Dominant Factor Analysis and Threshold Inflection Point Determination in Deep Learning-Based SWAT-LSTM Training Models with SHAP Interpretability Analysis

Climate change has intensified extreme hydrological risks, particularly in basins characterized by frequent seasonal streamflow interruptions and discontinuous hydrological records, where traditional process-based models exhibit limited capability for adaptive water resource management. This study develops a hybrid SWAT-LSTM framework that integrates SWAT-derived hydrological variables with meteorological factors and applies SHAP interpretability analysis to quantify dominant drivers and identify threshold inflection points of runoff variability. Using the upper and middle reaches of the Huolin River Basin as a case study, the coupled model outperformed the standalone SWAT model during the test period (NSE: 0.876 vs. 0.710; R2: 0.884 vs. 0.736) and more accurately reproduced extreme flood and drought events. Future projections (2026–2100), driven by the optimized FGOALS-g3 climate model under SSP2-4.5 and SSP5-8.5 scenarios, indicate increasing precipitation, accelerated minimum temperature rise, and a non-stationary runoff pattern characterized by a mid-century decline followed by a late-century increase. The SHAP results reveal strengthened meteorological dominance, particularly for precipitation and minimum temperature, while soil moisture, evapotranspiration, and percolation remain key hydrological controls. The upward shift in the minimum temperature threshold reflects strengthened temperature control on runoff dynamics under warming. The proposed framework improves extreme runoff prediction and provides a quantitative basis for climate-adaptive basin management.