A machine learning approach for estimating snow depth across the European Alps from Sentinel-1 imagery

Abstract

Seasonal snow plays a crucial role in society and understanding trends in snow depth and mass is essential for making informed decisions about water resources and adaptation to climate change. However, quantifying snow depth in remote, mountainous areas with complex topography remains a significant challenge. The increasing availability of high-resolution synthetic aperture radar (SAR) observations from active microwave satellites has prompted opportunistic use of the data to retrieve snow depth remotely across large spatial and frequent temporal scales and at a high spatial resolution. Nevertheless, these novel SAR-based snow depth retrieval methods face their own set of limitations, including challenges for shallow snowpacks, high vegetation cover, and wet snow conditions. In response, here we introduce a machine learning approach to enhance SAR-based snow depth estimation over the European Alps. By integrating Sentinel-1 SAR imagery, optical snow cover observations, and topographic, forest cover and snow class information, our machine learning retrieval method more accurately estimates snow depth at independent in-situ measurement sites than current methods. Further, our method provides estimates at 100 m horizontal resolution and is capable of better capturing local-scale topography-driven snow depth variability. Through detailed feature importance analysis, we identify optimal conditions for SAR data utilization, thereby providing insight into future use of C-band SAR for snow depth retrieval.