Runoff Simulation and Applicability Assessment of Process-Based and Explainable Machine Learning Models Across Basins with Contrasting Glacier Cover in the Upper Hotan River Basin

Abstract

Accurate runoff simulation in glacier-fed alpine basins is essential for sustainable water resource management and water security in downstream arid regions, particularly under climate warming and accelerating glacier retreat. To address the limited representation of glacier processes in the Soil and Water Assessment Tool (SWAT) and the challenges of runoff modelling in data-scarce alpine basins, this study compared monthly runoff simulations from SWAT, SWAT_Glacier, Light Gradient Boosting Machine (LightGBM), and Extreme Gradient Boosting (XGBoost) in the Yurungkash and Karakash River Basins, two glacier-fed tributaries of the upper Hotan River Basin, Xinjiang, China. Pearson correlation analysis was used to screen runoff-related predictors, and SHapley Additive exPlanations (SHAP) was applied to quantify model-specific feature attribution in runoff simulations. SWAT_Glacier improved validation-period Nash–Sutcliffe efficiency (NSE) from 0.50–0.57 to 0.80–0.81 and modified Kling–Gupta efficiency (KGE) from 0.57–0.66 to 0.76–0.82 relative to SWAT, while reducing root mean square error (RMSE) by 32.4–37.0%. LightGBM and XGBoost achieved comparable testing performance, with NSE and R2 of 0.81–0.82 and KGE of 0.80–0.86. Across the two basins, the relative performance of SWAT_Glacier and the machine-learning models varied by evaluation metric and basin characteristics, rather than showing a consistent model ranking. Temperature-related variables showed the strongest statistical associations and model attributions in both basins. These findings demonstrate that model selection should consider glacier cover, process representation, and data availability to improve runoff simulation reliability and support sustainable water resources management in glacier-fed basins.