Lhasa River Basin. Accurate and interpretable streamflow simulation in high-altitude regions remains a challenge due to the complex conditions of climate and land surface. This study aims to enhance the accuracy and interpretability of streamflow simulation by the proposed coupling framework of Snowmelt Runoff Model (SRM) and Machine Learning (ML). Two coupling strategies were developed (residuals from SRM simulations as Transformer features for error correction, and intermediate component of runoff by SRM as input features for physical constraints) to perform and understand driving mechanisms. The performance of the residual coupling strategy ( NSE : 0.97) is significantly better than that of individual SRM ( NSE : 0.68) and ML ( NSE : Transformer, 0.96; LSTM, 0.94), demonstrating its robustness in capturing complex streamflow dynamics. The precipitation was determined as the main driving factor for streamflow in the Lhasa River Basin (LRB) by SHapley Additive exPlanations (SHAP). It’s found that temperature above 11.2℃ has a significant positive promoting effect on streamflow, and solar radiation shows a pronounced positive contribution when it exceeds 245.69 W/m². There’s a similar result if the relative humidity is above 45%. In terms of underlying surface, when the proportion of barren and snow cover is high, they contribute to streamflow, with bare land threshold of about 10% and snow cover of about 26%, whereas wetlands show inhibitory effects. The study demonstrates the potential application of combining physical hydrological models with ML, revealing impact of climate change and land use on streamflow, and providing support for water resource management. • Construct a coupling framework of physical snow runoff model and machine learning. • Driver factor and influence on runoff was identified and quantified by using SHAP. • Coupling strategies were determined by objectives (accuracy or interpretability).
Wang et al. (Mon,) studied this question.