Wet Snow Detection From Satellite SAR Images by Machine Learning With Physical Snowpack Model Labeling

Abstract

The detection of wet snow by satellite imaging is currently done in an unsupervised way and lacks quantitative evaluation due to the difficulty of collecting ground truths in extreme environments. In this paper, we propose to take into account information associated with a physical model to label satellite data for the purpose of supervised learning of snow properties using Synthetic Aperture Radar (SAR) imagery. This dataset is constructed from Sentinel-1 SAR images, augmented with topographic information obtained from a Digital Elevation Model (DEM). The labelling of this data is done at the scale of the Northern Alps using the CROCUS physical snow model. Then, we trained, over thirteen combinations of labelled dataset, a wide range of machine learning models to quantitatively identify the most relevant learners for the wet snow detection task. The results demonstrate consistency among the different algorithms, with significant improvement observed when incorporating polarimetric combinations and topographic orientation data in the input of the model. The best algorithmic solution trained on this dataset is evaluated by comparing the obtained wet snow map over a validation area in the French massif of the Grandes Rousses with the existing Copernicus products, Fractional Snow Cover (FSC) and SAR Wet Snow (SWS). We also compare the temporal results obtained at one meteorological station located in the test area. The results show a better representation of wet snow during the melting period using the supervised learning approach, as well as a reduction in areas classified as wet during the winter season.