Runoff Evolution and Interpretable SWAT-RF-SHAP Modelling in the Upper Yangtze River: Integrating Hydrological Processes and Meteorological Drivers

Abstract

Intensified climate change and anthropogenic activities have increased the complexity and nonlinearity of watershed runoff, challenging accurate runoff simulation and input-factor importance analysis. Physical hydrological models are constrained by parameter uncertainty and structural limitations, whereas meteorology-driven data-driven models insufficiently capture the inherent characteristics of watershed hydrological processes. To address this gap, this study develops an interpretable SWAT–Random Forest–SHapley Additive exPlanations (SWAT-RF-SHAP) coupling framework with dual inputs of hydrological process variables and meteorological forcings, using monthly natural runoff data from 13 hydrological stations in the upper Yangtze River during 1980–2016. Seven SWAT-simulated hydrological variables (surface streamflow, groundwater streamflow, percolation, soil moisture, snowmelt, and potential and actual evapotranspiration) are combined with five meteorological variables (precipitation, air temperature, relative humidity, solar radiation and wind speed) to form a 12-dimensional feature set, and a random forest (RF) model optimised by the Sparrow Search Algorithm (SSA) integrates watershed hydrological features with meteorological forcing signals. The results show that: (1) annual runoff presents a three-stage evolution pattern and is decreasing at 11 of the 13 stations; the classical rescaled-range estimate of the mean Hurst exponent is 0.70, but because that estimator is biased upward for short records, the Anis–Lloyd bias-corrected estimate is used instead (mean 0.55, with 5 stations above 0.65), indicating moderate rather than strong long-term persistence, and no trend remains significant at the 1% level once serial correlation is accounted for. (2) The SWAT-RF model outperforms both standalone SWAT and the meteorology-driven RF model: the 13-station mean Nash–Sutcliffe Efficiency (NSE) is 0.951 during calibration (meteorology-driven RF: 0.909; SWAT: 0.679) and 0.723 during validation (RF: 0.678; SWAT: 0.587), and SWAT-RF attains the highest mean validation NSE in all 10 runs, by 0.044 over the meteorology-driven RF, so that its advantage exceeds the run-to-run variability of the optimisation. The margin over the error-correction model SWAT-RF-EC, which uses only SWAT-simulated runoff as input, is larger (validation NSE 0.640) but remains station-dependent, SWAT-RF-EC attaining the higher validation NSE at TZL and YC. (3) When the observed runoff of preceding months is added as a predictor, validation performance improves at all 13 stations relative to the re-optimised baseline, and the 12-month lag window yields the highest mean validation NSE of 0.795, a 10.4% gain; because these lag-augmented schemes are one-step-ahead predictors updated with observed runoff, their skill is not directly comparable with that of the simulation models. (4) SHAP analysis identifies six meteorological-dominant, six hybrid and one hydrological-process-dominant stations, revealing a longitudinal pattern in which hydrological processes dominate the source region while meteorological factors prevail downstream. The framework improves both simulation accuracy and interpretability, offering a scientific basis for regional water resources management in large river basins.