A physics-informed data-driven framework for 30m soil moisture estimation from multi-source datasets

Study region Hebei Plain, Hebei Province, China. Study focus Accurate, high-resolution soil moisture (SM) estimation is crucial for field-scale hydrological modeling, agricultural management, and climate research, yet current approaches remain limited by sparse in situ observations and coarse spatial resolution of satellite-based products. To address these challenges, this study develops a robust physics-informed data-driven framework for estimating 0–10 cm SM at 30 m resolution by integrating multi-source datasets through spatiotemporal data fusion and machine learning models. New hydrological insights The proposed framework consists of three key methodological advances. First, to bridge the spatial resolution gap and ensure high-quality inputs, we systematically evaluated four spatiotemporal data fusion algorithms, identifying Enhanced Spatial and Temporal Adaptive Reflectance Fusion Model (ESTARFM) as superior for maintaining consistent spectral and spatial representation of surface reflectance and temperature at 30 m resolution, achieving the highest r and the lowest RMSE, EDGE and LBP values across most cities. Second, we constructed a hybrid feature set coupling physically interpretable indices, such as the Temperature Vegetation Dryness Index (TVDI), with multi-source reanalysis and topographic data, ensuring an efficient predictor set through multicollinearity diagnosis. Third, using Optuna-driven automated hyperparameter optimization, we benchmarked ten machine-learning algorithms, finding that Random Forest (RF) exhibited the most robust performance with r of 0.921, R² of 0.835, RMSE of 2.512% and near-zero bias (−0.041%) against in situ SM observations. The resulting 30 m soil moisture maps effectively resolve fine-scale landscape heterogeneity. Overall, this study provides a robust, scalable, and application-ready ML-based framework for generating SM products with high spatiotemporal resolution, offering valuable support for regional water resource management and drought monitoring in complex agricultural landscapes.