Using Long-Short Term Memory Network to Estimate Snow Water Equivalent in Sierra Nevada

The amount of water in a snowpack, or snow water equivalent (SWE), is an important variable in hydrology because of its potential for storage and delayed water release, which is crucial for soil moisture, water supply and groundwater recharge during the dry season. Data-driven machine learning (ML) techniques are an alternative approach to traditional physically based models in finding solutions to problems in hydrological systems. We explore three machine learning methods - multiple linear regression (MLR), random forest regression (RFR) and a type of deep learning method called Long Short-term Memory Network (LSTM) - for their ability to predict historic SWE in the Tuolumne, Merced, American and Feather basins in the Sierra Nevada. These four basins were further divided into elevation zones: lower (-1500 m), middle (1500-2400 m), and upper ( +2400 m) zones. The SNOW-17 snow model was used as a hydrological model benchmark to test the ML methods’ performance as it is used to predict SWE by NOAA’s California Nevada River Forecasting Center. A set of meteorological forcing inputs - air temperature, precipitation, and elevation of the rain-snow line - were used to estimate SWE from the ML and SNOW-17 models. The historical SWE data used as a target estimate for this study was the retrospective SWE reanalysis dataset based on probabilistic Bayesian data assimilation method for California’s Sierra Nevada (Margulis et al. 2016, 2019). It was found that LSTM outperformed MLR, RFR and SNOW-17 in estimating historic SWE, shown from the Nash Sutcliffe Efficiency (NSE) and mean bias in peak SWE. Based on the NSE, performance was better for the upper elevation zones, least skillful for the lower elevation zones, and showed a wide range of skills for the middle elevation zones. Mean Biases in Peak SWE were generally higher for the lower elevation zones and lower as the elevation increased . The ephemeral nature of the snow might have added to the uncertainty in SWE estimation at the lower elevations, whereas the uncertainty in precipitation partitioning between rain and snow could be a factor at the mid elevation transition zones. These uncertainties are less prominent at the higher elevations.