Snowpack mapping and forecasting by blending deep-learning estimates with ground observations in the Northern Sierra Nevada

To provide snow-storage and snowmelt estimates in mountain basins, this study presents a blending approach based on a deep-learning Long Short-Term Memory (LSTM) model and a zonal-bias-correction method using ground snow measurements. We explored the blending approachs ability for Snow Water Equivalent (SWE) mapping and forecasting in a domain of the Northern Californias Sierra Nevada, including the Feather, Yuba-Bear, and American River basins. First, initial daily SWE estimates were obtained from the LSTM model trained with inputs of historical gridded SWE products (e.g., SWE reanalysis dataset), historical daily precipitation and temperature, and landscape attributes. Then, point-scale daily snow measurements from snow pillows and wireless-sensor networks were used to bias-correct the initial SWE estimates, using a Zonal-Inverse-Distance-Weighting (Zonal-IDW) method to account for SWE spatial variability. Remote-sensing snow-covered-area data were used to filter the bias-corrected SWE estimates, removing areas without snow cover. In such a way, the blending approach can provide daily SWE maps, and thus change to daily SWE. Comparing to the averaged SWE from snow pillows in mountain areas above 1500-m elevation, the Zonal-IDW-based bias correction improved LSTM-only-model R2 from 0.79 to 0.97. Results from the blending approach were evaluated by monthly snow-course data for the testing period of water years 2007-2016, showing a mean difference of 24 mm, which was better than other gridded SWE datasets. The Zonal-IDW-based bias-correction significantly reduced potential SWE over-correction from the IDW-only method in rain-shadow areas of the Feather River basin, where high spatial SWE variability was not encompassed by limited snow pillows. Further, we also tested the blending approach with forecasted one-day-ahead precipitation and temperature data as inputs, suggesting that basin-scale SWE estimates were not sensitive to precipitation and temperature data from either observation-based dataset or weather forecasts. The proposed blending approach provides important ground snowpack and snowmelt estimates, which are crucial to runoff forecasting and real-time decision making in mountain basins.