Estimation of phytoplankton pigments from ocean-color satellite observations in the Senegalo–Mauritanian region by using an advanced neural classifier

Abstract

Abstract. We processed daily ocean-color satellite observations to construct a monthly climatology of phytoplankton pigment concentrations in the Senegalo–Mauritanian region. Our proposed new method primarily consists of associating, in well-identified clusters, similar pixels in terms of ocean-color parameters and in situ pigment concentrations taken from a global ocean database. The association is carried out using a new self-organizing map (2S-SOM). Its major advantage is allowing the specificity of the optical properties of the water to be taken into account by adding specific weights to the different ocean-color parameters and the in situ measurements. In the retrieval phase, the pigment concentration of a pixel is estimated by taking the pigment concentration values associated with the 2S-SOM cluster presenting the ocean-color satellite spectral measurements that are the closest to those of the pixel under study according to some distance. The method was validated by using a cross-validation procedure. We focused our study on the fucoxanthin concentration, which is related to the abundance of diatoms. We showed that the fucoxanthin starts to develop in December, presents its maximum intensity in March when the upwelling intensity is maximum, extends up to the coast of Guinea in April and begins to decrease in May. The results are in agreement with previous observations and recent in situ measurements. The method is very general and can be applied in every oceanic region.

Evidencing the Impact of Climate Change on the Phytoplankton Community of the Mediterranean Sea Through a Bioregionalization Approach

“…We followed the algorithm implementation described in Ouattara ( 2014) and Yala et al (2020); we trained 80 different 2S-SOM rectangular maps of size 10 × 5 and chose the one that minimizes the within-group variance at best.…”

“…We used the 2S-SOM algorithm (Outtara, 2014;Yala et al, 2020), which is a modified version of the SOMs (Kohonen, 2013). The SOMs are unsupervised neural network classifiers, which have been commonly used to solve environmental problems (El Hourany, Abboud-Abi Saab, Faour, Aumont, et al, 2019;El Hourany, Abboud-Abi Saab, Faour, Mejia, et al, 2019;Jouini et al, 2013Jouini et al, , 2016Liu & Weisberg, 2005;Liu et al, 2006;Niang et al, 2003Niang et al, , 2006Richardson et al, 2003).…”

“…The block weighting facilitates the clustering procedure by considering the most relevant environmental variables for each cluster (or neuron). The different weights are estimated during the learning phase by minimizing a cost function similar to this used in the SOM estimation where, in addition to the classical cost, the entropy of each cluster is minimized (a complete presentation of the 2S-SOM can be found in Yala et al, 2020). We note that the weight attributed to every block of variable varies from one neuron c to another.…”

“…This block partition permits a better determination of the influence of each environmental block of variables (j) on the classification with respect to the geographical location. For each neuron c, the weights α jc (j = 1, 2, 3, c = 1, 2, …, 50) attributed to each block of variables are determined during the learning phase (Yala et al, 2020). We note that the values of α jc vary between 0 and 1 by construction and that…”

“…In this study, we used a modified Self-Organizing method, called 2S-SOM to provide a robust decomposition of the Mediterranean into bioregions using a well-documented set of variables (SST and Chla given by the satellite observations of MODIS Aqua; MLD from Argo floats in situ observations). The 2S-SOM has been proposed by Yala et al (2020) to retrieve phytoplankton pigment concentration from ocean color satellite data. The 2S-SOM algorithm differs from the classical SOM algorithm by introducing, during the learning phase, a system of weights acting on the most sensitive variables.…”