Soil moisture dynamics in heterogeneous tropical landscapes using active–passive Earth observations and machine learning

Accurate monitoring of soil moisture (SM) at high spatial resolution remains challenging in tropical landscapes, where interactions among soil, vegetation, and terrain drive complex hydrological dynamics. We developed a Relative Soil Moisture Index (RSMI) that integrates multi-temporal Sentinel-1 SAR, Sentinel-2 optical imagery, terrain indices, and pedological attributes within a Random Forest (RF) framework. Across ten field campaigns spanning a full seasonal cycle, we collected 1560 SM observations covering diverse pedomorphogeological units of the Cerrado biome. The RF model achieved strong predictive skill (mean validation R² = 0.79; RMSE ≈ 5.0 %), with performance peaking in dry months (R² up to 0.91) and slightly lower accuracy in wet periods (R² ≈ 0.68). Feature importance analysis consistently identified VH backscatter, NDVI, terrain roughness, and clay content as dominant predictors. Validation against Soil Moisture Active Passive (SMAP) products (downscaled to 250 m) confirmed the approach's robustness (R² = 0.65). By normalizing site-specific dynamics, the RSMI provided coherent cross-site assessments and revealed greater persistence of moisture in clay-rich and flatter terrains. This method delivers a scalable and transferable tool for operational SM monitoring, supporting applications in precision agriculture, drought risk management, and sustainable land use planning in tropical environments. • Remote sensing reveals accurate spatial patterns of soil water availability. • Radar backscatter and vegetation indices explain soil wetness variation. • Terrain and clay content control the persistence of water in landscapes. • Short-term seasonal monitoring supports drought risk assessment in agriculture. • Spatio-temporal maps highlight soil–landscape interactions in tropical regions.