Development of a hybrid multisource soil moisture estimation framework (HMSMEF) using machine learning and remote sensing data

Accurate, high-resolution soil moisture data is essential for optimizing agricultural water use and supporting environmental monitoring, particularly in rice cultivation areas with complex terrain and spatiotemporal variability. This study aimsto develop a Hybrid Multisource Soil Moisture Estimation Framework (HMSMEF) that integrates advanced machine learning techniques with diverse remote sensing and auxiliary datasets to downscale SMAP soil moisture data to 10 m resolution. The framework synthesizes optical, radar, land use, precipitation, and topographic data through permutation testing of nine synergy combinations and applies three regression models: Random Forest (RF), Gradient Boosting (GB), and Classification and Regression Trees (CART). The optimal configuration—Synergy 6, incorporating Sentinel-1 SAR, CHIRPS precipitation, and SRTM topography—achieved the best results using the CART model (R² = 0.927, RMSE = 0.005). Bias correction methods improved R² to 0.936, and a stacking ensemble of CART, RF, and GB further enhanced accuracy. Validation with two in-situ datasets showed strong performance: (1) 2018–2019 field data (R² = 0.91, RMSE = 5.35) for 10–40 cm depth, with a widely sensor distance of 73 m; and (2) 2021–2023 green manure plots (R² = 0.67 untreated, R² = 0.60 treated), with dense sensor spacing (6.8 m), reflecting local variability. HMSMEF demonstrates a scalable and robust solution for high-resolution soil moisture estimation, with significant implications for precision agriculture, sustainable water management, and environmental modeling.