Two different approaches were initially tested to estimate PPC from Rrs. For the first approach, SPM is estimated by the algorithm (referred to as HA16) which consists of semi-analytical relationships between SPM and Rrs in the red or NIR bands, according to the level of turbidity. The typical band–ratio relationship using the red to blue–green bands is used to assess POC from the algorithm of (referred to as TR19). Once POC and SPM are derived, the POC / SPM ratio is calculated, and PPC is estimated using the different threshold values (see Sect. ). However, showed that POC concentration can be overestimated in the presence of mineral waters. The result is that the POC / SPM ratio is also overestimated when mineral-dominated waters dominate (not shown). Moreover, the errors in both POC and SPM estimations are additive when the POC / SPM ratio is finally calculated. To limit this error propagation, we focus on the development of a single algorithm to derive the POC / SPM ratio from Rrs in one step.

For that purpose, a neural network approach has been selected as a second approach. We used a feed-forward network with log-sigmoid hidden neurons and linear output neurons coupled to a Levenberg–Marquardt algorithm allowing an efficient back propagation through the training procedure . The DS dataset was randomly divided into three datasets to develop, train, and validate this NN. Overall, 60 % of the observations were used to construct and train the NN, 20 % were used for its validation, and 20 % were used to test its performance independently. The training and validation phases are performed jointly, allowing us to stop the training procedure when the generalization of the NN stops improving. We tested several combinations of Rrs bands (412, 443, 490, 510, 560, and 665 nm) to best predict the POC / SPM ratio. The final NN architecture was best trained using Rrs at 412, 490, 510, and 560 nm as the input layer, two hidden layers (8 and 10 neurons), and one output layer (POC / SPM ratio). The metrics used to evaluate the performance of the NN optimization are described in .

As previously explained, the POC / SPM ratio is an indicator of the particle assemblage and can be used to partition in situ data into three water types as defined by . In addition, some theoretical and field studies showed that the variability in the ratio bbp/bp can be related to the particle composition . Low bbp/bp values are observed for a particle population dominated by low refractive-index material such as phytoplankton. In contrast, high bbp/bp values are generally observed in the presence of a relatively high concentration of inorganic particles. At 650 nm, cp is dominated by the scattering, and the optical ratio bbp/cp can be used instead of bbp/bp .

The objective here is to re-examine the pertinence of the POC / SPM threshold values of on a larger in situ coastal dataset (DS0) (covering a wider range of optical and biogeochemical variability) through the examination of the POC / SPM to bbp/cp relationship. As expected, bbp/cp decreases when POC / SPM increases, that is when we move from a mineral-dominated to an organically dominated environment, with bbp/cp values typically lower than 0.010–0.012 (Fig. ). The regression line (bbp/cp vs. POC / SPM; black line) is plotted, and the estimated regression coefficients are indicated with their standard error (Fig. ). The threshold values are first fixed according to the bbp/cp values (as a given range of bbp/cp values corresponds to a given range of the refractive index of the bulk particulate matter) and then adjusted, with a careful examination of each data point for which ancillary data (i.e., chlorophyll a, counted cells, phytoplankton to particulate absorption ratio, and Rrs spectra) are used to better characterize the bulk particulate matter. The first threshold value has been shifted to 0.08 (corresponding to a bbp/cp value of 0.012) to encompass data points collected in mineral-dominated environments (close to river mouths). An asymptote is reached at high POC / SPM values (above 0.2) and concerns data points with low bbp/cp (below 0.075) values typical of phytoplankton-dominated environments. The new threshold value of POC / SPM for organically dominated waters is set to 0.2 (i.e., bbp/cp<0.075) to encompass in situ data points collected during bloom events. This value is, however, very similar to the value previously obtained by (0.25). Data points located along the asymptote are therefore associated with organically dominated waters.

The thresholds are applied to monthly POC / SPM values derived from MERIS data, and the frequency of dominance is computed for each pixel of the scene over the 10 years (2002–2012) as detailed below. A pixel geographically located at a given latitude and longitude is named the k pixel (k=1, Stot). Stot is the total number of pixels over the selected geographical area. For each k pixel, we computed the number of valid pixels (Ntotk) over the period that corresponds to 120 months (Ntotk≤120). The term “valid pixel” means a non-flag pixel for which the Rrs value is provided in order to derive POC / SPM with the neural network. Then, for the k pixel, we calculate how many times the class i is identified over Ntotk: nik is the class occurrence, and i=1,2, or 3 with 1 for mineral-dominated, 2 for organically dominated, and 3 for mixed water classes. The frequency of dominance is defined as the ratio of the occurrence to the number of valid pixels: 1Dk=maximumn1k,n2k,n3kNtotk×100(%). Eq. () is repeated Stot times to compute the frequency of dominance over the whole scene and provide maps in Sect. .

In situ measurements of Rrs from DS are used as input for the neural network algorithm to test the performance of the estimation of algorithm-derived values, named NN POC / SPM (Fig. a). The NN algorithm achieves a good performance over the whole range of POC / SPM (Fig. a). The median absolute percentage difference (MAPD) is 24 %, the median ratio (MR) is 1.06, and the bias is 0.004. The slope of the type II (log-transformed) regression is 0.90. The same classification is obtained for 88.5 % of data points between the in situ and model-derived values. About 10.9 % of data points are misclassified in the adjacent group and only 0.62 % are misclassified in a non-adjacent group. SeaWiFS and MERIS match-ups on both the independent SOMLIT in situ dataset (for which no in situ Rrs values are available) and DS in situ data (for which in situ Rrs values are available) are shown in Fig. b. Compared to the results obtained using in situ measurements (Fig. a), the data points are much more scattered around the 1:1 line when POC / SPM is derived from satellite Rrs (Fig. b). This strongly emphasizes the impact of atmospheric correction on PPC, which is a common feature of OCR products. In Fig. b, all mineral-dominated data are from the SOMLIT dataset. As the in situ Rrs values are not available for SOMLIT data, we cannot apply the criteria described in Sect.  to remove inaccurate satellite Rrs retrievals. Removing such Rrs values will certainly allow us to identify strongly misclassified patterns. In Fig. b, 68.4 % of mineral-dominated data, 44.4 % of mixed data, and 73.7 % of organically dominated data are well-classified. Due to atmospheric correction uncertainties, a proper estimation of POC / SPM values from remote sensing is still very challenging, while the estimation of PPC can still be performed with a reasonable accuracy. Extra match-up data points, including both in situ Rrs and POC / SPM measurements, are, however, needed to definitely support this conclusion.

The PPC algorithm has been applied to MERIS monthly Rrs(λ) data over the 2002–2012 time period, and the maximum occurrence of PPC has been calculated over this 10-year time period as described in Sect.  (Fig. a). As expected, we observe a well-marked near-shore–off-shore gradient from mineral-dominated waters to organically dominated waters with a transition zone corresponding to mixed particulate assemblages. Large deltas and estuaries characterized by intense sediment discharge, such as the Amazon Estuary, Mekong Delta, Huanghe (Yellow River) Delta, Yangtze Estuary, or Ganges–Brahmaputra Delta, present a very high occurrence of mineral-dominated waters (70 %–90 %). These areas can extend much further from the river outlet, depending on the bathymetry and the surface currents, as for instance in the central part of the Yellow Sea at the Yangtze Estuary. Values of POC / SPM of 0.12–0.15, depicting mix-dominated particulate assemblages, have been measured in the East China Sea region (124–126∘ E, 30–32∘ N) in great agreement with our present finding . The dominance of mineral-dominated waters is generally persistent throughout the considered time period as depicted by the relatively low (20 %–30 %) coefficient of variation observed at the outlet of these different estuaries and deltas (Fig. b). In contrast, at the Amazon and Yangtze estuaries and the Huanghe Delta, some off-shore areas with a high (70 %) coefficient of variation depict the impact of seasonal and interannual interactions between the continental (river and sediment discharges) and oceanic forcing (tide, wind, surface current, sub-mesoscale structures) and the bathymetry. For example, close interaction between the Amazon plume dominated by mineral particles and the retroflection of the North Brazil Current carrying oceanic particulate organic matter occurs in the region at the Amazon Estuary where the slope in the continental shelf drastically changes , causing a significant temporal variation in the PPC. In contrast to the previously mentioned rivers and despite the Congo River being the world's second-largest river in terms of both drainage area and water discharge, only a small PPC signature (mainly organically dominated) can be seen at the outlet of its estuary (not visible in Fig. a), except in the coefficient of variation (Fig. b). This is coherent with the fact that the sediment discharge of the Congo River is very low, in contrast to the dissolved part of the discharge, and that most of the particulate assemble is organic . A specific zoom on this area reveals the dominance of organic matter even at the outlet of the estuary, in good agreement with in situ data reported in .

While an extensive investigation of the different spatiotemporal patterns presented in Fig.  are is beyond the scope of the present paper, a specific focus illustrating the pertinence of the PPC product is conducted over the eastern English Channel (EEC) and the southern North Sea (SNS), for which the biogeochemical and physical processes driving the suspended particulate matter composition are relatively well-documented. The eastern English Channel and southern North Sea coastal region is known for its strong hydrodynamics marked by intense tidal currents. This shallow-water region, therefore largely impacted by resuspension effects, is also largely influenced by freshwater discharges from different rivers, the most important ones being the Seine and the Somme along the French coast of the EEC and the Thames and the Scheldt along the SNS. Strong spring and early summer phytoplankton blooms of diatoms and Phaeocystis globosa occurred in this region at different times of the year, depending on the nutrient and light (i.e., season and turbidity level) availability . For these different reasons, the composition of the suspended particulate matter in this area is strongly variable, as already stressed by different field studies based on optical and biogeochemical measurements . The monthly occurrence of each PPC class is in excellent agreement with our present knowledge of this area (Fig. ). Between mid-fall and winter (from October to February), the particulate matter pool of the EEC and the SNS is characterized by mineral and mix-dominated assemblages, mainly due to sediment river inputs and resuspension effects from strong physical (i.e., waves) forcing . This is consistent with the results of , who observed that the relative contribution of the particulate matter assemblage with a strong proportion of mineral particles increases significantly during the winter time. In March, organically dominated pixels appear in the SNS coastal region and at the Somme River area in the EEC. Then, while the organically dominated particle assemblage areas are developing in the SNS from March to July, similar particulate assemblages start developing from May to July in off-shore waters of the EEC and along the EEC coast from June. The timing of the development of organic matter areas in the EEC and the SNS is in perfect agreement with phytoplankton bloom dynamics in these regions, which first develop in the Bay of the Somme area and the SNS, followed by another bloom which develops south in the area of the Bay of the Seine . Finally, areas covered by organically dominated pixels are at a maximum in June and July, and then mixed assemblages start developing in the northern part of the domain to fully cover the whole region in October. Mixed assemblages are more likely dominated by aggregates of mineral particles and detrital matter (non-living organic particles) which developed during the senescence of the bloom .