Explainable dynamic hybrid fusion method of remote sensing and land surface model data for soil water estimation in high mountainous areas

Study region Heihe River Watershed, a high mountainous region in Northwest China Study focus This study proposed an Explainable Dynamic Hybrid Data Fusion (XDHDF) method that leverages remote sensing products and model simulations to adapt to high mountainous areas, and employed Explainable Artificial Intelligence (XAI) to elucidate the nonlinear impacts of environmental factors on soil water dynamics. New hydrological insights for the regions The XDHDF method achieves a correlation coefficient ( R ) of 0.983, a root mean square error ( RMSE ) of 0.019 m³ /m³ , and a Kling-Gupta efficiency ( KGE ) of 0.982, significantly outperforming existing data fusion methods in high mountainous areas. The fused product accurately captures spatial patterns and temporal persistence, demonstrating a strong capability for geographic extrapolation. The SHapley Additive exPlanations (SHAP) analysis reveals distinct non-linear controls. Two critical elevation thresholds (~2900 m and 3700 m) govern changes in soil water content (SWC): SWC decreases with elevation below 2900 m due to transpiration offsetting precipitation, increases between 2900 m and 3700 m due to net water accumulation, and the impact of elevation weakens above 3700 m because of permafrost and ice effects. Additionally, a U-shaped relationship between SOC and SWC indicates higher SWC at both low (<0.6 g/kg) and high (>4.6 g/kg) SOC levels, highlighting SOC as an integrative indicator of soil hydraulic properties. These non-linear patterns cannot be captured by traditional linear correlation analysis.