[en] This study presents a deep learning-based methodology to estimate two critical hydrological parameters, soil moisture (SM) and inland surface water, using Global Navigation Satellite System-Reflectometry (GNSS-R) data. Leveraging 2D Delay Doppler Maps (DDMs) from NASA’s CYGNSS mission, along with auxiliary inputs, the approach addresses the limitations of sparse in-situ observations by offering high-resolution, all-weather alternatives. Reference datasets include SMAP-enhanced 9-km SM products and Landsat-based Global Surface Water maps, with surface water converted into binary classifications. The models were trained and tested across diverse geographies: SM over the Continental U.S. and surface water over the Amazon Basin. A tile-based architecture, comprising over 92,000 region-specific models using CNN and ANN layers, enables localized learning. Feature importance analysis was conducted via permutation methods. Results show robust performance, with an R value of 0.87 and ubRMSE of 0.06 m³/m³ for SM, and 0.88 accuracy for surface water classification. The final outputs were aggregated into monthly 3 km-resolution products, demonstrating the feasibility of GNSS-R and DL for global hydrological monitoring.
