Abstract
Mapping snow cover fraction (SCF) from optical remote sensing is a widely studied topic that still presents some challenges, especially in mountainous areas when using high-resolution time series. The generation of reliable maps under all acquisition conditions is limited by topographic effects such as shadows, sunglint on snow and atmospheric disturbances. If these effects are not corrected or taken into account also the most advanced methods fail. The errors usually present in these maps are a limitation to the full exploitation of these SCF for model assimilation or retrieval of snow parameters, e.g. Snow Water Equivalent (SWE) estimation. In this work, we propose a new unmixing method to SCF retrieval based on the Vapnik-Chervonenkis (VC) theory. The main idea is that is easier to find different model that locally adapt the specific problem, than a unique model able to generalize the problem in all the situations. The proposed method starts from two sets of endmembers of the pure class "snow" and "snowless" that are identified in an unsupervised manner considering the spectral signature adapted to the scene and finds the best separation hyperplane by maximizing the distance between the two classes. When an unknown pixel is considered, it is mapped into that same space and predicted to belong to a category based on which side of the gap it falls. If it falls in the tube defined by the end members, then is a mixed pixel with a abundance that is proportional to the SCF. To handle the non-linearity of the problem, an RBF (Radial Basis Function) kernel was considered instead of simply using the scalar product between trainings. The method was validated with the three WV3 images and acquired in several sites in the Alps, showing promising results both in terms of robustness under challenging situations and in terms of SCF sensitivity, reaching over test sites RMSE lower than 20 % and bias close to 0.