A Spatial Downscaling Method for Remote Sensing Soil Moisture Using Adaptive Weighted Stacking Strategy

Soil moisture (SM) derived from remote sensing plays a crucial role in understanding land-atmosphere interactions between water and carbon cycles. However, existing remotely sensed surface SM products (e.g., ESA CCI SM) have relatively coarse spatial resolutions (25 – 40 km), limiting their suitability for precision agriculture and ecological management. To address this limitation, this study proposes an adaptive weighted stacking strategy for soil moisture downscaling. A stacking framework integrating Random Forest (RF), Gradient Boosted Regression Trees (GBRT), and XGBoost was developed to downscale 25 km resolution ESA CCI SM data to a high-resolution 1km product. Key predictors, including surface albedo, apparent thermal inertia, clay content, and leaf area index, were identified through SHAP (SHapley Additive exPlanations) feature importance analysis. An adaptive weight strategy was then introduced to dynamically optimize the contributions of each base model. The downscaled SM was validated using in-situ SM measurements from the Murrumbidgee River Basin. Results indicate that both GBRT (R = 0.916, RMSE = 0.046 m³/m³) and XGBoost (R = 0.915, RMSE = 0.047 m³/m³) models significantly outperformed the RF model (R = 0.847, RMSE = 0.066 m³/m³). Notably, the stacking strategy method that combines a linear regression meta model with adaptive weighting achieved the best performance (R = 0.931, RMSE = 0.041 m³/m³). The downscaled SM exhibits finer spatial details of within-field variability compared to the original CCI SM. Further spatiotemporal analysis confirmed the downloaded SM effectively captures precipitation response and seasonal variations, particularly providing more detailed representation in farmland and pastureland regions. This study provides an effective method for high-resolution soil moisture monitoring in semi-arid areas, with significant applications in agricultural irrigation, water resource management, and climate change research.