Accurate prediction of multi-layer soil moisture in arid, saline agro-ecosystems is critical for sustainable water management but remains challenging due to pronounced vertical heterogeneity in hydrological processes. To address this, we developed a physically-consistent stratified hybrid modeling framework aligned with the vertical stratification mechanism of soil hydrological processes, which integrates deep learning with machine learning, tailored to the distinct drivers governing different soil layers. Leveraging high-frequency in-situ observations from eight depths (20–350 cm) in Xinjiang, China, along with meteorological and remote sensing data, our approach first identifies the dominant controls on soil moisture across the profile: energy and vegetation dynamics dominate the shallow layers (0–60 cm), while groundwater depth becomes the primary controller in deeper layers (>100 cm). We then propose a hierarchical TimesNet-Random Forest fusion model with layer-adaptive weighting, which synergistically captures complex temporal dynamics and multivariate environmental interactions. This framework achieves state-of-the-art predictive accuracy (e.g., MAE=0.132 m 3 m −3 at 20 cm; MAE=0.030 m 3 m −3 at 250 cm) and its SHAP-based interpretability confirms a physically consistent shift in dominant predictors with depth. This framework explicitly decouples the depth-dependent hydrological processes across the soil profile, achieving physically consistent multi-layer soil moisture predictions without relying on complex inter-layer hydraulic coupling calculations, providing a robust and easy-to-implement tool for data-scarce arid regions.
Xu et al. (Sun,) studied this question.