Abstract
Study region Tuolumne, Merced, American and Feather basins, Sierra Nevada. Study focus This study explores the ability of data-driven models and an operational, process-based SNOW-17 snow model to estimate historical snow water equivalent (SWE). Current operational streamflow forecasts issued by the California-Nevada River Forecast Center use SWE simulated from SNOW-17. Three machine learning methods of varying complexity: Multiple-Linear Regression, Random Forest Regression, and Long Short-term Memory (LSTM) were tested. SWE from LSTM, the best performing data-driven model, was then compared to SNOW-17 for operational context to understand their respective skills and biases, thereby highlighting opportunities to provide improved SWE information for streamflow forecasts. New hydrological insights for the region LSTM outperformed other data-driven methods and SNOW-17, with cross-basin median KGE of 0.76 and 0.45 in LSTM and SNOW-17, respectively during 2011–2016. This demonstrates LSTM’s ability to capture non-linear processes and delayed hydrologic responses, enabled by its internal memory structure. Both models performed better at higher elevations, more so in SNOW-17. Biases in SNOW-17 were similar across different test periods, while LSTM was sensitive to training data. However, basin-to-basin variability in KGE and bias was larger in SNOW-17. In addition, snowmelt duration was shorter and snow disappeared earlier in SNOW-17. Overall, LSTM’s relatively high skill in estimating basin-wide SWE indicates strong potential for integration into operational hydrologic forecasting systems, with possible benefits for improving streamflow simulations.
Showing the abstract — retrieve the full paper via the Exa API.