Performance evaluation of ML techniques in hydrologic studies: Comparing streamflow simulated by SWAT, GR4J, and state-of-the-art ML-based models

This study presents a comprehensive comparison between traditional hydrological models and advanced machine learning (ML) techniques in predicting streamflow dynamics. Traditional models, namely the Soil and Water Assessment Tool (SWAT) and G & eacute;nie Rural & agrave; 4 Param & egrave;tres Journalier (GR4J), are juxtaposed against ML models, including Random Forest (RF), Artificial Neural Network (ANN), Long Short-Term Memory (LSTM), and Bidirectional LSTM (BiLSTM). Both SWAT and GR4J demonstrated commendable performance, with GR4J displaying marginally superior predictive accuracy, evidenced by its tighter RMSE values. In the realm of ML, RF exhibited exceptional prowess in integrating diverse climatic features, especially in a scenario integrating comprehensive meteorological data. ANN showcased consistent performance across different input scenarios, emphasising its robustness. LSTM and BiLSTM, tailored for time series data, underscored the importance of precipitation's temporal dynamics in streamflow predictions. A notable revelation is the significance of choosing appropriate input data, with certain scenarios outperforming others based on the amalgamation of meteorological parameters. The flow duration curve (FDC) analysis further highlighted the model capabilities, with RF and BiLSTM excelling in capturing extreme flows, while traditional models resonated more with medium flow regimes. This research offers vital insights for hydrologists and decision-makers, aiding in informed model selection for streamflow predictions.