Machine learning-enhanced ecohydrological modelling of monthly-scale oasis dynamics driven by ecological water diversion: A case study of the Qingtu Oasis, China

Ecohydrological models support oasis restoration projects driven by ecological water diversion. However, standalone machine learning models fail to capture underlying ecohydrological mechanisms and interactions. To address this limitation, we introduce a framework that integrates machine learning with empirical ecohydrological equations to construct a mechanistically grounded modelling approach. Ten candidate machine learning models, including Support Vector Regression, Categorical Boosting, and Deep Deterministic Policy Gradient (DDPG), were evaluated to construct independent modules for lake area, water table depth, oasis area, and fractional vegetation coverage. Empirical ecohydrological equations for oasis evapotranspiration and the groundwater-based carrying capacities of oasis area and fractional vegetation coverage were integrated into these modules, allowing the framework to incorporate ecohydrological mechanisms, which were further validated using SHapley Additive exPlanations (SHAP) analysis. The enumeration method was used to integrate the four modules, enabling the representation of groundwater–vegetation interactions and the simulation of key ecohydrological processes. A recursive strategy was then applied to generate continuous, monthly simulations of lake area, water table depth, oasis area, and fractional vegetation coverage. The framework was tested in the Qingtu Oasis of the Shiyang River Basin, China, as a representative case study. Results demonstrate both the feasibility and methodological validity of the approach, indicating that combining the DDPG model with empirical ecohydrological equations provides a useful reference for ecohydrological simulation. Future research will aim to reduce data uncertainty, incorporate soil salinity into the modelling framework, and explore transferability across different arid endorheic basins. • A framework was proposed for the ecohydrological simulation of restored oases. • Combining machine learning models with empirical equations was practical. • Empirical equations inputted ecohydrological mechanisms into machine learning. • The enumeration method was effective for modelling ecohydrological interactions. • The proposed framework improved the accuracy of ecohydrological simulation.