ArticlePDF Available

A system to measure the data quality of spectral remote sensing reflectance of aquatic environments

  • NOAA/STAR; Global Science & Technology Inc

Abstract and Figures

Spectral remote sensing reflectance (Rrs, sr−1) is the key for ocean color retrieval of water bio-optical properties. Since Rrs from in-situ and satellite systems are subject to errors or artifacts, assessment of the quality of Rrs data is critical. From a large collection of high quality in situ hyperspectral Rrs datasets, we developed a novel quality assurance (QA) system that can be used to objectively evaluate the quality of an individual Rrs spectrum. This QA scheme consists of a unique Rrs spectral reference and a score metric. The reference system includes Rrs spectra of 23 optical water types ranging from purple blue to yellow waters, with an upper and a lower bound defined for each water type. The scoring system is to compare any target Rrs spectrum with the reference and a score between 0 and 1 will be assigned to the target spectrum, with 1 for perfect Rrs spectrum and 0 for unusable Rrs spectrum. The effectiveness of this QA system is evaluated with both synthetic and in situ Rrs spectra and it is found to be robust. Further testing is performed with the NOMAD dataset as well as with satellite Rrs over coastal and oceanic waters, where questionable or likely erroneous Rrs spectra are shown to be well identifiable with this QA system. Our results suggest that applications of this QA system to in situ datasets can improve the development and validation of bio-optical algorithms and its application to ocean color satellite data can improve the short- and long-term products by objectively excluding questionable Rrs data. This article is protected by copyright. All rights reserved.
Content may be subject to copyright.
A system to measure the data quality of spectral
remote-sensing reflectance of aquatic environments
Jianwei Wei
, Zhongping Lee
, and Shaoling Shang
Optical Oceanography Laboratory, School for the Environment, University of Massachusetts Boston, Boston,
Massachusetts, USA,
State Key Laboratory of Marine Environmental Science, Xiamen University, Xiamen, China
Abstract Spectral remote-sensing reflectance (R
) is the key for ocean color retrieval of water
bio-optical properties. Since R
from in situ and satellite systems are subject to errors or artifacts,
assessment of the quality of R
data is critical. From a large collection of high quality in situ hyperspectral
data sets, we developed a novel quality assurance (QA) system that can be used to objectively evaluate
the quality of an individual R
spectrum. This QA scheme consists of a unique R
spectral reference and a
score metric. The reference system includes R
spectra of 23 optical water types ranging from purple blue
to yellow waters, with an upper and a lower bound defined for each water type. The scoring system is to
compare any target R
spectrum with the reference and a score between 0 and 1 will be assigned to the
target spectrum, with 1 for perfect R
spectrum and 0 for unusable R
spectrum. The effectiveness of this
QA system is evaluated with both synthetic and in situ R
spectra and it is found to be robust. Further
testing is performed with the NOMAD data set as well as with satellite R
over coastal and oceanic waters,
where questionable or likely erroneous R
spectra are shown to be well identifiable with this QA system.
Our results suggest that applications of this QA system to in situ data sets can improve the development
and validation of bio-optical algorithms and its application to ocean color satellite data can improve the
short-term and long-term products by objectively excluding questionable R
1. Introduction
Remote-sensing reflectance (R
, units: sr
) is defined as the ratio of water-leaving radiance (L
, units: mW
) to downwelling irradiance just above the surface (E
, units: mWcm
). R
is a critical
optical property for deriving water’s optical and biogeochemical properties that include chlorophyll acon-
centration [O’Reilly et al., 1998], colored dissolved organic material (CDOM) absorption coefficient and partic-
ulate backscattering coefficient [IOCCG, 2006; Lee et al., 2002; Mannino et al., 2008; Wei and Lee, 2015], etc.
Reliable retrieval and appropriate interpretation of these properties, to the first order, demand accurate R
cannot be directly measured in the field or obtained remotely, rather derived from two properties (L
and E
) obtained independently. Errors in R
data from field measurements can be related to instrument
platforms, strategies of deployment, data processing as well as the ambient environments [Bailey et al.,
2008; Hooker et al., 2002; Mueller et al., 2003; Toole et al., 2000; Zibordi et al., 2002]. The in-water approach
measures the in-water downwelling irradiance (E
, units: mWcm
) and upwelling radiance (L
, units:
) separately, and then the in-water radiance is propagated to right above the water sur-
face to obtain the water-leaving radiance. The known sources of measurement uncertainty could originate
from instrument calibrations [Bailey et al., 2008; Wei et al., 2012], instrument self-shading effect [Gordon and
Ding, 1992], wave-focusing effects [Wei et al., 2014; Zibordi et al., 2004], propagation of L
(z) to L
[Wei et al.,
2015], and reflection or shading noise from the ship hull. On the other hand, R
data from the above-water
approach are likely subject to uncertainties due to sea surface reflection [Lee et al., 2010a; Mobley, 1999],
bidirectional reflectance distribution function (BRDF) [Lee et al., 2011; Morel and Gentili, 1996; Voss and Morel,
2005], sky cloudiness, sun glints, and white caps [Gordon and Wang, 1994a]. The skylight-blocked approach
(SBA) [Lee et al., 2013; Tanaka and Sasaki, 2006] can directly measure L
(then R
), but to some degree it is
also subject to self-shading error. In addition to these systematic errors, there are likely unidentified errors
and uncertainties or artifacts during field observations resulted from uncontrollable field environment.
Key Points:
A QA system is developed for
spectral remote-sensing reflectance
The system consists of a reference
and a score metric
It is applicable to both remotely
sensed and in situ ocean color data
Correspondence to:
J. Wei,
Wei, J., Z. Lee, and S. Shang (2016),
A system to measure the data quality
of spectral remote-sensing reflectance
of aquatic environments, J. Geophys.
Res. Oceans,121, doi:10.1002/
Received 5 JUL 2016
Accepted 19 OCT 2016
Accepted article online 25 OCT 2016
C2016. American Geophysical Union.
All Rights Reserved.
Journal of Geophysical Research: Oceans
Obtaining accurate R
spectra from space is even more challenging. The quality of R
spectra derived from a
satellite ocean color sensor is significantly impacted by the efficacy of atmospheric corrections [Gordon and
Wang, 1994b; IOCCG, 2010]. For example, at the top of atmosphere (TOA) about 90% of the total radiance
received by a satellite sensor is contributed from the atmosphere. Small errors in the atmospheric correction
will result in large errors in the derived R
. In coastal regions, atmospheric correction of satellite ocean color
images is a daunting task because of the complex atmospheric and water properties [IOCCG, 2010].
Methods to quality control (QC) R
data are diverse and system dependent [Bailey and Werdell, 2006;
McClain et al., 1992; Ruddick et al., 2005; Werdell and Bailey, 2005; Zibordi et al., 2009b]. The in situ R
urements usually go through certain QC steps during data conversion (e.g., absolute calibration) and subse-
quent data postprocessing (e.g., tilt filtering and data binning); these processes are often embedded in the
processing software provided by the instrument manufacturers. The difficulty is, however, that environmen-
tal disturbance to the measurements and artifacts cannot be fully accounted for in these postprocessing
steps. When compiling the NASA bio-Optical Marine Algorithm Dataset (NOMAD), Werdell and Bailey [2005]
defined and utilized rejection criteria to eliminate abnormal and spurious data. For instance, the observed
spectral surface irradiances were compared with modeled clear sky values, and those stations were consid-
ered questionable and discarded when the in situ value exceeded the modeled value by more than 33%.
Ruddick et al. [2005] discussed the QC of above-water R
measurements at near-infrared (NIR) bands based
on the spectral similarity in the NIR domain. Their method is more applicable to coastal turbid waters and
does not provide further information on the quality of R
spectra in the visible domain. In processing the
AERONET-OC data, Zibordi et al. [2009b] proposed a three-level scheme for the normalized water-leaving
radiance (L
, units: mWcm
) with thresholds including, e.g., L
(k)>20.01 mWcm
to ensure an exclusion of large negative values; L
(412) <L
(443) as commonly met in coastal waters
and L
(1020) <0.1 mWcm
to exclude reflecting obstacles along the optical path between the
instrument and the water surface; and then inspection of the spectral consistency using statistical methods
[D’Alimonte and Zibordi, 2006]. A QC procedure was also developed for satellite ocean color data [McClain
et al., 1992]. This method consists of a system of processing flags for Level-2 products. A total of 31 flags are
currently employed and applied pixel by pixel, including ATMFAIL, STRAYLIGHT, HIGLINT, COASTZ, etc. [Rob-
inson et al., 2003]. These flags use predetermined thresholds. For example, the glint flag is activated when
the normalized sun glint reflectance is found exceeding 0.005. Pixels with stray light contamination from
adjacent, bright sources such as coasts and clouds are flagged in Level-2 products and masked in Level-3
products. Further, to validate satellite R
against in situ data, Bailey and Werdell [2006] introduced a scheme
by considering the spatial uncertainty, coincidence determination, viewing and solar zenith angles, and sta-
tistical confidence of the mean pixel values. To reduce the R
errors from satellite sensor digitization noise,
Hu et al. [2005] recommended the median R
values over a box of 3 33 pixels be obtained to reduce
uncertainties in satellite ocean color products.
A common practice to evaluate the accuracy of the R
data obtained from remote-sensing platforms is to
compare R
data with in situ measurements wavelength-by-wavelength. Statistical measures (e.g., R
relative error, etc.) are often obtained as an indicator of the accuracy of the R
data obtained remotely.
Such an approach characterizes the overall quality of the remotely obtained R
data and provides an indica-
tion of the systematic performance of the instrument and data-processing system for each spectral band
[Bailey and Werdell, 2006; Zibordi et al., 2009a]. This way of evaluation treats the R
property similarly as
those of biogeochemical properties such as the chlorophyll aconcentration, and considers R
at each wave-
length as an independent property. Such scatter-plots or linear regressions wavelength by wavelength can-
not tell the quality of each R
spectrum, however. Note that for remote-sensing inversions [IOCCG, 2000,
2006], it seldom uses R
at one band to derive in-water properties, rather it uses multiple bands or the
entire R
spectrum. Therefore, it is important and imperative to measure objectively the quality of each R
spectrum obtained from any platforms.
In this study, a novel system is developed for objective quality assurance (QA) of an individual R
Based on in situ R
data obtained from coastal waters and clear oceanic waters, we define the domain of
variability of the spectral shapes and amplitudes of R
spectra in different water types using a unique opti-
cal water type classification scheme. A score metric ranging from 0 (lowest quality) to 1 (highest quality) is
further devised to measure the quality of any target R
spectrum. Examples of applications of this QA sys-
tem to synthetic, in situ, and satellite data are also presented.
Journal of Geophysical Research: Oceans 10.1002/2016JC012126
2. Development of the QA system
2.1. In situ Measurements
In situ hyperspectral R
spectra were collected from a wide range of marine environments, with chlorophyll
aconcentration ([CHL], units: mg m
) varying from as low as 0.02 mg m
in the subtropical gyres to tens
of mg m
in the coastal regions (Figure 1). The data were retrieved by above-water approach, SBA [Lee
et al., 2013] and depth-profiling approach, with a wavelength range of 400–800 nm and 10 nm resolution.
In total, there are 958 R
spectra adopted in subsequent analyses. The R
data from the three approaches
account for about 83%, 11%, and 6%, respectively, of the whole data sets. The aggregate data of these mea-
sured hyperspectral R
spectra are illustrated in Figure 2. In this study, R
of nine visible wavelengths (412,
443, 488, 510, 531, 547, 555, 667, and 678 nm) are extracted from the hyperspectral database, and are
referred to as the ‘‘reference’’ data hereafter. It is noted that many of these wavelengths are already incorpo-
rated in heritage and currently operational ocean color satellite sensors including the Sea-viewing Wide
Field-of-view Sensor (SeaWiFS), Moderate-resolution Imaging Spectroradiometer (MODIS), Medium Resolu-
tion Imaging Spectrometer (MERIS), Visible Infrared Imaging Radiometer Suite (VIIRS), and Landsat 8.
2.1.1. R
Data From Above-Water Approach
The primary data for this R
database were obtained from the Gulf of Mexico, Adriatic Sea, South China Sea, North
Pacific Gyre, Mississippi River plume, Yangtze River plume, as well as optically shallow waters in the West Florida
Shelf and around Key West (Florida) (Fig-
ure 1). The range of [CHL] is from 0.02 to
80 mg m
; the range of suspended sed-
iment matter is from the sensors’ lower
detection limit up to 100 g m
range of bottom depth is from 1mto
>1000 m with substrates of seagrass,
sand, and coral reef. The instruments
used are all calibrated commercial radio-
meters from ultraviolet bands to near-
infrared bands including Spectrix (400–
850 nm), GER (350–1050 nm), and RAM-
SES (320–950 nm) and have a fine pixel
dispersion of 2–3 nm with half-value-
bandwidth of 1–1.5 nm. The data-
processing scheme follows the NASA
ocean optics protocol [Muelleretal., 2003]
and the spectral optimization scheme of
Lee et al. [2010a] for refinement.
Figure 1. Locations of field observation of remote-sensing reflectance.
Figure 2. Spectra of remote-sensing reflectance used in this study.
Journal of Geophysical Research: Oceans 10.1002/2016JC012126
2.1.2. R
Data From Floating Instrument With the SBA
Another source of hyperspectral R
was obtained from a floating system incorporating the SBA [Lee et al.,
2013] in coastal waters including Massachusetts Bay, northern Gulf of Mexico, and the Caribbean Sea (south
of Puerto Rico) (Figure 1). The SBA system, as its name implies, consists of a hyperspectral radiometric profil-
er (350–800 nm, Satlantic Inc., Halifax, Canada) and a skylight-blocking apparatus attached to the radiance
sensor. It is thus able to directly measure L
in a continuous manner and provides a time series of R
allows a reliable evaluation of the R
uncertainty. The time series measurements of R
spectra were further
inspected based on the R
data at a single band of 698 nm following the procedure of Wei et al. [2015].
Basically, a density function was first determined for the R
(698) time series, and those deviating by 630%
from the first mode of the density function were then removed. This procedure can effectively eliminate
those data potentially contaminated by sea surface reflected light when the cone is suspended above the
surface and/or immersion of the radiance sensor head within the water at high seas. After filtering, the
median R
spectrum is derived for the remaining R
data and considered as true R
for that water body.
For the waters encountered and the spectral range (400–700 nm) considered, the instrument self-shading
error is found quite small (less than 5%) based on Gordon and Ding [1992].
2.1.3. R
Data From Hyperspectral Profiler
A third source of hyperspectral R
data was retrieved from the North Pacific Gyre and the South Pacific Gyre
using a free-fall hyperspectral profiler (HyperPro, Satlantic Inc., Halifax, Canada). The data-processing proce-
dure follows the NASA ocean optics protocols [Mueller et al., 2003]. Specifically, the depth profiles of upwell-
ing radiance are propagated to right above the water surface to derive the water-leaving radiance L
, using
the determined diffuse attenuation coefficient for upwelling radiance, and the upwelling radiance transmit-
tance [Austin, 1974; Wei et al., 2015]. Then the mean remote-sensing reflectance is derived as the ratio of L
to E
. Because of extreme clarity, the self-shading effect of this system for the 400–700 nm range is negligi-
ble in these waters.
2.2. Clustering of Optical Water Types
The reference R
spectra were first normalized by their respective root of sum of squares (RSS),
nRrsðkÞ5Rrs ðkÞ
where the index Nrepresents the total number of wavelengths, varying from 1 to 9 and k
corresponds to
the wavelengths of 412, 443, 488, 510, 531, 547, 555, 667, and 678 nm. The nR
spectra vary over the range
between 0 and 1, while it retains the ‘‘shapes’’ pertaining to the original R
spectra, i.e., the band ratios of
(k) remain the same as R
The number of data clusters kwas evaluated using the gap method [Tibshirani et al., 2001]. The gap value is
defined as:
nlog ðWkÞ½2log ðWkÞ(2)
where nis the sample size, kis the number of clusters being evaluated, and W
is the pooled within-cluster
dispersion measurement, with
where n
is the number of data points in cluster r, and D
is the sum of the pair-wise distances for all points
in cluster r. The expected value E
)] is determined by Monte Carlo sampling from a reference distri-
bution, and log(W
) is computed from the sample data. According to the gap method, the optimum cluster
number of the nR
data is determined as 23. Interestingly, this number is nearly the same as that of Forel-
Ule water type classes developed 100 years ago [Arnone et al., 2004].
The unsupervised method, K-means clustering technique, was further used to group the nR
K-means clustering, or Lloyd’s algorithm [Lloyd, 1982], is an iterative, data-partitioning algorithm that assigns
nobservations to exactly one of kclusters defined by the centroids, where kis chosen before the algorithm
Journal of Geophysical Research: Oceans 10.1002/2016JC012126
starts (as shown above, k523). We note that the cosine distance was used in the clustering, which is one
minus the cosine of the angle between the nR
spectrum x(a reference) and c(a target),
Each centroid is the mean of the nR
spectra in that cluster, after normalizing those points to unit Euclidean
length. The error function to be minimized is the total within-cluster sum of squares.
The nR
spectra as clustered into the 23 optical water types are illustrated in Figure 3, with the centroids of
spectra highlighted for each water type. Within each water type, the nR
spectra are very similar to
each other and tightly distributed about the centroids. The centroid nR
spectra of the 23 water types are
illustrated in Figure 4 (with values tabulated in Table 1). These ‘‘mean’’ reference spectra represent a large
Figure 3. Clustering of 23 optical water types. The gray-black curves represent the normalized R
spectra classified into each water type.
The mean R
spectra are highlight in red, with the error bars characterizing the maximum and minimum values. And in each water type,
the total numbers of available R
spectra are also given.
Journal of Geophysical Research: Oceans 10.1002/2016JC012126
range of waters with [CHL] varying
from 0.02 mg m
to tens mg m
Corresponding to Figure 3, the upper
limits and lower limits of all nR
tra at discrete bands have been identi-
fied for each water type and are
tabulated in appendix (see Appendix A
Tables A1 and A2). The identified nR
mean spectra, together with the upper
boundary spectra and lower boundary
spectra form the core of the QA system
and will be used in the following sec-
tions to perform the QA of an individu-
al R
It is necessary to emphasize that the
above clustering procedure is focused
on the spectral shapes, and is distinct
from other exercises of optical water
type clustering which often used the
Euclidean distance of R
spectra [Le
et al., 2011; M
elin and Vantrepotte,
2015; Moore et al., 2009, 2001]. The
Euclidean distance, including transformed and weighted forms such as Mahalanobis and likelihood distan-
ces, is inherently insensitive to the shapes of the spectral pattern [Sohn et al., 1999].
2.3. Water Type Matchup and Comparison
To evaluate or quantitatively measure the data quality of a target reflectance spectrum, R
rs, a four-step pro-
cedure is developed:
Step 1 is to match up R
rs(k0) with nR
(k) with regard to the wavelengths. If R
rs has more spectral bands than
that of nR
, we will only choose the same wavelengths with nR
for further analysis. If R
rs has fewer
Figure 4. Mean normalized remote-sensing reflectance spectra characteristics of
23 water types.
Table 1. Mean Normalized Remote-Sensing Reflectance Spectra (nR
) for the 23 Optical Water Types
Water Type
Wavelength (nm)
CHL (mg m
412 443 488 510 531 547 555 667 678
1 0.738 0.535 0.335 0.169 0.112 0.084 0.072 0.007 0.007 0.06
2 0.677 0.534 0.394 0.225 0.156 0.120 0.104 0.011 0.010 0.10
3 0.608 0.521 0.436 0.280 0.204 0.161 0.140 0.016 0.017 0.16
4 0.510 0.478 0.462 0.348 0.279 0.230 0.206 0.029 0.031 0.35
5 0.430 0.436 0.472 0.386 0.326 0.278 0.253 0.038 0.041 0.52
6 0.363 0.387 0.458 0.408 0.368 0.328 0.304 0.042 0.047 0.76
7 0.309 0.355 0.451 0.419 0.392 0.356 0.335 0.048 0.052 0.94
8 0.276 0.315 0.415 0.415 0.414 0.394 0.378 0.062 0.067 1.51
9 0.349 0.335 0.391 0.386 0.387 0.382 0.378 0.090 0.118 1.63
10 0.228 0.275 0.383 0.407 0.430 0.427 0.420 0.079 0.082 2.35
11 0.291 0.276 0.342 0.367 0.401 0.424 0.437 0.129 0.181 3.28
12 0.187 0.241 0.342 0.382 0.427 0.450 0.461 0.147 0.151 3.80
13 0.173 0.220 0.342 0.393 0.447 0.462 0.464 0.093 0.096 4.61
14 0.188 0.235 0.319 0.363 0.412 0.445 0.463 0.215 0.214 4.77
15 0.143 0.191 0.306 0.365 0.434 0.472 0.492 0.170 0.180 6.57
16 0.181 0.200 0.261 0.307 0.365 0.410 0.437 0.359 0.374 7.25
17 0.174 0.203 0.283 0.334 0.399 0.446 0.472 0.272 0.280 7.07
18 0.142 0.169 0.279 0.349 0.439 0.498 0.525 0.121 0.131 10.41
19 0.050 0.126 0.219 0.277 0.340 0.392 0.423 0.452 0.449 10.81
20 0.117 0.153 0.258 0.324 0.412 0.477 0.515 0.243 0.259 12.28
21 0.163 0.175 0.249 0.308 0.400 0.490 0.544 0.190 0.217 16.08
22 0.111 0.135 0.226 0.292 0.385 0.463 0.511 0.310 0.329 17.57
23 0.145 0.133 0.176 0.215 0.286 0.423 0.548 0.341 0.449 34.59
Median chlorophyll aconcentration.
Journal of Geophysical Research: Oceans 10.1002/2016JC012126
wavelengths than the nR
spectra (i.e., N(k0)<9), a subset of nR
(k0) and associated upper boundary spectra
(k0) and lower boundary spectra nR
(k0) will be extracted first for k0.
Step 2 is the normalization of R
rs spectra following equation (1). For the case of N(k0)<9, the new nR
spectra will be rescaled through the normalization procedure of equation (1) so that the RSS of nR
equal to one. Further, the new upper and lower boundary spectra nR
(k) and nR
(k0) will also be rescaled
by the newly rescaled nR
(k0) spectra as below,
nRrsUðkÞ5nRrs UðkÞ
nRrsLðkÞ5nRrs LðkÞ
Step 3 is to assign a water type to the target spectrum by comparing it with the reference nR
spectra. The
spectral similarity between the target spectrum nR
rs and reference spectra nR
are estimated using a spec-
tral angle mapper (SAM) [Kruse et al., 1993],
cos a5X
rs nRrs
where ais the angle formed between the reference spectrum nR
and the normalized target spectrum nR
As a spectral classifier, SAM is able to determine the spectral similarity by treating them as vectors in a space
with dimensionality equal to the number of bands, N. The water type of the target spectrum nR
rs is identi-
fied as the one with the largest cosine values (equivalent to the smallest angles).
Step 4 is the computation of QA scores by comparing the target spectrum nR
rs with the upper and lower
boundaries (nR
and nR
) of the corresponding water type. The number of wavelengths where nR
rs falling
within the boundaries is counted, and used to derive the total score (C
) for the nR
rs spectrum,
where C
is the wavelength-specific score, with Nthe total number of wavelengths for both R
rs and Rref
rs .At
wavelength k
, for example, if nR
) is found beyond either the upper (nR
)) or lower (nR
)) bound-
ary of nR
, a score of 0 will be assigned to this wavelength, i.e., C(k
)50; otherwise, C(k
)51. As suggested
by equation (7), the total score C
will vary within the range of [0, 1]. A higher score indicates higher data
To account for the measurement uncertainty and possible data-processing errors and likely insufficient data
coverage, the original upper boundary and lower boundary are slightly modified by 60.5%, nR
(1 10.005) and nR
3(1 20.005), respectively. Note that this added range of 0.5% is one order of
magnitude smaller than the projected accuracy for radiance measurement [Hooker et al., 1992].
3. Evaluation of the QA System
3.1. Synthesized R
3.1.1. Measurement-Error-Free Data
The R
spectra were synthesized with a database of inherent optical properties (IOP) presented in the Inter-
national Ocean Color Coordinating Group report [IOCCG, 2006]. The synthetic IOP data cover a wide range
of aquatic environments and have been extensively used for model development and validation [Chen
et al., 2015; Salama and Stein, 2009; Wang et al., 2005; Wei and Lee, 2013, 2015; Wei et al., 2016]. The remote-
sensing reflectance was simulated by Hydrolight 5.1 [Mobley and Sundman, 2008]. The inelastic scattering
Journal of Geophysical Research: Oceans 10.1002/2016JC012126
including Raman scattering and fluorescence was invoked and included in the model runs. For chlorophyll
fluorescence, a fluorescence efficiency of 0.02 was adopted by default in Hydrolight. The chlorophyll aspe-
cific absorption spectra a
ph(k) were derived as the ratio of a
(k)/[CHL]. Clear skies were assumed with the
solar zenith angles at 308. Sea surface roughness was relatively mild with wind speed of 5 m/s. The scatter-
ing phase function of particles was assumed to follow Fournier-Forand model, with changing backscattering
ratios. The water absorption coefficient from 400 to 550 nm was adopted from Lee et al. [2015]; while the
measurements by Pope and Fry [1997] were used for the other wavelengths. The scattering coefficients of
pure seawater were provided by Morel [1974].
The data quality of simulated R
spectra at nine wavelengths (412, 443, 488, 510, 531, 547, 555, 667, and
678 nm) is assessed with the-above designed score metric model. As shown in Figure 5a, 65% of the spectra
are identified with the highest quality with scores of 1. Another 18% of the spectra are scored as 8/9, mean-
ing the individual R
spectrum at one wavelength has gone beyond the domain defined by our metrics.
There are another 10% of the R
data found out of the range at two wavelengths. Figure 5b further charac-
terizes the statistics of simulated R
spectra in terms of the two spectral ratios, R
(443) and
(547). The IOCCG data represent a large range of waters varying from water types 1 to 23. And
the relatively low QA scores (in this case, scores 0.5) are found with the spectra in extremely turbid waters
and a few others in relatively clear oceanic waters. These low-score data points can be partly explained by
the limited number of reference spectra used in particular water types in the model development, such as
water types of 17, 19, 21, and 23 (Figure 3).
3.1.2. Error-Disturbed Data
Since the errors in R
data are hardly predictable, we only discussed two simple cases for assessment of the
performance of the QA system. First, we synthesized random-error-disturbed R
spectra. The wavelengths
were randomly chosen from the nine wavelengths of 412, 443, 488, 510, 531, 547, 555, 667, and 678 nm.
Then the percentage errors were randomly set to d55%, 10%, 20%, 30%, or 40%, which are assumed spec-
trally the same for all chosen wavelengths. So the error-disturbed spectra R
rs at wavelengths k
was repre-
sented as R
)* 5R
)3(1 1d). The data quality of R
rs was then examined with the score metric model
and is illustrated in Figure 6a. When the R
spectra are only subjected to small errors (5%, in this case), very
high scores (8/9) are found for 90% of the spectra; 70% of the R
spectra have found with the highest
scores of 1. As expected, with the increasingly larger errors being added more R
spectra show low scores.
It is also noted that a small portion of R
spectra are found assigned to the highest scores of 1 even though
the percentage errors are as high as up to 30% and 40%; this reflects the rare situation where the R
trum is subjected to same percentage errors at all nine wavelengths.
Unlike the random errors considered above, the R
spectra measured from space and in situ are often sub-
ject to spectrally dependent errors [Bailey and Werdell, 2006; Hooker et al., 2002; Hu et al., 2013; Voss et al.,
2010; Zibordi et al., 2009a]. In blue oceanic waters, for example, the percentage errors of R
are more
Figure 5. (a) Frequency distribution of the scores assigned to the simulated R
data. The scores are given in fractional numbers with the
numerators referring to the numbers of good-quality wavebands. (b) Ratio distribution of R
(443) against R
(547). The cor-
responding 23 water types are denoted in colors and the spectra identified with low QA scores (0.5) are denoted with cross symbols in
Journal of Geophysical Research: Oceans 10.1002/2016JC012126
significant in longer wavelengths because of their extremely small reflectance values [Hu et al., 2013]. Like-
wise, the percentage errors of R
spectra in green waters tend to be larger at shorter wavelengths. We syn-
thesized new R
spectra with spectrally related errors. For the data belonging to water types 1–7, a
relatively smaller error (5%, 10%, 15%, 20%, or 30%) and a larger error (correspondingly, 15%, 30%, 50%,
75%, or 100%) are universally added to the shorter wavelengths (412, 443, 488, 510, and 530 nm) and lon-
ger wavelengths (547, 555, 667, and 678 nm), separately. For other water types, we added a relatively larger
error to the shorter wavelengths (412, 443, and 488 nm), a relatively small error to the intermediate wave-
lengths (510, 531, and 547 nm), and relatively large but different error to the longer wavelengths (555, 667,
and 678 nm). Figure 6b illustrates the frequency distribution of the scores for the blue waters (in this case,
water types 1–7), which has shown an increasing number of low-score spectra identified with increasing
errors being added to R
spectra. Figure 6c describes the data quality for the green and yellow waters; the
same conclusion can also be reached. In particular, the last group of frequency distribution in Figure 6c
mimics the satellite ocean color in coastal waters when the absorbing aerosols are present. For such cases,
the model of score metrics works effectively in identifying them by finding low scores.
3.2. Satellite Measured R
The MODIS Aqua ocean color images were selected and retrieved from NASA ocean color data archive,
which were calibrated and processed by Ocean Biology Processing Group (OBPG) with the latest processing
(R2014). The R
spectra at six wavelengths of 412, 443, 488, 531, 547, and 667 nm were further obtained on
Figure 6. Frequency distribution of the scores for simulated error-disturbed R
spectra. (a) Same errors are added to randomly selected
wavelengths. (b) Blue water R
spectra with different errors added to shorter wavelengths (412, 443, 488, 510, and 530 nm) and longer
wavelengths (547, 555, 667, and 678 nm), separately; each pair of errors are denoted in the format of fractions in the plot. (c) Similar to
(b) but for green and yellow water R
spectra; three errors are added to shorter wavelengths (412, 443, and 488 nm), intermediate
wavelengths (510, 531, and 547 nm) and longer wavelengths (555, 667, and 678 nm), separately.
Journal of Geophysical Research: Oceans 10.1002/2016JC012126
a per-pixel basis. Quality of the R
spectrum of each pixel was then quantified by the score metrics, and
examples are provided below.
3.2.1. Level-2 Ocean Color on the U.S. East Coasts
An ocean color image collected over the northeastern U.S. coasts on 15 September 2015 was retrieved from
OBPG, and the sky was clear. As would be expected, the distribution of the QA scores reveals that the quali-
ty of R
data is very much location dependent (Figure 7a). The offshore waters in the Mid-Atlantic Bight and
Southern Atlantic Bight are generally characterized by high QA scores, except for those near the shorelines.
From the New York Bight, Gulf of Maine to St. Lawrence River and further north, however, the R
QA scores
are generally lower than 0.5, suggesting questionable R
data for at least half of the evaluated spectral
bands. In particular, the water masses in the center of the Gulf of Maine are of very low QA scores (less than
0.3), indicating highly questionable R
data for this ocean color image. Indeed, these low-score pixels are
generally flagged in Level-2 ocean color products. For this particular image, the flags of ‘‘production failure
(PRODFAIL)’’ are generally evoked and assigned to the pixels in the vast majority of the waters within the
Gulf of Maine.
On the same day, R
was measured with the SBA system in the southwest of Gulf of Maine to validate the
satellite ocean color data (Figure 7a). Comparison of the matchup data indicates that the MODIS Aqua R
spectrum (which has a QA score of 0.5) deviates from in situ matchup R
spectrum (which has a QA score of 1)
increasingly from 488 nm toward 412 nm (Figure 7b). This finding confirms and validates the low QA score
of the satellite R
spectrum. Note that the MODIS Aqua R
has apparently been underestimated at the
shorter wavelengths, with negative values found at 412 nm, but is fairly reasonable at green and red bands.
Such discrepancies suggest a failure of atmosphere correction for these measurements. The Gulf of Maine is
generally clear and not subject to significant contribution of suspended sediments. A likely cause for the
failure of atmospheric correction is the frequent occurrence of absorbing aerosols in the New England
region [Sierau et al., 2006].
It is interesting, and surprising, that the distribution of the QA score of this MODIS Aqua image is found
related to the mesoscale eddy systems of the Gulf Stream: the clear blue waters in the Sargasso Sea are
observed with relatively lower QA scores (<0.7), with extremely low QA score pixels (<0.1) found for a large
Figure 7. Quality evaluation of the MODISA ocean color data in US East Coasts and Mid-Atlantic Bight. (a) Scores of the satellite ocean col-
or data (September 2015, A2015258180000.L2_LAC). (b) Comparison of the satellite and in situ ocean color matchup in Massachusetts Bay
(in situ station is indicated by a ‘‘’’ symbol in Figure 7a; position: 42.40498N, 270.54698W; sampling time: 17:07 UTC). (c) MODISA mea-
sured ‘‘bad’R
spectra at the low-score pixels (indicated by ‘‘’’ in Figure 7a).
Journal of Geophysical Research: Oceans 10.1002/2016JC012126
water mass stretching northeastward in the east of North Carolina (Figure 7a). These pixels are flagged in
the Level-2 products as perturbation of moderate glints. We retrieved and plotted the R
spectra from this
region and presented Figure 7c, and found that the R
spectra at shorter wavelengths are generally under-
estimated, likely a result of overcorrection of the atmospheric contributions due to glint effects.
Another MODIS Aqua ocean color image was obtained for east U.S coastal and offshore waters on 11
December 2015. The QA score map is shown in Figure 8a. The score distribution has shown very good data
quality for most of the pixels in this scene. The coastal waters in west of the Atlantic Ocean are generally
assigned to very high scores (>0.9), including lower Chesapeake Bay. In situ R
spectrum was obtained on
the same day at a deep water location (off South Carolina) using the SBA system. A comparison of satellite
with in situ R
confirms the good quality of the satellite data for this pixel (Figure 8b).
On the other hand, low-score pixels are also easily found along the edges of cloud patches (the clouds are
masked out and shown in white, north of 288N in the image). Most of these pixels at the edges are flagged
as ‘‘product failure (PRODFAIL)’’ in the Level-2 products. According to the QA score distribution, however,
the pixels with a distance more than a few pixels away from the cloud edges still show very low QA scores
(<0.5). Figure 8c shows the R
spectra of these pixels, where the R
values at the blue wavelengths are
obviously underestimated while R
(670) is likely overestimated.
3.2.2. Level-3 Ocean Color Data in Global Oceans
As examples, the QA metrics was also applied to global Level-3 MODISA daily (A2002188_L3m_9km) and
monthly R
data (A20021822002212_L3m_9km) (Figure 9). Note all questionable pixels as flagged in Level-
2 products have been removed before the Level-3 composites. As a result, the percentage of pixels with
high scores in the Level-3 R
images is generally very high. Further, the frequency of high-score pixels in
monthly product is slightly higher than that of daily product (for brevity, the figures are not shown here),
which is a result of the Level-3 data binning procedure.
The satellite R
measurements in the majority of global oceans have QA scores close to 1, with the highest
scores generally found in the center of the subtropical gyres. Not only the deep oceans but most coastal
waters retained high QA scores for these Level-3 products. Exceptions do exist, however. Relatively lower
QA scores (<0.6) are found in the Mediterranean Sea and in the equatorial upwelling regions in west of
Figure 8. Quality evaluation of the MODISA ocean color data in US East Coasts and the Northwest Atlantic Ocean. (a) Scores of the satellite
ocean color data (11 December 2015, A2015345180500.L2_LAC). (b) Comparison of the satellite ocean color and in situ data measured at a
station indicated by ‘‘’ symbol in Figure 8a (Station number: 20; position: 32.49448N, 277.85818W; sampling time: 17:04 UTC). (c) MODISA
measured ‘‘bad’R
spectra from the pixels indicated by ‘‘’ in Figure 8a.
Journal of Geophysical Research: Oceans 10.1002/2016JC012126
Africa (see Figure 9). Surprisingly, the extremely clear oceanic waters in the North Atlantic Gyre (NAG) do
not show high QA scores either. It should be emphasized this phenomenon of low-score Level-3 R
data in
NAG and the equatorial upwelling region in west of Africa persists almost throughout the year (for brevity,
the figures for other months are omitted here). This low score could be due to underrepresentation of R
NAG in the reference database, where the spectral shapes of these R
might be different from that of other
oligotrophic waters because of likely impacts of the Saharan dusts [e.g., Claustre et al., 2002]. It is necessary
to carry out detailed in situ measurements over this region to pinpoint the exact reasons of the low QA
scores of these clear waters.
3.3. In Situ Measured R
The NOMAD data set was compiled from a large online data depository SeaBASS [Werdell and Bailey, 2005],
and is a widely used for bio-optical model development and validations [Hu et al., 2012; O’Reilly et al., 1998;
Szeto et al., 2011]. Currently the database consists of more than 4000 measurements. For testing and appli-
cation purposes, we only chose the R
spectra with available measurements at wavelengths of 412, 443,
488, 555, and 667 nm, and obtained a total of 2358 R
spectra (Figure 10). Among these spectra, about one-
third was measured in clear oceanic waters with [CHL] <0.25 mg m
; the rest were from turbid coastal
waters. Two-thirds of the spectra were measured by the in-water approach, the rest was obtained from the
above-water approach. After applying the QA system to these selected R
, we found low-quality R
Figure 9. Scores of the MODISA R
data in global oceans. (a) Level-3 daily data (A2002188_L3m_9km). (b) Level-3 monthly data
(A20021822002212_L3m_9km) for July 2002.
Journal of Geophysical Research: Oceans 10.1002/2016JC012126
accidentally included in the NOMAD
data set. For illustration, a few ques-
tionable R
spectra (QA scores <0.5)
are highlighted in Figure 10. In addi-
tion, we found that the QA scores are
not dependent on instruments or
[CHL], and R
with low QA appeared in
every ocean basins (For brevity, the
distribution map is not shown).
The inclusion of low-quality R
data in
NOMAD may impact the related ocean
color algorithm development and sub-
sequent model validations. For exam-
ple, if we only use the R
with the highest scores (QA score 51)
for the development of [CHL] algo-
rithm, the coefficients of the new algo-
rithm (here, called OC3m-QA) will be
quite different than the original OC3m
algorithm, with improved model accu-
racy (see Figure 11).
4. Discussion
4.1. Characteristics of the New Water Type Classification
The score metrics developed in this study relies on cosine-based classification scheme, which groups the R
spectra with similar spectral shapes. This optical classification is distinctive from other practices [Le et al.,
2011; M
elin and Vantrepotte, 2015; Moore et al., 2009, 2001]. The spectral angle mapper [Kruse et al., 1993] is
widely used in land remote-sensing community and its extension to ocean color application deserves more
First, it suppresses the disturbances of
absolute amplitudes of R
spectra to
the clustering, and only takes in
account the spectral shapes of R
direct consequence is that only the
waters with similar ‘‘water color’’ can
be grouped together. For example, the
blue water is by no accident combined
with green or yellow color waters
because of their contrasting R
tral shapes (Figure 3).
Second, the normalization in equation
(1) basically rescales the R
spectra to
the domain of (0,1), rendering the nR
spectra comparable with respect to
their ‘‘amplitudes.’’ Moreover, the nor-
malization retains the same spectral
ratios for nR
with the corresponding
spectra, while at the same time has
constrained the clustered nR
within a very narrow range. To illus-
trate, the coefficient of variation (CV)
of nR
and R
is compared for each
Figure 10. R
spectra extracted from NOMAD database for QA testing (n52358).
The spectra highlighted with symbols (circles, diamonds, square, and triangles)
exemplify the bad R
spectra present in NOMAD.
Figure 11. Comparison of estimated [CHL] with in situ measured [CHL]. The origi-
nal (not quality controlled) data are compared to OC3m modeled data (in green
open circles). The quality controlled data (in pink dots, with the highest scores)
are compared to OC3m-QA estimated data. OC3m-QA has new model coefficients
developed from these highest-score data (N51059, excluding those from the Arc-
tic and Southern Oceans) only. The OC3m-QA has the new model coefficients as
0.2053, 22.0159, 1.1159, 21.3885, and 0.5474.
Journal of Geophysical Research: Oceans 10.1002/2016JC012126
water type in Table 2. The CV’s of nR
spectra are much smaller than the CV’s of the corresponding R
tra in the visible domain. In blue waters, such a tendency is more evident at short wavelengths because of
higher R
values in this spectral domain. In contrast, it is more obvious at longer wavelengths in the green
or yellowish waters.
4.2. Applications to Satellite Ocean Color Validations
A common practice in evaluating the satellite ocean color data quality is through comparison with in situ
data set; with the latter assumed as the ‘‘ground truth’’ [Bailey and Werdell, 2006] by default. Graphically, the
ocean color data matchups can be illustrated in a manner of scatter-plots. Various error statistics can be
derived for the matchups, such as the relative percentage error (e), unbiased absolute percentage error (d),
and the root-mean-square error (RMSE). These estimators convey important information on the overall qual-
ity of various ocean color observations [IOCCG, 2006; Zibordi et al., 2009a]. However, the practices of scatter-
plots alone do not explicitly convey the quality of an individual R
spectrum. The limitation is further illus-
trated in Figure 12 with R
spectra. The newly developed QA system complements this common validation
practice in ocean color products, and assesses the data quality for individual R
The ocean color data measured by different sensors have different number of spectral bands and different
central wavelengths, including MODIS Aqua (412, 443, 488, 531, 547, 667, and 678 nm), SNPP VIIRS (412,
443, 488, 555, and 667 nm), SeaWiFS (412, 443, 488, 510, 531, 555, and 667 nm), MERIS (412, 443, 488, 510,
555, 667, and 678 nm), and Landsat 8 (443, 482, 561, and 655 nm). The QA score system developed here
includes nine spectral bands, which have more bands than any of these satellite ocean color sensors. It is
important to evaluate whether the developed QA system with nine wavelengths can be reliably applied to
data from sources with lower number of spectral bands. To test this impact, we extracted four subsets of
data from the reference data set used in section 2, according to different wavelengths of MODIS Aqua,
SeaWiFS, VIIRS and Landsat 8 (for Landsat 8, we assumed 443, 488, 555, and 667 nm as the central wave-
lengths). We first tested the repeatability of identified water types. It is found that the water type clustering
is fairly accurate, even though different spectral bands are employed. The statistics of the scores deter-
mined for every sensor are presented in Figure 13. For each case, about 90% of the evaluation spectra are
assigned to high QA scores of >0.8. The best performance is found with SeaWiFS data, then MODIS Aqua,
VIIRs and Landsat 8 data (Figure 13b). This is probably related to the fact that SeaWiFS has the highest num-
ber of wavelengths, while the simulated Landsat 8 data only have four bands being used in the QA system.
Note that in the comparisons we ignored the potential effects due to the differences in the spectral
Table 2. Coefficient of Variation (%) of the nR
Spectra and Corresponding R
Spectra for Each Water Type
Water Type
412 nm 443 nm 555 nm 667 nm 412 nm 443 nm 555 nm 667 nm
1 3 2 218625241377
2 3 1 114319181851
3 4 2 115521202078
4 6 4 8 34 27 26 25 46
5 7 4 8 38 37 38 44 96
6 8 3 5 20 43 44 44 48
7 8 4 4 29 60 60 60 66
8 9 5 4 21 61 63 63 64
9 9 4 6 29 82 95 91 145
10 13 6 4 25 60 66 64 87
11 11 8 6 16 108 119 103 129
12 13 7 3 14 58 61 58 65
13 18 11 3 25 39 40 39 44
14 13 8 3 11 42 46 39 46
15 18 9 2 10 52 54 48 53
16 15 11 4 6 105 111 108 111
17 14 9 4 7 86 92 84 91
18 22 13 3 24 55 51 41 50
19 21 17 3 5 58 49 47 46
20 25 15 4 10 52 53 44 48
21 23 14 4 13 51 46 40 46
22 28 16 5 8 67 71 74 74
23 20 18 6 7 46 46 32 36
Journal of Geophysical Research: Oceans 10.1002/2016JC012126
bandwidth (FWHM) of satellite sensors. Landsat 8 has a much wider bandwidth (15–60 nm) than the other
ocean color satellite sensors (10–20 nm) in the visible domain.
On the other hand, for hyperspectral data, a modification or expansion of this QA system can be readily
4.3. Bio-Optical Implications and Significance
The remote-sensing reflectance is probably the most commonly used apparent optical property (AOP) in
ocean color remote sensing. For example, the widely used band ratio [O’Reilly et al., 1998] or band differ-
ence [Hu et al., 2012] algorithms for [CHL] rely on R
data at both blue and green wavelengths. Semianalyti-
cal bio-optical algorithms require the R
spectrum at more wavelengths as inputs [Ciotti et al., 1999; IOCCG,
2000, 2006; Wei and Lee, 2015]. Thus, the QA of every individual R
spectrum is fundamental for accurate
remote sensing of water-column properties. When questionable or erroneous R
spectra are involved, the
bio-optical retrievals from the ocean color will mostly likely be undermined. Our preliminary analysis dem-
onstrates the significance of R
data quality to estimates of [CHL] (Figures 10 and 11). Errors or uncertainties
of R
measurements will propagate to the retrievals of inherent optical retrievals as well [Lee et al., 2010b].
The ‘‘low’’ data quality of MODIS R
found in the North Atlantic Gyre and in the Mediterranean Sea (Figure 9)
may be partly related to an insufficient representativeness of R
spectra in the database for these waters.
Indeed, the oligotrophic waters in North Atlantic and marginal seas could be optically different than that in
the Pacific or South Atlantic, because the former are likely subject to the impacts of dusts transported from
Sahara in the easterly trades and Asian dusts from the west [Uno et al., 2009]. The absorptive fine mineral
particles present in surface waters can attenuate the blue light, addition to the contribution of
Figure 12. Application and limitation of scatter-plot practice in evaluating the data quality of R
spectra. The R
spectrum (in green) rep-
resents true values. The R
and R
spectra refer to two example measurements. The scatter-plot of Figure 12b represents the measure-
ment scenario in Figure 12a, where the R
and R
spectra retain the same shapes as R
. Figure 12c represents another situation when
the R
and R
spectra are erroneously measured (for brevity, we simply switched R
values at some bands between R
and R
Interestingly, the resulted new scatter-plot in Figure 12d is exactly the same with Figure 12b, with the same error statistics.
Journal of Geophysical Research: Oceans 10.1002/2016JC012126
phytoplankton absorption and CDOM and detritus [Claustre et al., 2002]. On the other hand, the presence of
absorptive particles in the atmosphere poses a challenge to atmospheric correction, resulting in under-
estimated R
at blue bands in these waters [Schollaert et al., 2003]. The latter situation could have occurred
in the satellite retrieved ocean color data. A comprehensive investigation with accurate field measurement
of R
is required to completely understand the low QA scores of MODIS R
in those waters.
Nevertheless, it is imperative to QC and quality assure an individual R
spectrum measured from both in
situ and satellite platforms, in order to produce multisensor optical and biogeochemical products with high
5. Conclusions
We have developed a QA system to objectively measure the quality of each individual R
spectrum. The QA
system is based on classification of R
spectral similarity determined by the cosine distance, which is further
measured by the upper and lower boundaries obtained from a wide range of field observations. This scheme
and the spectral constraints together constitute a robust and quantitative QA system. Basically the metric sys-
tem provides information of likely reliability of a target R
spectrum. If a score of 0 is reached, it indicates the
target R
is highly questionable, although it could be true for that specific environment. The performance of
this QA scheme was tested with synthesized, in situ measured, and satellite obtained R
spectra and has
shown great reliability and applicability, where ‘‘bad’’ or highly questionable R
spectra were well identified.
The QA metric system does not require a priori knowledge of the optical properties of waters under study,
and it is applicable to multispectral or hyperspectral R
data obtained from any platform. Since it is based
on in situ R
measurements from a wide range of water types, the innovative QA system can be universally
used for QC of ocean color data from global oceans.
Figure 13. Testing results for the accuracy of score metrics model when applied to satellite ocean color data. (a) MODIS Aqua: 412, 443,
488, 531, 547, and 667 nm; (b) SeaWiFS: 412, 443, 488, 510, 531, 555, and 667 nm; (c) VIIRS: 412, 443, 488, 555, and 667 nm; (d) Landsat
8 (assumed 443, 488, 555, and 667 nm).
Journal of Geophysical Research: Oceans 10.1002/2016JC012126
Because all field measurements are discrete, it is likely that the database used here does not cover every
water types and/or there are situations where the range of R
variability goes beyond the domains defined
here. Such a limitation can be updated or revised when more high-quality in situ measurements are avail-
able. A MATLAB
script is made available ( to facilitate the evalu-
ation and refinement of the score metrics. Nevertheless, this QA scheme provides an easily applicable
system to quantitatively evaluate the quality of individual R
Table A1. Upper Boundary of the Normalized Remote-Sensing Reflectance Spectra for the 23 Water Types
OWT 412 443 488 510 531 54750 555 667 678
1 0.780 0.559 0.367 0.203 0.138 0.109 0.096 0.046 0.047
2 0.711 0.555 0.424 0.254 0.182 0.141 0.126 0.028 0.027
3 0.646 0.540 0.471 0.322 0.243 0.197 0.173 0.067 0.062
4 0.570 0.515 0.528 0.374 0.312 0.265 0.240 0.062 0.062
5 0.478 0.488 0.548 0.418 0.352 0.314 0.301 0.099 0.098
6 0.423 0.416 0.506 0.427 0.390 0.358 0.345 0.065 0.071
7 0.362 0.386 0.485 0.439 0.413 0.378 0.360 0.090 0.096
8 0.328 0.343 0.464 0.449 0.441 0.418 0.412 0.094 0.140
9 0.429 0.369 0.434 0.413 0.412 0.403 0.410 0.166 0.175
10 0.283 0.318 0.471 0.451 0.451 0.454 0.452 0.128 0.125
11 0.360 0.319 0.373 0.400 0.427 0.451 0.477 0.170 0.284
12 0.253 0.287 0.374 0.405 0.439 0.475 0.507 0.183 0.188
13 0.235 0.253 0.392 0.424 0.473 0.486 0.488 0.128 0.134
14 0.263 0.263 0.350 0.382 0.429 0.461 0.507 0.262 0.276
15 0.202 0.219 0.333 0.381 0.448 0.493 0.521 0.203 0.224
16 0.230 0.224 0.296 0.339 0.382 0.432 0.465 0.393 0.419
17 0.232 0.244 0.316 0.355 0.415 0.463 0.503 0.302 0.313
18 0.202 0.204 0.309 0.376 0.455 0.522 0.560 0.163 0.170
19 0.066 0.147 0.236 0.296 0.367 0.415 0.439 0.479 0.493
20 0.159 0.184 0.296 0.356 0.429 0.500 0.571 0.290 0.293
21 0.235 0.237 0.293 0.336 0.443 0.515 0.605 0.241 0.286
22 0.159 0.167 0.251 0.318 0.408 0.482 0.573 0.351 0.383
23 0.180 0.167 0.198 0.233 0.310 0.452 0.578 0.379 0.509
Table A2. Lower Boundary of the Normalized Remote-Sensing Reflectance Spectra for the 23 Water Types
OWT 412 443 488 510 531 547 555 667 678
1 0.709 0.512 0.271 0.119 0.073 0.053 0.044 0.002 0.002
2 0.638 0.509 0.364 0.198 0.132 0.100 0.084 0.003 0.003
3 0.553 0.497 0.412 0.246 0.179 0.140 0.119 0.007 0.007
4 0.436 0.438 0.419 0.310 0.241 0.193 0.169 0.010 0.011
5 0.365 0.390 0.417 0.366 0.287 0.232 0.202 0.016 0.015
6 0.307 0.360 0.405 0.387 0.347 0.297 0.272 0.029 0.028
7 0.251 0.315 0.415 0.403 0.373 0.334 0.306 0.016 0.021
8 0.195 0.266 0.375 0.386 0.390 0.371 0.345 0.023 0.025
9 0.295 0.316 0.367 0.362 0.359 0.352 0.341 0.058 0.066
10 0.131 0.234 0.336 0.381 0.407 0.390 0.376 0.022 0.032
11 0.247 0.240 0.311 0.345 0.366 0.370 0.377 0.085 0.118
12 0.148 0.207 0.302 0.336 0.409 0.425 0.427 0.110 0.115
13 0.092 0.161 0.313 0.375 0.423 0.438 0.436 0.024 0.023
14 0.158 0.200 0.265 0.311 0.382 0.427 0.438 0.154 0.179
15 0.066 0.149 0.273 0.334 0.418 0.455 0.466 0.135 0.143
16 0.156 0.161 0.226 0.282 0.356 0.394 0.417 0.328 0.332
17 0.137 0.176 0.252 0.310 0.388 0.418 0.437 0.244 0.243
18 0.058 0.116 0.249 0.321 0.419 0.480 0.499 0.050 0.054
19 0.032 0.080 0.183 0.246 0.324 0.378 0.411 0.417 0.409
20 0.036 0.096 0.218 0.293 0.395 0.464 0.490 0.204 0.217
21 0.107 0.141 0.199 0.246 0.347 0.464 0.508 0.149 0.171
22 0.073 0.098 0.200 0.249 0.330 0.450 0.485 0.264 0.292
23 0.093 0.095 0.146 0.194 0.265 0.382 0.485 0.301 0.383
Journal of Geophysical Research: Oceans 10.1002/2016JC012126
Arnone, R. A., A. M. Wood, and R. W. Gould Jr. (2004), Science box: The evolution of optical water mass classification, Oceanography,17(2),
Austin, R. W. (1974), The remote sensing of spectral radiance from below the ocean surface, in Optical Aspects of Oceanography, edited by
N. G. Jerlov and E. Steemann Nielsen, pp. 317–344, Academic, New York.
Bailey, S. W., and P. J. Werdell (2006), A multi-sensor approach for the on-orbit validation of ocean color satellite data products, Remote
Sens. Environ.,102(1–2), 12–23, doi:10.1016/j.rse.2006.01.015.
Bailey, S. W., S. B. Hooker, D. Antoine, B. A. Franz, and P. J. Werdell (2008), Sources and assumptions for the vicarious calibration of ocean
color satellite observations, Appl. Opt.,47(12), 2035–2045.
Chen, J., T. Cui, and W. Quan (2015), A neural network-based four-band model for estimating the total absorption coefficients from the
global oceanic and coastal waters, J. Geophys. Res.,120, 36–49, doi:10.1002/2014JC010461.
A. M., J. J. Cullen, and M. R. Lewis (1999), A semi-analytical model of the influence of phytoplankton community structure on the rela-
tionship between light attenuation and ocean color, J. Geophys. Res.,104(C1), 1559–1578.
Claustre, H., A. Morel, S. B. Hooker, M. Babin, D. Antoine, K. Oubelkheir, A. Bricaud, K. Leblanc, B. Queguiner, and S. Maritorena (2002), Is
desert dust making oligotrophic waters greener?, Geophys. Res. Lett.,29(10), 1469, doi:10.1029/2001GL01 4056.
D’Alimonte, D., and G. Zibordi (2006), Statistical assessment of radiometric measurements from autonomous systems, IEEE Trans. Geosci.
Remote Sens.,44(3), 719–728, doi:10.1109/TGRS.2005.862505.
Gordon, H. R., and K. Ding (1992), Self-shading of in-water optical instruments, Limnol. Oceanogr.,37(3), 491–500.
Gordon, H. R., and M. Wang (1994a), Influence ofoceanic whitecaps on atmospheric correction of ocean-color sensor, Appl. Opt.,33, 7754–7763.
Gordon, H. R., and M. Wang (1994b), Retrieval of water-leaving radiance and aerosol optical thickness over the oceans with SeaWiFS: A pre-
liminary algorithm, Appl. Opt.,33(3), 443–452.
Hooker, S. B., W. E. Esaias, G. C. Feldman, W. W. Gregg, and C. R. McClain (1992), An overview of SeaWiFS and Ocean Color, NASA Tech.
Memo. Rep. 10456, 24 pp., NASA Goddard Space Flight Cent., Greenbelt, Md.
Hooker, S. B., G. Lazin, G. Zibordi, and S. McLean (2002), An evaluation of above- and in-water methods for determining water-leaving radi-
ances, J. Atmos. Oceanic Technol.,19(4), 486–515.
Hu, C., F. E. M
uller-Karger, C. Taylor, K. L. Carder, C. Kelble, E. Johns, and C. Heil (2005), Red tide detection and tracing using MODIS fluores-
cence data: A regional example in SW Florida coastal waters, Remote Sens. Environ.,97, 311–321.
Hu, C., Z. P. Lee, and B. Franz (2012), Chlorophyll a algorithms for oligotrophic oceans: A novel approach based on three-band reflectance
difference, J. Geophys. Res.,117, C01011, doi:10.1029/2011JC007395.
Hu, C., L. Feng, and Z. P. Lee (2013), Uncertainties of SeaWiFS and MODIS remote sensing reflectance: Implications from clear water meas-
urements, Remote Sens. Environ.,133(2), 168–182.
IOCCG (2000), Remote sensing of ocean colour in coastal, and other optically-complex, waters, Int. Ocean-Colour Coord. Group Rep. 3, 140
pp, IOCCG, Dartmouth, Nova Scotia, Canada.
IOCCG (2006), Remote sensing of inherent optical properties: Fundamentals, tests of algorithms, and applications, Int. Ocean-Colour Coord.
Group Rep. 5, 126 pp., IOCCG, Dartmouth, Nova Scotia, Canada.
IOCCG (2010), Atmospheric correction for remotely-sensed ocean color products, Int. Ocean-Colour Coord. Group Rep. 10, 78 pp., IOCCG,
Dartmouth, Nova Scotia, Canada.
Kruse, F. A., A. B. Lefkoff, J. B. Boardman, K. B. Heidebrecht, A. T. Shapiro, P. J. Barloon, and A. F. H. Goetz (1993), The spectral image process-
ing system (SIPS)—Interactive visualization and analysis of imaging spectrometer data, Remote Sens. Environ.,44, 145–163.
Le, C., Y. Li, Y. Zha, D. Sun, C. Huang, and H. Zhang (2011), Remote estimation of chlorophyll a in optically complex waters based on optical
classification, Remote Sens. Environ.,115(2), 725–737.
Lee, Z. P., K. L. Carder, and R. Arnone (2002), Deriving inherent optical properties from water color: A multi-band quasi-analytical algorithm
for optically deep waters, Appl. Opt.,41(27), 5755–5772.
Lee, Z. P., Y.-H. Ahn, C. Mobley, and R. Arnone (2010a), Removal of surface-reflected light for the measurement of remote-sensing reflec-
tance from an above-surface platform, Opt. Express,18(25), 26,313–26,342.
Lee, Z. P., R. Arnone, C. Hu, P. J. Werdell, and B. Lubac (2010b), Uncertainties of optical parameters and their propagations in an analytical
ocean color inversion algorithm, Appl. Opt.,49(3), 369–381.
Lee, Z. P., K. Du, K. J. Voss, G. Zibordi, B. Lubac, R. Arnone, and A. Weidemann (2011), An inherent-optical-property-centered approach to
correct the angular effects in water-leaving radiance, Appl. Opt.,50(19), 3155–3167.
Lee, Z. P., N. Pahlevan, Y.-H. Ahn, S. Greb, and D. O’Donnell (2013), Robust approach to directly measuring water-leaving radiance in the
field, Appl. Opt.,52(8), 1693–1701.
Lee, Z. P., J. Wei, K. Voss, M. Lewis, A. Bricaud, and Y. Huot (2015), Hyperspectral absorption coefficient of ‘‘pure’’ seawater in the range of
350–550 nm inverted from remote sensing reflectance, Appl. Opt.,54(3), 546–558.
Lloyd, S. (1982), Least squares quantization in PCM, IEEE Trans. Inform. Theory,28(2), 129–137, doi:10.1109/TIT.1982.1056489.
Mannino, A., M. E. Russ, and S. B. Hooker (2008), Algorithm development and validation for satellite-derived distribu tions of DOC and
CDOM in the US Middle Atlantic Bight, J. Geophys. Res.,113, C07051, doi:10.1029/2007JC004493.
McClain, C. R., W. E. Esaias, W. Barnes, B. Guenther, D. Endres, S. B. Hooker, B. G. Mitchell, and R. Barnes (1992), Calibration and validation
plan for SeaWiFS, in NASA Tech. Memo. 104566, vol. 3, edited by S. B. Hooker and E. R. Firestone, 41 pp., NASA Goddard Space Flight
Cent., Greenbelt, Md.
elin, F., and V. Vantrepotte (2015), How optically diverse is the coastal ocean?, Remote Sens. Environ.,160, 235–251, doi:10.1016/
Mobley, C. D. (1999), Estimation of the remote-sensing reflectance from above-surface measurements, Appl. Opt.,38(36), 7442–7455.
Mobley, C. D., and L. K. Sundman (2008), Hydrolight 5 and Ecolight 5 User’s Guide, 99 pp., Sequoia Scientific, Inc., Bellevue, Wash.
Moore, T. S., J. W. Campbell, and H. Feng (2001), A fuzzy logic classification scheme for selecting and blending satellite ocean color algo-
rithms, IEEE Trans. Geosci. Remote Sens.,39(8), 1764–1776.
Moore, T. S., J. W. Campbell, and M. D. Dowell (2009), A class-based approach to characterizing and mapping the uncertainty of the MODIS
ocean chlorophyll product, Remote Sens. Environ.,113(11), 2424–2430.
Morel, A. (1974), Optical properties of pure water and pure sea water, in Optical Aspects of Oceanograp hy, edited by N. G. Jerlov and E. Stee-
mann Nielsen, 494 pp., Academic, New York.
Morel, A., and B. Gentili (1996), Diffuse reflectance of oceanic waters. III: Implications of bidirectionality for the remote-sensing problem,
Appl. Opt.,35(24), 4850–4862.
This study is funded by the National
Oceanic and Atmospheric
Administration (NOAA) JPSS VIIRS
Ocean Color Cal/Val Project (Lee), the
National Key Research and
Development Program of China
(2016YFC1400900, Shang), the
National Aeronautic and Space
Administration (NASA) Ocean Biology
and Biogeochemistry and Water and
Energy Cycle Programs (Lee), and the
Chinese Ministry of Science and
Technology (2016YFA0601201, Shang),
and the NSF-China (41576169, Shang).
We are indebted to Giuseppe Zibordi
(Joint Research Center, Italy),
Chuanmin Hu (University of South
Florida), Marlon Lewis (Dalhousie
University, Canada) for sharing in situ
radiometric measurements and many
investigators who submitted data to
SEABASS (data used in this effort are
downloadable from http://oceanoptics. We appreciate
the fine work of Kelly Luis in the
correction of the English text. We
thank an anonymous reviewer and
Giuseppe Zibordi for constructive
comments which have improved the
Journal of Geophysical Research: Oceans 10.1002/2016JC012126
Mueller, J. L., G. S. Fargion, and C. R. McClain (2003), Radiometric measurements and data analysis protocols, Ocean Opt. Protoc. for Satell.
Ocean Color Sensor Validation Rep. NASA/TM-2003-21621/Rev-Vol III, 78 pp., NASA Goddard Space Flight Cent., Greenbelt, Md.
O’Reilly, J., S. Maritorena, B. G. Mitchell, D. Siegel, K. L. Carder, S. Garver, M. Kahru, and C. McClain (1998), Ocean color chlorophyll algorithms
for SeaWiFS, J. Geophys. Res.,103, 24,937–24,953.
Pope, R. M., and E. S. Fry (1997), Absorption spectrum (380–700 nm) of pure water. II. Integrating cavity measurements, Appl. Opt.,36(33),
Robinson, W. D., B. A. Franz, F. S. Patt, S. W. Bailey, and P. J. Werdell (2003), Masks and flags updates, in Algorithm Updates for the Fourth
SeaWiFS Data Reprocessing, edited by S. B. Hooker and E. R. Firestone, pp. 34–40, NASA Goddard Space Flight Cent., Greenbelt, Md.
Ruddick, K., V. De Cauwer, and B. Van Mol (2005), Use of the near infrared similarity reflectance spectrum for the quality control of remote
sensing data, Proc. SPIE,5885, doi:10.1117/12.615152.
Salama, M. S., and A. Stein (2009), Error decomposition and estimation of inherent optical properties, Appl. Opt.,48(26), 4947–4962, doi:
Schollaert, S. E., J. A. Yoder, J. E. O’Reilly, and D. L. Westphal (2003), Influence of dust and sulfate aerosols on ocean color spectra and chlo-
rophyll a concentrations derived from SeaWiFS off the U.S. east coast, J. Geophys. Res.,108(C6), 3191, doi:10.1029/2000JC000555.
Sierau, B., D. S. Covert, D. J. Coffman, P. K. Quinn, and T. S. Bates (2006), Aerosol optical properties during the 2004 New England Air Quality
Study–Intercontinental Transport and Chemical Transformation: Gulf of Maine surface measurements—Regional and case studies,
J. Geophys. Res.,111, D23S37, doi:10.1029/2006JD007568.
Sohn, Y., E. Moran, and F. Gurri (1999), Deforestation in North-Central Yucatan(1985–1995)—Mapping secondary succession of forest and
agricultural land use in Sotuta using the cosine of the angle concept, Photogramm. Eng. Remote Sens.,65, 947–958.
Szeto, M., P. J. Werdell, T. S. Moore, and J. W. Campbell (2011), Are the world’s oceans optically different?, J. Geophys. Res.,116, C00H04, doi:
Tanaka, A., and H. Sasaki (2006), Alternative measuring method for water-leaving radiance using a radiance sensor with a domed cover,
Opt. Express,14(8), 3099–3105.
Tibshirani, R., G. Walther, and T. Hastie (2001), Estimating the number of clusters in a data set via the gap statistic, J. R. Stat. Soc. Ser. B,63,
part 2, 411–423.
Toole, D. A., D. A. Siegel, D. W. Menzies, M. J. Neumann, and R. C. Smith (2000), Remote-sensing reflectance determinations in the coastal
ocean environment: Impact of instrumental characteristics and environmental variability, Appl. Opt.,39(3), 456–469.
Uno, I., K. Eguchi, K. Yumimoto, T. Takemura, A. Shimizu, M. Uematsu, Z. Liu, Z. Wang, Y. Hara, and N. Sugimoto (2009), Asian dust trans-
ported one full circuit around the globe, Nat. Geosci.,2(8), 557–560.
Voss, K. J., and A. Morel (2005), Bidirectional reflectance function for oceanic waters with varying chlorophyll concentrations: Measure-
ments versus predictions, Limnol. Oceanogr. Methods,50(2), 698–705.
Voss, K. J., S. McLean, M. R. Lewis, C. Johnson, S. Flora, M. Feinholz, M. Yarbrough, C. C. Trees, M. S. Twardowski, and D. K. Clark (2010), An
example crossover experiment for testing new vicarious calibration techniques for satellite ocean color radiometry, J. Atmos. Oceanic
Technol.,27(10), 1747–1759.
Wang, P., E. Boss, and C. S. Roesler (2005), Uncertainties of inherent optical properties obtained from semianalytical inversions of ocean col-
or, Appl. Opt.,44(19), 4074–4085.
Wei, J., and Z. P. Lee (2013), Model of the attenuation coefficient of daily photosynthetically available radiation in the upper ocean, Meth-
ods Oceanogr.,8, 56–74.
Wei, J., and Z. P. Lee (2015), Retrieval of phytoplankton and color detrital matter absorption coefficients with remote sensing reflectance in
an ultraviolet band, Appl. Opt.,54(4), 636–649.
Wei, J., R. Van Dommelen, M. R. Lewis, S. McLean, and K. J. Voss (2012), A new instrument for measuring the high dynamic range radiance
distribution in near-surface sea water, Opt. Express,20(24), 27,024–27,038.
Wei, J., M. R. Lewis, R. Van Dommelen, C. J. Zappa, and M. S. Twardowski (2014), Wave-induced light field fluctuations in measured irradi-
ance depth profiles: A wavelet analysis, J. Geophys. Res.,119, 1344–1364, doi:10.1002/2013JC009572.
Wei, J., Z. P. Lee, M. Lewis, N. Pahlevan, M. Ondrusek, and R. Armstrong (2015), Radiance transmittance measured at the ocean surface, Opt.
Express,23(9), 11,826–11,837.
Wei, J., Z. P. Lee, M. Ondrusek, A. Mannino, M. Tzortziou, and R. Armstrong (2016), Spectral slopes of the absorption coefficient of colored
dissolved and detrital material inverted from UV-visible remote sensing reflectance, J. Geophys. Res.,121, 1953–1969, doi:10.1002/
Werdell, P. J., and S. W. Bailey (2005), An improved bio-optical data set for ocean color algorithm development and satellite data product
validation, Remote Sens. Environ.,98, 122–140.
Zibordi, G., S. B. Hooker, J. F. Berthon, and D. D’Alimonte (2002), Autonomous above-water radiance measurements from an offshore plat-
form: A field assessment experiment, J. Atmos. Oceanic Technol.,19(5), 808–819.
Zibordi, G., D. D’Alimonte, and J. F. Berthon (2004), An evaluation of depth resolution requirements for optical profiling in coastal waters,
J. Atmos. Oceanic Technol.,21, 1059–1073.
Zibordi, G., J.-F. Berthon, F. M
elin, D. D’Alimonte, and S. Kaitala (2009a), Validation of satellite ocean color primary products at optically
complex coastal sites: Northern Adriatic Sea, Northern Baltic Proper and Gulf of Finland, Remote Sens. Environ.,113, 2574–2591.
Zibordi, G., et al. (2009b), AERONET-OC: A network for the validation of ocean color primary products, J. Atmos. Oceanic Technol.,26(8),
1634–1651, doi:10.1175/2009JTECHO654.1.
Journal of Geophysical Research: Oceans 10.1002/2016JC012126
... Although QAA and GIOP were not originally designed or validated for turbid coastal and inland waters, many studies have performed validation approaches in globally distributed coastal and inland waters, and found regional adaptations to their algorithms can be made to improve performance (Najah and Al-Shehhi, 2021). Specifically, QAA has been applied to many turbid inland lakes in China (Le et al., 2009;Chu et al., 2020;Wei et al., 2016), where shifting the reference wavelength from the red to the NIR, or using multiple reference wavelengths (Pan et al., 2015), improved overall performance. This algorithm has also been applied, and the empirical relationships modified, for more accurate retrievals from coastal regions in the Yellow and East China Seas (Qing et al., 2011). ...
... Suspicious outliers were removed from the predominantly south and east U.S. coastal waters dataset (contributed by John F. Schalles). 84 suspicious outliers were removed, these samples had an Rrs(410)/Rrs(450) > 1.5 and at least one wavelength that did not agree with previously measured ocean color values (i.e., they had a quality assurance score of <1, a metric assessing the proportion of spectral R rs that agree with previously measured ocean color values (Wei et al., 2016)). ...
Full-text available
The simultaneous remote estimation of biogeochemical parameters (BPs) and inherent optical properties (IOPs) from hyperspectral satellite imagery of globally distributed optically distinct inland and coastal waters is a complex, unsolved, non-unique inverse problem. To tackle this problem, we leverage a machine-learning model termed Mixture Density Networks (MDNs). MDNs outperform operational algorithms by calculating the covariance between the simultaneously estimated products. We train the MDNs on a large (N = 8237) dataset of co-aligned, in situ measured, hyperspectral remote sensing reflectance (Rrs), BPs, and absorbing IOPs from globally representative optically distinct inland and coastal waters. The estimated IOPs include absorption due to phytoplankton (aph), chromophoric dissolved organic matter (acdom), and non-algal particles (anap). The estimated BPs include chlorophyll-a, total suspended solids, and phycocyanin (PC). MDNs dramatically reduce uncertainty in the retrievals, relative to operational algorithms, when using a 50/50 dataset split, where the MDNs are trained on a randomly selected half of the in situ dataset and validated on the other half. Our model is shown to have higher, or equivalent, generalization performance than the calculated operational algorithms available for all BPs and IOPs (except PC) via a leave-one-out cross-validation assessment. The MDNs are sensitive to uncertainties in the hyperspectral satellite Rrs, resulting from instrument noise and atmospheric correction; there is a difference of ∼37.4–62.8% (using median symmetric accuracy) between the MDNs' estimates derived from co-located satellite-derived Rrs and in situ Rrs. Of the IOPs, acdom and anap are less sensitive to uncertainties in hyperspectral satellite imagery relative to aph, with remote estimates of aph exhibiting incorrect spectral shape and magnitude relative to in situ measured IOPs. Despite the uncertainties in satellite derived Rrs, the spatial distributions of BPs and IOPs in MDN-derived product maps of Lake Erie and the Curonian Lagoon, based on imagery taken with the Hyperspectral Imager for the Coastal Ocean (HICO) and PRecursore IperSpettrale della Missione Applicativa (PRISMA), are confirmed via co-aligned in situ measurements and agree with the literature's understanding of these well-studied regions. The consistency and accuracy of the model on HICO and PRISMA imagery, despite radiometric uncertainties, demonstrate its applicability to future hyperspectral missions, such as the Plankton, Aerosol, Cloud, ocean Ecosystem (PACE) mission, where the simultaneous estimation model will serve as a key part of phytoplankton community composition analysis.
... Consequently, classes with questionable mean reflectances can also be defined. Some OWT frameworks are primarily used to evaluate the quality of R rs spectra (e.g., Wei et al., 2016). An independent control is the Quality Water Index Polynomial (QWIP) method of Dierssen et al. (2022). ...
... Considering the recommended flags, the suitability of IPF and ACOLITE-DSF in the investigated classifications is insufficient. The work by Liu et al. (2021) also compares IPF, C2RCC, POLYMER, and other OLCI AC methods in context with the optical water type and quality control framework of Wei et al. (2016), which differentiates 22 classes; they conclude that POLYMER has best performance followed by C2RCC and IPF. Figure 3A illustrates a remaining problem, namely that fundamentally different spectral shapes of the derived R rs can often occur in the transition from coast to sea, when the freshwater CDOM concentration is diluted. In some cases, there are features in the TOA signal that can be used to flag potential uncertainties, e.g., a red-edge enhancement ( Figure 3D). ...
Full-text available
Satellite remote sensing allows large-scale global observations of aquatic ecosystems and matter fluxes from the source through rivers and lakes to coasts, marginal seas into the open ocean. Fuzzy logic classification of optical water types (OWT) is increasingly used to optimally determine water properties and enable seamless transitions between water types. However, effective exploitation of this method requires a successful atmospheric correction (AC) over the entire spectral range, i.e., the upstream AC is suitable for each water type and always delivers classifiable remote-sensing reflectances. In this study, we compare five different AC methods for Sentinel-3/OLCI ocean color imagery, namely IPF, C2RCC, A4O, POLYMER, and ACOLITE-DSF (all in the 2022 current version). We evaluate their results, i.e., remote-sensing reflectance, in terms of spatial exploitability, individual flagging, spectral plausibility compared to in situ data, and OWT classifiability with four different classification schemes. Especially the results of A4O show that it is beneficial if the performance spectrum of the atmospheric correction is tailored to an OWT system and vice versa. The study gives hints on how to improve AC performance, e.g., with respect to homogeneity and flagging, but also how an OWT classification system should be designed for global deployment.
... R rs (B 0 ) was derived by assuming a bandwidth of 5 nm for the virtual band and interpolating hyperspectral data at 412 nm. Next, the 33 bands of each sample point were normalized by the arithmetic square root of B 0 to B 32 (Wei et al., 2016): ...
It can be challenging to accurately estimate the Chlorophyll-a (Chl-a) concentration in inland eutrophic lakes due to lakes' extremely complex optical properties. The Orbita Hyperspectral (OHS) satellite, with its high spatial resolution (10 m), high spectral resolution (2.5 nm), and high temporal resolution (2.5 d), has great potential for estimating the Chl-a concentration in inland eutrophic waters. However, the estimation capability and radiometric performance of OHS have received limited examination. In this study, we developed a new quasi-analytical algorithm (QAA716) for estimating Chl-a using OHS images. Based on the optical properties in Dianchi Lake, the ability of OHS to remotely estimate Chl-a was evaluated by comparing the signal-to-noise ratio (SNR) and the noise equivalent of Chl-a (NEChl-a). The main findings are as follows: (1) QAA716 achieved significantly better results than those of the other three QAA models, and the Chl-a estimation model, using QAA716, produced robust results with a mean absolute percentage difference (MAPD) of 11.54 %, which was better than existing Chl-a estimation models; (2) The FLAASH (Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes) atmospheric correction model (MAPD = 22.22 %) was more suitable for OHS image compared to the other three atmospheric correction models we tested; (3) OHS had relatively moderate SNR and NEChl-a, improving its ability to accurately detect Chl-a concentration and resulting in an average SNR of 59.47 and average NEChl-a of 72.86 μg/L; (4) The increased Chl-a concentration in Dianchi Lake was primarily related to the nutrients input, and this had a significant positive correlation with total nitrogen. These findings expand existing knowledge of the capabilities and limitations of OHS in remotely estimating Chl-a, thereby facilitating effective water quality management in eutrophic lake environments.
... QC of the reflectance spectra Reflectance spectra, while highly variable between different environments, are constrained in the spectral shapes that they exhibit. A tool based on a library of high-quality spectra has been created to assess the likelihood that a reflectance spectra is reasonable [47], assigning a quality score that can be used to filter nonphysical spectra and suggest which AC works best in a given environment. The tool is available for Sentinel-2 and Landsat-8 ( for which a score between 0 and 1 is assigned to the targeted spectrum, with 1 for likely good R rs spectra and 0 for likely unusable R rs spectra. ...
Full-text available
Satellites have provided high-resolution ( < 100 m) water color (i.e., remote sensing reflectance) and thermal emission imagery of aquatic environments since the early 1980s; however, global operational water quality products based on these data are not readily available (e.g., temperature, chlorophyll- a , turbidity, and suspended particle matter). Currently, because of the postprocessing required, only users with expressive experience can exploit these data, limiting their utility. Here, we provide paths (recipes) for the nonspecialist to access and derive water quality products, along with examples of applications, from sensors on board Landsat-5, Landsat-7, Landsat-8, Landsat-9, Sentinel-2A, and Sentinel-2B. We emphasize that the only assured metric for success in product derivation and the assigning of uncertainties to them is via validation with in situ data. We hope that this contribution will motivate nonspecialists to use publicly available high-resolution satellite data to study new processes and monitor a variety of novel environments that have received little attention to date.
... We quantitatively assessed the quality of the hyperspectral data by using the quality assurance system (QAS) proposed by Wei and Lee [52]. The QAS consists of four steps. ...
Full-text available
Inaccuracies in the atmospheric correction (AC) of data on coastal waters significantly limit the ability to quantify the parameters of water quality. Many studies have compared the effects of the atmospheric correction of data provided by the Sentinel−2 satellites, but few have investigated this issue for coastal waters in China owing to a limited amount of in situ spectral data. The authors of this study compared four processors for the atmospheric correction of data provided by Sentinel−2—the Atmospheric Correction for OLI ‘lite’(ACOLITE), Sea-viewing Wide Field-of-view Sensor (SeaWiFS) Data Analysis System (SeaDAS), Polynomial-based algorithm applied to MERIS (POLYMER), and Case 2 Regional Coast Colour (C2RCC)—to identify the most suitable one for water bodies with different turbidities along the coast of China. We tested the algorithms used in these processors for turbid waters and compared the resulting inversion of the remote sensing reflectance (Rrs) using in situ reflectance data from three stations with varying levels of coastal turbidity (HTYZ, DONG’OU, and MUPING). All processors significantly underestimated the results on data from the HTYZ station, which is located along waters with high turbidity, with the SeaDAS delivering the best performance, with an average band of 0.0146 and an average of 29.80%. It was followed by ACOLITE, with an average band of 0.0213 and an average of 43.43%. The performance of two AC algorithms used in ACOLITE, dark spectrum fitting (DSF) and exponential extrapolation (EXP), was also evaluated by comparing their results with in situ measurements at the HTYZ site. The ACOLITE-EXP algorithm delivered a slight improvement in results for the blue band compared with the DSF algorithm in highly turbid water, but led to no significant improvement in the green and red bands. C2RCC delivered the best performance on data from the DONG’OU station, which is located along water with medium turbidity, and from the MUPING station (water with low turbidity), with values of the of 18.58% and 28.41%, respectively.
... The definition of the number of OWTs can be assessed by classifications and/or indexes regarding the particularities of the water body. Knowing the OWT present in the study area is essential to improve the application and understanding of the algorithms used for ocean colour retrievals (Moore et al. 2014;Turner et al. 2022;Valerio et al. 2021;Wei, Lee, and Shang 2016). ...
Retrieving chlorophyll-a concentration (Chla) from remote-sensing data is especially challenging in coastal regions because of the presence of other co-dominant optically significant constituents (OSCs). In this study, we characterize bio-optical variability at a fixed coastal time series station in the South Brazil Bight (SBB), the ANTARES-Ubatuba site (São Paulo, Brazil) from radiometric measurements and water sample analyses, and evaluate the performance of ocean colour (OC) algorithms for Chla retrieval. In situ Chla and the spectral absorption of phytoplankton (aph), detritus, and coloured dissolved organic matter (CDOM), as well as the above-water remote-sensing reflectance (Rrs) were obtained from 2004 to 2019 at a quasi-monthly sampling frequency. Chla exhibited high variability, with a standard deviation twice as large as the mean (1.0 ± 1.83 mg m⁻³), and values ranging from 0.18 mg m⁻³ up to 18.09 mg m⁻³ during episodic bloom events. CDOM was the dominant OSC contributing to the absorption coefficient at 443 nm all-year round, suggesting the influence of continental sources, even though the station is located at the external limit of the inner shelf (40 m). Two optical water types were defined: one composed mainly of ‘blue’ waters with lower concentrations of the OSCs and another of ‘green’ waters with higher concentrations of OSCs. Despite the CDOM spectral dominance, the empirical OCx performed reasonably well with a positive bias for both in situ (17–25%) and satellite (25–31%) Chla retrievals. The performance was slightly better using the 3-band OC3 algorithm by selecting the best band ratio with lower influence of CDM (CDOM + detritus) absorption (Rrs (490/555)). The main sources of uncertainty were caused by higher CDM proportions and phytoplankton-specific absorption, yielding positive biases on the OC3 retrievals. The results suggest that standard satellite products of Chla can be used (with some caution) to monitor and study the dynamics of the phytoplankton biomass within the region, knowing the expected uncertainties.
... The eight wavelength dataset obtained from NOMAD is geographically limited. All in situ R rs data are filtered following a quality assurance (QA) system [55] to exclude any low quality R rs with a QA score smaller than 0.5. ...
We investigated the optimal number of independent parameters required to accurately represent spectral remote sensing reflectances ( R rs ) by performing principal component analysis on quality controlled in situ and synthetic R rs data. We found that retrieval algorithms should be able to retrieve no more than four free parameters from R rs spectra for most ocean waters. In addition, we evaluated the performance of five different bio-optical models with different numbers of free parameters for the direct inversion of in-water inherent optical properties (IOPs) from in situ and synthetic R rs data. The multi-parameter models showed similar performances regardless of the number of parameters. Considering the computational cost associated with larger parameter spaces, we recommend bio-optical models with three free parameters for the use of IOP or joint retrieval algorithms.
Full-text available
The spectral slope of the absorption coefficient of colored dissolved and detrital material (CDM), Scdm (units: nm−1), is an important optical parameter for characterizing the absorption spectral shape of CDM. Although highly variable in natural waters, in most remote sensing algorithms, this slope is either kept as a constant or empirically modeled with multi-band ocean color in the visible domain. In this study, we explore the potential of semi-analytically retrieving Scdm with added ocean color information in the ultraviolet (UV) range between 360-400 nm. Unique features of hyperspectral remote sensing reflectance in the UV-visible wavelengths (360-500 nm) have been observed in various waters across a range of coastal and open ocean environments. Our data and analyses indicate that ocean color in the UV domain is particularly sensitive to the variation of the CDM spectral slope. Here, we used a synthesized dataset to show that adding UV wavelengths to the ocean color measurements will improve the retrieval of Scdm from remote sensing reflectance considerably, while the spectral band settings of past and current satellite ocean color sensors cannot fully account for the spectral variation of remote sensing reflectance. Results of this effort support the concept to include UV wavelengths in the next generation of satellite ocean color sensors. This article is protected by copyright. All rights reserved.
Full-text available
The flags and masks used for the SeaWiFS level-2 and level-3 processing were updated for the recent fourth reprocessing. This chapter discusses the changes and why they were made. In many cases, underlying algorithms were changed. Some flags changed their states to either flagging (noting a condition), or masking (denoting data excluded from the product) to allow more data to be kept or to improve its quality. New flags were either introduced as a part of new algorithms or to denote the status of the data more clearly. The flag and mask changes significantly contributed to the improvement in the data quality and increased the amount of data retrieved.
Full-text available
A large data set containing coincident in situ chlorophyll and remote sensing reflectance measurements was used to evaluate the accuracy, precision, and suitability of a wide variety of ocean color chlorophyll algorithms for use by SeaWiFS (Sea-viewing Wide Field-of-view Sensor). The radiance-chlorophyll data were assembled from various sources during the SeaWiFS Bio-optical Algorithm Mini-Workshop (SeaBAM) and is composed of 919 stations encompassing chlorophyll concentrations between 0.019 and 32.79 μg L-1. Most of the observations are from Case I nonpolar waters, and ~20 observations are from more turbid coastal waters. A variety of statistical and graphical criteria were used to evaluate the performances of 2 semianalytic and 15 empirical chlorophyll/pigment algorithms subjected to the SeaBAM data. The empirical algorithms generally performed better than the semianalytic. Cubic polynomial formulations were generally superior to other kinds of equations. Empirical algorithms with increasing complexity (number of coefficients and wavebands), were calibrated to the SeaBAM data, and evaluated to illustrate the relative merits of different formulations. The ocean chlorophyll 2 algorithm (OC2), a modified cubic polynomial (MCP) function which uses Rrs490/Rrs555, well simulates the sigmoidal pattern evident between log-transformed radiance ratios and chlorophyll, and has been chosen as the at-launch SeaWiFS operational chlorophyll a algorithm. Improved performance was obtained using the ocean chlorophyll 4 algorithm (OC4), a four-band (443, 490, 510, 555 nm), maximum band ratio formulation. This maximum band ratio (MBR) is a new approach in empirical ocean color algorithms and has the potential advantage of maintaining the highest possible satellite sensor signal: noise ratio over a 3-orders-of-magnitude range in chlorophyll concentration.
Full-text available
A new spectral pattern matching approach that utilizes the spectral angle (the cosine of the angle) concept was used for mapping deforestation and successional stages of forest regrowth in Sotuta in the state of Yucatan, Mexico. By calculating spectral angles between finely defined spectral clusters and known reference signatures, and assigning each spectral cluster to one of the reference classes based on the minimum spectral angle rule, we were able to map forest regrowth stages and agricultural land-use classes. Our research shows that, by adapting a spectral pattern matching approach demonstrated in this paper, spectral clusters can be assigned into information classes precisely and objectively, and the inconsistency involved in visual interpretations can be avoided. The conceptual difference between the spectral distance and spectral angle in feature space is also reviewed. In the study area, the rate of deforestation is high and agricultural land use is intensifying increasingly. The limited amount of land granted to ejidos and rapid population growth seem to be major causes of deforestation in the study area.
Full-text available
The radiance transmittance (Tr) is the ratio of the water-leaving radiance (Lw(0+)) to the sub-surface upwelling radiance (Lu(0-)), which is an important optical parameter for ocean optics and ocean color remote sensing. Historically, a constant value (~0.54) based on theoretical presumptions has been adopted for Tr and is widely used. This optical parameter, however, has never been measured in the aquatic environments. With a robust setup to measure both Lu(0-) and Lw(0+) simultaneously in the field, this study presents Tr in the zenith direction between 350 and 700 nm measured in a wide range of oceanic waters. It is found that the measured Tr values are generally consistent with the long-standing theoretical value of 0.54, with mean relative difference less than 10%. In particular, the agreement within the spectral domain of 400-600 nm is found to be the best (with the averaged difference less than 5%). The largest difference is observed for wavelengths longer than 600 nm with the average difference less than 15%, which is related to the generally very small values in both Lu(0-) and Lw(0+) and rough environmental conditions. These results provide a validation of the setup for simultaneous measurements of upwelling radiance and water-leaving radiance and confidence in the theoretical Tr value used in ocean optics studies at least for oceanic waters.
Full-text available
Coastal regions are a resource for societies while being under severe pressure from a variety of factors. They also show a large diversity of optical characteristics, and the potential to optically classify these waters and distinguish similarities between regions is a fruitful application for satellite ocean color. Recognizing the specificities and complexity of coastal waters in terms of optical properties, a training data set is assembled for coastal regions and marginal seas using full resolution SeaWiFS global remote sensing reflectance RRS data that maximize the geographic coverage and seasonal sampling of the domain. An unsupervised clustering technique is operated on the training data set to derive a set of 16 classes that cover conditions from very turbid to oligotrophic. When applied to a global seven-year SeaWiFS data set, this set of optical water types allows an efficient classification of coastal regions, marginal seas and large inland water bodies. Classes associated with more turbid conditions show relative dominance close to shore and in the mid-latitudes. A geographic partition of the global coastal ocean serves to distinguish general optical similarities between regions. The local optical variability is quantified by the number of classes selected as dominant across the period, averaging 5.2 classes if the cases accounting for 90% of the data days are considered. Optical diversity is more specifically analyzed with a Shannon index computed with the class memberships. Regions with low optical diversity are the most turbid waters as well as closed seas and inland water bodies. Oligotrophic waters also show a relatively low diversity, while intermediate regions between coastal domain and open ocean are associated with the highest diversity, which has interesting connections with ecological features.
Since the advent of manned spaceflight, there have been many interesting examples of photography obtained from space over ocean and coastal regions. The potential applications of such space photography to the solution of oceanographic problems and the even greater potential of directly-obtained spectroradiometric data are manifestly obvious. The visible region of the electromagnetic spectrum is unique in that it allows information to be carried from below the ocean surface to a remote sensor. More specifically, it is only in the narrower spectral region from about 400 to 600 nanometers that sensible data can be obtained from water depths greater than a few meters. Beyond this wavelength, the remote sensing technique becomes more useful for the appraisal of surface than of subsurface phenomena. At the shorter wavelengths, that is below about 400 nm, the water absorption also increases rapidly in all but the clearest of oceanic waters. Perhaps even more significant at the shorter wavelengths is the rapid increase of the veiling effects of the atmosphere. As the wavelength decreases that portion of the radiation emanating from the ocean which reaches the remote sensor becomes insignificant when compared to the radiance of the scattered daylight in the path of sight. The remote optical sensor utilizes the spectral, spatial and temporal information encoded in the radiance signal which it receives. Inferences about such variables as the water depth, turbidity, or chlorophyll level are drawn after this encoded information has been processed in some manner. Unfortunately, the apparent signal with which the remote sensor must work usually differs markedly from the inherent signal which exists at or just below the water surface. As it is this latter signal which contains the desired information about the water or the ocean bottom, it is important to understand how it is attenuated and modified in travelling upward through the atmosphere to the sensor.