Explainable XGBoost-based models of root-zone soil moisture profiling using coupled Sentinel-2 and IoT data in loam and silt loam soil

Accurate prediction of soil moisture content (SMC) is crucial for sustainable irrigation management and enhancing resilience against climate change. However, in-situ sensing offers accurate point-scale measurements but lacks spatial representativeness, while satellite-based offer spatial coverage but are either too coarse at field scale or indirect and cloud -sensitive. Integrating satellite observation with ground-based monitoring and IoT meteorological data could exploit complementary strengths by linking canopy conditions and atmospheric drivers to reliable in-field reference measurements. This study predicts SMC at five depths (5 to 80 cm) for two soil texture classes (loam and silt loam) using Extreme Gradient Boosting (XGBoost) by integrating Sentinel-2 vegetation indices Normalized Difference Vegetation Index (NDVI) and Normalized Difference Moisture Index (NDMI) with Internet of Things (IoT)-derived meteorological data using two input scenarios and gravimetric SMC as reference for model training and evaluation. The model trained with combined inputs achieved higher accuracy compared with using only vegetation indices as predictors in both soil textures and all depths. This was especially evident in loam soil at 5 and 20 cm depth, with R² values of 0.95 and 0.79 and RMSE values of 0.88% and 1.46%, respectively, compared to R² values of 0.70 and 0.69 and RMSE values of 2.26% and 1.77% when using vegetation indices only. The model achieved near-perfect accuracy in silt loam with R² = 0.99 at 5–20 cm and RMSE = 0.49–0.40% at same depths. SHapley Additive exPlanations (SHAP) analysis identified NDVI as the most influential predictor in surface soil layers (mean SHAP = 0.11–0.22), reflecting its strong sensitivity to canopy vigor. In contrast, solar radiation emerged as key determinant in deeper soil layers (60 and 80 cm; SHAP = 0.12–0.18), highlighting the importance of atmospheric evaporative demand in controlling subsoil moisture dynamics. The model’s accuracy and interpretability enable depth-specific decision support for irrigation timing and water use efficiency under variable weather conditions, while providing actionable driver insights for climate-adaptive management aligned with SDGs 6 and 13. The approach is validated for loam and silt loam textures using optical Sentinel-2 indices, which are subject to cloud cover and revisit latency; therefore, the current framework is not suitable for real-time irrigation scheduling without accounting for these delays. Future integration with SAR and gap-filling strategies would be required for operational real-time applications. Workflow for depth-specific soil moisture prediction using Sentinel-2 vegetation indices and IoT meteorological data with explainable XGBoost (SHAP). XGBoost achieved high accuracy in predicting soil moisture across five depths. Integrating Sentinel-2-derived vegetation indices with IoT data improved prediction across five depths and two soil textures (loam and silt loam). Explainable AI (SHAP) analysis clarified depth-dependent drivers, with NDVI had the highest importance in shallower depths and solar radiation dominated in the deeper depths. The proposed approach supports depth specific irrigation decision-making and sustainable water management aligned with SDG 6 and 13.