MAPunet: High-resolution snow depth mapping through U-Net pixel-wise regression

Abstract

Accurate snow depth prediction is essential for hydrological risk assessment, flood prediction, water resource management, and weather forecasting. While previous studies have successfully applied deep learning techniques to generate snow depth maps, many have been constrained by geographical coverage or low data resolution. This work addresses these limitations by integrating high-resolution LiDAR maps, satellite imagery, digital elevation models, and three novel time-dependent variables. Additionally, the well-known U-Net architecture has been customized to perform pixel-wise regression and accurately predict snow depth over large geographic areas. The proposed method, called MAPunet, effectively models snow depth in the mountainous region of Davos, achieving an average error of 0.62 m at a 5 m resolution. The experimental results demonstrate the potential of combining high-resolution data with advanced deep learning techniques for enhanced snow depth mapping.