Improving AMSR2-derived Snow Depth Spatial Resolution Through Synergistic Combination of Machine Learning and Random Sets Over the European Alps

Abstract

Snow depth is a critical parameter for meteorological, climatological, and hydrological models, enhancing regional water resource information and weather prediction accuracy. AMSR2 passive microwave data with 10-kilometer spatial resolution results in high error and uncertainty in snow depth estimation. This study attempted to develop machine learning algorithms to downscale AMSR2 snow depth data to 1 kilometer spatial resolution. One of the primary factors inflicting errors in snow depth downscaling is the absence of proper identification of snow-covered area boundaries (mixed pixels), leading to extensional uncertainty. In the present study, a novel method based on the theory of random sets has been developed to resolve the issue of mixed pixels in the boundaries of snow-covered areas for the European Alps. In addition to snow cover data acquired from random sets, auxiliary data for developing and training machine learning algorithms comprises of the Digital Elevation Model (DEM), Land Surface Temperature (LST), and Land Cover Type data were processed. The machine learning algorithms employed includes Support Vector Regression (SVR), Multi-Layer Perceptron (MLP), and Random Forest (RF). The RF algorithm, when considering snow cover data derived from random sets method, confirmed the excellent performance among the SVR and MLP algorithms. Using RF, the RMSE decreased from 31.67–16.34 cm, the correlation coefficient increased from 0.312–0.81, and the relative error decreased from 9.85–0.52% (reduced error and increased accuracy). The results based on evaluation metrics indicates an outstanding performance of the proposed innovative method for downscaling of AMSR2 snow depth. © Deutsche Gesellschaft für Photogrammetrie, Fernerkundung und Geoinformation (DGPF) e.V. 2025.