Coupled SWAT, Stationary Wavelet Transform, and Interpretable Machine Learning to Improve Watershed Streamflow Simulation

Climate change and human activities have significantly impacted water resource availability, creating a critical challenge for accurately simulating high- and low-streamflow patterns in global hydrology. This study presents an innovative framework integrating the Soil and Water Assessment Tool (SWAT), Stationary Wavelet Transform (SWT), and interpretable machine learning models, including convolutional neural networks (CNN) and long short-term memory (LSTM) networks. Hydrometeorological features generated by SWAT were preprocessed with SWT and then input into the CNN-LSTM model, achieving R² and NSE values of 0.90 and 0.88 during testing. This approach outperformed traditional machine learning methods (e.g., support vector machines, random forests, LSTM) and the calibrated SWAT model, reducing high- and low-streamflow biases to within 1.3%. Interpretability analysis identified solar radiation and percolation as key drivers of streamflow, with their influence decreasing exponentially as lag time increased, while precipitation emerged as the most sensitive factor. The model demonstrated robust stability, maintaining R² and NSE values above 0.88 and 0.80 even with shorter sliding window lengths. Comparative experiments further validated the effectiveness of SWT in mitigating overfitting. In conclusion, this study integrates SWAT, SWT, and interpretable machine learning models, supported by interpretability tools such as individual conditional expectation (ICE), partial dependence plots (PDP), and Shapley additive explanations (SHAP). This comprehensive approach enhances simulation accuracy, strengthens the credibility of model results, and offers a reliable, efficient solution for long-term watershed streamflow modeling.