ArticlePDF Available

Improving Satellite Global Chlorophyll a Data Products Through Algorithm Refinement and Data Recovery


Abstract and Figures

A recently developed algorithm to estimate surface ocean chlorophyll a concentrations (Chl in mg m ⁻³ ), namely, the ocean color index (OCI) algorithm, has been adopted by the U.S. National Aeronautics and Space Administration to apply to all satellite ocean color sensors to produce global Chl maps. The algorithm is a hybrid between a band-difference color index algorithm for low-Chl waters and the traditional band-ratio algorithms (OCx) for higher-Chl waters. In this study, the OCI algorithm is revisited for its algorithm coefficients and for its algorithm transition between color index and OCx using a merged data set of high-performance liquid chromatography and fluorometric Chl. Results suggest that the new OCI algorithm (OCI2) leads to lower Chl estimates than the original OCI (OCI1) for Chl < 0.05 mg m ⁻³ , but smoother algorithm transition for Chl between 0.25 and 0.40 mg m ⁻³ . Evaluation using in situ data suggests that similar to OCI1, OCI2 has significantly improved image quality and cross-sensor consistency between SeaWiFS, MODISA, and VIIRS over the OCx algorithms for oligotrophic oceans. Mean cross-sensor difference in monthly Chl data products over global oligotrophic oceans reduced from ~10% for OCx to 1–2% for OCI2. More importantly, data statistics suggest that the current straylight masking scheme used to generate global Chl maps can be relaxed from 7 × 5 to 3 × 3 pixels without losing data quality in either Chl or spectral remote sensing reflectance (R rs (λ), sr ⁻¹ ) for not just oligotrophic oceans but also more productive waters. Such a relaxed masking scheme yields an average relative increase of 39% in data quantity for global oceans, thus making it possible to reduce data product uncertainties and fill data gaps.
This content is subject to copyright. Terms and conditions apply.
Improving Satellite Global Chlorophyll aData Products
Through Algorithm Renement and Data Recovery
Chuanmin Hu
, Lian Feng
, Zhongping Lee
, Bryan A. Franz
, Sean W. Bailey
P. Jeremy Werdell
, and Christopher W. Proctor
College of Marine Science, University of South Florida, St. Petersburg, FL, USA,
School for the Environment, University
of Massachusetts Boston, Boston, MA, USA,
NASA GSFC, Greenbelt, MD, USA,
Science Systems and Applications, Inc.,
Lanham, MD, USA
Abstract A recently developed algorithm to estimate surface ocean chlorophyll aconcentrations
(Chl in mg m
), namely, the ocean color index (OCI) algorithm, has been adopted by the U.S. National
Aeronautics and Space Administration to apply to all satellite ocean color sensors to produce global Chl
maps. The algorithm is a hybrid between a banddifference color index algorithm for lowChl waters
and the traditional bandratio algorithms (OCx) for higherChl waters. In this study, the OCI algorithm is
revisited for its algorithm coefcients and for its algorithm transition between color index and OCx
using a merged data set of highperformance liquid chromatography and uorometric Chl. Results suggest
that the new OCI algorithm (OCI2) leads to lower Chl estimates than the original OCI (OCI1) for
Chl < 0.05 mg m
, but smoother algorithm transition for Chl between 0.25 and 0.40 mg m
. Evaluation
using in situ data suggests that similar to OCI1, OCI2 has signicantly improved image quality and
crosssensor consistency between SeaWiFS, MODISA, and VIIRS over the OCx algorithms for oligotrophic
oceans. Mean crosssensor difference in monthly Chl data products over global oligotrophic oceans
reduced from ~10% for OCx to 12% for OCI2. More importantly, data statistics suggest that the current
straylight masking scheme used to generate global Chl maps can be relaxed from 7 × 5 to 3 × 3 pixels without
losing data quality in either Chl or spectral remote sensing reectance (R
(λ), sr
) for not just
oligotrophic oceans but also more productive waters. Such a relaxed masking scheme yields an average
relative increase of 39% in data quantity for global oceans, thus making it possible to reduce data product
uncertainties and ll data gaps.
Plain Language Summary There are generally two issues with any remote sensing data product:
data quality (accuracy) and data quantity (coverage). In this work, these two issues for global ocean data
products of chlorophyll aconcentrations (Chl in mg m
) are investigated through revisiting a recently
developed algorithm concept and statistical analyses of cloudadjacent data. The use of more data in
algorithm development leads to slightly different algorithm coefcients and smoother transition between
clear and turbid waters, and a new straylight masking scheme is proposed to recoversome of the
previously masked data in the global data products. The new algorithm leads to signicantly improved
crosssensor consistency (SeaWiFS, MODIS, VIIRS) as compared to the traditional bandratio algorithms,
with mean monthly difference reduced from 10% to 12%. The new straylight masking scheme leads to
relative increase of 39% in data quantity in the global ocean without losing data quality. These improvements
are expected to reduce uncertainties and ll gaps in the global data products.
1. Introduction
One of the most historically meaningful and utilized data products from ocean color satellites is surface con-
centration of chlorophyll a(Chl in mg m
) because this photosynthetic pigment plays a fundamental role in
affecting ocean biology and ecology. To many researchers, Chl and ocean colorhave become synonymous,
as Chl (plus accessary phaeopigment) has been the main data product since the proofofconcept Coastal
Zone Color Scanner (19781986) era. Indeed, all satellite ocean color missions have Chl as one of their main
data products, for example, the SeaWiFS mission (Seaviewing Wide Fieldofview Sensor; 19972010),
MODIS (Moderate Resolution Imaging Spectroradiometer; 1999 to present on Terra and 2002 to present
on Aqua), MERIS (Medium Resolution Spectroradiometer; 20022012), and the most recent VIIRS
(Visible Infrared Imaging Radiometer Suite; 2012 to present).
©2019. American Geophysical Union.
All Rights Reserved.
Key Points:
OCI chlorophyll aalgorithm
revisited for its algorithm
coefcients and transition
New qualitycontrol scheme
proposed to mask straylight
Both data quality and data quantity
improved for global oceans
Correspondence to:
C. Hu,
Hu, C., Feng, L., Lee, Z., Franz, B. A.,
Bailey, S. W., Werdell, P. J., & Proctor,
C. W. (2019). Improving satellite global
chlorophyll adata products through
algorithm renement and data
recovery. Journal of Geophysical
Research: Oceans,124, 15241543.
Received 3 JAN 2019
Accepted 9 FEB 2019
Accepted article online 14 FEB 2019
Published online 7 MAR 2019
HU ET AL. 1524
Deriving accurate Chl data products from measurements several hundreds of kilometers above the Earth
requires a thorough understanding of the entire system of ocean color measurements from the satellite to
the sea surface. This includes the development and implementation of algorithms for global routine proces-
sing such as onorbit temporal stability corrections, vicarious calibration, atmospheric correction (including
whitecap and Sun glint corrections), and biooptical algorithm development. All steps except the last one are
designed to obtain accurate spectral remote sensing reectance in every spectral band (R
(λ), sr
), and the
last one is to estimate Chl from R
(λ). Descriptions of all steps above are beyond the scope of this paper as
they can be found in the literature (e.g., SeaWiFS technical report series; Franz et al., 2007; Frouin et al.,
1996; Gordon, 1997; Gordon & Wang, 1994; McClain, 2009; McClain et al., 2004; Moore et al., 2015;
Stumpf et al., 2003; Wang & Bailey, 2001; Wang & Shi, 2007). Instead, the focus of this paper is on a Chl algo-
rithm optimized for global oceans (i.e., data quality) and also on data quantity through quality control.
A review of Chl algorithms for global oceans has been provided in recent papers including Dierssen (2010) and
Hu and Campbell (2014). Briey, there are two general approaches to estimate Chl from satellitederived
(λ). The rst is empirical, through regression of R
(λ) band ratios or band differences against Chl (e.g.,
Hu et al., 2012; Kahru & Mitchell, 1999; O'Reilly et al., 2000). The second is semianalytical, through solving
(λ) equations established from simplications of radiative transfer theory using certain biooptical
assumptions (e.g., Carder et al., 1999; IOCCG, 2006; Lee et al., 2002; Maritorena et al., 2002; Sathyendranath
et al., 1989). While in principle the latter should lead to more accurate results than the former because water
constituents affecting R
(λ) are explicitly separated, in practice both approaches have their own strengths and
weaknesses. For example, the latter relies on tuning of the assignment of assumptions in the biooptical rela-
tionships using global or local data sets. The former accounts for R
(λ) input errors through tuning of the
empirical coefcients, although the empirical design makes it impossible to differentiate explicitly the various
inwater constituents. On the other hand, empirical approaches through either band ratio or band difference
can partially compensate some of the R
(λ) input errors because these errors to a large extent are spectrally
related (Hu et al., 2013). Thus, following the tradition from the CZCS era, even though empirical Chl estimates
are known to contain large uncertainties (Dierssen, 2010) and such uncertainties may vary across different
ocean basins (Szeto et al., 2011), the U.S. National Aeronautics and Space Administration (NASA) still uses
empirical algorithms to produce default (also called standard) Chl products from the mainstream sensors.
Recently, NASA adapted a new Chl algorithm proposed by Hu et al. (2012) as the default global algorithm in
their reprocessing 2014.0, with new processing codes updated in the software package SeaDAS (version 7.2).
The new algorithm is a hybrid of a novel concept of threeband difference color index (CI) and the tradition-
ally band ratio algorithm (OC4 or OC3, termed as OCx in this paper; O'Reilly et al., 2000) to apply to lowChl
waters (Chl 0.25 mg m
) and higherChl waters (Chl > 0.3 mg m
), respectively. For intermediateChl
waters in the transition zone of 0.250.3 mg m
, a weighted mixture of the two algorithms is used.
Collectively, the algorithm is termed as Ocean Color Index (OCI) algorithm. Following this change, after
independent evaluation of the new algorithm, the U.S. National Oceanic and Atmospheric
Administration (NOAA) also made similar changes in their processing code to apply the new algorithm
approach (after further tuning of algorithm coefcients) for VIIRS global processing (Wang & Son, 2016).
This is the rst time since the CZCS era that the algorithm concept has been changed in global data proces-
sing from band ratio to band difference for lowChl waters. Along with extensive evaluations by the commu-
nity (e.g., Brewin et al., 2013, 2015) and by NASA, the reason behind these changes is mainly because of the
tolerance of the CI design to known and unknown errors from atmospheric correction and other sources;
such a tolerance leads to reductions in many image artifacts. For example, in addition to the documented
error tolerance in Hu et al. (2012), NASA found that the CI algorithm was also tolerant to an unknown error
of digital count round off in one of the atmospheric correction bands due to sensor degradation (Franz,
2014). Indeed, the CI algorithm led to (1) more spatially coherent and temporally consistent Chl patterns,
(2) better Chl retrievals than OCx as gauged by eld validations, and (3) more consistent crosssensor time
series for lowChl waters. Consequently, NASA has applied the OCI algorithm as the default Chl algorithm
for all ocean color sensors currently being used at NASA, including CZCS, SeaWiFS, MODIS, MERIS, VIIRS,
OCTS (Ocean Color and Temperature Sensor; 19961997), and GOCI (Geostationary Ocean Color Imager;
2011 to present). This led to a more consistent multisensor time series as shown in the overlapping periods
for more than one sensor (Franz, 2014; Franz et al., 2016). Furthermore, inspired by the algorithm design,
similar banddifference approaches have been developed to estimate surface concentrations of particulate
Journal of Geophysical Research: Oceans
HU ET AL. 1525
inorganic carbon (Mitchell et al., 2017) and particulate organic carbon (Le et al., 2018), both showing
improved performance over the original algorithms.
However, from a practical point of view, several weaknesses still exist in the NASA default global Chl data
products. Some of these originate from the algorithm itself while others are due to data processing quality
control when generating global products. Specically,
1. The original CI algorithm was tuned using highperformance liquid chromatography (HPLC) Chl data
only. As a result, there are much fewer data points in the CI algorithm training than in the original
OCx where both HPLC and uorometric Chl data were used. The difference between CI algorithm tun-
ing using different data sets needs to be evaluated;
2. During evaluation of the OCI algorithm using global data it was found that the threshold of 0.25 mg m
originally proposed by Hu et al. (2012) to transition between CI and OCx could lead to some discontinuity
in global data histogram distributions for intermediate Chl (0.250.3 mg m
). Therefore, the OCI imple-
mentation lowered this threshold to 0.175 mg m
. Consequently, the full advantage of the CI design was
compromised because for Chl between 0.175 and 0.25 mg m
the bandratio OCx algorithm still played a
role but OCx is less tolerant than CI to input errors. Because 15% of the global oceans have Chl between
0.175 and 0.25 mg m
according to satellite data statistics, it is desirable to extend CI to its original upper
bound of 0.25 mg m
to take its full advantage;
3. A major advantage of remote sensing is large data quantity (i.e., frequent and synoptic measurements).
Current quality control criteria within data processing, however, sometimes make valid retrievals dif-
cult even under cloudfree conditions when brighttarget adjacent stray light, Sun glint, large solar or
viewing angle, or other nonoptimal observing conditions exist (Feng & Hu, 2016a). Because the CI algo-
rithm has been shown to be tolerant to many residual correction errors in the algorithm inputs, it might
be possible to relax some of the quality control ags to recoversome of the previously masked data,
leading to more valid retrievals without diminishing data quality. Indeed, the problem of no valid retrie-
vals under cloudfree conditions not only reduces coverage, increases uncertainties (due to fewer data
points used in calculating the mean), but also presents an obstacle for realtime guidance of ship surveys
or monitoring of blooms. Therefore, this problem is briey reviewed below.
Cloud statistics from MODIS measurements showed that, on average, global ocean cloud fraction is between
70% and 75% (King et al., 2013). MODISA has neardaily coverage over subtropical and tropical oceans and
daily or more frequent coverage over highlatitude regions. Assuming daily coverage, on average at a given
location there should be one cloudfree measurement every three to four days (i.e., about 2530% chance).
However, after rigorous quality control with the various quality ags applied, global statistics indicated that
valid retrievals of Chl were only 5% (Feng & Hu, 2016a), meaning that for a random 1km ocean location
(before spatial or temporal data binning) there was only one valid retrieval every 20 days. In contrast, valid
retrievals for sea surface temperature from the same MODIS measurements nearly doubled to 10%. Clearly,
cloud is not the only major factor affecting the quantity of valid retrievals. Others factors such as Sun glint,
cloudadjacent stray light, large solar and/or view angles, and swath width also play important roles. Of
these, the straylight ag appeared to be one of the dominant factors.
Stray light near bright targets, such as clouds, is caused by light scattered by the target entering the eld of
view of the satellite instrument through the sensor's point spread function (Meister & McClain, 2010). In
practice, the MODIS straylight ag was dened as a 7 × 5 pixel dilation from any cloud pixel, where 7 is
in the crosstrack direction and 5 is in the alongtrack direction (Franz et al., 2005). In other words, a
cloudfree pixel within 3 pixels in crosstrack direction or 2 pixels in alongtrack direction from a nearby
cloud pixel was considered as being contaminated by clouds and therefore discarded in global data compo-
sites. MODIS statistics near clouds indicated that this was perhaps too conservative in masking lowquality
cloudfree data, leading to a signicant loss of data quantity. Feng and Hu (2016b) showed that although
MODIS total radiance even 10 pixels away from clouds may still have signicant stray light effect (i.e.,
>1% in total radiance), some of these effects were mitigated through atmospheric correction because the
atmospheric correction bands also suffered from the same adjacency effects with the same sign, leading to
only 12 pixels adjacent to clouds being contaminated by stray light in the retrieved R
(λ) values. Based
on these results, Feng and Hu (2016b) argued that the 7 × 5 straylight ag might be relaxed to increase data
quantity without reducing data quality. However, to what extent the ag can be relaxed is unknown. A thor-
ough evaluation using eld data and more statistics is required.
Journal of Geophysical Research: Oceans
HU ET AL. 1526
Therefore, given the potential weakness in the OCI algorithm and a desire to increase data quantity, the
objective of this paper is to revisit the OCI algorithm design with its parameterization as well as to improve
both data quality and data quantity for global oceans.
2. Data and Methods
As in Hu et al. (2012), the focus is on a unied algorithm that can be applied to all ocean color sensors rather
than on sensorspecic tuning of parameterization to account for crosssensor difference in sensor band set-
tings. Such a difference can be corrected through converting the sensorspecicR
(λ) to sensorindependent
(λ) for common, unied wavelengths using hyperspectral R
(λ) data collected from in situ measure-
ments. For this reason, the three wavelengths used in this work are 443, 555, and 670 nm as a heritage
of SeaWiFS.
In situ R
(λ) and Chl data from the NASA bioOptical Marine Algorithm Data set (NOMAD) version 2 were
used to rene the Chl algorithm (Werdell & Bailey, 2005). R
(λ) and Chl data were collected by many
research groups around the world, thus forming data pairs for algorithm development. To increase data
volume, Chl data determined using both HPLC and uorometric methods were used here. If both measure-
ments were conducted at a station, HPLC Chl was used. Several additional qualitycontrol criteria were
applied to the NOMAD data set: R
(λ) > 0.0 sr
, Chl > 0.0 mg m
, and latitude between 60°N and
60°S. A total of 2,306 data pairs were selected.
CI for each data pair was calculated using R
, and R
as in Hu et al. (2012):
CI ¼Rrs;555Rrs;443 þ555443ðÞ=670443ðÞxR
In traditional empirical algorithm design, blue/green band ratios have been used to estimate Chl because for
most oceanic waters, R
in blue bands decreases with increasing Chl but R
in green bands tends to be rela-
tively stable. Here the maximum ratio, R, between four wavelengths of SeaWiFS was estimated as (O'Reilly
et al., 2000)
R¼max Rrs;443;Rrs;490;Rrs;510
=Rrs;555 (2)
In the above equation, the maximum of R
is selected from three blue bands (443, 490, and 510 nm) to
improve signaltonoise when the water type changes from clear to more turbid.
Similarly, after the launch of MODIS and VIIRS sensors the maximum ratio Rbetween three wavelengths
was also estimated and used in the same fashion in the algorithm development.
Following the NASA approach to derive the OCx algorithm coefcients, Chl data were rst gridded into
logarithmic space, with the corresponding mean CI and mean Rcalculated for each grid as described below.
Then, for consistency with the NASA approach, such gridded data were gridded again according to Rin log
space, with the corresponding mean Chl and mean CI calculated and used in the regression to derive the CI
algorithm coefcients using the following algorithm formulation:
Log10 ChlðÞ¼aCI þb(3)
where a and b are the CI algorithm coefcients. In practice, the incremental steps in the gridded Chl space
were calculated as
L;Chl;i¼0:01x1:01 ;H;Chl;i¼0:01xiþ1ðÞ
where L
and H
are the lower and upper bounds for the i
Chl bin. The form in equation (4) is to
account for lognormal Chl data distributions (Campbell, 1995). For example, the range of the 200th Chl a
bin is 0.07320.0739 (i.e., 0.01 × 200
0.01 × 201
). For all data points whose Chl values fall within a
grid, their mean Chl, mean CI, and mean Rvalues were calculated (denoted as Chl,CI, and R, respec-
tively), resulting in 927 valid bins. These bins were then gridded into the Rspace again, where the incremen-
tal steps in the gridded R space were calculated as
Journal of Geophysical Research: Oceans
HU ET AL. 1527
1:01;H;R;i¼0:05 xiþ1ðÞ
where L
and H
are the lower and upper bounds for the i
R bin. The
form in equation (5) is to account for the fact that Ris not used linearly in
the OCx algorithm design, but log
(R) is used (equation (6)). For all 927
points whose Rvalues fall within a grid, the mean Chl and mean CI were
calculated (denoted as Chland CI, respectively), resulting in 245 nal
Of the 245 pairs of Chland CI, only 116 pairs with CI< 0.0001 (corre-
sponding to Chl~0.4 mg m
) were selected to determine the algorithm
coefcients using equation (3) (Figure 1).
The CI algorithm is only applicable to lowChl waters. For higherChl
waters, the original OCx algorithm is used. For SeaWiFS, the fourband
algorithm is specied as (O'Reilly et al., 2000)
y¼a0þa1χþa2χ2þa3χ3þa4χ4;χ¼log10 RðÞ (6)
where a
are the algorithm coefcients determined through nonlinear
regression using gridded Chl and Rdata as described above. The current
algorithm coefcients used operationally by NASA (version 6) are 0.3272, 2.9940, 2.7218, 1.2259, and
0.5683, respectively. The algorithm forms and algorithm coefcients for MODIS and VIIRS are determined
similarly. All algorithm forms and coefcients can be found from the NASA algorithm webpage (https://
Because the CI was designed for lowChl waters only, the global OCI algorithm is a hybrid between CI and
OCx, which is formulated as
ChlOCI ¼ChlCI for ChlCI ChlLmg m3
ChlOCX for ChlCI >ChlHmg m3
αxChlOCX þβx ChlCI for ChlL<ChlCI ChlHmg m3
where the weighting factors are α= (Chl
), β= (Chl
). Chl
and Chl
dene the lower and upper bounds of the algorithm transition zone. In Hu et al. (2012), they were
0.25 and 0.30 mg m
, respectively. In the NASA operational processing the lower bound was adjusted to
0.175 mg m
. In this study they are determined to be 0.25 and 0.40 mg m
to make the CI design applicable
for a higher range while maintaining a smooth transition between the CI and OCx algorithms (see below).
On average, SeaWiFS monthly data between 1998 and 2010 showed that 77.8 ± 1.0% of the global ocean had
Chl 0.25 mg m
and 5.1 ± 0.4% of the global ocean had Chl between 0.25 and 0.3 mg m
Because there are several algorithms used in this paper, for clarity the terminology is dened in Table 1.
Algorithm evaluation was performed using several ways. First, in situ data archived in the NASA SeaWiFS
Biooptical Archive and Storage System (SeaBASS; Werdell et al., 2003) were used to evaluate the Chl data
products after applications of the Chl algorithms. The following search criteria, most of which recommended
Figure 1. Scatterplot to show the CI2 algorithm. For comparison, the origi-
nal CI algorithm (CI1) is also presented. The red circle highlights the dif-
ference between CI2 and CI1 for lowChl waters.
Table 1
Algorithm Terminology Used in This Paper
Algorithm Name Meaning
CI1 Original CI algorithm with transition zone of 0.250.30 mg m
(Hu et al., 2012)
CI1Same as CI1, but the transition zone is adjusted to 0.250.40 mg
CI2 Adjusted algorithm with new parameterization, with transition zone of 0.250.40 mg m
OCI1 Original OCI algorithm as a hybrid between CI1 and OCx
OCI1Adjusted algorithm as a hybrid between CI1and OCx
OCI2 Adjusted algorithm as a hybrid between CI2 and OCx
Journal of Geophysical Research: Oceans
HU ET AL. 1528
by NASA, were used to nd the satellitein situ matching pairs: bathymetry >30 m, sensor zenith angle <56°,
solar zenith angle <70°, <3hr time difference between in situ and satellite measurements, median value of
coefcients of variation (calculated as standard deviation divided by mean) of several products (R
412 and 555 nm, aerosol optical thickness at 865 nm) <15% for the 5 × 5pixel window centered at the in situ
station, and difference between simulated and measured surface irradiance <100%; >50% pixels in the 5 × 5
box must be valid (i.e., not associated with any of the standard qualitycontrol ags such as straylight and
high glint; Bailey & Werdell, 2006). A total of 1,424 matching pairs were selected for SeaWiFS between
1998 and 2010, and 330 were obtained for MODIS Aqua between 2002 and 2010. These are basically the same
data sets as used in Hu et al. (2012).
Second, in addition to the approach above for product evaluation, additional effort was used to nd in situ
data collected under MODIS straylight mask. To achieve this, all concurrent and collocated MODIS and in
situ matching pairs were downloaded from NASA without applying any criteria above except that the time
window was relaxed from ±3 to ±4 hr. The same was performed to nd the corresponding SeaWiFS les. A
total of 3,728 MODIS level2les and 5,763 SeaWiFS MLAC level2les were downloaded for this purpose.
Then, local processing (instead of using the SeaBASS search engine) was used to apply all but the straylight
ag to determine which in situ data points were associated MODIS straylight ag. Unfortunately, there were
only <5 points qualied for such criteria, making it impossible to evaluate MODIS retrievals under the 7 × 5
straylight ag using in situ measurements. Thus, statistical analysis using satellite data alone is the only
approach to determine whether these 7 × 5 agged pixels are statistically different from the adjacent,
nonagged pixels.
3. Results
3.1. Rened OCI Algorithm (OCI2) and Its Field Validation
Figure 1 shows the new CI algorithm (CI2) based on all qualied NOMAD Chl data and R
(λ) data after
gridding in log space. For comparison, the original CI algorithm (hereafter CI1; Hu et al., 2012) is also
shown. Although they are similar between Chl of 0.10.2 mg m
, there are noticeable differences for
Chl < 0.05 mg m
(red circle) and for Chl > 0.3 mg m
. The differences are due to their different inputs
of Chl and R
(λ) data because CI1 used the limited HPLC data only while CI2 used all data including
HPLC and uorometric Chl. From a pure statistical point of view, with the limited data, however, it is dif-
cult to conclude which one is superior. This is especially true for Chl < 0.05 mg m
where data spread
around the tting line is signicantly higher than for other Chl values. Also note that for the same CI values
(i.e., same input R
(λ) data), Chl
is lower than Chl
for Chl < 0.05 mg m
. This will have signicant
impact on studies of ocean gyres where Chl is extremely low.
The OCI2 algorithm was applied to both SeaWiFS and MODISA data where concurrent eldmeasured Chl
data were available through NASA's SeaBASS archive. Such estimated Chl
were compared with eld
measured Chl in Figure 2 using the search criteria described above. For comparison, Chl data derived from
SeaWiFS and MODISA using the OCx algorithm are also presented in Figure 2, while statistical measures of
the algorithm performances for lowChl waters are all listed in Table 2.
The evaluation results show that compared with the OCx algorithm, OCI2 performance is signicantly bet-
ter in nearly all statistical measures for lowChl waters. Hu et al. (2012) showed that even if the OCx algo-
rithm coefcients were retuned using low Chl data (<0.4 mg m
) only, the CI algorithm design still led
to improved algorithm performance than OCx because of the former's tolerance to input R
(λ) errors (admit-
tedly that all in situ R
(λ) data still contain a few percent errors or uncertainties even after rigorous quality
control). Compared to OCI1, OCI2 performance measures are mixed when these discrete in situ measure-
ments were used to evaluate product uncertainties, but they are generally comparable to those of OCI1.
One obvious and also previously known nding is the signicant data spread for almost the entire data
range, highlighting the need for improved data quality in eld measurements. Most importantly, data are
extremely scarce from ocean gyres where Chl is low. For MODISA, there is not a single point with in situ
Chl < 0.03 mg m
. For SeaWiFS, there is not a single point with in situ Chl < 0.02 mg m
, and data for
Chl < 0.1 mg m
appear to be more spread than in other ranges. Clearly, the puzzle of which one is closer
Journal of Geophysical Research: Oceans
HU ET AL. 1529
to the truthfor lowChl waters, Chl
or Chl
in Figure 1, cannot be resolved from evaluations using the
current in situ data sets. Other measures, as shown below, may be used to determine which one is better.
3.2. OCI2 Performance by Other Measures
Although scatterplots have been widely used for algorithm evaluation, they only provide one side of the story
because (1) eld data are always limited, and they may not cover the full range under all possible measure-
ment scenarios, and (2) more importantly, potential artifacts in image quality and crosssensor inconsistency
cannot be revealed by scatterplots. Therefore, data statistics, image quality, and crosssensor consistency are
examined in order to further evaluate the OCI2 performance.
The rst check is on Chl continuity in the transition zone, which is 0.250.3 mg m
for OCI1 but 0.250.4
for OCI2. For an appletoapple comparison, an adjusted OCI1 (i.e., OCI1in Table 1) with the same transi-
tion zone of 0.250.4 mg m
was tested as well. This is because if a smoother transition zone is achieved it is
easier to verify whether it is due to algorithm change from CI1 to CI2 or due to transition zone change from
0.250.3 to 0.250.4 mg m
. Here the transition thresholds of 0.25 and 0.4 mg m
correspond to the clear
water denition given by Gordon and Clark (1981) and the upper limit of the CI algorithm applicability
(Figure 1), respectively. The comparison between these three schemes is shown in Figure 3. The discontinu-
ity in OCI1 largely disappeared in OCI1and OCI2, with the latter two showing slightly different histogram
distributions due to their different algorithm coefcients (Figure 1). Overall, OCI2 appears to have signi-
cantly reduced the discontinuity in Chl distribution across the transition zone for all three sensors examined.
Next, the three OCI algorithms are compared against the default OCx algorithm for the same images as used
in Figure 3, with results shown in Figure 4. Compared to OCI1, both OCI1(black dots) and OCI2 (red dots)
Figure 2. Evaluation of the OCI2 algorithm with SeaBASS data. For comparison, performances of the OCx algorithms are
also presented. The inset gures show the locations of the satellitein situ matching pairs. The statistical measures for low
Chl waters can be found in Table 2.
Table 2
Performance of the Updated Chl Algorithm (CI2) for Both MODISA and SeaWiFS for Chl < 0.25 mg m
, as Gauged by
FieldMeasured Chl From NASA's SeaBASS Data Archive
RMSE URMSE Mean Ratio Median Ratio MRE R
(Log) N
MODISA OCx 77.7% 44.2% 1.24 1.05 32.0% 0.42 0.66 63
CI1 43.9% 32.7% 1.15 1.04 25.4% 0.62 0.71 63
CI2 51.2% 37.6% 1.12 0.94 35.2% 0.59 0.71 63
SeaWiFS OCx 535.8% 54.2% 1.79 1.19 41.5% 0.01 0.33 357
CI1 91.8% 47.2% 1.40 1.16 36.8% 0.31 0.39 357
CI2 102.0% 49.6% 1.38 1.14 39.4% 0.28 0.39 357
Note. For comparison, performances of the original CI1 (Hu et al., 2012) and the OCx algorithms are also listed.
Denitions and meanings of the statistical terms can be found in Hu et al. (2012). RMSE: rootmeansquare error,
URMSE: unbiased RMSE, MRE: mean relative error.
Journal of Geophysical Research: Oceans
HU ET AL. 1530
showed smoother transition (i.e., no apparent bump) between low and
higher Chl ranges. However, compared to OCI1, OCI2 is closer to OCx
around the 1:1 line for Chl > 0.2 mg m
and especially for
Chl > 0.4 mg m
, suggesting that OCI2 is likely better in algorithm con-
sistency when considering that OCx is actually the algorithm used for
Chl > 0.4 mg m
. Subtle but noticeable difference was found for
Chl > 0.2 mg m
, which does not indicate poor OCI algorithm perfor-
mance because by algorithm design CI is different from OCx and this dif-
ference can also show up in the transition zone. Because in the transition
zone the weighted mixing (equation (7)) is based on Chl
, OCI2 and OCx
do not merge exactly on the 1:1 line even after Chl > 0.4 mg m
Statistically it is difcult to determine which algorithm is closer to the
truth due to signicant data spread in this range and due to the small dif-
ference between CI1 and CI2 (Figure 1). However, an algorithm should
have smooth transition between lowChl and higherChl waters. In this
regard, OCI2 appears to have better performance than either OCI1
(0.250.3) or OCI1 (0.250.4; i.e., OCI1).
The effect of OCI2 on image quality is signicant when the OCx image is
used as a reference to compare with (Figure 5). This is similar to the OCI1
effect shown in Hu et al. (2012). Most of the noise due to cloudadjacent
stray light contamination, shown clearly in the OCx image, disappeared
in both OCI1 and OCI2 images. This effect is what the algorithm concept
was designed for (Hu et al., 2012). For this gyre location OCI2 showed
lower Chl than OCI1, which was also closer to OCx (not necessarily more
accurate, however), a result of adjustment of the algorithm
coefcients (Figure 1).
The OCI2 and OCx algorithms are further compared in their global
monthly MODIS Chl maps in Figure 6. Visually, they appear similar in
all major spatial patterns as well as Chl magnitudes except that OCI2
Chl maps are more spatially coherent because most of the noiseinduced
errors for lowChl waters have been removed in the band difference
design (Hu et al., 2012; such a contrast can only be visualized in the
fullsize versions of the images). However, their relative difference images
clearly show disparity across major ocean basins. For example, during
winter months (January and March), Chl
is mostly lower than
for North Atlantic and North Pacic, but during the month of
June the opposite is observed. Similar to those from the scatterplot evalua-
tions, however, it is difcult to conclude from these results alone which
algorithm is closer to the truth.
The most signicant improvement of the CI design is on crosssensor con-
sistency, as shown in Hu et al. (2012). Such an improvement is illustrated
again in Figure 7, where histograms of monthly Chl distributions for oli-
gotrophic oceans (dened by SeaWiFS mission climatology
Chl < 0.1 mg m
), derived from OCx, OCI1, and OCI2 algorithms, are
compared between SeaWiFS and MODISA and between VIIRS and
MODISA. The distributions from OCx showed large crosssensor differences that are difcult to observe
from image inspection or scatterplot evaluations. Such differences are largely removed in both OCI1 and
OCI2, where they show nearly identical crosssensor distributions.
The improvement in crosssensor consistency is further shown in Figure 8, where global mean Chl of oli-
gotrophic oceans (dened by SeaWiFS mission climatology Chl < 0.1 mg m
) are compared between
SeaWiFS and MODISA and between VIIRS and MODISA using their ratios. A perfect crosssensor consis-
tency would result in a ratio of 1.0 for all months. The OCx algorithms are associated with relatively large
Figure 3. Improvements of OCI2 over OCI1 in the algorithm transition
zone of 0.250.4 mg m
for all three sensors. Global level3 daily R
composites in 2010 (365 daily composites) were used to generate the histo-
gram distributions for MODISA and SeaWiFS, and 365 global daily level3
composites in 2014 were used for VIIRS. The wavelengths of MODIS and
VIIRS green bands (547 and 551 nm, respectively) were adjusted to 555 nm
using empirical relationships between these bands derived from NOMAD
Journal of Geophysical Research: Oceans
HU ET AL. 1531
crosssensor discrepancies (about 10%), while both OCI1 and OCI2 reduced such discrepancies
signicantly to 12%.
In summary, similar to the original OCI1, the OCI2 algorithm showed improvements over the OCx algo-
rithms in production of equivalent or better Chl values for lowChl waters, but more importantly, far
improved image quality and crosssensor consistency. Furthermore, it reduced the discontinuity in the algo-
rithm transition zone. Therefore, future data processing may adopt OCI2 as an update of OCI1.
3.3. Data Recovery
The MODIS default 7 × 5 straylight mask was changed to 0 × 0 (i.e., no straylight), 3 × 3, 5 × 3, and 9 × 7.
Because in situ data could not be found from the entire SeaBASS archive corresponding to these masked pix-
els, statistics of image data alone was used to evaluate their potential differences.
Figure 9 shows the Chl histogram distributions from pixels with different masking schemes. The evaluation
was performed from pixels of 665 MODISA imagesfor all ringsnear cloud edges. The rst ring includes the
cloudfree pixels with 0distance to cloud edge, that is, those between the 0 × 0 and 3 × 3 masks. The second
ring includes cloudfree pixels between the 3 × 3 and 5 × 3 masks, and so on. Figure 9a shows that except for
the rst ring (black line), all other rings have similar histogram distributions, especially when considering
that they are comparable to those pixels traditionally considered as being not inuenced by stray light (i.e.,
between 7 × 5 and 9 × 7). This is not a surprise, as the tolerance of the OCI design to input errors has already
been demonstrated in Hu et al. (2012). What is surprising is that for the 665 evaluated MODISA images, OCx
also appears to be tolerant to stray light after a 3 × 3 masking (Figure 9b). Indeed, other than the rst ring (i.e.,
3 × 3 masked pixels), allother rings show similar histogram distributions. These results indicate that the 7 × 5
straylight mask might be relaxed to 3 × 3 in order to increase data quantity without losing data quality.
The statistics on Chl distributions were from 665 MODISA images over clear waters in the North Atlantic
gyre. To evaluate whether the same observations can also be made to R
(λ), the same statistical analysis
was performed for the 665 MODISA images, but with the focus on R
(λ) instead of Chl. Our rationale is that
conclusions drawn from the R
(λ) statistics over clear waters may be extended globally (including
coastal waters).
Figure 10 shows that similar to the Chl distributions, except for the pixels immediately adjacent to cloud
edges (i.e., 3 × 3 masking), all other rings have similar histogram distributions, meaning that statistically pix-
els in the ring between 3 × 3 and 7 × 5 are not different from those between 7 × 5 and 9 × 5 masking. The
importance of this nding is that because R
(λ) serves as the inputs for nearly all biooptical inversion algo-
rithm (including Chl algorithms), the straylight ag may be relaxed from the current 7 × 5 to the new 3 × 3
pixels for possibly all ocean color data products including Chl, of course pending further evaluations on spe-
cic data products. Indeed, results from the same analysis over all MODISA images over the Gulf of Mexico
Figure 4. Comparison between OCx (xaxis) and OCI1 and OCI2 (yaxis) for three ocean color sensors. The gray dashed line of each panel indicates the 1:1 line.
The results were generated using the entire global level3 daily R
data sets in 2010 for MODIS and SeaWiFS and in 2014 for VIIRS, respectively. Here OCII
(0.250.4) is termed as OCI1in Table 1. Note that OCI and OCx do not merge exactly on the 1:1 line even after the upper bound of the transition zone because the
weighted mixing in equation (7) is based on Chl
Journal of Geophysical Research: Oceans
HU ET AL. 1532
Figure 5. Comparison between Chl images (approximately 1,000 km wide) derived from three algorithms for MODIS data
collected on 30 August 2004 over the South Pacic gyre (subscene from the inset gure).
Journal of Geophysical Research: Oceans
HU ET AL. 1533
during 2010 indicate that even when pixels are restricted to waters with Chl > 0.4 mg m
, pixels in every
ring (except the 3 × 3 ring) around cloud edge have nearidentical histogram distributions in either spectral
(λ) or Chl
(Figure 11), suggesting that a 3 × 3 straylight mask is sufcient even for productive waters.
Based on these results, the relaxed 3 × 3 straylight ag was applied to global level2 OCI2 data for March and
July 2005 to generate new global mean Chl
maps with a 9km resolution grid, which were compared
with the current (default) Chl
maps. The results are shown in Figure 12. Some of the data gaps in the
default Chl
maps are lled in the new maps after relaxing the straylight mask from 7 × 5 to 3 × 3 pixels.
These are particularly apparent in the equatorial Pacic and Atlantic within the Intertropical Convergence
Zone, the North Atlantic, the North Indian Ocean, and the Southern Ocean (all circled in red in Figure 12).
Figures 12e and 12f show the comparison between the monthly mean Chl
derived from the default 7 × 5
masking and the new 3 × 3 masking. To reduce pepper noise from individual 9km pixels, spatial binning
was performed rst. Although the changes in Chl
between the two masking schemes are relatively small
(within a few percent), there are some spatially coherent patterns, for example, slightly lower Chl
in the
Southern Ocean and waters off Peru (in July 2005) from the new masking scheme.
The increased coverage and different monthly mean Chl from the relaxed straylight masking are all due to
the increased number of valid observations in every 9km grid cell. Figure 13 shows the comparison between
number of valid level2 pixels used to calculate the monthly mean from the 7 × 5 and 3 × 3 masking schemes.
Globally, the relative increases in the number of valid level2 pixels are ~39% for both months, but the
increases can be more signicant in many regions, often reaching >100% as indicated by the green and
colder colors in Figures 13e and 13f. In the North Indian Ocean, the increases reached 400% in July 2005
(i.e., the ratio is ~ <0.2 in the maps).
4. Discussion
4.1. Need for More Field Data From Ocean Gyres
Each time satellite data are reprocessed with new calibration or new algorithm the resulting data products
are changed, and most times with improvements. To date, data from all sensors have been reprocessed sev-
eral times by NASA, with the OCI1 algorithm incorporated during the most recent reprocessing (2014.0 and
Figure 6. Comparison of global monthly Chl maps derived from (left column) OCx and (middle column) OCI2 for different months of 2010, which were derived
using daily level3R
composites. The right column shows their relative differences.
Journal of Geophysical Research: Oceans
HU ET AL. 1534
2018.0) for its advantages over the traditional OCx algorithms. The results shown here demonstrate that
OCI2 may be a better choice. This is because that while its absolute accuracy is similar to OCI1 as gauged
by the in situ validation statistical measures and its crosssensor consistency is also comparable to OCI1,
its applicability range is extended to 0.4 mg m
with smoother algorithm transition between 0.25 and
0.4 mg m
than OCI1.
One large difference between OCI2 and OIC1 is their Chl retrievals for low concentration waters
(<0.05 mg m
), mostly over ocean gyres. These account for about 14.7 ± 2% of the global ocean area accord-
ing to SeaWiFS monthly statistics between 1998 and 2010. From a statistical point of view, however, it is
Figure 7. Comparison of Chl distributions from global oligotrophic oceans (geographic extent dened by SeaWiFS Chl climatology 0.1 mg m
, where actual Chl
during a specic month can be 0.4 mg m
(corresponding to log10(Chl) = 0.4) or higher) obtained from three algorithms. The comparison was made (left col-
umn) between SeaWiFS and MODISA and (right column) between VIIRS and MODISA.
Journal of Geophysical Research: Oceans
HU ET AL. 1535
difcult to know which algorithm yielded better (i.e., more accurate) results due to the extreme data scarcity
from eld measurements and the general uncertainties associated with collecting in situ Chl in these very
clear waters. The signicant data spread for Chl < 0.05 mg m
in Figure 1 also suggest possible errors in
the eldmeasured Chl or R
(λ) because biooptical properties over ocean gyres are expected to be very
stable between Chl and R
(λ). The reason for data scarcity is possibly due to the remote locations, while
the reason for erroneous data may possibly be due to difculty in obtaining reliable in situ R
(λ) data and
due to the requirement of large volume of waters to be ltered to obtain a measurable signal from these
extremely lowChl waters. Indeed, even after applying additional quality control to the NOMAD R
data using a recently developed scoring system (Wei et al., 2016), it is still impossible to determine which
Figure 8. Mean Chl ratios over global oligotrophic oceans (geographic extent dened by SeaWiFS Chl climatology
0.1 mg m
, where actual Chl during a specic month can be 0.4 mg m
(corresponding to log10(Chl) = 0.4) or
higher) between different sensors using three different algorithms. (a) MODISA/SeaWiFS. (b) VIIRS/MODISA. In (a), the
longterm mean monthly ratios are 0.91 ± 0.04, 0.99 ± 0.03, and 1.00 ± 0.03 (N= 102) for OCx, OCI1, and OCI2,
respectively. In (b), the longterm mean ratios are 0.89 ± 0.04, 0.98 ± 0.04, and 0.98 ± 0.05 (N= 50) for OCx, OCI1, and
OCI2, respectively.
Figure 9. Histograms of (a) Chl
and (b) Chl
from pixels near cloud edges after applying different masking meth-
ods. The statistics were derived from 665 MODISA images in 2010 over the North Atlantic gyre (60°40°W, 15°35°N).
Note that except for pixels in the rst ring(i.e., within 3 × 3 pixels of cloud edge), pixels from all other rings, including
those outside the default 7 × 5 straylight mask, have similar histogram distributions.
Journal of Geophysical Research: Oceans
HU ET AL. 1536
measurements from ocean gyres are more trustable than others, leading to no difference in the statistical
regressions. Therefore, more highquality Chl and R
(λ) data need to be collected from ocean gyres,
especially from waters with Chl < 0.05 (or even <0.03 mg m
), in order to determine the lower bound of
Chl in algorithm development. Currently, given the data scarcity and the slight difference between OCI1
and OCI2 for extremely clear waters, it is difcult to tell which algorithm is closer to the truth for these
waters and what is the lowest possible Chl in global natural waters. However, some indirect inference
Figure 10. Histograms of (a) R
, (b) R
, and (c) R
from pixels near cloud edges after applying different
masking methods. The statistics were derived from 665 MODISA images in 2010 over the North Atlantic gyre (60°40°W,
Figure 11. Histograms of (a) R
, (b) R
, (c) R
, and (d) Chl
from pixels near cloud edges after applying
different masking methods. The statistics were derived from all MODISA images in 2010 over the Gulf of Mexico
(98°80°W, 18°31°N) for Chl > 0.4 mg m
Journal of Geophysical Research: Oceans
HU ET AL. 1537
suggests that 0.01 mg m
might be more reasonable than 0.02 mg m
. For example, the maximum Secchi
depth measured in the South Pacic gyre was ~70 m (Doron et al., 2011). Using the Case1biooptical
model (Morel & Maritorena, 2001) and the new Secchi depth theory (Lee et al., 2015), this value more
likely corresponds to Chl value of ~0.01 mg m
rather than 0.02 mg m
. However, without more eld
measurements of highquality Chl, this can only be used as indirect evidence to support OCI2. Indeed, the
lowest Chl value in the algorithm development data set (NOMAD) is 0.012 mg m
, indicating the need
of more data around this value.
On the other hand, even without resolving the puzzle of which one is closer to the truth for lowChl waters,
OCI1 or OCI2, they both can be used to study changes induced by climate variability. This is because each
algorithm is selfconsistent, therefore even if the absolute Chl value may be slightly off the changes detected
by the algorithm are still trustable. In this regard, algorithm consistency is more important than algorithm
accuracy. This is exactly why the improved crosssensor consistency from either OCI1 or OCI2 (over OCx) is
important in order to form a seamless, multisensor climate data record, which can then be used to assess
longterm ocean changes in response to climate variability (e.g., Antoine et al., 2005; Gregg et al., 2005;
Polovina et al., 2008; Signorini et al., 2015).
4.2. Data Quality, Quantity, and Uncertainty
In addition to Chl algorithm renement, this work also proposed a practical scheme to recover some of the
previously masked cloudfree data through relaxing the cloudadjacent straylight mask. Although these data
exist at level2, they are discarded in global level3 composites using a 7 × 5 mask. Our results showed that
Figure 12. (a and b) Level3 monthly global Chl
(mg m
) at each 9km grid for March and July 2005, respectively, when the default 7 × 5 straylight masking
scheme was applied together with other quality controls to the individual level2les to derive the global composites. (c and d) Same as in (a) and (b) but
relaxed straylight masking (3 × 3) was applied together with other quality controls. The red eclipses highlight regions where considerable amount of missing data
from the former have been recovered in the latter. (e and f) Their relative (%) differences (relaxed/default 1.0) are shown, respectively. To remove pepper noise,
the images were spatially binned before calculating the difference, leading to the apparently reduced data gaps than shown in the defaultag maps.
Journal of Geophysical Research: Oceans
HU ET AL. 1538
such a mask was probably too conservative, and a reduction to a 3 × 3 mask would recover many pixels with
similar quality to those nearby pixels. Globally, these recovered pixels represent about 40% of the previous
valid retrievals, thus indicating a signicant increase in data quantity. In some regions during certain
seasons, the increases could reach several hundred percent.
Such increases have profound impacts on both data availability at daily scale and data product uncertainties
at longer (e.g., weekly, monthly, seasonal) scales. First, most global or regional studies have used monthly
mean products to study spatial/temporal changes. Because product uncertainties induced by independent
and normally distributed errors are inversely proportional to square root of number of observations, the
increased number of valid retrievals will lead to reduced data product uncertainties. This type of uncertainty
reduction has been demonstrated in Qi et al. (2017) using global ocean color data products. For a 39%
increase, uncertainties will be reduced by 15% (=1 1/sqrt(1.39)). For some regions, the reduction can be
much higher. For regions with signicant and persistent cloud cover (e.g., Intertropical Convergence
Zone, North Indian Ocean during summer), the new masking scheme may ll some of the monthly data
gaps (i.e., from 0 observation to at least 1 observation) to lead to more spatially coherent patterns. For the
same reason, when applied to level2les or shortterm (threeday, sevenday) image composites, the new
scheme would provide more spatial coverage to guide eld measurements in near real time.
The increases from the new masking scheme are in relative terms. Overall the number of valid retrievals is
very low (~5% for both MODISA and MODIST; Feng & Hu, 2016a). After a 39% increase this becomes ~7%.
Clearly, future studies should focus on whether there is additional room for further increases of data quan-
tity without losing quality, for example, through relaxing the Sun glint masking threshold, through new
Figure 13. (a and b) Number of valid MODISA Chl observations at each 9km grid for March and July 2005, respectively, when the default 7 × 5 straylight masking
scheme was applied together with other quality controls to the individual level2les to derive the global composites. A number of 200 indicates that
during the month there are 200 valid 1km data points in the 9km grid. (c and d) Same as in (a) and (b) but relaxed straylight masking (3 × 3) was applied together
with other quality controls. (e and f) Ratios between (a) and (c) and between (b) and (d), respectively. Several regions of low ratios (<0.2) are outlined, meaning
that the relative increase in valid observations can reach >400% if the relaxed masking scheme is used. Overall the global average increase is about 40%.
Journal of Geophysical Research: Oceans
HU ET AL. 1539
algorithms to dene straylight pixels (e.g., Jiang & Wang, 2013), or through machine learning using the
spectra of lowquality pixels (e.g., Chen et al., 2019).
Although the focus of this paper is on Chl data products, statistics of cloudadjacent pixels showed that the
7 × 5 straylight mask might be relaxed to 3 × 3 even for R
(λ). An extensive evaluation for the entire
Figure 14. Histograms of (a) R
, (b) R
, (c) R
, and (d) Chl
from pixels near cloud edges after applying
different masking methods. The statistics were derived from all VIIRS images in 2013 over the North Atlantic gyre for
Chl < 0.25 mg m
Figure 15. Histograms of (a) R
, (b) R
, (c) R
, and (d) Chl
from pixels near cloud edges after
applying different masking methods. The statistics were derived from all VIIRS images in 2013 over the Gulf of Mexico
(98°80°W, 18°31°N) for Chl > 0.4 mg m
Journal of Geophysical Research: Oceans
HU ET AL. 1540
MODISA time series of 20032016 indicated that although the new masking scheme resulted in slightly
different (up to 3%) global statistics in R
(667), the global statistics of R
(λ) for all MODISA ocean
bands remained nearly unchanged (; AT135
= 7 × 5 masking, AT140 = 3 × 3 masking). However, the distributions of regional R
(λ) have changed
due to increased number of observations (e.g., Figures 12 and 13). This has signicant implications for all
other ocean color data products (e.g., inherent optical properties such as absorption and backscattering coef-
cients), as R
(λ) is the common input to algorithms for these products. Therefore, similar analyses may be
conducted on these products in the future. Likewise, whether the new straylight masking scheme is applic-
able to sensors other than MODISA should also be evaluated. For example, analyses of VIIRS data over the
North Atlantic gyre and Gulf of Mexico for both clear waters (Chl < 0.25) and more productive waters
(Chl > 0.4 mg m
), respectively, showed that statistically, there is little difference between pixels within
the 3 × 3 7 × 5 rings and outside the 7 × 5 straylight mask (Figures 14 and 15). These results indicate that
VIIRS straylight mask may also be relaxed from 7 × 5 to 3 × 3 in order to increase data quantity without com-
promising data quality.
Unlike other data recovering schemes that rely on a model to change the R
(λ) values in order to improve
coverage (Chen et al., 2016) or a machine learning approach to use R
(λ; Rayleigh corrected reectance) to
improve coverage (Chen et al., 2019), the new masking scheme does not change R
(λ), therefore is straight-
forward to implement and test for Chl retrievals. Once proven feasible, the scheme is expected to have a
major impact on ocean color data records from either a single sensor or merged multiple sensors, for exam-
ple, by the Ocean Color Climate Change Initiative supported by the European Space Agency (Brewin et al.,
2015) or by a NASAfunded multisensor data merging effort (Maritorena et al., 2010).
Finally, regardless of all its advantages, the CI design is still empirical, and uncertainties due to variable
contributions from colored dissolved organic matter or other nonliving water constituents are embedded
in the empirical Chl data products, and such contributions may vary across different ocean basins (Szeto
et al., 2011; Figure 5 of Hu et al., 2013). Although the impact of such uncertainties may be small when
evaluating relative longterm Chl changes of individual ocean basins, they may represent a small but vari-
able bias in the absolute Chl values across ocean basins. Such uncertainties might be reduced through an
empirical correction scheme using the curvature from the 412, 443, and 488nm bands (e.g., Hu et al.,
2014), yet its efciency still remains to be tested.
5. Conclusion
The original OCI algorithm designed to retrieve Chl of global ocean surface waters from satellite ocean color
measurements has been revisited using more eld data. Compared to the original OCI algorithm (OCI1), the
new OCI2 algorithm shows similar improvements over the traditional OCx for oligotrophic waters in terms
of accuracy, image quality, and crosssensor consistency between SeaWiFS, MODISA, and VIIRS, but at the
same time leads to smoother transition in the intermediate Chl range (0.250.4 mg m
). Therefore, OCI2
may be a better choice than OCI1 for global data processing. More importantly, a relaxed straylight masking
scheme (from 7 × 5 to 3 × 3 pixels) leads to signicantly increased data quantity without losing data quality
for both Chl and R
(λ) when tested with MODISA and VIIRS over the North Atlantic (oligotrophic waters)
and Gulf of Mexico (more productive waters). The increases in data quantity over global oceans reach 39%.
Future efforts should focus on collection of highquality data in ocean gyres with Chl < 0.05 mg
in order to
further improve data product accuracy for these extremely clear waters, and focus on evaluation of other
ocean color data products near cloud edges from the mainstream ocean color sensors.
Antoine, D., Morel, A., Gordon, H. R., Banzon, V. F., & Evans, R. H. (2005). Bridging ocean color observations of the 1980s and 2000s in
search of longterm trends. Journal of Geophysical Research,110, C06009.
Bailey, S. W., & Werdell, P. J. (2006). A multisensor approach for the onorbit validation of ocean color satellite data products. Remote
Sensing of Environment,102(12), 1223.
Brewin, R. J. W., Raitsos, D. E., Pradhan, Y., & Hoteit, I. (2013). Comparison of chlorophyll in the Red Sea derived from MODISaqua and
in vivo uorescence. Remote Sensing of Environment,136, 218224.
Brewin, R. J. W., Sathyendranath, S., Müller, D., Brockmann, C., Deschamps, P. Y., Devred, E., et al. (2015). The ocean colour climate
change initiative: III. A roundrobin comparison on inwater biooptical algorithms. Remote Sensing of Environment,162, 271294.
Journal of Geophysical Research: Oceans
HU ET AL. 1541
This study was supported by the U.S.
algorithm renement and for data
product improvement. Satellite data
were provided by NASA Ocean Biology
Processing Group (OBPG) while VIIRS
level1 data were provided by NOAA.
We thank all researchers who
contributed to the NASA SeaBASS data
archive. All eld and satellite data can
be obtained from the NASA OB.DAAC
through http://oceancolor.gsfc.nasa.
gov. We thank the two anonymous
reviewers for their valuable comments
to improve this manuscript.
Campbell, J. W. (1995). The lognormal distribution as a model for biooptical variability in the sea. Journal of Geophysical Research,
100(C7), 13,23713,254.
Carder, K. L., Chen, F. R., Lee, Z. P., Hawes, S. K., & Kamykowski, D. (1999). Semianalytic moderateresolution imaging spectrometer
algorithms for chlorophyll aand absorption with biooptical domains based on nitratedepletion temperatures. Journal of Geophysical
Research,104(C3), 54035421.
Chen, J., Lee, Z., Hu, C., & Wei, J. (2016). Improving satellite data products for open oceans with a scheme to correct the residual errors in
remote sensing reectance. Journal of Geophysical Research: Oceans,121, 38663886.
Chen, S., Hu, C., Barnes, B. B., Xie, Y., Lin, G., & Qiu, Z. (2019). Improving ocean color data coverage through machine learning. Remote
Sens. Environ., Remote Sens. Environ.,222,286302.
Dierssen, H. M. (2010). Perspectives on empirical approaches for ocean color remote sensing of chlorophyll in a changing climate.
Proceedings of the National Academy of Sciences,107(40), 17,07317,078.
Doron, M., Babin, M., Hembise, O., Mangin, A., & Garnesson, P. (2011). Ocean transparency from space: Validation of algorithms using
MERIS, MODIS and SeaWiFS data. Remote Sensing of Environment,115(12), 29863001.
Feng, L., & Hu, C. (2016a). Comparison of valid ocean observations between MODIS Terra and Aqua over the global oceans. IEEE
Transactions on Geoscience and Remote Sensing,54(3), 15751585.
Feng, L., & Hu, C. (2016b). Cloud adjacency effects on topofatmosphere radiance and ocean color data products: A statistical assessment.
Remote Sensing of Environment,174, 301313.
Franz, B. A. (2014). NASA Ocean Biology Processing Group update (MODIS, VIIRS, MERIS, HICO, GOCI, SGLI, OLCI, LandSat8, Multi
Mission Processing). NASA Ocean Color Research Team Meeting, 57 May 2014. Washington, DC. Retrieved from https://oceancolor.
Franz, B. A., Bailey, S. W., Werdell, P. J., & McClain, C. R. (2007). Sensorindependent approach to the vicarious calibration of satellite
ocean color radiometry. Applied Optics,46(22), 50685082.
Franz, B. A., Behrenfeld, M. J., Siegel, D. A., & Signorini, S. R. (2016). Global ocean phytoplankton [in state of the climate in 2015].
Bulletin of the American Meteorological Society,97(8), SiS275.
Franz, B. A., Werdell, P. J., Meister, G., Bailey, S. W., Eplee, R. E., Feldman, G. C., et al. (2005). The continuity of ocean color
measurements from SeaWiFS to MODIS. Proc. SPIE 5882, Earth Observing Systems X, 58820W (2005/08/22).
Frouin, R., Schwindling, M., & Deschamps, P.Y. (1996). Spectral reectance of sea foam in the visible and nearinfrared: In situ mea-
surements and remote sensing implications. Journal of Geophysical Research,101(C6), 14,36114,371.
Gordon, H. R. (1997). Atmospheric correction of ocean color imagery in the Earth Observing System era. Journal of Geophysical Research,
102(D14), 17,08117,106.
Gordon, H. R., & Clark, D. K. (1981). Clear water radiances for atmospheric correction of coastal zone color scanner imagery. Applied
Optics,20(24), 41754180.
Gordon, H. R., & Wang, M. (1994). Retrieval of waterleaving radiance and aerosol optical thickness over the oceans with SeaWiFS: A
preliminary algorithm. Applied Optics,33(3), 443452.
Gregg, W. W., Casey, N. W., & McClain, C. R. (2005). Recent trends in global ocean chlorophyll. Geophysical Research Letters,32, L03606.
Hu, C., & Campbell, J. (2014). Oceanic chlorophyllacontent. In J. M. Hanes (Ed.), Biophysical Applications of Satellite Remote Sensing
(pp. 171203). Berlin Heidelberg: Springer Remote Sensing/Photogrammetry. SpringerVerlag.
Hu, C., Feng, L., & Lee, Z. (2013). Uncertainties of SeaWiFS and MODIS remote sensing reectance: Implications from clear water mea-
surements. Remote Sensing of Environment,133, 168182.
Hu, C., Lee, Z., & Franz, B. (2012). Chlorophyll aalgorithms for oligotrophic oceans: A novel approach based on threeband reectance
difference. Journal of Geophysical Research,117, C01011.
Hu, C., Qi, L., Lee, Z., & Franz, B. A. (2014). Minimize CDOM impact on the bandsubtraction chlorophyll algorithm through optical
weighting: Preliminary results. Ocean Optics XXII, Oct 2431, 2014, Portland, ME.
IOCCG (2006). Remote sensing of inherent optical properties: Fundamentals, tests of algorithms, and applications. Reports of the
International OceanColour Coordinating Group, No. 5. Z.P. Lee. Dartmouth, Canada, IOCCG.
Jiang, L., & Wang, M. (2013). Identication of pixels with stray light and cloud shadow contaminations in the satellite ocean color data
processing. Applied Optics,52, 67576770.
Kahru, M., & Mitchell, B. G. (1999). Empirical chlorophyll algorithm and preliminary SeaWiFS validation for the California Current.
International Journal of Remote Sensing,20(17), 34233429.
King, M. D., Platnick, S., Menzel, W. P., Ackerman, S. A., & Hubanks, P. A. (2013). Spatial and temporal distribution of clouds observed by
MODIS onboard the Terra and Aqua satellites. IEEE Transactions on Geoscience and Remote Sensing,51(7), 38263852.
Le, C., Zhou, X., Hu, C., Lee, Z., Li, L., & Stramski, D. (2018). A colorindexbased empirical algorithm for determining particulate organic
carbon concentration in the ocean from satellite observations. Journal of Geophysical Research: Oceans,123, 74077419.
Lee, Z. P., Carder, K. L., & Arnone, R. (2002). Deriving inherent optical properties from water color: A multiband quasianalytical algo-
rithm for optically deep waters. Applied Optics,41(27), 57555772.
Lee, Z. P., Shang, S., Hu, C., Du, K., Weidemann, A., Hou, W., et al. (2015). Secchi disk depth: A new theory and mechanistic model for
underwater visibility. Remote Sensing of Environment,169,139149.
Maritorena, S., d'Andon, O. H. F., Mangin, A., & Siegel, D. A. (2010). Merged satellite ocean color data products using a biooptical
model: Characteristics, benets and issues. Remote Sensing of Environment,114(8), 17911804.
Maritorena, S., Siegel, D. A., & Peterson, A. (2002). Optimization of a semianalytical ocean color model for global scale applications.
Applied Optics,41(15), 27052714.
McClain, C. R. (2009). A decade of satellite ocean color observations. Annual Review of Marine Science,1(1), 1942.
McClain, C. R., Feldman, G. C., & Hooker, S. B. (2004). An overview of the SeaWiFS project and strategies for producing a climate research
quality global ocean biooptical time series. DeepSea Research II,51(13), 542.
Journal of Geophysical Research: Oceans
HU ET AL. 1542
Meister, G., & McClain, C. R. (2010). Pointspread function of the ocean color bands of the Moderate Resolution Imaging
Spectroradiometer on Aqua. Applied Optics,49(32), 62766285.
Mitchell, C., Hu, C., Bowler, B., Drapeau, D., & Balch, W. M. (2017). Estimating particulate inorganic carbon concentrations of the global
ocean from ocean color measurements using a reectance difference approach. Journal of Geophysical Research: Oceans,122, 87078720.
Moore, T. S., Campbell, J. W., & Feng, H. (2015). Characterizing the uncertainties in spectral remote sensing reectance for SeaWiFS and
MODISAqua based on global in situ matchup data sets. Remote Sensing of Environment,159,1427.
Morel, A., & Maritorena, S. (2001). Biooptical properties of oceanic waters: A reappraisal. Journal of Geophysical Research,106(C4),
O'Reilly, J. E., Maritorena, S., O'brien, M. C., Siegel, D. A., Toole, D., Menzies, D., et al. (2000). SeaWiFS Postlaunch calibration and
validation analyses, Part 3. In S. B. Hooker & E. R. Firestone (Eds.), NASA Tech. Memo. 2000206892 (Vol. 11, 49 pp.). Greenbelt, MD:
NASA Goddard Space Flight Center.
Polovina, J. J., Howell, E. A., & Abecassis, M. (2008). Ocean's least productive waters are expanding. Geophysical Research Letters,35,
Qi, L., Lee, Z., Hu, C., & Wang, M. (2017). Requirement of minimal signaltonoise ratios of ocean color sensors and uncertainties of ocean
color products. Journal of Geophysical Research: Oceans,122, 25952611.
Sathyendranath, S., Prieur, L., & More, A. (1989). A three component model of ocean colour and its application to remote sensing of
phytoplankton pigments in coastal waters. International Journal of Remote Sensing,10(8), 13731394.
Signorini, S., Franz, B. A., & McClain, C. R. (2015). Chlorophyll variability in the oligotrophic gyres: Mechanisms, seasonality and trends.
Frontiers in Marine Science,2.
Stumpf, R. P., Arnone, R., Gould, R. W., Martinolich, P. M., & Ransibrahmanakul, V. (2003). A PartiallyCoupled OceanAtmosphere Model
for Retrieval of WaterLeaving Radiance From SeaWiFS in Coastal Waters (Vol. 22). Maryland: NASA Goddard Space Flight Center,
Szeto, M., Werdell, P. J., Moore, T. S., & Campbell, J. W. (2011 ). Are the world's oceans optically different? Journal of Geophysical Research,
116, C00H04.
Wang, M., & Bailey, S. W. (2001). Correction of Sun glint contamination on the SeaWiFS ocean and atmosphere products. Applied Optics,
40(27), 47904798.
Wang, M., & Shi, W. (2007). The NIRSWIR combined atmospheric correction approach for MODIS ocean color data processing. Optics
Express,15(24), 15,72215,733.
Wang, M., & Son, S.H. (2016). VIIRSderived chlorophyllausing the ocean color index method. Remote Sensing of Environment,182,
Wei, J., Lee, Z., & Shang, S. (2016). A system to measure the data quality of spectral remote sensing reectance of aquatic environments.
Journal of Geophysical Research: Oceans,121, 81898207.
Werdell, P. J., & Bailey, S. W. (2005). An improved insitu biooptical data set for ocean color algorithm development and satellite data
product validation. Remote Sensing of Environment,98(1), 122140.
Werdell, P. J., Bailey, S. W., Fargion, G., Pietras, C., Knobelspiesse, K., Feldman, G. C., & McClain, C. R. (2003). Unique data repository
facilitates ocean color satellite validation. Eos, Transactions of the American Geophysical Union,84(38), 377.
Journal of Geophysical Research: Oceans
HU ET AL. 1543
... CEAN color satellite missions have provided continuous observations of the global oceans in recent decades. These satellite observations have been used to derive various bio-optical parameters of the upper ocean, such as the chlorophyll concentration (Chl) [1][2][3], the content of particulate organic carbon (POC) [4], [5], and net primary productivity (NPP) [6], [7], the outcomes of which benefit scientific research and environmental protection measures. However, an accurate atmospheric correction (AC) is required to remove interference signals from the atmospheric path radiance before any of these parameters can be retrieved [8]. ...
... MODISA Level-1A data were processed with NASA's SeaDAS software version 7.5 using the standard iterative NIR algorithm [13], [28], [29] to obtain the Rayleigh scattering reflectance and R rs , as well as the solar zenith angle (SZA), viewing zenith angle (VZA), and relative azimuth angle (RAA). The TOA reflectance ( ρ t ) measured by the sensor can be expressed as ρ t (λ) = ρ r (λ) + [ρ a (λ) + ρ ra (λ)] + T(λ) × ρ g (λ) + t v × ρ wc (λ) + t v × ρ w (λ), (1) where λ is the wavelength, ρ r (λ) is the reflectance contributed by Rayleigh scattering due to atmospheric molecules, ρ a (λ) is the reflectance contributed by aerosol scattering, ρ ra (λ) is the reflectance contributed by aerosol-Rayleigh interaction scattering, ρ g (λ) is the reflectance contributed by sunglint, ρ wc (λ) is the reflectance contributed by sea surface whitecap, and T(λ) and t v are the beam and diffuse transmittances, respectively, of the atmosphere from the sea surface to the sensor. ρ w (λ) is the water-leaving reflectance, and its relation to the desired quantity, R rs , can be expressed as ...
... We obtained a total of 1.04×10 9 high-quality ρ rc -R rs pairs, but such a data volume is excessively large considering the limited model parameters for deep learning and the simple spectral shapes in the open ocean (mostly belonging to the first three optical water types in the QA system). To exclude duplicate data, we divided the data according to eight Chl bins ranging from 0 to 0.4 mg m -3 (with log-spaced intervals) and randomly selected 600,000 matchups for each bin (4,800,000 in total); the Chl values were estimated with the improved OCI algorithm [1]. The ranges of the training dataset and validation dataset are shown in Appendix Fig. Ⅰ. ...
Full-text available
Although the 5% mission goal for NASA’s standard atmospheric correction (AC) algorithm (i.e., the near-infrared (NIR) algorithm) for oligotrophic oceans has been met, this algorithm applies only to blue bands and is highly sensitive to contamination from cloud straylight and sunglint. Here, we developed an AC algorithm for clear waters based on deep learning (namely, DLAC). The algorithm was trained using 3.6 million pairs of MODIS-Aqua high-quality Rrs from the NIR algorithm and Rayleigh-corrected reflectances selected across the global oceans and from all seasons. Validations using in situ data and a chlorophyll (Chl) constraint-based approach showed that the uncertainties in the R <sub xmlns:mml="" xmlns:xlink="">rs</sub> retrievals for DLAC are lower than those for the NIR algorithm, especially for the green and red bands. More importantly, the DLAC algorithm is more tolerant to cloud adjacency effects and moderate sunglint. As a result, the number of valid observations increased by ~50%, and the coverage of monthly global Level-3 R <sub xmlns:mml="" xmlns:xlink="">rs</sub> composites increased by up to 20%. More spatially and temporally consistent patterns were also found for the Level-3 R <sub xmlns:mml="" xmlns:xlink="">rs</sub> and Chl products, and large changes in their magnitudes (up to 20% for R <sub xmlns:mml="" xmlns:xlink="">rs</sub> and 30% for Chl) were detected in some oceanic regions. With these improvements in the quality and quantity of data, our DLAC algorithm may be valuable as another option for processing global data.
... The empirical coefficients of different sensors have also been further optimized to verify the accuracy of OCx algorithms [19]. Hu et al. found that the performance of OCx algorithms was not good at low Chl-a concentrations; hence, they proposed the color index (CI) algorithm for Chl-a concentrations < 0.15 mg m −3 and set Chl-a concentration thresholds for the application of the OCx and CI algorithms, creating the ocean color index (OCI) algorithm [20,21]. However, the OCx and OCI algorithms were both designed to achieve global ocean observations and may cause errors at regional scales due to the different optical properties caused by differences in colored dis-solved matter and phytoplankton community structure [22]. ...
... The in situ Chl-a data (Chl in situ ) and log 10 (MBR) were plugged into the polynomials with the original empirical coefficients as the starting coefficients to improve the original algorithms of the three sensors. Previous studies have shown that the performance of OCx algorithms is not good at low Chl-a concentrations and that systematic deviation exists [20,21,28,34]. In this study, according to the fitting results for the Chl-a data and variation of bias, the in situ data and satellite data were divided into two groups for each sensor (a low-Chl-a concentration group and a high-Chl-a concentration group). ...
... In different application ranges, the new algorithms or default algorithms of each sensor were used to generate Chl-a concentration data. Algorithms using blending windows are called blending algorithms [21,28]. ...
Full-text available
Chlorophyll-a (Chl-a) is an important marine indicator, and the improvement in Chl-a concentration retrieval for ocean color remote sensing is always a major challenge. This study focuses on the northwest Pacific fishing ground (NPFG) to evaluate and improve the Chl-a products of three mainstream remote sensing satellites, Himawari-8, MODIS-Aqua, and VIIRS SNPP. We analyzed in situ data and found that an in situ Chl-a concentration of 0.3 mg m−3 could be used as a threshold to distinguish the systematic deviation of remote sensing Chl-a data in the NPFG. Based on this threshold, we optimized the Chl-a algorithms of the three satellites by data grouping, and integrated multisource satellite Chl-a data by weighted averaging to acquire high-coverage merged data. The merged data were thoroughly verifified by Argo Chl-a data. The Chl-a front of merged Chl-a data could be represented accurately and completely and had a good correlation with the distribution of the NPFG. The most important marine factors for Chl-a are nutrients and temperature, which are affected by mesoscale eddies and variations in the Kuroshio extension. The variation trend of merged Chl-a data is consistent with mesoscale eddies and Kuroshio extension and has more sensitive responses to the marine climatic conditions of ENSO
... High-frequency hyperspectral optical data can complement relatively scarce in situ measurements. This allows improving the knowledge about short-term processes in lakes and could be linked with Earth Observation (EO) measurements to increase knowledge in spatial scale (Siegel et al., 2013;Binding et al., 2018;Hu et al., 2019). EO data provides a frequent, large-scale synoptic overview of lakes and has been increasingly integrated operationally into inland water algal bloom monitoring (Binding et al., 2021). ...
Full-text available
Phytoplankton and its most common pigment chlorophyll a (Chl-a) are important parameters in characterizing lake ecosystems. We compared six methods to measure the concentration of Chl a (CChl-a) in two optically different lakes: stratified clear-water Lake Saadjärv and non-stratified turbid Lake Võrtsjärv. CChl-a was estimated from: in vitro (spectrophotometric, high-performance liquid chromatography); fluorescence (in situ automated high-frequency measurement (AHFM) buoys) and spectral (in situ high-frequency hyperspectral above-water radiometer (WISPStation), satellites Sentinel-3 OLCI and Sentinel-2 MSI) measurements. The agreement between methods ranged from weak (R 2 = 0.26) to strong (R 2 = 0.93). The consistency was better in turbid lake compared to the clear-water lake where the vertical and short-term temporal variability of the CChl-a was larger. The agreement between the methods depends on multiple factors, e.g., the environmental and in-water conditions, placement of sensors, sensitivity of algorithms. Also in case of some methods, seasonal bias can be detected in both lakes due to signal strength and background turbidity. The inherent differences of the methods should be studied before the synergistic use of data which will clearly increase the spatial (via satellites), temporal (AHFM buoy, WISPStation and satellites) and vertical (profiling AHFM buoy) coverage of data necessary to advance the research on phytoplankton dynamics in lakes.
... To follow the convention adopted by NASA in determining the coefficients for OCx algorithms (Hu et al., 2019), we binned the data according to different levels of Chl-a and R rs,645 (598 in total for pairs with R rs,645 above 0.006). Chl-a data were first binned in logarithmic space (Eq. ...
Extensive human activities and climate change in recent decades have triggered severe eutrophication problems in the coastal oceans in the Greater Bay Area (GBA) of China. However, a comprehensive characterization of the spatial and temporal patterns of chlorophyll-a (Chl-a, a major indicator of phytoplankton biomass) in this region is not available. Our study attempts to fill this gap by using long-term satellite observations. With massive in situ datasets from underway sampling systems, we developed a novel hybrid Chl-a retrieval algorithm combining the recalibrated OC3 and line-height-based (BL443) algorithms for waters with different turbidity levels. Satellite-retrieved Chl-a values with the hybrid algorithm agreed well with in situ measurements, with an uncertainty level of 33.8 %. Long-term analysis revealed significant decreasing trends over the inner Pearl River Estuary (averaged at 0.054 μg/L yr⁻¹), while significant increasing trends were found in eastern Daya Bay (averaged at 0.035 μg/L yr⁻¹). The developed algorithm is expected to aid routine Chl-a monitoring in the adjacent oceans of the GBA, and the long-term datasets here can serve as critical information for further coastal conservation and management efforts.
... Atmospheric correction algorithms are typically applied to remove the contribution of the atmosphere from TOA radiances, which can exceed 90% of the total signal, and produce estimates of spectral remote-sensing reflectances (R rs ), the light exiting the water normalized to the downwelling surface irradiance (Mobley et al., 2016). Bio-optical algorithms are then applied to the derived R rs to produce estimates of geophysical and optical quantities, such as the near surface concentration of the phytoplankton pigment chlorophyll-a (Hu et al., 2019;O'Reilly and Werdell 2019), spectral inherent optical properties (IOCCG 2006;Werdell et al., 2018), and metrics of phytoplankton community composition Mouw et al., 2017). The performance requirements needed to accurately perform ocean color atmospheric correction and geophysical inversions ultimately dictated how OCI was built, operated, and tested. ...
Full-text available
This paper summarizes the results from the system level test campaign of the Engineering Test Unit (ETU) of the ‘Ocean Color Instrument’ (OCI), the primary payload of NASA’s ‘Plankton, Aerosol, Cloud and ocean Ecosystem’ (PACE) mission. The main goals of the test campaign were to optimize characterization procedures and evaluate system level performance relative to model predictions. Critical performance parameters such as radiometric gain, signal-to-noise ratio, polarization, instantaneous field-of-view, temperature sensitivity, relative spectral response and stability were evaluated for wavelengths from 600 to 2,260 nm and are in line with expectations. We expect the OCI flight unit to meet the PACE mission performance requirements. Building and testing the ETU has been extremely important for the development of the OCI flight unit (e.g. improved SNR by increasing the aperture, optimized thermal design), and we strongly recommend the inclusion of an ETU in the development of future spaceborne sensors that rely on novel technological designs. ETU testing led to the discovery of a hysteresis issue with the SWIR bands, and a correction algorithm was developed. Also, the coregistration of the SWIR bands relative to each other is worse than expected, but this was discovered too late in the schedule to remediate.
Full-text available
The number of cloud droplets per unit volume (Nd) is a fundamentally important property of marine boundary layer (MBL) liquid clouds that, at constant liquid water path, exerts considerable controls on albedo. Past work has shown that regional Nd has a direct correlation to marine primary productivity (PP) because of the role of seasonally varying, biogenically derived precursor gases in modulating secondary aerosol properties. These linkages are thought to be observable over the high-latitude oceans, where strong seasonal variability in aerosol and meteorology covary in mostly pristine environments. Here, we examine Nd variability derived from 5 years of MODIS Level 2-derived cloud properties in a broad region of the summer eastern Southern Ocean and adjacent marginal seas. We demonstrate latitudinal, longitudinal and temporal gradients in Nd that are strongly correlated with the passage of air masses over high-PP waters that are mostly concentrated along the Antarctic Shelf poleward of 60∘ S. We find that the albedo of MBL clouds in the latitudes south of 60∘ S is significantly higher than similar liquid water path (LWP) clouds north of this latitude.
The biological pump transports organic matter, created by phytoplankton productivity in the well-lit surface ocean, to the ocean's dark interior, where it is consumed by animals and heterotrophic microbes and remineralized back to inorganic forms. This downward transport of organic matter sequesters carbon dioxide from exchange with the atmosphere on timescales of months to millennia, depending on where in the water column the respiration occurs. There are three primary export pathways that link the upper ocean to the interior: the gravitational, migrant, and mixing pumps. These pathways are regulated by vastly different mechanisms, making it challenging to quantify the impacts of the biological pump on the global carbon cycle. In this review, we assess progress toward creating a global accounting of carbon export and sequestration via the biological pump and suggest a potential path toward achieving this goal. Expected final online publication date for the Annual Review of Marine Science, Volume 15 is January 2023. Please see for revised estimates.
Sea surface chlorophyll-a concentration (Chl-a) is a key proxy for phytoplankton biomass. Spatio-temporal continuous Chl-a data are important to understand the mechanisms of chlorophyll occurrence and development and track phytoplankton changes. However, the greatest challenge in utilizing daily Chl-a data is massive missing pixels due to orbital position and cloud coverage. This study proposes the application of a spatial filling method using the machine learning-based Extreme Gradient Boosting (BST) to reconstruct missing pixels of daily MODIS Chl-a data from 2007 to 2018. The approach is applied to different trophic biogeographical subregions of the Northwestern Pacific where it has complex phytoplankton dynamics and frequent data missing. Various environmental variables are taken into consideration, including meteorological forcing, geographic and topographic features, and oceanic physical components. The BST-reconstructed Chl-a (BST Chl-a) is validated using in-situ Chl-a measurements, VIIRS and Himawari-8 Chl-a products. The results show that the BST model is highly adaptive in reconstructing Chl-a data, and it performs well in pelagic, offshore and coastal with the best performance in pelagic. BST Chl-a improves coverage without significant quality degradation compared to the original MODIS Chl-a. BST Chl-a agrees better with in-situ data than that of MODIS, with CC of 0.742, RMSE of 0.247, MAE of 0.202 and Bias of 0.089. Cross-satellite validation using VIIRS and Himawari-8 Chl-a also shows promising results with the CC of 0.861 and 0.765, respectively, suggesting the high accuracy of BST Chl-a. The inter-annual trend of BST Chl-a decreases in coastal and increases in offshore and pelagic. BST Chl-a images present similar spatial patterns to MODIS Chl-a under different missing rates, with gradual decreases from coastal to pelagic. It indicates that phytoplankton bloom patterns can be identified by daily BST Chl-a images.
The main objective of this study is to improve the retrieval of phytoplankton absorption coefficients using ocean and land color instrument (OLCI) bands (413, 443, 490, 510, 560, and 665 nm). In this study, the Raman-scattering correction was considered in an analytical forward model, and the corresponding Raman excitation OLCI band centers were calculated. The coefficients and exponents for the power law model of phytoplankton absorption were determined by a neural network classifier using a combination of sun elevation, photosynthetically active radiation, and remote sensing reflectance at OLCI bands (413, 443, 490, 510, 560, 620, 665, and 683 nm). Two optimizations were executed. The shape of the colored detrital matter (CDM) spectrum was allowed to change during the second optimization. Based on comparisons of the phytoplankton absorption coefficients at 443 nm ( a Φ [443]) predicted by the improved inversion with field measurements taken from cruise surveys of the Pearl River estuary and Daya Bay and the SeaWiFS Bio-optical Archive and Storage System dataset and NASA bio-Optical Marine Algorithm Dataset, the modified inversion procedure could provide a good performance ( r ² = 0.86). The results showed that the contribution of Raman scattering to the remote sensing reflectance at 665 nm exceeded 22% in the open ocean. Moreover, the residual from the first optimization was compared with that from the second optimization, demonstrating that the variable CDM spectral slope in the analytical forward model could improve the accuracy of the forward model.
The present study aims to address the effect of spatial resolution and retrieval algorithms on the performance of operational chlorophyll-a (chl-a) product from different sensors. To meet the objective, chl-a product at different spatial resolutions derived from MODIS, SNPP-VIIRS, OC-CCI, and Sentinel-3A OLCI were compared with in-situ measurements of the coastal waters of Karnataka, south-eastern Arabian Sea and errors were quantified using statistical criteria. MODIS-derived chl-a concentrations at 4 km tends to overestimate in the nearshore waters as shown by a high mean relative percentage difference (MRPD) of 77.09%, and a low coefficient of determination (R²) of 0.35. A significant improvement was seen for MODIS 1 km chl-a product, (R² = 0.64, MRPD = 22.30%) and Mean Absolute Percentage Deviation (MAPD) = 47.08%. Validation of SNPP-VIIRS derived chl-a of 4 km with in-situ measurements showed moderate correlation (R² = 0.54, MAPD = 49.21%), whereas 1 km SNPP-VIIRS showed substantially better result than 4 km with majority of data points aggregated along the 1:1 line (R² = 0.86, MAPD = 43.66%, MRPD = 20.97%). Chlorophyll-a product at original resolution (750 m) showed less error than 4 km and 1 km resolution (MAPD = 39.18%, R² = 0.75, MRPD = 10.875%). In contrast, OC-CCI underestimated field measured chl-a concentrations as indicated by MRPD (-36.72%), with moderate R² = 0.58. All 4 km data showed poor results than higher spatial resolution data. Sentinel-3A OLCI 300 m full-resolution (FR) derived chl-a, showed good agreement with sea-truth data (R² = 0.80, MAPD = 33.07%, and MRPD = 23.89%), whereas Sentinel-3A reduced-resolution (RR-1.2 km) exhibited slight overestimation (MRPD = 28.43%, MAPD = 39.88%, R² = 0.78) than Sentinel-3A(FR). Comparison of operational chl-a product from OC-CCI, MODIS, SNPP-VIIRS, and Sentinel-3A OLCI illustrated that Sentinel-3A OLCI performed best in coastal waters of Karnataka indicating the importance of higher spatial resolution and neural network-based algorithms.
Full-text available
A new algorithm for estimating particulate inorganic carbon (PIC) concentrations from ocean color measurements is presented. PIC plays an important role in the global carbon cycle through the oceanic carbonate pump, therefore accurate estimations of PIC concentrations from satellite remote sensing are crucial for observing changes on a global scale. An extensive global data set was created from field and satellite observations for investigating the relationship between PIC concentrations and differences in the remote sensing reflectance (Rrs) at green, red, and near-infrared (NIR) wavebands. Three color indices were defined: two as the relative height of Rrs(667) above a baseline running between Rrs(547) and an Rrs in the NIR (either 748 or 869 nm), and one as the difference between Rrs(547) and Rrs(667). All three color indices were found to explain over 90% of the variance in field-measured PIC. But, due to the lack of availability of Rrs(NIR) in the standard ocean color data products, most of the further analysis presented here was done using the color index determined from only two bands. The new two-band color index algorithm was found to retrieve PIC concentrations more accurately than the current standard algorithm used in generating global PIC data products. Application of the new algorithm to satellite imagery showed patterns on the global scale as revealed from field measurements. The new algorithm was more resistant to atmospheric correction errors and residual errors in sun glint corrections, as seen by a reduction in the speckling and patchiness in the satellite-derived PIC images.
Full-text available
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.
Oceanic chlorophyll-a concentration (Chl, mg m⁻³) maps derived from satellite ocean color measurements are the only data source which provides synoptic information of phytoplankton abundance on global scale. However, after excluding data collected under non-optimal observing conditions such as strong sun glint, clouds, thick aerosols, straylight, and large viewing angles, only ~5% of MODIS ocean measurements lead to valid Chl retrievals, regardless of the fact that about 25–30% of the global ocean is cloud free. A recently developed ocean color index (CI) is effective in deriving relative ocean color patterns under most non-optimal observing conditions to improve coverage, but these patterns cannot be interpreted as Chl. In this study, we combine the advantage of the high-quality, low-coverage Chl and lower-quality, higher-coverage CI to improve spatial and temporal coverage of Chl through machine learning, specifically via a random forest based regression ensemble (RFRE) approach. For every MODIS scene, the machine learning requires CI, Rayleigh-corrected reflectance (Rrc (λ = 469, 555, 645 nm), dimensionless), and high-quality low-coverage Chl from the common pixels where they all have valid data to develop an RFRE-based model to convert CI and Rrc (λ) to Chl. The model is then applied to all valid CI pixels of the same scene to derive Chl. This process is repeated for each scene, and the model parameterization is optimized for each scene independently. The approach has been tested for the Yellow Sea and East China Sea (YSECS) where non-optimal observing conditions frequently occur. Validations using extensive field measurements and image-based statistics for 2017 show very promising results, where coverage in the new Chl maps is increased by ~3.5 times without noticeable degradation in quality as compared with the original Chl data products. The improvement in Chl coverage without compromising data quality is not only critical in revealing otherwise unknown bloom patterns, but also important in reducing uncertainties in time-series analysis. Tests of the RFRE approach for several other regions such as the East Caribbean, Arabian Sea, and Gulf of Mexico suggest its general applicability in improving Chl coverage of other regions.
An empirical algorithm for estimating particulate organic carbon (POC) concentration in the surface ocean from satellite observations is formulated and validated using in situ POC data and remote-sensing reflectance (Rrs) data obtained from match-up satellite ocean color measurements. The algorithm builds upon the band-difference algorithm concept, which was originally developed for estimating chlorophyll-a concentration in clear waters. This algorithm utilizes three spectral bands centered approximately at 490, 550, and 670 nm to determine a color index (CIPOC), from which POC can be estimated from satellite measurements. For comparison, the blue-green band-ratio algorithm is also formulated using the same data set of in situ POC and satellite-derived Rrs. Results show that the statistical parameters characterizing the differences between the satellite-derived POC and matchup in situ POC are similar when the CIPOC and band ratio algorithms are applied to open ocean waters where the values of CIPOC are relatively low. In coastal waters where the values of CIPOC are generally higher, the statistical parameters of algorithm performance are better for the CIPOC algorithm. In addition, because the CIPOC algorithm is less sensitive to errors and noise in the satellite-derived Rrs, the image quality obtained with this algorithm can be improved for both open-ocean and coastal waters.
Using simulations, error propagation theory, and measurements from the Moderate Resolution Imaging Spectroradiometer (MODIS), we determined the minimal signal-to-noise ratio (SNR) required for ocean color measurements and product uncertainties at different spatial and temporal scales. First, based on typical top-of-atmosphere (TOA) radiance over the ocean, we evaluate the uncertainties in satellite-derived Rrs in the visible wavelengths (ΔRrs(vis)) due to sensor noise in both the near-infrared (NIR) and the visible bands. While the former induces noise in Rrs(vis) through atmospheric correction, the latter has a direct impact on Rrs(vis). Such estimated uncertainties are compared with inherent ΔRrs(vis) uncertainties from in situ measurements and from the operational atmosphere correction algorithm. The comparison leads to a conclusion that once SNR(NIR) is above 600:1, an SNR(vis) better than 400:1 will not make a significant reduction in product uncertainties at pixel level under typical conditions for a solar zenith angle of 45°. Then, such uncertainties are found to decrease significantly in data products of oceanic waters when the 1-km pixels from individual images are binned to lower spatial resolution (e.g., 4-km) or temporal resolution (e.g., monthly). Although these findings do not suggest that passive ocean color sensors should have SNR(vis) around 400:1, they do support the argument for more trade space in higher spatial and/or spectral resolutions once this minimal 400:1 SNR(vis) requirement is met. This article is protected by copyright. All rights reserved.
In 2015, the dominant greenhouse gases released into Earth's atmosphere carbon dioxide, methane, and nitrous oxide all continued to reach new high levels. At Mauna Loa, Hawaii, the annual CO2concentration increased by a record 3.1 ppm, exceeding 400 ppm for the first time on record. The 2015 global CO2 average neared this threshold, at 399.4 ppm. Additionally, one of the strongest El Niño events since at least 1950 developed in spring 2015 and continued to evolve through the year. The phenomenon was far reaching, impacting many regions across the globe and affecting most aspects of the climate system. Owing to the combination of El Niño and a long-term upward trend, Earth observed record warmth for the second consecutive year, with the 2015 annual global surface temperature surpassing the previous record by more than 0.1°C and exceeding the average for the mid- to late 19th century commonly considered representative of preindustrial conditions by more than 1°C for the first time. Above Earth's surface, lower troposphere temperatures were near-record high. Across land surfaces, record to near-record warmth was reported across every inhabited continent. Twelve countries, including Russia and China, reported record high annual temperatures. In June, one of the most severe heat waves since 1980 affected Karachi, Pakistan, claiming over 1000 lives. On 27 October, Vredendal, South Africa, reached 48.4°C, a new global high temperature record for this month. In the Arctic, the 2015 land surface temperature was 1.2°C above the 1981-2010 average, tying 2007 and 2011 for the highest annual temperature and representing a 2.8°C increase since the record began in 1900. Increasing temperatures have led to decreasing Arctic sea ice extent and thickness. On 25 February 2015, the lowest maximum sea ice extent in the 37-year satellite record was observed, 7% below the 1981-2010 average. Mean sea surface temperatures across the Arctic Ocean during August in ice-free regions, representative of Arctic Ocean summer anomalies, ranged from ~0°C to 8°C above average. As a consequence of sea ice retreat and warming oceans, vast walrus herds in the Pacific Arctic are hauling out on land rather than on sea ice, raising concern about the energetics of females and young animals. Increasing temperatures in the Barents Sea are linked to a community-wide shift in fish populations: boreal communities are now farther north, and long-standing Arctic species have been almost pushed out of the area. Above average sea surface temperatures are not confined to the Arctic. Sea surface temperature for 2015 was record high at the global scale; however, the North Atlantic southeast of Greenland remained colder than average and colder than 2014. Global annual ocean heat content and mean sea level also reached new record highs. The Greenland Ice Sheet, with the capacity to contribute ~7 m to sea level rise, experienced melting over more than 50% of its surface for the first time since the record melt of 2012. Other aspects of the cryosphere were remarkable. Alpine glacier retreat continued, and preliminary data indicate that 2015 is the 36th consecutive year of negative annual mass balance. Across the Northern Hemisphere, late-spring snow cover extent continued its trend of decline, with June the second lowest in the 49-year satellite record. Below the surface, record high temperatures at 20-m depth were measured at all permafrost observatories on the North Slope of Alaska, increasing by up to 0.66°C decade-1 since 2000. In the Antarctic, surface pressure and temperatures were lower than the 1981-2010 average for most of the year, consistent with the primarily positive southern annular mode, which saw a record high index value of +4.92 in February. Antarctic sea ice extent and area had large intra-annual variability, with a shift from record high levels in May to record low levels in August. Springtime ozone depletion resulted in one of the largest and most persistent Antarctic ozone holes observed since the 1990s. Closer to the equator, 101 named tropical storms were observed in 2015, well above the 1981-2010 average of 82. The eastern/central Pacific had 26 named storms, the most since 1992. The western north Pacific and north and south Indian Ocean basins also saw high activity. Globally, eight tropical cyclones reached the Saffir-Simpson Category 5 intensity level. Overlaying a general increase in the hydrologic cycle, the strong El Niño enhanced precipitation variability around the world. An above-normal rainy season led to major floods in Paraguay, Bolivia, and southern Brazil. In May, the United States recorded its all-time wettest month in its 121-year national record. Denmark and Norway reported their second and third wettest year on record, respectively, but globally soil moisture was below average, terrestrial groundwater storage was the lowest in the 14-year record, and areas in "severe" drought rose from 8% in 2014 to 14% in 2015. Drought conditions prevailed across many Caribbean island nations, Colombia, Venezuela, and northeast Brazil for most of the year. Several South Pacific countries also experienced drought. Lack of rainfall across Ethiopia led to its worst drought in decades and affected millions of people, while prolonged drought in South Africa severely affected agricultural production. Indian summer monsoon rainfall was just 86% of average. Extremely dry conditions in Indonesia resulted in intense and widespread fires during August-November that produced abundant carbonaceous aerosols, carbon monoxide, and ozone. Overall, emissions from tropical Asian biomass burning in 2015 were almost three times the 2001-14 average.
An implementation approach using the ocean color index (OCI)-based chlorophyll-a (Chl-a) algorithm for the Visible Infrared Imaging Radiometer Suite (VIIRS) on the Suomi National Polar-orbiting Partnership (SNPP) has been developed. The OCI Chl-a algorithm for satellite-derived Chl-a data was originally developed by Hu, Lee, and Franz (2012) (J. Geophys. Res., 117, C01011, doi: 01010.01029/02011JC007395) for the Moderate Resolution Imaging Spectroradiometer (MODIS). It uses two Chl-a algorithms, i.e., the color index (CI)-based (reflectance difference-based) algorithm for oligotrophic waters and the usual ocean chlorophyll-type (OCx)-based (reflectance ratio-based) algorithm (e.g., OC3M for MODIS and OC3V for VIIRS), and merges the two algorithms for different Chl-a range applications (named OCI algorithm). In this study, we use the in situ Marine Optical Buoy (MOBY) optics data to demonstrate conclusively that using the CI-based Chl-a algorithm can significantly improve VIIRS Chl-a data over oligotrophic waters with much reduced data noise from instrument calibration and the imperfect atmospheric correction. Using the VIIRS-measured global Chl-a data derived from the Multi-Sensor Level-1 to Level-2 (MSL12) ocean color data processing system, we have developed the CI-based algorithm specifically for VIIRS, and further improved the two Chl-a algorithms merging method using the blue-green reflectance ratio values. Extensive evaluation results show that the new OCI Chl-a algorithm for VIIRS can produce consistent Chl-a data compared with those from the OC3V algorithm. In particular, the data transition between the CI-based and OC3V-based Chl-a algorithm is quite smooth, and there are no obvious discontinuities in VIIRS-derived Chl-a data. The new OCI-based Chl-a algorithm has been implemented in MSL12 for routine production of VIIRS global Chl-a data.