Authors: Ryan Johnson – University of Utah
Title: Combining Large-Domain meteorological datasets and remote sensing products in a Machine Learning framework to create high spatial resolution snow-water-equivalent maps
Presentation Type: Lightning Talk
Abstract: Accurately representing the heterogeneous distribution of snow in montane regions continues to challenge the hydrological modeling community, especially in the western US where it is commonly the driver of streamflow and the overall water supply. Addressing the challenge, we develop a regionally agnostic machine learning (ML) workflow that leverages NLDAS precipitation, NASA VIIRS normalized difference snow index, NRCS Snow Telemetry SWE, California Data Exchange Center SWE, Sturm regional snow classification, North American Land Change Monitoring System vegetation classification, Copernicus DEM, and NASA Airborne Snow Observatory (ASO) SWE data to estimate SWE for NASA ASO domains. Model development evaluated deep neural networks and tree-based ML architectures using nearly 2.7 million data points in the ML training workflow, with the multi-model evaluation identifying the XGBoost algorithm as demonstrating the greatest model skill. The Snow-Water-Equivalent ML v2.0 (SWEMLv2.0) model with XGBoost produced 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 high model skill in capturing the spatial distribution of SWE in montane catchments highlights the potential for high spatial resolution, spatially continuous SWE products and increasing hydrological modeling capabilities, such as the Next Generation National Water Model, by updating the state of snowpack for reducing the uncertainty in seasonal water supply