Résumé
We propose a statistical algorithm to assess chlorophyll-a concentration ([chl-a]) using remote sensing reflectance (Rrs) derived from MODerate Resolution Imaging Spectroradiometer (MODIS) data. This algorithm is a combination of two models: one for low [chl-a] (oligotrophic waters) and one for high [chl-a]. A satellite pixel is classified as low or high [chla] according to the Rrs ratio (488 and 555 nm channels). If a pixel is considered as a low [chl-a] pixel, a log-linear model is applied; otherwise, a more sophisticated model (Support Vector Machine) is applied. The log-linear model was developed thanks to supervised learning on Rrs and [chl-a] data from SeaBASS and more than 15 campaigns accomplished from 2002 to 2010 around New Caledonia. Several models to assess high [chl-a] were also tested with statistical methods. This novel approach outperforms the standard reflectance ratio approach. Compared with algorithms such as the current NASA OC3, Root Mean Square Error is 30% lower in New Caledonian waters.