Item

Abstract

Seasonal snow plays a crucial role as a water reservoir and energy balance component, yet accurately estimating the depth of the snowpack remains a challenge, particularly in remote areas. ICESat-2 laser satellite altimetry offers the potential to provide precise snow depth measurements by comparing satellite-based snow surface elevation profiles with a high-quality Digital Elevation Model (DEM) of the snow-free ground. However, the satellite’s acquisition pattern is sparse in time and space, raising the need for additional data to produce a spatially complete snow depth map. In this study, we generate snow depth maps for mainland Norway by employing machine-learning methods to combine snow depths derived from the ICESat-2 ATL08 product (2018-2022) with ERA-5 Land data. Our methodology involves careful data co-registration, benchmarking of DEM uncertainties using ICESat-2 surface elevations from snow-free conditions as a reference, and applying a machine learning-based bias correction on the derived snow depths. Subsequently, snow depth maps are generated in an XGBoost regressor by statistically downscaling ERA-5 snow depth time series with the derived snow depth and incorporating terrain, vegetation and wind parameters. Our results suggest that while ERA-5 alone overestimates snow depth on a national scale, our approach removes this bias and reproduces snow depth patterns at the hillslope scale when compared to lidar-based snow depth maps acquired in the field. The approach is applicable globally wherever accurate snow-free DEMs are available.