EVALUATING LAKE PHYTOPLANTON RESPONSE TO HUMAN DISTURBANCE AND
CLIMATE CHANGE USING SATELLITE IMAGERY
By
Linda Nicole Novitski

A DISSERTATION
Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of
Zoology – Doctor of Philosophy
2014

ABSTRACT
EVALUATING LAKE PHYTOPLANTON RESPONSE TO HUMAN DISTURBANCE AND
CLIMATE CHANGE USING SATELLITE IMAGERY
By
Linda Nicole Novitski
Accurate and cost-effective assessment of water quality is necessary for proper
management and restoration of inland water bodies susceptible to algal bloom conditions.
Landsat and MODIS satellite images were used to create chlorophyll and Secchi depth
predictive models for algal assessment of Great Lakes and other lakes of the United States.
Boosted regression tree (BRT) models using satellite imagery are both easy to use and can
have high predictive performance. BRT models inferred chlorophyll and Secchi depth more
accurately than linear regression models for all study locations. Inferred chlorophyll of
inner Saginaw Bay was subsequently used in ecological models to help understand the
ecological drivers of algal blooms in this ecosystem.
For small lakes (non-Great Lakes), the best national Landsat model for lntransformed chlorophyll was the BRT model and had a cross-validation R2 of 0.44 and a
0.76 ln-transformed µg/L RMSE. The best national Landsat model for Secchi depth was also
a BRT model that had an adjusted R2 of 0.52 and a 0.80 m RMSE. We assessed the
applicability of the national chlorophyll model for ecological analysis by comparing the
total phosphorus- chlorophyll relationship with chlorophyll determined from sampling or
remote sensing, which showed the total phosphorus-chlorophyll relationship had an
adjusted R2 = 0.58 and 1.02 ln-transformed µg/L RMSE with sampled chlorophyll versus an
adjusted R2 = 0.56 and 1.04 ln-transformed µg/L RMSE with chlorophyll determined by the

boosted regression tree remote sensing model. For Great Lakes models, the MODIS BRT
model predicted chlorophyll most accurately of the three BRT models and compared well
to other models in the literature. BRT models for Landsat ETM+ and TM more accurately
predicted chlorophyll than the MSS model and all Landsat models had favorable results
when compared to the literature. BRT chlorophyll predictive models are useful in helping
to understand historical, long-term chlorophyll trends and to inform us of how climate
change may alter ecosystems in the future. In inner Saginaw Bay, annual average and upper
quartile Landsat-derived chlorophyll decreased from 7.44 to 6.62 and 8.38 to 7.38 µg/L
between 1973-1982, and annual upper quartile of 8-day phosphorus loads increased from
5.29 to 6.79 kg between 1973-2012. Simple linear and multiple regression models and
Wilcoxon rank test results for MODIS and Landsat-derived chlorophyll indicate that
distance from the Saginaw River mouth influences chlorophyll concentration in Saginaw
Bay; Landsat-derived surface water temperature and phosphorus loads to a lesser extent.
Mixed-effect models for MODIS and Landsat-derived chlorophyll were related to
chlorophyll better than simple linear or multiple regressions, with random effects of pixel
and sample date contributing substantially to predictive power (NSE=0.35-70), though
phosphorus loads, distance to Saginaw River mouth, and water were significant fixed
effects in most models. Water quality changes in Saginaw Bay between 1972-2012 were
influenced by phosphorus loading and distance to the Saginaw River’s mouth.
Landsat and MODIS imagery are complementary platforms because of the long
history of Landsat operation and the finer spectral resolution and image frequency of
MODIS. Remote sensing water quality assessment tools can be valuable for limnological
study, ecological assessment, and water resource management.

Copyright by
LINDA NICOLE NOVITSKI
2014

ACKNOWLEDGEMENTS

Many thanks are due to the research group in which this research was created; my
main advisor, Dr. R. Jan Stevenson, Dr. Jiaguo Qi, Dr. Dave Hyndman, Dr. Peter Esselman, Dr.
Anthony Kendall, Dr. Sherry Martin, Dr. Siam Lawawirojwong, Dr. Tanita Suepa, and Dr.
Shengpan Lin. Thanks are also due to committee members Dr. Elena Litchman, Dr. Orlando
Sarnelle, and Dr. Mantha Phanikumar. Thanks to Dr. Narumon Wiangwang and Zhen Zhang
for their assistance. Finally, I am deeply appreciative of the continuous support of my
loving parents, sisters, and husband through this process.

v

TABLE OF CONTENTS

LIST OF TABLES ...................................................................................................................................................... viii
LIST OF FIGURES .................................................................................................................................................... ix
KEY TO ABBREVIATIONS ....................................................................................................................... xi
CHAPTER 1
REMOTE SENSING OF CHLOROPHYLL-A IN THE GREAT LAKES USING LANDSAT AND
MODIS SATELLITE IMAGERY; A HISTORY ................................................................................................. 1
Literature Review ................................................................................................................................... 1
LITERATURE CITED................................................................................................................................ 5
CHAPTER 2
NATIONAL CHLOROPHYLL AND SECCHI DETECTION BY LANDSAT 7 ETM+ SATELLITE
IMAGERY FOR WATER QUALITY ASSESSMENT ...................................................................................... 11
Abstract ........................................................................................................................................................ 11
Introduction ............................................................................................................................................... 13
Methods ........................................................................................................................................................ 16
Ecological data, satellite imagery, and pixel extraction ................................................. 16
Remote sensing images and data processing...................................................................... 17
Linear regression analyses ........................................................................................................ 18
Boosted regression tree analyses ............................................................................................ 20
Model diagnostics ..............................................................................................................................21
Results........................................................................................................................................................... 22
National and lake type models to predict chl .........................................................................22
National and lake type models to predict Secchi depth .................................................. 24
Discussion ................................................................................................................................................... 25
Evaluation of national chl model ............................................................................................ 25
Evaluation of national Secchi depth model ......................................................................... 25
Possible improvement of models ............................................................................................. 26
Robustness and application of national models ................................................................ 28
Summary and concluding remarks ........................................................................................ 29
APPENDIX .................................................................................................................................................... 31
LITERATURE CITED................................................................................................................................ 47
CHAPTER 3
MODIS AND LANDSAT BOOSTED REGRESSION TREE CHLOROPHYLL PREDICTIVE
MODELS FOR THE GREAT LAKES................................................................................................................... 52
Abstract ........................................................................................................................................................ 52
Introduction ............................................................................................................................................... 52
Methods ........................................................................................................................................................ 55
vi

Study locations and ecological data ...................................................................................... 55
MODIS satellite image processing and pixel extraction for BRT model training ... 55
Landsat image processing and pixel extraction for BRT model training .................. 56
Data preparation and statistics for MODIS and Landsat data ..................................... 58
MODIS image processing for model application ................................................................ 59
Landsat image processing for model application ............................................................. 60
Model evaluation .......................................................................................................................... 61
Results........................................................................................................................................................... 61
BRT models and partial correlation plots............................................................................ 61
Discussion ................................................................................................................................................... 63
APPENDIX .................................................................................................................................................... 67
LITERATURE CITED................................................................................................................................ 80
CHAPTER 4
IMPACTS OF CLIMATE CHANGE AND LAND-USE/LAND-COVER CHANGE ON
PHYTOPLANKTON IN INNER SAGINAW BAY USING LANDSAT AND MODIS SATELLITE
IMAGERY, 1973-2012 .......................................................................................................................................... 84
Abstract ........................................................................................................................................................ 84
Introduction ............................................................................................................................................... 85
Research objectives ...................................................................................................................... 88
Methods ........................................................................................................................................................ 88
Data ..........................................................................................................................................................88
Statistics for MODIS and Landsat ........................................................................................... 90
Results........................................................................................................................................................... 92
Temporal trends ........................................................................................................................... 92
Relationships between algal blooms and global change variables ............................. 93
Discussion ................................................................................................................................................... 94
APPENDIX .................................................................................................................................................... 98
LITERATURE CITED................................................................................................................................ 115
CHAPTER 5
SUMMARY AND SYNTHESIS: USE OF REMOTE SENSING IN ECOLOGIAL ASSESSMENT ...... 121
National Lakes Assessment chlorophyll and Secchi depth predictive models....... 121
Landsat and MODIS boosted regression tree models for chlorophyll prediction
in the Great Lakes ................................................................................................................................... 122
Using Landsat- and MODIS-derived chlorophyll to assess water quality in
Saginaw Bay from 1972-2012 .......................................................................................................... 123
LITERATURE CITED................................................................................................................................ 125

vii

LIST OF TABLES

Table 2-1: Landsat 7 ETM+ wavelength band features (USGS 2011a). ..................................... 32
Table 2-2: US EPA National Lakes Assessment lake types. .............................................................. 33
Table 2-3: National boosted regression tree chl and Secchi depth models. .......................... 34
Table 2-4: National and lake type linear regression chl detection models. ........................... 35
Table 2-5: National and lake type linear regression chl and Secchi model coefficients. 36
Table 2-6: National and lake type linear regression Secchi depth detection models....... 37
Table 2-7: Chlorophyll and Secchi depth models from the literature. ...................................... 38
Table 3-1: MODIS wavelength band features (NASA 2013). .......................................................... 68
Table 3-2: Landsat 7 ETM+ wavelength band features (USGS 2011)........................................ 69
Table 3-3: Landsat 1-3 MSS wavelength band features (USGS 2011). ...................................... 70
Table 3-4: Boosted regression tree results for MODIS, Landsat ETM+, TM, and MSS. ...... 71
Table 3-5: Comparison of BRT chl predictive models to models in the literature. ............ 72
Table 4-1: Linear regression results for temporal trends in chlorophyll, water
temperature, and phosphorus loads in inner Saginaw Bay............................................................ 99
Table 4-2: Wilcoxon Rank test results using near and far groupings of MODIS or
Landsat chl or water temperature data. ................................................................................................... 100
Table 4-3. MODIS (2000-2012) linear regression results. .............................................................. 101
Table 4-4: MODIS linear mixed-effect model results. ........................................................................ 102
Table 4-5: Landsat (1973-2012 or 1984-2012) linear regression results. ............................ 103
Table 4-6: Landsat linear mixed-effect model results. ...................................................................... 104
Table 4-7: Comparison of MODIS- and Landsat-derived inner Saginaw Bay
chlorophyll values and values from the literature.............................................................................. 105

viii

LIST OF FIGURES

Figure 2-1: Histograms of untransformed and ln-transformed chlorophyll and Secchi
depth. ........................................................................................................................................................................... 42
Figure 2-2: Remotely sensed (linear regression predicted) ln-transformed
chlorophyll a using band ratio 1/3 versus measured ln-transformed chlorophyll a
(R2adj=0.19, 1.38 ln-transformed µg/L). .................................................................................................. 43
Figure 2-3: Remotely sensed (boosted regression tree predicted) natural lntransformed chlorophyll versus measured ln-transformed chlorophyll a (R2adj=0.44,
0.76 ln-transformed µg/L). ................................................................................................................... 44
Figure 2-4: Remotely sensed (linear regression predicted) natural ln-transformed
Secchi depth using band ratio 1/3 versus measured ln-transformed Secchi depth.
(R2adj=0.49, 0.77 ln-transformed m RMSE). .......................................................................................... 45
Figure 2-5: Remotely sensed (boosted regression tree predicted) Secchi depth versus
measured untransformed Secchi depth (R2adj=0.52, 0.80 m RMSE). ....................................... 46
Figure 3-1: Measured chl versus MODIS BRT predicted chl for 2007-2009 GLNPO
training data with 1:1 line. ............................................................................................................................... 73
Figure 3-2: Partial dependence plots for the MODIS BRT................................................................ 74
Figure 3-3: Measured chl versus Landsat ETM+ and TM BRT predicted chl for 20072009 GLNPO training data with 1:1 line.................................................................................................... 76
Figure 3-4: Partial dependence plots for the Landsat ETM+ and TM BRT. ............................. 77
Figure 3-5: Measured chl versus Landsat MSS predicted chl for 2007-2009 GLNPO
training data with 1:1 line. ............................................................................................................................... 78
Figure 3-6: Partial dependence plots for the Landsat MSS BRT. .................................................. 79
Figure 4-1: Saginaw Bay’s inner bay with depth ≥ 5 m...................................................................... 107
Figure 4-2: Inner Saginaw Bay study area and river mouth inputs into Saginaw Bay. .... 108
Figure 4-3: Annual average and upper quartile MODIS and Landsat-derived
chlorophyll concentration (µg/L) in inner Saginaw Bay, 2000-2012. ...................................... 109
Figure 4-4: Annual average and upper quartile Landsat-derived chlorophyll
concentration (µg/L) in inner Saginaw Bay, 1973-2012.................................................................. 110
ix

Figure 4-5: Annual average and upper quartile 8-day phosphorus load (kg) in inner
Saginaw Bay, 1973-2012. .................................................................................................................................. 111
Figure 4-6: Annual average and upper quartile Landsat-derived surface water
temperature (˚C) in inner Saginaw Bay, 1984-2012. .......................................................................... 112
Figure 4-7: MODIS-derived chlorophyll vs. mixed-effect model chlorophyll (µg/L) in
inner Saginaw Bay, 2000-2012 with 1:1 line. ......................................................................................... 113
Figure 4-8: Landsat-derived chlorophyll vs. mixed-effect model chlorophyll (µg/L) in
inner Saginaw Bay, 1984-2012 with 1:1 line. ......................................................................................... 114

x

KEY TO ABBREVIATIONS

A=August
AIC = Akaike information criterion
avg. = average
BIC = Bayesian information criterion
BRT = boosted regression tree
chl = chlorophyll a
D = distance between satellite image pixel and Saginaw River mouth (km)
diff. days = difference between in-lake sample date and date of satellite overpass
DOC = dissolved organic carbon
DOY = day of year
ETM+ = enhanced thematic mapper
GLNPO = Great Lakes National Program Office
JL=July
JN=June
L1G = Landsat image processing level
L1T = Landsat image processing level
LOESS = locally weighted scatterplot smoothing
max = maximum
MCKT = MODIS conversion toolkit
MERIS = Medium Resolution Imaging Spectrometer
min = minimum

xi

MOD02 = MODIS satellite data product
MOD09 = MODIS satellite data product
MODIS = Moderate Resolution Imaging Spectroradiometer
Month=sample month
MRT = MODIS reprojection tool
MSS = multispectral scanner
NASA = National Aeronautics and Space Administration
NLA = United States Environmental Protection Agency National Lakes Assessment
NOAA = National Oceanic and Atmospheric Administration
NS = not statistically significant
NSE = Nash-Sutcliffe efficiency
OT=one-time
P = 8-day phosphorus load (kg)
R=range
RMSE = root mean square error
S=September
SA=seasonal average
SeaWiFS = Sea-Viewing Wide Field-of-View Sensor
TM = thematic mapper
TOA = top of atmosphere
TOTS=two one-time samples
TP = total phosphorus
Type=sample type
xii

UQ = upper quartile
USGS = United States Geological Survey
WT = Landsat-derived water temperature (˚C)

xiii

CHAPTER 1
REMOTE SENSING OF CHLOROPHYLL-A IN THE GREAT LAKES USING LANDSAT AND
MODIS SATELLITE IMAGERY; A HISTORY
Literature Review
Since the 1970s, improving water quality has been a national priority, however, in
the Great Lakes, concerns, such as fish population declines (Burkholder 1930) and
eutrophication (Beeton 1965), were noted much earlier. One retrospective study of diatom
records indicates nutrient enrichment as early as European settlement and forest clearing
in some part of the Great Lakes (Schelske et al. 1983). In 1972, guidelines were established
by the Clean Water Act that mandated regulation of point source pollution into water
bodies, including municipal and industrial sources of nutrients (Copeland 2010).
Amendments to the Clean Water Act in 1987 added regulations for nonpoint nutrient
sources, including agriculture (Copeland 2010). In addition to the protection provided to
the Great Lakes by the Clean Water Act, independent remediation action plans have been
created and carried out for numerous areas of concern, including Saginaw Bay (USEPA
2013). In the years following remediation, nutrients and chemicals from municipal,
industrial, and agricultural sources decreased significantly and led to improvements in
water quality in some areas (Auer et al. 1976, Bierman et al. 1984). Meanwhile, other
perturbations, including invasive species introductions and climate change, have caused
additional ecological challenges (Fahnenstiel et al. 1995, Lofgren and Gronewold 2012).
To enable effective assessment, remediation, and management of affected water
bodies, it is necessary to understand how the water quality varies in space and time and
how it is related to sources of pollution. Chlorophyll a (chl) concentration, a proxy of algal
1

biomass, is often used to help quantify ecological health of water bodies, as it can be
indicative of high nutrient levels and water temperatures (Carlson 1977, Paerl and
Huisman 2008). In some cases, high algal biomass may have deleterious effects on fish and
other aquatic organisms, as well as humans who touch or ingest water with algae (Paerl et
al. 2001). Numerous studies indicate that algal blooms have the potential to increase
(Hauer et al. 1997) and, in some areas, are already increasing in frequency, which could be
caused by increased nutrient input, higher water temperatures and increasing
atmospheric CO2 (Shapiro 1997, Hunter 2003).
Water quality assessments are usually done by physically taking water samples and
analyzing them in the lab. However, advancements in remote sensing technology and
development of modeling tools in recent decades have allowed ecological assessment from
a distance, enabling evaluation with larger spatial and temporal scales and providing a
more comprehensive view at, potentially, less cost. Landsat satellite imagery has often been
used to detect chl in inland lakes (Lillesand et al. 1983, Mayo et al. 1995) and coastal ocean
waters (Erkkila and Kalliola 2004, Han and Jordan 2005). Landsat also has a long history in
water quality detection in the Great Lakes, from observation of colors in Lake Erie (Strong
1974), calcium carbonate precipitation in Lake Michigan and Lake Ontario (Strong 1978),
chl detection in central Lake Michigan and Green Bay (Lathrop and Lillesand 1986), total
suspended solids or Secchi depth in Green Bay (Lathrop et al. 1991), or phycocyanin
detection in western Lake Erie (Vincent et al. 2004). While some Landsat studies have
looked back at water quality over time (Olmanson et al. 2008), none have looked back at
chl trends over the lifetime of Landsat.
Since its inception, chl algorithms have also been created, tested, and used for the
2

Moderate Resolution Imaging Spectroradiometer (MODIS) satellite in both ocean waters
(O'Reilly et al. 2000, Carder et al. 2004) and in the Great Lakes (Shuchman et al. 2006,
Kerfoot et al. 2008, Weghorst 2008, Becker et al. 2009, Kerfoot et al. 2010, Binding et al.
2012, Lesht et al. 2013). One of the early MODIS chl predictive models, the OC3 model that
was created for ocean waters (O'Reilly et al. 2000) was applied with some success in
freshwater lakes with little non-algal color producing agents (CPAs), or when chl was the
dominant CPA at the time (Kerfoot et al. 2008), but not as well when other CPAs are
present. More recent MODIS algorithms for freshwater lakes have tried to take account of
the other CPAs, such as colored dissolved organic matter (CDOM) and suspended particular
matter (Shuchman et al. 2006, Binding et al. 2012).
Whether using chl models for Landsat or MODIS, complications arise when
detecting chl in shallow areas, where reflectance from the bottom of the lake could
interfere with the chl signal detected by the satellite. Numerous studies have acknowledged
and tried to address this issue (Højerslev et al. 1977, Lyzenga 1981, Rundquist et al. 1995,
Lee et al. 1998, Tolk et al. 2000, Lodhi and Rundquist 2001, Cannizzaro and Carder 2006),
however no standard approach has been widely accepted and applied.
Historically, two general approaches have been use to develop models for predicting
chl with satellite images: 1) deriving empirical relationships between chl and satellite
bands or band ratios, alone or separately from other CPAs using linear or polynomial
regression analyses (Carpenter and Carpenter 1983, O'Reilly et al. 1998, Giardino et al.
2001) or 2) deriving relationships between two or more CPAs (including chl) at once
(Carder et al. 2004, Pozdnyakov et al. 2005). Many of the current models are either not
widely applicable, have not been tested for wide use, are challenging to implement without
3

expert statistical and programming knowledge, or require costly computer programs.
Other algorithms have been created and tested for the Great Lakes using other
satellites, including the medium resolution imaging spectrometer (MERIS)(Gons et al.
2008, Wynne et al. 2008), CZCS and SeaWiFS (Bukata et al. 1981, Shuchman et al. 2006,
Binding et al. 2007), or hyperspectral devices (Ali 2011). For additional information, there
are comprehensive summaries of remote sensing of cyanobacteria (Kutser 2009), of case II
waters (Bukata 2005, Odermatt et al. 2012, Lesht et al. 2013), or ocean waters (Stumpf and
Tomlinson 2005).
In my dissertation I used field data in conjunction with satellite imagery to test
models for chl and Secchi depth using existing remote sensing models and using a new
approach with boosted regression tree (BRT) model-building techniques. In the first
project, US Environmental Protection Agency field measurements and Landsat satellite
imagery were used to create chl concentration and Secchi depth predictive models for use
in small inland lakes across the United States. The second chapter explorers the use of BRT
models for chl prediction in the Great Lakes, using both Landsat and MODIS satellite
imagery. Landsat and MODIS complement each other in that Landsat allows for long-term
analysis of algal biomass at a finer spatial scale, and MODIS allows for chl detection with
finer spectral wavelengths at more frequent intervals. The third chapter uses Landsat and
MODIS-derived chl concentrations in mixed-effects models to better understand the
influence of river discharge and nutrient loading on chl concentration in Saginaw Bay. I
expect that developing remote sensing tools for water quality assessment can be useful for
ecological analysis to not only understand the natural dynamics of the ecosystem, but also
help predict future trends caused by climate change.
4

LITERATURE CITED

5

LITERATURE CITED

Ali, K. A. 2011. Prediction of water qualitly parameters from VIS-NIR radiometry: Using
Lake Erie as a natural laboratory for anallysis of case 2 waters. Kent State
University.
Auer, M. T., R. P. Canale, and P. L. Freedman. 1976. The limnology of Grand Traverse Bay,
Lake Michigan. Michigan Sea Grant Program.
Becker, R. H., M. I. Sultan, G. L. Boyer, M. R. Twiss, and E. Konopko. 2009. Mapping
cyanobacterial blooms in the Great Lakes using MODIS. Journal of Great Lakes
Research 35:447-453.
Beeton, A. M. 1965. Eutrophication of the St. Lawrence Great Lakes. Limnology and
Oceanography 10:240-254.
Bierman, V. J., Jr., D. M. Dolan, and R. Kasprzyk. 1984. Retrospective analysis of the response
of Saginaw Bay, Lake Huron, to reductions in phosphorus loadings. Environmental
Science & Technology 18:23-31.
Binding, C. E., T. A. Greenberg, and R. P. Bukata. 2012. An analysis of MODIS-derived algal
and mineral turbidity in Lake Erie. Journal of Great Lakes Research 38:107-116.
Binding, C. E., J. H. Jerome, R. P. Bukata, and W. G. Booty. 2007. Trends in water clarity of the
lower Great Lakes from remotely sensed aquatic color. Journal of Great Lakes
Research 33:828-841.
Bukata, R. P. 2005. Satellite monitoring of inland and coastal water quality: Retrospection,
inspection, and future directions. Taylor & Francis, Boca Raton, FL.
Bukata, R. P., J. E. Bruton, J. H. Jerome, S. C. Jain, and H. H. Zwick. 1981. Optical water quality
model of Lake Ontario. 2: Determination of chlorophyll a and suspended mineral
concentrations of natural waters from submersible and low altitude optical sensors.
Applied Optics 20:1704-1714.
Burkholder, P. R. 1930. A biological survey of Lake Erie. Science 71:288-289.
Cannizzaro, J. P. and K. L. Carder. 2006. Estimating chlorophyll a concentrations from
remote-sensing reflectance in optically shallow waters. Remote Sensing of
Environment 101:13-24.

6

Carder, K. L., F. R. Chen, J. P. Cannizzaro, J. W. Campbell, and B. G. Mitchell. 2004.
Performance of the MODIS semi-analytical ocean color algorithm for chlorophyll-a.
Advances in Space Research 33:1152-1159.
Carlson, R. E. 1977. A trophic state index for lakes. Limnology and Oceanography 22:361369.
Carpenter, D. J. and S. M. Carpenter. 1983. Modeling inland water quality using Landsat
data. Remote Sensing of Environment 13:345-352.
Copeland, C. 2010. Clean Water Act: A Summary of the Law. Congressional Researc Service.
Erkkila, A. and R. Kalliola. 2004. Patterns and dynamics of coastal waters in multi-temporal
satellite images: support to water quality monitoring in the Archipelago Sea,
Finland. Estuarine, Coastal, and Shelf Sciences 60:165-177.
Fahnenstiel, G. L., T. B. Bridgeman, G. A. Lang, M. J. McCormick, and T. F. Nalepa. 1995.
Phytoplankton productivity in Saginaw Bay, Lake Huron: Effects of zebra mussel
(Dreissena polymorpha) colonization. Journal of Great Lakes Research 21:465-475.
Giardino, C., M. Pepe, P. A. Brivio, P.Ghezzi, and E.Zilioli. 2001. Detecting chlorophyll, Seechi
disk depth and surface tempterature in a sub-alpine lake using Landsat imagery.
The Science of the Total Environment 268:19-29.
Gons, H. J., M. T. Auer, and S. W. Effler. 2008. MERIS satellite chlorophyll mapping of
oligotrophic and eutrophic waters in the Laurentian Great Lakes. Remote Sensing of
Environment 112:4098-4106.
Han, L. and K. J. Jordan. 2005. Estimating and mapping chlorophyll-a concentration in
Pensacola Bay, Florida using Landsat ETM+ data. International Journal of Remote
Sensing 26:5245-5254.
Hauer, F. R., J. S. Baron, D. H. Campbell, K. D. Fausch, S. W. Hostetler, G. H. Leavesley, P. R.
Leavitt, D. M. McKnight, and J. A. Stanford. 1997. Assessment of climate change and
freshwater ecosystems of the Rocky Mountains, USA and Canada. Hydrological
Processes 11:903-924.
Højerslev, N., N. G. Jerlov, and G. Kullenberg. 1977. Colour of the ocean as an indicator in
phtosynthetic studies. Journal du Conseil international pour l'Exploration de la Mer
37:313-316.
Hunter, P. R. 2003. Climate change and waterborne and vector-borne disease. Journal of
Applied Microbiology, Journal of Applied Microbiology Symposium Supplement
94:37S-46S.

7

Kerfoot, W. C., J. W. Budd, S. A. Green, J. B. Cotner, B. A. Biddanda, D. J. Schwab, and H. A.
Vanderploeg. 2008. Doughnut in the desert: Late-winter production pulse in
southern Lake Michigan. Limnology and Oceanography 53:589-604.
Kerfoot, W. C., F. Yousef, S. A. Green, J. W. Budd, D. J. Schwab, and H. A. Vanderploeg. 2010.
Approaching storm: Disappearing winter bloom in Lake Michigan. Journal of Great
Lakes Research 36:30-41.
Kutser, T. 2009. Passive optical remote sensing of cyanobacteria and other intense
phytoplankton blooms in coastal and inland waters. International Journal of Remote
Sensing 30:4401-4425.
Lathrop, R. G., Jr. and T. M. Lillesand. 1986. Use of Thematic Mapper data to assess water
quality in Green Bay and central Lake Michigan. Photogrammetric Engineering and
Remote Sensing 52:671-680.
Lathrop, R. G., T. M. Lillesand, and B. S. Yandell. 1991. Testing the utility of simple multidate Thematic Mapper calibration algorithms for monitoring turbid inland waters.
International Journal of Remote Sensing 12:2045-2063.
Lee, Z., K. L. Carder, C. D. Mobley, R. G. Steward, and J. S. Patch. 1998. Hyperspectral remote
sensing for shallow waters. I. A semianalytical model. Applied Optics 37:6329-6338.
Lesht, B. M., R. P. Barbiero, and G. J. Warren. 2013. A band-ratio algorithm for retrieving
open-lake chlorophyll values from satellite observations of the Great Lakes. Journal
of Great Lakes Research 39:138-152.
Lillesand, T. M., W. L. Johnson, R. L. Deuell, O. M. Lindstrom, and D. E. Meisner. 1983. Use of
Landsat data to predict the trophic stat of Minnesota lakes. Photogrammetric
Engineering and Remote Sensing 49:219-229.
Lodhi, M. A. and D. C. Rundquist. 2001. A spectral analysis of bottom-induced variation in
the colour of Sand Hills lakes, Nebraska, USA. International Journal of Remote
Sensing 22:1665-1682.
Lofgren, B. and A. Gronewold. 2012. Water Resources. Great Lakes Integrated Sciences and
Assessments Center.
Lyzenga, D. R. 1981. Remote sensing of bottom reflectance and water attenuation
parameters in shallow water using aircraft and Landsat data. International Journal
of Remote Sensing 2:71-82.
Mayo, M., A. Gitelson, Y. Z. Yacobi, and Z. Ben-Avraham. 1995. Chlorophyll distribution in
Lake Kinneret determined from Landsat Thematic Mapper data. International
Journal of Remote Sensing 16:175-182.
8

O'Reilly, J. E., S. Maritorena, B. G. Mitchell, D. A. Siegel, K. L. Carder, S. A. Garver, M. Kahru,
and C. McClain. 1998. Ocean color chlorophyll algorithms for SeaWiFS. Journal of
Geophysical Research 103:24,937-924,953.
O'Reilly, J. E., S. Maritorena, M. C. O'Brien, D. A. Siegel, D. Toole, D. Menzies, R. C. Smith, J. L.
Mueller, B. G. Mitchell, M. Kahru, F. P. Chavez, P. Strutton, G. F. Cota, S. B. Hooker, C.
R. McClain, K. Carder, F. Müller-Karger, L. Harding, A. Magnuson, D. Phinney, G. F.
Moore, J. Aiken, K. R. Arrigo, R. Letelier, and M. Culver. 2000. SeaWiFS postlaunch
calibration and validation analyses, Part 3.in S. B. a. E. R. F. Hooker, editor. SeaWiFS
Postlaunch Techinical Report Series. NASA.
Odermatt, D., A. Gitelson, V. E. Brando, and M. Schaepman. 2012. Review of constituent
retrieval in opticaly deep and complex waters from satellite imagery. Remote
Sensing of Environment 118:116-126.
Olmanson, L. G., M. E. Bauer, and P. L. Brezonik. 2008. A 20-year Landsat water clarity
census of Minnesota's 10,000 lakes. Remote Sensing of Environment 112:40864097.
Paerl, H. W. and J. Huisman. 2008. Blooms like it hot. Science 320:57-58.
Paerl, H. W., I. R.S. Fulton, P. H. Moisander, and J. Dyble. 2001. Harmful freshwater algal
blooms, with and emphasis on cyanobacteria. The Scientific World 1:76-113.
Pozdnyakov, D., R. Shuchman, A. Korosov, and C. Hatt. 2005. Operational algorithm for the
retrieval of water qualilty in the Great Lakes. Remote Sensing of Environment
97:352-370.
Rundquist, D. C., J. F. Schalles, and J. P. Peake. 1995. The response of volume reflectance to
manipulated algal concentrations above bright and dark bottoms at various depth in
an experimental pool. Geocarto International 10:5-14.
Schelske, C. L., E. F. Stoermer, D. J. Conley, J. A. Robbins, and R. M. Glover. 1983. Early
eutrophication in the lower Great Lakes: New evidence from biogenic silica in
sediments. Science 222:320-322.
Shapiro, J. 1997. The role of carbon dioxide in the initiation and mantenance of blue-green
dominance in lakes. Freshwater Biology 37:307-323.
Shuchman, R., A. Korosov, C. Hatt, D. Pozdnyakov, J. Means, and G. Meadows. 2006.
Verification and application of a bio-optical algorithm for Lake Michigan using
SeaWiFS: A 7-year inter-annual analysis. Journal of Great Lakes Research 32:258279.
Strong, A. E. 1974. Remote sensing of algal blooms by aircraft and satellite in Lake Erie and
Utah Lake. Remote Sensing of Environment 3:99-107.
9

Strong, A. E. 1978. Chemical whitings and chlorophyll distributions in the Great Lakes as
viewed by Landsat. Remote Sensing of Environment 7:61-72.
Stumpf, R. P. and M. C. Tomlinson. 2005. Remote sensing of harmful algal blooms. Pages
277-296 in R. L. Miller, C.E. Del Castillo, and B.A. McKee, editor. Remote Sensing of
Coastal Aquatic Environments. Springer, Ah Dordrecht, The Netherlands.
Tolk, B. L., L. Han, and D. C. Rundquist. 2000. The impact of bottom brightness on spectral
reflectance of suspended sediments. International Journal of Remote Sensing
21:2259-2268.
USEPA. 2013. Great Lakes Areas of Concern. United States Environmental Protection
Agency.
Vincent, R. K., X. Qin, R. M. McKay, J. Miner, K. Czajkowski, J. Savino, and T. Bridgeman. 2004.
Phycocyanin detection from Landsat TM data for mapping cyanobacterial blooms in
Lake Erie. Remote Sensing of Environment 89:381-392.
Weghorst, P. L. 2008. MODIS algorithm assessment and principal component analysis of
chlorophyll concentration in Lake Erie. Kent State University.
Wynne, T. T., R. P. Stumpf, M. C. Tomlinson, R. A. Warner, P. Tester, J. Dyble, and G. L.
Fahnenstiel. 2008. Relating spectral shape to cyanobacterial blooms in the
Laurentian Great Lakes. International Journal of Remote Sensing 29:3665-3672.

10

CHAPTER 2
NATIONAL CHLOROPHYLL AND SECCHI DETECTION BY LANDSAT 7 ETM+ SATELLITE
IMAGERY FOR WATER QUALITY ASSESSMENT
Abstract
We developed remote sensing models for chlorophyll a (chl) and Secchi depth for
use at national and regional scales with US EPA’s National Lake Assessment data (NLA) and
Landsat 7 ETM+ satellite images. Using linear regression, separate Landsat models for chl
and Secchi depth were developed at the national scale and for each of seven NLA lake
classes to determine if climatic as well as naturally varying physical and chemical attributes
could reduce variation in Landsat models. National boosted regression tree (BRT) models
were also made and compared to national models made with linear regression. For all
models, Landsat bands and band ratios were used to infer chl or Secchi depth, and adjusted
R2 and RMSE were used to select the best models. For the linear regression models, the
national chl model performed better than the lake type models, except for lake type B, and
the national Secchi depth model performed better than all lake types except for lake types F
and G. Overall, the best national Landsat model for ln-transformed chl was the boosted
regression tree model and had a cross-validation R2 of 0.44 and a 0.76 ln-transformed µg/L
RMSE. The best national Landsat model for Secchi depth was also a boosted regression tree
model that had an adjusted R2 of 0.52 and a 0.80 m RMSE. Potential sources of error in the
national chl boosted regression tree model were indicated by residuals being positively
correlated with DOC (adjusted R2=0.14) and cyanobacterial density (adjusted R2=0.18), and
negatively correlated with lake depth (adjusted R2=0.25). Lake area, difference between
image and sample date, and percent cloud cover were not correlated with national chl
boosted regression tree model residuals. National Secchi boosted regression tree model
11

residuals were weakly, negatively correlated with DOC (adjusted R2=0.08) and
cyanobacterial density (adjusted R2=0.06) and weakly, positively correlated with difference
between image and sample dates (adjusted R2=0.02) and lake depth (adjusted R2=0.17).
Percent cloud cover of Landsat image and lake area were not correlated with national
Secchi boosted regression tree model residuals. We assessed the applicability of the
national chl model for ecological analysis by comparing the total phosphorus-chl
relationship with chl determined from sampling or remote sensing, which showed the total
phosphorus-chl relationship had an adjusted R2=0.58 and 1.02 ln-transformed µg/L RMSE
with sampled chl versus an adjusted R2=0.56 and 1.04 ln-transformed µg/L RMSE with chl
determined by the boosted regression tree remote sensing model. Remote sensing water
quality assessment tools can be valuable for limnological study, ecological assessment, and
water resource management.

12

Introduction
Lakes are important sources of water for drinking, agriculture, industrial processes,
recreation, and aesthetic well being. Water quality is highly threatened by withdrawals,
pollution, and climate change-related alterations in precipitation and temperature (Parry
et al. 2007). Toxic or nuisance algal growth in lakes can result from human activities
generating nitrogen and phosphorus pollution (Carpenter et al. 1998) and may be
exacerbated by increasing temperature with climate change (Paerl and Huisman 2008).
Excessive algal biomass causes a variety of problems ranging from decreasing oxygen as
the algae decompose (Wetzel 2001), harboring pathogens (Ishii et al. 2006, Byappanahalli
et al. 2009), and generating toxins that harm aquatic organisms (Carmichael 1996) and
cause sickness or death in terrestrial vertebrates (Briand et al. 2003) and humans (Watson
et al. 2008, Backer et al. 2010). Therefore, developing and refining tools to assess algae at
spatial and temporal scales that will advance ecological understanding for resource
management is imperative as human needs and climate change stress lake ecosystems.
Satellite imagery has been used to assess water quality of freshwater ecosystems for
several decades (Strong 1974, Scarpace et al. 1979, Lillesand et al. 1983, Gitelson et al.
1993, Kloiber et al. 2002a, Bustamante et al. 2009) Remote water quality assessment has
the potential to be more time and cost-effective than traditional sampling techniques
because the cost of acquiring some satellite imagery and time for processing are relatively
low. Remotely sensed imagery also has advantages over more traditional sampling
techniques by incorporating spatial variation over the size of a pixel and providing spatial
detail throughout a lake. In addition, many historic satellite images are available, which
enables temporal study of climate and land use change.
13

Landsat has greater potential for historical analyses than other satellites with nearly
four decades of available data. Perhaps most importantly, unlike other satellites (e.g.,
MODIS, SeaWiFS), Landsat’s small pixel size allows for the assessment of small lakes and
fine resolution of spatial patterns within bigger lakes (Chipman et al. 2004, Han and Jordan
2005, Olmanson et al. 2008, Torbick et al. 2008). The disadvantages of Landsat are the
relatively low number of bands compared to other satellites and relatively wide
bandwidths, which produce some limitations on spectral resolution. However, Landsat
imagery was used as early as Strong (1974) to detect the presence of algae in freshwater
lakes. Since Strong’s study, many models using spectral band measurements from Landsat
images have been developed for inference of chl and Secchi depth using data from
relatively small regions, such as states, and using only a subset of the available Landsat
bands (Kloiber et al. 2002b, Nelson et al. 2003, Chipman et al. 2004) .
While Landsat’s visible wavelengths capture peak absorption and reflectance
wavelengths for chl, they also capture changes in reflectance due to the increased
scattering of light by particles near the water’s surface (Lathrop et al. 1991, Han et al.
1994). Therefore, we might expect the visible wavelength bands to be the most useful for
inferring chl and Secchi depth. However, infrared bands can also capture the scattering
caused by either particles in the surface water (Lavery et al. 1993, Kloiber et al. 2002a), or
by floating biomass, which reflects more light than clear water (Strong 1974). Additionally,
the thermal infrared band can detect increases in surface water temperature that can both
cause increased algal biomass and indicate warming by energy absorption and heat
dissipation by organic and inorganic matter (Wetzel 2001). Furthermore, using bands in
ratios can help minimize the relative influence of atmosphere between images (Jensen
14

2005). Despite this demonstrated utility, many studies have excluded all or most of the
infrared bands in Landsat models of water quality parameters (Brivio et al. 2001, Kloiber et
al. 2002b, Nelson et al. 2003, Chipman et al. 2004, Olmanson et al. 2008).
Most Landsat remote sensing studies have been limited in geographic range, or in
the ranges of chl or Secchi depth they use to develop the model, for example, not covering
both the lower and upper range of chl concentrations found in nature (Baban 1993, Cox et
al. 1998, Allee and Johnson 1999, Giardino et al. 2001, Duan et al. 2007). Often the
geographic ranges of studies are single lakes or states (Carpenter and Carpenter 1983,
Mayo et al. 1995, Brivio et al. 2001, Nelson et al. 2003, Mishra and Mishra 2010). Having
one or a strategically designed set of remote sensing models that could be used throughout
the country would enhance lake studies by providing data at spatial and temporal scales
that practically cannot be achieved by water sampling and also provide a standard method.
The goal of our study was to develop and test remote sensing models of chl and
Secchi depth using Landsat 7 ETM+ imagery for application across large spatial and
temporal scales by using field measurements from the National Lakes Assessment (NLA) of
the United States Environmental Protection Agency (USEPA). We developed national chl
and national Secchi models and compared them to a set of models for different types of
lakes found across the US, using linear regression. National chl and Secchi models were also
made with boosted regression tree statistical techniques and compared to the national
models made with linear regression. In this effort, all Landsat bands and band ratios,
including infrared and thermal, were used in model development. Sources of error, such as
confounding environmental factors and non-linearities of relationships were evaluated.
Finally, to determine if the national remote sensing model for chl was accurate enough to
15

be useful for ecological and management studies, we compared relationships between total
phosphorus (TP) concentrations in lakes and either chl measured with water samples or
chl measured with Landsat imagery.
Methods
Ecological data, satellite imagery, and pixel extraction
Algal and water chemistry data were collected during the USEPA’s NLA from more
than 1000 lakes in the 48 contiguous states (USEPA 2009). Lakes were sampled between
May and October of 2007 (USEPA 2009). Landsat 7 ETM+ images with level 1G processing
were downloaded from the US Geological Survey’s Global Visualization Viewer website
(USGS 2011b) for dates as close as possible to when lakes were sampled. Images were
reprojected to a WGS-84 projection system in Erdas Imagine 9.3. Exact location of lakes
was determined by overlaying a shapefile (made in Arc Map 9.3) of NLA sampling sites on
each Landsat image. Landsat 7 ETM+ images are comprised of seven layers, each layer
representing a different range of wavelengths, called “bands” (Table 2-1). Each layer is
comprised of 30 m2 pixels except for the thermal infrared, which has 60 m2 pixels. Each of
the pixel’s seven layers has an associated, unitless numeric value that represents the
amount of reflectance emitted by those Landsat band wavelengths. Pixel values were
extracted from Landsat images using Erdas Imagine 9.3 software.
Lakes were included in our study if the sample location could be found in an image
without being covered by clouds and images were taken within eight days of the NLA
sample date. We picked eight days because Landsat images are taken every sixteen days.
Thus, the maximum time between sample collection and a Landsat image would be eight
days. Due to a mechanical malfunction, some Landsat 7 ETM+ images contain black stripes
16

in areas where information has been lost. If a sample location coincided with a black stripe,
a nearby pixel (when possible in the same row or column as the sample pixel) was selected.
The same method was used when sample locations were located in a mixed pixel that had
both water and land signatures. We considered using all 1028 lakes in the 2007 NLA,
however, clear Landsat images could not be found for some lakes within a month of the
sample date and so were not used in the analysis. Also, lakes with maximum depth less
than three times Secchi depth were removed due to the possible influence of lake bottom
detection on Landsat images. This criterion allowed for retention of lakes in the dataset
with a range of depth, from shallow to deep. The final dataset had 447 lakes. Chl ranged
from 0.14–684 µg/L. TP ranged from 1-2147 µg/L in these lakes. Secchi depth ranged from
0.04 meters to 13.7 m.
Remote sensing images and data processing
Radiometric calibration, provided by conversion to top of the atmosphere (TOA)
reflectance values, is necessary to account for image differences in sun angle, in Earth-Sun
distance, and in solar irradiance for different wavelengths that occur above the earth’s
atmosphere (Chander et al. 2009). Therefore, after pixel extraction, raw pixel values were
converted to spectral radiance values with the equation:
Radiance =

(LMAX − LMIN )
(pixel digital number) + LMIN
255

Where pixel digital number is the raw pixel value, and LMAX and LMIN correspond to
known spectral radiances for each of the Landsat 7 ETM+ spectral bands (Chander 2009).
Radiance values were further converted to TOA reflectance values with this equation:

Pp =

(Π)(Lλ )(d 2 )
(E λ )(cosθ s )
17

Here Pp is unitless planetary reflectance, Lλ is spectral radiance at the sensor’s aperture, d
is the earth-sun distance in astronomical units based on the image acquisition date
(Chander et al. 2009), and E λ is mean solar exoatmospheric irradiance. Solar zenith ((θs)
values were obtained from NOAA’s Solar Position Calculator (NOAA 2011). High gain
thermal band radiance values were converted directly to temperature (˚C) with this
equation:

K2
-273.15
 K1 
 +1
 Lλ 
Where, T is the effective at-satellite temperature in Kelvin, K2 is calibration constant,
T=

1282.71, K1 is calibration constant, 666.09, and Lλ is the spectral radiance in
watts/((m2)(steradians)(µm)) (Chander 2009).
If the TOA value was 0.3 or greater, pixel values for that lake were excluded from
further analyses due to the likelihood of nearby cloud interference. No additional
atmospheric correction was done due to lack of information on atmospheric conditions for
the sampling date and location.
Linear regression analyses
Natural log transformations were formed on both chl and Secchi depth data to
obtain normal distributions before use in linear regression analyses (Figure 2-1). Linear
regression analyses were run using Landsat’s visible bands, infrared bands (including
thermal infrared), and all possible two-band ratios to determine the best Landsat 7 ETM+
model. It should also be noted that, although including inverse ratios (e.g., 1/3 and 3/1) in
the same model would seem to cancel effects of each other, inverse ratios have different
rates of change along the same band ranges, and therefore have different weights in

18

different band ranges. Linear regression analyses were run, first, with each of the 49
possible band or band ratios, individually, to elucidate the relationship between each band
or ratio and chl or Secchi depth. Testing all bands/ratios also formed a base for comparison
to previous studies that often did not test the same subset of bands or ratios for their
respective predictive power. Bootstrapping was run on each model to assess the
robustness of the model. The bands/ratios that yielded the highest R2 and lowest RMSE
were chosen for multiple linear regression analysis. Multiple linear regression was done
starting with the best predictive bands from the first round of linear regression and adding
all other bands or ratios one at a time. The model was retained if it improved the linear
regression model R2 by at least 0.05. Next, bootstrapping was done on the multiple linear
regression model to determine if R2 value of the multiple linear regression retained the
0.05 increase in predictive power over the linear regression model. None of the
bootstrapped multiple linear regression models improved the R2 linear regression model
by at least 0.05, so no further work was done on multiple linear regression models. We
checked for non-linearities by running a LOESS analysis and visually comparing model fits.
To assess model robustness, we looked for consistencies in band and band ratios used in
the national models, the lake type models, and models in the literature.
We also developed chl and Secchi depth linear regression models by lake type to test
whether grouping lakes by natural features allows for better predictive power. Lake types
were determined by a cluster analysis of variables that are relatively unaffected by human
activities (Herlihy et al. 2013), but do influence the lakes’ ecology (Table 2-2). These
variables included: surface area, maximum depth, morphology, elevation, maximum annual
air temperature in the watershed, annual maximum of monthly precipitation, calcium
19

concentration, latitude, and longitude. Variables were transformed to reduce skewness,
and they were standardized to provide similar weights. The cluster analysis was
constrained by level three ecoregions, which were: Eastern Highlands; Plains and
Lowlands; and Xeric and Mountain West. Seven lake types resulted from the cluster
analysis, labeled A through G. As with national models, all bands and ratios were
considered in the lake type models. We compared national model and lake type models by:
1) comparing the adjusted R2 and RMSE of relationships between chl and Secchi depths
measured by sampling and Landsat imagery using national and lake type specific models,
and 2) comparing the resulting R2 and RMSE values to remote sensing water quality
models in the literature.
Boosted regression tree analyses
To test whether simultaneously incorporating multiple Landsat spectral
bands/ratios and accounting for non-linearities could improve predictive power, national
boosted regression tree models were created using the same Landsat ETM+ pixel data as
the linear regression models and all 49 band and band ratios that can be created from
Landsat ETM+ bands 1-7, rather than previous analyses with one or two bands or ratios (as
with linear or multiple linear regression models). Although response variables should,
generally, not require transformation for use in boosted regression tree analysis, we found
it necessary to ln-transform chl for with this dataset. This may be due to the extreme
skewness of the chl data (Figure 2-1). Secchi depth was left untransformed for boosted
regression tree analysis. Initial boosted regression tree models were simplified using the
gbm.simplify program in R that determines the number of variables that can be removed
without increasing predictive deviance (Elith et al. 2008). Root mean square error (RMSE)
20

values, calculated using predicted and measured chl, were compared for final model
selection. Since boosted regression tree models were already cross-validated with data
internal to the dataset we did cross-validate them a second time with bootstrapping. All
statistics were run in gbm package 2.9 in R. R 2.15.1.
Model diagnostics
An internal 10-fold cross-validation procedure, incorporated into the boosted
regression tree program code, is designed to avoid overfitting the data. The crossvalidation process began with a division of the data into 10 subsets. Ten boosted regression
trees were run, simultaneously. Model training was done on 90% of the data and testing it
on the remaining 10% until each of the 10 subsets has been used as a test dataset. Next, a
new set of 10 boosted regression trees was run and tested with a larger number of trees.
This process was continued until, at a given number of trees, the average predictive power
of the first five boosted regression trees was surpassed in predictive power by the second
set of five boosted regression trees. At the point, the minimum (optimal) number of trees
for the model had been passed. From this, we determined optimal tree number and
continued to refine other model parameters, including learning rate, bag fraction, and tree
complexity. Learning rate, or relative contribution of each tree to the final model, was
determined to be 0.005 for the chl model and 0.011 for the Secchi model; optimal bag
fraction, or amount of data chosen at random from the data set for tree construction, was
determined to be 0.50 for both chl and Secchi models; optimal tree complexity, or
approximate order of the model, was determined to be three for both chl and Secchi
models; number of trees in final boosted regression tree models was 3050 for the chl
model and 1800 for the Secchi model (Table 2-3).
21

To determine if the national boosted regression tree models could be improved with
available information, we analyzed national model residuals to determine if other
ecological variables were confounding the inference of chl or Secchi depth. Non-remotely
sensed predictor variables used in the residuals analysis included dissolved organic carbon
(DOC), cyanobacterial density (cells/cm2), “diff. days” (the number of days between
sampling and Landsat image acquisition), percent cloud cover of the Landsat image, lake
area, and maximum measured lake depth. Any variables that were correlated to national
model residuals were further analyzed to elucidate trends. Partial correlation coefficients
were calculated to determine the relative influence of band ratios vs. depth on chl or Secchi
depth detection. Finally, to determine if the national boosted regression tree models are
useful for ecological and management studies, we used linear regression to relate TP and
actual, measured chl and Landsat-inferred chl. All statistics were run with R, version
2.15.0.
Results
National and lake type models to predict chl
For the linear regression national model, band ratios 1/2 and 1/3 were selected for
the national model (Tables 2-4 and 2-5, Figure 2-2). Bootstrap adjusted R2 values for all
linear regression chl models were about the same as the original model adjusted R2 (Table
2-4), perhaps, due to the simplicity of the model and the size of the training data set. No
two-band/ratio multiple linear regression models had bootstrap R2 values that improved
the R2 of the original linear regression model by more than 0.05. Band ratio 1/2 was also
the best predictor of chl for lake types B, C, and D, while other bands or ratios were best

22

predictors for other lake types. Lake type chl models had lower adjusted R2, but also lower
RMSE values than the national model in all cases, except for lake types B (Table 2-4).
Lake type B yielded the best model, while lake type D yielded the weakest (Table 24). Bands ratio 1/2 was most frequently the best predictor of chl in the lake type ratios and
band 1 was used in at least one of the best predictor ratios in all lake types except for F.
Band 3 was the next most common band used in best predictor band ratios. Band 4, near
infrared, was used in a best predictor ratio in lake type E; band ratio 7/5 was just as good a
predictor of chl as ratio 1/2 in lake type D; band 6, thermal infrared, was the best predictor
in lake type F. Using ratio 1/2 on lake types for which it was not the best predictor, did not
yield substantially poorer results, except for lake type A (Table 2-4). Some of the lake type
models with smaller sample size also had some of the highest R2 and lowest RMSE values
(Table 2-4). Upon visual inspection, the national linear regression chl model did not display
substantial non-linearities.
The national chl boosted regression tree model had the highest R2 and lowest RMSE
of all models (0.44 cross-validation R2 and 0.76 ln-transformed µg/L RMSE; Table 2-3,
Figure 2-3). National chl boosted regression tree model residuals were positively
correlated with DOC (adjusted R2 = 0.14) and cyanobacterial density (adjusted R2 = 0.18),
and negatively correlated with lake depth (adjusted R2 = 0.25). Lake area, difference
between image and sample date, and percent cloud cover were not correlated with national
chl boosted regression tree model residuals. The relationship between TP and chl using
NLA data had an adjusted R2 = 0.58 and 4.90 µg/L RMSE, whereas the relationship between
TP and Landsat inferred chl had an R2 of 0.40 and 5.64 µg/L RMSE.

23

National and lake type models to predict Secchi depth
For the national linear regression Secchi model, band ratio 1/3 produced the best
model, while ratios 3/1, 1/2, and 2/1 produced similarly predictive models (Table 2-5 and
2-6 and, Figure 2-4). Bootstrap adjusted R2 values for all linear regression Secchi depth
models were about the same as the adjusted R2 of the original model (Table 2-6). As with
the chl models, this similarity may be due to the simplicity of the model and the size of the
training data set. As with chl models, no two-band/ratio multiple linear regression models
had bootstrap R2 values that improved the original linear regression model ≤ 0.05. The
best Secchi depth models were for lake types F and G and the poorest in lake types D, E, and
A. Band ratio 1/3 was also the best predictor of chl for lake types C, D, and E, while ratio
3/1 was equally predictive for lake type E. Ratio 2/1 yielded the best model for lake types
B and F. Band 1 was used in all best predictor models for Secchi depth. No infrared bands
were found to be useful in a best predictive model. Lake type models generally had lower
adjusted R2, but higher RMSE values than the national model in all cases, except for lake
types F and G (Table 2-6). Using ratio 1/3 on lake types for which it was not the best
predictor, yielded similar results for best predictive models using band ratio 3/1 and more
poorly for best predictive models using band ratio 2/1 (Table 2-6). All results for national
and lake type models were statistically significant at p≤ 0.05. As with the lake type chl
models, some of the lake type Secchi depth models with smaller sample size also had some
of the highest R2 and lowest RMSE values (Table 2-6). Upon visual inspection, the national
linear regression Secchi depth model did not display any non-linearities.
The national Secchi boosted regression tree model had the highest R2 and lowest
RMSE (0.52 cross-validation R2 and 80 m RMSE; Table 2-6, Figure 2-5) of all models.
24

National Secchi boosted regression tree model residuals were weakly, negatively
correlated with DOC (adjusted R2 = 0.08) and cyanobacterial density (adjusted R2 = 0.06)
and positively correlated with lake area (adjusted R2 = 0.02) and lake depth (adjusted R2 =
0.17). Percent cloud cover of Landsat image and diff. days were not correlated with
national Secchi boosted regression tree model residuals.
Discussion
Evaluation of national chl model
The national chl boosted regression tree model made using a combination of
Landsat bands/ratios performed well based on our evaluation criteria. The adjusted R2 of
our national and lake type models was lower than many other studies, and those studies
were done at smaller scales or with satellite imagery with narrower bandwidths (Table 27). The linear regression model for lake type B also had good performance compared to
others from smaller geographic areas. Therefore, grouping eastern highlands lakes and
reservoirs may be useful for improvement of chl prediction. Alternatively, small RMSE
values and correspondingly high adjusted R2 values of some lake type models may be a
result of a smaller sample size and/or smaller chl range. These results may also indicate
that the sampling structure of the NLA, with one chl sample per lake, cannot adequately
capture the spatial variability needed for model chl development. Having a gradient of chl
within one lake could help to calibrate the model in relation to other lakes attributes, such
as DOC. Chl inferred from Landsat imagery was a good predictor of TP when compared to
measured chl.
Evaluation of national Secchi depth model

25

The national boosted regression tree Secchi depth model made using a combination
of Landsat bands/ratios performed adequately, while the lake type models performed well
based on our criteria, falling well within the performance range seen in the literature for
smaller regions, based on R2 values (Table 2-7). As with chl model results, smaller RMSE
values with corresponding higher adjusted R2 values for Secchi depth lake type models
may, in some cases, be a result of a smaller Secchi depth range. Overall, our national
boosted regression tree model performs well compared to the lake type linear regression
models, and by incorporating spatial variation across the United States, it can be a valuable
tool for larger-scale ecological studies. Interestingly, at least one each of the best predictor
band ratios for national and lake type chl and Secchi models overlapped in best predictor
band ratios, except for lake types C and D. This may be due to the overlapping information
in chl and Secchi depth measurements, as Secchi depth includes algal-related turbidity.
Possible improvement of models
Investigation of possible non-linear trends in the relationship between measured
chl in samples and Landsat inferred chl or measured Secchi depth and Landsat-inferred
Secchi depth were not evident upon visual comparison of linear and LOESS models.
However, testing a variety of boosted regression tree models indicated a third order trend,
which agrees with previous chl models made for MODIS (O'Reilly et al. 2000). Analysis of
national boosted regression tree chl and Secchi model residuals elucidated additional
trends. The correlation between maximum depth and both chl and Secchi depth, and the
partial correlation coefficients suggested that Landsat may detect differences in color
intensity based on the volume of water. For example, it is possible that more chl is detected
in deeper water bodies due to the increased volume of water for algae to inhabit. The fact
26

that depth explained so much variation in chl agrees with the idea that shallower lakes are
more susceptible to algal blooms, due to lower phosphorus processing capacity and higher
water temperatures. A similar explanation can be made for the correlation between
national Secchi model residuals and lake depth. With increasing depth of a lake, there is
more cumulative suspended matter throughout the water column that Landsat may be able
to detect, but that is unaccounted for in field measurements that only represent a small
portion of the water column.
The weak, positive and negative correlation of cyanobacterial density and boosted
regression tree model residuals of the Landsat chl and Secchi depth models, respectively,
was likely driven by differences in the taxonomic composition of algal communities, which
affects algal color because of accessory pigments, such as phycocyanin. Alternatively, this
positive correlation could be due to different cell mass/volume relationships of taxa (Morel
and Bricaud 1981, Gitelson et al. 2000, Wetzel 2001). The fact that there was a correlation
with cyanobacterial density indicates that Landsat’s utility for cyanobacteria density
prediction should be tested.
The weak, positive and negative correlation of DOC to national boosted regression
tree chl and Secchi models, respectively, indicates that the varying levels of reddish-brown
color reflected by DOC may be introducing noise into the signals reaching Landsat. For the
chl model, the positive correlation suggests that the reflectance in the red band that DOC
creates may overwhelm the subtle absorption changes we expect to see detecting chl.
A weak, positive relationship between Secchi boosted regression tree model
residuals and lake area may indicate that surface area-related ecological attributes were
not all account for in the model. This is also suggested by the lack of significance in the
27

relationship between chl boosted regression tree model residuals and lake area. Lack of
correlation between diff. days and boosted regression tree model residuals indicates that
number of days between sampling and image acquisition was not a factor in ability to
predict chl and Secchi. We conclude that 8-day differences in sampling and imagery had
little effect on model development.
Percent cloud cover was not related to either boosted regression tree chl or Secchi
model residuals. Non-normality of percent cloud data, even after transformation, could
have contributed to this result, however, precautions in choosing pixels for data extraction
could have prevented any major cloud interference in chl and Secchi prediction.
Robustness and application of national models
Since each Landsat 7 ETM+ band provides different biological or physical
information about the ecosystem, it is not surprising that the best boosted regression tree
models included all of the visible bands in various combinations as well as infrared bands.
The visible bands represent pigments or colors of matter suspended near the water’s
surface, infrared bands detecting biological structure, and thermal infrared detecting
differences in temperature. Although there are differences in the bands and ratios in
national models compared to those included in the lake type models (Tables 2-4 and 2-5),
the similarities in the linear regression models agree with previous studies (Table 2-7) that
visible bands and sometimes infrared bands alone or in ratios are important for chl and
Secchi depth inference.
The boosted regression tree models, but especially the linear regression national
and lake type models emphasize the importance of the visible bands as part of the chl and
Secchi depth models. As there is variation in the bands used in the literature, there is also
28

variation in exactly which bands used in our national and lake type models. Such variation
probably reflects the variety of algal species or inorganic suspended sediments in the water
in those areas. For example, Lake type D includes relatively shallow lakes, which might
become stratified more quickly. Lake type A, whose lakes are not particularly shallow, but
whose temperature (similarly to lake type D) is significantly higher than other lakes types
might also stratify more quickly, allowing high temperature tolerant phytoplankton species
to thrive, and change overall pigment content or amount of floating biomass on the water’s
surface (Gitelson et al. 2000). Lake type E has a much larger range of DOC than other lake
types which may flood the more subtle changes in red wavelength absorptance and
reflectance of chl or colors of various sediment types. Lake depth, itself, may affect which
bands/ratios are most useful in detecting chl due to change in color at different depths
(Provost et al. 2004). In short, differences in algal pigments, DOC, and sediment types may
explain variation in bands and ratios for each lake type.
Another reason for the variety of models is that the spectral signatures of certain
pigments or particular matter may lie at the edge of Landsat wavelength bands. For
example, Gitelson et al. (2000) found that chl has a subtle absorption feature starting
around 690 nm, which coincides with the upper edge of the red Landsat 7 ETM+ detection
band (Gitelson et al. 2000). In addition, algae have many accessory pigments with
absorption peaks varying across the visible spectrum. Slight differences in environmental
conditions can alter algal physiology, accessory pigment concentration, and color of the
algae (Reynolds 2006).
Summary and concluding remarks

29

The national boosted regression tree models to infer chl and Secchi depth using
Landsat imagery performed well and shows substantial potential for use in ecological
studies. Confounding environmental variables limited the precision of national models, but
future studies could incorporate natural landscape feature that can help predict the
confounding variables. Refinement of national or regional-scale remotely sensed water
quality might include accounting for different algal types, or wind speed near the lake.
Other refinements may also be possible with the launch of the Landsat 8 (Landsat Data
Continuity Mission) that occurred in February, 2013. Landsat 8 includes a second blue
wavelength band, a new shortwave infrared band, and a narrower near infrared band that
better cover the absorption features of chl (USGS 2011a). Continued refinement of remote
sensing tools for characterizing water quality will be valuable for water resource
management.

30

APPENDIX

31

Table 2-1: Landsat 7 ETM+ wavelength band features (USGS 2011a).
Landsat bands

Band wavelength (nm)

Resolution (m)

Blue

450-520

30

Green

520-600

30

Red

630-690

30

Near Infrared

770-900

30

1550-1750

30

10400-12500

60

2090-2350

30

Short Wave Infrared
Thermal Infrared
Short Wave Infrared

32

Table 2-2: US EPA National Lakes Assessment lake types. Lake types, based on cluster
analysis of US EPA National Lakes Assessment water chemistry.
Lake type
A

Lake type description
Eastern highlands: Low to moderate-elevation,
warm reservoir

B

Eastern highlands: Low to moderate-elevation,
northern lakes and reservoirs

C

Northern central plains and lowlands: Lowelevation lakes and reservoirs with small surface
area

D

Southern and coastal plains and lowlands: Lowelevation, shallow, warm lakes and reservoirs with
large surface area

E

Northern central plains and lowlands: Low to
moderate-elevation, shallow lakes and reservoirs

F

North-western mountains: Moderate to highelevation, cold, deep lakes and reservoirs

G

South-western mountains: Moderate to highelevation, cold lakes and reservoirs

33

Table 2-3: National boosted regression tree chl and Secchi depth models. National
boosted regression tree models for ln-transformed chl and untransformed Secchi depth
made with Landsat 7 ETM+ bands and ratios. Units of RMSE are ln-transformed µg/L for
chl and m for Secchi, CV=cross-validation, NT=number of trees in final BRT model, LR=BRT
learning rate, BF=bag fraction, TC=tree complexity.

Model
Ln Chl

N
447

Training
R2
0.78

RMSE
0.76

CV R2
0.44

NT
3050

LR
0.005

BF
0.50

TC
3

Secchi

447

0. 87

0.80

0.52

1800

0.008

0.50

3

34

Table 2-4: National and lake type linear regression chl detection models. National and
lake type chl detection models made with Landsat 7 ETM+ bands and ratios. All results for
national and lake type models were statistically significant at p≤ 0.05. Both national models
using band ratio 1/2 and 1/3 were tested on each lake type, consecutively. Units of RMSE
are ln-transformed µg/L, LT = lake type, NS = not statistically significant, NA = not
applicable.

Best
model
1/2
1/3

Adj
R2
0.18
0.19

RMSE
1.39
1.38

Adj R2
Bootstrap
0.19
0.20

Nat.
Model on
LT RMSE
NA

Model
Nat.

N
447

A

44

2/3
3/1
3/2

0.14
0.14
0.15

1.38
1.38
1.37

0.15
0.15
0.16

1.46, 1.40

B

48

1/2
2/1

0.31
0.31

1.17
1.16

0.33
0.33

1.17, 1.26

C

102

1/2

0.11

1.00

0.12

1.01, 1.04

D

44

1/2
7/5

0.09
0.09

0.94
0.94

0.10
0.09

1.60 1.52

E

102

4/1

0. 24

1.31

0.25

1.54

F

45

6

0.14

1.04

0.15

1.55, 1.60

G

62

1/3

0.17

1.17

0.18

1.47, 1.40

35

Table 2-5: National and lake type linear regression chl and Secchi model coefficients.
BR=band ratios., coeff. = coefficients, int.= intercept.
Model
Chl BR
Chl coeff.
Chl int.
Secchi BR
Secchi coeff.
Secchi int.
National

1/2

-3.64

7.21

1/3

-1.42

5.18

2/3

-3.28

7.19

3/1

6.54

-1.06

3/2

7.50

-2.80

1/2

-4.31

8.12

2/1

9.35

-4.68

C

1/2

-2.51

D

1/2

A

B

E

1/3

1.58

-3.04

3/1

-5.62

3.15

2/1

-6.61

5.31

5.65

1/3

1.38

-2.56

-2.20

6.53

1/3

0.90

-2.30

7/5

-1.62

4.59

4/1

3.26

1.36

1/3

1.53

-3.29

3/1

-4.28

1.95

F

6

0.14

-1.78

2/1

-8.04

6.77

G

1/3

-1.12

3.92

3/1

-4.56

2.90

36

Table 2-6: National and lake type linear regression Secchi depth detection models.
National and lake type Secchi depth detection models made with Landsat 7 ETM+ band
ratios. All results for national and lake type models were statistically significant at p≤ 0.05.
Units of RMSE are ln-transformed meters. LT = lake type, NA = not applicable.

Adj
R2
0.49

RMSE
0.77

Adj R2
Bootstrap
0.49

Nat. Model
on LT
RMSE
NA

Model
Nat.

N
447

Best
model
1/3

A

44

3/1

0.36

0.66

0.36

5.19

B

48

2/1

0.40

0.69

0.41

6.44

C

102

1/3

0.47

0.55

0.47

4.91

D

44

1/3

0.34

0.41

0.34

4.12

E

102

1/3
3/1

0. 35
“

0.77
“

0.35
“

3.38

F

45

2/1

0.65

0.53

0. 63

7.26

G

62

3/1

0.61

0.75

0.60

4.38

37

Table 2-7: Chlorophyll and Secchi depth models from the literature. Previous studies
using various satellites and bands or band ratios to infer chl and Secchi depth. Dashes
indicate duplicate information from the line above.
Geographic
Bands
Year
Authors
Satellite
Chl/Secchi
R2
extent
used

1983

Lillesand et al.

Landsat MSS

Twenty-eight
Minnesota
lakes

Green,
red, nearinfrared

Chl

0.84

Secchi

0.880.94

--

--

--

--

Green,
red, nearinfrared

1983

Carpenter and
Carpenter

Landsat MSS

Lake Burley
Griffin,
Australia

Green,
red, nearinfrared

Chl

0.500.85

1989

Lathrop and
Lillesand

Spot-1

Lake
Michigan,
USA

Nearinfrared

Secchi

0.83

1991

Khorram et al.

Landsat TM

August Bay,
Italy

Blue,
green

Chl

0.84

--

--

--

--

Blue,
green

Secchi

0.83

1992

Mittenzwey et
al.

Ship-based
spectrometer

Several lakes
and rivers in
Germany

Two red
bands

Chl

0.98

1994

Pattiaratchi et
al.

Landsat TM

Cockburn
Sound,
Australia

Blue,
green,
red

Chl

0.730.77

38

Table 2-7 (cont’d)
Year

Authors

Satellite

Geographic
extent

Bands
used

Chl/Secchi

R2

--

--

--

--

Red

Secchi

0.720.78

1995

Gitelson et al.

Landsat TM
and
ship-based
radiometer

Haifa Bay,
Israel

Blue,
green,
red

Chl

0.370.93

1995

Mayo et al.

Landsat TM

Lake
Kinneret,
Israel

Blue,
green,
red

Chl

0.49

2000

Gitelson et al.

LI-1800,
Ocean Optics
ST1000, ASD
spectrometers

Several water
bodies in
Israel and
several lakes
in Iowa, USA

Two red
bands

Chl

0830.96

2001

Brivio et al.

Landsat TM

Lake Garda,
Italy

Blue,
green,
red

Chl

0.680.82

2001

Giardino et al.

Landsat TM

Lake Iseo,
Italy

Blue,
green

Chl, Secchi

2001

Pulliainen et
al.

Airborne
imaging
spectrometer

Eleven lakes
in Finland

Two red
bands

Chl

0.410.97

Landsat MSS

Five hundred
Minnesota
lakes

Blue,
green

Secchi

0.600.79

2002b

Kloiber et al.

39

Table 2-7 (cont’d)
Year

Authors

Satellite

Geographic
extent

Bands
used

Chl/Secchi

R2

--

--

Landsat TM

Five hundred
Minnesota
lake

Blue, red

Secchi

0.720.93

2003

Nelson et al.

Landsat ETM+

State of
Michigan

Blue, red

Secchi

0.430.82

2004

Chipman et al.

Landsat TM
and ETM+

State of
Wisconsin

Blue, red

Secchi

0.420.88

Blue,
green,
red, near
infrared

Chl

0.88

2005

Brezonik et al.

Landsat TM

Fifteen
Minnesota
lakes

2008

Kabbara et al.

Landsat ETM+

Coast of
Tripoli

Blue,
green,
red

Chl

0.72

--

--.

--

--

Blue,
green

Secchi

0.54

2008

Olmanson et
al.

Landsat MSS,
TM, and ETM+

State of
Minnesota

Blue, red

Secchi

0.710.95

ASTER

Lake
Beysehir,
Turkey

Green,
red,
nearinfrared

Chl

0.86

2009

Nas et al.

40

Table 2-7 (cont’d)

2010

Mishra and
Mishra

MODIS

Lake
Pontchatrain,
USA

41

Green,
red

Chl

0.65

100
60
0

20

40

Frequency

80

300
200
100
0

Frequency

0

100

300

500

700

-2

2

4

6

Ln-Transformed Chlorophyll (ln(µg/L))

60
20

40

Frequency

150
100

0

0

50

Frequency

200

80

Chlorophyll Concentration (µg/L)

0

0

2

4

6

8

10

12

14

-3

Secchi Depth (m)

-2

-1

0

1

2

3

Ln-Transformed Secchi Depth (ln(m))

Figure 2-1: Histograms of untransformed and ln-transformed chlorophyll and Secchi
depth.

42

6
4
2
0
-2

RS Predicted Chlorophyll ln(µg/L)

-2

0

2

4

6

Measured Chlorophyll ln(µg/L)

Figure 2-2: Remotely sensed (linear regression predicted) ln-transformed
chlorophyll a using band ratio 1/3 versus measured ln-transformed chlorophyll a
(R2adj=0.19, 1.38 ln-transformed µg/L). Line indicates model’s fit.

43

6
5
4
3
2
0

1

RS Predicted Chlorophyll ln(µg/L)

-2

0

2

4

6

Measured Chlorophyll ln(µg/L)

Figure 2-3: Remotely sensed (boosted regression tree predicted) natural lntransformed chlorophyll versus measured ln-transformed chlorophyll a (R2adj=0.44,
0.76 ln-transformed µg/L). Line indicates model fit.

44

3
2
1
0
-2

-1

RS Predicted Secchi ln(m)

-2

-1

0

1

2

3

Measured Secchi ln(m)

Figure 2-4: Remotely sensed (linear regression predicted) natural ln-transformed
Secchi depth using band ratio 1/3 versus measured ln-transformed Secchi depth.
(R2adj=0.49, 0.77 ln-transformed m RMSE). Line indicates model’s fit.

45

10
8
6
4
0

2

RS Predicted Secchi Depth (m)

0

2

4

6

8

10

12

14

Measured Secchi Depth (m)

Figure 2-5: Remotely sensed (boosted regression tree predicted) Secchi depth versus
measured untransformed Secchi depth (R2adj=0.52, 0.80 m RMSE). Line indicates
model’s fit.

46

LITERATURE CITED

47

LITERATURE CITED

Allee, R. J. and J. E. Johnson. 1999. Use of satellite imagery to estimate surface chlorophyll a
and Secchi disc depth of Bull Shoals Reservoir, Arkansas, USA. International Journal
of Remote Sensing 20:1057-1072.
Baban, S. M. J. 1993. Detecting water quality parameters in the Norfolk Broads, U.K., using
Landsat imagery. International Journal of Remote Sensing 14:1247-1267.
Backer, L. C., S. V. McNeel, T. Barber, B. Kirkpatrick, C. Williams, M. Irvin, Y. Zhou, T. B.
Johnson, K. Nierenberg, M. Aubel, R. LePrell, A. Chapman, A. Foss, S. Corum, V. R. Hill,
S. M. Kieszak, and Y.-S. Cheng. 2010. Recreational exposure to microcystins during
algal blooms in two California lakes. Toxicon 55:909-921.
Briand, J.-F., S. Jacquet, C. Bernard, and J.-F. Humbert. 2003. Health hazards for terrestrial
vertebrates from toxic cyanobacteria in surface water ecosystems. Veterinary
Research 34:361-377.
Brivio, P. A., C. Giardino, and E. Zilioli. 2001. Determination of chlorophyll concentration
changes in Lake Garda using an image-based radiative transfer code for Landsat TM
images. International Journal of Remote Sensing 22:487-502.
Bustamante, J., F. Pacios, R. Diaz-Delgado, and D. Aragonés. 2009. Predictive models of
turbidity and water depth in the Doñana marshes using Landsat TM and ETM+
images. Journal of Environmental Management 90:2219-2225.
Byappanahalli, M. N., R. Sawdey, S. Ishii, D. A. Shively, J. A. Ferguson, R. L. Whitman, and M. J.
Sadowsky. 2009. Seasonal stability of Cladophora-associated Salmonella in Lake
Michigan watersheds. Water Research 43:806-814.
Carmichael, W. W. 1996. Toxic Microcystis in the environment. Pages 1-11 in M. F. Wantabe,
K. H. Harada, W. W. Carmichael, and H. Fujiki, editors. Toxic Microcystis. CRC Press,
Boca Raton, FL.
Carpenter, D. J. and S. M. Carpenter. 1983. Modeling inland water quality using Landsat
data. Remote Sensing of Environment 13:345-352.
Carpenter, S. R., N. F. Caraco, D. L. Correll, R. W. Howarth, A. N. Sharpley, and V. H. Smith.
1998. Nonpoint pollution of surface waters with phosphorus and nitrogen.
Ecological Applications 8:559-568.

48

Chander, G., B. L. Markham, and D. L. Helder. 2009. Summary of current radiometeric
calibration coefficients for Landsat MSS, TM, ETM+, and EO-1 ALI sensors. Remote
Sensing of Environment 113:893-903.
Chipman, J. W., T. M. Lillesand, J. E. Schmaltz, J. E. Leale, and M. J. Nordheim. 2004. Mapping
lake water clarity with Landsat images in Wisconsin, U.S.A. Canadian Journal of
Remote Sensing 30:1-7.
Cox, R. M., R. D. Forsythe, G. E. Vaughan, and L. L. Olmsted. 1998. Assessing water quality in
Catawba River reservoirs using Landsat Thematic Mapper satellite data. Lake and
Reservoir Management 14:405-416.
Duan, H., Y. Zhang, B. Zhang, K. Song, and Z. Wang. 2007. Assessment of chlorophyll-a
concentration and trophic state for Lake Chagan using Landsat TM and field spectral
data. Environmental Monitoring and Assessment 129:295-308.
Elith, J., J. R. Leathwick, and T. Hastie. 2008. A working guide to boosted regression trees.
Journal of Animal Ecology 77:802-813.
Giardino, C., M. Pepe, P. A. Brivio, P. Ghezzi, and E. Zilioli. 2001. Detecting chlorophyll,
Secchi disk depth and surface temperature in a sub-alpine lake using Landsat
imagery. The Science of the Total Environment 268:19-29.
Gitelson, A., G. Garbuzov, F. Szilagyi, K.-H. Mittenzwey, A. Karnieli, and A. Kaiser. 1993.
Quantitative remoe sensing methods for real-time monitoring of inland waters
quality. International Journal of Remote Sensing 14:1269-1295.
Gitelson, A., Y. Z. Yacobi, J. F. Schalles, D. C. Rundquist, L. Han, R. Stark, and D. Etzion. 2000.
Remote estimation of phytoplankton density in productive waters. Archiv für
Hydrobiologie Special Issues in Advanced Limnology 55:121-136.
Han, L. and K. J. Jordan. 2005. Estimating and mapping chlorophyll-a concentration in
Pensacola Bay, Florida using Landsat ETM+ data. International Journal of Remote
Sensing 26:5245-5254.
Han, L., D. C. Rundquist, L. L. Liu, R. N. Fraser, and J. F. Schalles. 1994. The spectral response
of algal chlorophyll in water with varying levels of suspended sediment.
International Journal of Remote Sensing 15:3707-3718.
Herlihy, A. T., N. C. Kamman, J. C. Sifneos, D. Charles, M. D. Enache, and R. J. Stevenson. 2013.
Using multiple approaches to develop nutrient criteria for lakes in the conterminous
USA. Freshwater Science 32:367-384.
Ishii, S., T. Yan, D. A. Shively, M. N. Byappanahalli, R. L. Whitman, and M. J. Sadowsky. 2006.
Cladorphora (Chlorophyta) spp. harbor human bacterial pathogens in nearshore
water of Lake Michgian. Applied and Environmental Microbiology 72:4545-4553.
49

Jensen, J. R. 2005. Introductory Digital Image Processing. A Remote Sensing Perspective.
Prentice Hall, Upper Saddle River, New Jersey.
Kloiber, S. M., P. L. Brezonik, and M. E. Bauer. 2002a. Application of Landsat imagery to
regional-scale assessments of lake clarity. Water Research 36:4330-4340.
Kloiber, S. M., P. L. Brezonik, L. G. Olmanson, and M. E. Bauer. 2002b. A procedure for
regional lake water clarity assessment using Landsat multispectral data. Remote
Sensing of Environment 82:38-47.
Lathrop, R. G., T. M. Lillesand, and B. S. Yandell. 1991. Testing the utility of simple multidate Thematic Mapper calibration algorithms for monitoring turbid inland waters.
International Journal of Remote Sensing 12:2045-2063.
Lavery, P., C. Pattiaratchi, A. Wyllie, and P. Hick. 1993. Water quality monitoring in
estuarine waters using the Landsat Thematic Mapper. Remote Sensing of
Environment 46:268-280.
Lillesand, T. M., W. L. Johnson, R. L. Deuell, O. M. Lindstrom, and D. E. Meisner. 1983. Use of
Landsat data to predict the trophic state of Minnesota Lakes. Photogrammetric
Engineering and Remote Sensing 49:219-229.
Mayo, M., A. Gitelson, Y. Z. Yacobi, and Z. Ben-Avraham. 1995. Chlorophyll distribution in
Lake Kinneret determined from Landsat Thematic Mapper data. International
Journal of Remote Sensing 16:175-182.
Mishra, D. R. and S. Mishra. 2010. Plume and bloom: Effect of the Mississippi River
diversion on the water quality of Lake Pontchartrain. Geocarto International
25:555-568.
Nelson, S. A. C., P. A. Soranno, K. S. Cheruvelil, S. A. Batzli, and D. L. Skole. 2003. Regional
assessment of lake water clarity using satellite remote sensing. Journal of Limnology
62:27-32.
NOAA. 2011. Solar Position Calculator National Oceanic and Atmospheric Administration.
O'Reilly, J. E., S. Maritorena, M. C. O'Brien, D. A. Siegel, D. Toole, D. Menzies, R. C. Smith, J. L.
Mueller, B. G. Mitchell, M. Kahru, F. P. Chavez, P. Strutton, G. F. Cota, S. B. Hooker, C.
R. McClain, K. Carder, F. Müller-Karger, L. Harding, A. Magnuson, D. Phinney, G. F.
Moore, J. Aiken, K. R. Arrigo, R. Letelier, and M. Culver. 2000. SeaWiFS postlaunch
calibration and validation analyses, Part 3.in S. B. a. E. R. F. Hooker, editor. SeaWiFS
Postlaunch Techinical Report Series. NASA.

50

Olmanson, L. G., M. E. Bauer, and P. L. Brezonik. 2008. A 20-year Landsat water clarity
census of Minnesota's 10,000 lakes. Remote Sensing of Environment 112:40864097.
Paerl, H. W. and J. Huisman. 2008. Blooms like it hot. Science 320:57-58.
Provost, J.-N., C. Collet, P. Rostaing, P. Pérez, and P. Bouthemy. 2004. Hierarchial Markovian
segmentation of multispectral images for the reconstruction of water depth maps.
Computer Vision and Image Understanding 93:155-174.
Reynolds, C. S. 2006. Ecology of Phytoplankton. Cambridge University Press, Cambridge,
United Kingdom.
Scarpace, F. L., K. W. Holmquist, and L. T. Fisher. 1979. Landsat analysis of lake quality.
Photogrammetric Engineering and Remote Sensing 45:623-633.
Strong, A. E. 1974. Remote sensing of algal blooms by aircraft and satellite in Lake Erie and
Utah Lake. Remote Sensing of Environment 3:99-107.
Torbick, N., F. Hu, J. Zhang, J. Qi, H. Zhang, and B. Becker. 2008. Mapping chlorophyll-a
concentrations in West Lake, China using Landsat 7 ETM+. Journal of Great Lakes
Research 34:559-565.
USEPA. 2009. National Lakes Assessment: A collaborative survey of the nation's lakes.in O.
o. W. a. O. o. R. a. D. Environmental Protection Agency, editor., Washington, D.C.
USGS. 2011a. Landsat Missions. United States Geological Survey.
USGS. 2011b. US Geological Survey Gloval Visualization Viewer United States Geological
Survey.
Watson, S. B., J. Ridal, and G. L. Boyer. 2008. Taste and odour and cyanobacterial toxins:
impairment, prediction, and management in the Great Lakes. Canadian Journal of
Fisheries and Aquatic Sciences 65:1779-1796.
Wetzel, R. G. 2001. Limnology: Lake and Reservoir Ecosystems. Third edition. Academic
Press, New York, New York.

51

CHAPTER 3
MODIS AND LANDSAT BOOSTED REGRESSION TREE CHLOROPHYLL PREDICTIVE
MODELS FOR THE GREAT LAKES
Abstract
Numerous models for predicting chlorophyll (chl) based on satellite images have
been tested on the Great Lakes, and some have been built specifically for the Great Lakes.
However many of these models either lack the ability to predict chl or are technically
complicated or expensive to employ. Also, few studies have taken advantage of archived
satellite images to assess spatial and temporal dynamics over chlorophyll during past
decades. Boosted regression tree (BRT) models of chl using satellite imagery are both easy
to use and can have high predictive performance. Landsat and MODIS imagery are
complementary platforms because of the long history of Landsat operation and the finer
spectral resolution and image frequency of MODIS. Three BRT models were created: 1) a
MODIS model; 2) a Landsat ETM+ and TM model; and 3) a Landsat MSS model. MODIS BRT
model predicted chl most accurately of the three models and compared well to other
models in the literature. BRT models for Landsat ETM+ and TM more accurately predicted
chl than the MSS model and all Landsat models had favorable results when compared to the
literature. BRT chl predictive models are useful in helping to understand historical, longterm chl trends and to inform us of how climate change may alter ecosystems in the future.
Introduction
Early water quality assessments were done by physically taking samples. However,
advancements in remote sensing technology and development of modeling tools in recent
decades have allowed ecological assessment from a distance, enabling evaluation from
larger spatial and temporal scales and providing a more comprehensive view at potentially
52

less cost. Historically, the two main approaches to remote sensing chl predictive model
development in the Great Lakes have been: 1) deriving empirical relationships between chl
and satellite bands or band ratios, alone or separately from other color producing agents
using linear or polynomial regression analyses (Carpenter and Carpenter 1983, O'Reilly et
al. 1998, Giardino et al. 2001) or 2) deriving relationships between two or more color
producing agents (including chl) at once (Carder et al. 2004, Pozdnyakov et al. 2005). Many
of the current chl models are either not widely applicable, have not been tested for wide
use, or are challenging to implement without expert statistical and programming
knowledge, or without costly computer programs. One model-building approach that has
potential to alleviate these challenges is boosted regression tree (BRT) analyses, which is
relatively easy and fast to run with free software, the R program. Code for models can be
easily shared and applied with models to output from standard remote sensing software.
BRT analysis is a machine learning approach for building models (De'ath 2007, Elith
et al. 2008) with advantages of both regression trees and stochastic boosting. In regression
tree analysis, models are fit by recursive partitioning of the response variable, such as chl,
into the two most homogenous groups possible at each split in a way that maximally
reduces error in chl prediction. Regression trees are able to simultaneously handle
predictor variables of different types and scales, they can function in the presence of
outliers and missing data, they can model non-linear relationships and interactions
between predictors, and they do not require that the response variable is normally
distributed (Elith et al. 2008). Boosting is the process of iteratively fitting the regression
tree models to explain the largest error in residuals, in contrast to common modeling
approaches where only one best model is created and chosen (Friedman 2002). Stochastic
53

boosting can decrease overfitting the data and increase predictive power in the final model
by randomly selecting subsets of training data to use in each stage of the tree fitting
process. The combination of regression tree analyses and boosting further increases
predictive power (Elith et al. 2008). In addition, BRT analysis is relatively easy to perform
and apply to satellite imagery. Results of BRT models can be used alone or in conjunction
with other models depending on the study goals. While BRT models often have increased
predictive performance, they can be harder to interpret due to their complexity, causing
them to be more “black box” in nature when compared to simple linear or polynomial
regression models. For both increased predictive performance and interpretability, a
multimodel approach could be employed, combining BRTs and more simple regression
models.
Our research objectives were to develop and test boosted regression tree (BRT)
predictive models for chl in the Laurentian Great Lakes using Landsat and MODIS imagery
and chl measured in water samples from 2007-2009 by the Great Lakes National Program
Office. Landsat and MODIS imagery are complementary platforms due to the fine spatial
resolution and long history of Landsat operation and the finer spectral resolution and
image frequency of MODIS. To take advantage of the strengths of both satellites, we
developed separate models for Landsat and MODIS. A total of two models were made for
Landsat due to significant differences in number and location of spectral bands of Landsat
ETM+ and TM compared to the spectral bands of Landsat MSS. The three models developed
include: 1) a MODIS model; 2) a Landsat ETM+ and TM model; and 3) a Landsat MSS model.
In this effort, all MODIS and Landsat bands and band ratios, including infrared and thermal,
were used in model development. Models were validated by k-fold cross validation. Root
54

mean square error (RMSE) was the final determinant of model function. BRT models were
created and compared to algorithms from the literature. BRT models were applied to
historical Landsat (1972-2012) and MODIS (2000-2012) images for long-term analysis of
chl concentration in the Great Lakes. Historical trend data can also help elucidate how algal
biomass may change in the Great Lakes due to climate change.
Methods
Study locations and ecological data
Study locations include the five Great Lakes, with a focus on three targeted areas:
Saginaw Bay (43˚55’N 83˚35’W), Grand Traverse Bay (43˚3’53.24”N, 83˚29’9.56”W), and
the eastern shore of Lake Michigan, near the Grand River mouth (43˚3’47”N 86˚13’42”).
Great Lakes chl data were collected by the Great Lakes National Program Office (GLNPO) in
August and September of 2007-2009 across the Great Lakes. Data for the MODIS model
ranged from 7.2-306.1 m in sample station depth and 0.31-12.64 µg/L chl from 226 chl
observations. Data for the Landsat ETM+ and TM model and the Landsat MSS model ranged
from 7.9-306.1 m in sample station depth and 0.29-12.64 µg/L chl from 177 chl
observations.
MODIS satellite image processing and pixel extraction for BRT model training
MOD021km (1 km resolution) images, available free of charge, were downloaded
from http://ladsweb.nascom.nasa.gov (Masuoka and Horrocks 2010) and converted to
spectral reflectance using the MODIS Conversion Toolkit (MCTK). Images were chosen
based on proximity of image date to GLNPO sample date. When sampling stations were
located on pixels covered by clouds or cloud shadows, an image from a previous or
subsequent day would be assessed. This process continued until a cloud-free pixel was
55

found. Number of days between image capture and sample collection ranged from 0-6 days.
A shapefile containing GLNPO chl sample stations was created in ArcMap 9.3 and used in
conjunction with MODIS images in Erdas Imagine 9.3 to extract pixels values closest to the
sample station from bands 1-16 (for a total of 18 bands, with bands 13 and 14 each having
a low and high gain bands, Table 3-1).
Landsat image processing and pixel extraction for BRT model training
Landsat 7 ETM+ (Table 3-2) images with level L1T processing were downloaded
from the US Geological Survey (USGS) Global Visualization Viewer website:
http://glovis.usgs.gov/ (USGS 2013c) for dates as close as possible to when lakes were
sampled. Exact location of lakes was determined by overlaying a shapefile (made in Arc
Map 9.3) of GLNPO sampling sites on each Landsat image. When sampling stations were
located on pixels covered by clouds or cloud shadows, an image from a previous or
subsequent image would be assessed. This process continued until a cloud-free pixel was
found. Number of days between image date and sample date ranged from 0-19 days for
Landsat training data. Due to mechanical malfunction, some ETM+ images contain black
stripes in areas causing loss of information. If a sample location coincided with a black
stripe, a nearby pixel (when possible in the same row or column as the sample pixel) was
selected. The same method was used when sample locations were located in a mixed pixel
that had both water and land signatures.
Radiometric calibration, provided by conversion to top of the atmosphere (TOA)
reflectance values, is necessary to account for image differences in sun angle, in Earth-Sun
distance, and in solar irradiance for different wavelengths that occur above the earth’s
atmosphere (Chander et al. 2009). Therefore, after pixel extraction, raw pixel values were
56

converted to spectral radiance values with the following equation:
Radiance =

(LMAX − LMIN )
(pixel digital number) + LMIN
255

Where pixel digital number is the raw pixel value, and LMAX and LMIN correspond to
known spectral radiances for each of the Landsat 7 ETM+ spectral bands (Chander et al.
2009). Radiance values were further converted to TOA reflectance values with this
equation:

(Π)(Lλ )(d 2 )
Pp =
(E λ )(cosθ s )
Here Pp is unitless planetary reflectance, L is spectral radiance at the sensor’s aperture, d is
the earth-sun distance in astronomical units based on the image acquisition date (Chander
et al. 2009), and Eλ is mean solar exoatmospheric irradiance. Solar zenith ((s) values were
obtained from NOAA’s Solar Position Calculator at:
http://www.esrl.noaa.gov/gmd/grad/solcalc/ (NOAA 2013). High gain thermal band
radiance values were converted directly to temperature (˚C) with this equation:

T=

K2
-273.15
 K1 
 +1
 Lλ 

Where, T is the effective at-satellite temperature in Kelvin, K2 is calibration constant,
1282.71, K1 is calibration constant, 666.09, and L is the spectral radiance in watts/(m2 μm
steridians)(Chander et al. 2009).
If the TOA value was 0.3 or greater, pixel values for that lake were excluded from
further analyses due to the likelihood of nearby cloud interference. No additional
atmospheric correction was done due to lack of information on atmospheric conditions for

57

the sampling date and location.
Data preparation and statistics for MODIS and Landsat data
To avoid the effects of lake bottom reflectance, stations for which depth was less
than twice their Secchi depth were removed from the training data set before performing
statistical analyses. For stations where Secchi was not available, it was estimated using
Carlson’s (1977) Secchi-chl relationship:
‫݊ܮ‬ሺ݄ܵ݁ܿܿ݅ሻ ൌ െ0.68 ‫݊ܮ כ‬ሺ݄݈ܿሻ ൅ 2.04
Pixel value data were analyzed in separate BRT analyses for MODIS, for Landsat ETM+ and
TM, and for Landsat MSS. For the Landsat MSS model, three of the ETM+ bands from the
training data were used that corresponded most closely with MSS bands (Table 3-3). All
statistics were run in gbm package 2.9 in R 2.15.1. For each BRT, all available visible and
infrared bands, including thermal infrared were used, along with all possible band ratios. A
total of 49 bands and ratios were used in BRT analyses for Landsat ETM+ and TM, ten
bands and ratios were used for Landsat MSS, and 324 were used for MODIS. Although
including inverse ratios (e.g., 1/3 and 3/1) in the same model would seem to introduce
redundancy into the model, inverse ratios have different rates of change along the same
band ranges, and therefore have different weights in high and low ranges of band ratios.
Also, since each waveband covers a slightly different part of the electromagnetic spectrum
and because chl detection is influenced by algal community structure and time of day
sampled (Owens et al. 1980, Lebouteiller and Herbland 1982, Reynolds 2006), even bands
in the same general category, such as “red bands,” can give us additional information about
the presence of algae.
Bag fractions (the percent of data taken from the training data set, at random) for
58

every BRT model were set to 0.5 and, learning rate, which controls the contribution of each
tree to the model (Elith et al. 2008), was set to 0.01 for every model. Initial BRT models
were simplified by looking at the partial correlation plots of each band or ratio and
measured chl. Final BRT models were chosen using 10-fold internal cross validation
procedure, incorporated into the gbm package. Top bands or ratios were kept, the number
of which was based on the total number of data rows, so that the ratio of observations
(rows of data) to predictors (bands/ratios) was no lower than 10:1 (Peng et al. 2002,
Babyak 2004), which reduced the chance of overfitting the data. Root mean square error
(RMSE) values, calculated between predicted and measured chl, were compared for final
model selection. For the final MSS BRT model, measured (not remotely sensed) water
depth was used as a predictor variable in conjunction with Landsat bands and ratios for
better chl predictive performance.
MODIS image processing for model application
MODIS MOD02L2 images (1km resolution) were downloaded
http://ladsweb.nascom.nasa.gov (Masuoka and Horrocks 2010), covering July 20 September 6, 2000 to 2012 across the Great Lakes. Level 1B of MOD02L2 images were
selected because they contain calibrated and geolocated at-aperture radiances for bands 116, generated from MODIS Level 1A sensor counts (MOD01). Spectral radiance, in units of
watts/(m2 μm steridians), was calculated for MOD02L2 bands 1-16 using the MODIS
Conversion Toolkit (MCTK) and the images were mosaicked in the same date to generate
the daily images at 1 km spatial resolution.
MODIS MOD09A1 images (500 m resolution) were also downloaded
http://e4ftl01.cr.usgs.gov/MOLT (USGS 2013b), covering July 20 - September 6, 2000 to
59

2012 across the Great Lakes. After processing, MOD09A1 images were reprojected, and
converted to the GeoTIFF format using the MODIS Reprojection Tool (MRT) to provide the
day of year (DOY) and cloud images. Since MOD09A1 images only include bands 1-7, we did
not use the MOD09A1 8-day composite image product for our study. Instead, metadata
from MOD09A1 images were used to create 8-day composite images from MOD02L2
images. We assumed that that the best possible TOA reflectance of MOD02L2 bands 1-16
during 8-day composite periods was the same as for the MOD09A1 product. For each pixel,
the best possible observation of spectral radiance obtained during the 8-day composite
period was assumed to be the DOY provided by MOD09A1. The criteria for selecting the
DOY of MOD09A1 is based on high observation coverage, low view angle, absence of clouds,
shadow, and aerosol loading, and minimum value of blue band (Toller et al. 2006). This
compositing method minimized cloud effect as well as view angle, bi-directional effects,
and other atmospheric effects. DOY and corresponding name of the final MOD02L2 8-day
composite images, therefore, represent the first day of the 8-day period when the best
observation is recorded. Cloud masking information, also found in the MOD09A1 product,
was used to perform cloud masking of 8-day composite MOD02L2 images before applying
the BRT chl model.
Landsat image processing for model application
Landsat 7 ETM+, Landsat 4 and 5 TM, and Landsat 1-3 MSS images (Tables 3-2 and
3-3) with level L1T or L1G processing were downloaded from the US Geological Survey
(USGS) Global Visualization Viewer website: http://glovis.usgs.gov/ (USGS 2013c) or the
USGS Earth Explorer website: http://earthexplorer.usgs.gov/ (USGS 2013a). Landsat
images were processed from digital number to TOA reflectance (Chander et al. 2009) using
60

Python 2.6. Images downloaded at L1G processing were georectified using four control
points with RMSE of less than 30 m. The reference layer was Landsat TM image with 0%
cloud cover and quality of 9. Cloud mask algorithm (Oreopoulos et al. 2011) was performed
for all Landsat TM and ETM+ images. This method was modified from a cloud/clear
masking algorithm initially developed for MODIS clear-image compositing. This cloud
masking technique applies a simple threshold scheme by selection four bands (band 1, 3, 4,
and 5) in Landsat TM and ETM+ to remove cloud pixels.
Model evaluation
When possible, we tested models in the literature with our data to standardize the
evaluation process (Lathrop and Lillesand 1986, O'Reilly et al. 2000, Lesht et al. 2013).
When either the full algorithm was not available to test with our data (Shuchman et al.
2006, Becker et al. 2009) or we did not have measurements of other color producing agents
to use with dual retrieval approaches (Binding et al. 2012), we based the performance of
our model on reported R2 or RMSE values from those models.
Results
BRT models and partial correlation plots
The final MODIS BRT model had 27 predictors, 0.10 µg/L RMSE and 0.85 crossvalidation R2 (Table 3-4, Figure 3-1). Chl was accurately predicted through the range of
measured chl values (Figure 3-1). As previously shown (Table 3-1), MODIS has multiple
bands that fall into each general band category. Many of the bands were used in the final
model, and some of those bands are from the same general band category. For example,
1/13H is a band ratio comprised of two red bands. MODIS band ratios with the most
predictive power included blue/green, green/blue, red/red and red/blue, using bands 1,10,
11, and 13H (Figure 3-2). The MODIS BRT model had a lower RMSE and higher R2 than
61

other MODIS Great Lakes chl models (Table 3-5).
The final Landsat ETM+ and TM BRT model had 17 predictors, 0.55 µg/L RMSE and
0.69 cross-validation R2 (Table 3-4, Figure 3-3). Chl was accurately predicted between 0-8
µg/L chl, above which higher measured chl are inaccurately estimated to be approximately
8 µg/L. Bands and ratios with the most predictive power included short wave infrared
(SWIR), green/blue, blue/green, thermal, and blue, using bands 1, 2, 6, and 7 (Figure 3-4).
The Landsat ETM+ and TM BRT model had a lower RMSE and higher R2 than other Landsat
Great Lakes chl models (Table 3-5).
The final Landsat MSS BRT model had 10 predictors, 0.73 µg/L RMSE and 0.66
cross-validation R2 (Table 3-4, Figure 3-5). As with the ETM+ and TM model, chl is fairly
accurately predicted up until about 8 µg/L chl, above which measured chl are inaccurately
predicted around 8 µg/L. Variables that had the most predictive power in this model
included depth, green, and red; bands 2 and 3 (Figure 3-6). This is the first Landsat MSS chl
predictive model created for the Great Lakes.
Overall, MODIS was able to detect chl more accurately than Landsat, and Landsat
ETM+ and TM models outperformed MSS models. Green/blue and blue/green band ratios
were the most common predictors of chl, however the most powerful predictor of chl was
different for each of the three models, including blue/green, SWIR, and depth (Figures 3-2,
3-4, 3-6, 3-8, 3-10). The blue/green and green/blue ratios were commonly found among
the top predictors, except for the MSS models. MSS does not contain a blue band.
Interestingly, the red band was a common variable in MODIS and MSS models, but not in
the Landsat ETM+ and TM models.

62

Discussion
To date, our BRT models are the some of the most accurate chl predictive models for
the Great Lakes (Table 3-5). Both MODIS and Landsat BRT models were able to predict chl
well compared to Great Lakes models in the literature as indicated by R2 and RMSE values
(Table 3-5). Bands and ratios that are often used by models in the literature were also
found among the top predictors in the BRT models, including blue-green ratios (O'Reilly et
al. 2000, Lesht et al. 2013), red-green ratios (Lillesand et al. 1983), and red-blue ratios
(Kabbara et al. 2008, Torbick et al. 2008). This was expected, since chl reflects light in the
green wavelengths and has strong absorption features in the blue and red wavelengths.
Infrared bands were also found in all BRT models, which may indicate that incorporation of
physical features, such as algal scums on the surface of the water, is a relevant component
to algal biomass detection. Predictive power of physical variables such as the thermal band
and measured water depth in the Landsat models indicate that detection of chl can be
improved by accounting for physical features of the sample location. Though in this study
station depth was obtained by in-lake measurement, future values could be obtained by
bathymetric map.
One notable difference between BRT models is the incorporation of multiple band of
the same general category (for example, red bands), in the MODIS model. This is
compatible with the idea of satellites with more numerous and finer bands being able to
detect different amount of chl or chl that comingle with other algal pigments more
accurately (Gower and Borstad 2004). Great Lakes algal communities span over six
divisions that include variety of non-chl a pigments, including chlorophylls b, c, and d, as
well as carotenoids, xanthophylls, and phycobilins (Vollenweider et al. 1974, Reynolds
63

2006). Finally, pixel size of the remote sensing image influences the spatial scale at which
chl is being summarized and possibly which bands would be most useful in chl prediction.
In areas of the Great Lakes where chl values are fairly consistent, a larger pixel size may be
adequate. A smaller pixel size becomes more important in areas with highly variable chl
concentrations throughout the summer.
Reasons that MODIS outperformed Landsat may include that MODIS is able to
capture more subtle spectral features with greater temporal frequency (NASA 2013). Also,
MODIS has more wavelength bands with which to build BRT models that have the potential
to explain more color-related or physical attributes of water. As with MODIS to Landsat
comparisons, the favorable performance of ETM+ and TM to MSS models may have to do, in
part, with more numerous spectral bands, slightly finer spectral and temporal resolution.
Application of BRT models to the 40-year Landsat image archive can help elucidate
changes in algal biomass across the Great Lakes for as long as the Clean Water Act has been
in place, with high spatial resolution. These patterns can help discern regular, seasonal
fluctuations from climate change, invasive species, and habitat alteration effects on algal
biomass. With its high chl prediction accuracy, application of the MODIS BRT model to the
12-year MODIS image archive can give us additional information on the seasonal dynamics
of algal biomass within each year. The notably high accuracy of the MODIS BRT model
throughout the range of measured chl values indicates its applicability across the Great
Lakes in areas with both somewhat high and low chl values. The Landsat models are less
accurate at detecting higher chl values than MODIS, but can still capture algal conditions
above 5 µg/L. Therefore, application of the MODIS BRT models would be appropriate
across the Great Lakes and at times when chl values range from 0-12 µg/L. Landsat models
64

would be most appropriate for use in areas or during times of year where chl does not
reach go much beyond 8 µg/L. The advantages of Landsat and MODIS complement each
other and enable future research and management programs to be focused on parts of the
Great Lakes that are most vulnerable to stress, including climate change related impacts.
BRT predictive models could also be used conjunction with other chl models. As
previously mentioned, a main advantage of BRT models is that they have high predictive
accuracy. However, they are less easily interpretable than simple linear or polynomial
algorithms, leading BRTs to be more “black box” in nature (Elith et al. 2008). If predictive
accuracy and not interpretability is the objective, then the black box nature of the models is
not as much of a drawback. However, if the study goal is both predictive accuracy and
interpretability, then using a multimodel approach that incorporates results from BRT and
other models may be best. With a multimodel approach one may gain benefits of both, for
results that may be more accurate and more widely applicable than either model alone. Yet
another approach to chl prediction might use different models in different parts of the
Great Lakes. Where the dominant color producing agent is chl and concentrations are
relatively low and constant, the ocean models may work well enough. In areas with higher
or more variable chl, or with more reflective influence by the bottom in nearshore zones,
other models may be more appropriate.
Chl BRT predictive models, used alone or in conjunction with other models, could be
used to establish a standard method in chl prediction in the Great Lakes. Advantages to a
well-tested standard method include being able to compare chl concentrations across
broader spatial and temporal scales and efficient use of resources for the implementation
and improvement of one model rather than production of additional models. Such efforts
65

can advance our understanding of ecosystem changes and allow managers to reliably
assess and target regions of the Great Lakes that experience the largest algae-related
stresses. Improving water quality in these parts of the Great Lakes would allow for
recovery of habitat for aquatic species that are negatively impacted by algal bloom
conditions that often include, low oxygen levels, presence of toxins, and low light levels
(Wetzel 2001). People, too, would benefit from more reliable chl prediction in Great Lakes
waters, to ensure safe recreation, water and fish consumption, and aesthetic wellbeing.
Our study sites may have been variably affected by color producing agents, such as
dissolved organic carbon. Since we did not have total inorganic carbon or dissolved organic
carbon data, we were not able to separate the chl signal from these color producing agents.
Also, the range of chl provided by the training data is not representative of the upper range
of chl experienced during bloom events (Freedman 1974). A larger data set, with more
frequent sampling over the summer might allow for more bloom events to be sampled.
Possible problems with developing or applying Landsat and MODIS model across lakes,
time periods include turbulence caused by wind on the lake surface (Singh 1994). Such
weather patterns can change within a short period, and can be hard to determine with
accuracy after the sampling period in the vicinity of the sample station. Time delays in
image and sample collection date in the training data could have introduced some
additional error, particularly with Landsat images. For the Landsat models, better results
might have been somewhat dependent on sample size.

66

APPENDIX

67

Table 3-1: MODIS wavelength band features (NASA 2013).
Band Number
Band
Band wavelength (nm)

Resolution (m)

1

Red

620-670

250

2

Near Infrared

841-876

250

3

Blue

459-479

500

4

Green

545-565

500

5

Near Infrared

1230-1250

500

6

Short Wave Infrared

1628-1652

500

7

Short Wave Infrared

2105-2155

500

8

Blue

405-420

1000

9

Blue

438-448

1000

10

Blue

483-493

1000

11

Green

526-536

1000

12

Green

546-556

1000

13 L/H

Red

662-672

1000

14 L/H

Red

673-683

1000

15

Near Infrared

743-753

1000

16

Short Wave Infrared

1550-1750

1000

68

Table 3-2: Landsat 7 ETM+ wavelength band features (USGS 2011). Values are the
same for Landsat 4-5 TM, unless indicated by and asterisk. * = 760 for TM, ** = 120 m for
TM, and *** = 2080 for TM .
Band number Landsat bands
Band wavelength (nm)
Resolution (m)
1

Blue

450-520

30

2

Green

520-600

30

3

Red

630-690

30

4

Near Infrared

770*-900

30

5

Short Wave Infrared

1550-1750

30

6

Thermal Infrared

10400-12500

60**

7

Short Wave Infrared

2090***-2350

30

69

Table 3-3: Landsat 1-3 MSS wavelength band features (USGS 2011).
Band
Corresponding Landsat bands
Band wavelength Resolution
number

ETM+/TM band

4

2

Green

500-600

60

5

3

Red

600-700

60

6

4

Near Infrared

700-800

60

7

--

Near Infrared

800-1100

60

(nm)

70

(m)

Table 3-4: Boosted regression tree results for MODIS, Landsat ETM+, TM, and MSS.
Satellite
RMSE
R2
Cross-Val
Number of
(µg/L)

R2

BRT Trees

MODIS

0.10

0.98

0.85

8300

ETM+ & TM

0.55

0.93

0.69

1800

MSS

0.73

0.88

0.66

1350

71

Table 3-5: Comparison of BRT chl predictive models to models in the literature.
Satellite
Source
Model
R2
RMSE
(µg/L)
TM & ETM+ Novitski et al. in

BRT

0.69

0.55

Lathrop &

Band ratio

0.16

5.55

Lillesand 1986

(Green bay)

Lathrop &

Band ratio

0.16

4.39

Lillesand 1986

(Lake Michigan)

Novitski et al. in

BRT

0.66

0.73

BRT

0.85

0.10

OC3

0.75

0.47

OC3-refine

0.75

0.49

prep.
TM & ETM+

TM & ETM+

MSS

prep.
MODIS

Novitski et al. in
prep.

MODIS

O’Reilly et al.
2000

MODIS

Lesht et al.
2013

72

8
6
4
0

2

MODIS Predicted Chl (µg/L)

10

12

Great Lakes
Measured Chl versus MODIS Predicted Chl

0

2

4

6

8

10

12

Measured Chl (µg/L)

Figure 3-1: Measured chl versus MODIS BRT predicted chl for 2007-2009 GLNPO
training data with 1:1 line.

73

74

75

10
8
6
4
2
0

Landsat ETM+ and TM Predicted Chl (µg/L)

12

Great Lakes
Measured Chl versus Landsat ETM+ and TM Predicted Chl

0

2

4

6

8

10

12

Measured Chl (µg/L)

Figure 3-3: Measured chl versus Landsat ETM+ and TM BRT predicted chl for 20072009 GLNPO training data with 1:1 line.

76

77

8
6
4
0

2

Landsat MSS Predicted Chl (µg/L)

10

12

Great Lakes
Measured Chl versus Landsat MSS Predicted Chl

0

2

4

6

8

10

12

Measured Chl (µg/L)

Figure 3-5: Measured chl versus Landsat MSS predicted chl for 2007-2009 GLNPO
training data with 1:1 line.

78

2.0

0.02

0.5

0.7

0.9

b4_toaOb3_toa (2.2%)

0.06

0 1 2 3

fitted function

0 1 2 3

0.06

0.08

1.4

2.2

2.0

3.0

4.0

b2_toaOb4_toa (3%)

0 1 2 3

b3_toaOb2_toa (3.3%)

1.8

b2_toaOb3_toa (3.8%)

fitted function
0.40 0.50 0.60 0.70

b4_toa (3.6%)

fitted function

0 1 2 3

fitted function

b3_toaOb4_toa (3.7%)

0.04

0.04

b3_toa (5.5%)

0 1 2 3

fitted function
1.6

0.02

b2_toa (12.4%)

0 1 2 3

fitted function

1.2

fitted function
0.05 0.07 0.09 0.11

0 1 2 3

250

0 1 2 3

150

depth (60.6%)

fitted function

0 50

0 1 2 3

fitted function

0 1 2 3

fitted function

surface_chl - page 1

0.3

0.4

0.5

0.6

b4_toaOb2_toa (1.9%)

Figure 3-6: Partial dependence plots for the Landsat MSS BRT. Each plot shows the
relationship of chl (response variable) to the individual band or ratio (predictor variables)
after accounting for the effects of all other predictor variables in the final BRT model.

79

LITERATURE CITED

80

LITERATURE CITED

Babyak, M. A. 2004. What you see may not be what you get: A brief, nonethical introduction
to overfitting in regression-type models. Psychosomatic Medicine 66:411-421.
Becker, R. H., M. I. Sultan, G. L. Boyer, M. R. Twiss, and E. Konopko. 2009. Mapping
cyanobacterial blooms in the Great Lakes using MODIS. Journal of Great Lakes
Research 35:447-453.
Binding, C. E., T. A. Greenberg, and R. P. Bukata. 2012. An analysis of MODIS-derived algal
and mineral turbidity in Lake Erie. Journal of Great Lakes Research 38:107-116.
Carder, K. L., F. R. Chen, J. P. Cannizzaro, J. W. Campbell, and B. G. Mitchell. 2004.
Performance of the MODIS semi-analytical ocean color algorithm for chlorophyll-a.
Advances in Space Research 33:1152-1159.
Carpenter, D. J. and S. M. Carpenter. 1983. Modeling inland water quality using Landsat
data. Remote Sensing of Environment 13:345-352.
Chander, G., B. L. Markham, and D. L. Helder. 2009. Summary of current radiometric
calibration coefficients for Landsat MSS, TM, and ETM+, and EO-1 ALI sensors.
Remote Sensing of Environment 113:893-903.
De'ath, G. 2007. Boosted trees for ecological modeling and prediction. Ecology 88:243-251.
Elith, J., J. R. Leathwick, and T. Hastie. 2008. A working guide to boosted regression trees.
Journal of Animal Ecology 77:802-813.
Freedman, P. L. 1974. Saginaw Bay: An evaluation of existing and historical conditions.
Pages 1-137 in U. S. E. P. Agency, editor. Environmental Protection Agency,
University of Michigan.
Friedman, J. H. 2002. Stochastic gradient boosting. Computational Statistics & Data Analysis
38:367-378.
Giardino, C., M. Pepe, P. A. Brivio, P.Ghezzi, and E.Zilioli. 2001. Detecting chlorophyll, Secchi
disk depth and surface temperature in a sub-alpine lake using Landsat imagery. The
Science of the Total Environment 268:19-29.
Gower, J. F. R. and G. A. Borstad. 2004. On the potential of MODIS and MERIS for imaging
chlorophyll fluorescence from space. International Journal of Remote Sensing
25:1459-1464.
81

Kabbara, N., J. Benkhelil, M. Awad, and V. Barale. 2008. Monitoring water qualitly in the
coastal area of Tripoli (Lebanon) using high-resolution satellite data. ISPRS Journal
of Photogrammetry & Remote Sensing 63:488-495.
Lathrop, R. G., Jr. and T. M. Lillesand. 1986. Use of Thematic Mapper data to assess water
quality in Green Bay and central Lake Michigan. Photogrammetric Engineering and
Remote Sensing 52:671-680.
Lebouteiller, A. and A. Herbland. 1982. Diel variation of chlorophyll-a as evidenced from a
13-day station in the equatorial atlantic-ocean. Oceanologica Acta 5:433-441.
Lesht, B. M., R. P. Barbiero, and G. J. Warren. 2013. A band-ratio algorithm for retrieving
open-lake chlorophyll values from satellite observations of the Great Lakes. Journal
of Great Lakes Research 39:138-152.
Lillesand, T. M., W. L. Johnson, R. L. Deuell, O. M. Lindstrom, and D. E. Meisner. 1983. Use of
Landsat data to predict the trophic state of Minnesota lakes. Photogrammetric
Engineering and Remote Sensing 49:219-229.
Masuoka, E. and K. Horrocks. 2010. Level 1 and atmospheric archive and distribution
system. NASA Goddard Space Flight Center.
NASA. 2013. MODIS Specifications.in N. A. a. S. Administration, editor., National Aeronautics
and Space Administration.
NOAA. 2013. NOAA solar calcuator. U.S. Department of Commerce, National Oceanic and
Atmospheric Adminitration.
O'Reilly, J. E., S. Maritorena, B. G. Mitchell, D. A. Siegel, K. L. Carder, S. A. Garver, M. Kahru,
and C. McClain. 1998. Ocean color chlorophyll algorithms for SeaWiFS. Journal of
Geophysical Research 103:24,937-924,953.
O'Reilly, J. E., S. Maritorena, M. C. O'Brien, D. A. Siegel, D. Toole, D. Menzies, R. C. Smith, J. L.
Mueller, B. G. Mitchell, M. Kahru, F. P. Chavez, P. Strutton, G. F. Cota, S. B. Hooker, C.
R. McClain, K. Carder, F. Müller-Karger, L. Harding, A. Magnuson, D. Phinney, G. F.
Moore, J. Aiken, K. R. Arrigo, R. Letelier, and M. Culver. 2000. SeaWiFS postlaunch
calibration and validation analyses, Part 3.in S. B. a. E. R. F. Hooker, editor. SeaWiFS
Postlaunch Techinical Report Series. NASA.
Oreopoulos, L., M. J. Wilson, and T. Varnai. 2011. Implementation on Landsat data of a
simple cloud mask algorithm developed for MODIS land bands. . IEEE Geoscience
and Remote Sensing Letters 8:597-601.
Owens, T. G., P. G. Falkowski, and T. E. Whitledge. 1980. Diel periodicity in cellular
chlorophyll content in marine diatoms. Marine Biology 59:71-77.
82

Peng, C.-Y. J., K. L. Lee, and G. M. Ingersoll. 2002. An introduction to logistic regression
analysis and reporting. The Journal of Educational Reserach 96:3-14.
Pozdnyakov, D., R. Shuchman, A. Korosov, and C. Hatt. 2005. Operational algorithm for the
retrieval of water qualilty in the Great Lakes. Remote Sensing of Environment
97:352-370.
Reynolds, C. 2006. Ecology of Phytoplankton. Cambridge University Press, New York.
Shuchman, R., A. Korosov, C. Hatt, D. Pozdnyakov, J. Means, and G. Meadows. 2006.
Verification and application of a bio-optical algorithm for Lake Michigan using
SeaWiFS: A 7-year inter-annual analysis. Journal of Great Lakes Research 32:258279.
Toller, G. N., A. Isaacman, J. Kuyper, and V. Salomonson. 2006. MODIS Level 1B Product
User's Guide. NASA/Goddard Space Flight Center.
Torbick, N., F. Hu, J. Zhang, J. Qi, H. Zhang, and B. Becker. 2008. Mapping chlorophyll-a
concentrations in West Lake, China using Landsat 7 ETM+. Journal of Great Lakes
Research 34:559-565.
USGS. 2011. Landsat Missions. United States Geological Survey.
USGS. 2013a. EarthExplorer. United States Geological Survey.
USGS. 2013b. USGS FTP server parent directory. USGS.
USGS. 2013c. USGS Global Visualization Viewer. United States Geological Survey.
Vollenweider, R. A., M. Munawar, and P. Stadelmann. 1974. A comparative review of
phytoplankton and primary production in the Laurentian Great Lakes. Journal of
Fisheries Research Board of Canada 31:739-762.
Wetzel, R. G. 2001. Limnology: Lake and Reservoir Ecosystems. Adacemic Press, New York,
New York.

83

CHAPTER 4
IMPACTS OF CLIMATE CHANGE AND LAND-USE/LAND-COVER CHANGE ON
PHYTOPLANKTON IN INNER SAGINAW BAY USING LANDSAT AND MODIS SATELLITE
IMAGERY, 1973-2012
Abstract
Saginaw Bay has experienced significant fluctuations in water quality over the last four
decades. Remote sensing of chlorophyll (chl) can help to track water quality over large spatial
and temporal scales. Our research objectives were to determine: 1) if water temperature and
nutrient inputs have changed in inner Saginaw Bay from 1973-2012; and 2) how any changes in
water temperature and nutrient inputs as well as distance from Saginaw River inputs affect algal
blooms in Saginaw Bay. Chl concentrations were predicted from Landsat 1973-2012 and
MODIS 2000-2012 maps with boosted regression tree analysis. Phosphorus loading was
estimated with a mixed-effect model based existing phosphorus, precipitation, discharge, and
land use/land cover data from various sources. Surface water temperature was derived from the
thermal band aboard Landsat’s ETM and TM satellites.
In inner Saginaw Bay, annual average and upper quartile Landsat-derived
chlorophyll decreased from 7.44 to 6.62 and 8.38 to 7.38 µg/L between 1973-1982, and
annual upper quartile of 8-day phosphorus loads increased from 5.29 to 6.79 kg between
1973-2012. Wilcoxon rank tests comparing chl values in groups, “near” and “far” from the
Saginaw River mouth showed increases in minimum, average, and upper quartile chl values for
MODIS (2000-2012) and Landsat- (1984-2012) derived chl near the Saginaw River mouth.
Simple linear and multiple regression models for MODIS-derived chl indicate that distance from
the Saginaw River mouth may influence chl concentration in Saginaw Bay; Landsat-derived
surface water temperature and phosphorus loads to a lesser extent. Mixed-effect models for
MODIS and Landsat-derived chl were related to chl better than simple linear or multiple
84

regressions, with random effects of pixel and sample date contributing substantially to predictive
power (NSE=0.35-70), though phosphorus loads, distance to Saginaw River mouth, and water
were significant fixed effects in most models. Water quality changes in Saginaw Bay between
1972-2012 may be influenced by phosphorus loading and distance to the Saginaw River’s mouth.
Other factors, such as the zebra mussel invasion may have had an impact on chl concentration.
Introduction
Saginaw Bay has been the subject of great ecological interest for decades. This shallow,
elongated bay in southwestern Lake Huron, with surface area of approximately 2,960 km2,
shows marked differences in water chemistry, chlorophyll (chl) concentration, and bacterial
activity (Schelske et al. 1974, Moll et al. 1980) than the rest of Lake Huron. Even the inner and
outer portions of Saginaw Bay that are separated by a broad shoal (Freedman 1974) differ in
circulation patterns (Danek and Saylor 1977) and algal assemblages (Stoermer and Theriot
1985). This may be due, in part, to differences in mean depth; 4.6 and 14.6 m for inner and outer
bays, respectively (Freedman 1974).
Degradation of water quality in Saginaw Bay over the years has been influenced not only
by natural basin characteristics, such as water depth, but also by direct municipal, agricultural,
and industrial inputs, and invasive species introductions. In the 1970s, approximately 50% of the
land in the watershed was agricultural (Freedman 1974), and 55% of the phosphorus in Saginaw
Bay was coming from agricultural runoff (MDNR 1988). However, large quantities of nutrient
and chemical inputs were also coming from industrial and municipal sources from the Saginaw
River, which provides over 90% of Saginaw Bay’s drainage inputs (Bierman et al. 1984).
Around this time, phytoplankton studies indicated that cyanobacteria, including potential toxinforming Anabaena and Microcystis, were the dominant algal taxonomic group present in

85

Saginaw Bay between July and September (Munawar and Munawar 1982). High cyanobacterial
abundance even caused taste and odor problems in municipal drinking water supplies until the
late 1970s (MDNR 1988).
After Saginaw Bay earned an area of concern designation (MDNR 1988), improvements
in water quality followed reductions in nutrient inputs, including from a phosphorus ban in
laundry and other detergents in the Great Lakes (Hartig and Horvath 1982, Bierman et al. 1984)
and additional targeted remediation efforts. However, by 1991 other water quality changes
occurred after zebra mussels (Dreissena polymorpha) became established (Nalepa and
Fahenestiel 1995), namely a 59% decrease in phytoplankton biomass as chlorophyll a
(Fahnenstiel et al. 1995b) allowing planktonic cyanobacterial species to thrive relative to diatoms
and chlorophytes (Bridgeman et al. 1995, Heath et al. 1995, Lavrentyev et al. 1995, Vanderploeg
et al. 2001, Bierman et al. 2005). As planktonic algae decreased, benthic algal production
increased (Fahnenstiel et al. 1995a) along with macrophytes (Skubinna et al. 1995). Benthic
community composition shifted from diatoms to filamentous green algae (Lowe and Pillsbury
1995).
Ecosystem changes that involve algal abundance and community composition have a
number of possible causes and implications, including changes in nutrient cycling and presence
of toxins. Some studies indicate that zebra mussels control algal abundance and community
composition directly by selective feeding (Vanderploeg et al. 2001) or indirectly by controlling
the amount of phosphorus in the water column through assimilation, or sedimentation and
resuspension of feces (Johengen et al. 1995, Vanderploeg et al. 2002). Nicholls et al. (1999)
found that chl:phosphorus ratios after zebra mussel invasion were significantly different in
different parts of the Great Lakes, having implications for the efficacy of phosphorus reduction

86

efforts (Nicholls et al. 1999). Zebra mussels may even indirectly affect their own filtering rate by
selective feeding on certain phytoplankton types, leaving the less edible taxa (Pillsbury et al.
2002). Along with the increased presence of Microcystis, is increased chance for toxins. One
study showed that the gene for microcystin production was more prevalent in Saginaw Bay than
in Lake Erie (Dyble et al. 2008). Another study found that phosphorus has a large influence on
microcystin based on its influence on Microcystis growth rates and abundance (Fahnenstiel et al.
2008).
Remote sensing is a useful tool for broad-scale, long-term assessment of water quality
indicators, such as chl concentration. Early remote sensing studies of Saginaw Bay, using
Landsat MSS satellite imagery, were done for the purpose of creating maps of water features,
including chlorophyll, Secchi depth, and temperature (Rogers 1975, Rogers et al. 1976, McKeon
et al. 1977). More recently, the AVHRR satellite has been used to evaluate differences in Secchi
depth and total suspended solids before and after zebra mussel invasion in Saginaw Bay (Budd et
al. 2001). Other satellites that have been used to study various parameters of Saginaw Bay, such
as, chl concentration or concentration of cyanobacterial species, include the SeaWiFS
(Pozdnyakov et al. 1999), and MERIS (Wynne et al. 2008). Little work has been done in
Saginaw Bay with Landsat and MODIS satellites, both which are currently functional and have
images available to the public free of charge. In addition, the constellation of Landsat satellites
has an image archive dating back to 1972, the year the Clean Water Act was enacted. MODIS
has an image archive that goes back to 2000. Therefore, these satellites have the potential to be
extremely valuable for water quality assessment of water bodies, such as Saginaw Bay, that have
variable ecological history.

87

Research objectives
Using inferred-chl from MODIS and Landsat images and satellite boosted regression tree
models, USGS river discharge data, and distance-to-river mouth data, our research objectives
were to determine: 1) if water temperature and nutrient inputs have changed in inner Saginaw
Bay from 1973-2012; and 2) how changes in water temperature and nutrient inputs as well as
distance from Saginaw River inputs affect algal blooms in Saginaw Bay using satellite-derived
chl concentrations.
Methods
Data
Chl concentration was inferred from boosted regression tree (BRT) models (Novitski et
al. in prep) applied to MODIS and Landsat MSS, TM, and ETM+ images using Python with
GDAL and RPy2 (Lawawirojwong et al. in prep). Chl data were extracted from images with
pixels that corresponded to a portion of Saginaw Bay’s inner bay (Figure 4-1) deemed deep
enough to avoid the influence of reflection from the lake bottom, using Python 2.5 software
(Lawawirojwong et al. in prep). This subset of pixels from the inner bay was chosen for the
analysis based on a depth to chl relationship, derived from Carlson’s Secchi to chl relationship
(Carlson 1977).
‫݊ܮ‬ሺ݄ܵ݁ܿܿ݅ሻ ൌ െ0.68 ‫݊ܮ כ‬ሺ݄݈ܿሻ ൅ 2.04
If depth was at least as deep as Secchi disk depth, we assumed that reflectance from the bottom
of the bay would not interfere with detection of chl in the water column. Most of the predicted
chl concentrations for Saginaw Bay were > 2 µg/L, which agrees with historical data of the inner
bay (CCIW 1972, Schelske and Roth 1973, Schelske et al. 1974, Smith et al. 1977, Fahnenstiel et
al. 1995b, Nalepa et al. 1995, Suzuki et al. 1995, Fanslow et al. 2001, Vanderploeg et al. 2001,

88

Lehman et al. 2004). Based on Carlson’s equation, with at least 2 µg/L, the minimum depth we
used in our analyses is 4.8 m. For practicality, we rounded this number to 5 m. MODIS-inferred
chl data consists of pixel values from 8-day composite MODIS images, starting from the last
week of July through the first week of September, for a total of six images per year, except for
the year 2000 in which only 5 composite images were created because images between
08/04/2000-08/11/2000 were not available. Therefore, a total of 77 MODIS images were used in
these analyses. The Landsat-inferred chl data consist of 209 images, taken from July through
September of 1972-2012. Due to cloud interference or low image quality for some images, the
number of images ranges from 0-11 per year, 1972 and 1983 having no images. To reduce the
Landsat data to a more manageable size for statistical analyses, we found the latitude and
longitude of each MODIS pixel centroid using ArcGIS 9.3 and extracted the Landsat-inferred chl
value from the Landsat pixel at that same location using Python 2.5 software. Therefore, the
resulting number of chl values per Landsat image, 1887, was the same for the MODIS data.
Locations of river mouth in Saginaw Bay data were determined using hydrological
information from NHDPlus (USEPA and USGS 2005). Once MODIS pixel centroids were
determined in ArcGIS 10.1, distance-to-river mouth data were created in Microsoft Excel 2007
by applying the haversine formula. For the purposes of this study, we selected the distance to the
Saginaw River mouth for each pixel.
Phosphorus loads at river mouths emptying into Saginaw Bay (Figure 4-2) were
estimated for 1972-2012 using linear mixed-effects models (Esselman et al. in prep). Land
use/land cover information used to inform phosphorus load models included 2001 and 2006
National Land Cover Database data, 1978 Michigan Resource Inventory System data, and 1996
Coastal Change Analysis Program data. Discharge data used in phosphorus load model

89

development were, themselves, created using precipitation data from NOAA North American
Regional Reanalyis and National Centers for Environmental Prediction and from a watershed
characteristics database of Dr. Dave Hyndman’s lab. Measured phosphorus data used in the
phosphorus load model building came from three general sources: EPA STORET, Kendall et al.
(unpublished), and Dr. R.J. Stevenson. To obtain phosphorus load values at a level that was
spatially relevant to the chl data, phosphorus loads were weighted by the distance data using two
weighting approaches (Van Sickle and Johnson 2008) inverse distance weighting f(d) = d-α and
exponential weighting f(d) = exp(-αd) where d= distance between the pixel centroid and the river
mouth. In early analyses, inverse distance weighting resulted in better model outcomes, so it was
used for the remainder of analyses.
Surface water temperature data were obtained from Landsat ETM+ and TM images by
extracting thermal band pixel values corresponding to the location of MODIS pixel centroids and
converting the digital number to degrees Celsius (Chander et al. 2009) using Python 2.5.
Statistics for MODIS and Landsat
Eighteen MODIS composite images and seven Landsat images had missing pixel values
due to clouds. For one of the MODIS images (8/4/2000), no pixel values were available, and for
Landsat, images from 1972 and 1983 missing due to cloud coverage or image quality problems.
For the images with missing pixel values, temporal analyses of these pixels or sample dates
could not be carried out.
Simple linear regressions were done to determine temporal trends of chl, water
temperature, and phosphorus loads in inner Saginaw Bay over the lifetime of each satellite.
However, since Landsat MSS did not have a thermal band with which to detect water
temperature, Landsat-derived water temperature only goes back to 1984. Landsat-derived chl

90

over the 1973-2012 had a non-normal, bimodal distribution. A category and regression tree
analysis was run to determine if there were two statistically distinct chl groups (1973-1982 and
1984-2012). Based on the category and regression tree results and the distribution of chl data in
the resulting groups, linear regressions were run on the 1973-1982 and 1984-2012 data,
individually. Wilcoxon rank test was done on “near” and “far” pixels (below and above 43.8˚N,
respectively) to further discern the influence of the Saginaw Bay river mouth.
Simple linear regression analyses were also run to model effects of water temperature
(X1), phosphorus loads (X2) on chl concentration (Y). The effects on chl of distance to the
Saginaw River mouth was tested, but never in conjunction with phosphorus loads due to the
redundancy of the distance-weighted phosphorus loads.
Finally, mixed-effects models were calculated to determine if the predictive power can be
improved by adding the unexplained or random variation of time or space using the following
equation:
ܻ௜௝ ൌ ߚ଴ ൅ ߚଵ ܺଵ௜ ൅ ߚଶ ܺଶ௝ ൅ ߙ௜ ൅ ߛ௝ ൅ ߝ௜௝

where ߚ are the coefficients for the X predictors (fixed effects), the site index is
represented by ݅ ൌ 1,2, … , ܰ, the time index is represented by ݆ ൌ 1,2, … , ܶ. Also, the
residual term is ߝ௜௝ ~ܰሺ0, ߪ ଶ ሻ, the random effect on sample site is ߙ௜ ~ܰሺ0, ߬ ଶ ሻ, and the
random effect on sample date is ߛ௝ ~ܰሺ0, ߜ ଶ ሻ, where ߪ ଶ , ߬ ଶ and ߜ ଶ represent variability
parameters.
Criteria used to evaluate the mixed-effects models included Bayesian information
criterion (BIC) and Akaike information criterion (AIC). Both BIC and AIC evaluate the
performance of the model while penalizing for increasing complexity. Root mean square
error (RMSE) was calculated to help understand the error associated with the final predicted
91

values of the model, and Nash-Sutcliffe efficiency (NSE) was used determine the proportion of
variation in the chl data that were explained, much like R2 values in linear regression models.
Smaller BIC, AIC, and RMSE values and larger NSE values are preferable. All statistics were
run in R version 2.15.1. Linear regression models were run in R the package, stats, and mixedeffect models were run in the lme4 package. Satellite-inferred chl was then compared to values
in the literature from inner Saginaw Bay.
Results
Temporal trends
Although there was a slight increasing trend in annual average and maximum MODISderived chl in the 2000-2012 time period during which MODIS has been operational, the results
were not statistically significant (Table 4-1, Figure. 4-3). Annual average and upper quartile
Landsat-derived chl decreased between 1972-2012, with a notable drop in the early 1980s,
decreasing in one year from 6.5 to 4.5 µg chl/L and 8.5 to 6.5 µg chl/L in average and upper
quartile values, respectively (Figure 4-4). Regression tree analysis showed a significant
changepoint producing two distinct periods for Landsat-derived chl: 1973-1983 and 1984-2012
(complexity parameter=0.95). It seemed likely that this changepoint was due to differences in
differences in band ranges and widths for sensors on Landsat satellites, so the two periods were
analyzed separately. Linear regressions run on each group of Landsat-derived chl showed a
decrease in average annual chl between 1973-1982 from 7.44 to 6.62 µg/L, but no change from
1984-2012 (Table 4-1, Figure 4-4). For both MODIS and Landsat (1894-2012), Wilcoxon test of
“near” and “far” derived chl, average, upper quartile, and minimum values were statistically
significantly higher near the Saginaw River mouth.

92

Between 2000-2012, MODIS-derived chl concentrations were consistently higher than
Landsat concentrations, although trends over time had some similarities with relatively low
concentrations in 2000 and 2003, and higher concentrations in 2001, 2005, and 2007 (Figure 43). Annual upper quartile 8-day phosphorus load increased between 1973-2012 from about from
5.29 to 6.79 kg/8 days, while temporal trends in annual average 8-day phosphorus were not
significant (4-1, Figure 4-5). Upper quartile values of 8-day phosphorus loading are notably
smaller than average values (Figure 4-5). Between 1984-2012, annual upper quartile water
temperature increased from about 19˚C to 20˚C, however trends in annual average water
temperature were not statistically significant (Table 4-1, Figure 4-6).
Relationships between algal blooms and global change variables
MODIS-derived chl was significantly (p<0.05) related to phosphorus loading, distance to
the mouth of the Saginaw River. Landsat-derived surface water temperature and interactions
among these variables, but phosphorus loading and water temperature or other combinations of
those predictor variables had almost no predictive power (Table 4-3) Interestingly, the best
mixed-effect models for MODIS-derived chl were any that included random effects of sample
site and sample date, explaining 56% (RMSE=1.10 µg/L) of the variation in the satellite-inferred
chl data (Table 4-4; Figure 4-7). Water temperature was the fixed effect that was most commonly
not statistically significant in mixed effect models, while phosphorus loads and distance were
always significantly (p<0.05) related to MODIS-derived chl (Table 4-4).
For 1973-2012, phosphorus loads and distance were negatively correlated to Landsatderived chl (p<0.05, Table 4-5). For 1984-2013, and phosphorus loads, water temperature were
positively correlated to Landsat-derived chl, while distance was negatively correlated with chl
(p<0.05, Table 4-5). Similarly to the MODIS mixed-effect model results, the best 1973-1982

93

Landsat mixed-effect model results were with models that included either sample date or both
random effects of pixel and sample date, accounting for 70% (RMSE=0.80 µg/L) of the variation
in the chl data (Table 4-6). Models that included sample date as the only random effect were
almost as strong as those with both random effects (Table 4-6).
The 1984-2012 mixed-effect models (some of which included water temperature as a
fixed effect) showed similar trends to other Landsat and MODIS mixed-effect models with those
that included both random effects being the strongest predictors of chl, explaining about 35%
(RMSE=0.77 µg/L) of the variation in the chl data, and those with just sample date as explaining
29-34% (RMSE=0.78-0.81 µg/L) of the variation (Table 4- 6). Only two Landsat mixed-effect
models with non-significant fixed effect of phosphorus loads (Table 4-6). As with MODIS mixed
effects models, the strongest of all Landsat mixed models was one that included distance as the
only fixed effect, as well as pixel and sample date as random effects (Table 4-6; Figure 4-8).
Discussion
Long-term trends in satellite-derived chl, phosphorus loading, and surface water
temperature, as well as and relationships among these variables and to distance to the Saginaw
River mouth, indicate that land use change, climate change, and invasive species may be
influencing water quality in inner Saginaw Bay. Although, it appears that chl concentration has
decreased, overall, since the Clean Water Act was enacted in 1972, a possible weakening of that
trend in recent years indicates a continued need for water quality assessment and management.
Increasing upper quartile values of 8-day phosphorus load in the last 40 years, and substantially
larger average vs. upper quartile phosphorus load values indicate that sizable phosphorus inputs,
though possibly infrequent, could be one factor in the lack of chl decrease in recent years (Paerl
and Huisman 2008). This occurred despite efforts to decrease phosphorus loads into Saginaw

94

Bay (Fahnenstiel et al. 1995b) since the mid-1970s. Possible explanations for increasing
phosphorus could be a changing urban or agricultural landscape in the Saginaw River Watershed,
or more frequent precipitation events in the Midwest (Lofgren and Gronewold 2012) that wash
phosphorus directly into Saginaw Bay, or through nearby rivers. Also, though trends in surface
water temperature were not apparent using Landsat-derived values, recent studies of warming in
the Great Lakes indicate that warming may still be a factor in Saginaw Bay (Austin and Colman
2007).
Indirect or interactive effects of phosphorus, water temperature and chl may be important
relationships to take into account. For example, the indirect effects of increasing water
temperature on chl may be mediated by physical factors, such as length of water column
stratification, and may effect phosphorus cycling and accessibility of phosphorus to biological
organisms (Scavia 1979). We did not see an overall effect of zebra mussels on chl. This
contradicts other findings of decreased phytoplanktonic chl after the zebra mussel invasion
(Fahnenstiel et al. 1995b), which may be accounted for by the inclusion of shallow sampling
stations. Zebra mussels, themselves, could be mediating the effects of phosphorus or water
temperature on phytoplankton (Sarnelle et al. 2005) or they could be moderating effects of
phosphorus or surface water temperature on chl by removing algae from the water column
(Fahnenstiel et al. 1995b). Algal communities that have been grazed by zebra mussel tend to
have more buoyant cyanobacterial species, such as Microcystis, which are thought to be harder
for zebra mussels to ingest (Vanderploeg et al. 2001). Shifts toward more toxin-producing algal
taxa is a major concern for managers and restoration efforts.
Both MODIS and Landsat-derived chl were consistent with chl values observed in
Saginaw Bay (Table 4-7). The July-September seasonal average of MODIS-inferred chl fit

95

within a range of average and one-time chl values from the literature (Table 4-7). Seasonal
average Landsat-inferred chl for 1991 and 1994 were within 1 µg/L chl of average values from
the literature for these years (Table 4-7). Overall, MODIS-inferred chl shows a more dynamic
range than Landsat-inferred chl. and is possibly more accurate due to narrower band widths and
larger numbers of spectral bands. MODIS also has finer temporal resolution to capture the chl
temporal signature. However, lakes that are too small for MODIS will still require the use of a
Landsat model or a model that hybridizes MODIS and Landsat-derived chl values. In such a
hybrid model, the bias between MODIS and Landsat models could be determined and used
to recalibrated Landsat models to reflect a more accurate surface chl concentration.
Mixed-effects models were able to capture relationships between predictor variables and
satellite-derived chl much more effectively than simple linear or multiple regression models.
Continued refinement of mixed-effect models with existing data can enable us to understand
which ecosystem attributes are most influential determinants of algal abundance over space and
through time. Similar models can be applied to other water bodies for improved management.
Future models could take into account the lag in the response of chl concentration to nonpoint
source pollution, internal nutrient cycling, or water circulation effects. We would expect this
connection to be particularly important in extreme weather events, including heavy storms or
drought. Currently, the data used for chl model calibration do not necessarily include a spring or
summer bloom event for each year of data collected. Having more chl data as well as
information on the extent and duration of blooms events would help us better capture bloom
timing and extent in our models, and it would also help us understand the timing of these events
in relation to other ecological factors, such as storm events. Having satellite inferred-chl data for
shallower depths could help to elucidate any relationships due to the proximity of river water and

96

nutrient inputs. In addition, the Landsat-inferred water temperature may have noise associated
with it due to atmospheric conditions that could not be accounted for. Modeled phosphorus loads
were extrapolated from numerous data sources. Refining both of these data sets could potentially
allow for more accurate chl prediction. Finally, accounting for interactions not just within each
pixel or sample date, but also between pixels and sample dates may help to improve model
outcomes. Additional years of data may help to elucidate trends in MODIS-derived chl, which
has the potential to be a powerful ecological assessment tool with its daily overpass and
relatively fine spectral bands (NASA 2013).
Our results show the utility of using satellite-derived chl data to monitor and understand
water quality trends in ecological models. Though changes in the data over the 40 year study
period are modest, they show a distinct trend that could become more prominent in the future
with possible tipping points after which management or restoration of Saginaw Bay is not
practical or feasible.

97

APPENDIX

98

Table 4-1. Linear regression results for temporal trends in chlorophyll, water
temperature, and phosphorus loads in inner Saginaw Bay. The 2000-2012 results
correspond to MODIS-derived chl. Landsat-derived chl results are broken into two year
groups (1973-1982 and 1984-2012) based on regression tree groupings. Landsat-derived
water temperature values span 1984-2012. Chl = chlorophyll concentration (µg/L), P load
= 8-day phosphorus load (kg), water temp. = surface water temperature (˚C), Avg = annual
mean value, UQ = annual upper quartile value, NS = not statistically significant.
Satellite
Years
Variable Annual
Increase/ Coefficients Adjuste
pAvg/Max Decrease
d R2
value
MODIS 2000-2012
Chl
Avg
---0.06
NS
MODIS 2000-2012
Chl
UQ
---0.08
NS
Landsat 1973-1982
Chl
Avg
Decrease
-0.09
0.33
0.05
Landsat 1973-1982
Chl
UQ
--0.15
NS
Landsat 1984-2012
Chl
Avg
---0.03
NS
Landsat 1984-2012
Chl
UQ
---0.04
NS
Landsat 1973-2012
P
Avg.
--0.04
NS
Landsat 1973-2012
P
UQ
Increase
0.02
0.26
<0.01
Landsat 1984-2012
WT
Avg
--0.00
NS
Landsat 1984-2012
WT
UQ
--0.07
NS

99

Table 4-2: Wilcoxon Rank test results using near and far groupings of MODIS or
Landsat chl or water temperature data. Chl=chlorophyll, WT=water temperature,
Max=annual maximum, Min=annual minimum, Avg=annual mean, UQ=annual upper
quartile. * indicates ties in ranking for which an exact p-value cannot be calculated.
Satellite
Chl/WT
Index
W
p-value
MODIS 2000-2012
Chl
Max
80
NS
Min
27
<0.01
Avg
24
<0.01
UQ
24
<0.01
WT
Max
76
NS*
Min
51
NS*
Avg
65
NS
UQ
61
NS*
Landsat 1973-1982
Chl
Max
53
NS*
Min
36.5
NS*
Avg
55
NS
UQ
49
NS
Landsat 1984-2012
Chl
Max
529.5
NS*
Min
117.5
<0.01
Avg
119
<0.01
UQ
177
<0.01
WT
Max
393.5
NS *
Min
339.5
NS *
Avg
367
NS *
UQ
337.5
NS *

100

Table 4-3. MODIS (2000-2012) linear regression results. Linear regression results
using combinations of variables (water temperature, phosphorus load, and distance-toriver mouth to predict chlorophyll concentration in inner Saginaw Bay. P=8-day
phosphorus load (kg), D=distance to Saginaw River mouth from MODIS pixel centroid (km),
WT=Landsat-derived surface water temperature (˚C), pixel=sample site, date=date of
satellite-derived chl. Phosphorus load and distance were never used in the same model
because phosphorus load is inverse-distance weighted for regression analyses.
Predictor Variables
Adjusted R2
p-value
P
0.01
< 2x10-16
WT
0.01
< 2x10-16
D
0.12
< 2x10-16
P + WT
0.01
< 2x10-16
WT + D
0.13
< 2x10-16

101

Table 4-4: MODIS linear mixed-effect model results. P=8-day phosphorus load (kg),
D=distance to Saginaw River mouth from MODIS pixel centroid (km), WT=Landsat-derived
surface water temperature (˚C), pixel=sample site, date=date of satellite-derived chl,
RMSE=root mean square error, NSE=Nash-Sutcliffe efficiency, AIC=Akaike information
criterion, BIC=Bayesian information criterion. Statistical significance is at p≤0.05 level.
Fixed-effects variables
Significant fixed RMSE NSE
AIC
BIC
effect variables
P + pixel
P
1.52
0.17
466231 466270
P + date
P
1.21
0.47
403940 403979
P + pixel + date
WT
1.10
0.56
387856 387905
WT + pixel
WT
1.53
0.15
467454 467493
WT + date
WT
1.28
0.41
417343 417382
WT + pixel + date
-1.10
0.56
387859 387908
D + pixel
D
1.54
0.15
464877 467916
D + date
D
1.14
0.53
388388 388427
D + pixel + date
D
1.11
0.56
384771 384820
P + WT + pixel
P
1.52
0.17
465758 465807
P + WT + date
P
1.21
0.47
403757 403805
P + WT + pixel + date
P
1.10
0.56
387867 387926
D + WT + pixel
D,WT
1.53
0.15
464435 464483
D + WT + date
D, WT
1.14
0.53
388398 388447
D + WT + pixel + date
D
1.10
0.56
384782 384840

102

Table 4-5: Landsat (1973-2012 or 1984-2012) linear regression results. Linear
regression results using combinations of variables (water temperature, phosphorus load,
and distance-to-river mouth to predict chlorophyll concentration in inner Saginaw Bay.
P=8-day phosphorus load (kg), D=distance to Saginaw River mouth from Landsat pixel
centroid (km), WT=Landsat-derived surface water temperature (˚C), pixel=sample site,
date=date of satellite-derived chl. Phosphorus load and distance were never used in the
same model because phosphorus load is inverse-distance weighted for regression analyses.
Year Range
Predictor Variables
Adjusted
p-value
2
R
1973-2012
P
0.05
< 0.01
1973-2012
D
0.01
< 0.01
1984-2012
P
0.00
< 0.01
1984-2012
D
0.05
< 0.01
1984-2012
WT
0.00
< 0.01
1984-2012
P + WT
0.00
< 0.01
1984-2012
D + WT
0.06
< 0.01

103

Table 4-6: Landsat linear mixed-effect model results. P=8-day phosphorus load (kg),
D=distance to Saginaw River mouth from Landsat pixel centroid (km), WT=Landsatderived surface water temperature (˚C), pixel=sample site, date=date of satellite-derived
chl, RMSE=root mean square error, NSE=Nash-Sutcliffe efficiency, AIC=Akaike information
criterion, BIC=Bayesian information criterion. Statistical significance is at p≤0.05 level.
Year Range Fixed-effects
Signif.
RMSE NSE AIC
BIC
variables
fixed
effect
variables
1973-2102 P + pixel
P
1.39
0.09 1291662 1291705
1973-2102 P + date
P
0.81
0.69 890991
891035
1973-2102 P + pixel + date
P
0.80
0.70 885052
885106
1973-2102 D + pixel
D
1.45
0.14 1314953 1314996
1973-2102 D + date
D
0.80
0.70 883763
883807
1973-2102 D + pixel + date
D
0.80
0.70 882928
882982
1984-2012 P + pixel
P
0.92
0.07 777789
777831
1984-2012 P + date
P
0.80
0.31 687537
687579
1984-2012 P + pixel + date
-0.80
0.35 676255
676308
1984-2012 WT + pixel
WT
0.93
0.07 778085
778127
1984-2012 WT + date
WT
0.81
0.29 697233
697276
1984-2012 WT + pixel + date
WT
0.77
0.35 675714
675767
1984-2012 D + pixel
D
0.93
0.06 776769
776811
1984-2012 D + date
D
0.78
0.34 674787
674829
1984-2012 D + pixel + date
D
0.77
0.35 673112
673165
1984-2012 P + WT + pixel
P,WT
0.92
0.08 776176
776229
1984-2012 P + WT + date
P,WT
0.80
0.31 687340
687393
1984-2012 P + WT + pixel + date WT
0.77
0.35 675733
675797
1984-2012 D + WT + pixel
D,WT
0.93
0.07 774963
775016
1984-2012 D + WT + date
D, WT
0.78
0.34 674239
674292
1984-2012 D + WT + pixel + date D, WT
0.77
0.35 672553
672617

104

Table 4-7: Comparison of MODIS- and Landsat-derived inner Saginaw Bay chlorophyll values and values from
the literature. Pub=publication,1=Schelske & Roth 1973, 2=CCIW 1872*, 3=Schelske et al. 1974, 4=Smith et al. 1977,
Fahnenstiel et al. 1995, 6=Fanslow et al. 2001, 7=Nalepa et al. 1995, 8=Lehman et al. 2004, 9=Suzuki et al. 1995,
10=Vanderploeg et al. 2001, 11=Landsat-derived chlorophyll, 12=MODIS-derived chlorophyll. *Cited in Freedman et al.
1974, **samples taken from wetland area, †some values estimated from graph and some exact numbers taken from
Fanslow et al. 2001. Month=sample month, JN=June, JL=July, A=August, S=September. Type=sample type, Avg.=average,
R=range, Max=maximum, SA=seasonal average, OT=one-time, TOTS=two one-time samples. Last two digits used to
indicate sample year (e.g., 70=1970, 00=2000).
Pub
1

2*
3

Month Type
JL
Avg.

A
S

Avg.;
R

4
5
6
6
7
7
7
7
8
8

R;
Max
Avg.

JN-S
JL
A
S
JL
A
S
A
S

SA

70
7.26

72

74

89

90

91

92

93

11.0

7.5

6.0

3.0

94

95

96

16.1
9.01
6.28

3.49
13.45
16.79

99

00

4.012;
20.0
15.0
16.26;
2.3345.7

OT
OT
OT
OT
OT
OT

9.1
2.6
2.6

7.8
11.9
16.4

1.1
2.6
2.9
10.22
17.4,
17.5

TOTS

105

Table 4-7 (cont’d)
Pub

Month Type

9**

A

R

10†
10†

JL
A

OT
OT

10†

S

OT

11
12

JL-S
JL-S

SA
SA

70

72

74

7.03

89
0.542.0

4.58

90

91

92

93

94

95

96

9.0
2.5

8.0
12.0

1.0
2.5

8.0
15.5

13.0
7.64

1.85
16.79

14.0

2.5

16.5

3.0

11.0

3.74

7.68

4.69

4.32

4.43

4.52

4.40

4.56

4.40

106

99

4.45

00

4.47
6.68

Figure 4-1:
1: Saginaw Bay’s inner bay with depth ≥ 5. The gray dots indicate the portion
Saginaw Bay’s inner bay with depth ≥ 5 m that was used for statistical analyses. In the
lower left is the Saginaw River.

107

44.4
44.2
44.0
43.6

43.8

Latitude

-84.0

-83.8

-83.6

-83.4

-83.2

-83.0

-82.8

Longitude

Figure 4-2: Inner Saginaw Bay study area and river mouth inputs into Saginaw Bay.
Inner Saginaw Bay study area (gray dots) and river mouth inputs into Saginaw Bay (black
dots).

108

9
8
7
6
5
4

MODIS
Landsat
3

Annual upper quartile satellite-derived chlorophpyll (µg/L)

9
8
7
6
5
4

Annual average satellite-derived chlorophpyll (µg/L)

3

MODIS
Landsat

2000

2002

2004

2006

2008

2010

2012

2000

Year

2002

2004

2006

2008

2010

2012

Year

Figure 4-3. Annual average and upper quartile MODIS and Landsat-derived
chlorophyll concentration (µg/L) in inner Saginaw Bay, 2000-2012. Annual average
MODIS and Landsat-derived chl concentration (µg/L) in inner Saginaw Bay, 2000-2012
(left). Upper quartile MODIS and Landsat-derived chl concentration in inner Saginaw Bay,
2000-2012 (right).

109

8
7
6
5
3

4

Upper quartile of Landsat-derived chl concentration (µg/L)

9

9
8
7
6
5
4

Average Landsat-derived chl concentration (µg/L)

3

1980

1990

2000

2010

1980

Year

1990

2000

2010

Year

Figure 4-4: Annual average and upper quartile Landsat-derived chlorophyll
concentration (µg/L) in inner Saginaw Bay, 1973-2012. Annual average Landsatderived chl concentration (µg/L) in inner Saginaw Bay, 1973-2012 (left). Upper quartile
Landsat-derived chl concentration in inner Saginaw Bay, 1973-2012 (right).

110

150
100
0

50

Annual upper quartile of 8-day phosphorus loads(kg)

200

200
150
100
50

Annual average 8-day phosphorus loads (kg)

0

1980

1990

2000

2010

1980

Year

1990

2000

2010

Year

Figure 4-5: Annual average and upper quartile 8-day phosphorus load (kg) in inner
Saginaw Bay, 1973-2012. Annual average 8-day phosphorus load (kg) in inner Saginaw
Bay, 1973-2012 (left). Upper quartile 8-day phosphorus load (kg) in inner Saginaw Bay,
1973-2012 (right).

111

20
10

15

Upper quartile of Landsat-derived water temperature (˚C)

25

25
20
15

Average Landsat-derived water temperature(˚C)

10

1985

1990

1995

2000

2005

2010

1985

Year

1990

1995

2000

2005

2010

Year

Figure 4-6: Annual average and upper quartile Landsat-derived surface water
temperature (˚C) in inner Saginaw Bay, 1984-2012. Annual average Landsat-derived
surface water temperature (˚C) in inner Saginaw Bay, 1984-2012 (left). Upper quartile
Landsat-derived surface water temperature (˚C) in inner Saginaw Bay, 1984-2012 (right).

112

Figure 4-7: MODIS-derived chl vs. mixed-effect model chl (µg/L) in inner Saginaw
Bay, 2000-2012 with 1:1 line.

113

Figure 4-8: Landsat-derived chl vs. mixed-effect model chl (µg/L) in inner Saginaw
Bay, 1984-2012 with 1:1 line.

114

LITERATURE CITED

115

LITERATURE CITED

Austin, J. A. and S. M. Colman. 2007. Lake Superior summer water temperatures are
increasing more rapidly than regional air temperature: A positive ice-albedo
feedback. Geophysical Research Letters 34.
Bierman, V. J., Jr., D. M. Dolan, and R. Kasprzyk. 1984. Retrospective analysis of the response
of Saginaw Bay, Lake Huron, to reductions in phosphorus loadings. Environmental
Science and Technology 18:23-31.
Bierman, V. J., Jr., J. Kaur, J. V. DePinto, T. J. Feist, and D. W. Dilks. 2005. Modeling the role of
zebra mussels in the proliferation of blue-green algae in Saginaw Bay, Lake Huron.
Journal of Great Lakes Research 31:32-55.
Bridgeman, T. B., G. L. Fahnenstiel, G. A. Lang, and T. F. Nalepa. 1995. Zooplankton grazing
during the zebra mussel (Dreissena polymorpha) colonization of Saginaw Bay, Lake
Huron. Journal of Great Lakes Research 21:567-573.
Budd, J. W., T. D. Drummer, T. F. Nalepa, and G. L. Fahnenstiel. 2001. Remote sensing of
biotic effects: Zebra mussels (Dreissena polymorpha) influence on water clarity in
Saginaw Bay, Lake Huron. Limnology and Oceanography 46:213-223.
Carlson, R. E. 1977. A trophic state index for Lakes. Limnology and Oceanography 22:361369.
CCIW. 1972. Canadian Center for Inland Waters: Biological and chemical water quality dat
available from their various Great Lakes cruises.
Chander, G., B. L. Markham, and D. L. Helder. 2009. Summary of current radiometeric
calibration coefficients for Landsat MSS, TM, ETM+, and EO-1 ALI sensors. Remote
Sensing of Environment 113:893-903.
Danek, L. J. and J. H. Saylor. 1977. Measurements of the summer currents in Saginaw Bay,
Michigan. Journal of Great Lakes Research 3:65-71.
Dyble, J., G. L. Fahnenstiel, R. W. Litaker, D. F. Millie, and P. A. Tester. 2008. Microcystin
concentrations and genetic diversity of Microcystis in the lower Great Lakes.
Environmental Toxicology 23:507-516.
Fahnenstiel, G. L., T. B. Bridgeman, G. A. Lang, M. J. McCormick, and T. F. Nalepa. 1995a.
Phytoplankton productivity in Saginaw Bay, Lake Huron: Effects of zebra mussel
(Dreissena polymorpha) colonization. Journal of Great Lakes Research 21:465-475.
116

Fahnenstiel, G. L., G. A. Lang, T. F. Nalepa, and T. H. Johengen. 1995b. Effects of zebra mussel
(Dreissena polymorpha) colonization on water quality parameters in Sagainaw Bay,
Lake Huron. Journal of Great Lakes Research 21:435-448.
Fahnenstiel, G. L., D. F. Millie, J. Dyble, R. W. Litaker, P. A. Tester, M. J. McCormick, R.Rediske,
and D. Klarer. 2008. Microcystin concentrations and cell quotas in Saginaw Bay,
Lake Huron. Aquatic Ecosystem Health & Management 11:190-195.
Fanslow, D. L., T. F. Nalepa, and T. H. Johengen. 2001. Seasonal changes in the respiratory
electron transport system (ETS) and respiration of the zebra mussel, Dreissena
polymorpha in Saginaw Bay, Lake Huron. Hydrobiologia 448:61-70.
Freedman, P. L. 1974. Saginaw Bay: An evaluation of existing and historical conditions.
Pages 1-137 in U. S. E. P. Agency, editor. Environmental Protection Agency,
University of Michigan.
Hartig, J. H. and J. Horvath. 1982. A preliminary assessment of Michigan's phosphorus
detergent ban. Journal (Water Pollution Control Federation) 54:193-197.
Heath, R. T., G. L. Fahnenstiel, W. S. Gardner, J. F. Cavaletto, and S.-J. Hwang. 1995.
Ecosystem-level effects of zebra mussels (Dreissena polymorpha): An enclosure
experiement in Saginaw Bay, Lake Huron. Journal of Great Lakes Research 21.
Johengen, T. H., T. F. Nalepa, G. L. Fahnenstiel, and G. Goudy. 1995. Nutrient changes in
Saginaw Bay, Lake Huron, after the establishment of zebra mussel (Dreissena
polymorpha). Journal of Great Lakes Research 21:449-464.
Lavrentyev, P. J., W. S. Gardner, J. F. Cavaletto, and J. R. Beaver. 1995. Effects of the zebra
mussel (Dreissena polymorpha Pallas) on protozoa and phytoplankton from
Saginaw Bay, Lake Huron. Journal of Great Lakes Research 21:545-557.
Lehman, J. T., A. Bazzi, T. Nosher, and J. O. Nriagu. 2004. Copper inhibition of phytoplankton
in Saginaw Bay, Lake Huron. Canadian Journal of Fisheries and Aquatic Sciences
61:1871-1880.
Lofgren, B. L. and A. Gronewold. 2012. Water Resources.
Lowe, R. L. and R. W. Pillsbury. 1995. Shifts in benthic algal community structure and
function following the appearance of zebra mussels (Dreissena polymorpha) in
Saginaw Bay, Lake Huron. Journal of Great Lakes Research 21:558-566.
McKeon, J. B., R. H. Rogers, and V. E. Smith. 1977. Production of a water quality map of
Saginaw Bay by computer processing of Landsat-2 data.1045-1054.
MDNR. 1988. Saginaw River and Saginaw Bay. Page 588 in M. D. o. N. Resources, editor.
MDNR, Lansing, Michigan.
117

Moll, R. A., C. O. Davis, and C. L. Schelske. 1980. Phytoplankton productivity and standing
crop in the vicinity of the Lake Huron-Saginaw Bay front. Journal of Great Lakes
Research 6:232-246.
Munawar, M. and I. F. Munawar. 1982. Phycological studies in lakes Ontario, Erie, Huron,
and Superior. Canadian Journal of Botany 60:1837-1858.
Nalepa, T. F. and G. L. Fahenestiel. 1995. Driessina polymorpha in Saginaw Bay, Lake Huron
ecosystem: Overiew and perspective. Journal of Great Lakes Research 21:411-416.
Nalepa, T. F., J. A. Wojcik, D. L. Fanslow, and G. A. Lang. 1995. Initial colonization of the
zebra mussel (Dreissena polymorpha) in Saginaw Bay, Lake Huron: Population
recruitment, density, and size structure. Journal of Great Lakes Research 21:417434.
NASA. 2013. MODIS Specifications.in N. A. a. S. Administration, editor., National Aeronautics
and Space Administration.
Nicholls, K. H., G. J. Hopkins, and S. J. Standke. 1999. Reduced chlorophyll to phosphorus
ratios in nearshore Great Lakes waters coincide with the establishment of
dreissenid mussels. Canadian Journal of Fisheries and Aquatic Sciences 56:153-161.
Paerl, H. W. and J. Huisman. 2008. Blooms like it hot. Science 320:57-58.
Pillsbury, R. W., R. L. Lowe, Y. D. Pan, and J. L. Greenwood. 2002. Changes in the benthic
algal community and nutrient limitation in Saginaw Bay, Lake Huron, during the
invasion of the zebra mussel (Dreissena polymorpha). Journal of the North American
Benthological Society 21:238-252.
Pozdnyakov, D. V., A. V. Lyaskovsky, F. J. Tannis, and D. R. Lyzenga. 1999. Modeling of
apparent hydro-optical properties and retrievals of water quality in the Great Lakes
for SeaWiFS: A comparison with in situ measurements. Pages 2742-2744
Geoscience and Remote Sensing Symposium, 1999. IGARSS'99 Proceedings. IEEE
1999 International.
Rogers, R. H. 1975. Application of Landsat to the surveillance and control of eutrophication
in Saginaw Bay. National Aeronautics and Space Administration, Ann Arbor,
Michigan.
Rogers, R. H., N. J. Shah, J. B. McKeon, and V. E. Smith. 1976. Computer mapping of water
quality in Saginaw Bay with Landsat digital data. Proceedings of the American
Society of Photogrammetry, Annual Meeting:584-596.

118

Sarnelle, O., A. E. Wilson, S. K. Hamilton, L. B. Knoll, and D. F. Raikow. 2005. Complex
interactions between zebral mussel, Dreissena polylmorpha, and the harmful
phytoplankter, Microcystis aeruginosa. Limnology and Oceanography 50:896-904.
Scavia, D. 1979. Examination of phosphorus cycling and control of phytoplankton dynamics
in Lake Ontario with an ecological model. journal of the Fisheries Research Board of
Canada 36:1336-1346.
Schelske, C. L., L. E. Feldt, M. S. Simmons, and E. F. Stoermer. 1974. Storm induced
relationships among chemical conditions and phytoplankton in Saginaw Bay, Lake
Huron. Proceedings of the 17th Conference on Great Lakes Research:78-91.
Schelske, C. L. and J. C. Roth. 1973. Limnological survey of Lakes Michigan, Superior, Huron,
and Erie. University of Michigan, Ann Arbor, MI.
Skubinna, J. P., T. G. Coon, and T. R. Batterson. 1995. Increased abundance and depth of
submersed macrophytes in response to decreased turbidity in Saginaw Bay, Lake
Huron. Journal of Great Lakes Research 21:476-488.
Smith, V. E., K. W. Lee, J. C. Filkins, K. W. Hartwell, K. R. Rygwelski, and J. M. Townsend.
1977. Survey of chemical and biological factors in Saginaw Bay (Lake Huron). EPA
Contract Number R802685, Cranbrook Institute of Science; Environmental Research
Laboratory
Stoermer, E. F. and E. Theriot. 1985. Phytoplankton distribution in Saginaw Bay. Journal of
Great Lakes Research 11:132-142.
Suzuki, N. N., S. Endoh, M. Kawashima, Y. Itakura, C. D. McNabb, F. M. D'Itri, and T. R.
Batterson. 1995. Discontinuity bar in a wetland on Lake Huron's Saginaw Bay.
Journal of Freshwater Ecology 10:111-123.
USEPA and USGS. 2005. National Hydrography Dataset Plus - NHDPlus Version 1.0.
Van Sickle, J. and C. B. Johnson. 2008. Parametric distance weighting of landscape influence
on streams. Landscape Ecology 23:427-438.
Vanderploeg, H. A., J. R. Liebig, W. W. Carmichael, M. A. Agy, T. H. Johengen, G. L.
Fahnenstiel, and T. F. Nalepa. 2001. Zebra mussel (Dreissna polymorpha) selective
filtration promoted toxic Microcystis blooms in Saginaw Bay (Lake Huron) and Lake
Erie. Canadian Journal of Fisheries and Aquatic Sciences 58:1208-1221.
Vanderploeg, H. A., T. F. Nalepa, D. J. Jude, E. L. Mills, K. T. Holeck, J. R. Liebig, I. A.
Grigorovich, and H. Ojaveer. 2002. Dispersal and emerging ecological impacts of
Ponto-Caspian species in the Laurentian Great Lakes. Canadian Journal of Fisheries
and Aquatic Sciences 59.
119

Wynne, T. T., R. P. Stumpf, M. C. Tomlinson, R. A. Warner, P. A. Tester, J. Dyble, and G. L.
Fahnenstiel. 2008. Relating spectral shape to cyanobacterial blooms in the
Laurentian Great Lakes. International Journal of Remote Sensing 29:3665-3672.

120

CHAPTER 5
SUMMARY AND SYNTHESIS: USE OF REMOTE ESNSING IN ECOLOGICAL ASSESSMENT
National Lakes Assessment chlorophyll and Secchi depth predictive models
National Landsat boosted regression tree (BRT) models of chlorophyll (chl)
concentration and Secchi depth are able to predict chl better than linear regression models
for inland lakes in the United States (chl: R2 of 0.44 and a 0.76 ln-transformed µg/L RMSE;
Secchi depth: R2 of 0.52 and a 0.80 m RMSE). Applicability of the national BRT chl model for
ecological analysis was demonstrated with a comparable total phosphorus-chl relationship
whether using measured chl (R2 = 0.58 and 1.02 ln-transformed µg/L RMSE) or remotely
sensed chl (R2 = 0.56 and 1.04 ln-transformed µg/L RMSE). Using multiple bands and ratios
in the model increased predictive power. There is, however, much improvement to be
made in broad-scale chl and Secchi depth models for lakes with widely variable physical,
chemical, and biological attributes, including differing depth, dissolved organic carbon
levels, and algal community composition. Correlations of national BRT model residuals
with algal community composition, lake depth, and dissolved organic carbon levels, as well
as time between sampling and image capture indicate that other factors may need to be
addressed to further model development. Overall, subsetting the lakes into groups by
similar regional characteristics did not improve chl or Secchi model performance.
Next steps in models for satellite measures of chl in inland lakes using BRT should
consider incorporation of natural landscape variables, such as land use, into the model.
Also, sampling each lake at multiple locations and over several times may help understand
fluctuations algal biomass in each individual lake more clearly. Matching sample date as
closely as possible with Landsat passover date may also help to reduce the noise in the
relationship between predicted vs. measured chl. Finally, including field measurements of
121

water quality attributes, such as inorganic carbon, can help distinguish the chl signal from
other substances in the water. Remote sensing water quality assessment tools can be
valuable for limnological study, ecological assessment, and water resource management.

Landsat and MODIS boosted regression tree models for chlorophyll prediction in the
Great Lakes
Landsat and MODIS satellites are complementary platforms because of the long
history of Landsat operation and the finer spectral resolution and image frequency of
MODIS. Landsat and MODIS BRT models are better predictors of chl in the Great Lakes than
other, readily available chl models. BRT models are also relatively easy to produce and to
share. Of the three chl BRT predictive models that were made for the Great Lakes, the
MODIS model was most accurate (0.85 cross-validation R2 and 0.10 µg/L RMSE) and
compared well to other models in the literature. BRT models for Landsat ETM+ and TM
more accurately predicted chl (0.69 cross-validation R2 and 0.55 µg/L RMSE) than the MSS
model (0.66 cross-validation R2 and 0.73 µg/L RMSE). All our Landsat models had favorable
results when compared to models in the literature.
Though several of the current Great Lakes remote sensing models involving MODIS
and Landsat were not publicly available for a direct comparison with our models, a
comparison of this kind would be informative. Next steps should also include validation
with external data and tests to see how widely the models can be applied. Model
improvement may involve accounting for biological factors, such as position of the algae in
the water column (Kutser et al. 2008). More consistent time of day for sampling may allow
for more accurate detection with each image, and better means for comparison of longterm ecological trends. Also, application of BRT models could assist with discerning
122

different algal groups in the Great Lakes, whether they be used with Landsat, MODIS, or
with satellite products with higher spectral or spatial resolution (Bergmann et al. 2004,
Hunter et al. 2008, Wynne et al. 2008, Bracher et al. 2009) various algal types and
simulated responses from Landsat and other less-spectrally sensitive satellites. Accounting
for other color producing agents will elucidate the benefits of using all bands and ratios in
model creation. Also, further tests should be done to determine if models that use all bands
and ratios are less sensitive to shallow water due to incorporation of more spectral
information. Application of BRT chl predictive models to the lifetime of both Landsat and
MODIS could be useful understanding historical, long-term chl trends and to inform us of
how climate change may alter ecosystems in the future.

Using Landsat- and MODIS-derived chlorophyll to assess water quality in Saginaw
Bay from 1972-2012
Deriving remotely-sensed chl concentrations from Landsat and MODIS satellite
imagery enabled a long-term historical analysis of water quality trends in Saginaw Bay,
which has a history of water quality problems and continues to experience regular,
summertime algal blooms. Chl concentrations were predicted from Landsat 1973-2012 and
MODIS 2000-2012 with boosted regression tree analysis. In a mixed-effect model,
remotely-sensed chl concentration can be used with other physical, chemical, or biological
parameters to test specific hypotheses about what drives algal biomass. Linear regression
analyses show that Landsat-inferred chl decreased between 1973-1982 and that 8-day
phosphorus load increased between 1972-2012. Simple linear regression showed
influences of 8-day phosphorus load, distance to the Saginaw River mouth, and water
temperature on satellite-inferred chl. Mixed effects models also show significant effects of
123

8-day phosphorus load, distance to the Saginaw River mouth, and Landsat-inferred water
temperature on MODIS and Landsat-inferred chl, however, the results suggest that a large
proportion of the random spatial effects are unaccounted for.
In future studies, variables such as water circulation patterns or internal nutrient
cycling could be added to the mixed-effects model to help elucidate the cause for spatial
correlations in chl. Future models could also take into account the lag in the response of chl
concentration to precipitation events and nutrient exposure. We would expect this
connection to be particularly important in extreme weather events, including heavy storms
or drought. Having more chl data for spring blooms might also be informative in seeing
long-term trends. Finally, having satellite inferred-chl data for shallower depths could help
elucidate any relationships due to the proximity of external river water and nutrient inputs.
Using remotely-sensed chl in ecological analyses can help determine which factors have,
historically, been the largest drivers of algal biomass production in Saginaw Bay, and,
therefore, which may be the most critical for management in the face of climate change.

124

LITERATURE CITED

125

LITERATURE CITED

Bergmann, T., G. Fahnentiel, S. Lorhenz, D. Millie, and O. Schofield. 2004. Impacts of a
recurrent resuspension event and variable phytoplankton community composition
on remote sensing reflectance. Journal of Geophysical Research 109:1-12.
Bracher, A., M. Vountas, T. Dinter, J. P. Burrows, B. Schmitt, R. Röttgers, and I. Peeken. 2009.
Quantitative observation of cyanobacteria and diatoms from space using
PhytoDOAS on SCIAMACHY data. Biogeosciences 6:751-764.
Hunter, P. D., A. N. Tyler, M. Présing, A. W. Kovács, and T. Preston. 2008. Spectral
discrimination of phytoplankton colour groups: The effect of suspended particulate
matter and sensor spectral resolution. Remote Sensing of Environment 112:15271544.
Kutser, T., L. Metsamaa, and A. G. Dekker. 2008. Influence of the vertical distribution of
cyanobacteria in the water column on the remote sensing signal. Estuarine, Coastal,
and Shelf Sciences 78:649-654.
Wynne, T. T., R. P. Stumpf, M. C. Tomlinson, R. A. Warner, P. Tester, J. Dyble, and G. L.
Fahnenstiel. 2008. Relating spectral shape to cyanobacterial blooms in the
Laurentian Great Lakes. International Journal of Remote Sensing 29:3665-3672.

126