Exploring the Potential of Long Short-Term Memory Networks for Improving Understanding of Continental- and Regional-Scale Snowpack Dynamics

Accurate estimation of the spatio-temporal distribution of snow water equivalent is essential given its global importance for understanding climate dynamics and climate change, and as a source of fresh water. Here, we explore the potential of using the Long Short-Term Memory (LSTM) network for continental and regional scale modeling of daily snow accumulation and melt dynamics at 4-km pixel resolution across the conterminous US (CONUS). To reduce training costs (data are available for ∼0.31 million snowy pixels), we combine spatial sampling with stagewise model development, whereby the network is first pretrained across the entire CONUS and then subjected to regional fine-tuning. Accordingly, model evaluation is focused on out-of-sample predictive performance across space (analogous to the prediction in ungauged basins problem). We find that, given identical inputs (precipitation, temperature, and elevation), a single CONUS-wide LSTM provides significantly better spatio-temporal generalization than a regionally calibrated version of the physical-conceptual temperature-index-based SNOW17 model. Adding more meteorological information (dew point temperature, vapor pressure deficit, longwave radiation, and shortwave radiation) further improves model performance, while rendering redundant the local information provided by elevation. Overall, the LSTM exhibits better transferability than SNOW17 to locations that were not included in the training data set, reinforcing the advantages of structure learning over parameter learning. Our results suggest that an LSTM-based approach could be used to develop continental/global-scale systems for modeling snow dynamics.