Assistant Professor The University of Utah, Utah, United States
Abstract Submission: Snowmelt is the primary driver of streamflow in the western US. Yet, its heterogeneous distribution, year-to-year variability, and lack of spatially representative observations at the appropriate cadence and temporal resolution continue challenging the hydrological modeling community. Addressing the novel challenge, we build on our machine learning-based snow modeling workflow to create SWEMLv2.0, or version 2.0 of the Snow Water Equivalent Machine Learning model, to map spatially continuous SWE distribution in mountainous terrain. We train SWEMLv2.0 with inputs consisting of large-scale retrospective climate datasets, large-scale land cover mapping and terrain products, in situ snow motoring system observations, and a target of the spatially continuous NASA Airborne Snow Observatory lidar-derived SWE product within the respective watersheds – an essential target that bypasses the bias and limitations of in situ observations from SNOTEL that many traditional snow models rely upon. We train SWEMLv2.0 as one XGBoost model on 2.7 million spatially distributed data points between 2013 and 2019, yielding high model performance with a KGE of 0.923, RMSE of 13.3 cm, and a PBias of 0.023% over the western US modeling domain at a 300 m resolution. The blend of operational remote sensing products, in situ observations, and key terrain features in SWEMLv2.0’s modeling workflow presents a novel advancement for near-real-time high spatial resolution SWE mapping in the mountainous western US – critical for advancing season-to-season water supply estimates and informing water resources management decision-making.
Learning Objectives/Expected Outcome (Optional) : Machine Learning, Snow Hydrology, SWE
