Pflug et al. (2025) -- Process based and machine learning model outputs

Snow reanalyses that combine process-based models and remote sensing observations of snow provide estimates of seasonal snow water equivalent (SWE) evolution that surpass the accuracies of traditional modeling approaches. However, snow reanalyses are only available over smaller subregions, and sometimes use computationally expensive modeling approaches. We investigate whether 1 km-resolution and daily SWE from a popular reanalysis could be learned by connecting only the most-trusted meteorological fields (multidecadal precipitation patterns and daily air temperature) and remotely sensed snow cover using a deep learning model. Relative to point observations of SWE evolution in the western United States, the lightweight deep learning model was able to reproduce the spatial and temporal evolution estimated by the snow reanalysis. Further, we found that the deep learning model could be trained in the western United States and then reused to estimate SWE evolution in the European Alps, demonstrating a high average coefficient of correlation (0.81) and low peak-SWE bias (< 1%) versus point estimates of SWE. SWE from the deep learning model also outperformed SWE estimates from physically based land surface simulations, capturing elevation-driven impacts on SWE spatial heterogeneity and interannual differences in seasonal SWE magnitudes important for water resources, climate regulation, and local ecology. This study demonstrates how deep learning approaches could be used to mine connections between daily SWE evolution, snow cover remote sensing, and limited meteorological information to generate and expand the geographical extent of fine-resolution historical snow estimates in complex terrain.