25% of the human developed land in 1972 and increased to approximately 50% by 1992.
For the natural land cover category, bottomland hardwood forest and freshwater marsh
decreased whereas swamp forest increased throughout the entire study period. I further
combined two wetland categories (freshwater marsh and swamp forest) to examine total
wetland area changes in the upper basin (Figure 1.5). In 1972, the total wetland area was
64,718 ha (38% of the upper basin). By 1992 this area had increased to 78.488 ha (46%),
gaining 13,770 ha over the entire study period.

Table 1.2 shows the rate of land cover change between each time step. The
largest overall rates of change across the whole study period was an increase of 937 ha/yr
of swamp forest from 1972-1992, and an increase of 993 ha/yr for urban land cover from
1972-1992. However, the rate of change of urban land significantly decreased over the
next time step to 288 ha/yr (1985-1992). Freshwater marsh showed a net loss at a rate of
363 ha/yr (4,720 ha) between 1972 and 1985, but only 36 ha/yr (249 ha) between 1985
and 1992.

To examine in more detail how land classes changed, I calculated land cover

transition matrices (Tables 1.3-l.5). Two land cover conversions were most prominent:
the conversion of land in the upland perimeter of the basin from agriculture to urban land,
and the conversion in the lower elevations from bottomland hardwood forest to swamp
forest land cover. All land cover categories changed as might be expected: agriculture
land cover primarily changed to urban; freshwater marsh to swamp forest; and
bottomland hardwood forest to swamp forest. However, although most urban land cover
was converted from agricultural land there were also unlikely transitions from urban land
to freshwater marsh, swamp forest, and bottomland hardwood forest. In most cases these
transitions occurred at a rate of 1 percent or less, which is within the error of the data.
Another unlikely transition was from urban land cover to agriculture land cover (3 - 6%
of the total urban area of the upper basin landscape) (Table 1.5). This transition is also

likely to be a misclassification, and it also falls within the error of the data.

Discussion and Conclusions

Studies of wetland change in the Barataria basin as well as other coastal estuaries
have focussed on wetland loss primarily in lower coastal zones (Turner 1997, Martin and
others 2000, Day and others 2000). However, because the upper freshwater zones of
coastal estuaries are subject to different human influences and physical drivers than the
lower coastal zones, different trends in land cover changes are likely to occur. I found
that important changes have occurred in the upper freshwater parts of the basin, which
are quite different from trends observed for lower coastal zones. In particular, I found
two major land cover trends in the upper Barataria Basin. The first is a change within

human developed land cover (agriculture and urban land cover) in the upland regions of

20

C C
1111

102

C01

111C
[tau
fine
don
litre
Sub.
Hm:

"10d

the upper basin, and the second is a change within the natural land cover (swamp forest,
bottomland forest and freshwater marsh) in the low elevation regions of the upper basin.

I explore both trends in detail below.

For the upper Barataria Basin, Hopkinson and Day (1980) predicted a 321%
increase in urban land from 1975-1995 based on projections from the Louisiana State
Planning Office. I found only a 108% increase in urban land from 1972-1992, which is
lower than predicted, but still large. Based on data from 1985 to 1992, it appears that
urban lands are increasing at a rate of 288 ha/yr at the expense of agricultural lands. The
conversion from agricultural lands in 1972 to urban lands in the upper basin in the 1990’s
and beyond may have significant impacts on nutrient loading to the adjacent swamp
forest, the open water areas, and the lower coastal zones, which warrants further research.
Any changes that are observed in natural land cover in the lower elevations must be

considered in light of these changing patterns in upland areas.

In the low elevation regions of the upper basin. 1 found that swamp forest
increased at the expense of bottomland forest and freshwater marsh, although the
transition from bottomland forest was the dominant change. My results suggest an
encroachment of the wetter swamp forest areas into higher elevation edge areas that are
dominated by bottomland hardwood forest. I found that 46% of bottomland hardwood
forest was converted to swamp forest over this period. It is very likely that land
subsidence is the cause of these changes in natural land cover (Kress and others 1996).
However, I must also consider other human drivers of change. Although I found only
moderate transitions from bottomland forest to urban land cover (1 15 ha/yr), this rate

may increase if urbanization continues, and as agricultural lands are diminishing, creating

21

a further threat to bottomland hardwood forests. Based on my results, if present trends
in the reduction of bottomland forest land cover were to continue, the upper Barataria
Basin may have no bottomland hardwood forests left by the year 2025, as it is subjected
to multiple stressors both in the higher elevations (from urbanization) and lower
elevations (most likely from land subsidence). Protections should be put in place now to

prevent further loss of this important habitat type of the upper basin.

Previous studies suggest that high rates of land subsidence within the Louisiana-
Mississippi River deltaic plain are one of the primary causes of wetland loss within
coastal regions of Louisiana (Scaife and others 1983, Conner and Day 1988, Visser and
others 1999). Land subsidence within the deltaic plain results primarily from the
compaction of sediments within the basin over time. The complete levying of the basin
has largely eliminated overflow of the Mississippi River, which was the primary source
of marsh-building sediments (Visser and others 1999). Land subsidence may also be
related to increases in water level, which has been measured at approximately 8.5 mm/yr
within the upper Barataria Basin (Conner and Day 1988). Conner and Day (1988)
suggested that water levels should increase over time and, along with decreasing
sedimentation rates (vertical accretion), lead to a reduction in forested wetlands as a
result of prolonged, deep flooding. I did not find a reduction in swamp forest wetlands in
the Barataria upper basin; rather I found that swamp forest area increased 37%. One
possible explanation for the increased swamp forest in my study is the existence of
numerous oil, gas and drainage canals throughout the basin. These canals may allow
swamp forest areas to periodically drain during drier seasons, providing the necessary dry

seed recruitment periods for the baldcypress and water tupelo species that dominate this

22

land cover class (Conner and others 1986, Souther and Shaffer 2000). However, it is
quite possible that the presence of these canals have only slowed changes predicted by
Conner and Day (1988). A second possible explanation for the increased swamp forest
would be if rainfall had changed in any way over time. However, annual rainfall data
measured at the New Orleans, Louisiana Weather Data Center shows no trend through

time.

Current management plans, such as the Coastal 2050 management plan (Bourne
2000), call attention to the alarming rate of shoreline erosion in coastal zones, and to a
lesser degree, the impacts of inland wetland and marsh change higher in the watershed.
Management strategies include reintroducing Mississippi River sediments over portions
of the Barataria coastal areas through freshwater diversion projects and filling some
canals that have led to saltwater intrusion and tidal action in the middle and lower basins.
However, these strategies may do little to affect land cover changes in the upper basin.
Freshwater diversions, which target rebuilding coastal margins of the basin that are
subject to erosion and sea level rise, may be instituted at points too low in the watershed
to aid in rebuilding land that could eventually be reclaimed by bottomland forest areas in
the upper basin. The levee system, which completely isolates both the upper and lower
Barataria Basin from Mississippi River sediments, will continue to inhibit vertical
accretion, allowing subsidence to continue to increase the lower wetland area in the upper

basin at the cost of reductions in bottomland hardwood forests.

Although bottomland hardwood forests provide important habitat for wildlife.
migratory bird species, and sport and commercial resources, the expanding wetlands in

the upper region may provide benefits as well. My results suggest that urbanization

within this region of the Barataria Basin may continue which increases the threat of
nonpoint source pollution. Expanding wetland areas in the upper basin may provide an
increased capacity to buffer excess upland nutrient loading to the inland lakes within this
region of the basin, and reduce sediment and nutrient export into the Gulf of Mexico.
Furthermore, with the alarming rates of coastal wetland loss in this region, upper basin
wetlands may prove to be critical in maintaining the rich ecological habitats of southern
Louisiana. However, further research needs to be conducted to assess the impacts of

losing bottomland forest in this important ecosystem.

Quantifying historical land cover change provides a reference for investigating
current landscape change. However, large-scale studies can be limited by the ability to
acquire timely, cost effective, and accurate data. Remote sensing using sensors that span
decades allows us to quantify historic land cover patterns for large areas such as the
upper Barataria Basin. My classification of Landsat MSS data produced an overall
estimated accuracy of 89% among the six land use categories, which is an error rate that
is lower than the changes that I detected. Failure to understand the type of changes that
are taking place within a watershed, and the degree to which these changes occur can lead
to serious errors in assessing ecological risks. Remote sensing provides both temporal
and spatial information on the structure of the landscape. Further study is needed to
examine the effects of these observed land cover changes on water quality within both the

upper and lower Barataria basins.

24

Acknowledgements

This research was supported by funds from the Earth System Science Office at the
NASA Stennis Space Center, MS, provided to S.A.C.N. from the NASA Graduate
Student Research Fellowship Program and the Michigan State University Minority
Competitive Doctoral Fellowship. 1 thank Richard Miller, Armond Joyce, Marco
Giardino, Greg Booth, Patrick Jones and Ramona Pelletier-Travis for technical
assistance, laboratory space and computing resources at Stennis Space Center. Partial
funds were provided by a NASA rangeland project at MSU/USDA. I are grateful to
J ianguo Liu, David Lusch, David Skole and Walter Chomentowski for their contributions
to this research. I would also like to thank Kendra Spence Cheruvelil and Marc
Linderman for providing comments on earlier drafts that greatly improved the
manuscript. Finally, I would like William Taylor and James Jay for their invaluable

support and assistance.

Literature Cited

Bourne, J. 2000. Louisiana’s vanishing wetlands: going, going... Science 289: 1860-1863.

Conner, W.H., J .R. Toliver, and EH. Sklar. Natural regeneration of baldcypress
(Taxodium distichum) in a Louisiana swamp. Forest Ecology and Management
14:305-317.

Conner, W.H., J.W. Day, J.G. Gosselink, C.S. Hopkinson, and WC. Stowe. 1987.
Chapter 4 vegetation: composition and production. Pages 31-47 in W.H. Conner
and J.W. Day. (eds.) The ecology of Barataria Basin, Louisiana: an estuarine

profile. US. Fish and Wildlife Service. Biological Report 85(7.13).

25

Conner, W.H. and J.W. Day. 1988. Rising water levels in coastal Louisiana:
Implications for two coastal forested wetland areas in Louisiana. Journal of
Coastal Research 4(4):589-596.

Dahl, T.E., Johnson, C.E., Frayer, W.E. 1991. Wetlands status and trends in the
conterminous United States, mid- 1970's to mid-1980's. US. Department of the
Interior, Fish and Wildlife Service, Washington. DC, 1-15 pp.

Day, J.W., L.D. Birtsch. S.R. Hawes, G.P. Shaffer, D. J. Reed, and D. Cahoon. 2000.
Patter and process of land loss in the Mississippi Delta: A spatial and temporal
analysis of wetland habitat change. Estuaries 23(4):425-438.

Feierabend, J .S., and J.M. Zelazny. 1987. Status report on our nation's wetlands. National
Wildlife Federation, Washington, DC, 22-46 pp.

Haaek, B. 1996. Monitoring wetland changes with remote sensing: an east African
example. Environn‘zental Management 20(3):4l 1—419.

Hopkinson, CS, and J .W. Day, Jr. 1980. Modeling hydrology and eutrophication in a
Louisiana swamp forest ecosystem. Environn-zental Management 42325-335.

Jensen, J.R., Hodgson. M.E., Christensen. E.. Mackey. H.E., Tinney, L.R., Sharitz, R.
1986. Remote sensing inland wetlands: a multispectral approach.
Photogrammetric Engineering and Remote Sensing 52(1):87-100.

Jensen, J.R., Rutchey. K, Koch, MS, and Narumalani, S. 1995. Inland wetland change
detection in the Everglades Water Conservation Area 2A using a time series of
normalized remotely sensed data. Photogrammetric Engineering and Remote

Sensing 61(2): 199-209.

26

Jensen, John R., 1996. Introductory digital image processing: a remote sensing
perspective. Upper Saddle River, New Jersey. 318 pp.

Kress, M.R., M.R. Graves, and SC. Bourne. 1996. Loss of bottomland hardwood forests
and forested wetlands in the Cache River Basin. Arkansas. Wetlands 16(3): 258-263.

Martin, J.F., M.L. White. E. Reyes, G.P. Kemp, H. Mashriqui, and J.W. Day. 2000.
Evaluation of coastal management plans with a spatial model: Mississippi Delt,
Louisiana, USA. Environmental Management 26(2):l 17-129.

Miller, R.L., M. Giardino, B.A. McKee, J .F. Cruise, G. Booth, R. Rovansek, D. Muirhead.
W. Cibula, K. Holiday, R.E. Pelletier, W. Hudnall, C. Bergeron, J. Ioup, G. Ioup,
and G. Love. 1995. Processes and fate of sediments and carbon in Barataria Bay,
LA. Pages 18-20 in Proceedings from the Third Thematic Conference on Remote
Sensing for Marine and Coastal Environments, 18-20 September 1995. Seattle,
Washington.

Mitsch, W. J. and J. G. Gosselink. 1986. Wetlands. Van Nostrand Reinhold. New York,
New York, 456 — 458 pp.

Munyati, C. 2000. Wetland change detection on the Kafue Flats, Zambia, by
classification of a multitemporal remote sensing image dataset. International
Journal of Remote Sensing 21(9): 1787-1806.

Ricketts, Peter J. 1992. Current approaches in geographic, information systems for
coastal management. Marine Pollution Bulletin 25(1-4):82—87.

Sasser, C.E., M.D. Dozier, J.G. Gosselink, J.M. Hill. 1986. Spatial and temporal changes
in Louisiana’s Barataria Basin marshes. 1945—1980. Environmental Management

10(5):67 1-680.

27

Tu

Scaife, W.W., R.E. Turner, R. Costanza. 1983. Coastal Louisiana recent land loss and
canal impacts. Environmental Management 72433-442.

Sklar, F.H., and J .A. Browder. 1998. Coastal environmental impacts brought about by
alterations to freshwater flow in the Gulf of Mexico. Environmental Management
22:547-562.

Souther, RF, and GP. Shaffer. 2000. The effects of submergence and light on two age
classes of baldcypress (Taxodium Distichum ( L.) Richard) seedlings. Wetlands
20(4):697-706.

Stow, C.A., R.D. De Lune, W.H. Patrick Jr. 1985. Nutrient fluxes in a eutrophic coastal
Louisiana freshwater lake. Environmental Management 9(3):243-252.

Templet, RH, and K.J. Meyer-Arendt. 1988. Louisiana wetland loss: a regional water
management approach to the problem. Environmental Management 12(2): 181-
192.

Tiner, R.W., I. Kenenski, T. Nuenninger, J. Eaton, D.B.'Foulis, G.S. Smith, and WE
Frayer. 1994. Wetlands of the Chesapeake Watershed: status and, trends from
1982 to 1989. Technical Report RSES, Fish and Wildlife Service. Region 5,
Ecological Services, Hadley, MA, 71 pp.

Turner, RE. 1997. Wetland loss in the northern Gulf of Mexico: multiple working
hypotheses. Estuaries 20: 1-13.

Visser, J .M., C.E. Sasser, R.H. Chabreck. and RC. Linscombe. 1999. Long-term vegetation

change in Louisiana tidal marshes. 1968-1992. Wetlands 19(1):168-175.

28

Chapter 2: Assessing regional lake water clarity using Landsat and the role of

inter-lake variability

Introduction

Monitoring the status of inland lakes is critical because they provide an important
recreational, commercial and aesthetic resource to the public. The water quality
monitoring of lakes often includes the monitoring of water clarity using a Secchi disk.
Although more sophisticated measurement techniques exist, Secchi disk transparency
(SDT) appears to be a relatively good coarse measure of water clarity, partly because it is
inexpensive and easy to use. The use of SDT has been widely adopted in many lake
monitoring programs worldwide (Bukata et al., 1995). Several studies have found SDT
to correlate well with a number of other water quality variables, such as trophic status,
phosphorus loads, planktonic chlorophyll concentrations, hypolimnetic oxygen
concentrations, suspended sediment concentrations, and fish yields (Oglesby, 1977;
Lorenza, 1980; Lee et al., 1995). Despite the ease of obtaining Secchi data, collecting
SDT on large numbers of lakes can be costly and challenging for monitoring agencies,
especially in states with thousands of lakes. However, in the past several decades,
citizen-volunteer lake monitoring programs using SDT measurements have allowed data
collection over larger regions than capable through local and state agencies alone. Many
studies have shown that the accuracy of volunteer collected is comparable to data
collected by professional monitoring personnel with no statistical difference in summer
averages of SDT between volunteer and professional measurements (Heiskary et al.,

1994; Kerr et al., 1994: Obrecht et al., 1998; Canfield et al., 2002). However, extending

29

volunteer collection programs to thousands of lakes, on a statewide level, is still costly
and logistically prohibitive. Thus, other approaches are needed to collect SDT on large
numbers of lakes.

Satellite remote sensing using Landsat Thematic Mapper (TM) has been explored
in several studies as a method of reducing the cost and labor of sampling water clarity in
the field (Khorram and Cheshire, 1985; Lathrop, 1992: Kloiber et al., 2000; Dewider and
Khedr, 2001). An advantage of using remote sensing is that data for multiple lakes
within a single image can be collected quickly and relatively inexpensively. Remote
sensing technology has been used for several years in oceanography to measure
chlorophyll, water color, and suspended sediments over large areas, but has only recently
been explored in lake studies (Packard and Emery, 1982; Kloiber et al., 2000). Although
sensors such as Landsat TM were primarily designed for detecting land features, recent
improvements now provide better spatial and spectral resolutions for aquatic studies
(Zilioli, 2001). Recently, remotely sensed data has been shown to correlate well with
lake SDT values (Khorram and Cheshire. 1985; Lathrop, 1992; Kloiber et al., 2000;
Dewider and Khedr, 2001). However, to effectively implement remote sensing into a '
state monitoring program for inland lakes, there still remain some unanswered questions.

For inland lakes, the relationship between remotely sensed images and SDT has
been examined through simple linear regression analysis of in-situ measurements of SDT
and spectral brightness values from the sensor (Lillesand et al., 1983; Lathrop and
Lillesand, 1986; Kloiber et al., 2002). Regression models developed for single water
bodies using a single image have resulted in high r2 values (0.80 — 0.90, p < 0.05)

(Lathrop and Lillesand, 1986; Lavery et al., 1993; Giardino et al., 2001 ). These strong

30

relationships suggest that remote sensing may be useful for measuring SDT in a large
number of lakes across larger spatial scales.

There have been fewer studies that have used Landsat data to measure SDT of
multiple lakes, over a large region, or covering multiple images (Table 2.1). The few
studies that have been conducted using multiple lakes from a single image have resulted
in relatively high r2 values (Table 2.1). However, to apply these approaches to the
statewide scale, such as for the lakes of Michigan, I must test this approach in cases
where I have data from multiple images and a wider diversity of lakes than is often
present in a single image. My objectives in this study were to determine ifI could use
Landsat-7 data to measure lake water clarity across a large region of lakes from three
images in a single path (94,350 kmz), and to determine how the regional distribution of

SDT affects models of regional lake water clarity.

Materials and Methods
Study area

The state of Michigan has approximately 3,500 inland lakes > 10 ha in surface
area and many thousands of smaller lakes (unpublished data from the Michigan
Department of Environmental Quality). My study lakes included 93 lakes in the lower
peninsula of Michigan (Figure 2.1). The lakes were distributed throughout an area of
94,350 kmz, which made up approximately 80% of the lower peninsula of the state. My
lakes ranged in surface area from 12 — 4,125 ha with a mean Secchi depth of 3.1 m (Table

2.2).

31

 

'—

\l'

Field Observations

Field observations of SDT data were obtained for the 93 lakes from three
sampling programs: (1 ) the Michigan Citizens Lake Monitoring Program (CLMP), (2) the
Michigan Department of Environmental Quality’s (MDEQ) Lake Water Quality
Assessment (LWQA) Monitoring Program, (3) Michigan State University (MSU; Table
2.2). I selected lakes that: (l) were sampled from late July through August to ensure
samples were taken during the summer stratified period; (2) had a surface area of 2 10 ha
within the lower peninsula of Michigan; and (3) were sampled within i 7 days of the
satellite imagery.
Satellite Image Data

I used three Landsat-7 Enhanced Thematic Mapper Plus (ETM+) scenes from
August 28, 2001 from ground track Path 21 (Worldwide Reference System-2) that
covered the lower peninsula of Michigan (Figure 2.1 ). Spectral digital number (DN)
values were extracted within the pelagic region of each lake, which was identified by
creating an area of interest (AOI) using ERDAS Imagine version 8.4 image processing
software. The pelagic region of each lake was defined within an area > 4.5 m in depth. I
used bathymetric maps to identify the 4.5 m depth contour within the 80 sample lakes for
which I had maps. For the 13 lakes without bathymetric maps, 1 created an A01 for each
lake from groups of pixels within the center of each lake, avoiding any shoreline or
shallow areas to ensure that each AOI would be free of reflections from the lake bottom
or submerged macrophytes. From each lake AOI. I extracted the mean DN for each
spectral band to use in all subsequent analyses. My AOI values ranged from 8 to 1,012

pixels (_-+_- 183 pixels).

32

Statistical Analysis

All statistical analyses were done using SYSTAT (SPSS Software, Inc., 2001).
Probability distribution plots of SDT indicated that a natural log-transformation was
necessary for SDT. I developed a linear regression model using the band ratio of
ETMl/ETM3 as the independent variable and the natural log of SDT as the dependent
variable. Previous studies found these regression variables to be the best predictor of
SDT (Lathrop, 1992; Pattiaratchi et al., 1994; Kloiber et al., 2000). Outliers and lakes
having large leverages within the SDT and Landsat datasets were removed from the final
regression analysis, reducing the number of lakes for the final analysis from 96 to 93.

To examine the role of lake SDT distribution in My regression model, I developed
an additional regression model with a different SDT distribution based on a subsample of
My complete dataset. I used a manual selection technique to create the subsampled
dataset with a similar SDT distribution to match a previous study that found a strong
relationship between multiple lake SDT values and Landsat spectral values (Kloiber et
al., 2000). To compare the regressions of the complete and subsampled datasets, I used a
slope heterogeneity test and analysis of covariance (ANCOVA). I also performed an F-
test on the residual values to determine if there was a significant difference between the

variances of the residuals of each dataset.

SDT Distributions
I extracted values of SDT in Figure 2.5 of Kloiber et al. (2000) to compare their
SDT distribution to my complete and subsampled dataset. I also compiled a distribution

of statewide Michigan summer lake SDT data available from the US EPA Storage and

b.)
DJ

Retrieval (STORET) database. These data included 675 lakes (>20 ha) sampled during
the 1970’s and 1980’s. Data from some lakes were averaged across sample years if
multiple years were sampled. Otherwise the data represented average summer SDT

values from biweekly samples in June, July, and August.

Results

My regression of SDT and ETMl/ETM3 produced an r2 value of 0.43 (p < 0.001;
Figure 2.2). To examine why my models explained so much less variation than previous
studies, I compared the SDT distribution of the 47 lakes included in the Kloiber et al.
(2000) study to the distribution of my 93 lakes. 1 found that approximately 62% of the
lakes used in the Kloiber et al. (2000) study had a SDT of 1.5 m or less (Figure 2.3). For
the 93 lakes used in my study, only 9% had a SDT of 1.5 m or less (Figure 2.3). Thus,
the majority of lakes included in Kloiber et al. (2000) are predominantly eutrophic
(Forsberg and Ryding, 1980). The lakes included in my study would be considered
largely mesotrophic to oligotrophic (Figure 2.3). To examine the effect these different
distributions may have on SDT prediction. I manually adjusted my SDT sample
distribution to have a mean Secchi depth of 1.8 meters and a distribution with 47% of the
lakes having a SDT of g 1.5m. This subsample included 17 lakes and resembled the
sample distribution found in the Kloiber et a1. (2000) study (Figure 2.4). The regression
for this subsampled dataset had an r2 value of 0.82 (p < 0.001; Figure 2.5).

I compared the trend lines from the regressions of my two models, the complete
dataset model (n = 93) and the subsampled dataset model (n = 17), and found that the

difference in the slopes of the two models was not statistically significant (ANCOVA p =

34

0.247; Figure 2.6). I also examined the residuals against their predicted Landsat ETM
values (Figure 2.7). This analysis suggests that there is less variability in the residuals
about the mean for the subsampled dataset than the complete dataset, thus leading to the
lower r2 value of the complete dataset model. I used an F-test to determine if the
variance of the residuals of the two datasets were significantly different. This test
indicated that there was a significant probability that the residual variances in the two
datasets are statistically different (F = 2.86. p = 0.0099).

Finally, I further explored the models by plotting the residuals against the
observed Secchi values (log transformed) to determine if the residual errors differed at
different SDT values (Figure 2.8). This analysis suggests that the complete dataset model
may provide better SDT predictions, as its confidence intervals for the predicted mean are

narrower for both shallower and deeper Secchi depths than for the subsampled dataset.

Discussion and Conclusions

My results show that I can develop a standard regression model for many lakes
within a large geographic region, encompassing multiple Landsat scenes, to predict SDT.
These results are in agreement with authors such as Lathrop and Lillesand (1986) who
recommended the establishment of common statistical models (e. g. standardized
regressions) to use remote sensing for monitoring water quality and enhance the
development of future sensors tailored to freshwater lake systems. Several previous
studies have developed successful statistical relationships between a single Landsat
image and water quality of a single lake or water body. However, to my knowledge only

Kloiber et al. (2000) has examined this relationship on a larger scale, using 47 lakes

35

within a 7,700 km2 area. The results from Kloiber et al. (2000) show promise in
developing larger scale satellite-based water clarity monitoring protocols using my
current sensors. My study expands on this work by examining almost twice the number
of lakes within an area of approximately twelve times the areal extent of the Kloiber et al.
(2000) study. However, I had much more inter-lake variation in my study leading to
substantially lower r2 values than those in the Kloiber et al. (2000) study. Previous
single-lake studies using similar Landsat bands also have higher regression results than I
did (Lathrop, 1992; Pattiaratchi et al., 1994; Cox et al., 1998). This discrepancy was
likely a result of the SDT distribution of my study lakes. My complete dataset consisted
of a broad range of SDT values, including deeper SDT lakes. Lakes with deeper SDT
values return less signal reflectance from algal turbidity in the water column to a remote
sensor and, thus, may not be well detected using current sensors. In contrast, the Kloiber
et al. (2000) study was based on a cluster of lakes within the metropolitan area of
Minneapolis and St. Paul, Minnesota, USA, in which over 60% of the lakes included
were eutrophic (Forsberg and Ryding, 1980). My subsampled dataset incorporated a
larger percentage of eutrophic lakes and yielded a comparable r2 result to the Kloiber et
a1. (2000) study. My results suggest that inter-lake variably plays an important role in
influencing the strength and reliability of regression models of Landsat and lake SDT.
Two questions remain: ( 1) what are the best ranges of SDT that can be predicted
from Landsat; and (2) what are the best regression models for regions containing lakes
with a wide range of SDT values? To address the first question, I explored the residual
plots of each dataset against the observed SDT (Figure 2.8). This analysis is useful to

determine how the residual errors differed in closeness of fit about the mean for both

36

models. The confidence intervals of the subsampled dataset were wider for both

shallow and deep SDT values than for the complete dataset. Although a wider interval in
the subsampled dataset may be related to the much smaller sample size of the subsampled
dataset, the wider interval may also indicate a higher overall degree of uncertainty, or
possibly inconstant variance throughout the residual terms (Van Belle, 2002). Both the
complete and subsampled dataset plots show a distinct cluster of residual values within
the 2 - 3 m Secchi range and a narrowing of the confidence intervals within this range.
This result may suggest that my models are better suited for predicting Secchi depths
within these ranges than they are at predicting Secchi depth at the shallow or deep end of
the distribution. Overall, my residual analyses showed wider confidence intervals in the
subsampled dataset compared to the complete dataset. This suggests the results of the
complete dataset model provides a better estimate of the range of the true distribution, but
that less of the variation in the dataset is explained.

To address the second question of which regression model is better for predicting
SDT in regions with a wide range of SDT values. we first need to consider the goal of the
study in question. For shallow SDT lake studies within a smaller region, models such as
my subsampled dataset and the Kloiber et al (2000) study should provide desirable
results. This type of investigation could be particularly useful for studies of eutrophic
lakes in urbanized watersheds. However. if the goal is to examine SDT at a statewide
level, a broader range of inter-lake variability must be considered. For example, my
complete model, with a lower r2 value, incorporated a wide range of SDT. A wider range
of SDT should be expected when modeling many lakes throughout a large geographic

region, such as the state of Michigan. In fact. the range of the SDT data included in my

37

complete dataset model is comparable to the range of SDT collected from 675 lakes in
Michigan (US EPA STORET; Figure 2.9). The data for Secchi in the state of Michigan
ranges from 0.5 - 8.0 m, with a mean of 3.0 m. My complete dataset captures this range,
from 0.9 — 7.6 m with a mean 3.1 m (Figure 2.9).

To use remote sensing for statewide assessments of water clarity, there still must
be an investment in some degree of statewide sampling efforts that coincide with satellite
overflights. Although I included ground observations from a number of reliable sources,
I was still somewhat limited in the number of lakes to include in the model. For example,
during this study, I experienced a large number of days with high cloud cover that
resulted in few cloud free scenes. This made it difficult to match field SDT data from
more than 93 lakes across the state within i 7 days of the satellite over flight. In
addition, I experienced difficulties in creating a single regression relationship that
included images from multiple paths. Further research is necessary to explore the effect
of path to path atmospheric variations. Within the three contiguous Landsat scenes used
in this study, atmospheric corrections were not necessary for sampling SDT because the
images were free from atmospheric haze and all of the images were taken on a single
date.

In conclusion, although a standard equation relating Landsat data and water
clarity over a large area can be used to predict SDT, the sensitivity of the regression
model to the distribution of lake calibration data must be taken into consideration.
Further research is required to determine the optimal ranges in which Secchi depth can be
predicted from Landsat data, the reliability of predictive models in shallow as well as

deep SDT lakes, and the best regression estimates for states with a wide range of Secchi

38

depth values. Remote sensing can play a vital role in reducing the cost, labor, and time
required to monitor inland lake water clarity. It has the potential to be used as a tool to
develop statewide assessments that are currently impossible using traditional field
operations. However, caution should be used in assessing water quality based solely on
remotely sensed data. Current methods and sensors may be able to effectively measure
shallow SDT lakes, but may not do as well for lakes with deeper SDT. Until our method
can be improved or possibly better sensors are developed for aquatic applications, inland
lake remote sensing will be most useful as a supplement to existing volunteer and agency

monitoring programs, rather than a replacement of these programs.

Acknowledgements

This research was funded in part by a Michigan State University Minority
Competitive Doctoral Fellowship to SACN. Landsat-7 ETM+ imagery was provided by
the Basic Science and Remote Sensing Initiative (BSRSI, now known as the Center for
Global Change and Earth Observation) and through the N ASA-funded Upper Midwest
RESAC project (Regional Earth Science Applications Center). Additional funding was
provided through a USGS/DEQ grant to BSRSI. The CLMP field observation data were
provided by Ralph Bednarz, MDEQ. I thank Kevin Hoffman, John Whitlock, Tamara
Brunger-Lipsey, Kristy Rogers, Sarah Wills. and Konrad Kulacki for field data collection
and processing. I am also grateful to J iaguo Qi, J ianguo Liu, David Lusch, and Walter
Chowmentowski for their contributions to this research. Additionally, I would like to
thank Jonathan Chipman, Michael Belligan, Narumon Wiangwan g, Marc Linderman, and

Chris Barber for technical support and database assistance. Finally. I thank Daniel Hayes

and Orlando “Ace” Sarnelle for statistical advice.

Literature Cited

Bukata, R. P., Jerome J. H., & Burton J. E. (1988). Relationships among Secchi disk
depth, beam attenuation coefficient. and irradiance attenuation coefficient for
Great Lakes waters. Journal of Great Lakes Research, 14(3), 347-355.

Canfield, D. E., Jr., Brown. C. D., Bachmann, R. W., & Hoyer, M. V. (2002). Volunteer
lake monitoring: testing the reliability of data collected by the Florida
LAKEWATCH program. Lake and Reservoir Management, 18, 1-9.

Cox. R. M., Forsythe, R. D., Vaughan, G. E.. & Olmstead, L. L. (1998). Assessing water
quality in the Catawba River reservoirs using Landsat Thematic Mapper satellite
data. Lake and Reservoir Management, 14. 405-416.

Dekker, A.G., & Peters S.W.M. (1993). The use of the Thematic Mapper for the analysis
of eutrophic lakes: a case study in the Netherlands. International Journal of
Remote Sensing, 14(5), 799-821.

Dewider, K. & Khedr, A. (2001). Water quality assessment with simultaneous Landsat-5
TM at Manzala Lagoon, Egypt. Hydrobiologia, 457, 49-58.

Forsberg, C. & Ryding S. O. (1980). Eutrophication parameters and trophic state indices
in 30 Swedish waste-receiving lakes. Archiv fur Hydrobiologie. 88: 189-207.

Giardino, C., Pepe, M., Brivio, P. A., Ghezzi, P., & Zilioli, E. (2001). Detecting
chlorophyll, Secchi disk depth and surface temperature in a sub-alpine lake using

Landsat imagery. Science of the Total Environment, (268'). 19-29.

40

Heiskary, S., Lindbloom, J., & Wilson, C. B. (1994). Detecting water quality trends
with citizen volunteer data. Journal of Lake and Reservoir Management, 9(1), 4-
9.

Kerr, M., Ely, E., Lee, V., & Mayio, A. (1994). A profile of volunteer environmental
monitoring: National survey results. Journal of Lake and Reservoir Management,
9(1), 1-4.

Khorram, S. & Cheshire H. M. (1985). Remote sensing of water quality in the Neuse
River Estuary, North Carolina. Photogrammetric Engineering and Remote
Sensing, 51(3), 329-341.

Kloiber, S.M., Anderle, T.H., Brezonik, P.L., Olmanson, L., Bauer, M.E. & Brown, DA.
(2000). Trophic state assessment of lakes in the Twin Cities (Minnesota, USA)
region by satellite imagery. Archive Hydrobiologie Special Issues Advances in
Limnology. 55,137-151.

Kloiber, S. M., Brezonik, P. L., Olmanson, L. G., & Bauer, M. E. (2002). A procedure for
regional lake water clarity assessment using Landsat mutispectral data. Remote
Sensing of Environment (in press).

Lathrop, R. G. (1992). Landsat Thematic Mapper monitoring of turbid inland water
quality. Photogrammetric Engineering and Remote Sensing, 58(4), 465 - 470.

LathrOp, R. G., & Lillesand, T. M. (1986). Utility of Thematic Mapper data to assess
water quality. Photogrammetric Engineering and Remote Sensing, 52, 671-680.

Lavery, P., Pattiaratchi, C., Wyllie, A., & Hick, P. (1993). Water quality monitoring in
estuarine waters using the Landsat Thematic Mapper. Remote Sensing of

Environment, 46, 268-280.

41

 

St;

Lee, G. F., Jones-Lee, A. F., & Rast, W. (1995). Secchi depth as a water quality
parameter. Repon of Fred Lee & Associates, El Macero, CA. (15 pp.).

Lillesand, T. M., Johnson, W.L., Deuell, R. L., Linstrom, O. M., & Meisner, D. E.
(1983). Use of Landsat data to predict the trophic state of Minnesota lakes.
Photogrammetric Engineering and Remote Sensing, 49(2), 219 - 229.

Lorenza, M. W. (1980). Use ofchlorophyll-Secchi disk relationships. Limnology and
Oceanography, 25(3), 371-372.

Obrecht, D. V., Milanick, M., Perkins B. D., Ready, D., & Jones, J. R. (1998).
Evaluation of data generated from lake samples collected by volunteers. Journal
of Lake and Reservoir Management, 14(1), 21-27.

Oglesby, R. T. (1977). Relationships of fish yield to lake phytoplankton standing crop,
production, and morphedaphic factors. Journal of Fisheries Research Board of
Canada, 34, 2271-2279.

Packard, G. L., & Emery, W. J ., (1982). Descriptive physical oceanography: an
introduction. Pergamon Press. New York, New York. 4th ed. (47-58 pp.).

Pattiaratchi, C. , Lavery, P., Wyllie, A., Hick, P. (1994). Estimates of water quality in
costal waters using multi-date Landsat Thematic Mapper data. International
Journal of Remote Sensing, 15(8). 1571 -1584.

Song, C., Woodcock, C. E., Seto, K. C., Lenney, M. P., & Macomber. S. A. 2001.
Classification and change detection using Landsat TM data: when and how to
correct atmospheric effects? Remote Sensing of Environment. 75, 230-244.

Van Belle, G. (2002). Statistical rules of thumb. Wiley and Sons. New York, New York.

(16-19 pp.).

42

Wallin, M. L., & Hakanson, L. (1992). Morphometry and sedimentation as regulating
factors for nutrient recycling and trophic level in coastal waters. Hydrobiologia,
235, 33-45.
Zilioli, E., (2001). Lake water monitoring in Europe by means of remote sensing. Science

of the Total Environment. 268:1-2.

43

Chapter 3: Remote sensing of freshwater macrophytes and the influence of lake

characteristics

Introduction

Macrophytes play several important roles in lake ecosystems, such as stabilizing
shorelines, reducing erosion, and improving water clarity. Macrophytes also provide
many benefits for aquatic organisms, such as providing refuge and nursery habitat for
young and small species, increasing dissolved oxygen in water, and providing surface
area for the growth of invertebrates and algae, which provide food for larger organisms
(Carpenter and Lodge, 1986). For example, of particular interest to many lake managers
is the role that macrophytes play in fish growth. Based on a qualitative review of the
literature, Schneider (2000) reported that 25-36% macrophyte cover is optimal for
largemouth bass (Micropterus salmoides) and bluegill (Lepomis macrochirus) growth
rates. Thus, knowing information about the macrophyte community, such as the amount
of macrophyte cover in a lake can provide valuable information.

Information on the species composition of plants can also be important, especially
the presence or absence of exotic plants. Nuisance exotic macrophyte species, such as
Eurasian watermilfoil (Myriophyllum spicatum, EWM), are now widespread throughout
much of North America (Couch and Nelson, 1985). EWM forms dense surface canopies
that suppress native plant growth, which can lead to homogeneous macrophyte beds
(Aiken et al., 1979; Madsen et al., 1988, Madsen et al., 1991). High percent cover of
these dense canopies and loss of macrophytes diversity can influence littoral zone food

webs in many important ways (Lyons 1989; Trebitz et al., 1996; Cheruvelil et al., 2002).

44

IE

311

These concerns have prompted much research into EWM ecology and management
(Chilton, 1990, Smith and Barko, 1990, Trebitz et al.. 1993). However, management
agencies in many affected areas do not have good data on the extent of the EWM
problem in their state. This lack of information is largely due to the expense and
challenges associated with sampling macrophytes.

Assessing macrophyte distribution and abundance in an individual lake takes a
significant amount of time, expense and labor. Assessments are further complicated in
regions containing thousands of lakes. For example. the state of Michigan has
approximately 3,500 inland lakes > 10 ha in surface area and many thousands of smaller
lakes (unpublished data. Michigan Department of Natural Resources). Currently, it is not
possible to inventory this many lakes using traditional field sampling techniques such as
sampling along transects, within quadrants, or subsampling randomly-stratified lake
points. Although these techniques can give good estimates of local macrophyte biomass
and species composition, and results from these approaches are sometimes extrapolated
to the whole lake, these methods cannot capture the patchy distribution of aquatic plants
in an entire lake (Zhang, 1998). Remote sensing has the potential to help improve
estimates of plant cover because of its ability to provide a synoptic view of pant
distribution within an entire lake and for multiple lakes within a single image. Using
remote sensing to measure plant cover in general and especially EWM, would be an
invaluable tool for lake managers to monitor the cover of macrophytes and the presence
and distribution of nuisance species.

Although satellite sensors such as Landsat Thematic Mapper (TM) were primarily

designed for detecting land features, recent improvements now provide better spatial and

45

spectral resolutions that may be applicable for aquatic studies (Zilioli, 2001). However,
remotely sensed images have been used to measure macrophyte cover by mapping the
areal distributions of macrophytes along coastal margins from visual interpretations of
aerial photographs (Orth and Moore. 1983; Marshall and Lee, 1994). Although this
technique provides a quick and cost effective method of qualitatively and quantitatively
assessing aquatic vegetation, studies have reported that these techniques correctly record
submersed coastal macrophytes only 56-70% of the time (Scholoesser et al., 1987;
Marshall and Lee, 1994). Additionally, this technique can be fairly labor intensive.
Some of the advantages of using remote sensing for lake studies include acquiring
instantaneous sampling of multiple lakes within a single image that can be collected
relatively inexpensively. However, satellite remote sensing of aquatic plants, especially
submersed plants, has been less studied as a result of the difficulties inherent in
interpreting reflectance values of water (Penuelas et a1. 1993, Lehmann et al., 1997).
Clear water provides little atmospheric reflectance and either absorbs or transmits the
majority of incoming radiation (Lillesand and Kiefer 1994, Verbyla 1995). For mapping
purposes, unless a sharp contrast exists between the water body and adjacent shore, soils,
or vegetation, even boundary delineations can be challenging using remote sensing
(Verbyla 1995). As a result, researchers have relied on using remotely sensed data to
detect primarily emergent vegetation or homogenous clusters of submersed vegetation
(Ackleson and Klemas, 1987; Armstrong, 1993). Other studies have reported successful
detection of submersed aquatic plants using high-resolution aircraft or hand sensors
(Penuelas et a1. 1993, Lehmann et al., 1997, Ullah et al., 2000). However, these studies

are mostly limited to few or individual macrophytes species and are often restricted to a

46

 

Sl

single water body. Despite the potential limitations of using current sensors such as
Landsat to detect submersed aquatic plants, more research clearly needs to be done to
determine whether Landsat can be used to map macrophytes in large numbers of lakes
across a state such as Michigan.

Several studies have shown that remotely sensed images can measure lake
characteristics such as chlorophyll, Secchi disk transparency, and suspended sediments
(Lathrop and Lillesand 1986, Jensen et al., 1993; Narumalani et al., 1997; Lillesand et al.,
1983; Khorram and Cheshire, 1985; Dekker and Peters. 1993; Kliober et al., 2000).
However, lakes within a region may vary widely in several of the above characteristics,
as well as others that may strongly influence how submersed plants are remotely sensed
such as watercolor and lake water depth. It is likely that it will be necessary to
incorporate such factors into models predicting aquatic macrophytes. For example, water
depth has also been incorporated into models to detect submersed plants using sensors
such as Landsat in individual water bodies (Raitala and Lampinen, 1985; Ackleson and
Klemas, 1987; Armstrong, 1993; Narumalani. 1997). However, to my knowledge, no
studies have linked the influence of these physical and chemical lake features to the
ability to detect submersed plants on a regional scale, across multiple lakes.

My objectives in this study were: 1) to determine if different aquatic plant cover
types and growth forms (overall littoral plant cover, emergent plants, floating leaf plants,
submersed vegetation, and submersed EWM) could be detected using the Landsat-5 TM
sensor and 2) to determine if I could improve predictions of macrophyte distributions in
lakes by including lake characteristics (Secchi disk transparency, chlorophyll a,

phytoplankton biovolume, water color, and sediment color) in the models. I hypothesize

47

that including these additional lake characteristics will improve the strength of

relationships between macrophyte cover and Landsat-5 TM spectral values.

Materials and Methods
Study Area
The 13 study lakes, located in the lower peninsula of Michigan, U.S.A., were

selected using a random-stratified procedure (Figure 3.1). These lakes are distributed
throughout an area of approximately 70,000 kmz. The lakes were stratified by surface
area (20 - 140 ha), mean depth (2 - 10 m), and Secchi disk transparency depth (0.6 — 8.2
m). A total of 36 lakes were selected and sampled. However, due to cloudy weather
during late summer 2001, the final analyses were conducted on only 13 lakes for which
cloud-free imagery could be obtained. One lake was selected randomly and withheld
from the model development for use as a validation dataset (Table 3.1a).
Macrophyte Sampling

The lakes were sampled during the summer-stratified season during peak plant
biomass (July 17 — September 12, 2001). Plants were sampled using a modification of the
point intercept method (Madsen, 1999; Figure 3.2). Each lake was mapped using a
geographic information system (GIS) and then overlain with a grid of points which
represented the macrophyte survey points in the field. Plant cover was assessed at each
point for an area of 40 x 40 m or 50 x 50 m depending on lake area. This technique
resulted in 138-467 points per lake. The sample points were located in the field with a
global positioning system (GPS). At each point, I measured water depth, assessed plant

cover of the site, and recorded plant presence in each of four categories: emergent,

48

as
1115
C0
CO
CO

p1:

1111'
16:
It‘s

rel

lllt
pl:

.11

La

floating leaf, total submersed, and submersed EWM. Plant cover at each site was
assessed by qualitatively assigning a ‘plant cover level’, which was done by visual
inspection in shallow/clear water, or by throwing a two—sided rake in deeper water. Plant
cover levels were: 0 (0 — 20% plant cover), 1 (21 - 40% plant cover), 2 (41 - 80% plant
cover), and 3 (81 - 100% plant cover). An additional category of total littoral zone plant
cover was developed by combining each of the 4 levels recorded at each point for each
plant category, which represents plant presence or absence at each point as 0 (0 - 20%

plant cover) or 1 (21 - 100% plant cover).

I also calculated littoral and total lake percent plant cover for each lake as the number
of points sampled with any plant category greater than level 0 (i.e., > 20% cover at an
individual site), divided by the total number of points in the littoral zone or lake,
respectively. The littoral zone was defined 5 4.5 m water depth. Depths > 4.5 m were
regarded as pelagic in which volume reflectance of the water column would dominate the
reflectance spectra necessary for submersed plant detection by Landsat (Lillesand et al.,

1983; Kloiber et al., 2000).

Examples of the plant species included in the emergent category were Typha latifolia,
Pontederia cordata, and Scirpus spp. Representative species of the floating leaf category
included Nuphar advena, Nymphaea odorata, and Brasenia schreberi. Examples of
plants in the total submersed category included Chara spp., Potamogeton spp., as well as
Myriophyllum spicatum. The submersed EWM category consisted of only the exotic,

Myriophyllum spicatum.

Lake Characteristics

In the pelagic zone, water samples were taken from the deepest area of each lake

49

for nutrients, chlorophyll a, total alkalinity, and total phytoplankton biovolume. For
each lake, the depth of the epilimnion was calculated from a temperature profile and an
integrated tube sampler was used to take an integrated epilimnetic sample. Total
alkalinity (mg/L CZICO3) was determined on site with a titration test kit (LaMotte). For
chlorophyll analysis, water was filtered on site through a glass fiber filter (Whatman GF-
C) and stored in dark containers until being returned to the lab and frozen. Chlorophyll a
concentrations were determined fluorometrically with phaeopigment correction following
24 hour extraction in methanol (Nusch, 1980). Total nitrogen was determined using a
persulfate digestion followed by second derivative spectroscopy (Crumpton et al., 1992).
Total phosphorus was determined using a persulfate digestion (Menzel and Corwin,
1965) followed by standard colorimetry (Murphy and Riley 1962). Secchi disk depth
was measured for each lake off of the shady side of the boat. Phytoplankton were
preserved with 1% Lugol’s solution and stored in the dark until samples were counted
and identified using an inverted microscope. For each lake, a 50 ml phytoplankton
sample was settled for 5 days except for 3 lakes (Nevins, Little Whitefish, and Round
(J ackson)), which were only settled for 48 hours. The counts from these 3 lakes were
then extrapolated to 5-day counts based on comparisons of the 2 settling lengths for 6
lakes that were counted after being settled for both 48 hours and 5 days. For each lake,
more than 400 individuals were counted using multiple magnifications (40, 100, 400, and
1000X), and phytoplankton densities (um3/ml) were then calculated.

I developed sediment type estimates in the littoral zone of each lake by visually

assessing an average of 6-10 sites per lake. I categorized the estimates of sediment type

into one of 4 categories: sand, gravel/rock/cobble, marl, or silt/much/peat.

50

Satellite Imagery

I used three Landsat-5 TM scenes: one from August 4, 2001 and two from
September 5, 2001. All three scenes were from the Landsat ground track Path 21
(Worldwide Reference System-2; WRS) (Figure 3.1). Image rectification and
geoprocessing was conducted using ERDAS Imagine 8.4 image processing software.
Spectral digital number (DN) values for all single pixels located at the same position of
each lake sample point were extracted from the imagery using the field-recorded GPS
coordinates within each lake. This extraction was performed using the ArcInfo GRID
module of ArcGIS 8.1 (ESRI). Spectral DN values for the pelagic region of each lake
(sample points with depth measurements > 4.5 m) were also extracted to analyze the
relationship between pelagic zone lake characteristics and spectral values (see below).
Because the pelagic zone is more homogeneous than the littoral zone, I averaged the
spectral DN values for all pixels within the pelagic zone to come up with one spectral
value.
Statistical Analysis

I regressed the satellite imagery DN values against macrophyte data using
binomial and multinomial logistic regression (logit models) in SAS/STAT software (SAS
Institute, Inc. 2001). Spectral DN values for TM bands 2 and 4 were selected as the best
bands to include in the models as the independent variables based on numerous attempts
to fit several individual and combined bands to the data (Nelson, unpublished data). In
all of the models, I also included lake depth at each sampling point as an interaction term
with each main effect variable (band 2 and band 4).

For all analyses, I included sample points from multiple lakes in the logit models

51

(unless noted otherwise). However, because I had images from two different sample
dates, I developed three separate models to assess whether including results from
different sample dates (and potentially different atmospheric levels, i.e. haze) would
change model predictions. Thus, I grouped the 12 sample lakes into three datasets to
coincide with the image dates: August, September, and a combined dataset (both August
and September lakes).

Each of the five plant categories served as individual response variables for each
of the three models. The logit model uses the explanatory and interaction covariates to
predict the probability that the response variable will take on a given value (SAS Institute
Inc., 1995). For binomial logistic regression, the logit model indicates how increases or
decreases in the explanatory variable (TM values) affect the probability of the event
being observed versus not being observed. For the multinomial logistic regression,
probable outcomes of observations are calculated by analyzing a series of binomial
submodels that represent the overall model’s ability to map changes in the explanatory
variable, which are related to the plant cover response variable. For all logit model
analyses, I used the descending option to select the highest plant category level as the
response variable reference (level 3 for emergent. floating leaf, total submersed, and
submersed EWM and level 1 for littoral plant cover). This selection ensured that the
results were based on the probabilities of modeling an event (plants present), rather than a
nonevent (plants not present).

I assessed the models by examining the percent concordant values and the Wald
test statistic. The percent concordant values provide an indication of overall model

quality through the association of predicted probabilities and observed responses. These

52

values are based on the maximum likelihood estimation of the percent of paired
observations of which values differ from the response variable (Kleinbaum, 1994; SAS
Institute, Inc. 2001). The higher the predicted event probability of the larger response
variable (based on the highest plant category level), the greater the percent concordant
value will be. The chi-square level of significance of the Wald test statistic tests the
hypothesis that the coefficients of the independent variables are significantly different
from zero by fitting the model using the intercept terms (Kleinbaum, 1994; Pampel,
2000). I used the Wald test because it is regarded as being more accurate than other test
statistics when large sample sizes are used (Kleinbaum, 1994).

To examine whether there were significant differences between data from the two
image dates, and whether it was valid to include lakes in a combined model, I developed
separate logit models for individual lakes. I then compared the model output and model
coefficients from lakes in the August image to lakes in September image. First, I used a
two sample t-test to test the difference between the means of the model coefficients (log
transformed) of the September lakes to the August lakes. The p-values reflected the p-
values of the pooled variance, and significance was determined at the 0.1 level.
Insignificant results from these tests suggest that the variance of the means of the 12
individual lakes are not statistically different, thus atmospheric effects may not
significantly affect the datasets and it may be valid to include lakes in a combined model
across image dates. Second, I compared the means of the percent concordant values from
the individual models for the September lakes and the August lakes. In this analysis, the

absence of large differences between the September and August percent concordant

53

values may also suggest that it may be valid to include lakes in a combined model
across image dates.

Using the logit model for individual lakes, 1 also examined whether various lake
characteristics can help improve predictions of plant cover using Landsat imagery. Using
ordinary least squares regression, I regressed each of the four model coefficients (TM2,
TM4 and the interaction terms with lake depth for both) from the individual lake logit
models against each of the measured lake characteristics individually: Secchi disk
transparency, water color, chlorophyll a, total phytoplankton biovolume, and sediment
type.

Logit Model Validation

I validated the results of the logit models by using field- collected data for
Winfield Lake, a lake not included in model development. Because the date of data
collection for Winfield Lake fell within the range of dates for the lakes sampled in the
September model, I validated only the September and the combined logit models. The
validation was done by calculating the logit values for each sample point in Winfield
Lake from the logistic (for plant cover) and bionomial (for all other plant variables)
regression equations for both the September and combined models. The logit values
represented the cumulative probability of each sample point being in each plant cover
level (0, l, 2, and 3). The cumulative probability value of the logit was then used to
calculate the actual probability of each sample point being in each plant cover level. The
actual probabilities were then averaged to determine the overall probability of sample

points belonging in plant cover level for each plant category.

54

(I)

Results

The 13 study lakes (including the validation lake) had some sites with each of the
four plant categories, although there was a fair amount of variation across all lakes as
well as among categories (Figure 3.3). In particular, few lakes had very dense amounts
of emergent and floating leaf vegetation compared to submersed macrophytes. In
addition, although EWM was present in relatively low amounts compared to total
submerged plants, it was present in all lakes. The overall percent cover of plants in the
study lakes ranged from 5% to 42%, and the percent cover of plants in the littoral zone
was fairly high in most lakes (Figure 3.4). Four lakes had 5 50% plant cover in the
littoral zone; so in these lakes, the littoral zone was dominated by exposed lake sediments
rather than plants. Although the presence of plants was quite high in all lakes, the
‘density’ (or cover) of plants at each sample point was quite low (Figure 3.5). The
majority of sample points had only 0-20% plant cover (level 0) at a site, which led to
skewed distributions of plant levels across the four categories. The total submersed plant
category was the most evenly distributed category.

The highest concordance values for predicting the plant cover measures from
Landsat DN values using the logit models were for the floating leaf and emergent (74-
89%) plant categories (Table 3.2). Lower concordance values were found for the two
submersed categories, total submersed (62-71%) and submersed EWM (39-62%). These
results suggest it may be easier to predict plant categories that are above the water surface
than below the water surface. Interestingly, when all plant categories were aggregated
into a single category (littoral plant cover), the logit models had relatively high percent

concordance (65-72%). The p-values for the Wald test were highly significant (p<0.001)

55

lllt

rel

six

51;

V:

ft

ll

(*1

a1

for all categories except for submersed EWM (p=0.067) in the September model. This
result may be due to the slightly lower amount of plant level 3 in for the lakes in the
September model, compared to the other models (Figure 3.5).

To compare the validity of combining lakes into a combined August-September
model, I developed individual logit models for each. I found no significant differences in
the model coefficients between the August and September lakes, suggesting that the
relationships between the means of the spectral DN values and plant cover estimates are
similar across different images, despite any atmospheric differences between the two
images (Table 3.3). Additionally, the comparisons of the percent concordant values from
the individual lake logit models suggest that the August and September models produced
similar model fits to the data (Figure 3.6). Given the above results, I assumed that there
are no statistically significant differences between the two images and therefore, it is
valid to combine all lakes into a single ‘combined’ model, even though data for the
spectral DN values come from two different images from two different dates.

The ordinary least squares regressions of the logit model coefficients estimated
for the 12 individual lake datasets versus lake characteristics produced very low r2 values
and insignificant p-values for the majority of the variables (Table 3.4). Total
phytoplankton biovolume produced the highest r2 values and four of the plant categories
were significant at the 0.10 level (littoral plant cover, emergent, total submersed,
submersed EWM; Table 3.4). This result suggests that knowing phytoplankton
biovolume may be helpful to improve predictions of plant cover in a lake using Landsat.
However, although some relationships were significant, this result was not consistent for

all independent variables tested and the significant relationships were not strong (r2 =

56

0.25 — 0.42). Because I found it surprising that these lake characteristics did not appear
to improve model predictions, I examined how well I could predict these lake
characteristics from Landsat DN values from the pelagic zone. I examined the pelagic
zone because variables such as chlorophyll a, Secchi disk transparency, phytoplankton
biovolume, and water color are better assessed in the pelagic zone than in the littoral
zone. I found significant correlations between all variables and spectral DN values of the
pelagic zone except for water color (Table 3.5). Thus. I was able to detect differences in
these variables across the 12 lakes sampled. However the r:2 values were only moderately
high for the significant variables (r2 = 0.56 — 0.65) and perhaps there was too much
variation in these relationships to help predict plant abundance.

I validated the September and the combined models using field-collected data
from Winfield Lake. The overall plant characteristics (Figures 3.3-3.4) and the
distribution of plant cover levels in Winfield Lake (Figure 3.7) were similar to the
distributions found in the 12 lakes used to develop the models. Skewed distributions of
plant levels across the four categories also occurred in Winfield Lakes because the
majority of sample points had only 0-20% plant cover at most site sampled. The total
submersed plant category was the most evenly distributed category.

The validation of the September and the combined models showed little
difference (Table 3.6). I found relatively low, yet similar probabilities of correctly
classifying each multilevel plant cover category (emergent, floating leaf, total submersed,
and submersed EWM) and a somewhat higher result for the aggregated category of plant
cover presence or absence (littoral plant cover), although this result is skewed by the fact

that I correctly classified level 0 at a very high probability level, but I did not correctly

be

\tl

C3

SE

llt:

El

fr

classify level 1 at all. However, within each of the multilevel categories, I obtained the
best probability of correct classification from the higher levels of plant cover (level 2 and
3). High classification levels were also seen in the submersed EWM category for level 2,
which was the highest level for this category since a level 3 was not observed for this
category in Winfield Lake. The most evenly distributed range of probabilities occurred
in the total submersed category with probabilities of classification ranging from 0.17 —

0.31 for the September model, and 0.18 — 0.35 for the combined model.

Discussion and Conclusions

I found relatively strong relationships between the five plant categories and
Landsat spectral TM values using logistic regression models. For each plant cover type, I
obtained significant model fits and reasonably high percent concordant values. All
models showed the highest percent concordant values for the plant categories with plants
above water: floating leaf and emergent plants. Lower percent concordant values were
seen for the submersed categories (total submersed and submersed milfoil). This result is
not surprising, considering the difficulties inherent in remotely sensing the aquatic
environment (Lillesand and Kiefer 1994, Verbyla 1995). However, the fact that the
models resulted in percent concordant values for the two submersed categories ranging
from 39 — 71%, suggests that remote sensing may prove to be a valuable tool for
measuring plant distributions in lakes and that further research is necessary to improve
the fit of the models. Interestingly, I did not find an effect of lake characteristics on the
ability to detect aquatic plants in lakes. I found this result surprising given the wide

range of water clarity that was present in the study lakes. However, it is possible that

58

with only 12 lakes for the combined model (and 6 lakes for the single-date models), the
sample size was too low to detect the effects of lake characteristics. Additionally,
although studies by Lathrop (1992), Cox et al. ( 1998), and Kloiber et al. (2000) have
found Landsat to be significantly correlated to a variety of water clarity and water quality
parameters, there may be enough variation around these relationships that it is not
possible to quantify the effect that they have on model predictions of aquatic plants,
which themselves have a significant amount of variation as well. Although the
regressions of three of the four lake characteristics against Landsat spectral values were
significant, the regressions had relatively low r2 values (0.56-0.65; Table 3.5).

The model validation analysis adds additional insights into the use of remote
sensing to detect aquatic vegetation. The validation resulted in low probabilities (0.23-
0.49) of correctly classifying plants in all plant categories. This result was surprising
because the models produced high overall percent concordant values and significant
Wald test statistics, suggesting a good model fit of the data. I can only speculate as to
why the validation and the model statistics resulted in different interpretations. For
example, because the validation analysis was based on only one lake, some feature of the
validation lake that was not present in the model lakes may have influenced the validation
results. However, using the categories and levels of plant cover, the validation lake
seemed very similar to the model lakes. In addition, the models themselves, although
significant, may be sensitive to the distributions of plants in each category. Therefore, if
the plant-cover level distributions of both the model datasets and the validation datasets
had been more equally balanced between levels, I may have found higher probabilities in

the validation results. In fact, low validation probabilities seemed to be related to low

59

numbers of actual observed sites in those categories (Table 3.7). For this study, I
sampled a total of 36 randomly selected lakes. The 13 lakes I was able to incorporate
into the models were not necessarily representative of the overall 36 lakes. For example,
9 out of the 36 lakes had total plant cover > 42%. However, no clear images were
available for these higher plant cover lakes. Therefore, additional clear images, would
have increased the number of lakes included in the models, and may have allowed the
distributions to be more evenly distributed. Finally, it is possible that I did not
experience adequate positional accuracy with the GPS unit on the days I sampled either
any of the model lakes or of the validation lake. Although greater GPS accuracy is
currently possible now that selective availability has been turned off, horizontal accuracy
of hand-held GPS units is still from 7— 15m depending on environmental and satellite
signal reception conditions. Perhaps sampling at an even finer grid resolution in the field
(< 40 m) would solve the potential problem of sample point/pixel misalignment.

It is commonly accepted that values sampled from images collected on different
dates should not be combined because they are affected by different atmospheric effects
(such as haze) specific to the date of image acquisition (Song, 2001). However, I did not
find this to be the case. In fact, any atmospheric differences produced no statistically
significant differences between the datasets. This conclusion is also evident in the
validation analysis. The application of the validation data to models developed based on
the coefficients of the both the September and combined models produced very similar
results for all plant cover categories and category levels. This result is important because
it allowed us to increase the sample size by including all lakes in a combined model;

which also allowed us to examine the effect of lake characteristics on the models.

60

Although I sampled 36 lakes for this study, I was only able to find clear images for 12
lakes, thus being able to use images from multiple dates will increase the likelihood in
future studies that clear images and field data can be matched.

The use of remote sensing in freshwater lake studies can play a vital role in
reducing the cost, labor, and time required to monitor these systems over a large
geographic area. Remote sensing also has the potential to be used as a tool for statewide
assessments that are currently impossible using traditional field operations. Here, I have
provided a general method for detecting littoral zone plant cover in inland freshwater
lakes using satellite imagery. However, the attempts to predict plant categories in
unsampled lakes (i.e. validation analysis) produced varying results, despite the relatively
good fit of the models.

I conclude that more research is necessary to conclusively assess the utility of
using Landsat TM data to predict macrophyte cover in lakes. I offer the following
suggestions. First, one option would be to include more sample lakes into the model
calibration to provide representative distributions of regional plant cover within the lakes
modeled. A second possibility is to use a finer spatial resolution of the field-collected
data to match more exactly the spatial resolution of Landsat data. Third, perhaps a
different method of assigning plant levels could be developed. The method of sampling
macrophytes and assigning category levels was relatively coarse, which may have
influenced to the ability to detect plants using Landsat data. However, more detailed
plant sampling methods, such as linear shoreline transects, are often very limited in
spatial scale, which would also be difficult to relate to remotely sensed data. Fourth,

current sensors such as Landsat, may be limited in providing the appropriate spectral

61

resolution necessary to more effectively measure macrophyte distributions and the water
quality parameters that affect these distributions. We need to determine the best spectral
band configurations of sensors to more accurately measure aquatic plants. Until our
methods can be improved or possibly better sensors are developed for aquatic
applications, inland lake remote sensing may be most useful as a supplement to existing

volunteer and agency monitoring programs, rather than a replacement of these programs.

Acknowledgements

This research was funded in part by a Michigan State University Minority
Competitive Doctoral Fellowship to SACN and an EPA STAR Fellowship to KSC.
Landsat-5 ETM+ imagery was provided by the Basic Science and Remote Sensing
Initiative (BSRSI, now known as the Center for Global Change and Earth Observation)
and through the NASA-funded Upper Midwest RESAC project (Regional Earth Science
Applications Center). I thank Kevin Hoffman, John Whitlock, Tamara Brunger-Lipsey,
Kristy Rogers, Sarah Wills, and Konrad Kulacki for field data collection and processing.
1 are also grateful to David Skole, Jiaguo Qi, J ianguo Liu, David Lusch, and Walter
Chowmentowski for their contributions to this research. Additionally, I would like to
thank Michael Belligan, Marc Linderman, and Chris Barber for technical support and
database assistance. Finally, I thank Daniel Hayes. Orlando “Ace” Sarnelle, Ken Frank,

Jim Bence, and Emily Szali for statistical advice.

62

Literature Cited

Aiken, S. G., Newroth, P. R., and Wile, I., 1979. The biology of Canadian weeds. 34.
Myriophyllum spicatum L. Canadian Journal of Plant Science. 59, 201-215.

Ackleson S.G., Klemas V., 1987. Remote-sensing of submerged aquatic vegetation in
lower Chesapeake bay: a comparison of Landsat MSS to TM imagery. Remote
Sensing of Environment. 22, 235-248.

Armstrong, R., 1993. Remote sensing of submerged vegetation canopies for biomass
estimation. International Journal of Remote Sensing. 14, 621-627.

Carpenter, SK, and Lodge, D.M., 1986. Effects of submersed macrophytes on
ecosystem processes. Aquatic Botany. 26, 341-370.

Cheruvelil, K.S., Soranno, P.A., Madsen, J.D., and Roberson, M.J., 2002. Plant
architecture and epiphytic macroinvertebrate communities: the role of an exotic
dissected macrophyte. Journal of the North American Benthological Association.
21, 261-277.

Chilton, E. W. 1990. Macroinvertebrate communities associated with three aquatic
macrophytes (Ceratophyllum demersum, Myriophyllum spicatum, and Vallisneria
americana) in Lake Onalaska, Wisconsin. Journal of Freshwater Ecology. 5, 455-
466.

Couch, R., and Nelson, E., 1985. Myriophyllum spicatum in North America. In:
Proceedings of the First International Symposium on Watermilfoil (Myriophyllum
spicatum) and related Haloragaceae species. L. W. J. Anderson editor. Aquatic

Plant Management Society, Washington, DC. 8—18.

63

Dr

161

Kc:

Kle;

K101

Cox, R. M., Forsythe, R. D., Vaughan, G. E., and Olmstead, L. L.. 1998. Assessing
water quality in the Catawba River reservoirs using Landsat Thematic Mapper
satellite data. Lake and Reservoir Management. 14, 405-416.

Crumpton, W.G., T.M. Isenhart, and PD. Mitchell. 1992. Nitrate and organic N analysis
with 2nd-derivative spectroscopy. Limnology and Oceanography 37 (4): 907-913.

Dekker, A.G., and Peters S.W.M., 1993. The use of the Thematic Mapper for the
analysis of eutrophic lakes: a case study in the Netherlands. International Journal

of Remote Sensing. 14, 799-821.

 

Jensen, J.R., Narumanlani, S., Weatherbee, O., Mackey, H.E., Jr., 1993. Measurement of
seasonal and yearly cattail and waterlily changes using multidate SPOT
panchromatic data. Photogrammetric Engineering and Remote Sensing. 52, 31-
36.

Khorram, S., and Cheshire H. M., 1985. Remote sensing of water quality in the Neuse
River Estuary, North Carolina. Photogrammetric Engineering and Remote
Sensing. 51, 329-341.

Keast, A., 1984. The introduced aquatic macrophyte, Myriophyllum spicatum, as habitat
for fish and their invertebrate prey. Canadian Journal of Zoology. 62, 1289-1303.

Kleinbaum, D.G., 1994. Logistic Regression A Self-learning Text. Springer-Verlag
Publishing. New York, NY.

Kloiber, S.M., Anderle, T.H., Brezonik, P.L., Olmanson, L., Bauer, ME. and Brown,
D.A., 2000. Trophic state assessment of lakes in the Twin Cities (Minnesota,
USA) region by satellite imagery. Archive Hydrobiologie Special Issues

Advances in Limnology. 55. 137-151.

64

Kloiber, S. M., Brezonik, P. L., Olmanson, L. G., and Bauer, M. E., 2002. A procedure
for regional lake water clarity assessment using Landsat multispectral data.
Remote Sensing of Environment (in press).

Lathrop, R. G., 1992. Landsat Thematic Mapper monitoring of turbid inland water
quality. Photogrammetric Engineering and Remote Sensing. 58, 465 - 470.

Lathrop, R. G., and Lillesand, T. M., 1986. Utility of Thematic Mapper data to assess
water quality. Photogrammetric Engineering and Remote Sensing, 52, 671-680.

Lehmann, A., and Lachavanne, J.B., 1997. Geographic information systems and remote
sensing in aquatic botany. Aquatic Botany. 58. 195-207.

Lillesand, T. M., Johnson, W.L., Deuell, R. L.. Linstrom, O. M., and Meisner, D. E.,
1983. Use of Landsat data to predict the trophic state of Minnesota lakes.
Photogrammetric Engineering and Remote Sensing. 49, 219 - 229.

Lillesand, T.M., and Kiefer, R.W., 1994. Remote Sensing and Image Interpretation. John
Wiley and Sons, Inc. New York, New York. 430-501.

Lyons, J., 1989. Changes in the Abundance of Small Littoral—zone Fishes in Lake
Mendota, Wisconsin. Canadian Journal of Zoology. 67, 2910-2916.

Madsen, J.D., Eichler. L.W. and Boylen, C.W., 1988. Vegetative spread of Eurasian
watermilfoil in Lake George, New York. Journal Aquatic Plant Management. 26,
47-50.

Madsen, J. D., J. W., Sutherland, J. A. Bloomfield, L. W. Eichler, and C. W. Boylen.,
1991. The decline of native vegetation under dense Eurasian watermilfoil

canopies. Journal of Aquatic Plant Management. 29, 94-99.

65

Madsen, J. D. 1999. Point intercept and line intercept methods for aquatic plant
management. Army Corps of Engineers Waterways Experiment Station Technical

Note MI-02.

Marshall, TR. and Lee, PF, 1994. Mapping aquatic macrophytes through digital image
analysis of aerial photographs: assessment. Journal of Aquatic Plant
Management. 32. 61-66.

Menzel, D. W., and N. Corwin. 1965. The measurement of total phosphorus in seawater
based on the liberation of organically bound fractions by persulfate oxidation.
Limnology and Oceanography 10:280-282.

Murphy, J., and L. P. Riley. 1962. A modified single solution method for the
determination of phosphate in natural waters. Analytica Chimica Acta 27:31-36.

Narumalani, S., Jensen. J.R., Althausen, J.D., Burkhalter, S.G., Mackey, H.E., Jr., 1997.
Aquatic macrophyte modeling using GIS and logistic multiple regression.
Photogrammetric Engineering and Remote Sensing. 63, 41-49.

Nusch, EA, 1980. Comparison of different methods for chlorophyll and phaeopigment
determination. Archives of Hydrobiolia. 14. 14—36.

Orth, R.J. and Moore, K.A., 1983. Submerged vascular plants: techniques for analyzing
their distribution and abundance. Marine Technology Society Journal. 17, 38-52.

Penuelas, J., Gamon, J.A., Griffen, KL, and Field. CB, 1993. Assessing coummunity
type, plant biomass, pigment composition, and photosyntheic efficiency of aquatic

vegetation on spectral reflectance. Remote Sensing of Environment. 46, 1 10-1 18.

66

Rui

SAS

Srhl

 

Schn

Snnfl

SOng

Inbi

Pampel, EC, 2000. Logistic Regression: A Primer. Sage University Papers Series
Quantitative Applications in the Social Sciences. series no. 07-132. Sage
Publications. Thousand Oaks, CA.

Raitala, J., and Lampinen, J ., 1985. A Landsat study of the aquatic vegetation of the lake
Luodonjarvi reservoir, Western Finland. Aquatic Botany. 21, 321-346.

SAS Institute Inc., 1995. Logistic Regression Examples Using SAS System. Version 6,
first edition. SAS Institute Inc. Cary, NC.

Schloesser, D.W., Manny, B.A., Brown, CL, and Jaworski, E., 1987. Use of low-
altitude aerial photography to identify submersed aquatic macrophytes. In: Color
Aerial Photography in the Plant Sciences and Related fields. Proceedings of the
10‘h Biennial Workshop, University of Michigan, Ann Arbor. 19-28.

Schneider, J .C., 2000. Evaluation of the effects of the herbicide Sonar on sport fish
populations in Michigan lakes. Michigan Department of Natural Resources,
Fisheries Technical Report No. 2000-2. 1-1 1.

Smith, C. S., and Barko, J. W., 1990. Ecology of Eurasian watermilfoil. Journal of
Aquatic Plant Management. 28, 55-64.

Song, C., Woodcock, C. E., Seto, K. C., Lenney, M. P., and Macomber, S. A., 2001.
Classification and change detection using Landsat TM data: when and how to
correct atmospheric effects? Remote Sensing of Environment. 75, 230-244.

Trebitz, AS, and Nibbelink, N., 1996. Effect of Pattern of Vegetation Removal on
Growth of Bluegill: A Simple Model. Canadian Journal Fish and Aquatic Science.

53, 1844-1851.

67

 

[Ha

Veil

Zhan

Zfliol

Trebitz, A. S., Nichols. S. A., Carpenter, S. R., and Lathrop, R. C., 1993. Patterns of
vegetation change in Lake Wingra following a Myriophyllum spicatum decline.
Aquatic Botany. 46, 325-340.

Ullah, A., Rundquist, DC, and Derry, DP. 2000. Characterizing spectral signature for
three selected emergent aquatic macrophytes: a controlled experiment. Geocarto
International. 15, 29-39.

Verbyla, D.L., 1995. Satellite Remote Sensing of Natural Resources. CRC Lewis
Publishers, New York, New York. 2-21.

Zhang, X., 1998. On the estimation of biomass of submerged vegetation using Landsat
Thematic Mapper (TM) imagery: a case study of the Honghu Lake, PR China.
International Journal of Remote Sensing. l9, 1 1-20.

Zilioli, E., 2001. Lake water monitoring in Europe by means of remote sensing. Science
of the Total Environment. 268, 1-2.

68

 

re
ha

1m

Lair

CONCLUSIONS

The strength of remote sensing techniques for large-scale ecosystem studies lie in
its ability to provide spatial measurements of features across large areas that are typically
not possible with in situ sampling procedures. In addition, sensors such as Landsat
provide an excellent tool to examine changes over time, since this imagery has been
collected over large areas of the earth's surface for several decades. To date, Landsat and
other such sensors have been used extensively for land-based studies to measure such
things as land use change, leaf area, forest inventories, etc. (Wilkie and Finn 1996).
However, sensors such as these have been less well applied to wetland and inland lake
studies. Because Landsat sensors were primarily designed for the detection of terrestrial
features, the detection of aquatic features has been more problematic for some
applications. For example, the band placement within the Landsat sensor is such that the
band wavelength with the maximum capability of penetrating water (blue, 0.45-0.52 um)
is also the band which is most scattered by the atmosphere, making it extremely sensitive
to atmospheric conditions during image capture. In contrast, the longer wavelength
bands are more absorbed by clear water, thus limiting the amount of spectral energy
returned to the sensor for feature detection. Despite these potential limitations, Landsat
has been used successfully in aquatic studies for the remote sensing of Secchi disk
transparency (SDT), Chl a, and suspended sediments (Khorram and Cheshire 1985,
Lathrop 1992, Lathrop and Lillesand 1986).

However, I believe despite these limitations, remote sensing in general, and

Landsat in particular, can still provide valuable information on aquatic ecosystems such

69

 

H1
mi
Sm
31h

Ella

as wetlands and lakes. In my dissertation, I explore how far we can take the Landsat
sensor in detecting features of aquatic ecosystems that span the range from above water-
measures such as wetland and emergent plant cover, the littoral zone features of lakes
that includes submerged plants, and finally the pelagic zones of lakes that includes water
clarity. I have been interested in exploring what we can and cannot detect using the
Landsat sensor. In this section, I will discuss: (1) some of my significant overall findings
and applications of this research to monitoring and management, (2) some of the
logistical problems that I encountered during this research, and (3) what I believe is

needed for future development of wetland and inland lake remote sensing capabilities.

Major conclusions and applications of this research

In chapter 1, I examined historical changes within the Barataria Basin, LA. This
study was unique for two reasons. First, this study used historical imagery from the
Landsat MSS sensor across a 20-year period. Landsat MSS has been in orbit since 1972
and currently provides our longest available historical record of earth surface remote
sensing, however this historical record has only rarely been examined (Jensen et al. 1995,
Haaek 1996, Munyati 2000). The spatial resolution of MSS data is approximately 80
meters. For smaller-scale studies, this coarse resolution may be a limiting factor.
However, in my investigation of the upper Barataria Basin (200,000 ha), an area of this
magnitude necessitated larger resolution imagery. Additionally, in comparison to
smaller-scaled field studies or aerial surveys. this large-scale analysis using MSS,
allowed me to identify changes in the upper basin that simply may not have been obvious

at a local scale or ground level investigation. Secondly, Landsat MSS allowed me to

70

IE

tlu
C01
E851
dart:
Clea
the 1
Mod

"‘1’ s

assess land cover changes through time within the upper basin. 1 related observed
increases in urban and swamp forest to known drivers of change affecting the basin as a
whole (i.e. land subsidence and water level rise). The bottomland forest land cover
occupied the mid-elevation land cover within the basin. As the swamp forest in the lower
elevations increase, and the human land use in the higher elevations increase, the
bottomland forest is subjected to multiple stressors. As a result, this land cover has
shown a remarkable decline and has the potential to be lost to the system within the next
25 years. The main management implication in this study is that there are major changes
occurring in the more inland regions of this coastal estuary, namely, the loss of
bottomland forest and gains in total wetland area; and these changes have been
previously overlooked because of the focus on the coastal wetland losses, which are some
of the highest losses in the world. Changes in the upper freshwater portion of this
important estuary may have important additional implications for the lower brackish
regions and is an important area for future study.

In my second chapter, I investigated the use of Landsat-7 to detect lake water
clarity using SDT which incorporates the range of lake SDT values found within a region
comparable to 80% of the lower peninsula of Michigan, USA. To my knowledge, an
assessment of water clarity of this scale has never been attempted using remotely sensed
data. In this study, I identified an important consideration necessary for modeling water
clarity on a large scale: models developed for large regions such as this, must incorporate
the full range of lake SDT values found throughout the region in the model calibration.
Models that incorporate this complete range of SDT values, such as the one developed in

my study, may not produce r2 results as high as models that are developed for groups of

71

lakes that don’t incorporate the full regional range of SDT values (Lathrop and Lillesand
1986, Dekker and Peters 1993, Kliober et al. 2000). In addition, my results suggest that
Landsat is better at predicting shallow SDT values than deep SDT values. Although the
model that I developed has a lower r2 than some previously published studies, I believe
that it can still be used my state management agencies for selected purposes. For
example, it could provide managers with a method to inventory shallow SDT lakes
(eutrophic) across the entire state where in situ data is difficult to obtain. In particular,
using the current model produced in my study, state agencies (i.e Michigan Department
of Environmental Quality, MDEQ) would be able to monitor private lakes without
spending resources on these lakes and without requiring the permission of the property
owner. Additionally, using my model to detect eutrophic lakes may also be useful in the
historical analysis of watershed changes that may have caused lake eutrophication to
occur over time. This application is a very attractive use of remote sensing since we can
always obtain historical imagery for lakes we know that have become eutrophic over
time.

In my third chapter, I investigated the use of Landsat-5 to monitor aquatic plants
in inland lakes within the lower peninsula of Michigan. The remote sensing of aquatic
plants in lakes has been less well studied than SDT. Previous studies have focused on
detecting emergent vegetation or homogenous clusters of submersed vegetation
(Ackleson and Klemas, 1987; Armstrong, 1993). Also, many of these studies have been
limited to few or individual macrophytes species and restricted to a single water body
(Penuelas et al. 1993, Lehmann et al., 1997, Ullah et al., 2000). My study demonstrates

that multiple types of plant cover, both emergent and submersed, can be detected using

111

L0

remote sensing. I developed a model to detect five different types of macrophyte cover
that was calibrated using twelve different lakes across a large region of the state of
Michigan. The models developed in my study fit the calibration data very well and
resulted in statistically significant model fits. However, when I tried to validate the model
using a single lake, the outcomes were disappointing. This suggested further research
might be necessary to validate these models, since perhaps there was something unique
about our single validation lake. In the immediate future, I will explore other options to
validate this model. For example, I will attempt to validate the complete model, which
has a larger number of total lakes (N: l 3). I will randomly choose nine lakes to calibrate
the model and four lakes to validate the model. This would reduce the total number of
lakes in the complete model dataset. However, because the validation analysis would be
based on more than one lake, it would reduce the influence of features of the lakes used
for validation that may not possibly be present in the model lakes, and vice versa. In
general, the significant value of this research is that it lays the foundation for more
research to improve on our ability to detect submerged plants in lakes. Further
development of these models will benefit lake managers interested in assessing several
lakes within a large region for the monitoring and assessment of aquatic plants and

nuisance species.

Logistical issues involved in this research
There are several logistical issues in using remote sensing for aquatic studies that
I found throughout my research. For example, in chapter 1, I found the storage and

retrieval of the large Landsat MSS image files to be one of the most limiting factors.

73

 

(55

ea
[tel
hilt:
be”.

man;

Each Landsat MSS image required approximately 340 megabytes of hard drive storage
space. This not only made file storage a problem, but also combined with limited CPU
processing power, file manipulation and processing in ERDAS Imagine was very slow.
Another problem I found involved the use of high-resolution digital aerial photographs
for classification accuracy assessment. Approximately 1 15 individual air photos had to

be mosaicked and georectified to the 1992 MSS scene. This was a time consuming

. . . . r“
procedure because detailed ground maps that matched the spatial resolution of the arr
photos (3 meters) were not available. Thus, the each photo required manual “stitching”
in ERDAS until a larger mosaicked scene was created for overlay and georectifying with ;
a.

 

the MSS imagery. A final problem area in this study occurred in the preprocessing of the
imagery itself. Linear spectral “bandings” were very visible in the 1972 image data.
Further observation revealed the same spectral distortions were also present in the 1985
and 1992 imagery, however not as prominent as in the 1972 image. After performing a
series of supervised classifications without correcting these distortions, I found that the
accurate separation between the bottomland forest and the swamp forest land cover was
impossible to achieve, even with numerous training samples selected from the
corresponding land cover. By using a band ratio, developed by dividing MSS band 1 by
MSS band 2, the scene distortion was reduced and allowed for accurate classification of
each land cover type. Some of the above issues will be reduced as newer hardware
technology and software advances become available. For instance, personal computers
have more than tripled in processing power and storage capabilities, almost daily, since I
began this study. Additionally, software improvements have also produced easier

management of larger files in the ERDAS Imagine program. However, historical data

74

obtained from these sensors in the earlier 1970’s to mid-1980’s still remain archived on
magnetic tape throughout many datacenters in the US. This media has a tendency to
begin oxidizing after 20 years or so, resulting in partial or total loss of the stored data.
This could severely limit our ability in developing historical land cover change studies.
In Chapter 2, one of the greatest challenges to this research was locating cloud-
free Landsat TM-7 images for a large number of lakes throughout the lower peninsula of
Michigan. I was only able to obtain a total of six images which had no or limited cloud
cover over lake within the lower peninsula of Michigan, for the sample period of mid-
July to mid-September. The three path 21 images, used in chapter 2, proved to have the
least amount of could cover allowing for cloud-free extraction of the pelagic spectral
values of the 93 lakes used in this study. Additionally, during the summer of 2001, our
sampling program sampled 36 randomly stratified lakes, however only one lake had
available cloud-free imagery and was within +/- 7 days of the satellite overflight.
Interestingly, I found that SDT data from the Michigan Citizens Lake Monitoring
Program (CLMP) provided the most usable data source because the volunteers sampled
most lakes weekly or biweekly. I received a dataset of 195 lakes from CLMP and was
able to incorporate 79 lakes within +/- 7 days of the satellite overflight. However, one
limitation of using CLMP data was that no geographic coordinate information is provided
with this data, other than county, or sometimes township names, in which the lake was
located. Of the total CLMP 195 lakes, 42 lakes had to be discarded because accurate
identification of the lake could not be made as a result of incorrect county or township
names, or multiple lakes within the county or township bearing the same name. The

above limitations maybe overcome by continued used of the CLMP data, as volunteers

75

are collectively able to sample a large number of lakes throughout the state and
throughout the summer season. This frequency will allow for lake samples to be
obtained without the need of synchronizing sampling operations will cloud-free satellite
acquisition periods. The collective volunteer data will inevitably contain dates that meet
the 7 day criteria necessary for detecting SDT. Additionally, I would recommend that the
volunteers be provided with handheld GPS units, which would allow researchers to
accurately identify the sampled lake. These units have become relatively inexpensive
over the last few years and, in states with thousands of lakes, it may actually be more
economical for the agencies to supply GPS units and training to volunteers than to
conduct an agency statewide inventory of lakes themselves.

In chapter 2, I also explored a combined regression model of all available SDT
lake data (+/- 56 days of the satellite overflight) from lakes within the three adjacent
ground track paths in the lower peninsula of Michigan (paths 20, 21, and 22). This
regression model yielded an r2=0.05 (p<0.01, n=142). I then explored a combined
regression model based on all available SDT lake data. +/- 27 days of the satellite
overflight, from lakes within path 21. This regression model yielded an r220.27 (p<0.01,
n=104). Reducing this model to 93 lakes, within path 21, and within +/- 7 days of the
satellite overflight produced my highest r2 result (r3=0.43, p<0.01) using the largest
number of lakes possible. I did not atmospherically correct the lake images because the
spectral values extracted for each lake were so low that the likelihood of finding a lower
dark object (in which to base a dark body subtraction procedure) within each image was

highly unlikely.

76

One possible source of error in the model developed in chapter 2 is the use of
‘averaged’ spectral DN values to represent the pelagic zone. I do not believe that this is a
large source of error however. First, it is similar to the protocol that has been developed
and used in other state programs (Minnesota RESAC project; Kloiber et al. 2000). I
tested the validity of this approach by examining the variation in pelagic spectral DN
values in the pelagic zone and the variation was found to be very low. Additionally, using
the averaged value would have removed the chance that individual pixels may have been
spuriously elevated from a spurious bottom reflection or dense plant reflectance from a
shoal area occurring within the lake’s pelagic region.

In my final chapter (chapter 3), again the greatest logistical challenge was finding
cloud- free satellite images. During the late summer of 2001, 36 lakes were sampled for
macrophyte abundance. However, I was only able to obtain cloud free imagery for 13 of
these lakes using Landsat-5 TM. The days between lake sampling and image acquisition
were 5-l8(+/—). Using Landsat—7 ETM+ I would have been limited to one sample lake
within 7 days of overflight and 2 lakes within 27 days of overflight. An additional
concern was that the TM band 2 values in the August dataset were slightly higher than
the TM2 values in the September dataset, which led me to believe that this dataset was
affected by atmospheric haze. Several models have been developed for the correction of
atmospheric effects for improved image interpretations (Chavez 1988, Campbell and Ran
1993). However, there still exists a large discrepancy in the literature concerning a
universal model, or how one (or several) correction models should be used (Brivio et al.
2001, Rahman et al. 2001, Song et al. 2001). To correct for haze, I attempted an image-

based method of atmospherically correcting for haze by regressing the visible bands

77

 

(TMl, TM2, TM3) each against the TM4 band. In the absences haze scattering, the
regression Y-intercepts should pass through the X-Y origin (Wilkie and Firm 1996). The
visible bands in both the August and September datasets were adjusted by the value at
which the regression line crossed the Y-axis and the origin (0). The adjusted TM values
were then tested in the Secchi model developed in chapter 2 using pelagic TM values
(TMlzTM3). I felt this was a sufficient test since the SDT model had been established
within the literature. The Y-intercept model was then compared to an uncorrected Secchi '
model using the lakes from both the August and September datasets. I found that the Y-

intercept adjusted models failed to improve the regression results when compared to the

 

uncorrected models. This may indicate that other effects that are not influenced by
atmospheric conditions elevate the TM values in these lakes. My further statistical
analyses (two sampled t-test) support this conclusion, which showed no statistically
significant difference between the mean TM coefficient values produced from the 12
individual lakes.

One additional problem in chapter 3 occurred in determining the “best” bands or
band combinations to use in developing my logistic regression models. I explored
several series of binomial and logistic stepwise regression models for 42 individual and
ratioed band combinations for all datasets. These analyses produced varying results.
Actually, several bands appeared to be very similar. Box plots for each band revealed
few differences across the different band combinations. However, several bands were
still identified as significant variables. Additionally, box plot failed to reveal any large
degree of separation between bands. Thus, we selected bands TM2 and TM4, which

appeared to be the most consistent significant bands when tested against all plant cover

78

types and levels of plant abundance. Additionally, my decision to select 2 bands versus

several, allowed for the development of a simpler model with fewer parameters to fit.

Future use of remote sensing for freshwater studies

Currently, Landsat provides one of our richest data sources for the remote sensing
of historical changes in large areas of wetlands, water quality and aquatic plants. This is
evident by the high revisit cycle, spatial resolutions, and 30 years of operation provided
by these sensors. Based on my research, and the research of others, it appears that the
spectral resolution of these sensors (4 in MSS and 6 in TM) currently allow us to identify
broad vegetation land cover types in wetland areas, SDT estimates (at least for shallow
SDT lakes), emergent and floating aquatic plants, and past and present rates of change of
each. However, there are features that Landsat has problems in detecting. For example,
individual plant species within highly mixed vegetation patches or within patches which
have an areal extent below the spatial resolution of the sensor. To improve these
problems, sensors with finer spectral and spatial band resolutions may be necessary to
fine-tune these models for improved feature detection within the aquatic environment.
However, Landsat-7 data now has the capability of being ordered with adjusted high or
low gain settings. The low gain setting is what has traditionally been used by the Landsat
sensor series and is optimized for highly reflective features, such terrestrial vegetation.
However, the new option of obtaining the high gain settings may have the potential for
improving aquatic feature detection. I may also explore this capability of Landsat-7 in

the near future.

79

 

Finally, I will conclude by posing my thoughts on the future prospects of remote
sensing in wetland and inland lake studies. Given some of the limitations I have detailed
above I think that new sensors for aquatic studies should include the following
specifications. First, it would be helpful to have self calibrating sensors with lowered or
compensated calibrations for water absorption in IR bands and lowered or compensated
atmospheric scattering in the visible bands because these are the most problematic
regions for aquatic remote sensing. Additionally, hand-held or boom-held, multi-band, F
hyperspectral sensor systems have become popular for ecological studies within the last

10 years. However, only a few studies have applied these systems to the aquatic

 

I."'"IQJ.:

environment (Penuelas et a1 1993, Ullah et a1. 2000). Also, because these sensors have
such fine spatial resolutions, which depends solely on the height on the sensor above the
feature of interest, they cannot be used across large regions as I have used Landsat in my
research. Additionally, the imagery cost is still prohibitive for most large scale research
applications. For these hyperspectral sensors to become widely usable for regional
wetland and inland lake studies, I suggest a configuration similar to the Landsat orbit,
revisit period, and spatial resolution. More specifically, I would like to see a sensor
system with incremental narrow band placements in the near-IR where water absorption
lessens and vegetative reflection begins to slightly increase, as well a several incremental
narrow bands within the green and red portions of the visible spectrum, as these regions
are known to strongly match vegetation and chlorophyll reflectance peaks. Finally, it
would helpful to have a multi-satellite constellation systems, real-time acquisition

systems, and user-definable/employable micro-satellites which would be directly geared

80

towards improving the capability of remote sensing for monitoring small and large

areas, wetland systems, inland waters, as well as global landscape change.

Literature Cited

Ackleson S.G., Klemas V., 1987. Rernote-sensing of submerged aquatic vegetation in
lower Chesapeake bay: a comparison of Landsat MSS to TM imagery. Remote
Sensing of Environment. 22, 235-248.

Armstrong, R., 1993. Remote sensing of submerged vegetation canopies for biomass
estimation. International Journal of Remote Sensing. 14, 621-627.

Brivio, P.A., C. Giardino, E. Zilioli. 2001. Validation of satellite data for quality
assurance in lake monitoring applications. Science of the Total Environment.
26823-18.

Cambell, J .B. and L. Ran. 1993. CHROM: A C Program to evaluate the application of the
dark object subtraction technique to digital remote sensing data. Computers and
Geosciences. 19: 1475-1499.

Chavez, P.S. An improved dark-object subtraction technique for atmospheric scattering
correction of multispectral data. Remote sensing of environment 24:459-479.

Dekker, A.G., and Peters S.W.M. 1993. The use of the Thematic Mapper for the analysis
of eutrophic lakes: a case study in the Netherlands. International Journal of
Remote Sensing, 14(5):799-821.

Haaek, B. 1996. Monitoring wetland changes with remote sensing: an east African

example. Environmental Management 20(3):41 1-419.

81

lensc

 

Khorr

Kloibc‘

Lathro

Luthrc

Penue

Jensen, J.R., Rutchey, K, Koch, MS, and Narumalani, S. 1995. Inland wetland change
detection in the Everglades Water Conservation Area 2A using a time series of
normalized remotely sensed data. Photogrammetric Engineering and Remote
Sensing 61(2):199-209.

Khorram, S. and Cheshire H. M. 1985. Remote sensing of water quality in the Neuse
River Estuary. North Carolina. Photogrammetric Engineering and Remote
Sensing, 51(3):329-34l.

Kloiber, S.M., Anderle, T.H., Brezonik, P.L.. Olmanson, L., Bauer, ME. and Brown,
DA. 2000. Trophic state assessment of lakes in the Twin Cities (Minnesota,
USA) region by satellite imagery. Archive Hydrobiologie Special Issues
Advances in Limnology. 55:137-151.

Lathrop, R. G. 1992. Landsat Thematic Mapper monitoring of turbid inland water
quality. Photogrammetric Engineering and Remote Sensing, 58(4):465 - 470.

Lathrop, R. G., & Lillesand, T. M. 1986. Utility of Thematic Mapper data to assess water

quality. Photogrammetric Engineering and Remote Sensing, 52:671-680.

Lehmann, A., and Lachavanne, J.B., 1997. Geographic information systems and remote
sensing in aquatic botany. Aquatic Botany. 58. 195-207.

Munyati, C. 2000. Wetland change detection on the Kafue Flats, Zambia, by
classification of a multitemporal remote sensing image dataset. International
Journal ofRemote Sensing 2 l (9): 1787-1806.

Penuelas, J., Gamon, J.A., Griffen, K.L., and Field, CB, 1993. Assessing coummunity
type, plant biomass, pigment composition, and photosyntheic efficiency of aquatic

vegetation on spectral reflectance. Remote Sensing of Environment. 46, 1 10-1 18.

82

Rah

Song

111::

Will

Rahman, H. 2001. Influence of atmospheric correction on the estimation of biophysical
parameters of crop canopy using satellite remote sensing. International Journal of
Remote Sensing. 22(7):1245-1268.

Song, C., C.E. Woodcock, KC. Seto, M.P. Lenney, and SA. Macomber. 2001.
Classification and change attention using Landsat TM data: when and how to
correct atmospheric effects? Remote Sensing of Environment. 75:230-244.

Ullah, A., Rundquist, DC, and Derry, DP, 2000. Characterizing spectral signature for r"
three selected emergent aquatic macrophytes: a controlled experiment. Geocarto

International. 15. 29-39. :1

 

Wilkie, D.S. and Firm, J.T. 1996. Remote sensing imagery for natural resources

monitoring: a guide for first-time users. Columbia University Press. Pg 295.

 

APPENDICES

84

moSmE wee $3.8. _ SEED

_ Oahu/mime.

85

l 1 u . C cc Illitvt‘

OHF—UJI. UJ’FhJ‘v- ‘ App-Avmv 2 s. —V v . l I I. I 0. -AV— .u-.'- .f’.“ U n .V- - V “3 VA -A Vic. -1A-~‘AVJ ‘ V x O I n a
c n. I. ~ ~ F J . . f I 1 I It i . v t f l A ‘ .L --§ .AV_~V~ I V A ~'
I f t u .H _ i— 1 ~ ~H OH 1. nt‘. ~ Nyla~ ‘~ ~ ~ n I s

NI

 

5.2: 533505 m 3.88 mEaBm N 328 ecu—Eozom _

 

 

8m 2 _ _ a: R me E .33
S : o _ o o o 55$
5.5.:

3 o 3 _ o o m 533:8;
m: o o 2: 4 o _ :28 9.35
328

E _ o o «a c _ 32588,.

R o o o c S m 2.5.3:?

me o o o _ _ G 52.5

33H. 5855 n :95: «an—«3m _ $981.— o..5.:u_..w< 53.5

.U—OD Gm Ecowwmfi on“ :0 0.8 HD>OU U=m_ Sumo uOh— £0me UDwfimwN—U ch—uovboo wO uDDEflfi OFF .m00~ HODOHUO ON SOC

05% 00:28“: 350 SE. Ham—3:3 3:288 Loam“: 05 8 consensus—o .82 2: 50382 :oflSmEoo e8 #:ng 3.2m .43 .~ meK.

86

£233 3:: «Samoa—aim GOES—37.0.. ..D:%.£ 0:. O~ n..e......a0_..~..zx.:.u N130— ...z: 2003....15 z...3v.._.:wi_:©.u E: >.U.....:..fl:.. 2..E....u.c.\.;.v.:>.\ .2. ~ «:45?

 

 

 

e a «an EN emn .38.
e R : S m. 33$
ea o2 _ _ E _ _ 5.5:. 33.5.8.5
as as 2: m: 2: ate 95%
as mm mm E R :28 Engage
e we a. E a. 23.823.
as on is me am 55.5
GE .8852... 8230 nun—=52 £33. commode—U £85. 2.on 8:233—

.wwd co ease

imam :Eo>o an £3, and _ m we; EoEmmommm £5 :5: moosuoa .3232” cosmoEmmfio :Eo>o 2:. .530 use. some 5L uoEmmm—o

388.50 Evoqu 2: E 33 >oE=uu< .288. SE. 3%ch 2: :82: a5 connoEmmm—o So E £82m .8 copes: 2: 2 8250 HonEsc

of. .258 cogocmmmflo Noe So :8: eo>oo ecu. some 5 282m Co 59:5: 2: 8a 239 voEmma—o 2t. .8on 8:882 SE. SE53

2.: :F—uma floats—06m HO>OU flan: SUMO Sm 20%?” .«O .5963: 0:“ O» muons“ GED—00 m—wHOu DCOUm Quay—Duo.— Ofirfi .ocoom Racy—0mm.— AcomHS—OmOH

E00 3:. 3.855 coca—88 82$: 2: 9 2235683 N02 2t 5953 381368 8% .8838 anagrams—U MS.— 033.

87

 

 

was- $3.- om- atm- mcm- 85.- see: Basie:

So 32: B... S; E w $2: :28 95%

E- 23 _ - ME- 83. E- was - see gasses.

o8- seem _ - as- $3.. 08- $3. 8333.?

8; 22: ea 23 as 08.2 53.5
2%.: as 2%.: $5 3a.: 35
mung—U .3 33— no.2 owes—5 he 83— 3.3. owaano we 83— 3.2

N832: Has... 22.me mans-E

.550 9:2 5 meme-o— vum 33:5: gamma: 28 33 830 new. 5 mEmw 8m 3025:: vision ..Cowofio 525 us: some .8 new 2::

some 5933 £3.23 omcmno $50 was we 28 2: use 35 no: 2:: some amnion omega :58 was. we was :38 of. .m._ 2an

88

 

58:: 83323:”: n :85: 953m m 20:8 228033: _

 

 

$3: 826 33 30.: 8 R Maw-.32
£33 :m 935% Q? N: Sm N955
£3... 02 5.: : 3:38 $3 $0.: $83,:
”3.: a :8 $0.: 38.3 8:: 23.823
33: w 2: No: 33 8va :25 3:
13°F Nba— nun—mun: «QEQ3W :umohch 9:515th< GNP—D
mag:

do; 2:: 52:8 2: E mac-a em: :5: on: 3:823: 5:28 :30: of- .88553: E

8a owcazo o: :0 mug: of- .mwfl :5: N8. :836: :8: on: c: :38 9:2 0:0 80:: @8528 9:2 :0 35 «Ba :30: 2E- .<m.: 2an

89

 

 

 

33: 88.8 20.: so 0 0 E20:

05:. 03.: 6.00.8 0%.: 8 E- «0:55

0mm...»- 00: 03.: $on 0 mm:- ovo 083-:

8.0... m0 8: £3 a8: 02.: : 25:30?

03.00 :2 00.: Km 8 : d 200.: 0 53:: £0:
:55 as:

$12 300.: $3 00: .m S E. «is:

$5: 0? Awe-E 9:0 mm 2 m «050%

3.0.3 :2 $00: $0.: 0 0%.: 003 088:

8.: a SN 88 8005 E00: 2:280?

50.: 0:- 3: 0% 02.0 8000: :25 NS:
130,—- N3— 0:802 ~0E03m $080k 0:::_=0_:w< :08:

~00:

.00; 0E: :0::00 0:: E 000:0 00: 0:0: 0:: 3:000:00: 5:28 :30: 0F:- .m0m0:::0:00 E 0:0 0m:0:0 0: :0 000:0 0f- .82

0:0 30: 803:0: 0:0 .80: 0:0 N3: 803:0: :x0: 0:: 0: :0>00 0:0: 0:0 ::0:-: 00::0>:8 0:0: :0 05 00:0 .90: 0::- .mm.: 030-:-

90

 

:m:0:: 88350:”: m 50:0: QE03m N 30:0: 0:0_::0::0m _

91

 

:30E 8:03:85 m :m0:0: 9:026 m :85: 0:0_::0::0m _

 

 

0:: 33 mm w _v _v E30:
0:: _ $0: a :v : 0:503:
00: : mm :5 n m 0020:
2: :v : m A : 0: mm 05:82}:
0:: :v : : mm G0: :25 NE:
00:0-F ~23 0:80;: «9:0»; $00.50: 0:=:_:0_:w< 53.5
00::

.8; 0E: :0::00 0:: E 000:0 03: 0:0: 0:: 388:8: :::=_8

00:0: 0: :- 8:05:20: E 0:0 0w:0:0 0: :0 m00:< 6:3 3 £0: E 0:0: :0::: 0: 00::0>:8 003 :0:: N3: E 8000:: 0:0: 3528800 :0

::00:0: 0:: 03:58 :0”: .30: 0:0 m3: 803:0: :08: 0:: 0: :0>00 0:0: 0:0 80:: 008800 0:0: :0 3:00:08 00:0 0: E .31 030,:-

92

 

 

0:: m 60: 0 :v _ «050%

00: : 00 am: : m :88:

2: :v : \- A mm: 0m 0:330:84:

Ax: : : m m m 50: 53:3 mwa—
130,—. mwafi

00: :0: 00 0: :v :v m:30:

2: : 30: m :v : 09:03m

00: : 00 :00: 0 0 08.8:

2: :v :v m @m: ow 0:530:84:

Ax: : V : N mm 300: 53:: Nham
:30:- NS: 0:0-:02 N9:03m $88,: 0::—8:84 8:: D

Naafi

d0?- 0:::: 8::00 0:: E 000:0 03: 0:0: 0:: 388:8: :E28 :0:0: 0: :- .88:::0:0: E:::3 0:0 0w:0:0 0: :0 000:0 0: :- .0000 3

Nom— Cm UGG— amps: O: GOP—02:00 mag an: th— Cm :CDmOHQ UCN— Qua—Sotwm MO acoouom on: .EQEGXO :On— .Nm©~ O: mwam USN NBQ— .N00—

0:0 39: 803:0: 0:0 .32 0:0 m3: 803:0: :8: 0:: 0: :0>00 08— 0:0 50:: 00:83:00 0:0: :0 9:00:03 00:0 0:::k .mv. : 0305

93

50:00: 6:03:02”: 0 30:0: 08026 0 :00:0.: 0:0_E0::0m _

 

E:

85

o:

:

_v :v M40:0:

94

20:00: 6:03:08": 0 :m0:0: aE03m N 30:0: 0:0—E033: _

 

 

0 6 m _ o o 2:02

:0 o as m o o ”050%

mu 0 : a: N : 0.88,:

M11.

00 c o _ G: o 25.823

0 o o o m 6 53.5 NS:

139—. Nhaw nflmhflz «QEa3m amuhcrm cyan—flumhw< flank:
m00_

00005500 E 0:0 005050 0: :0 000:0 0::. .005

0::: :0::00 0;: E 000:0 :0>00 0:0: 05: 3:000:00: 5:200 :08: 0::. .30: 0: N3: 50:: 0:0: 000:: 0: 0:=:_=0::m0 80:: 00:02:00 003

00:0 5000 :0:0: 0:: :0 000 .0EE0x0 :0": 00:03:00 :050 0::: 00000050 :05 0:023 0 00 5000 0;: E :0>00 0:0: £000 :0 0000:0080: :05:

0;: 30% 20.: 0000050 .30: 2:0 «8— :0033 :00: 0::: 0: :0>00 0:0: 0:0 60:: :00::0> :00 0:0: :0 3:00:00: 00:0 0::. (6.: 0305

0: o o 6 o o 008:8:

 

 

:0 o o _ a: 0 23.80.00

0: o o _ 0 a: 53.5 000:
.050 000:

0 6 m _ o o 2802

:0 o as N o c "05030

00 o N: 2 c _ _ n:83:

00 o o _ 3: 0. 82.3000

0 o o o 0 6 :25 NS:
180,—. N53 0:050: «9:035 _:00:0,m 0:=:_=0::w< 000:3

~00:

000050800 E 0:0 0wc0n0 0: :0 000:0 0::. .080 08: :0::00 0::: E 000:0 :0>00 0:0:
0::: 0:500:00: 5:200 :08: 0.; .30: 0: who: 60:: 0:0: 53:: 0: 0:::_:0_:w0 80:: 02:02:00 003 00:0 E000 :0:0: 0::: :0 00:: .0EE0x0

:0”: .00_:0w0:00 :050 BE 00wc0E0 :05 20:3 0 00 E000 0::: E :0>00 000: £000 :0 0w0:=00:0m :0:0: 0::: 30:0 0:0: 00050.5 .80:

0:0 32 00:03:00: 000 .mmo: 000 m8: 503:0: :8: 0::: 0: :0>00 0:0: 0:0 80:: 00:03:00 :50: :0 3:00:00: 00:0 0:: .mm: 030.:

96

 

:0:0E 5:03:02”: 0 :00:0: 9:026 N :00:0: 0:0—E033: _

 

mm

6

8m:

0

o c .0802

o c «9.03%

97

....._.. ..............::_........-. .. '7'

 

 

 

 

 

 

 

 

f) '1

Mean Elevation ; Louisiana
0 - 1 meters 2
‘ "”"‘fi"""3
E -~ 2 - 3 meters }
~--: {3‘- - - f
- 4 - 10 meters “kw. . 10‘s ..

 

 

STUDY ARE: " K
Swamp forest! ‘\
freshwater marsh zone

Barataria Basin

 
   
 

Brackish \
marsh zone ‘,

Saltwater coastal \
,1
marsh zone K Gulf of
{‘7

Mexico

3 i2 fig g2 80 Kilometers

Figure 1.1.

98

 

 

Urban

Agriculture
Bottomland forest
Swamp forest
Freshwater marsh

Water

 

 

 

 

20 40 Kilometers

N
O
O

5
m

S

Figure 1.2.

 

99

 

 

 

 

 

 

I 1972

 

 

.......
ii... ..
.333}.
. .3»:
r . .

E u t . ...\

fr... .. .01 0%., ..
w..........a...... . rot.
4V... . 03...... . w
, . . .l..r7~;{.!

 

 

50

_ _ :
0 O O
4 3 2

E020:

Figure 1.3.

100

 

 

 

I Urban
E] Agriculture

 

 

1972 1985 1992

 

120,000
100,000 I
80,000
60,000

 

 

 

 

 

[I Bottomland forest

[1 Freshwater marsh
Swamp forest

 

Hectare

/?

 

 

 

 

 

 

40,000 / /
20,030 /// /

1972 1985 1992

 

 

0\\\\\

 

_.

01

Figure 1.4.

 

 

 

 

 
   

 

 

90,000
0) 75,000
a 60,000 I Freshwater marsr
'5' 45,000’ ' [:1 Swamp forest
I 30,000
15,000 a \

 

 

 

1972 1985 1992

Figure 1.5.

102

 

moonwE can $.an N SEED

m fiQZmnE/w

103

 

owe 2%?

7:2 52< sax—3862

82o 55 as

$2

 

mod - 3.0 :- mo .32 33. sewom 3%; 32:20 .3 8 958:5
$2 33 EmEBmE< mm? 3800
Ed - cod _.m - cad m # .owfl 2:; £83 38680: can Exxon
22 .85 ooom
m8 3 - no 3 52 .aom 550882 8:6 55 .3 a Egg

:5 owsfi woo—E b6
mg Eooom 89:52 325 295m ozm beam 20:85

 

.803 95:: 29:58 was: 596 Eooom Ho 36:3 museum 8088 “3934 ._.N Bank

104

 

 

 

 

 

Sty
€2o>ED 23m 53:22».
>530 _§:oE:o:>:m ho EoEtmaoQ :nwEBEN
2:2on w::o:=o§ 813 2350.
as. I a... fin mN; I N~ EN ma .35,
I- _.m I- wm _ 392
E I Z 3” cm? I 8 2m 2 Nomi:
as I ad Nm mm; I m_ mom on 1:240
owcmm =32 owcam S82 80:3 :osaficww.0
£5 Eooom $5 «24 885m Mo 59:52 wE—QEam

 

.mcocmtomno 22.. 82.8 8 wow: moo—2 3:5 mm 05 a8 :ocmctfifi .N.N 2an

105

  
   
   
   
   
  

61

Sample lakes
(11:93)

 

 

County borders

 

WRS-2 Path 21

 

 

 

Fl 31
g 360 Kilometers

Figure 2.1.

106

 

 

ln(Secchi depth)

 

 

 

 

 

3 I I
Complete Dataset
2
1
0 ° ° p < 0.001
rmse = 0.31
l I
1.8 2.1 2.4
ETMl/ETM3

107

2.7

Figure 2.2.

 

(ID-h
010

Percent of lakes (%)

001

 

010

001

‘n 1 1-

Kloiber et al. 2000 1,
Mean Secchi depth =

D 1.5m [j

s
Q
8
K o
x. (/o of lakes _<_1.5m =
O
s:
\
O

 

61.7%)

This study
(Compete dataset)

Mean Secchi depth =
3.1 m

(% of lakes 515m =
8.6%)

Eutroph/c
Mesa; r 0.0/7/0

 

 

 

 

 

 

 

 

 

 

 

 

 

0.51 1.52 3 3.54 4.55 6.57 7.5
Secchi depth (m)

Figure 2.3.

108

44mm 4>
o 0

Percent of lakes (%)

OU'I

 

com
.019"?

0.01

 

ll

 

 

 

 

EU! r 010th

 

 

 

M680] r 0.0/7/0

 

é)

8

sh
g,
Q

 

Kloiber et al. 2000
Mean Secchi depth =
1.5m

(% of lakes 51.5m =
61.7%)

This study

(Subsampled dataset)
" ' Mean Secchi depth =

1.8m
(% of lakes _<_1.5m =

I 47.1%)

 

0.51 1.52 3 3.54 4.55 6.57 7.58
Secchi depth (m)

109

Figure 2.4.

 

 

ln(Secchi depth)

N

_L

O

 

 

l

Subsampled Dataset

    

r2 = 0.821

 

 

p < 0.001 -
rmse = 0.19
1.8 2.1 2.4
ETM1/ETM3

NO

2.7

Figure 2.5.

 

lt-

ln(Secchi depth)

 

dataset

’
fl
0’ I’
I I
l' I
I
" I
’ ”
I
r’
I
r’
I

I
O
f
f
"
C
’

"
"
’

Subsampled

 

Complete I’ ,

    
     

’
l ‘

’
I ”

’4

 

 

I dataset
1 .8 2.1 2.4 2.7
ETM 1 /ETM3

Ill

Figure 2.6.

 

1.0

0.5

Residuals

 

 

 

 

 

 

 

 

 

 

I I l I
Complete dataset Subsampled dataset
0
O
.. 00 ° 0 ° — _ ,4
o o :0 o ,x’/
o m ”””” o ,x’”
~~~~~~~~~~~~~~~ 5 “~-~~‘ _-—" x' o
‘B“‘°---& {58°00 o o ""6 Q,
o -3 (”F
--w mio- ...... °. ,,--. .......
vvvv 3° 00° o
o 0 <9 0 ~- 0 O O ““““
_D 0 00 0 O o _ .— “““““ e:
O 0
O
l l l 1
1.8 2.1 2.4 2.71.8 2.1 2.4

Predicted ETM1/ETM3

l12

Figure 2.7.

2.7

Residuals

 

l l I

Complete dataset

 

 

 

 

 

1 3 7

20013 7

 

Sutisampied dataset

 

 

 

 

Observed Secchi depth (m)

113

Figure 2.8.

 

 

 

Percent of lakes (%)
o 01 '4 '4 N N 03 (JD h
. .0. 01 O. 01. O. U? 0

 

Statewide MI dataset
Mean Secchi depth =
3.0m

(°/o of lakes 51.5m =
14.5%)

 

f, __ This study
fl (Compete dataset)

Mean Secchi depth =
3.1m

(% of lakes 51.5m =
8. 6%)

Jim _,,.

0.511.52 3 3.54 4.55 6.57 7.58
Secchi depth (m)

 

 

 

 

 

 

 

 

 

 

Figure 2.9.

114

mocswi new 333. m 5520

m fiQmemxx

ll5

 

23 _.m 0.5m Z- SON .m ..onEBmmm EEoEoE 8:820:

 

_oidw
\ 0398 \ 868 vs Now a- Son .m eon—Eamom EESEE 53835
23 ad 0.3. n- _oom .m conanom owxaaoz “madam

322w
\ 6:38 \ 338 we osm _ _- 50m .v “Isms/x .5032 5.0082353
tag. me wdm N _ - 58 .v Hmsw=< acme—03 283m
use am 4.3 w a - Sam .4 same... Egg 28m
:9: m.m w.mm a. 58 .v Saws< 826:3 venom
2:3 Wm _dm m- 5cm .v Hm=w=< confines—mm 2me
:2: 2 EN o _ - 58 .4 2&3 8323 8.5

:35?an mums—g
-'..28 c5 35 can wEEEam

“cos—meow 59% :82 no..." 385m 8:“— :353 when 3a.. «wasp: .3580 8:3

 

$83 .453 Sm Swag—£5 owns: was 35:88.35 Bambi .m_.m 3an

H6

 

.HCUEQO_D>DU _OUOE ..Om “yum: HO: mm; 9:: .EOCNU:~N> _DUOE uOm Gum: mag Dv—wtu U-D—mfi—g n
7

.EcQEBEb—B do ..SE ._o>2m\o_nnoo\m5_oe .23 $5328 a

wocwimw beigfizmsv 203 :02? 8:2 8m 35 29:3 :8 2 33 mo 355% cm so 393 Boa 33:53 8.8 22:68 econ 3823 a

 

9:; We 03 _ _- 58 .m baEBmom 838:0: LEocEB
_oiuw
\ 2.38 \ 838 ix 1mm 5+ _oom .m Bap—Baum «3080 39:5
:9: m8 o.~m a. Son .m eBEoEom 8.3252 mE>oZ

Ea Wm nfi E- _oomwgeoaom 58:52 fiche; 2e:

 

 

 

oo . .Nvo; . com mmv hm 5m. . . 0.. _oonR 55585...

85.334 cm. mwm 5. Wm om EN 5on3 558.55

8m..om.o. am new mm Wm m men 5on3 3.5%.

ooodmw. m owm 0mm m. m.m om m .m 585% V.5525.V§>

oo. .w . mK com .35 cm o.m o. 5.. .oonE 5.95m

$55.95

83.3 am. New m. 5.. n o.m 585% .553.

A 83551..

8.5. ..m 30. cwv 5. ad a m.m 585% .553.

85.2.... .8. a: 2 Q m em 5855 23m

839... com Sm w M: E em 53% 805

Eases 5558 33 Sue 3-6. :5

2.5.5.5.: ARE. 55.5.: 35:55.... SE a. a55.55 5555535.:

535.55.555.25 fizz—5.2 .35... .855 a 51.0 .533 um... Eooom 855 own—E 3.5..

 

.85.... >93 .5 85.55825 85.5.2.8... .5. .m 5.5a...

118

55552555 .5555 5. 55m: 5: v.53 55 55.25:? .5555 5. 55.5 553 5.51. 5.5.5.? o

 

£5825 ..

8558.53 5.5:; .5 55.585855 gm... 55: £553 £55.53 won 555.53 .2355; 5. 5-50 Gem 5. 555.5. 55.5 b5> E o :55.

85.5: 55.53 .325 5.50. 5.85 525.55.55.58 5:. 5 55585.2 .HMMOHm . . M3735. OmQ-..>. .3 855:8 5523 5.8 53>? a

 

oofigqm
ooo€oofi

oov.ow..m

coo. _ mum

3:

3

OS

No.

mwv

wvm

Pm

wmm

w.

v.

m...

Qm

ON

..v

Om

o.

0..

..m

Wm

..v

.ooN\m\o

.ooQQm

.ooQQm

.ooQQa

55.5.5.3

55.555

5357.

52:55.55

5.35

119

 

 

 

m. c 5 .52
.oodv .oodv .oodv .525. 33:
Q5 00 0K5 N0 0&5 —h ambtfifikbbtbb gtmbkmm U0m..®ED~.—m 139.1
.. o m .52
.oodV _oodv .oodv a :55 ES:
eke cw 5x9 mw ox» mm = 552355525 35.55% 955.”.
. . o m .3.
.oodv .oodv .oodv 5:5. .53:
Q5 Vb 55 mm Re mm 35555255 355.5% 55325..
N. o 5 .2
.oodv .oodv .oodv 5:5. .53:
0R9 —h 05 Nb 0R5 m0 : wUQQVKQUQQQ .295ka ..D>Oo Hrs—Q 3.03.1—
.55:. 5:35.50 .555 52.55:.5m .55.: Hm=w=< “m5. Roumzfim .5355 Eur.

 

5.558 55:: 55 5. 5.5.3555. :wo. 55 EB. 3.35M. .m.m 533‘

I20

..5508 5... c. 85.5. .0 .5588 .55. 5... 0. 5.5.5. Z 5
0.55 80.. 85.5...5 3852.558

5. 555...; 8558585... :5 .0 85.5...505 5... .5... £55809... 5... 5.8. .. .... .5508 ..5.5>0 .0. 55555.5 5.55.8.8 5... m. .55. 5.5 >9 5.. H .
5.55.5.8 .55... ..555 5. 50.3555. ..wo. 25.55 5;. ..5 555.58 55.53 ..550853. .55. 5.58.55 :58 88.3 80553530
.0 .5588 5... 80.. 55.5.8.8 5.5 8.5.. 855.0555 5. ..5.. 5553550 5... ..5... .555 .55... obom v .0 855.5: d. 5:.5> 55.5555.
5:585 5... :5... 53:55.5... 85>5 55.5.55... .55... 5 55.. C555 .55.. .0 .55. 5 .0. o A. 55.5.. 55:095. .5555. 5... .. “5055.55 .55...
5 8555.55 .0 5.5.3550... 5... 8.55... A5.5508 5... 555555.. .5...85x5 5n. 6.55.5.8 85.... 5.55..5> 55.5.55. 5... .0 m5:.5> 85.5...5

5.? 8059.550 .0 8.5.. ..5 .0 5055858 500....5v... 8585.58 5... ..0 5555.. 5.5.55.5 .5508 ..5.5>0 5... 55555555 555555588 855.5. a

 

m. 5 5 .52
.oodv 550.5 .oodv 5.55. 5.3...

55 mm 55 mm 55 N5 5 5525555205 855.5. 23m. 55858.56

12]

 

at!
'l
5
J
7

 

 

 

L . .
mm... om... mm... 8... :35 53.55.35
.5... mm... mm... mm... 5355.3 .30.
om... Om... 98 cm... 5:58..
oi. 5.... .55 8... 555555
a... 95... .5... cm... .28 :5... .225
5.55%.): 5555525 3,: as... .8523 5.5.5

 

5.5508 55.0. 5585.55m 555 .msw5< 5... 5. 5555 550.5. .555.>.55.

N. 5... .0. 5585.555: mo: 5.55.5...505 5.55 5... .0 55558 55. 5553.55 5555.5...5 5... 55$. 80.. m55.5> 5. .55.-.. .m.m 5.55 H

122

 

 

mm .5 55.5 5. .5 w. .5 m w .5 55.5 5... .5 55.5 55 .5 m5.5 5.555.554...
5m .5 5. .5 55 .5 N55 55 .5 N55 5N .5 m. .5 55.5 N55 5.555.554...
5N .5 m . .5 m 5 .5 55.5 . 5 .5 55.5 5. .5 5N.5 m. .5 5N5 5.)...
NM .5 . . .5 5~ .5 5N... . m .5 . . .5 55 .5 55.5 55 .5 No.5 NS... 555.55%.
55 .5 55.5 5. .5 ..N... 5.5 .5 55.5 55.5 55.5 55.5 .55 5555.55.25
55 .5 55.5 55 .5 55.5 55.5 55.5 55.5 55.5 55 .5 55.5 5.55Q*N.>...
5. .5 mN.5 mm .5 5. .5 55 .5 55.5 mm .5 mod N 5 .5 m5.5 5.2... .5>05
55.5 55.5 55.5 55.5 5N .5 m . .5 5m .5 55.5 55 .5 55.5 NS... .55... .5555
555.5.»
a 5.. 5 N. m 5.. a. N. 5 N. 85.55500 5.03.50 .55...
5.05 585.05... 5555.59.55...
858.55m M052... 5 ..5550.0.5U 5.00 5.53 55.5 .5855

 

.5555 5..w..5.55.5..5 55.5. .0. .5. 5.55. 55m ..5>5. 5. .5 5... .5 5555.555 5.5 5.05

5. 5525., .m5.55..5> 8.5. 55.55.55 5.5.5.. 5.5.5.55 55.5. .555.>.55. 50 55.58555 5.55.5...505 .5508 5... .0 3.555. 50.55555”. .5.m 5.55 ...

123

55.5

55.5

5N .5

5N5

.55

55 .5

N..5

5N5

55.5

55 .5

55.5

.55

55.5

55.5

55.5

NN.5

m..5

55.5

55.5

55.5

m55

5.5

5.5

5N .5

5N5

5N5

55.5

55 .5

N55

5N .5

N55

5N5

N55

5.5

5N5

:5

5.5

5.5

5.5

55.5

..5

N55

5.5

...5

5.5

..5

J.
bl nulll .L.

55.5
.55
55 .5
N55
N55
N55
55 .5

5N5

55 .5
5m .5
.5 .5

55 .5

55.5

55.5

55.5

.55

55.5

55.5

N55

555

5.5

N55

.55

m55

N55

55 .5

N55

55.5

5N5

5.. .5

55 .5

55.5

55.5

.55

55 .5

55.5

55 .5

55.5

55.5

5.5

w5.5

..5

.N.5

N55

55.5

55.5

N55

N55

N55

m55

55.5

.55

55.5

55 .5

55 .5

55.5

55.5

55 .5

55.5

55 .5

55 .5

55.5

55.5

55.5

55.5

55.5

.55

55.5

55.5

55.5

.55

55.5

.55

55.5

55.5

 

N. 55555><

5559525.
5509.525
525
525
55.5%.)...
5895.)...
525

NE...

58532..
5509525
525

NE...

503.55.505.55 *

2.5m 55555:...5m

555558555 .55...

5.55507.

l24

 

 

8.0 one 082055 coggaoai
:3 3o 3 ifiosEu
5.0 8.0 58 533
56 No.0 558%:th v36 Eooom

a nu 85282255 813

 

.mzcs ucmtoaofimnu 8:2 _8 3 033 com .AQ u Zv 3%:

=« ‘3 “8483 3:588 2: m5? 85380820 83 Ema—om 3823 £59» 32% van NEH H3234 ko 3:52 coamewom .m.m 2an

125

 

 

 

0 mod cod ~
mg v0.0 86 o “:84 wage—m
36 3.: mum—«334‘
_ mwd --- m
m _ _d mud N
m 8.0 00.0 _
ms 86 So o “Swag
mud and owauo><
mi cod cad _
E cod 3.0 o 8>oU Ear. 128:1—
_o>o_ _u>o~
some wcrmzmmflo 380:8 some wEmem—o bamboo 654
8% @3530 mo # ho bzfimpoa uwfio>< h_o 3:53on owfiu>< 5>oU bowamu :85
6on 355800 $on BQEBQom

 

Sm H m_ 8:3 22%;? E was 0:8 .803: be 598:: :38 2: 9.3 23.53 was: 2228:? 688 05 Mo £33m .c.m 033.

126

m

.53.... _'

 

 

mm

mo

5

Va

2

C

mmd

86

mod

No.0

m~.¢

mmd

omd

w_.o

omd

mmé

cod

Ed

and uwaho><

wad

mod

cod

mNd

3.0

Cd

mmd

Omd

mNd

oed

mmd

uwuuo><

uwmao><

23mm me._®~:£3m

nomfiEnzm 50%

127

  
   
 
      
 

-‘ '
-__-_‘,

   
 

N

A

A August sample lakes
(n=6)

0 September sample
lakes(n=7)

 

, County borders

   

was-2 Path 21

—---J

 

 
  

,A R 31
H60 Kilometers

Figure 3.1.

128

 

    
 

liliili
liiiliii 1

w-‘i‘
1

none none. ..lll "
Iona-ouooouoluol

Figure 3.2.

129

 

 

 

Emergent

I Floating leaf

D Total submersed
l Submersed EWM

 

 

 

 

 

 

 

 

 

 

 

  

130

VA

 

 

W

 

 

 

 

Figure 3.3.

VA

 

 

 

 

 

 

 

 

1

 

’1

1114

1
”$050. 0?)
o '89 o
0
00

 

 

 

I Lake percent plant cover
1 [:1 Littoral percent plant cover

1

 

 

 

 

 

 

1

 

11:31:11

\
0Q

 

momunw
086

.o>oo Ema €00qu

40

20
0

Q?

Figure 3.4.

131

 

A. Aug st (N = 6 lakes)

  
     

g
1 a. 1'“

 

 

 

B. September (N = 6 lakes)

3O
20
1 0

O

 

Number of sites

 

 

 

 

1 000
800‘
600‘
400‘
200‘

 

 

C. Combined model (N = 12 lakes)

 

Leve|0
I

(020%)

Level1

(21 -40%)

. Level2
VA
(41 -80%)

Level 3
(81 -1 00%)

 

 

 

Figure 3.5.

132

 

 

 

I August

d

 

D Septem

 

 

 

 

 

 

 

 

 

 

 

 

 

Figure 3.6.

133

 

60‘

 

Number of
3.

 

 

 

 

 

 

LevelO

(020%)
[j Levell

(21 40%)

Level 2
(41 -80%)

Level3
(81-10096)

 

Figure 3.7.

134

 

 

IIIIIIIIIIIIIIIIIIIIIIIIIIIIIII

' llllllllllllllllllllllllllllllllllllljlllll

1293 02356 3