Predictive Performance of Ensemble Learning Boosting Techniques in Daily Streamflow Simulation

Although many modeling techniques are available, accurate runoff predictions, particularly for high-flow events, are challenging on a daily time scale. Ensemble learning is a powerful machine learning technique that has recently gained attention in hydrologic studies. This study investigates the suitability of Boosting Ensemble Models for daily rainfall-runoff simulations, incorporating hyperparameter tuning through optimization, a step that has rarely been undertaken in previous studies. Adaptive Boosting (AdaBoost), Gradient Boosting (GB), Stochastic Gradient Boosting (SGB), eXtreme Gradient Boosting (XGB), Categorical Boosting (CatBoost), Light Gradient Boosting (LGB), and Natural Gradient Boosting (NGB) are the models developed for the Meenachil River Basin, India. These are Interpretable AI (IAI) models, designed to offer transparency and insights into the decision-making process. Hyperparameter optimization is performed using GridSearchCV for AdaBoost and Tree Parzen Estimators (TPE) for the remaining models. This significantly enhances model performance, particularly for AdaBoost and NGB. Model performance evaluation indicates excellent predictive ability, with NGB (Kling-Gupta Efficiency = 0.98, Nash-Sutcliffe Efficiency = 0.98) outperforming the other models. The performance follows the order NGB > CatBoost > LGB > XGB > GB > SGB > AdaBoost. Model performance under high flow is analyzed separately, and the results highlight the superiority of NGB over the other models. Feature importance extraction is performed to interpret the models, identifying the most influential features that drive the outcomes. This study thus provides insight into the applicability of boosting ensemble models for rainfall-runoff simulation on a daily scale.