Soil moisture retrieval based on CYGNSS and Fengyun-3E GNOS-II dual-constellation observations

Soil moisture (SM) is a key variable in the Earth’s water and energy cycles, and its accurate large-scale monitoring is important for climate research, water resource management, and disaster early warning. Global Navigation Satellite System Reflectometry (GNSS-R) provides a promising approach for SM retrieval because of its all-weather capability and dense sampling potential. In this study, we developed a unified machine-learning framework for SM retrieval using CYGNSS and FY-3E GNOS-II multi-constellation observations. First, coarse-resolution SMAP data were transferred to GNSS-R specular points using spatial interpolation, where Random Forest achieved the best performance among the tested methods (R = 0.981, RMSE = 0.024 cm3/cm3). Second, multiple machine-learning retrieval models were trained and compared consistently for CYGNSS and FY-3E GNOS-II observations, and FY-3E GNOS-II showed the best retrieval performance, reaching R = 0.943 and RMSE = 0.029 cm3/cm3. To further assess robustness at a scale consistent with gridded products, specular-point retrievals were aggregated to the 36-km SMAP grid, where performance became more conservative, with R ranging from 0.851 to 0.893 and RMSE ranging from 0.055 to 0.068 cm3/cm3. Finally, an error-weighted fusion of the optimal CYGNSS and FY-3E retrievals was performed to improve both spatial coverage and retrieval accuracy, producing a quasi-global SM product with R = 0.952 and RMSE = 0.029 cm3/cm3. This research underscores the critical importance of high-quality reference data and multi-source GNSS-R data fusion via advanced ML for achieving high-precision soil moisture retrieval.