Improving forecasts of snow water equivalent with hybrid machine learning

Abstract. Accurate characterization of snow water equivalent (SWE) is important for water resource management in large parts of the Northern Hemisphere, but its large spatio-temporal variability and limited observational data make it difficult to quantify. Complex physically-based models have been developed that allow long-term SWE prediction, including scenarios without snowpack observations or in future events. However, those still suffer from large errors in their simulations, have long run times at large scales and provide challenges for integrating observational data. There have been attempts at using machine learning (ML) to improve SWE forecasting from meteorological data with promising results, but the data scarcity issue and concerns about the ability to extrapolate in time and space remain. In this study, we evaluate two hybrid setups that integrate physically-based simulations and ML. The first setup, referred to as post-processing, follows a common approach in which the simulated outputs from a numerical snow model, Crocus, are used as predictors to the ML component in addition to the meteorological data. The second setup, named data augmentation, involves an ML model trained not only on measured SWE but also on Crocus-simulated SWE at additional locations. These approaches are deployed using in-situ meteorological and SWE measurements available at ten stations throughout the Northern Hemisphere, and compared to Crocus and a ML setup using measured data only. The results show that the post processing setup outperforms all other approaches when predicting on left-out years in the training stations, but performs poorly when extrapolating to other locations compared to Crocus. The addition of a large set of Crocus-simulated variables besides SWE in the post-processing setup results in similar performance for left-out years but exacerbates the spatial extrapolation issue. On the other hand, the data-augmentation setup performs slightly worse on the left-out years, but showed much better transferability to new locations, improving the other ML-based setups greatly and reducing the RMSE in Crocus by more than 10%. The feature importances of the ML-models are consistent with physical knowledge, despite having unusual deviations at extreme values, which could be further improved with the data-augmentation setup. Lastly, the addition of lagged variables results in improved results, but they are only relevant for up to a week. These results prove the usefulness of hybrid models and particularly the data-augmentation setup for SWE prediction even in data-scarce domains, which has the potential to improve forecasts of SWE at unprecedented spatio-temporal scales.