Spatial downscaling of snow water equivalent estimates for hydrological applications in Alpine Europe using machine learning

Accurate Snow Water Equivalent (SWE) data is essential for hydrological modelling, flood forecasting, estimating terrestrial water storage, and understanding climate change impacts on water systems. The high horizontal and vertical heterogeneity of snowfall, snow accumulation and snowmelt restrict the usage of ground-based SWE observations for region-scale estimations. Climate reanalysis products like ERA5-Land provide SWE estimates globally but are often unable to capture local snow processes due to their limited spatial resolution, especially in mountain areas with heterogeneous topography.

To address these limitations, this study presents a Random Forest Regression (RFR)-based approach to downscale ERA5-Land SWE data to a finer spatial resolution using open-source global datasets and in situ SWE measurements. The RFR model was trained on a dataset of SWE observations at 383 snow weather stations between 1999 and 2019. Predictor datasets included climate reanalysis of ERA5-Land SWE and DEM-derived topographical covariates. The SWE downscaling methodology was trained and validated for the Upper Danube River Basin and its applicability in hydrological models is investigated in two case studies in Alpine Europe: CWatM model simulations for the Upper Danube and PCR-GLOBWB simulations for the Rhine-Meuse Basin.

ERA5-Land significantly overestimated SWE with a PBIAS of 444% at snow weather station locations in the Upper Danube River Basin. Applying the downscaling approach significantly reduced this bias to -11%. Downscaled SWE strongly correlated with the observations with an R² of 0.81 and an RMSE of 17.87 mm for the Upper Danube. The downscaled SWE showed improved temporal dynamics of snow accumulation and melt, and enhanced spatial distribution. These initial results highlight the potential of the RFR downscaling approach for improving snowmelt runoff calibration in the two case studies. In the Rhine-Meuse study, we validate the applicability of the RFR model to regions outside the training domain. The open-source and easily accessible nature of the predictor datasets ensures accessibility and adaptability across diverse landscapes.