Improving satellite-derived soil moisture estimates in a data-scarce watershed using strategically designed sampling and machine learning

Satellite-derived soil moisture products provide valuable large-scale information but often exhibit reduced accuracy in heterogeneous watersheds with limited monitoring data. This study evaluated whether a strategically designed monitoring network combined with machine learning could refine Soil Moisture Active Passive (SMAP) soil moisture estimates in the Vacacaí River basin (11,177 km²), southern Brazil. Monitoring locations were selected using Conditioned Latin Hypercube Sampling (cLHS) to maximize the representation of the watershed’s main topographic controls on soil moisture variability. Random Forest (RF) and Extreme Gradient Boosting (XGB) models were developed by integrating SMAP observations with topographic, climatic, and soil covariates. XGB achieved R² = 0.886, r = 0.941, KGE = 0.901, and RMSE = 0.0275 m³ m⁻³ during the test period, compared to R² = 0.887, r = 0.942, KGE = 0.882, and RMSE = 0.0274 m³ m⁻³ for RF. Leave-One-Site-Out cross-validation (LOSO-CV) showed that both models retained useful predictive skill at locations excluded from calibration, although performance varied among sites and environmental conditions. Variable importance and Shapley Additive Explanations (SHAP) analyses showed that SMAP provided the dominant large-scale temporal signal, while environmental covariates were strongly associated with local spatial variability. These findings demonstrate that combining strategically designed sampling with machine learning provides a transferable and physically consistent framework for refining satellite-derived soil moisture in heterogeneous and monitoring-limited watersheds.