Robust Algorithm For Estimating Total Suspended Solids (TSS) in Inland and Nearshore Coastal Waters

Version of Record: https://www.sciencedirect.
com/science/article/pii/S0034425720301383
Manuscript_72b546bf345dfd46af4d40cc8c86d22b
1 Robust Algorithm for Estimating Total Suspended Solids (TSS) in

2 Inland and Nearshore Coastal Waters
3
4 Sundarabalan V. Balasubramaniana,c,d, Nima Pahlevana,b,*, Brandon Smitha,b , Caren Bindinge,
5 John Schallesf, Hubert Loiselg, Daniela Gurlinh, Steven Grebi, Krista Alikasj, Mirjam Randlaj,
6 Matsushita Bunkeik, Wesley Mosesl, Hà Nguyễnm, Moritz K. Lehmannn, David O’Donnello,
7 Michael Ondrusekp, Tai-Hyun Hanq, Cédric G. Fichotr, Tim Moores, and Emmanuel Bosst
8
a
9 NASA Goddard Space Flight Center, Greenbelt, MD USA
b
10 Science Systems and Applications, Inc. (SSAI), Lanham, MD, USA
c
11 University of Maryland, Department of Geographical Sciences, College Park, MD, USA
d
12 Geo-Sensing and Imaging Consult, Trivandrum, Kerala, India
e
13 Environment and Climate Change Canada, Burlington, ON, Canada
f
14 Creighton University, Department of Biology, Omaha, NE, Nebraska
g
15 Univ. Littoral Côte d’Opale, Univ. Lille, CNRS, UMR 8187, LOG, Laboratoire d’Océanologie et de
16 Géosciences, Lille, France
h
17 Wisconsin Department of Natural Resources, Madison, WI, USA
i
18 University of Wisconsin-Madison, Space Science and Engineering, Madison, WI, USA
j
19 Tartu observatory of University of Tartu, Tartu, Estonia
k
20 University of Tsukuba, Ibaraki, Japan
l
21 Naval Research Laboratory, Washington, D.C., USA
m
22 VNU University of Science, Vietnam National University, Hanoi, Vietnam
n
23 Xerra Earth Observation Institute and the University of Waikato, Hamilton, New Zealand
o
24 Upstate Freshwater Institute, Syracuse, NY, USA
p
25 NOAA Center for Satellite Applications and Research, College Park, MD, USA
q
26 Korea Ocean Satellite Center KIOST, Busan, South Korea
r
27 Department of Earth and Environment, Boston University, Boston, MA, USA
s
28 Ocean Process Analysis Laboratory, University of New Hampshire, Durham, NH, USA
t
29 University of Maine, School of Marine Sciences, Orono, ME, USA
30
31
32 ABSTRACT:
33 One of the challenging tasks in modern aquatic remote sensing is the retrieval of near-surface
34 concentrations of Total Suspended Solids (TSS), which indirectly decrease the rate of primary
35 production by attenuating the propagation of the underwater light field. This study aims to
36 present a Statistical, inherent Optical property (IOP) -based, and muLti-conditional Inversion
37 proceDure (SOLID) for enhanced retrievals of satellite-derived TSS under a wide range of in-
38 water bio-optical conditions in rivers, lakes, estuaries, and coastal waters. In this study, using a
39 large in situ database (N > 3500), the SOLID model is devised using a three-step procedure: (a)
40 water-type classification of the input remote sensing reflectance ( ), (b) retrieval of particulate
© 2020 published by Elsevier. This manuscript is made available under the Elsevier user license
https://www.elsevier.com/open-access/userlicense/1.0/
41 backscattering ( ) in the red or near-infrared (NIR) regions using semi-analytical, machine-
42 learning, and empirical models, and (c) estimation of TSS from via water-type-specific
43 empirical models. Using an independent subset of in situ data (N = 2729) with TSS ranging from
44 0.1 to 2626.8 [g/m3], the SOLID model is thoroughly examined and compared against the state-
45 of-the-art algorithms (Miller and McKee 2004; Nechad et al. 2010; Novoa et al. 2017; Ondrusek
46 et al. 2012; Petus et al. 2010). We show that SOLID outperforms all the other models to varying
47 degrees (from 10 to > 100%) based on global and water-type-specific statistical attributes. For
48 demonstration purposes, the model is implemented for images acquired by the MultiSpectral
49 Imager aboard Sentinel-2A/B over the Chesapeake Bay, San-Francisco-Bay-Delta Estuary, and
50 Lake Okeechobee. To enable the generation of consistent, multimission TSS products, its
51 performance is extended, and evaluated further for missions, such as the Ocean and Land Color
52 Instrument (OLCI), Moderate Resolution Imaging Spectroradiometer (MODIS), Visible Infrared
53 Imaging Radiometer Suite (VIIRS), and Operational Land Imager (OLI). Sensitivity analyses on
54 uncertainties induced by the atmospheric correction indicate that 10% uncertainty in leads to
55 < 20% uncertainty in TSS retrievals. While this study suggests that SOLID has the potential for
56 producing TSS products in global coastal and inland waters, our extensive analysis certainly
57 verifies that there is still a need for improving retrievals across the wide spectrum of particle
58 loads.
59
60 1. Introduction
61 Suspended sediments in coastal and inland waters are introduced by various sources,
62 including river runoffs, dredging activities, resuspension events, and tidal currents. The sediment
63 composition is, in general, a mixture of organic and inorganic particles in the water column
2
64 (Miller and McKee 2004). High sediment loads can lead to poor water quality and potentially
65 increase temperature in the upper layer of the water column (Turner and Millward 2002).
66 Suspended solids can carry heavy metals, pollutants, and nutrients, and, therefore, contribute to
67 adverse environmental conditions in the water column. The excessive near-surface accumulation
68 of this optically significant water constituent affects light propagation into deeper layers and the
69 benthos diminishing the productivity and altering the ecosystem functioning. Conversely,
70 sediment loads are critical for maintaining sediment accretion rates and protecting the integrity
71 of sediment-maintained geomorphic features such as river deltas and marsh and mangrove
72 wetland platforms. These coastal features are damaged when sediment dynamics are
73 compromised by upstream trapping of suspended sediments by reservoirs, river channelization
74 and disturbance of delta distributary flows, and sea level rise. These geomorphic features support
75 and protect blue-carbon storages of coastal wetland sediments and are important defenses against
76 storm surge and high spring tide flooding events (Barbier et al. 2011; Weston 2014). Monitoring
77 sediment fluxes is, thus, essential for the sustainable management of coastal and inland water
78 ecosystems.
79 The sediment concentration is commonly quantified via laboratory analyses of grab samples
80 and expressed as the concentration of total suspended solids (TSS; [g/m3]), which is also referred
81 to as suspended particulate matter (SPM) and total suspended matter (TSM) in the literature.
82 Note that TSS includes living and detrital (non-living) particulate organic matter, such as
83 phytoplankton, and inorganic matter like clay and other suspended minerals. Due to the dynamic
84 nature of its spatial and temporal distribution, TSS quantified through field sampling is often
85 considered inadequate (Doxaran et al. 2014); hence, aquatic (optical) remote sensing is used as
86 an efficient proxy for its monitoring at local, regional, and global scales (Ahn et al. 2006;
3
87 Bowers and Binding 2006; Doxaran et al. 2009a; Feng et al. 2014; Forget and Ouillon 1998;
88 Loisel et al. 2014; Woźniak and Stramski 2004). One of the main products of aquatic color
89 remote sensing is the spectral remote sensing reflectance ( ), which carries information about
90 the bulk optical properties of near-surface water constituents, including TSS. The is defined
91 as the ratio of water leaving radiance to the downwelling irradiance just above the water and can
92 be related to aquatic reflectance ( ) assuming an isotropic upwelling radiance field, i.e., =
93 / (Mobley 1999). from satellite observations is computed following the removal of
94 atmospheric effects from the top of atmosphere (TOA) reflectance/radiance measurements made
95 by remote sensors (Gordon and Wang 1994). Aquatic biogeochemical products like near-surface
96 concentration of chlorophyll-a (Chla) and TSS together with the inherent optical properties
97 (IOPs) are obtained from products.
98 Over the last decades, several analytical, empirical, and semi-empirical relationships have
99 been devised to retrieve near-surface TSS (hereafter, TSS) using (or another equivalent
100 representation termed as the normalized water-leaving radiance, nLw) at either a single band or a
101 combination of bands (Binding et al. 2010; Dekker et al. 2002; Doxaran et al. 2009a; Han et al.
102 2016; Nechad et al. 2010; Novoa et al. 2017; Tassan 1993; Zhang et al. 2016). Site-specific
103 studies often utilize a single red band for TSS retrievals, which can provide reasonably accurate
104 estimates within limited TSS ranges. For example, Ouillon et al. (2004) applied a linear
105 regression analysis to map TSS within the 0 ~10 [g/m3] range in the southwest lagoon of New
106 Caledonia. Similarly, Miller and McKee (2004) estimated TSS in the Mississippi River Delta,
107 where TSS ranged from 0 to 60 [g/m3]. Kumar et al. (2016) and Ondrusek et al. (2012) utilized
108 higher order polynomials to generate spatial distributions of TSS in Chilika Lake, India (0 ~ 100
109 [g/m3]) and in the Chesapeake Bay (0 ~ 60 [g/m3]), respectively. These single band models are
4
110 easy to implement and straightforward for operational satellite services. However, the
111 performance of these algorithms degrades in areas with extreme sediment loads, where
112 radiometric measurements in the red band no longer correlate with increases in TSS, i.e., a
113 saturation effect is present (Feng et al. 2014; Luo et al. 2018; Ritchie et al. 2003; Shi and Wang
114 2009). Under such circumstances, ( > 700 nm) is commonly relied upon (e.g., 865, 1020,
115 and 1071 nm) (Knaeps et al. 2015). A major limitation of this NIR-based single band model is
116 that it generally performs poorly for low to moderate TSS (i.e., < 50 [g/m3]) (Han et al. 2016)
117 owing to the negligible contribution of particulate backscattering ( ) relative to the pure water
118 absorption (Doxaran et al. 2012), and lack of radiometric sensitivity in this spectral region. To
119 overcome this limitation, various multiband models based on band ratio or other band arithmetic
120 operations applied in empirical or semi-analytical models have been developed (Chen et al.
121 2015; Dekker et al. 2001; Doxaran et al. 2003; Feng et al. 2014; Novoa et al. 2017; Oyama et al.
122 2009). The main drawback of these methods is that each band is not only sensitive to TSS but
123 also to other optically significant constituents in the water column (e.g., Chla), which leads to
124 overestimation or underestimation of TSS. Alternatively, TSS may be derived from turbidity, a
125 proxy water quality parameter, at regional scales (Dogliotti et al. 2015) as empirical relationships
126 may vary seasonally due to phytoplankton growth (Hannouche et al. 2017).
127 In addition to approaches that derive TSS directly from , there are methods that
128 approximate TSS directly from either the particulate absorption coefficient (ap) (Babin et al.
129 2003; Zhang et al. 2010; Zheng and DiGiacomo 2017) or (Binding et al. 2010; Volpe et al.
130 2011; Woźniak et al. 2010) by approximating mass-specific particulate absorption (a∗ ) and
131 backscattering (b∗ ) coefficients, respectively. Semi-analytical methods that approximate
132 physics-based bio-optical parameterization provide another avenue for TSS retrievals by solving
5
133 for IOPs. For instance, widely used inversion models such as the Generalized IOP (GIOP)
134 (Werdell et al. 2013), the Quasi Analytical Algorithm (QAA) (Lee et al. 2002), and the Garver-
135 Seigel-Maritorena (GSM) (Maritorena et al. 2002) provide fairly accurate estimates of in
136 clear and/or moderately turbid waters but are less accurate in highly turbid/eutrophic waters
137 (Shanmugam et al. 2010; Zheng and DiGiacomo 2017). Under intense algal bloom conditions,
138 for example, phytoplankton backscattering dominates (600 nm < < 800 nm) and, as a
139 result, the overall magnitude/shape of in this region (Binding et al. 2010; Shi et al. 2018).
140 This leads to ambiguities in TSS models developed in the absence of high phytoplankton
141 concentrations when they are applied to waters with high concentrations of phytoplankton.
142 Several studies have attempted to fine-tune QAA for IOP retrievals in a few inland and coastal
143 waters (Joshi and D'Sa 2018; Mitchell et al. 2016; Mouw et al. 2013); nevertheless, these
144 algorithms and their performances require further independent verifications. Most methods
145 neglect impacts of the composition and size-distribution of particles on IOP spectra resulting in
146 inaccurate retrievals of TSS (Bowers and Binding 2006; Long and Pavelsky 2013; Neukermans
147 et al. 2012; Novo et al. 1989). Nonetheless, TSS remains a parameter of interest to estimate via
148 remote sensing.
149 Recognizing the lack of a) a global dataset for a thorough assessment of existing TSS
150 algorithms and b) a robust algorithm applicable to waters with a wide range of near-surface
151 particle load or waters with different particle types, this article offers an innovative hybrid
152 approach termed the Statistical, IOP-based muLti-conditional Inversion proceDure (SOLID),
153 which employs (600 nm < < 800 nm) retrieved from remote sensing as a proxy for TSS
154 retrievals. The retrieval begins by assigning input to one of three water types according to its
155 shape/magnitude (Section 3). For each of the assigned water types, a corresponding procedure is
6
156 followed to retrieve . A novel machine learning algorithm is proposed for estimating
157 (600 nm < < 700 nm) over a broad range of turbidity and trophic conditions (Section
158 4.1.1), whereas (700 nm < < 800 nm) is analytically solved for in waters mainly
159 dominated by suspended sediments (Section 4.1.2). In aquatic ecosystems (e.g., coastal waters),
160 where low TSS and/or Chla (i.e., < 2 units of concentrations) is commonly found, QAA is
161 applied for estimating (600 nm < < 700 nm). Then, the retrieved b (600 nm < λ <
162 800 nm) is empirically attributed to TSS via water-type-specific expressions (Section 4.2). The
163 algorithm is developed using a large dataset consisting of synthetic data and in situ
164 measurements, and is evaluated with > 2700 paired in situ and TSS samples. This approach
165 is further compared against several state-of-the-art algorithms (Miller and McKee 2004; Nechad
166 et al. 2010; Novoa et al. 2017; Ondrusek et al. 2012; Petus et al. 2010), and is demonstrated for a
167 handful of satellite missions to allow for seamless retrievals of TSS via a single blended
168 algorithm (SOLID). While the performance of the algorithm is mainly demonstrated for the
169 MultiSpectral Instrument (MSI) aboard Sentinel-2A/B, we will further extend our analysis to
170 other satellite missions (Section 5.4), including the Ocean and Land Color Instrument (OLCI),
171 Moderate Resolution Imaging Spectroradiometer (MODIS), Visible Infrared Imaging
172 Radiometer Suite (VIIRS), and Operational Land Imager (OLI) (Section 5.4).
173 2. Datasets
174 The data consisted of simulated data (Pahlevan et al. 2017d), in situ measurements, and
175 satellite images, representing a broad range of turbidity and trophic conditions. In situ
176 measurements (Fig. 1) represented waters with intense algal blooms (e.g., Lake Erie, Lake
7
177 Champlain, Lake Taihu) and very high turbidity (e.g., Red River and San-Francisco-Bay-Delta
178 Estuary).
179
180
181
182
183
184 Fig.1. The global distribution of in situ datasets (N = 3288) in addition to the NOMAD datasets (N=222)
185 used for testing, training, and validating (red squares) the SOLID model. The yellow boxes indicate areas
186 with training/testing datasets only, whereas validation data are specified with black boxes.
187
188
189 These ecosystems represented in the dataset varied in the amount of nutrients, organic matter
190 concentrations, productivity, biodiversity, climate, and watershed biogeochemical/physical
191 characteristics, which enable a comprehensive assessment of SOLID and other widely used TSS
192 algorithms (Miller and McKee 2004; Nechad et al. 2010; Novoa et al. 2017; Ondrusek et al.
193 2012; Petus et al. 2010). These in situ datasets include the remote sensing reflectance, [1/sr],
194 as well as bio-optical and biogeochemical data (Table 1).
195
196 2.1 Development data
197 Synthetic data: The radiative transfer model Hydrolight (Mobley 2008) was used to simulate a
198 large database (N=915,000) of hyperspectral (350 nm < < 800 nm) at 5-nm intervals
8
199 associated with various optical conditions. To do so, various specific absorptions of
∗ ∗ ∗
200 phytoplankton ( ! ), specific absorptions/scattering of non-algal particles ( "# , "# ),
201 particulate backscattering fraction ( $ ) (0.01 < $ < 0.03) along with uniformly distributed
202 Chla (0.01 ~ 100 [mg/m3]), TSS (0.0 ~ 98.4 [g/m3]), and the absorption by colored dissolved
203 organic matter '()* (440) (0.0009 ~ 6 [1/m]) with exponents in the range of 0.009 ~ 0.031
∗
204 [1/nm] were supplied as model inputs (Pahlevan et al. 2017d). The ! (440) ranged from 0.06
∗
205 to 0.09 [m2 /mg] while ! (667) varied between 0.02 and 0.06 [m2 /mg]. For NAP specific
206 absorption and scattering spectra, the five default spectra (e.g., (Bukata et al. 2018)) available in
207 Hydrolight were employed. This simulated dataset was used to train a machine learning model
208 for retrievals (Section 4.1.).
209 Field data: In situ data, such as , , and TSS required for algorithm development and
210 testing collected in various inland and coastal waters (Table 1) were compiled. The were
211 estimated from measurements made by various above- or in-water radiometers manufactured by
212 Ocean Optics, Inc., Trios, Sea-Bird Scientific, and ASD, Inc. The measurements were post-
213 processed according to instrument-specific protocols and standard procedures developed by the
214 manufacturing companies or the scientific community (Lee et al. 2013; Mobley 1999; Mueller et
215 al. 2004). Multispectral measurements were taken with ECO BB9 backscattering sensors
216 (WetLabs Inc.; Sea-Bird Scientific). These data, available for narrow spectral bands centered at
217 variable spectral positions (Binding et al. 2019; Moore et al. 2017; Mouw et al. 2013; Reynolds
218 et al. 2016) were linearly oversampled to 1-nm spectral spacing, allowing for resampling
219 measured spectra to corresponding satellite sensor spectral response functions (see Section
220 4.1.2). Overall, 246 pairs of and spectra were available to validate estimates (Section
221 5.1). TSS was determined gravimetrically using the standard technique and represent the dry
9
222 mass particles per unit volume of water. In situ TSS - measurement pairs (N = 607) were
223 further utilized to develop empirical relationships for TSS retrievals (Table 1). The data
224 originated from Lake Erie, Green Bay (WI), Lake Michigan (MI) from US inland waters (N =
225 150), Lake Taihu (N = 164) in China (Shi et al. 2018), and the SeaWiFS Bio-optical Archive and
226 Storage System (SeaBASS) (N = 293). In sediment-rich waters (where TSS is commonly > 50
227 [g/m3]), we applied all available , TSS and data (N = 112), acquired in European waters
228 (Knaeps et al. 2018) (N = 81) and in the Red River (N=31) in Vietnam (Pham et al. 2018).
229
230 Table 1. Summary of statistical attributes associated with in situ datasets for training, testing, or
231 validating (665) and TSS retrievals applied for the three water types. Note that the validation dataset
232 is employed for all three water types. See Section 3 for the definition of water types.
Mean Median Max Min N
Training & Testing Dataset (Type I & II)
3
TSS [g/m ] 14.6 1.94 113.1 0.039 607
Chla [mg/m3] 6.11 1.93 161 0.20 460
(665)* [1/m] 0.17 0.004 2.03 0.0006 607
(665) * [1/sr] 0.008 0.0071 0.03 0.001 151
Training & Testing Dataset (Type III)
TSS [g/m3] 150.0 106.8 1190 49.17 112
Chla [mg/m3] 10.2 8.1 24.7 1.3 35
(740) [1/m] 1.05 0.67 3.92 0.26 112
(740) [1/sr] 0.019 0.014 0.05 0.006 112
Validation Dataset
TSS [g/m3] 30.7 7.94 2626.8 0.102 2729
Chla [mg/m3] 19.9 6.52 490 0.02 1916
(665) [1/sr] 0.0103 0.006 0.084 0.001 2729
233 *
The band center refers to Sentinel-2A red channel.
234
235 For validating the TSS retrievals, a large independent database (N = 2729) of – TSS
236 measurements, whose frequency distributions are illustrated in histograms in Fig. 2, were
237 applied. These datasets were collected in various open ocean/coastal/inland regions over the
238 globe, such as Río de La Plata and French Guiana (South America), the Gulf of Mexico, the San-
239 Francisco-Bay-Delta Estuary, Chesapeake Bay, and the Plum Island Estuary (North America),
10
240 the south Atlantic Bight, the English Channel and French nearshore coastal waters, the Estonian
241 inland waters and Baltic Sea (Europe), coastal waters of South Korea, lakes in New Zealand, and
242 inland and bay waters of Vietnam (Asia) (Fig. 1). In addition, we used a subset of data available
243 in the NASA bio-Optical Marine Algorithm Dataset (NOMAD) (Werdell and Bailey 2005).
244
245
246 Fig. 2. Frequency distribution of validation dataset (N = 2729): TSS (a) and Rrs (665).
247
248 2.2 Image data
249 Optical remote sensing images over select coastal and inland waters from the MSI were
250 obtained from the United States Geological Survey (USGS) image portal
251 (https://earthexplorer.usgs.gov/) and processed to using the SeaWiFS Data Analysis System
252 (SeaDAS v7.5.3) with the 1609-2200 nm band combination for aerosol removal (Pahlevan et al.
253 2017b; Pahlevan et al. 2017c) to allow for the use of 865 nm for TSS retrievals in highly turbid
254 waters. To extend the performance analysis in extremely turbid waters, MSI images were also
255 processed using ACOLITE (v20190326.0), which is commonly used in such environments
256 (Vanhellemont and Ruddick 2014). Corrected images were processed to TSS fields for
257 assessment of realistic spatio-temporal patterns according to expert expectations.
258 3. Water-type classification
11
259 Following Novoa et al. (2017) and our objective to devise an algorithm applicable to a broad
260 spectrum of TSS conditions, we conducted a simple water-type classification using available in
261 situ spectra. We classified waterbodies into three basic types determined by the shape of .
262 The classification rules, designed experimentally following trials and errors, are summarized in
263 Table 2.
264 Table 2. Classification rules for water type determination

Order Rule Class
1 If (665) < ./ (560) & (665) > (492) ./ ( ) ∈ Type II
2 If (665) > (560) & (740) > 0.01 [1/sr] ./ ( ) ∈ Type III
3 If (560) < (492) ./ ( ) ∈ Type I
4 If #3 is false ./ ( ) ∈ Type II
265
266 These rules are applied in a specific order (e.g., input spectrum is first examined for its
267 assignment to Type II; if the condition is not met then the spectrum is assessed for Type III). The
268 classification scheme is based only upon three bands (blue, green, and red) to assure its utility for
269 heritage Landsat-class missions not equipped with red-edge bands. The three broad categories
270 include: Blue-green waters (Type I); falling into this category commonly exhibit very low
271 magnitudes within the red region, i.e., ./ (560) < ./ (492). These are normally optically
272 mixed waters with no single dominant water constituent. In clearer waters, is characterized
273 with peaks in the blue region whereas in more turbid waters a peak in the blue-green region
274 commonly exists (Fig. 3). Green waters (Type II); The characteristic peak in for these water
275 types is within the green region due to the elevated total absorptions in the blue and red portions
276 of the spectrum (Fig. 3). In these water types, the presence of CDOM further increases
277 absorption within the blue region lowering upwelling radiance compared to that in the red, i.e.,
278 ./ (665) > ./ (492). Brown water (Type III); The primary peak in is shifted to the red
279 region i.e., ./ (665) > ./ (λ). Here, the magnitude of in the red region is almost always
12
280 greater than those in the blue and green regions. This enhanced response is mainly because of the
281 increased backscattering in presence of sediments or non-algal particles, as well as the
282 dampening effects of CDOM at lower wavelengths (Schalles 2006). These spectral features are
283 expected in river runoffs, estuaries, and bays with significant sediment loads (Doxaran et al.
284 2003; Gernez et al. 2014) (Fig. 3). Further analysis of spectra suggested the need for an
285 additional criterion to enhance discrimination of water types II and III.
286
13
287
288 Fig. 3. A block diagram showing the sequence of our hybrid retrieval strategy (i.e., SOLID). The input
289 spectrum is first assigned to a water type followed by applying retrieval algorithms, i.e., the Quasi
290 Analytical Algorithm (QAA), the Mixture Density Network (MDN), and a NIR-based Inversion
291 Approach (NIR-IA). The TSS is then estimated from using empirical relationships. is normalized
292 by the peak value.
293
294 The misclassification arises from enhanced backscatter in the vicinity of 740-nm channel in
295 hypereutrophic waters (Gitelson et al. 2008). Hence, another condition, i.e., (740) >
296 0.01 [1/sr], was added to avert this misclassification. Given our paired and TSS dataset, TSS
14
297 > 65 [g/m3] commonly falls in Type III category.
298
299 4. Methodology
300 To estimate TSS from , the methodology followed in this study is explained below in two
301 major steps: (a) deriving from , and (b) estimating TSS from . The schematic
302 diagram for the three steps is shown in Fig. 3. Recognizing that commonly used IOP retrieval
303 algorithms are designed to function in oceanic waters (Werdell et al. 2018), we devise new
304 approaches for retrievals in Type II and Type III waters.
305 4.1 89: inversion
306 The in situ measured data available in this study can be expressed as the subsurface
307 remote sensing reflectance (; ) by accounting for air-water transmittance as below (Lee et al.
308 2002):
<=>
309 ; = (1)
?.@ABC.D<=>
310
311 This quantity is analytically related to the ratio of total backscattering ( ) to the sum of and
312 total absorption (a) (Gordon et al. 1988; Lee et al. 2002). Using the constant model parameters
313 (EC = 0.0949 and EA = 0.0794), the relationship can be expressed as follows:
(I) (I) A
314 ; ( ) ≈ EC GJ(I)B
H
(I)
K + EA GJ(I)B
H
(I)
K (2)
H H
315
316 where and can be further decomposed into:
317
318 ( )= M( ) + !( ) + '()* ( )+ "# ( ) (3)
319 ( )= M( ) + ( ) (4)
320
321 In above equations, M is the pure water absorption (Pope and Fry 1997), ! stands for
322 phytoplankton absorption, and "# represents absorption due to non-algal particles. Similarly,
15
323 includes M, the pure water backscattering, which is half of the pure water scattering ( /2)
324 (Smith and Baker 1978), and is the particulate backscattering due to algal and non-algal
325 constituents. These fundamentals will be referenced when deriving for Type III waters in
326 Section 4.1.2.
327 4.1.1 Machine learning approach: Mixture density network
328 In the past, artificial neural networks have been relied upon for component IOP inversions in
329 optically complex waters (D'Alimonte et al. 2012; Ioannou et al. 2011). Here, for Type II waters,
330 using the synthetic data described in Section 2.1, we train a machine learning model termed the
331 Mixture Density Network (MDN) for retrievals (Section 3 and Table 1). MDNs are a class of
332 neural networks for modeling a mixture of Gaussian functions (Bishop 1994). Instead of directly
333 outputting the target variable (e.g., ), this network generates a set of three variables per
334 Gaussian (mean, standard deviation, and mixing coefficients), where the number of Gaussians is
335 a tuning parameter. The Gaussians are then combined to form the final output estimation, via
336 either a probabilistic combination, or the maximum likelihood. MDNs are intended to model a Y
337 to X mapping; that is, in contrast to many standard machine learning models, whose main intent
338 is to find a function mapping X to Y. The primary difference between the two is that in the latter
339 case, there may be many X values for any particular Y value (e.g. a sine wave). In the former
340 case, modeled by an MDN, the situation is reversed: there may be many Y values for any
341 particular X value (e.g., an arcsine wave). The implemented MDN model learns the full
342 covariance matrix between spectral bands, avoiding ambiguities in retrievals (Sydor et al.
343 2004; Yang et al. 2011). To state this another way, a given value for an arbitrary channel
344 may be consistent with multiple different possible values; without more spectral information,
345 the probability of all of these values are potentially equal. A standard machine learning model
16
346 may, in the worst case, simply take the average of all these disparate values. The MDN,
347 however, learns to associate the spectrum with multiple parameters; thus, enabling a choice
348 of a spectrum, which is more likely, conditional upon all spectral bands. A schematic
349 diagram of the model is shown in Fig. 4.
350 Input to the model consists solely of the spectra, which are normalized based on median
351 centering and interquartile scaling. The learned output variables are subject to the same scaling
352 method, but also subsequently scaled to a [0.1, 1] range. This normalization method is robust to
353 outliers which may be present in the data, while also bringing network outputs into a range,
354 which is amenable to activation functions, such as rectified linear units (ReLU); thus, helping to
355 avoid the dying ReLU problem (Agarap 2018). The output of the model consists of spectra.
356
357
358 Fig. 4. Block diagram of the MDN network used for the retrieval of particulate backscattering ( ) from
359 satellite-derived for Type II waters.
360
361 There are a number of hyperparameters to tune, including the number of Gaussian distributions
362 which are modeled, the Gaussian mixing coefficients, as well as all standard neural network
363 hyperparameters. These choices appear to be fairly robust to changes within the current
364 implementation, especially with regard to the MDN-specific options. The current default model
17
365 uses a five-layer neural network with 25 neurons per layer, which is trained to output the
366 parameters of five Gaussians. From this mixture model, the overall estimate is selected via the
367 Gaussian with the maximum prior likelihood. The described model is trained a number of times
368 with random initializations of the underlying neural network, in order to ensure a stable final
369 output of the estimates. The median estimate is taken as the final result of these trials, with
370 convergence occurring within some small margin of error after approximately ten trials. Note
371 that none of these values are required at test time – they are used as output variables, and so are
372 produced by the model. The only input variables required are the values at five different
373 bands (443, 492, 560, 665, 705nm).
374
375 4.1.2 NIR-based Inversion Approach (NIR-IA)
376 For Type III waters, due to the absence of available in situ ( > 700 nm) driven by
377 saturation effects associated with backscattering sensors (Doxaran et al. 2016), we adopt a
378 different strategy. Applying the component models to Eqs.1-4, a modeled spectrum for a
379 NIR band of MSI can be constructed as
M (740) + (740)
380 ./ (740) = (N/O) (5)
M (740) + "# (740) + M (740) + (740)
381 Here, M (740) = 2.72 [1/m] (Pope and Fry 1997) and M (740) = 0.00025 [1/m] (Smith and
382 Baker 1978) are the pure water absorption and backscattering coefficients, and the N/O factor is
383 assumed constant, i.e., 0.105 (Morel and Prieur 1977), although variability in N/O does exist in
384 coastal and inland waters (Loisel and Morel 2001; Morel et al. 1995). According to the literature
385 (Babin and Stramski 2004; Doxaran et al. 2009b; Estapa et al. 2012), we adopted ∗
"# (740) =
386 0.011 [mA ⁄g], which gives rise to "# (740) = 1.65 [1/m] given the mean TSS value of 150
387 [g/m3] for our dataset (Table 1). Under these assumptions, (740) can be obtained by re-
18
388 arranging the above equation. The discussion on the uncertainties introduced via this assumption
389 is provided in Section 6.1. The band centers of the relative spectral response (RSR) functions
390 used for retrievals for all the satellite missions considered here are summarized in Table 3.
391 For Landsat-class missions, where a 740-nm channel is absent, we develop a very similar
392 approach applicable to the 865-nm channel. A power-law model fitted to the global statistics
∗
393 corresponding to an average "# spectrum (see Table 5 in Doxaran et al. (2009b)) was
394 employed to extrapolate ∗

"# (740) to 865 nm channel. This extrapolation resulted in
395 "# (865) = 0.55. Together with M (865) = 4.6 [1/m] and M (740) = 0.00014 [1/m], Eq. 5
396 was solved for (865).
397
398 Table. 3 Band centers (nm) of RSR functions of satellite mission considered in this study
MSI VIIRS MODIS OLCI OLI
Type I 665 672 667 665 655
Type II 665 672 667 665 655
Type III 740 748 748 754 865
399
400
401 4.2 TSS retrieval
402 With the knowledge of , we construct an empirical model to obtain TSS estimates
403 applicable to both Type I and II waters, and one model for Type III. We found a strong power-
404 law correlation (R2 = 0.97) between (665) and TSS in both Type I and II waters from Lake
405 Erie (N = 150; with 0.13 < TSS < 43.1 [g/m3] and 0.2 < Chla < 57 [mg/m3]). Fig. 5a illustrates
406 this strong correlation on a log-log plot. The regression equation is as follows:
407 TUU = 53.736 × (665)?.W@@X (6)
408 To affirm this relationship for a wider dynamic range, additional paired (665) and TSS data
409 from highly productive waters of Lake Taihu (N = 164, Shi et al., 2018), the Arctic region (N =
19
410 125, (Reynolds et al. 2016)), and U.S. coastal waters (Casey 2020; Wei et al. 2016) were applied.
411 The data overlaid onto Fig. 5a further indicate that the relationship, to a large extent, holds (R2 =
412 0.82) in intense algal bloom conditions (Type II).
413
414 Fig. 5. (a) A power-law function (R2 = 0.97) for TSS - (665) in Type I (blue-green) and II (green)
415 waters (Type II) (n=607). (b) Scatterplot corresponding to in situ TSS and bbp(740) datasets (N = 112)
416 collected in sediment-dominated waters (Type III). Note that the fitted lines are displayed in log-log
417 scale.
418
419
420 In Type III waters, TSS was estimated via (740), which is derived directly from in situ
421 measured (740) (Section 4.1.2.; Eq. 5) using SEASWIR data (N = 78) (Knaeps et al. 2018)
422 in sediment-rich waters (i.e., Scheldt River, Gironde River, Rio de La Plata). Fig. 6b shows the
423 (log-log) scatterplot for the derived TSS - (740) relationship demonstrating a fairly strong
424 relationship, i.e., R2 = 0.83, expressed as follows:
425 TUU = G207.57 × (740)K Y 46.78 (7)
426 This least square fit suggests the utility of this relationship for a wide range of TSS and
427 (740) from 65 ~ 1300 [g/m3] and from 0.3 to 5.0 [1/m], respectively. In order to further
428 corroborate this relationship, another dataset (N = 34) from highly turbid waters (13.6 < TSS <
429 178.3 [g/m3]) of the Red River, Vietnam, are overlaid onto the scatterplot (Fig. 5b). Adding these
20
430 data left R2 unchanged, indicating robustness of the model to different hydro-geomorphological
431 conditions. Note that (740) was similarly derived from (740) for this dataset.
432 Analogously, we developed a relationship for TSS retrievals from (865) for OLI (Section
433 4.1.2; Fig. A.1)
434 TUU = G224.43 × (865)K Y 12.575 (8)
435 It should be further noted that due to the absence of a NIR channel in the 740-nm region among
436 OLI suite of measurements, the corresponding conditional rule for Type III waters (Table 2)
437 cannot be evaluated; hence, elevated uncertainties are expected.
438 In summary, to perform TSS retrievals, SOLID selects inversion models according to a
439 select water type. For Type I and II waters, where QAA and MDN are utilized, SOLID applies
440 Eq. 6 for which TSS is expected to be within the 0.1 - 65 [g/m3] range. In extremely turbid
441 waters (Type III), SOLID uses the NIR-based Inversion Approach (NIR-IA) followed by Eq. 7
442 or Eq. 8 for TSS estimations. This hybrid approach has been implemented in Python to produce
443 TSS products over a very broad range of TSS, i.e., 0.1 to 2626 [g/m3], in coastal and inland
444 waters (Section 5.3).
445 4.3 Performance metrics
446 In this study, we examine both linear and log-transformed metrics for evaluations of
447 estimated (E) quantities ( or TSS) against those measured (M) in situ. The evaluation of all
448 TSS algorithms is carried out using an in situ validation set (Z > 2700) independent of the
449 training set. The performance metrics are as follows
C⁄A
∑h
dia(_)`ab (cd )e_)`ab (fd ))
g
450 [U\ = ] j
k (9)
451 [lm\ = 100 × median [|Er Y Mr |⁄Mr ] for i = 1, …, N (10)
21
452 tu v = 10w where { = |∑jr•C}~•EC? (\r ) Y ~•EC? ([r )€⁄Z‚ (11)
453 [l\ = 10ƒ where „ = [∑jr•C|~•EC? (\r ) Y ~•EC? ([r )|⁄Z ] (12)
454 where RMSE is the root mean squared logarithmic error, MAPE is the median absolute
455 percentage error, Bias represents log-transformed residuals, and MAE stands for the mean
456 absolute error computed in log-space. The metrics computed in log-transformed space are
457 believed to provide a better assessment of the algorithms owing to the log-normal distribution of
458 TSS data (Fig. 2). The interpretation of Bias and MAE are as follows: Bias of 1.5 or 0.8 implies
459 that TSS estimations are 50% and 20% overestimated or underestimated, respectively. Similarly,
460 MAE of 1.2 suggests 20% overall error; however, MAE takes on values equal to or greater than
461 unity and is a very robust metric to gauge the overall performance of an algorithm. In addition to
462 the above metrics, we will also include slope associated with the linear regression fits to facilitate
463 comparisons with previous publications.
464 5. Results
465 The results are presented in four subsections. First, an assessment of inversion is
466 presented. This is followed by a full evaluation of the SOLID model and comparing it against
467 five existing algorithms. TSS maps generated from 13 MSI images are then qualitatively
468 examined. Lastly, the performance assessment is extended to OLCI, MODIS, VIIRS, and OLI to
469 demonstrate the utility of SOLID for multimission production of TSS products.
470 5.1 89: retrievals
471 To choose an appropriate (665) retrieval method, we compared the performances of our
472 inversion techniques with those of widely used state-of-the-art IOP algorithms, namely QAA
473 (Lee et al. 2002) and the GIOP (Werdell et al. 2013), with its default configuration (GIOP-DC).
474 For Type I waters, given a small validation sample size (N = 35), we found that QAA provides
22
475 better (665) retrievals than GIOP (Table 4). Our statistical attributes (Table 4), however,
476 confirm that both QAA and MDN are proper candidates with QAA showing better performance
477 according to MAPE and slope and MDN providing superior estimates given MAE and RMSE.
478 Consequently, there was no strong evidence to choose one over the other, likely owing to the
479 lack of representative validation data. We, therefore, decided to choose the better approach via
480 appraising TSS retrievals (Section 5.2). Our analysis suggested that QAA provides more accurate
481 retrievals of TSS (N = 436) and, as a result, better (665) in Type I waters (Table 4).
482 The performance of MDN for Type II waters was also benchmarked against those of
483 QAA and GIOP (Fig. 6 & Table 4) using in situ (665) ranging from 0.05 to 0.5 [1/m] (N =
484 211; mean of 0.069, median of 0.041, and standard deviation of 0.072). The GIOP and QAA
485 both approximate (665) with relatively large biases and RMSEs larger than that of MDN
486 (e.g., QAA returns overestimated quantities). Overall, the MDN produces better results than
487 these two state-of-the-art algorithms, albeit GIOP is a competitive algorithm in such eutrophic
488 waters (Figs. 3 and 5). The fairly low Bias and RMSE (i.e., 0.21) together with a slope close to
489 unity are evidence for its strong performance for a wide range of (665) in Type II waters.
490 Note that MDN produces fairly robust across the rest of the visible bands in Type II waters.
491 The performance assessment associated with the four visible MSI bands is summarized in Table
492 A.1 (Appendix A). Due to their intended design targeting robust retrievals in clear ocean waters
493 (Mitchell et al. 2016; Wang et al. 2009; Zheng et al. 2014), QAA and GIOP are expected to have
494 limited use in Type III waters (Section 1). Hence, we adopted our NIR-IA, developed using
495 published in situ data (Knaeps et al. 2018), as the optimal approach for estimating (740).
496
23
497 Table 4. The performance of GIOP, QAA, and MDN as compared to in situ (665). The NIR-IA
498 retrievals were not assessed due to absence of in situ (740). Global statistics are associated with all
499 Type I and Type II data. Best statistical attributes in each category are boldfaced.
500
Model MAPE [%] MAE RMSE Slope Bias* N
Global
GIOP 33.2 1.58 0.26 0.98 0.74
QAA 57.7 1.74 0.29 1.04 1.44 246
MDN 32.4 1.47 0.22 0.93 1.07
Type I
GIOP 39.37 1.97 0.38 1.03 0.52
QAA 21.58 1.57 0.29 1.02 0.85 35
MDN 31.63 1.49 0.24 0.75 0.85
Type II
GIOP 28.94 1.52 0.24 0.85 0.78
QAA 62.70 1.76 0.29 0.90 1.58 211
MDN 33.15 1.47 0.21 0.85 1.11
501 *
Bias = 1 is considered ideal (zero-bias) – see Eq. 11.
502
503
504
505
506
507 Fig. 6. Comparison of the modeled (y axis) and in situ measured (665) from GIOP, QAA, and MDN
508 for Type II waters (N=211); see Table 4 for accompanying statistics.
509
510
511 5.2 TSS validation
512 Here, the performance of the state-of-the-art TSS models (Miller and McKee 2004; Nechad
513 et al. 2010; Novoa et al. 2017; Ondrusek et al. 2012; Petus et al. 2010) is compared against that
514 of SOLID (see Table A.2 for expressions). Fig. 7 shows the performances of these models
515 through an independent in situ dataset (N = 2729; Table 1). The Miller model (Fig. 7a) shows a
24
516 fair performance in Type II waters, while it performs poorly in Type III waters and produces
517 negative TSS values in Type I waters. Using (665), the Nechad model (Fig. 7b) performs
518 well in Type II waters, contrasted with its poor performance in Type I and III waters. The Petus
519 model retrieves TSS values better than the Nechad model for Type I waters and exhibits
520 slightly poorer performance in Type II waters (Fig. 7c). From Fig. 7a, b, c, it is evident that
521 (665) tends to saturate for TSS > 50 [g/m3] (Doxaran et al. 2014; Feng et al. 2014; Han et
522 al. 2016; Shi and Wang 2009). The Ondrusek model (Fig. 7d) surmounts the saturation effect
523 (non-linearity) via a 3rd order polynomial fit, with (640) as the independent variable. Yet,
524 the TSS estimates by this model are largely overestimated for all the three types (Fig. 7d).
525 Note, however, that, in this study, (665) was supplied to this model. The Novoa model
526 clearly shows robust retrievals (Fig.7e) across all water types compared to the other existing
527 models. That said, according to the global statistics provided in Table 5, SOLID outperforms
528 all the state-of-the-art models by a large margin (Table 5; Fig. 7f). The improvements made
529 possible by SOLID compared to second and third best models, i.e., Novoa and Nechad, are 10-
530 20% in MAPE, 6-30% in MAE, 10-43% in RMSE, and 15% in Bias and slope. The one-to-one
531 agreement with significantly small errors across this wide dynamic range (0.10 < TSS< 2626.8
532 g/m3) suggests that SOLID has the potential for producing reasonably accurate/precise TSS
533 products at global scales in coastal and inland waters.
534
25
535
536 Fig. 7. Performance evaluation of multiple TSS retrieval methods shown alongside our proposed model
537 (SOLID) using in situ data (N =2729) for the models. The benchmark algorithms are (Miller and McKee
538 2004)(a), (Nechad et al. 2010)(b), (Petus et al. 2010)(c), (Ondrusek et al. 2012)(d), and (Novoa et al.
539 2017)(e).
540
541
542 The water-type-specific performance statistics are elaborated upon in Table 5. It can be inferred
543 that, in blue-green waters, SOLID significantly outperforms all the state-of-the-art algorithms,
544 suggesting its viability for moderately turbid coastal waters. For Type II waters, the performance
545 of SOLID closely resembles those of Nechad and Novoa. The main difference stems from near-
546 zero bias and near-unity slope that renders SOLID superior. Further, SOLID exhibits a notably
547 better performance than the rest of the models in Type III waters. The second-best model is the
548 Novoa model, which exhibits comparable MAPE to that of SOLID, but performs less well when
549 other metrics are examined. It should be noted that the performances of three other models (Han
550 et al. 2016; Siswanto et al. 2011; Zhang et al. 2010) were also examined. Of these, the models by
26
551 Siswanto et al. (2011) and Zhang et al. (2010) performed poorly for the full range of TSS. On
552 the other hand, the model of Han et al. (2016) performed fairly well in Type II waters but
553 exhibited saturations in Type III waters.
554
555 Table. 5. Statistical analysis of TSS retrievals from existing algorithms and our proposed approach
556 (SOLID) using a comprehensive in situ dataset for different types of water. Note that negative retrievals
557 by the Miller model were excluded from calculations. The statistical characteristics for SOLID are
558 boldfaced. The negative retrievals of the Miller model in Type I and II waters were removed from
559 statistical computations; thus, lower number are reported here.
560
TSS Model MAPE [%] MAE RMSE Slope Bias* N
Global
SOLID 48.94 1.81 0.32 0.97 1.09 2729
Miller 58.82 2.56 0.53 1.05 0.50 2152
Nechad 59.74 2.31 0.46 0.53 1.26
Petus 57.71 2.17 0.41 0.72 0.70
2729
Ondrusek 68.38 2.28 0.46 0.90 1.60
Novoa 52.73 1.92 0.35 0.84 1.27
Type I
SOLID 52.86 1.76 0.31 0.93 1.14 430
Miller 49.37 2.80 0.60 1.49 0.49 44
Nechad 544.92 5.23 0.77 0.32 5.12
Petus 110.33 2.20 0.37 0.52 1.73
430
Ondrusek 233.15 3.36 0.58 0.89 3.21
Novoa 95.66 2.29 0.42 0.98 1.96
Type II
SOLID 50.15 1.86 0.33 0.98 1.09 2122
Miller 57.10 2.43 0.51 1.24 0.54 1931
Nechad 50.48 1.87 0.33 0.64 1.08
Petus 50.99 2.05 0.38 0.84 0.64
2122
Ondrusek 55.95 2.08 0.42 0.95 1.38
Novoa 50.08 1.87 0.34 0.83 1.22
Type III
SOLID 31.41 1.48 0.22 0.86 1.02
Miller 73.94 4.43 0.73 0.40 0.23
Nechad 71.44 4.09 0.70 0.37 0.24
177
Petus 73.37 4.29 0.71 0.53 0.23
Ondrusek 73.74 2.64 0.54 1.37 1.71
Novoa 33.21 1.72 0.34 1.17 0.73
561 *
562
563
564
27
565 5.3 Applications to Sentinel-2 Imagery
566 For demonstration purposes, several satellite images with a wide range of optical regimes
567 were processed. The goal is to ensure that the model is capable of producing reasonable spatial
568 distribution maps according to expert expectations; therefore, the absolute retrievals are not
569 examined due to the absence of reliable in situ matchups. Multiple cloud-free images over the
570 Chesapeake Bay (MD, USA), San-Francisco-Bay Delta Estuary (CA, USA), Lake Okeechobee
571 (FL, USA), and Lake Taihu (China) were selected to assess spatial distributions of TSS. Fig. 8
572 demonstrates ACOLITE-processed MSI-derived TSS products (29th Aug 2018, 27th December
573 2018, 25th February 2019, and 6th April 2019) obtained from SOLID, the Ondrusek model
574 (Ondrusek et al. 2012), and the Nechad model (Nechad et al. 2010) in the upper Chesapeake Bay
575 area. This region is recognized for its high spatio-temporal variability in sediment loads due to
576 tidal forcing and freshwater inputs from different watersheds (Fugate et al. 2007; Kemp et al.
577 2005; Ondrusek et al. 2012). High sediment loads along with bio-optical variability in CDOM
578 and Chla pose challenges in TSS retrievals in this region (Aurin and Dierssen 2012; Werdell et
579 al. 2010). The TSS variability obtained from the SOLID model over different seasons was
580 closely matched with the regional maps produced by the Ondrusek model, specifically designed
581 for the Chesapeake Bay. Similar variabilities are not captured in the products created via the
582 Nechad model, which is expected to fail to retrieve high TSS values in the upper Chesapeake
583 Bay area (Fig. 7b). Although there are discrepancies in the absolute values, the relative spatial
584 distributions of the two products, in particular, in the northern sections of the bay resemble that
585 of the SOLID model, which is expected to produce more accurate concentrations (Fig. 7 and
586 Table 4), i.e., the Ondrusek model tends to overestimate TSS and Nechad model is saturated
587 (Fig. 7b & d). Additionally, it is worth noting that the SOLID model produces smooth transitions
28
588 between the three water types (regions) and no artifacts are introduced as a result of the water-
589 type classification. Fig. A.2 illustrates water-type maps identified from MSI-derived products
590 (Table 2).
591 To analyze the spatial and temporal TSS variations at finer scales, SOLID was also applied to
592 three SeaDAS-processed MSI-A images over the San Pablo Bay (Northern San-Francisco-Bay-
593 Delta Estuary, CA) (Fig. 9). The patterns of TSS over the San Pablo Bay are often affected by
594 tides, river discharges, wind, and dredging activities, thereby providing a suitable testbed (Hestir
595 et al. 2013) to evaluate our new retrieval method. Fig. 9a shows an image captured on August
596 18th 2017 during high tide (GMT 19:03:26). At this time, suspended sediments were all mixed in
597 offshore areas and deposited near the coastline. In contrast, large gradients in sediment
598 concentrations caused by high discharge from the Petaluma River and the Sonoma creek are
599 observed during spring (Fig. 9b March 26th 2018). Finally, another MSI-A image collected on
600 July 24th 2018 two hours before high tide (GMT 19:04:56) showed lower concentrations within
601 the Bay (Fig. 9c), in line with what is expected during flood tide. Prior to high tides, seawater
602 approaches the coast and high sediments are observed to the north. This analysis indicates that
603 the SOLID model provides realistic TSS distribution and is capable of capturing the intra-annual
604 variability in sediment load.
29
605
606 Fig. 8. ACOLITE-processed MSI-A images show the spatial distribution of TSS over the Chesapeake
607 Bay in 2018 (29th August and 27th December) and 2019 (25th February and 6th April). (Top row) TSS
608 concentration retrieved from the SOLID model, (Middle row) TSS concentration retrieved from the
609 regional model (Ondrusek et al. 2012), (Bottom row) TSS concentration retrieved from the regional
610 model (Nechad et al. 2010). See Fig. A.2 for maps of water-types.
611
612
30
613
614 Fig. 9. SeaDAS processed Sentinel-2A MSI images show the spatio-temporal patterns of TSS (derived
615 from SOLID) over the San Pablo Bay (CA). a) Products during summer at high tide on August 18th 2017
616 (GMT 19:03:26), (b) Image captured during winter season at ebb tide on March 26th 2018 (GMT
617 19:04:52), and (c) Image obtained during flood tide (two hours before reaching the high tide zone) on July
618 24th 2018 (GMT 19:04:56).
619
620
621 Fig. 10. Spatio-temporal distribution of SOLID retrieved TSS concentrations over highly productive lakes
622 such as: (top row) Lake Okeechobee (USA) and (bottom row) Lake Taihu (China) using SeaDAS
623 processed Sentinel-2A/2B MSI images. (Top row) Transitional variations of TSS are captured during the
624 spring season of 2019 (29th January (a), 22nd March(b) and 21st April (c)). (Bottom row) Spatial
625 dynamics of TSS over Lake Taihu were captured in three different images during 2018 (11th October (d)
626 and 13th February (e), and 23rd February (f)). The images were processed using SeaDAS through a SWIR
627 based atmospheric correction procedure (Pahlevan et al., 2017d).
628
31
629 TSS retrievals in eutrophic lakes like Lake Okeechobee and Lake Taihu are further
630 qualitatively assessed. The images were atmospherically corrected using SeaDAS. Due to the
631 presence of high sediment loads, aerosol contributions were compensated for using two SWIR
632 bands (Pahlevan et al. 2017b). The disadvantage of this method is noisy retrievals in areas with
633 lower TSS. Fig. 10 (top row) shows our TSS products derived from MSI-A images over
634 productive waters of Lake Okeechobee. With a mean depth of 3 m together with high-wind
635 events particle resuspension events are quite common (Jin and Ji 2004). Higher concentrations in
636 the late January map are attributed to prior rain events and commonly high winds in wintertime.
637 The concentrations are found to be lower in March and April. The spatial distribution of TSS
638 across Lake Taihu is found to be very consistent with previous regional analyses (Zhang et al.,
639 2014), with higher loads in the central and southern sections of the basin. Evidently, the SOLID
640 model is able to produce spatially congruent TSS maps from MSI images, enabling sound spatio-
641 temporal analyses.
642 5.4 Extension to other satellite missions
643 The performance of SOLID for other satellite missions, such as MODIS, VIIRS, OLCI, and
644 OLI is shown in scatterplots. The corresponding global and water-type specific statistics are
645 provided in Table 6. In general, the statistical metrics are fairly consistent with those reported for
646 MSI (Table 5), indicating potentially interconsistent TSS products derived from various satellite
647 missions made possible through SOLID. The primary difference is attributed to the Bias, which
648 was lower for MSI than the rest of the missions. The largest errors are attributed to OLI-derived
649 TSS products, owing to the absence of spectral information within the 700-800 nm region for
650 approximating spectral (Section 4.1.1). In essence, the overall performance of the MDN
32
651 model in estimating at each individual band is better when more relevant spectral features
652 are
653
654 Fig. 11. Scatterplots illustrating the SOLID performance for four different satellite missions. The spectral
655 bands incorporated for retrievals are listed in Table 3.
656
657 Table. 6. Statistical analysis of TSS retrievals of SOLID using a comprehensive in situ dataset for
658 different satellite missions.
659
MAPE [%] MAE RMSE Slope Bias* N
MODIS 47.33 1.82 0.33 0.96 1.18
VIIRS 49.71 1.84 0.34 0.96 1.24 2729
OLCI 46.58 1.82 0.33 0.96 1.17
OLI 52.30 1.87 0.35 0.97 1.34 2633
Type I
MODIS 48.69 1.76 0.30 0.95 1.07
VIIRS 52.33 1.77 0.31 0.94 1.08
430
OLCI 51.23 1.77 0.31 0.93 1.08
OLI 51.20 1.76 0.31 0.94 1.12
Type II
MODIS 48.46 1.86 0.34 0.88 1.23
VIIRS 51.58 1.90 0.35 0.90 1.30
2122
OLCI 47.17 1.86 0.35 0.89 1.21
OLI 54.14 1.91 0.36 0.88 1.42
Type III
MODIS 30.56 1.48 0.22 0.86 0.99 177
VIIRS 31.25 1.47 0.22 0.87 1.00
33
OLCI 28.05 1.47 0.22 0.87 0.97
OLI 25.05 1.39 0.19 0.98 0.85 81
660 *
661
662
663 supplied (Pahlevan 2020), i.e., more accurate is possible via MSI or MODIS than that
664 through OLI (analysis not shown here). A secondary factor that may contribute to the reduced
665 performance of the OLI model is that OLI’s red channel, in contrast to other missions, does not
666 fully capture Chla absorption peak at ~ 670 nm (Table 3). Further, note that spectra that
667 lacked spectral coverage up to 900 nm were excluded for the performance assessment of OLI-
668 like spectra in Type III waters, reducing the number of spectra by nearly one half.
669
670 6. Discussion
671 Following a full assessment of SOLID, it is critical to gauge its sensitivity to the choice of
672 ∗
"# in the (740) retrieval scheme and to the uncertainties in the atmospheric correction.
673 While the former can partly explain algorithm uncertainties in Type III waters, the latter will
674 shed light on uncertainties in TSS products derived from satellite observations. In this section,
675 we will further address the steps required for implementing SOLID as a standard, global TSS
676 algorithm in coastal and inland waters.
677 6.1 Non-algal particle absorption (…†‡: )
678 Using empirical relationships or empiricism in bio-optical modeling or water constituent
679 retrievals is common-practice in ocean color or aquatic remote sensing. For example, best-
680 practice algorithms (Lee et al. 2002; Maritorena et al. 2002; Nechad et al. 2010; Werdell et al.
681 2013) incorporate various assumptions on the spectral shape and variability of component
682 mass-specific IOPs leading to loss of performance in aquatic environments where empirical
34
683 relationships may no longer hold (Werdell et al. 2018). Making such assumptions is inevitable
684 due to limited knowledge of the optical properties of myriads of particles in the water column
685 for the design of global algorithms. Although it is recognized that ∗

"# (740) varies with type,
686 composition, and size distribution of non-algal particles (Babin and Stramski 2004; Doxaran et
687 al. 2009b), for simplicity, we chose a constant value for ∗

"# (740) = 0.011 [m2/g] (Section
688 4.1.2). However, by definition "# (740) (= ∗

"# (740) × TSS) also changes with
689 concentration of particles. To assess the sensitivity of retrieved TSS to variability in "# (740)
690 in Type III waters, we supplied a broad range of "# (740), i.e., from zero to 10 [1/m] @
691 increments of 0.5 [1/m], to the SOLID model. We then analyzed the errors in TSS retrievals
692 expressed as MAE (Eq. 12) using Type III – TSS pairs (N =177, Table 5). MAE (a unitless,
693 log-based error metric) is a preferred measure because of its straightforward interpretation.
694 From Fig. 12, it can be inferred that MAE reaches a minimum for 1.2 < "# (740) <
695 1. 6 [1/m], i.e., the total error in TSS varies from 20 to 60%. Further, MAE tends to increase
696 monotonically for "# (740) > 2 [1/m]. Note that very similar behavior was found when
697 assessing RMSE and MAPE. This assessment suggests that our specific choice of "# (740)
698 results in small errors in TSS estimations given our validation dataset. A dataset representing
699 extremely higher TSS (>> 150 [g/m3]) will likely carry larger uncertainties in TSS retrievals
700 when using SOLID. Hence, care must be taken when interpreting the model outcomes in highly
701 sediment-rich waters.
35
702
703 Fig.12. The estimated error (MAE) in TSS retrievals for Type III waters as a function of non-algal
704 particle absorption at 740 nm.
705
706 6.2 Atmospheric correction
707 Uncertainties in the atmospheric correction are a major source of errors in and the
708 downstream products. These uncertainties emanate from highly turbid/eutrophic conditions,
709 complex aerosols, land/cloud adjacency effects, cloud shadows, and thin clouds (Ngoc et al.
710 2019; Pahlevan et al. 2017a; Sterckx et al. 2011). Here, the sensitivity to uncertainties in the
711 atmospheric correction is analyzed using a Monte Carlo simulation (n=1000). This simulation
712 was carried out by adding random Gaussian noise to the most relevant channels and
713 assessing the impacts on downstream TSS products. The MSI’s (665) and (740)
714 corresponding to the three water types (Fig. 13; top row) are perturbed by δ =10%, which is
715 assumed to be a realistic expected uncertainty in these bands over coastal and inland waters
716 (Pahlevan et al. 2017b). For completeness and accounting for larger uncertainties, we also
717 introduce δ =50% to the relevant bands to address the model tolerance to such degrees of
718 uncertainties. The simulated perturbed are then supplied to SOLID for TSS retrievals (Fig.
719 13; bottom row). Note that, here, we assume that the shape of is valid and the water types are
720 correctly assigned. Fig. 13 illustrates the output histograms with the statistics included in Table
721 6. Our analysis suggests that the SOLID model produces TSS products with < 21% uncertainties
36
722 when uncertainties in the corresponding channels are assumed 10%. With δ =50%, the
723 uncertainties become too large (Table 7) rendering the derived products unreliable for most
724 purposes. Thus, it is critical to ensure high-quality products retrieved through the
725 atmospheric correction process. Further analyses may be required to assess the performance of
726 SOLID considering biases in the absolute radiometric measurements made by satellite sensors.
727
728 Fig.13. Random noise distributions (δ =10%) used to perturb MSI-derived (top panels) in the red and
729 NIR regions for the three water types. The output frequency distribution of TSS (lower panels) retrieved
730 from SOLID given the distributions.
731
732
733
734
735
736
737
738 Table. 7. Statistics associated with the sensitivity analysis of SOLID to uncertainties in MSI-derived
739 induced by the atmospheric correction. Here, µ and σ are the mean and the standard deviation of the
740 Gaussian distributions. σRD(%) is the standard deviation of the relative difference.
741
37
δ [%] [1/sr] Output (TSS) distribution
TSS [g/m3]
µ [g/m3] σ [g/m3] σRD (%)
µ = 0.00025
10 0.272 0.027 10.1
σ = 0.000025
Type I 665 nm 0.275
µ = 0.00025
50 0.263 0.134 51.0
σ = 0.00013
µ = 0.0165
10 15.77 3.16 21.1
σ = 0.0016
Type II 665 nm 15.04
µ = 0.0165
50 15.93 19.27 125.1
σ = 0.0082
µ = 0.0312
10 265.45 54.73 19.5
σ = 0.0032
Type III 740 nm 260.49
µ = 0.0312
50 266.43 1039.9 399.0
σ = 0.0156
742
743
744 6.3 Implications for long-term global monitoring of TSS
745 Owing to its comparably favorable performance across a wide range of particle loads, the
746 SOLID model has the potential for global retrievals of TSS from a suite of satellite missions
747 that make measurements in the red and NIR spectral regions. Our statistical analysis, however,
748 showed that although the SOLID model outperforms the state-of-the-art models, the overall
749 error, i.e., ~ 80% gauged via MAE, is yet to be reduced to permit rigorous scientific
750 investigations, where precise estimations of near-surface particle loads are sought. Such
751 magnitudes of error may be ascribed to the complex populations of particles represented by this
752 commonly assessed variable, i.e. ranges of particle sizes and shape and complex optical
753 interactions of the particles. For example, the pigment absorbances of different algal taxa and
754 the formation of flocculant clay-detrital aggregates in the oligohaline reaches of estuaries; these
755 aggregates, in turn, can support a microbial biofilm community of algae and bacteria – imagine
756 the challenges of modeling this phenomenon).
757 One way to improve SOLID is to use in situ data for training MDN to enhance retrievals of
758 . The lack of adequate training data was the main reason for training MDN with simulated
38
759 data, which may not offer the true representation of bio-optical conditions given limited
760 knowledge of specific IOPs. Alternatively, possible improvements in QAA and GIOP for
761 retrievals in Type I and Type II waters, respectively, may eliminate the necessity for a machine
762 learning model. In addition, uncertainties in the atmospheric correction further precludes
763 precise retrievals of TSS. The performance of SOLID and likely other methods is expected to
764 degrade for large uncertainties in products.
765 To enable production of interconsistent, multimission TSS retrievals at a global scale, one
766 needs to ensure minimal biases exist among top-of-atmosphere observations. Pahlevan et al.
767 (2019) showed that ~ 4% difference in MSI- and OLI-derived (665) translate to ~ 10%
768 difference in TSS products derived via the Nechad model. Prior to generating global TSS
769 products, a comprehensive assessment of image-derived TSS products derived from multiple
770 missions (e.g., MODIS, VIIRS, OLCI, MSI, and OLI) against existing in situ databases (e.g.,
771 the SeaWiFS Bio-optical Archive and Storage System (SeaBASS), United States Geological
772 Survey’s National Water Information System (NWIS)) is required. Given sufficient agreements
773 with in situ data, per-retrieval uncertainty should be computed and produced alongside TSS
774 values. This can be achieved through Monte Carlo simulations and the training of a model to
775 provide uncertainty estimates in a computationally inexpensive fashion.
776 7. Conclusion
777 The primary goal of this study was to introduce a hybrid scheme for TSS retrievals that would
778 advance the state-of-the-art for TSS retrievals in both inland and coastal waters. This strategy
779 referred to as SOLID applies water-type-specific algorithms to provide an estimation of TSS for
780 a given ( ) by retrieving in the red/NIR region as intermediate products. The water types
781 determined given ./ (400 nm < < 700 nm) constitute blue-green waters (Type I), green
39
782 waters (Type II), and sediment-laden brown waters (Type III). For inversion in Type II
783 waters, we apply a machine learning model that enhances estimates compared to widely used
784 semi-analytical methods (e.g., QAA and GIOP). Through an extensive validation exercise, we
785 show that SOLID outperforms the state-of-the-art algorithms across a wide range of TSS, i.e.,
786 0.02 ~ 2626.8 [g/m3], suggesting its potential utility for global mapping of TSS. The global
787 statistics, including MAPE (49 %), RMSE (0.32), MAE (1.81), and Bias (1.09), corroborate that
788 the SOLID model improves retrieval performances offered by five widely used TSS retrieval
789 methods. In particular, the performance of SOLID is superior to that of Nechad et al. (2010) and
790 Novoa et al. (2017) in both Type I and Type III waters by a noticeable margin. We show that the
791 model is anticipated to perform well for various missions, such as MSI, OLCI, MODIS, VIIRS,
792 and OLI. The primary confounding factor for mass production of TSS products is believed to be
793 the uncertainties in the atmospheric correction, i.e., we show that 10% uncertainties in
794 (600 nm < < 800 nm) result in ~ 20% uncertainties in TSS. Considering its fairly
795 consistent performance across multiple missions, this model can be utilized operationally for
796 generating products subject to an extensive satellite matchup assessment in coastal and inland
797 waters. Although the SOLID model was found in our analysis to perform better than the existing
798 algorithms, there still remains an ~ 80% error in global TSS retrievals. Contingent upon the
799 availability of high-fidelity in situ data, machine-learning approaches shall be explored to further
800 reduce such uncertainties.
801
802
803
804
40
805
806
807
808
809
810
811
812
813
814 Appendix A
815
816 Fig. A.1. Scatterplot corresponding to in situ TSS and bbp(865) datasets developed for OLI observations in
817 Type III waters. Note that the fitted lines are displayed in log-log scale.
818
819
820
821
41
2018-08-29 2018-12-27 2019-02-25 2019-04-06
Land Type-I Type-II Type-III

822
823 Fig. A.2. Chesapeake Bay water-type maps corresponding to MSI images shown in Fig. 8. As expected,
824 the upper bay area is commonly assigned to Type III and the mid and lower bay regions are classified as
825 Type II waters. Type I waters are not found in the bay for these four dates.
826
827
828
829
830
831
832
833
834
835
42
836 The performance of MDN against that of GIOP and QAA is further elaborated. Evidently,
837 MDN-based retrievals are consistently better than those determined by both GIOP and QAA;
838 hence, MDN may be regarded as an alternative to these heritage algorithms in turbid and/or
839 eutrophic ecosystems in coastal and inland waters with (665) ranging from 0.05 to 0.5 [1/m]
840 (see Section 5.1).
841 Table A1. The performance of GIOP, QAA, and MDN for retrievals across all MSI visible bands
842 for Type II waters. Best statistical descriptors in each category are boldfaced.
843
MAPE [%] MAE RMSE Slope Bias* N
443 nm
GIOP 42.55 1.81 0.33 0.76 0.67
211
QAA 50.72 1.74 0.30 0.80 1.36
MDN 36.51 1.57 0.26 0.76 1.08
490 nm
GIOP 41.18 1.74 0.30 0.78 0.70
211
QAA 50.01 1.71 0.29 0.83 1.40
MDN 36.34 1.53 0.24 0.80 1.10
560 nm
GIOP 36.01 1.61 0.26 0.82 0.74
211
QAA 58.74 1.72 0.28 0.86 1.50
MDN 37.37 1.53 0.23 0.81 1.20
665 nm
GIOP 28.94 1.52 0.24 0.85 0.78
211
QAA 62.70 1.76 0.29 0.90 1.58
MDN 33.15 1.47 0.21 0.84 1.11
844 *
845
846 Table A2. Expressions used in this study for the assessment of state-of-the-art TSS models
Model Relation Source
(1140.25 Rrs(668)) -1.91 (Miller and McKee 2004)
Miller-McKee
Nechad 1.74+(355.85 w(665))/(1-( w(665)/1728)) (Nechad et al. 2010)
Petus {12450Rrs(668)2} + 666.1{Rrs(668)} + 0.4 (Petus et al. 2010)
Ondrusek 3.8813nLw(665)3-13.822nLw(665)2+ 19.61nLw(665) (Ondrusek et al. 2012)

531.5 x w(665) (Type-I&II) (Novoa et al. 2017)
Novoa
{37150 x w(865)2} +. {1751 x w(865)}(Type-III)
847
848
43
849 Acknowledgement
850 We gratefully acknowledge the PIs, scientists and contributors to SeaBASS, including
851 Zhongping Lee, Steve Ackleson, Richard Reynolds (ICESCAPE), Colleen Mouw, and are also
852 very thankful to the NASA Ocean Biology Processing Group (OBPG), David Doxaran, and
853 Kevin Ruddick for distributing in situ datasets. Nima Pahlevan is funded under NASA ROSES
854 Awards #80NSSC18K0077, #80HQTR19C0015, and the USGS Landsat Science Team Award
855 #140G0118C0011. Emmanuel Boss was funded under ONR grant N000141612218. John
856 Schalles was funded under NOAA grant #NA11SEC4810001 and NSF grant # 1832178. Further,
857 we are grateful to the two anonymous reviewers for their fruitful comments and meticulous
858 reviews, which ultimately led to the enhanced quality of this manuscript.
859
860
861 References
862 Agarap, A.F. (2018). Deep Learning using Rectified Linear Units (ReLU). arXiv preprint
863 arXiv:1803.08375
864 Ahn, Y.H., Shanmugam, P., & Moon, J.E. (2006). Retrieval of ocean colour from high
865 resolution multi-spectral imagery for monitoring highly dynamic ocean features. International
866 Journal of Remote Sensing, 27, 367-392
867 Aurin, D.A., & Dierssen, H.M. (2012). Advantages and limitations of ocean color remote
868 sensing in CDOM-dominated, mineral-rich coastal and estuarine waters. Remote Sensing of
869 Environment, 125, 181-197
870 Babin, M., & Stramski, D. (2004). Variations in the mass‐specific absorption coefficient
871 of mineral particles suspended in water. Limnology and Oceanography, 49, 756-767
872 Babin, M., Stramski, D., Ferrari, G.M., Claustre, H., Bricaud, A., Obolensky, G., &
873 Hoepffner, N. (2003). Variations in the light absorption coefficients of phytoplankton, nonalgal
874 particles, and dissolved organic matter in coastal waters around Europe. Journal of Geophysical
875 Research: Oceans, 108
44
876 Barbier, E.B., Hacker, S.D., Kennedy, C., Koch, E.W., Stier, A.C., & Silliman, B.R.
877 (2011). The value of estuarine and coastal ecosystem services. Ecological monographs, 81, 169-
878 193
879 Binding, C., Jerome, J., Bukata, R., & Booty, W. (2010). Suspended particulate matter in
880 Lake Erie derived from MODIS aquatic colour imagery. International Journal of Remote
881 Sensing, 31, 5239-5255
882 Binding, C., Zastepa, A., & Zeng, C. (2019). The impact of phytoplankton community
883 composition on optical properties and satellite observations of the 2017 western Lake Erie algal
884 bloom. Journal of Great Lakes Research, 45, 573-586
885 Bishop, C.M. (1994). Mixture Density Networks. NCRG/94/004. Aston University
886 Birmingham http://www.ncrg.aston.ac.uk.
887 Bowers, D., & Binding, C. (2006). The optical properties of mineral suspended particles:
888 A review and synthesis. Estuarine, Coastal and Shelf Science, 67, 219-230
889 Bukata, R.P., Jerome, J.H., Kondratyev, A.S., & Pozdnyakov, D.V. (2018). Optical
890 properties and remote sensing of inland and coastal waters. CRC press
891 Casey, K.A.R., Cecile S; Gregg, Watson W; Boss, Emmanuel; Chase, Alison P; Craig,
892 Susanne E; Mouw, Colleen B; Reynolds, Rick A; Stramski, Dariusz; Ackleson, Steven G;
893 Bricaud, Annick; Schaeffer, Blake; Lewis, Marlon R; Maritorena, Stéphan (2020). A global
894 compilation of in situ aquatic high spectral resolution inherent and apparent optical property data
895 for remote sensing applications. Earth System Science Data
896 Chen, S., Han, L., Chen, X., Li, D., Sun, L., & Li, Y. (2015). Estimating wide range Total
897 Suspended Solids concentrations from MODIS 250-m imageries: An improved method. ISPRS
898 Journal of Photogrammetry and Remote Sensing, 99, 58-69
899 D'Alimonte, D., Zibordi, G., Berthon, J.-F., Canuti, E., & Kajiyama, T. (2012).
900 Performance and applicability of bio-optical algorithms in different European seas. Remote
901 Sensing of Environment, 124, 402-412
902 Dekker, A.G., Vos, R., & Peters, S. (2002). Analytical algorithms for lake water TSM
903 estimation for retrospective analyses of TM and SPOT sensor data. International Journal of
904 Remote Sensing, 23, 15-35
905 Dekker, A.G., Vos, R.J., & Peters, S.W.M. (2001). Comparison of remote sensing data,
906 model results and in situ data for total suspended matter (TSM) in the southern Frisian lakes.
907 Science of the Total Environment, 268, 197-214
908 Dogliotti, A.I., Ruddick, K., Nechad, B., Doxaran, D., & Knaeps, E. (2015). A single
909 algorithm to retrieve turbidity from remotely-sensed data in all coastal and estuarine waters.
910 Remote Sensing of Environment, 156, 157-168
45
911 Doxaran, D., Ehn, J., Bélanger, S., Matsuoka, A., Hooker, S., & Babin, M. (2012).
912 Optical characterisation of suspended particles in the Mackenzie River plume (Canadian Arctic
913 Ocean) and implications for ocean colour remote sensing. Biogeosciences, 9, 3213-3229
914 Doxaran, D., Froidefond, J.-M., & Castaing, P. (2003). Remote-sensing reflectance of
915 turbid sediment-dominated waters. Reduction of sediment type variations and changing
916 illumination conditions effects by use of reflectance ratios. Applied Optics, 42, 2623-2634
917 Doxaran, D., Froidefond, J.-M., Castaing, P., & Babin, M. (2009a). Dynamics of the
918 turbidity maximum zone in a macrotidal estuary (the Gironde, France): Observations from field
919 and MODIS satellite data. Estuarine, Coastal and Shelf Science, 81, 321-332
920 Doxaran, D., Lamquin, N., Park, Y.-J., Mazeran, C., Ryu, J.-H., Wang, M., & Poteau, A.
921 (2014). Retrieval of the seawater reflectance for suspended solids monitoring in the East China
922 Sea using MODIS, MERIS and GOCI satellite data. Remote Sensing of Environment, 146, 36-48
923 Doxaran, D., Leymarie, E., Nechad, B., Dogliotti, A., Ruddick, K., Gernez, P., & Knaeps,
924 E. (2016). Improved correction methods for field measurements of particulate light
925 backscattering in turbid waters. Optics express, 24, 3615-3637
926 Doxaran, D., Ruddick, K., McKee, D., Gentili, B., Tailliez, D., Chami, M., & Babin, M.
927 (2009b). Spectral variations of light scattering by marine particles in coastal waters, from the
928 visible to the near infrared. Limnology and Oceanography, 54, 1257-1271
929 Estapa, M.L., Boss, E., Mayer, L.M., & Roesler, C.S. (2012). Role of iron and organic
930 carbon in mass‐specific light absorption by particulate matter from Louisiana coastal waters.
931 Limnology and Oceanography, 57, 97-112
932 Feng, L., Hu, C., Chen, X., & Song, Q. (2014). Influence of the Three Gorges Dam on
933 total suspended matters in the Yangtze Estuary and its adjacent coastal waters: Observations
934 from MODIS. Remote Sensing of Environment, 140, 779-788
935 Forget, P., & Ouillon, S. (1998). Surface suspended matter off the Rhone river mouth
936 from visible satellite imagery. Oceanologica acta, 21, 739-749
937 Fugate, D.C., Friedrichs, C.T., & Sanford, L.P. (2007). Lateral dynamics and associated
938 transport of sediment in the upper reaches of a partially mixed estuary, Chesapeake Bay, USA.
939 Continental Shelf Research, 27, 679-698
940 Gernez, P., Barillé, L., Lerouxel, A., Mazeran, C., Lucas, A., & Doxaran, D. (2014).
941 Remote sensing of suspended particulate matter in turbid oyster‐farming ecosystems. Journal of
942 Geophysical Research: Oceans, 119, 7277-7294
943 Gitelson, A.A., Dall'Olmo, G., Moses, W., Rundquist, D.C., Barrow, T., Fisher, T.R.,
944 Gurlin, D., & Holz, J. (2008). A simple semi-analytical model for remote estimation of
945 chlorophyll-a in turbid waters: Validation. Remote Sensing of Environment, 112, 3582-3593
46
946 Gordon, H.R., Brown, O.B., Evans, R.H., Brown, J.W., Smith, R.C., Baker, K.S., &
947 Clark, D.K. (1988). A semianalytic radiance model of ocean color. Journal of Geophysical
948 Research: Atmospheres, 93, 10909-10924
949 Gordon, H.R., & Wang, M. (1994). Retrieval of water-leaving radiance and aerosol
950 optical thickness over the oceans with SeaWiFS: a preliminary algorithm. Appl. Opt., 33, 443-
951 452
952 Han, B., Loisel, H., Vantrepotte, V., Mériaux, X., Bryère, P., Ouillon, S., Dessailly, D.,
953 Xing, Q., & Zhu, J. (2016). Development of a semi-analytical algorithm for the retrieval of
954 suspended particulate matter from remote sensing over clear to very turbid waters. Remote
955 Sensing, 8, 211
956 Hannouche, A., Joannis, C., & Chebbo, G. (2017). Assessment of total suspended solids
957 (TSS) event load and its uncertainties in combined sewer system from continuous turbidity
958 measurements. Urban Water Journal, 14, 789-796
959 Hestir, E.L., Schoellhamer, D.H., Morgan-King, T., & Ustin, S.L. (2013). A step
960 decrease in sediment concentration in a highly modified tidal river delta following the 1983 El
961 Niño floods. Marine Geology, 345, 304-313
962 Ioannou, I., Gilerson, A., Gross, B., Moshary, F., & Ahmed, S. (2011). Neural network
963 approach to retrieve the inherent optical properties of the ocean from observations of MODIS.
964 Applied Optics, 50, 3168-3186
965 Jin, K.-R., & Ji, Z.-G. (2004). Case study: modeling of sediment transport and wind-wave
966 impact in Lake Okeechobee. Journal of Hydraulic Engineering, 130, 1055-1067
967 Joshi, I.D., & D'Sa, E.J. (2018). An estuarine-tuned quasi-analytical algorithm (QAA-V):
968 assessment and application to satellite estimates of SPM in Galveston Bay following Hurricane
969 Harvey. Biogeosciences, 15, 4065-4086
970 Kemp, W.M., Boynton, W.R., Adolf, J.E., Boesch, D.F., Boicourt, W.C., Brush, G.,
971 Cornwell, J.C., Fisher, T.R., Glibert, P.M., & Hagy, J.D. (2005). Eutrophication of Chesapeake
972 Bay: historical trends and ecological interactions. Marine Ecology Progress Series, 303, 1-29
973 Knaeps, E., Doxaran, D., Dogliotti, A., Nechad, B., Ruddick, K., Raymaekers, D., &
974 Sterckx, S. (2018). The SeaSWIR dataset. Earth System Science Data, 10, 1439-1449
975 Knaeps, E., Ruddick, K., Doxaran, D., Dogliotti, A.I., Nechad, B., Raymaekers, D., &
976 Sterckx, S. (2015). A SWIR based algorithm to retrieve total suspended matter in extremely
977 turbid waters. Remote Sensing of Environment, 168, 66-79
978 Kumar, A., Equeenuddin, S.M., Mishra, D.R., & Acharya, B.C. (2016). Remote
979 monitoring of sediment dynamics in a coastal lagoon: Long-term spatio-temporal variability of
980 suspended sediment in Chilika. Estuarine, Coastal and Shelf Science, 170, 155-172
47
981 Lee, Z., Carder, K.L., & Arnone, R.A. (2002). Deriving inherent optical properties from
982 water color: a multiband quasi-analytical algorithm for optically deep waters. Applied Optics, 41,
983 5755-5772
984 Lee, Z., Pahlevan, N., Ahn, Y.-H., Greb, S., & O'Donnell, D. (2013). Robust approach to
985 directly measuring water-leaving radiance in the field. Applied Optics, 52, 1693-1701
986 Loisel, H., Mangin, A., Vantrepotte, V., Dessailly, D., Dinh, D.N., Garnesson, P.,
987 Ouillon, S., Lefebvre, J.-P., Mériaux, X., & Phan, T.M. (2014). Variability of suspended
988 particulate matter concentration in coastal waters under the Mekong's influence from ocean color
989 (MERIS) remote sensing over the last decade. Remote Sensing of Environment, 150, 218-230
990 Loisel, H., & Morel, A. (2001). Non-isotropy of the upward radiance field in typical
991 coastal (Case 2) waters. International Journal of Remote Sensing, 22, 275-295
992 Long, C.M., & Pavelsky, T.M. (2013). Remote sensing of suspended sediment
993 concentration and hydrologic connectivity in a complex wetland environment. Remote Sensing of
994 Environment, 129, 197-209
995 Luo, Y., Doxaran, D., Ruddick, K., Shen, F., Gentili, B., Yan, L., & Huang, H. (2018).
996 Saturation of water reflectance in extremely turbid media based on field measurements, satellite
997 data and bio-optical modelling. Optics express, 26, 10435-10451
998 Maritorena, S., Siegel, D.A., & Peterson, A.R. (2002). Optimization of a semianalytical
999 ocean color model for global-scale applications. Appl. Opt., 41, 2705-2714
1000 Miller, R.L., & McKee, B.A. (2004). Using MODIS Terra 250 m imagery to map
1001 concentrations of total suspended matter in coastal waters. Remote Sensing of Environment, 93,
1002 259-266
1003 Mitchell, C., Cunningham, A., & McKee, D. (2016). Derivation of the specific optical
1004 properties of suspended mineral particles and their contribution to the attenuation of solar
1005 irradiance in offshore waters by ocean color remote sensing. Journal of Geophysical Research:
1006 Oceans, 121, 104-117
1007 Mobley, C.D. (1999). Estimation of the remote-sensing reflectance from above-surface
1008 measurements. Appl Opt, 38, 7442-7455
1009 Mobley, C.D., Sundman, L.K. (2008). Hydrolight 5, Ecolight5 User Guide. Sequoia
1010 Scientific, Inc.
1011 Moore, T.S., Mouw, C.B., Sullivan, J.M., Twardowski, M.S., Burtner, A.M., Ciochetto,
1012 A.B., McFarland, M.N., Nayak, A.R., Paladino, D., Stockley, N.D., Johengen, T.H., Yu, A.W.,
1013 Ruberg, S., & Weidemann, A. (2017). Bio-optical Properties of Cyanobacteria Blooms in
1014 Western Lake Erie. Frontiers in Marine Science, 4
1015 Morel, A., & Prieur, L. (1977). Analysis of variations in ocean color 1. Limnology and
1016 Oceanography, 22, 709-722
48
1017 Morel, A., Voss, K.J., & Gentili, B. (1995). Bidirectional reflectance of oceanic waters: a
1018 comparison of modeled and measured upward radiance fields. Journal of Geophysical Research:
1019 Oceans, 100, 13143-13150
1020 Mouw, C.B., Chen, H., McKinley, G.A., Effler, S., O'Donnell, D., Perkins, M.G., &
1021 Strait, C. (2013). Evaluation and optimization of bio‐optical inversion algorithms for remote
1022 sensing of Lake Superior's optical properties. Journal of Geophysical Research: Oceans, 118,
1023 1696-1714
1024 Mueller, J.L., Fargion, G.S., McClain, C.R., Mueller, J., Brown, S., Clark, D., Johnson,
1025 B., Yoon, H., Lykke, K., & Flora, S. (2004). Ocean Optics Protocols For Satellite Ocean Color
1026 Sensor Validation, Revision 5, Volume VI: Special Topics in Ocean Optics Protocols, Part 2.
1027 NASA Report 211621, 2003, 1-36.
1028 Nechad, B., Ruddick, K., & Park, Y. (2010). Calibration and validation of a generic
1029 multisensor algorithm for mapping of total suspended matter in turbid waters. Remote Sensing of
1030 Environment, 114, 854-866
1031 Neukermans, G., Ruddick, K.G., & Greenwood, N. (2012). Diurnal variability of
1032 turbidity and light attenuation in the southern North Sea from the SEVIRI geostationary sensor.
1033 Remote Sensing of Environment, 124, 564-580
1034 Ngoc, D.D., Loisel, H., Jamet, C., Vantrepotte, V., Duforêt-Gaurier, L., Minh, C.D., &
1035 Mangin, A. (2019). Coastal and inland water pixels extraction algorithm (WiPE) from spectral
1036 shape analysis and HSV transformation applied to Landsat 8 OLI and Sentinel-2 MSI. Remote
1038 Novo, E., Hansom, J., & Curran, P. (1989). The effect of sediment type on the
1039 relationship between reflectance and suspended sediment concentration. Remote Sensing, 10,
1040 1283-1289
1041 Novoa, S., Doxaran, D., Ody, A., Vanhellemont, Q., Lafon, V., Lubac, B., & Gernez, P.
1042 (2017). Atmospheric corrections and multi-conditional algorithm for multi-sensor remote sensing
1043 of suspended particulate matter in low-to-high turbidity levels coastal waters. Remote Sensing, 9,
1044 61
1045 Ondrusek, M., Stengel, E., Kinkade, C.S., Vogel, R.L., Keegstra, P., Hunter, C., & Kim,
1046 C. (2012). The development of a new optical total suspended matter algorithm for the
1047 Chesapeake Bay. Remote Sensing of Environment, 119, 243-254
1048 Ouillon, S., Douillet, P., & Andréfouët, S. (2004). Coupling satellite data with in situ
1049 measurements and numerical modeling to study fine suspended-sediment transport: a study for
1050 the lagoon of New Caledonia. Coral Reefs, 23, 109-122
1051 Oyama, Y., Matsushita, B., Fukushima, T., Matsushige, K., & Imai, A. (2009).
1052 Application of spectral decomposition algorithm for mapping water quality in a turbid lake (Lake
1053 Kasumigaura, Japan) from Landsat TM data. ISPRS Journal of Photogrammetry and Remote
1054 Sensing, 64, 73-85
49
1055 Pahlevan, N., Chittimalli, S.K., Balasubramanian, S.V., & Vellucci, V. (2019). Sentinel-
1056 2/Landsat-8 product consistency and implications for monitoring aquatic systems. Remote
1058 Pahlevan, N., Roger, J.-C., & Ahmad, Z. (2017a). Revisiting short-wave-infrared (SWIR)
1059 bands for atmospheric correction in coastal waters. Optics express, 25, 6015-6035
1060 Pahlevan, N., Sarkar, S., Franz, B.A., Balasubramanian, S.V., & He, J. (2017b). Sentinel-
1061 2 MultiSpectral Instrument (MSI) data processing for aquatic science applications:
1062 Demonstrations and validations. Remote Sensing of Environment, 201, 47-56
1063 Pahlevan, N., Schott, J.R., Franz, B.A., Zibordi, G., Markham, B., Bailey, S., Schaaf,
1064 C.B., Ondrusek, M., Greb, S., & Strait, C.M. (2017c). Landsat 8 remote sensing reflectance (R
1065 rs) products: Evaluations, intercomparisons, and enhancements. Remote Sensing of Environment,
1066 190, 289-301
1067 Pahlevan, N., Smith, B., Binding, C., & O’Donnell, D.M. (2017d). Spectral band
1068 adjustments for remote sensing reflectance spectra in coastal/inland waters. Optics express, 25,
1069 28650-28667
1070 Pahlevan, N., Smith, B., Schalles, J., Binding, C.,Cao, Z., Ma, R., Alikas, K., Kangro, K.,
1071 Gurlin, D., Hà, N., Matsushita, B., Moses, W., Greb, S., Lehmann, M., Ondrusek, M., Oppelt,
1072 N., Stumpf, R. (2020). Seamless retrievals of chlorophyll-a from Sentinel-2 (MSI) and Sentinel-3
1073 (OLCI) in inland and coastal waters: A machine-learning approach. Remote Sensing of
1074 Environment
1075 Petus, C., Chust, G., Gohin, F., Doxaran, D., Froidefond, J.-M., & Sagarminaga, Y.
1076 (2010). Estimating turbidity and total suspended matter in the Adour River plume (South Bay of
1077 Biscay) using MODIS 250-m imagery. Continental Shelf Research, 30, 379-392
1078 Pham, Q., Ha, N., Pahlevan, N., Oanh, L., Nguyen, T., & Nguyen, N. (2018). Using
1079 Landsat-8 Images for Quantifying Suspended Sediment Concentration in Red River (Northern
1080 Vietnam). Remote Sensing, 10, 1841
1081 Pope, R.M., & Fry, E.S. (1997). Absorption spectrum (380–700 nm) of pure water. II.
1082 Integrating cavity measurements. Applied Optics, 36, 8710-8723
1083 Reynolds, R.A., Stramski, D., & Neukermans, G. (2016). Optical backscattering by
1084 particles in Arctic seawater and relationships to particle mass concentration, size distribution,
1085 and bulk composition. Limnology and Oceanography, 61, 1869-1890
1086 Ritchie, J.C., Zimba, P.V., & Everitt, J.H. (2003). Remote sensing techniques to assess
1087 water quality. Photogrammetric Engineering & Remote Sensing, 69, 695-704
1088 Schalles, J.F. (2006). Optical remote sensing techniques to estimate phytoplankton
1089 chlorophyll a concentrations in coastal. Remote sensing of aquatic coastal ecosystem processes
1090 (pp. 27-79): Springer
50
1091 Shanmugam, P., Ahn, Y.-H., Ryu, J.-H., & Sundarabalan, B. (2010). An evaluation of
1092 inversion models for retrieval of inherent optical properties from ocean color in coastal and open
1093 sea waters around Korea. Journal of Oceanography, 66, 815-830
1094 Shi, W., & Wang, M. (2009). An assessment of the black ocean pixel assumption for
1095 MODIS SWIR bands. Remote Sensing of Environment, 113, 1587-1597
1096 Shi, W., Zhang, Y., & Wang, M. (2018). Deriving Total Suspended Matter Concentration
1097 from the Near-Infrared-Based Inherent Optical Properties over Turbid Waters: A Case Study in
1098 Lake Taihu. Remote Sensing, 10, 333
1099 Siswanto, E., Tang, J., Yamaguchi, H., Ahn, Y.-H., Ishizaka, J., Yoo, S., Kim, S.-W.,
1100 Kiyomoto, Y., Yamada, K., & Chiang, C. (2011). Empirical ocean-color algorithms to retrieve
1101 chlorophyll-a, total suspended matter, and colored dissolved organic matter absorption
1102 coefficient in the Yellow and East China Seas. Journal of Oceanography, 67, 627-650
1103 Smith, R.C., & Baker, K.S. (1978). Optical classification of natural waters 1. Limnology
1104 and Oceanography, 23, 260-267
1105 Sterckx, S., Knaeps, E., & Ruddick, K. (2011). Detection and correction of adjacency
1106 effects in hyperspectral airborne data of coastal and inland waters: the use of the near infrared
1107 similarity spectrum. International Journal of Remote Sensing, 32, 6479-6505
1108 Sydor, M., Gould, R.W., Arnone, R.A., Haltrin, V.I., & Goode, W. (2004). Uniqueness in
1109 remote sensing of the inherent optical properties of ocean water. Applied Optics, 43, 2156-2162
1110 Tassan, S. (1993). An improved in-water algorithm for the determination of chlorophyll
1111 and suspended sediment concentration from Thematic Mapper data in coastal waters.
1112 International Journal of Remote Sensing, 14, 1221-1229
1113 Turner, A., & Millward, G. (2002). Suspended particles: their role in estuarine
1114 biogeochemical cycles. Estuarine, Coastal and Shelf Science, 55, 857-883
1115 Vanhellemont, Q., & Ruddick, K. (2014). Turbid wakes associated with offshore wind
1116 turbines observed with Landsat 8. Remote Sensing of Environment, 145, 105-115
1117 Volpe, V., Silvestri, S., & Marani, M. (2011). Remote sensing retrieval of suspended
1118 sediment concentration in shallow waters. Remote Sensing of Environment, 115, 44-54
1119 Wang, W.-q., Dong, Q., Shang, S.-l., Wu, J.-y., & Lee, Z.-p. (2009). An evaluation of
1120 two semi-analytical ocean color algorithms for waters of the South China Sea. J Trop oceanogr,
1121 28, 36-42
1122 Wei, J., Lee, Z., & Shang, S. (2016). A system to measure the data quality of spectral
1123 remote‐sensing reflectance of aquatic environments. Journal of Geophysical Research: Oceans,
1124 121, 8189-8207
51
1125 Werdell, P.J., & Bailey, S.W. (2005). An improved in-situ bio-optical data set for ocean
1126 color algorithm development and satellite data product validation. Remote Sensing of
1127 Environment, 98, 122-140
1128 Werdell, P.J., Franz, B.A., & Bailey, S.W. (2010). Evaluation of shortwave infrared
1129 atmospheric correction for ocean color remote sensing of Chesapeake Bay. Remote Sensing of
1130 Environment, 114, 2238-2247
1131 Werdell, P.J., Franz, B.A., Bailey, S.W., Feldman, G.C., Boss, E., Brando, V.E., Dowell,
1132 M., Hirata, T., Lavender, S.J., & Lee, Z. (2013). Generalized ocean color inversion model for
1133 retrieving marine inherent optical properties. Applied Optics, 52, 2019-2037
1134 Werdell, P.J., McKinna, L.I., Boss, E., Ackleson, S.G., Craig, S.E., Gregg, W.W., Lee,
1135 Z., Maritorena, S., Roesler, C.S., & Rousseaux, C.S. (2018). An overview of approaches and
1136 challenges for retrieving marine inherent optical properties from ocean color remote sensing.
1137 Progress in Oceanography, 160, 186-212
1138 Weston, N.B. (2014). Declining sediments and rising seas: an unfortunate convergence
1139 for tidal wetlands. Estuaries and Coasts, 37, 1-23
1140 Woźniak, S.B., & Stramski, D. (2004). Modeling the optical properties of mineral
1141 particles suspended in seawater and their influence on ocean reflectance and chlorophyll
1142 estimation from remote sensing algorithms. Applied Optics, 43, 3489-3503
1143 Woźniak, S.B., Stramski, D., Stramska, M., Reynolds, R.A., Wright, V.M., Miksic, E.Y.,
1144 Cichocka, M., & Cieplak, A.M. (2010). Optical variability of seawater in relation to particle
1145 concentration, composition, and size distribution in the nearshore marine environment at
1146 Imperial Beach, California. Journal of Geophysical Research: Oceans, 115
1147 Yang, W., Matsushita, B., Chen, J., & Fukushima, T. (2011). Estimating constituent
1148 concentrations in case II waters from MERIS satellite data by semi-analytical model optimizing
1149 and look-up tables. Remote Sensing of Environment, 115, 1247-1259
1150 Zhang, M., Tang, J., Dong, Q., Song, Q., & Ding, J. (2010). Retrieval of total suspended
1151 matter concentration in the Yellow and East China Seas from MODIS imagery. Remote Sensing
1152 of Environment, 114, 392-403
1153 Zhang, Y., Zhang, Y., Shi, K., Zha, Y., Zhou, Y., & Liu, M. (2016). A Landsat 8 OLI-
1154 Based, semianalytical model for estimating the total suspended matter concentration in the
1155 slightly turbid Xin’anjiang reservoir (China). IEEE journal of selected topics in applied earth
1156 observations and remote sensing, 9, 398-413
1157 Zheng, G., & DiGiacomo, P.M. (2017). Uncertainties and applications of satellite-derived
1158 coastal water quality products. Progress in Oceanography, 159, 45-72
1159 Zheng, G., Stramski, D., & Reynolds, R.A. (2014). Evaluation of the Quasi-Analytical
1160 Algorithm for estimating the inherent optical properties of seawater from ocean color:
1161 Comparison of Arctic and lower-latitude waters. Remote Sensing of Environment, 155, 194-209
52

Robust Algorithm For Estimating Total Suspended Solids (TSS) in Inland and Nearshore Coastal Waters

Uploaded by

Copyright:

Available Formats

Robust Algorithm For Estimating Total Suspended Solids (TSS) in Inland and Nearshore Coastal Waters

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Robust Algorithm For Estimating Total Suspended Solids (TSS) in Inland and Nearshore Coastal Waters

Uploaded by

Copyright:

Available Formats

Version of Record: https://www.sciencedirect.

1 Robust Algorithm for Estimating Total Suspended Solids (TSS) in

52 Instrument (OLCI), Moderate Resolution Imaging Spectroradiometer (MODIS), Visible Infrared

73 compromised by upstream trapping of suspended sediments by reservoirs, river channelization

92 be related to aquatic reflectance ( ) assuming an isotropic upwelling radiance field, i.e., =

93 / (Mobley 1999). from satellite observations is computed following the removal of

97 (IOPs) are obtained from products.

131 backscattering (b∗ ) coefficients, respectively. Semi-analytical methods that approximate

148 remote sensing.

171 Moderate Resolution Imaging Spectroradiometer (MODIS), Visible Infrared Imaging

190 concentrations, productivity, biodiversity, climate, and watershed biogeochemical/physical

194 as well as bio-optical and biogeochemical data (Table 1).

208 for retrievals (Section 4.1.).

257 assessment of realistic spatio-temporal patterns according to expert expectations.

258 3. Water-type classification

264 Table 2. Classification rules for water type determination

281 increased backscattering in presence of sediments or non-algal particles, as well as the

285 additional criterion to enhance discrimination of water types II and III.

304 approaches for retrievals in Type II and Type III waters.

305 4.1 89: inversion

326 Section 4.1.2.

327 4.1.1 Machine learning approach: Mixture density network

349 diagram of the model is shown in Fig. 4.

373 bands (443, 492, 560, 665, 705nm).

379 NIR band of MSI can be constructed as

394 employed to extrapolate ∗

396 was solved for (865).

407 TUU = 53.736 × (665)?.W@@X (6)

412 0.82) in intense algal bloom conditions (Type II).

424 relationship, i.e., R2 = 0.83, expressed as follows:

425 TUU = G207.57 × (740)K Y 46.78 (7)

433 4.1.2; Fig. A.1)

434 TUU = G224.43 × (865)K Y 12.575 (8)

437 cannot be evaluated; hence, elevated uncertainties are expected.

444 waters (Section 5.3).

445 4.3 Performance metrics

449 training set. The performance metrics are as follows

451 [lm\ = 100 × median [|Er Y Mr |⁄Mr ] for i = 1, …, N (10)

463 comparisons with previous publications.

470 5.1 89: retrievals

533 products at global scales in coastal and inland waters.

553 exhibited saturations in Type III waters.

590 (Table 2).

604 variability in sediment load.

641 temporal analyses.

642 5.4 Extension to other satellite missions

676 algorithm in coastal and inland waters.

677 6.1 Non-algal particle absorption (…†‡: )

678 Using empirical relationships or empiricism in bio-optical modeling or water constituent

685 for the design of global algorithms. Although it is recognized that ∗

687 al. 2009b), for simplicity, we chose a constant value for ∗

688 4.1.2). However, by definition "# (740) (= ∗

701 sediment-rich waters.

756 the challenges of modeling this phenomenon).

764 degrade for large uncertainties in products.

775 provide uncertainty estimates in a computationally inexpensive fashion.

800 reduce such uncertainties.

Land Type-I Type-II Type-III

840 (see Section 5.1).