Reconstruction of satellite chlorophyll images under heavy cloud coverage using a neural classification method

We present a method able to fill large data gaps in satellite chlorophyll (CHL) images, based on the principle of Self-Organizing Maps (SOM) classification methods. The method makes use of complementary oceanic remote sensing observations: sea surface temperature (SST) and sea surface height (SSH). It relies on the assumption that a state of the ocean can be locally defined by its values of SST, SSH and CHL, and that a codebook of possible (SST, SSH, CHL) situations, if large enough, allows the reconstruction of incomplete situations of CHL. To account for the spatio-temporal context, each situation is characterized by the values of the three variables CHL, SST and SSH over a 3 × 3 spatial window surrounding the pixel, and at three successive times. In a first step, the feasibility and robustness of the method was assessed on a synthetic data set derived from a high-resolution coupled bio-physical ocean model. The SOM reconstruction method was particularly powerful in the case of intense cloud cover and to reconstruct CHL at scales larger than ~ 10 km. Subsequently, the method was adapted to real satellite data, which required introducing iterative learning and dealing with multi-modal determination of the referent vectors. The proposed improvements of the method proved satisfactory in retrieving CHL missing values under heavy cloud coverage and in improving the quality of weekly CHL products.