Machine Learning and Explainable AI for Sentinel-1–Based Snow Depth Estimation: Insights from Three Different Mountainous Regions

Abstract. Mountain snow plays an important role in regulating freshwater resources and climate processes, making accurate monitoring of its depth and distribution essential for water resource planning and climate modelling. However, monitoring snow depth (SD) in mountainous regions remains challenging due to the extreme topographic and climatic conditions. Synthetic Aperture Radar (SAR)-based remote sensing is particularly useful in such regions due to its high spatial and temporal resolution and its ability to penetrate deep snow. This study focuses on improving Sentinel-1 C-band SAR backscatter-based SD estimation, which supports large-scale, continuous monitoring but faces limitations in vegetated, shallow, or wet snow conditions. To address these limitations, the present study developed a machine learning (ML) framework that integrates change detection-based backscatter indices, direct backscatter variables, and auxiliary datasets to improve SD estimation over the existing ML models. The framework is designed for implementation using preprocessed Sentinel-1 data from Google Earth Engine, thereby reducing computational requirements and enabling scalability across regions and time periods. It is implemented across three different mountainous regions, the Colorado Rocky Mountains, the European Alps and the Indian Western Himalayas, each representing diverse characteristics. Results demonstrate that the developed ML model significantly outperforms existing approaches in all regions. As the physical scattering mechanisms underlying the backscatter-based methods are not yet fully understood, explainable AI techniques are employed to interpret model behaviour and investigate how different input variables contribute under varying conditions. The analysis reveals distinct patterns across seasons, snow types, vegetation conditions, and regions, providing physical insights into C-band backscatter-based SD estimation.