Open Access
Add to BibSonomy
Abstract
The ocean color algorithm ONNS (OLCI Neural Network Swarm) is designed to process remote-sensing reflectances at 11 Sentinel-3/OLCI bands in order to derive water quality parameters for most natural waters [Hieronymi et al., Front. Mar. Sci. 4, 140 (2017)]. The present work introduces a spectral band-shifting procedure, which allows exploitation of atmospherically corrected input from SeaWiFS, MODIS, MERIS, OCM-2, VIIRS, SGLI, GOCI-2, EnMAP, or PACE/OCI and corresponding utilization of ONNS or other ocean color algorithms. The performance of the band adapters is evaluated in view of diverse optical water types. In the spectral range between 400 and 600 nm, the mean percentage retrieval error is mostly <5%. The band adaptation is a tool for cross-mission Earth observation and uncertainty estimate, as well as for extended possibilities of algorithm validation.
Published by The Optical Society under the terms of the Creative Commons Attribution 4.0 License. Further distribution of this work must maintain attribution to the author(s) and the published article's title, journal citation, and DOI.
1. Introduction
The OLCI Neural Network Swarm (ONNS) is a biogeo-optical algorithm for the retrieval of water quality parameters from atmospherically corrected satellite imagery or in situ radiometric measurements [1]. The aim of the development is to provide a single algorithm that is suitable to all natural waters, from oligotrophic ocean waters to very turbid coastal or highly absorbing inland waters. For this purpose, a fuzzy logic-based optical water type (OWT) classification scheme is applied in conjunction with a set of specific neural networks with task-optimized architectures. The algorithm retrieves concentrations of chlorophyll and suspended matter, the linkage to marine biomass and particulate organic carbon. In addition, ONNS retrieves different inherent optical properties (IOPs) such as the absorption coefficient of colored dissolved organic matter (CDOM) at 440 nm, which can be a tracer for terrestrially originated dissolved organic carbon in seawater. The estimated diffuse attenuation coefficient for downwelling irradiance at 490 nm can be used to alternatively classify water types and quantify the light availability in the upper ocean [2]. Moreover, ONNS provides the Forel-Ule color index, which is a bridge to historical data of the pre-satellite age. All products form optical closure with respect to the underlying training data, which is the basis for an algorithm-inherent uncertainty estimate.
The ONNS algorithm has been designed for data processing from the Ocean and Land Colour Instrument (OLCI) onboard Sentinel-3. Indeed, remote-sensing reflectances, Rrs, at 11 out of 21 OLCI wavebands are input to the algorithm, namely at 400, 412.5, 442.5, 490, 510, 560, 620, 665, 755, 777.5, and 865 nm. The algorithm is a very complex construct, which consists of a customized OWT classification scheme and altogether 58 specialized neural network algorithms (ONNS version 0.8). The comprehensive validation of all ONNS products is a work in progress and corresponding future improvements of the algorithm are likely. Furthermore, all ocean color (OC) algorithms are sensitive to the upstream atmospheric correction; in case of ONNS, this sensitivity is subject to ongoing research too. The question raised if ONNS could be applied to other ocean color sensors too, for example to MERIS, the precursor of OLCI. Ten of the used wavebands are identical, only the 400 nm band is missing in case of MERIS. Other ocean color sensors use partly the same spectral bands, but some bands are displaced or missing. Sensor-specific algorithms are preferable; however, ONNS processor adjustments are very expensive. An appropriate band shifting procedure requires far less efforts. The applicability of ONNS, or different OLCI- or MERIS-specific algorithms, to other OLCI-unlike ocean color data potentially opens new opportunities for an all-water-type-embracing and comprehensive validation, e.g., based on compiled global in situ data [3,4]. Moreover, the multi-sensor usability of an algorithm may increase our understanding of the retrieval uncertainties and could be used to better merge multi-satellite products. Merging time series of ocean color products provided by independent satellite missions supports related biogeochemical and environmental applications by combining temporally overlapping data sets and by increasing data coverage [5].
The present work introduces a set of ten ONNS adapters, which ingest in situ radiometric data or atmospherically corrected reflectance data from different ocean color sensors and provide remote-sensing reflectance at the 11 needed OLCI bands. The band-shifting procedure is based on simulated hyperspectral remote-sensing reflectance data, which are representative of many different water types. The adaptation is realized by means of different neural networks. The adapter performance is statistically assessed and discussed.
2. Methods
2.1 Data basis
Radiative transfer simulations with the commercial software Hydrolight (version 5.2; Sequoia Scientific, USA [6]) are the basis for the ONNS development. Hydrolight computes Rrs and other light field-related quantities from optical specifications of the water body, such as specific absorption and scattering properties. The underlying model assumptions as well as the ranges and co-variances of optically active components are described in Hieronymi et al. [1]. Additional insights regarding chlorophyll-specific absorption of five different phytoplankton groups and their impact on Rrs in the C2X (“Extreme Case-2 Waters”) database are provided by Xi et al. [7]. In brief, the simulations cover the spectral range from 380 to 1,100 nm in 2.5 nm steps, i.e. hyperspectral over the full visible (VIS) and near infrared range (NIR); they comprise clearest natural waters but also extreme absorbing or scattering waters. Considered chlorophyll concentrations, Chl, reach up to 200 mg m−3, total suspended matter concentrations, TSM, up to 1,500 g m−3, and the absorption coefficient of CDOM at 440 nm, acdom(440), is up to 20 m−1. The used reflectance data bear on aligned sun and viewing direction, i.e. fully normalization.
Within the framework of ONNS, 13 optical water types are differentiated [1]. Figure 1 shows their average remote-sensing reflectance spectra. The different spectral shapes and amplitudes are visible and Rrs varies over several orders of magnitude for all used OLCI bands. The phytoplankton-related fluorescence peak is visible in the spectral range between 660 and 710 nm. Indeed, this range must be seen carefully, because the fluorescence quantum yield efficiency may vary significantly and therefore the fluorescence line height varies too. Because of this uncertainty, ONNS does not make use of bands around the 680 nm fluorescence peak; moreover, the proposed band adapter does not expect input from the different sensors from this spectral range. However, for sensitivity tests simulated Rrs data with and without consideration of inelastic scattering effects are utilized.
Fig. 1 Mean hyperspectral remote-sensing reflectance from 13 optical water types, which are differentiated within the ONNS framework [1]. The used OLCI wavebands are marked with dashed lines. Below: Waveband characteristics of historical, current, and future OC sensors.
Apart from the reflectance-based OWT classification, it can be useful to differentiate waters according their tropic state and optical dominance of a water constituent. With regard to the chlorophyll concentration, the trophic classification of lakes differs usually oligotrophic (<2.6 mg m−3), mesotrophic (2.6 to 20 mg m−3), and eutrophic/hyper-eutrophic (>20 mg m−3) waters [8], whereby it should be noted that the value ranges are often smaller for corresponding definitions of oceanic waters, e.g [9]. Indeed, in the case of open ocean, 2 mg m−3 for Chl concentration is high and may be improper to be considered oligotrophic. In addition, distinction is made with respect to optical predominance of water constituents in terms of non-water absorption coefficient at 440 nm, namely 1) phytoplankton-dominated, 2) CDOM-dominated, and 3) sediment-dominated waters. For this study, 1000 simulated reflectance spectra were selected from the C2X database for each of the 3x3 sub-classes according their trophic state and water constituents. Histograms of reflectance spectra per sub-class and ranges are shown in Fig. 2; they illustrate the range of magnitude and approx approximate sub-class mean.
Fig. 2 Sub-classification of remote-sensing reflectance differentiated according to tropic state and absorption dominance of water constituents at 440 nm. The color indicates the occurrence in the test data. Wavebands used by ONNS are marked by red lines. Phytoplankton-dominated waters have phytoplankton absorption coefficients ap(440) of a) 0.005-0.42, d) 0.063-0.9, and g) 0.2-36.2 m−1. CDOM-dominated waters have acdom(440) of b) 0.006-1, e) 0.05-1, and h) 0.2-20 m−1 and NAP-rich waters have aNAP(440) of c) 0.006-6.2, f) 0.067-79.9, and i) 0.22-92.2 m−1, corresponding to TSM of up to 1500 g m−3.
2.2 Independent test data
The performance of the band adapters is demonstrated using in situ reflectance measurements and an atmospherically corrected OLCI matchup. In situ data of similar quality have been collected on six clear sky days between July 2016 and April 2018 aboard a ferry between Cuxhaven and the Island Helgoland in the German Bight, North Sea. Two independent sets of each three hyperspectral RAMSES sensors (TriOS, Germany) have been utilized and installed at the bow of the ship. Individually measured downwelling irradiance, Ed, sky radiance, Lsky, and upwelling surface radiance, Lsurf, have been averaged over each two sensors, 30 seconds in fast sampling mode and were based on the same wavelength, λ, range of 350 to 950 nm in 2.5 nm steps. The radiance sensor viewing zenith angles were 40° and 135° away from the sun. Remote-sensing reflectance above water has been determined from:
where ρ is the surface reflectance factor, which varies significantly depending on wind speed, sea state and sun-viewing geometry [10–12]. Indeed, the factor varied between 2 and 4% already under moderate wind conditions. This variance together with measurement uncertainties are summarized in the provided standard deviation of Rrs. Altogether 746 Rrs spectra have been used for testing.Sentinel-3A/OLCI has taken a matching recording between 09:34:21 and 09:36:21 UTC of the ship-undisturbed area. As the coastal zone is highly heterogeneous, a macro-pixel of only 3x3 pixels has been selected for comparison. Three different atmospheric correction (AC) models have been applied to the scene in order to determine remote-sensing reflectance at the sea surface: 1) the standard AC, IPF in the processing baseline version 2.23, 2) POLYMER version 4.7 [13], and 3) the AC of C2RCC version 1.2 [14].
2.3 Ocean color sensors
In the following, satellite sensors are specified, whose wavebands are primarily used for ocean color purposes. All of the satellites are polar-orbiting and sun-synchronous, only GOCI-2 is on a geostationary space-borne platform. The chronological list includes historical, current, as well as future ocean color missions, partly intended to be integrated in the European Copernicus Marine Environment Monitoring Service (CMEMS) or multi-sensor time series like the ESA Ocean Colour – Climate Change Initiative (OC-CCI) data set [15]. Figure 1 and Table 1 provide an overview of the sensor’s central wavebands, swaths, and spatial resolutions as listed by the International Ocean Colour Coordinating Group [16]. Different bandwidths and sensor-specific spectral response functions are not taken into account for this study, although this might be a potential source of uncertainty. Note that the given central wavelength in some cases deviate in diverse publications.
Table 1. Band adapters with sensor-specific inputa and output at OLCI bands.
SeaWiFS – One of the first ocean-dedicated sensors was the Sea-Viewing Wide Field-of-View Sensor on board of OrbView-2 developed by the US National Aeronautics and Space Administration (NASA) [17]. SeaWiFS was in operational service between 1997 and 2010. The swath of SeaWiFS was 2806 km and the spatial resolution 1100 m. It was the first sensor that contributed data to the OC-CCI time series; reflectances of all follow-up sensors were band-shifted to the SeaWiFS reference bands [15,18]. For band shifting purposes, two setups with input at SeaWiFS bands are investigated: 1) all bands without 670 nm at the fluorescence peak and 2) only the first five SeaWiFS bands until 555 nm, which are utilized in the framework of OC-CCI and often provided within in situ databases.
MODIS – The Moderate Resolution Imaging Spectrometer is developed by NASA too. Two sensors of this kind were launched into space, on board of the satellites Terra (1999) and Aqua (2002) [19]. Up to date, both sensors provide Earth observation data. The swath of MODIS is 2330 km with a resolution of 1000 m.
MERIS – The European Space Agency (ESA) developed the Medium Resolution Imaging Spectrometer on board of Envisat (Environmental Satellite), which served from 2002 to 2012 [20]. The image resolution of MERIS was 300 m for inland and coastal waters and 1200 m for the open sea. The swath was 1150 km.
OCM-2 – The Ocean Colour Monitor onboard the OCEANSAT-2 spacecraft is an Earth observation program of the Indian Space Research Organization (ISRO). The program was initialized through the start of OCM-1/OCEANSAT-1 in 1999; the successor is in service today [e.g 21.]. The sensor’s swath is 1420 km and the resolution is 360 m.
VIIRS – The Visible Infrared Imaging Radiometer Suite is a system developed by NASA and the US National Oceanic and Atmospheric Administration (NOAA). The first VIIRS was launched in 2011 onboard the satellite Suomi NPP (Suomi National Polar-orbiting partnership) and the second in 2017 onboard JPSS (Joint Polar Satellite System) [e.g 22,23.]. VIIRS has a swath of 3000 km and a resolution of 750 m.
OLCI – Based on the heritage of MERIS, ESA developed the Ocean and Land Colour Instrument for the Sentinel-3 series, which is operated by EUMETSAT (Europe) [24]. Sentinel-3A was launched in 2016, Sentinel-3B in 2018, and further Sentinels will follow. OLCI has the same spectral bands as MERIS but beyond that, six additional bands. OLCI generally operates in full resolution mode (300 m) with a swath of 1270 km.
SGLI – The Second-Generation Global Imager on board of the satellite GCOM-C (Global Change Observation Mission – Climate) by the Japanese Aerospace Exploitation Agency (JAXA) operates since 2017 [25]. The swath of SGLI is 1150 km and the resolution is 250 m. It should be noted that in case of the SGLI adapter, a nearly 300 nm wide gap in the red and NIR must be bridged.
GOCI-2 – The first Geostationary Ocean Color Imager onboard the Communication, Ocean and Meteorological Satellite (COMS) has been developed and operated by the Korean Ocean Satellite Center (KOSC) and Korea Institute of Ocean Science and Technology (KIOST) (South Korea) [26]. The launch of a similar satellite, GEO-KOMPSAT-2B, is scheduled for 2020; it will carry GOCI-2. In comparison to its precursor, GOCI-2 has an increased number of bands (13), an improved spatial resolution, and allows more observations (one per hour, 10 times per day). More precisely, the spatial resolution is 250 to 300 m in the local area mode and 1 km in the full disk mode with the center near the equator and 130°E.
EnMAP – The Environmental Mapping and Analysis Program is a German hyperspectral satellite mission that shall be launched in 2020 [27]. The VNIR instrument on board of EnMAP covers the range from 423 to 1000 nm with average spectral sampling interval of 6.5 nm; to be more precise, the minimum bandwidth is 4.8 nm around 490 nm and the maximum is 8.2 at 960 nm. A second spectrometer onboard images the shortwave infrared SWIR from 900 in 10 nm steps to 2450 nm. With respect to the band adapter, the ten bands close to the desired OLCI bands are chosen; although band extrapolation of up to 23 nm towards the 412.5 and 400 nm bands is critical. The image swath is 30 km with 30 m spatial resolution at ground and a revisit time of 27 days (off-nadir four days).
PACE/OCI – The Ocean Color Instrument is one of the primary science spectrometers on board of NASA’s future Plankton, Aerosol, Cloud, ocean Ecosystem (PACE) mission [28]. The launch is scheduled for 2022. The swath of PACE will be 1500 km with a 1 km resolution. The hyperspectral sensor covers the nominal range from 350 to 885 nm at 5 nm resolution. In view of ONNS, OCI/PACE data cover the full range with many same bands; only in three cases, a maximum band shift of 2.5 nm is needed.
2.4 Band shifting procedure
Different approaches for band shifting of multi-spectral reflectance data exist, for example by Mélin and Sclep [18], Pahlevan et al. [29] and Li et al. [30]. The challenges with reference to ONNS are the spectral width of input from 400 to 865 nm and the wanted high optical diversity due to different phytoplankton groups and high concentrations towards extreme waters. Band shifting can be done by means of linear or spline interpolation partly with acceptable accuracy. The band adaptation for some of the sensors or setups, however, requires an extrapolation to 400 nm or into the near infrared, which may lead to unpredictable results.
We propose using multilayer feedforward-backpropagation neural networks (NN) for the band adaptation, a similar approach as used by Pahlevan et al. [29]. This is a multi-dimensional regression method that is based on training data (the C2X data set with simulated hyperspectral reflectances) to predict reflectance at the 11 desired OLCI bands for ONNS applicability. Schiller [31] developed the utilized neural net training software. The algorithm ONNS is based on the same procedure. All input and output reflectances for the NNs are log-transformed, log10(Rrs(λ) + 0.001). The added offset enables the inclusion of Rrs(λ) = 0. Various NN architectures were trained, whereby nets with varying levels of convergence, i.e. over- or under-fitting, have been saved and statistically assessed. The ten selected NN architectures, i.e. input wavelengths of the sensors and hidden layers with numbers of nodes, are specified in Table 1.
2.5 Performance assessment
A selection of 25 promising NNs per sensor setup were statistically evaluated and compared using a relative scoring system similar to the OC-CCI approach [32–34]. Three assessment criteria were considered: 1) How good is the individual Rrs(λ) retrieval?, 2) Which model does preserve the spectral shape of Rrs best?, and 3) Which model does achieve the slightest deviation in the subsequent application of the ONNS algorithm? The first point mirrors the main purpose of the band adapter; in the final scoring, this point is weighted with 50%, whereas the two other points contribute 25% each. Moreover, value was set on good performance at globally common conditions and not extreme Case-2 waters.
Performances of various NN retrievals were considered using Pearson’s correlation coefficient (r), the root mean square error (RMSE), the bias, the slope (m) and intercept (b) of linear regression, the mean absolute error (MAE), and the mean percentage error (MPE) [32,33]. These statistical indicators were subdivided for the comparison with Rrs(λ) per sub-class (trophic state and optical dominance). The same statistics were applied to assess the impacts of the band adaptation on the ONNS retrieval of the diffuse attenuation coefficient at 490 nm, Kd(490), and the total particulate backscattering coefficient at 510 nm, bbp(510), both relatively accurate OC parameters. For this purpose, two “background NNs” for “Case-1” and “Case-2” waters were employed (from the further development of ONNS version 0.8 [1]).
In addition to the tests described above, a chi-square (χ2) test was also used to compare the goodness of fit and performance in preserving the spectral shape of Rrs. The lower the χ2 is, the better the band adapter reproduces the observed reflectance data [33]. A distinction is made here between three spectral ranges: 1) the full range, 2) only bands in the visible range of 400 to 700 nm, and 3) the blue-green region of 442 to 560 nm. The latter range is of particular interest for ONNS and other OC algorithms, like typical band ratio algorithms.
The sensor-specific band adapter NNs with best individual quality according to the scoring scheme were selected (see Table 1). Their retrieval capacity is presented by means of the correlation coefficients and the following statistical metrics:
where M, O, and N represent the modeled value, the observation, and the sample size, respectively. Nλ stands for the number of included wavebands.3. Results
3.1 Reflectance retrieval
Figure 3 and Table 2 document the efficiency of all ten sensor-specific adapters in comparison to simulated reflectances. The sample size is 9000 and the retrieval success was 100% for all models. The data shown in Table 2 refers to the complete set of spectra, yet statistics can vary significantly looking at individual water types. This has to do with the fact that the possible solution of a Rrs-retrieval problem varies over four orders of magnitude.
Fig. 3 Mean Percent Error of the sensor-specific band adapters for Rrs retrieval at ONNS bands (400, 412.5, 442.5, 490, 510, 560, 620, 665, 755, 777.5, and 865 nm). The triangles and circles represents mean under- or overestimation (negative or positive MPE), colors signify the sub-classes shown in Fig. 2, black is the mean value for all waters. The 5% line is marked.
Table 2. Performance (correlation coefficient, mean absolute error, bias, and goodness-of-fit) of band adapters for different sensors. The values correspond to average over all sub-classes (N = 9000) for simulated data.
The most challenging task, and therefore poorest statistical results expected, is the band adaptation from the SeaWiFS OC-CCI setup, where Rrs input is only provided until 555 nm and extrapolation is needed until 865 nm (over 310 nm). In this case, the correlation coefficient seen over all water types is >0.999 for Rrs between 400 and 560 nm, but r decreases to 0.817 at 865 nm, which is still a strong linear relationship. Looking at separate water types, correlation can be near zero (or even slightly negative r) for NIR wavebands, namely in cases with phytoplankton-dominance or extreme absorbing cases (Figs. 2(a), 2(d), 2(g), and 2(h)). Here, the expected Rrs is usually <10−4, but the prediction can be an order of magnitude higher. Similarly, the mean absolute error is worst at the NIR bands. But largest MAE values are found in NAP-dominated high chlorophyll waters (Figs. 2(f) and 2(i)) with MAE of Rrs(865) of up to 0.0076; seen over all waters, this value is only 0.0011.
The retrieval statistics are better for all the other setups. The mean absolute error of the Rrs(490) retrieval is usually <10−4. Even the extrapolation towards 400 nm from the VIS is achievable with good accuracy. In case of EnMAP, a distance of 23 nm must be bridged to retrieve Rrs at 400 nm. Here, the lowest correlation for a subclass is 0.982 (Fig. 2(d)), whereby r = 0.998 seen over all water types and MAE of Rrs(400) is <10−4. The percent deviation for the Rrs(400) retrieval can be can be 27% in eutrophic cases with high Rrs(400) variability (Figs. 2(g) and 2(h)). The bias of the reflectance retrievals is overall very small and mostly negative. Moreover, the goodness-of-fit yields very low values, in particular for the blue-green spectral range, showing that the spectral shape of Rrs is generally well preserved.
Figure 3 shows the mean percentage retrieval error per band, water sub-class, and sensor setup (a similar comparison with linear y-axis is provided in Fig. 4(d)). Red dashed lines mark the desired uncertainty threshold of 5%, originally defined for the blue-green spectral region and oligotrophic waters [e.g 35.]. Again, the magnitudes of Rrs(λ) vary significantly just as the individual error distributions; thus, relative mean errors can be misinterpreted. However, the black symbols indicate the average over all water types and the values are below the 5% threshold for most of the bands for all setups. Strictly speaking, MPE for Rrs in the VIS is often smaller than 1% for various water types. The SeaWiFS OC-CCI setup exhibits significant exceeding of MPE for wavelengths >560 nm (Fig. 3(b)). Moreover, the threshold is noticeably often exceeded in cases with eutrophic waters.
Fig. 4 Band adapter performance at in situ radiometric data and OLCI matchup (July 20, 2016; 09:37 UTC; 53.9905°N, 8.3774°E; Sun zenith angle was 44°; wind speed was 4 m s−1). a) Remote-sensing reflectance at OLCI bands measured in situ (red) and in comparison with three atmospheric correction algorithms for the corresponding matchup: IPF, POLYMER, and C2RCC. The lines show the mean of a 3x3 macro-pixel and the crosses show the estimate from the MERIS band adapter. b) Shows the in situ measured hyperspectral Rrs (red line) with standard deviation (dashed lines) and corresponding estimate of Rrs at ONNS bands from all ten band adapters. c) Shows the percentage error of the retrieved versus initial atmospherically corrected or in situ reflectance. The 5% threshold is marked with dashed lines. d) Shows the corresponding percentage errors for all band adapters.
The in situ measured remote sensing reflectances represent the transition from turbid coastal to relatively clear North Sea waters. The hyperspectral maximum is between 480 and 580 nm with Rrs(560) between 0.004 and 0.0192 sr−1. The retrieval performance of the band adapters in comparison with the in situ data is listed in Table 3. The overall performance is comparable with the simulated data, e.g. for wavelengths between 400 and 560 nm, the correlation is very high with r > 0.99. The two worst correlations, also in comparison with the simulated data, are achieved with the SeaWiFS OC-CCI setup for 865 nm with r = 0.73 and the SGLI setup for the 620 nm band with r = 0.886. These two cases also mark the extreme values of the achieved mean percentage errors of −39 to + 70% (based on low absolute values, see MAE). However, the mean deviation is mostly noticeably lower than 5% for Rrs between 400 and 560 nm.
Table 3. Corresponding statistics of all band adapters applied to in situ data (N = 746)..
3.2 Ocean color product retrieval
The results show that the reflectance shape is well reproduced by the band shifting procedures and that correct absolute values in particular in the visible range are generated (Table 2). This is of prime importance for subsequent in-water algorithms like ONNS, but also any other band ratio, algebraic or optimization model [33]. Like all models, ONNS is sensitive to changes of the input. Uncertainties of the Rrs-retrieval are directly transferred to the ocean color retrieval, but potentially on other scales. For example, Kd(490) varies between 0.02 and 0.17 m−1 when the “Case-1” background neural network of the ONNS processor is applied and between 0.02 and 12 m−1 for the corresponding “Case-2” NN. In the worst case of the SeaWiFS OC-CCI setup, the correlation coefficients are 0.942 and 0.729 for the Kd(490) retrieval in Case-1 and Case-2 respectively (Table 2). Accordingly MAE of Kd(490) are higher than those of Rrs, namely 0.009 and 1.066 m−1. The biases of all examples with bbp(510) and Kd(490) are mostly moderately positive, suggesting potential slight overestimation of these values. Overall, the results suggest that the band adaptation procedure is an additional source of uncertainties for the ocean color retrieval, whereby the significance depends on the region (water type) of interest and individual cases.
4. Discussion
4.1 Context with atmospheric correction
The main purpose of the band adapters is providing reflectance at needed bands. Impacts on in-water retrieval of water constituents and optical properties are subordinated and possibly subject to algorithm tuning. However, it is in the nature of things that the adapters modify the spectra and the question is, how weighty the changes are in comparison with other sources of uncertainties. In order to illuminate this, we compare in situ data with satellite-derived reflectance (Fig. 4(a)). The in situ reflectance measurements were conducted in the heavily navigated coastal region at the mouth of the Elbe, where tides, river plume, and shallow waters control the mixing of water masses. This is typical “Case-2” water, usually with chlorophyll concentrations of 0.8 to 10 mg m−3, suspended particle concentrations of 0.1 to 25 g m−3 (in the river much higher), and CDOM absorption at 440 nm of 0.15 to 0.8 m−1. In view of the water type classification, this corresponds to NAP-dominated oligotrophic to mesotrophic waters (Figs. 2(c) and 2(f)), or in view of the ONNS optical water type scheme, OWT 5 (Fig. 1).
Figure 4(a) shows a comparison of in situ measured remote-sensing reflectance, Rrs(40°,135°) at ONNS bands (red line) with satellite-derived atmospherically corrected reflectance (gray lines). For the sake of clarity, only the mean spectra of the 3x3 macro-pixel are shown. Note that in this case, OLCI’s spectral response functions are applied to the in situ spectrum, whereas this is not done in Fig. 4(b), where the original hyperspectral Rrs is displayed, which shows a spectral maximum around 570 nm and remarkable spectral features with local minima at 610, 667.5, and 735 nm. Overall, the agreement of AC reflectance with in situ data is not bad for this kind of waters [e.g 32,36.], although only at the spectral maximum at band six and AC-wise at some red/NIR bands, the percentage deviation complies with the uncertainty threshold (Fig. 4(c), circles). IPF overcorrects in the blue spectral range and partly provides physically wrong negative reflectances, but the deviations are small (<5% error) and partly lowest from band six on. POLYMER retrieves the reflectances most accurately in the range from 400 to 510 nm, but still with deviations of up to 16.5%. Seen over the whole spectrum, the percentage errors are between −105 and 2% for IPF, between −19 and 16.5% for POLYMER, and between −21.5 and 42.5% for C2RCC.
The band adapter-generated reflectances from atmospherically corrected and in situ data are marked with plus signs in Figs. 4(a) and 4(c). Here the MERIS adapter is used, where only the 400 nm band is missing, i.e. input to the adapter starts at band two. The band adapter can deal with negative Rrs to some extent (IPF), but has in this case larger deviations in the green bands. If these negative values are simply set to 10−3 (sr−1), then the band adapter produces a spectrum that is mostly closer to the in situ spectrum with lower mean deviation. In the range between 510 and 777.5 nm the mean deviation is <5% after band-shifting in comparison with the original IPF spectrum. In this one example, it seems that the band adapter-treated spectra are generally closer to the in situ spectrum over the full spectrum and for all three AC processors (compare circles and plus signs in Fig. 4(c). Nevertheless, exceptions exist for example with band 17, where the deviation increases after band adaptation until −16% based on an in situ value of Rrs(865) < 0.0004 sr−1. In comparison with the in situ spectrum, the retrieved Rrs(400) is underestimated by 6%, the following bands until 753 have lower deviation than 5%, and the other NIR bands exhibit again larger deviations until −22% at 865 nm. These are similar deviations as expected for NAP-dominated waters (Figs. 2(c), 2(f), and 3(d)).
The comparison suggests that the intermediate step with band adapter application to atmospherically corrected data could potentially improve the Rrs spectrum by weakening of unfavorable assumptions underlying the atmospheric correction. This thesis is supported by the fact that, in this example, the retrieved chlorophyll estimate by ONNS with input from the three correction methods lies closer to each other after band adapter application (the mean Chl is approximately 2 mg m−3 and the NAP-dominance is confirmed). Anyway, the comparison clearly shows that uncertainties due to selection of an atmospheric correction model is much higher than uncertainty introduced by application of the band adapter. Table 2 provides a first hint on the changes in the ONNS products Kd(490) and bbp(510) due to band adapter application. The choice of an atmospheric correction model generates larger deviations. The net errors and uncertainty propagation due to atmospheric correction plus band shifting should be focus of future investigations.
Figures 4(b) and 4(d) show the performance of all band adapters, each with input at different bands, in comparison with the hyperspectral match up in situ spectrum. The red dashed lines mark the standard deviation of the measurements. The correction of reflected blue skylight increases the uncertainty in the blue spectral range and beyond. The maximum retrieval deviation for Rrs(400) can be found in the EnMAP setup with −8.3%, where the largest extrapolation has to be made from the VIS; this error is still significantly smaller than the deviation due to measurement uncertainties. The percent error is generally lower than 5% for OLCI bands two to eight, with exception of the SGLI and SeaWiFS OC-CCI setups, which show significant deviations beyond 560 nm, where both setups have no supporting input. The magnitude of errors for the NIR bands is comparable with the errors due to atmospheric correction. The comparison with all in situ data yields similar conclusions (Table 3). All in all, the achieved retrieval accuracy is acceptable for most water types (Fig. 3) and in particular for the visible range; furthermore, the statistics are comparable with other approaches [18,29].
4.2 Spectral blurring
Application of the band adapters on atmospherically corrected satellite data may be of benefit for the ocean color retrieval; on the one hand through possible removal of biogeo-optically unjustified spectral features and on the other hand by providing a rather known spectrum to the in-water algorithm. The approach could have, however, disadvantages for the potential discrimination of phytoplankton groups due to blurring of real spectral features and possible forcing of the provided reflectance shape into the range of variability displayed in the training data. This regards in particular hyperspectral input from EnMAP or PACE/OCI. However, most of the multi-spectral sensors do not provide sufficient spectral information for statistically firm retrieval of phytoplankton functional types [37]. The range between 500 and 600 nm, where the band adapters perform generally well, is particularly affected by the influence of chlorophyll-specific absorption and therefore interesting for phytoplankton group differentiation. That applies for moderately turbid waters too, where sediment absorption and scattering do mask chlorophyll features over wide range of the visible [7,38]. High variability of chlorophyll-specific absorption is actually included in the training (and test) database, including that of cyanobacteria and green algae with counteractive impact on Rrs in the visible range [1,38]. This diversity is partly responsible for the worse performance of all band adapters for eutrophic waters (Fig. 3). The spectral essence of the retrieved reflectance at ONNS bands resembles an average phytoplankton absorption of the “brown algae group” [1], a globally very common absorption shape, similar to the standard phytoplankton absorption spectrum used in Hydrolight and therefore, likely basis for other band shifting approaches [18,29]. With the presented work, it is generally possible to better reproduce the original spectral shape, even in more turbid waters. In view of algae detection, this can be superior to approaches, where only one chlorophyll-specific absorption is considered. This will be subject to future investigations.
4.3 Uncertainties assessment
Looking at the total sum of squared errors, which is also an important indicator of NN training progress, yields the order of best performance of the band adapters: 1) PACE/OCI (0.7), 2) MERIS and GOCI-2 (both 0.8), 4) MODIS and OCM-2 (both 1.2), 6) EnMAP, SeaWiFS, and SGLI (1.4 each), 9) VIIRS (1.7), and 10) SeaWiFS OC-CCI (56.4). This basically mirrors the waveband distances to bridge and the potential uncertainties due to extrapolation. With regard to hyperspectral reflectance from PACE/OCI, fairly stated very good band shifting for ONNS usage can be achieved by simple linear or spline interpolation, which obviously takes much less effort; only three bands must be interpolated, the others are directly provided. Similarly, only the 400 nm band must be provided in case of MERIS input. However, the band adapting approach provides input for the ONNS algorithm with rather familiar spectral shapes and corresponding co-variances of IOPs. This capacity can be of advantage in case of inaccurate atmospheric correction. Sensitivity tests of the ONNS performance with options of replacing 1) only missing bands or 2) the full spectrum will be subject of future investigations.
The spectral (back-) scattering behavior of marine particles is poorly explored in comparison with absorption [e.g 39.]; this regards in particular the NIR spectral region. Basis for the used radiative transfer simulations are idealized spectral IOPs. Whilst living phytoplankton has no significant absorption beyond 700 nm, non-living particles do have; NAP absorption in the NIR can be in the order of the reflectance-dominating seawater absorption [40]. The vague model-assumption for (back-) scattering on the contrary foresees exponential decay (without spectral features) for all particles, but significant contribution to the NIR total attenuation. This can be seen in the reflectance signal for sediment-rich or highly productive waters (Fig. 2), where according the simulations, NIR reflectance varies over four orders of magnitude. Apart from that, all in situ radiometric measurements exhibit increased uncertainties in the red and near-infrared spectral range [35]. The dropping statistics of NIR Rrs retrieval for waters with higher concentrations of marine particles and for all sensors must be seen in this context. The statistical deviation and uncertainty increases in particular for far spectral extrapolation of up to 310 nm (from 555 to 865 nm) as in case of the SeaWiFS OC-CCI setup. For some subclasses, MAE of Rrs(865) <10%, but for other waters, MAE exceeds 1000% (Fig. 3(b), related to Fig. 2(d)). For the above reasons, cases with optical dominance of phytoplankton absorption are most problematic. Obviously, the retrieval is easier, if the spectral maximum of Rrs is within the initially provided spectrum; in case of the OC-CCI setup, input is gathered until 555 nm, which for example is not provided in the in situ spectrum. However, the retrieved values from 560 nm on are in a reasonable shape (Fig. 4b). The optimum performance of this adapter is achieved in oligotrophic waters, with negligible low reflectance in the NIR, i.e. quasi “black pixel” assumption. If substantial NIR values are still present in such waters, this is a sign of insufficient atmospheric correction, i.e. the “bright pixel correction” part, or residual sky reflection. Non-coastal waters are the primary application area of the OC-CCI merged satellite data set [15]. Nevertheless, this particular band adapter needs to be treated with caution and further investigations are needed. Anyway, looking over all reference data, a strong positive linear relationship can be achieved for the retrieval of Rrs at all bands (r is mostly >0.8).
Beside the already mentioned sources (Fig. 4), it is supposed that the following points induce similar or larger uncertainties as the band adaptation procedure: 1) insufficient averaging of measured underwater light field quantities due to wave-focusing effect and extrapolation towards the sea surface [35,41], 2) wind-depending sea surface roughness and corresponding diffusion of the underwater light field in comparison with the moderate-wind-assumptions underlying the ocean color and atmospheric correction algorithm development [e.g 1,13,42.], 3) enhanced spectral reflectance (in particular in the NIR) due to whitecaps, foam, and bubbles in the water column [43–45], 4) impacts of spatial sampling and adjacency effects in coastal waters [46,47], and 5) mission-specific biases, degradation, vicarious and cross-calibration, and related regional differences [48–51]. In the end, all contributing uncertainties should be transferred to the overall uncertainty provided to each ocean color product. This needs corresponding improvement in the ONNS processing scheme.
Looking at enhanced uncertainties at the edge of the used spectrum raises the question, why use all of these bands for retrieval of water quality parameters? Figure 2 shows the high natural variability of marine reflectance. The objective of ONNS was to retrieve information for all possible water types, independent of technical issues to providing reliable input. Light at 400 nm is at the edge of the visible range und part of the photosynthetically available radiation (PAR). For example, in clear oligotrophic waters, half of the surface Ed(400) is still available at 10 m depth with corresponding information content of the water column to Rrs(400). Whereas, the red and NIR spectral ranges are of relevance for biomass estimate in productive waters and possibly the only range with significant Chl-related signal, e.g. in CDOM-rich “black waters” [52]. Furthermore, the NIR range contains valuable information for turbid waters on the concentration of suspended particulate matter. It is not said that the ONNS algorithm is fully matured or that the used bands contain best or independent information. However, this work is done for a larger comparability and exploitability of synergetic multi-mission data. Band shifting is a powerful tool for providing necessary in situ data at the right wavebands for validation purposes. The use of band adapters will increase the number of available validation data [e.g 3,4.].
5. Conclusions
A series of ten band adapters has been developed in order to provide remote-sensing reflectance from different satellite missions at 11 Sentinel-3/OLCI wavebands from 400 to 865 nm. Input is possible from historical, current, and future ocean color sensors: SeaWiFS, MODIS, MERIS, OCM-2, VIIRS, SGLI, GOCI-2, EnMAP, PACE/OCI, and an additional setup with only the first five SeaWiFS bands, as provided in long-term data sets of the OC-CCI. Bands around the phytoplankton fluorescence peak are not utilized due to presumed uncertainties. The band adapters are based on neural networks and simulated hyperspectral reflectances, valid for a large variability of optical water types, from clear oceanic waters to extreme absorbing or scattering Case-2 waters. Looking at all natural waters, Rrs between 400 and 600 nm (visible range without red) varies typically over two orders of magnitude, whereas reflectance beyond 600 nm (in the red and NIR) varies over more than four orders of magnitude. This mirrors in the retrieval performance of the band adapters, which is more accurate in the visible range but drops towards the NIR. Between 400 and 600 nm, all band adapters provide Rrs usually with mean percentage error below the 5% uncertainty threshold, viewed over all water types. However, there are increased uncertainties for eutrophic/hyper-eutrophic waters with Chl >20 mg m−3, e.g. intense algae blooms or partly lake conditions. This is connected above all with the intensified spectral response due to high variability of chlorophyll-specific absorption of diverse phytoplankton groups. Overall, the accuracy of all band adapters is acceptable with good statistic agreement and preservation of the spectral input shape. Moreover, the performance (in the VIS) is comparable with other band-shifting approaches [18,29] but valid for extended optical water types. By means of these tools, it will now be possible to process multi-mission input with the in-water algorithm ONNS or alternatively, other MERIS or OLCI-specific ocean color algorithms. Beyond that, the adapters facilitate the exploitation of huge in situ databases [3,4] for validation purposes of OLCI algorithms.
Funding
The German Aerospace Center (DLR) provided funding for this work in the framework of the EnMAP scientific preparation program (50EE1718). Furthermore, this work is a contribution to the European Copernicus Marine Environment Monitoring Service (77-CMEMS-TAC-OC_HZG) and potentially for ESA’s Ocean Colour – Climate Change Initiative (OC-CCI).
Acknowledgments
I thank my colleagues Hajo Krasemann and Rüdiger Röttgers for helpful discussions. The HZG Remote Sensing Team and Crew of “MS Helgoland” are gratefully acknowledged for assisting in data collection during our regular Helgoland transects. I also acknowledge Vittorio Brando (CNR, Italy) for initialization of this work in the framework of CMEMS. Moreover, I thank two anonymous reviewers for their detailed and conscientious comments.
References
1. M. Hieronymi, D. Müller, and R. Doerffer, “The OLCI Neural Network Swarm (ONNS): A bio-geo-optical algorithm for open ocean and coastal waters,” Front. Mar. Sci. 4, 140 (2017). [CrossRef]
2. N. G. Jerlov, “Optical oceanography,” (Elsevier, 1968 (Vol. 5)).
3. B. Nechad, K. Ruddick, T. Schroeder, K. Oubelkheir, D. Blondeau-Patissier, N. Cherukuru, V. Brando, A. Dekker, L. Clementson, A. C. Banks, S. Maritorena, P. J. Werdell, C. Sá, V. Brotas, I. Caballero de Frutos, Y.-H. Ahn, S. Salama, G. Tilstone, V. Martinez-Vicente, D. Foley, M. McKibben, J. Nahorniak, T. Peterson, A. Siliò-Calzada, R. Röttgers, Z. Lee, M. Peters, and C. Brockmann, “CoastColour round robin data sets: A database to evaluate the performance of algorithms for the retrieval of water quality parameters in coastal waters,” Earth Syst. Sci. Data 7(2), 319–348 (2015). [CrossRef]
4. A. Valente, S. Sathyendranath, V. Brotas, S. Groom, M. Grant, M. Taberner, D. Antoine, R. Arnone, W. M. Balch, K. Barker, R. Barlow, S. Bélanger, J.-F. Berthon, S. Beşiktepe, V. Brando, E. Canuti, F. Chavez, H. Claustre, R. Crout, R. Frouin, C. García-Soto, S. W. Gibb, R. Gould, S. Hooker, M. Kahru, H. Klein, S. Kratzer, H. Loisel, D. McKee, B. G. Mitchell, T. Moisan, F. Muller-Karger, L. O’Dowd, M. Ondrusek, A. J. Poulton, M. Repecaud, T. Smyth, H. M. Sosik, M. Twardowski, K. Voss, J. Werdell, M. Wernand, and G. Zibordi, “A compilation of global bio-optical in situ data for ocean-colour satellite applications,” Earth Syst. Sci. Data 8, 235–252 (2016). [CrossRef]
5. F. Mélin, G. Zibordi, and S. Djavidnia, “Merged series of normalized water leaving radiances obtained from multiple satellite missions for the Mediterranean Sea,” Adv. Space Res. 43(3), 423–437 (2009). [CrossRef]
6. C. D. Mobley, “Light and water: radiative transfer in natural waters,” (Academic, 1994).
7. H. Xi, M. Hieronymi, H. Krasemann, and R. Röttgers, “Phytoplankton group identification using simulated and in situ hyperspectral remote sensing reflectance,” Front. Mar. Sci. 4, 272 (2017). [CrossRef]
8. R. E. Carlson and J. Simpson, J., “A coordinator’s guide to volunteer lake monitoring methods,” North American Lake Management Society 96, 305 (1996).
9. A. Bricaud, H. Claustre, J. Ras, and K. Oubelkheir, “Natural variability of phytoplanktonic absorption in oceanic waters: Influence of the size structure of algal populations,” J. Geophys. Res. Oceans 109, C11 (2004).
10. C. D. Mobley, “Estimation of the remote-sensing reflectance from above-surface measurements,” Appl. Opt. 38(36), 7442–7455 (1999). [CrossRef] [PubMed]
11. C. D. Mobley, “Polarized reflectance and transmittance properties of windblown sea surfaces,” Appl. Opt. 54(15), 4828–4849 (2015). [CrossRef] [PubMed]
12. M. Hieronymi, “Polarized reflectance and transmittance distribution functions of the ocean surface,” Opt. Express 24(14), A1045–A1068 (2016). [CrossRef] [PubMed]
13. F. Steinmetz, P. Y. Deschamps, and D. Ramon, “Atmospheric correction in presence of sun glint: application to MERIS,” Opt. Express 19(10), 9783–9800 (2011). [CrossRef] [PubMed]
14. C. Brockmann, R. Doerffer, M. Peters, K. Stelzer, S. Embacher, and A. Ruescas, “Evolution of the C2RCC neural network for Sentinel 2 and 3 for the retrieval of ocean colour products in normal and extreme optically complex waters,” In Proc. Living Planet Symposium, ESA SP 740, 54 (2016).
15. S. Sathyendranath, B. Brewin, D. Müeller, R. Doerffer, H. Krasemann, F. Mélin, C. Brockmann, N. Fomferra, M. Peters, M. Grant, F. Steinmetz, P.-Y. Deschamps, J. Swinton, T. Smyth, J. Werdell, B. Franz, S. Maritorena, E. Devred, Z. P. Lee, C. Hu, and P. Regner, “Ocean Colour Climate Change Initiative - Approach and initial results,” Proc. IEEE Int. Geoscience and Remote Sensing Symposium (IGARSS), 2024–2027 (2012). [CrossRef]
16. International Ocean Colour Coordinating Group, http://www.ioccg.org.
17. S. B. Hooker, E. R. Firestone, W. E. Esaias, G. C. Feldman, W. W. Gregg, and C. R. McClain, “An overview of SeaWiFS and ocean color,” SeaWiFS Technical Report Series, Vol. 1, (1992).
18. F. Mélin and G. Sclep, “Band shifting for ocean color multi-spectral reflectance data,” Opt. Express 23(3), 2262–2279 (2015). [CrossRef] [PubMed]
19. W. E. Esaias, M. R. Abbott, I. Barton, O. B. Brown, J. W. Campbell, K. L. Carder, D. K. Clark, R. H. Evans, F. E. Hoge, H. R. Gordon, W. M. Balch, R. Letelier, and P. J. Minnett, “An overview of MODIS capabilities for ocean science observations,” IEEE Trans. Geosci. Remote Sens. 36(4), 1250–1265 (1998). [CrossRef]
20. M. Rast, J. L. Bezy, and S. Bruzzi, “The ESA Medium Resolution Imaging Spectrometer MERIS a review of the instrument and its mission,” Int. J. Remote Sens. 20(9), 1681–1702 (1999). [CrossRef]
21. P. V. Nagamani, P. Chauhan, and R. M. Dwivedi, “Development of chlorophyll-a algorithm for ocean colour monitor onboard OCEANSAT-2 satellite,” IEEE Geosci. Remote Sens. Lett. 5(3), 527–531 (2008). [CrossRef]
22. C. Cao, J. Xiong, S. Blonski, Q. Liu, S. Uprety, X. Shao, Y. Bai, and F. Weng, “F, “Suomi NPP VIIRS sensor data record verification, validation, and long‐term performance monitoring,” J. Geophys. Res. Atmos. 118(20), 11–664 (2013). [CrossRef]
23. M. Wang, X. Liu, L. Tan, L. Jiang, S. Son, W. Shi, K. Rausch, and K. Voss, “Impacts of VIIRS SDR performance on ocean color products,” J. Geophys. Res. Atmos. 118(18), 10–347 (2013). [CrossRef]
24. C. Donlon, B. Berruti, A. Buongiorno, M. H. Ferreira, P. Féménias, J. Frerick, P. Goryl, U. Klein, H. Laur, C. Mavrocordatos, J. Nieke, H. Rebhan, B. Seitz, J. Stroede, and R. Sciarra, “The global monitoring for environment and security (GMES) sentinel-3 mission,” Remote Sens. Environ. 120, 37–57 (2012). [CrossRef]
25. K. Ogata, M. Toratani, and H. Murakami, “GCOM-C/SGLI Level-2 ocean color products generation,” In Proc. IEEE Geoscience and Remote Sensing Symposium (IGARSS), 5648–5649 (2017). [CrossRef]
26. J. K. Choi, Y. J. Park, J. H. Ahn, H. S. Lim, J. Eom, and J. H. Ryu, “GOCI, the world’s first geostationary ocean color observation satellite, for the monitoring of temporal variability in coastal water turbidity,” J. Geophys. Res. Oceans 117, C9 (2012).
27. L. Guanter, H. Kaufmann, K. Segl, S. Foerster, C. Rogass, S. Chabrillat, T. Kuester, A. Hollstein, G. Rossner, C. Chlebek, C. Straif, S. Fischer, S. Schrader, T. Storch, U. Heiden, A. Mueller, M. Bachmann, H. Mühle, R. Müller, M. Habermeyer, A. Ohndorf, J. Hill, H. Buddenbaum, P. Hostert, S. Van der Linden, P. J. Leitão, A. Rabe, R. Doerffer, H. Krasemann, H. Xi, W. Mauser, T. Hank, M. Locherer, M. Rast, K. Staenz, and B. Sang, “The EnMAP spaceborne imaging spectroscopy mission for earth observation,” Remote Sens. 7(7), 8830–8857 (2015). [CrossRef]
28. NASA’s PACE mission, http://www.pace.oceansciences.org/oci.htm.
29. N. Pahlevan, B. Smith, C. Binding, and M. O’Donnell, “Spectral band adjustments for remote sensing reflectance spectra in coastal/inland waters,” Opt. Express 25(23), 28650–28667 (2017). [CrossRef]
30. Z. Li, D. Sun, Z. Qiu, H. Xi, S. Wang, W. Perrie, Y. Li, and B. Han, “A reconstruction method for hyperspectral remote sensing reflectance in the visible domain and applications,” J. Geophys. Res. Oceans 123(6), 4092–4109 (2018). [CrossRef]
31. H. Schiller, “Feedforward-Backpropagation Neural Net Program ffbp1. 0,” Report GKSS 2000/37 (2000).
32. D. Müller, H. Krasemann, B. Brewin, C. Brockmann, P.-Y. Deschamps, R. Doerffer, N. Fomferra, B. Franz, M. Grant, S. Groom, F. Mélin, T. Platt, P. Regner, S. Sathyendranath, F. Steinmetz, and J. Swinton, “The ocean colour climate change initiative: I. A methodology for assessing atmospheric correction processors based on in-situ measurements,” Remote Sens. Environ. 162, 242–256 (2015). [CrossRef]
33. R. J. W. Brewin, S. Sathyendranath, D. Müller, C. Brockmann, P.-Y. Deschamps, E. Devred, R. Doerffer, N. Fomferra, B. Franz, M. Grant, S. Groom, A. Horseman, C. Hu, H. Krasemann, Z. P. Lee, S. Maritorena, F. Mélin, M. Peters, T. Platt, P. Regner, T. Smyth, F. Steinmetz, J. Swinton, J. Werdell, and G. N. White III, “The ocean colour climate change initiative: III. a round-robin comparison on in-water bio-optical algorithms,” Remote Sens. Environ. 162, 271–294 (2015). [CrossRef]
34. B. N. Seegers, R. P. Stumpf, B. A. Schaeffer, K. A. Loftin, and P. J. Werdell, “Performance metrics for the assessment of satellite data products: an ocean color case study,” Opt. Express 26(6), 7404–7422 (2018). [CrossRef] [PubMed]
35. G. Zibordi and K. J. Voss, “In situ optical radiometry in the visible and near infrared,” In Experimental Methods in the Physical Sciences, Vol. 47, pp. 247–304 (Academic, 2014).
36. P. Qin, S. G. Simis, and G. H. Tilstone, “Radiometric validation of atmospheric correction for MERIS in the Baltic Sea based on continuous observations from ships and AERONET-OC,” Remote Sens. Environ. 200, 263–280 (2017). [CrossRef]
37. A. Wolanin, M. A. Soppa, and A. Bracher, “Investigation of spectral band requirements for improving retrievals of phytoplankton functional types,” Remote Sens. 8(10), 871 (2016). [CrossRef]
38. H. Xi, M. Hieronymi, R. Röttgers, H. Krasemann, and Z. Qiu, “Hyperspectral differentiation of phytoplankton taxonomic groups: a comparison between using remote sensing reflectance and absorption spectra,” Remote Sens. 7(11), 14781–14805 (2015). [CrossRef]
39. IOCCG. Protocol Series, “Inherent optical property measurements and protocols: absorption coefficient,” A. R. Neeley, and A. Mannino, (eds.), IOCCG Ocean Optics and Biogeochemistry Protocols for Satellite Ocean Colour Sensor Validation, Volume 1.0, IOCCG, Dartmouth, NS, Canada (2018).
40. R. Röttgers, C. Dupouy, B. B. Taylor, A. Bracher, and S. B. Woźniak, “Mass‐specific light absorption coefficients of natural aquatic particles in the near‐infrared spectral region,” Limnol. Oceanogr. 59(5), 1449–1460 (2014). [CrossRef]
41. M. Hieronymi, A. Macke, and O. Zielinski, “Modeling of wave-induced irradiance variability in the upper ocean mixed layer,” Ocean Sci. 8(2), 103–120 (2012). [CrossRef]
42. M. Hieronymi and A. Macke, “On the influence of wind and waves on underwater irradiance fluctuations,” Ocean Sci. 8(4), 455–471 (2012). [CrossRef]
43. D. Stramski and J. Tegowski, “Effects of intermittent entrainment of air bubbles by breaking wind waves on ocean reflectance and underwater light field,” J. Geophys. Res. Oceans 106(C12), 31345–31360 (2001). [CrossRef]
44. K. D. Moore, K. J. Voss, and H. R. Gordon, “Spectral reflectance of whitecaps: Their contribution to water‐leaving radiance,” J. Geophys. Res. Oceans 105(C3), 6493–6499 (2000). [CrossRef]
45. H. M. Dierssen, “Hyperspectral measurements, parameterizations, and atmospheric correction of whitecaps and foam from visible to shortwave infrared for ocean color remote sensing,” Front. Earth Science 7, 14 (2019). [CrossRef]
46. N. Pahlevan, S. Sarkar, and B. A. Franz, “Uncertainties in coastal ocean color products: Impacts of spatial sampling,” Remote Sens. Environ. 181, 14–26 (2016). [CrossRef]
47. B. Bulgarelli and G. Zibordi, “On the detectability of adjacency effects in ocean color remote sensing of mid-latitude coastal environments by SeaWiFS, MODIS-A, MERIS, OLCI, OLI and MSI,” Remote Sens. Environ. 209, 423–438 (2018). [CrossRef] [PubMed]
48. E. J. Kwiatkowska, B. A. Franz, G. Meister, C. R. McClain, and X. Xiong, “Cross calibration of ocean-color bands from moderate resolution imaging spectroradiometer on Terra platform,” Appl. Opt. 47(36), 6796–6810 (2008). [CrossRef] [PubMed]
49. G. Meister, B. A. Franz, E. J. Kwiatkowska, and C. R. McClain, “Corrections to the calibration of MODIS Aqua ocean color bands derived from SeaWiFS data,” IEEE Trans. Geosci. Remote Sens. 50(1), 310–319 (2012). [CrossRef]
50. J. H. Ahn, Y. J. Park, W. Kim, B. Lee, and I. S. Oh, “Vicarious calibration of the geostationary ocean color imager,” Opt. Express 23(18), 23236–23258 (2015). [CrossRef] [PubMed]
51. F. Mélin, G. Sclep, T. Jackson, and S. Sathyendranath, “Uncertainty estimates of remote sensing reflectance derived from comparison of ocean color satellite data sets,” Remote Sens. Environ. 177, 107–124 (2016). [CrossRef]
52. T. Kutser, B. Paavel, C. Verpoorter, M. Ligi, T. Soomets, K. Toming, and G. Casal, “Remote sensing of black lakes and using 810 nm reflectance peak for retrieving water quality parameters of optically complex waters,” Remote Sens. 8(6), 497 (2016). [CrossRef]
