A Framework for High-Resolution Soil Moisture Mapping Using Sentinel-1/2 Predictors and a Stacking Ensemble

Soil moisture (SM) governs land–atmosphere exchanges and strongly influences agricultural management and hydrological assessment, yet high-resolution mapping remains challenging due to sensor-specific confounding effects and limited field observations. This study develops a practical workflow for point-scale SM estimation and wall-to-wall mapping by integrating multi-sensor remote sensing predictors with ensemble learning. A compact predictor set was constructed from Sentinel-2 optical indices, Sentinel-1 SAR descriptors (σVV and the polarization ratio σVH/σVV), and topographic information, collocated with in situ SM measurements along a transect in the study area. Three tree-based regressors—Random Forest, XGBoost, and CatBoost—were trained under an identical feature configuration and evaluated using R2, Root Mean Square Error (RMSE), and Mean Absolute Error (MAE) together with predicted–observed diagnostics. A stacking ensemble was then implemented using leakage-controlled K-fold out-of-fold predictions to generate meta-features, with a Decision Tree as the meta-learner tuned via a grid search. Results show that base learners achieve comparable skill (R2 ≈ 0.60–0.62; RMSE ≈ 0.038–0.039), while stacking improves test accuracy (RMSE = 0.0346) and provides a stable mapping-ready model. The trained framework was transferred to stacked raster predictors to produce spatially continuous SM maps, revealing coherent moisture heterogeneity across the region. Accordingly, the objective of this study is to develop a compact and application-oriented point-to-map workflow for high-resolution soil moisture estimation by integrating Sentinel-1/2-derived predictors with stacking-based model fusion, rather than to propose a new physically based retrieval model.