The Influencing Factors of Root Water Uptake by Salix psammophila in Semiarid Regions in China Based on Deep Learning

ABSTRACT

Root water uptake (RWU) is a critical process regulating the hydrological cycle and sustaining ecosystems, particularly in arid and semiarid regions. However, the spatiotemporal dynamics of RWU and the external factors controlling it have not been fully characterised. In this study, a combination of in situ experiments and a deep neural network model was adopted to characterise the RWU dynamics of Salix psammophila over a complete growing season in the Mu Us Sandy Land. Furthermore, SHapley Additive exPlanations (SHAP) was applied to identify its influencing factors. The results show that the proposed Res‐CNN‐LSTM model effectively integrates the vertical profile feature extraction capability of CNN and the time series modelling capability of LSTM, while incorporating a residual structure that enhances the network's stability and interpretability in RWU simulation. The model demonstrates superior performance in modelling the nonlinear spatiotemporal dynamics of vegetation RWU, achieving R ² values no less than 0.93. Unlike traditional numerical simulations that rely on complex physical parameters, it enables end‐to‐end prediction based on easily accessible meteorological, soil water and groundwater data. SHAP‐based interpretation further reveals that groundwater, potential evapotranspiration, precipitation and soil water collectively account for 81.23% of the total contribution to water stress, with groundwater identified as the most influential factor, contributing 30.61%. Leveraging deep learning to identify key drivers of RWU establishes a new pathway for ecohydrological research. The findings offer valuable guidance for supporting ecological restoration efforts in water‐limited ecosystems.