A Hybrid Physics-Guided Machine Learning for Soil Moisture Monitoring and Drought Assessment From CYGNSS in Complex Terrain

Accurate soil moisture estimation is critical for drought monitoring, water resource management, and agricultural planning from the Cyclone Global Navigation SatelliteSystem (CYGNSS). However, conventional approaches often struggle with spatial heterogeneity and seasonal variability in complex landscapes. This study develops a three-layer hybrid framework integrating physics-guided (PG) principles, machine learning, and spatial optimization to improve soil moisture estimation across Eastern China’s diverse hydroclimatic regions. The framework sequentially combines a PG empirical baseline, a machine-learning correction module, and a spatial coherence optimization layer to leverage complementary strengths across modeling paradigms. Performance was evaluated using internal CYGNSS consistency checks and external Soil Moisture of China (SMCI) in situ datasets, covering 41 948 samples across 4025 locations during the 2021 summer monsoon season. The hybrid framework achieved R₂ = 0.653 forinternal consistency checks and R₂ = 0.474 for external valida- tion. It significantly outperformed individual model components (p <0.001) and maintained temporal stability across the monsoon season. The framework successfully captured the drought-to-wettransition while reducing RMSE by 26.4% when compared to PG baseline models. Spatially, it showed consistent performance across diverse regions, with optimal results in agricultural areas (RMSE/0.046 m3/m3). Error distribution improved substantially,with 60.3% of predictions within ±0.05 m3/m3 of reference values. The framework also demonstrated improved consistency with multiscale drought indices, supporting its potential for operational soil moisture monitoring in complex landscapes andagricultural drought early warning.