A1415
Title: A snow water equivalent retrieval method from satellite-based passive microwave observations using machine learning
Authors: Jorge Arevalo - Universidad de Valparaiso (Chile) [presenting]
Josemanuel Escobar - Universidad de Valparaiso (Chile)
Abstract: Snow water equivalent (SWE) is a key variable for hydrology, water resources management, and climate studies, yet its large-scale estimation remains challenging due to the limitations of current remote sensing approaches. ML4SWEr is presented. ML4SWEr is a new machine learning-based SWE retrieval method, combines satellite passive microwave (PMW) observations with ancillary geospatial variables, including topography, land cover, and Forest density. The proposed algorithm is based on Random Forest regression and was trained over the contiguous United States (CONUS) using the UA-SWE dataset as both training target and validation reference. The methodology is designed to support the generation of a global daily SWE dataset spanning more than 30 years at 3.125 km spatial resolution. Major challenges addressed include the processing of massive multi-source datasets and the harmonization of different geospatial projections and resolutions. Results show strong agreement with the UA-SWE dataset and promising performance when evaluated against independent SWE observations from different regions worldwide, suggesting good transferability and global generalization capability. Additionally, constraining the retrievals using a conceptual mass-conservation framework improves the physical consistency and reliability of the estimates, particularly in regions where existing SWE products exhibit high uncertainty.
