Abstract ID: 3.11114 | Accepted as Talk | Talk/Oral | TBA | TBA

In mountain regions, today’s snow is tomorrow’s water. Indeed, the snowpack functions as a natural freshwater reservoir and is a primary source of streamflow, particularly during spring and summer,regulating the hydrological cycle. Accurate estimates of snow water equivalent and snow depth are essential for characterizing seasonal water storage, predicting water availability, and managing cascading socio-hydrologic impacts—particularly in an era of increasing climate variability and drought events.
However, operational snow hydrology models are subject to significant uncertainties, including structural deficiencies, meteorological inputs, and parameter variability. Moreover, ground-based snow measurements are often affected by instrumental noise and lack of representativeness, while remote sensing products suffer from coarse spatial resolution and retrieval uncertainties. To address these limitations, data assimilation methods, such as the Ensemble Kalman Filter, optimally combine models and observations to enhance snow state estimates. Despite their effectiveness, such ensemble-based data assimilation approaches can be computationally expensive, limiting operational feasibility. To mitigate this issue, we propose a deep learning-based framework that leverages Long Short-Term Memory networks. Trained on Ensemble Kalman Filter outputs from seven diverse study sites, our model achieves comparable accuracy while significantly reducing computational costs by 70% compared to a parallelized Ensemble Kalman Filter. On average, the network introduces only a minor root mean square error increase for snow water equivalent (+6 mm) and snow depth (+6 cm), with 12 out of 14 site-specific cases outperforming open-loop estimates. Building on these results, we extend the framework to a 2D semi-distributed implementation. To improve the spatial interpolation of the analysis correction, we incorporate recent developments in spatial snow data assimilation leveraging an abstract distance metric defined by topographical features and melt-out climatology to enhance spatial information propagation. This approach is validated across three hydrological basins in California, Norway, and Italy, leveraging high-quality observations from INARCH sites. By integrating deep learning into data assimilation, this research enhances hydrological forecasting in high mountain environments, enabling scalable and efficient snowpack modeling with uncertainty quantification for decision making.