Machine Learning to Advance National-Scale Snow Water Equivalent Modeling
Research Team Members
Objective:
Leverage advancements in ML to develop a spatially distributed modeling framework adapted to the dominant regional snow processes observed across the western US.
Approach:
Build on existing basin-scale ML snow models and methodology developed for the Valle d'Aosta region in the Italian Alps in collaboration with CIMA and the University of Utah. Analogous to SNOTEL limitations, in situ snow depth monitoring in the Valle d'Aosta region exhibits a bias of subalpine snow sensors due to installation challenges and reliable operations at high elevations. This includes climate variability, orographic heterogeneity, and the limited number of in situ measurement stations that challenge accurate basin-scale and larger SWE estimation, exhibiting an underprediction bias at elevations greater than 2000 m. Addressing the limitations of the existing catchment SWE model, the group developed a Python-based ML model leveraging intelligent feature engineering and the Random Forest algorithm to account for the non-linear relationships between orographic features and catchment SWE. The conceptual ML modeling framework employed in the Valle d'Aosta region formed the preliminary platform to build the western US ML-SWE model.
Impact:
The project kickstarted the University of Alabama ML snow modeling research focus. The work produced several conference presentations, including AGU, AMS, EWRI, and one manuscript in Environmental Modeling and Software. Research over the duration of the project identified conceptual, programmatic, and overall modeling limitations that spurred new research projects and objectives aiming to continue the advancement of AI/ML in spatial snow modeling. Lastly, the modeling work was semi-operationalized, being implemented on the University of Alabama Pantarhei High-Performance computing platform where it assimilates near-real time in situ and remote sensing observations, processes all data for modeling, makes a SWE prediction for over 20,000km2 of key water supplying catchments in the western US, pushes the data to a secure and publicly available Amazon Web Services S3 storage bucket, and supports visualization on the Tethys platform.Abstract:
Snow-derived water is a critical hydrological component for characterizing the quantity of water available for domestic, recreation, agriculture, and power generation in the western United States. Advancing the efficiency and optimization of these aspects of water resources management requires an enhanced characterization of the snow state variable, particularly the essential global inputs of snow-water-equivalent (SWE), peak SWE, and snowmelt onset for hydrological models. While physically-based models that characterize the feedbacks and interactions between influencing factors predict SWE values well in homogeneous settings, these models exhibit limitations for CONUS-scale deployment due to challenges attributed to spatial resolution, landscape heterogeneity, and computational intensity. Addressing these limitations, we developed the Snow Water Equivalent Machine Learning (SWEML) modelign workflow as a full stack data-driven modeling platform with a modular structure to account for the heterogeneity of climate and topographical influences on SWE across the western United States. The SWEML pipeline assimilates nearly 700 snow telemetry (SNOTEL) and California Data Exchange Center (CDEC) sites and combines with processed lidar-derived terrain features for the prediction of a 1 km x 1 km SWE inference in critical snowsheds. We complete the full stack product by pushing all inputs and model outputs to a publicly accessible Amazon Web Services S3 storage and use the Tethys interface to support the interactive use of the results. With preliminary regional testing displaying high model skill in key water supply catchments, SWEML demonstrates the potential of ML to advance SWE inputs into hydrological models such as the National Water Model to effectively improve supply and flood forecasting.