Evaluating a Long Short-Term Memory (LSTM) approach for Snow Water Equivalent (SWE) downscaling and hydrological modeling in mountainous terrain

In the context of the global hydrological cycle, runoff generated from snowmelt plays a key role in water availability, particularly in cold and mountainous regions. In many parts of Canada and the western United States, mountain snowpacks act as natural reservoirs by storing precipitation during cold seasons and releasing it during spring and summer. Accurate estimation of snow water equivalent (SWE) is therefore essential for hydropower reservoir operation, snow-related hazard assessment, and hydrological modeling. However, the coarse spatial resolution of widely available SWE products in northern latitudes, combined with complex mountain topography, introduces substantial uncertainty in their direct application to hydrological models. High-resolution SWE mapping remains a major challenge in these environments. In this study, we propose a multifactor SWE downscaling framework based on a Long Short-Term Memory (LSTM) deep learning approach, applied to the Nechako River watershed in British Columbia, Canada. The framework uses ERA5-Land SWE at 10 km resolution as the target variable, with predictor variables including precipitation, minimum and maximum temperature, solar radiation, and 2-m dewpoint temperature, together with static physiographic information such as elevation and land cover. Daily data from 1981 to 2024 are considered. The model is trained and evaluated at the 10 km resolution before being applied to generate SWE at 5 km resolution, corresponding to the spatial resolution of the CEQUEAU hydrological model. The downscaled SWE fields are designed to retain the large-scale snow patterns provided by ERA5-Land, while adding more spatial detail based on local elevation and land cover. Current work focuses on incorporating these downscaled SWE estimates into the CEQUEAU hydrological model and comparing the resulting runoff simulations with those obtained using CEQUEAU’s internal SWE representation. Rather than aiming to demonstrate clear improvements at this stage, the goal is to better understand how different SWE inputs influence the simulated hydrological response. Preliminary results suggest that LSTM-based downscaling offers a flexible and promising way to generate intermediate-resolution SWE fields in mountainous regions. This approach shows potential as a practical link between coarse-resolution reanalysis products and distributed hydrological models used for water resources and hydropower studies.