Authors: Engela Sthapit, William R. Currier, Mimi Hughes, Rob Cifelli – NOAA, Physical Sciences Laboratory
Title: LSTM Shows Promise in Estimating Snow Water Equivalent in Sierra Nevada
Abstract: Sierra Nevada snowpack is an important water source for the state of California, feeding into streams, lakes and reservoirs during the spring and summer snowmelt. Runoff forecasts rely on the proper estimation of snow water equivalent (SWE) at the basin-scale. Timing of the seasonal snowmelt also has important implications for water availability, reservoir management, and flooding. However, estimating SWE and timing of the melt in mountainous terrain is challenging due to complex topography and vegetation interactions that create considerable heterogeneity in snow distribution. Remote sensing, numerical modeling and in-situ observations have not been able to provide fully reliable spatiotemporal SWE estimates for this region. This study explores the ability of a data-driven method, the Long Short-Term Memory (LSTM) network, to estimate historic SWE in Tuolumne, Merced, American and Feather basins of the Sierra Nevada – each basin divided into smaller sub-basins by elevation zones: upper (+2000 m), middle (1500-2000 m) and lower (-1500 m) zones. LSTM used dynamic inputs from meteorological forcing variables and static elevation inputs to predict historic SWE. In each sub-basin, SWE was predicted with LSTM model trained and evaluated against a SWE reanalysis dataset developed by the Margulis research group. SWE from the LSTM model was then compared to the estimates from the benchmark SNOW-17 snow model, a temperature indexed hydrological model, that is used by the River Forecast Centers for streamflow forecasting. The results indicate that overall, the LSTM had better performance skill (KGE and RMSE) and smaller biases in peak SWE than the SNOW-17. Where the biases were present, the SNOW-17 mostly underpredicted and the LSTM mostly overpredicted the peak SWE. Both models showed better performance in the upper elevation zones compared to the middle and lower elevation sub-basins. Overall, the snow melted faster and earlier in the SNOW-17 than in LSTM; fastest in the middle and lower elevation sub-basins, and the slowest in the upper elevation sub-basins. This study has potential for incorporating SWE generated by data-driven models such as LSTM to improve streamflow forecasts by operational models.