Daily Rainfall-Runoff Modeling at Watershed Scale: A Comparison Between Physically-Based and Data-Driven Models

In the last decades, data-driven (DD) machine-learning models have been rapidly developed and widely applied to solve hydrologic problems. To explore DD approaches' capability in rainfall-runoff modeling compared to knowledge-driven models, we conducted a thorough comparison between Soil & Water Assessment Tool (SWAT) and Random Forest (RF) models. They were implemented to simulate the daily surface runoff at Santa Lucia Chico watershed in Uruguay. Aiming at making a fair comparison, the same input time series for RF and SWAT models were considered. Both approaches are able to represent the daily surface runoff adequately. The RF model shows a higher accuracy for calibration/training, while the SWAT model yields better results for validation/testing, indicating that the latter has a better generalization capacity. Furthermore, RF outperforms SWAT in terms of computational time needed for a proper calibration/training. Strategies to improve RF performance and interpretability should include feature selection, feature engineering and a more sophisticated sensitivity analysis technique.