Snow Water Equivalent Estimation Via Machine Learning in the Mountainous Region of British Columbia, Canada

Good estimates of snow water equivalent (SWE) in regions of significant seasonal accumulation are critical to understanding hydrologic states and forecasting future streamflow. Complex topography and heavy forest cover, conditions common in British Columbia, Canada, can make these assessments challenging. A number of readily available gridded products that include surface SWE (ERA-Interim, MERRA, GLDAS and GlobSnow) have been used to build a statistical SWE estimation model using machine learning methods. Evaluated methods include artificial neural networks, Bayesian neural networks, support vector regression and random forests. Cross-validated SWE estimates from the statistical model and the individual data products were compared against in-situ snow measurements at manual snow course stations throughout BC. In addition, SWE values simulated by the Variable Infiltration Capacity (VIC) macroscale hydrologic model were also evaluated against these in-situ data. Mean station RMSEs for the data products ranged from 319 to 431 mm SWE, while that of the VIC runs was 211 mm. Runs of the statistical model achieved a mean station RMSE as low as 190 mm SWE, an improvement of 40% to 56% over the individual products and 10% over VIC. Nonlinear machine learning methods outperformed linear regression by 16% to 19%. These results demonstrate that the skill of SWE estimates in mountainous regions may be increased by employing a fusion of available gridded products and relevant covariates.