Abstract. Seasonal mountain snow is an indispensable resource, but accurate estimates of this water storage remain limited, even in the European Alps, where there is a dense network of in situ monitoring stations. In this study, we address Alpine snow depth estimation at a 100 m spatial resolution and sub-weekly temporal resolution over the 2015–2024 period using multiple input configurations within an extreme gradient boosting (XGBoost) machine learning (ML) model. We explore the potential of Sentinel-1 C-band dual-polarized synthetic aperture radar polarimetry (PolSAR) observations, and include either regionally downscaled meteorological forcing data or modeled snow depth as additional inputs to further explain interannual and spatial variability. A threefold nested cross-validation scheme is used to account for the spatio-temporal dependencies present in the snow depth data. XGBoost's internal booster and Shapley additive explanation (SHAP) values are used to relate the input features with the predictions for both dry and wet snow conditions. Our results indicate that the inclusion of PolSAR observations leads to modest improvements over a backscatter-intensity-based configuration, whereas the SHAP-based feature attribution reveals a high reliance of XGBoost on the polarimetric scattering angle and co-polarized (VV) backscatter intensity. Next, incorporating either meteorological forcing data or modeled snow depth substantially enhances predictive performance, particularly when spatially distributed training data, proven to be essential for capturing topographic controls on snow depth variability, are included. When supplemented with spatial training data and either meteorological forcing data or modeled snow depth estimates, XGBoost shows good agreement with nine snow surveys conducted in the Dischma valley (Switzerland), achieving correlation coefficients (R) of 0.76 and 0.78 and mean biases of 0.07 and 0.17 m, respectively. When applied to unseen locations across the Alps, the performance remains high, with R=0.80 and biases of −0.04 and −0.03 m, respectively.
Boeykens et al. (Fri,) studied this question.