Sea-ice cloud screening for Copernicus sentinel-3 sea and land surface temperature radiometer (SLSTR)

Bayesian approach to classify SLSTR pixels over polar regions in clear ocean, clouds and sea-ice is presented. The approach is based on Look-Up-Tables estimating the probability distribution function (PDF) for a pixel, given a set of measured values for selected variables. PDF’s have been generated by analysing archived MODIS AQUA and TERRA products. MODIS data have been selected because of the long available time series, the quality of cloud mask products and possibility to simulate the SLSTR observation including the dual view capability. A first set of candidate input variables in the PDF’s, defined based on review relevant literature, has been optimized both in terms of classification skills and computational efficiency. Different combinations of variables have been considered together with ancillary data SST and observation geometry to get the final set of variables to be used for classification. The optimization process based on: visual analysis, quantitative comparison against SAR ice concentration products is presented. The method has been applied to SLTR L1 data showing improvement respect to the current operational method of cloud classification. In addition, classification of pixels covered by sea ices is provided which consequently improves the SST final product.