A Dynamical System Approach to Wet Snow Retrieval Using Sentinel-1 SAR Images

In this study, we develop an original method for segmenting Sentinel-1 SAR images as a discrete-time dynamical system for estimating the extent of wet snow. The system considers different variables and parameters, including radar amplitude images and terrain information (altitudes, slopes, orientations). The dynamical variable is the SAR amplitude ratio (ratio between a SAR and reference image without snow), whose evolution follows a law described by a function applied to a set of connected pixels. The set of corresponding pixels and the strength of their connection are determined from a digital terrain model. The application of the dynamical system is nonlinear and is inspired by models of interacting particles from biology (neural network models). As the iterations increase, the dynamics leads to a segmentation of the image taken as an initial condition. A significant advantage of our system is that it considers the physical information of the terrain to determine couplings between neighbouring pixels. This algorithm has been tested over the season 2017-2018 and provides estimates of wet snow maps that are perfectly consistent with reference products, opening up new prospects for their use in different contexts (for example, for improving existing segmentation methods and for enhancing the quality of gaps filling approaches in case of missing data, etc.).