Evaluation of a High-Resolution Operational Snow Cover Area Classification Algorithm

The Let-It-Snow (LIS) algorithm generates 20m resolution maps of the snow cover area (SCA) including snow, no-snow and cloud pixels from Sentinel-2 level-2A images for the Copernicus Snow & Ice Monitoring Service. The pixel classification relies on two Normalized Difference Snow Index (NDSI) threshold parameters and two red band reflectance threshold parameters. In this study we used the Active Learning for Cloud Detection (ALCD) software to generate via supervised classification reference products from 10 Sentinel-2 images representing the snow and weather conditions over one year of the T31TCH tile (Pyrenees). Those reference products were used to evaluate both the algorithm's snow-cloud discrimination performances and the sensitivity of its threshold parameters. While it shows good performances in snow-land differentiation, the algorithm classifies transparent clouds pixels as cloud when a human eye could see the snow underneath. We tried 24167 values combinations of the threshold parameters and compared the resulting products with the reference products. The results indicate that the default parameters already give near-optimal performances and that they could be modified for a slight increase in snow detection.