Resampling-driven machine learning models for enhanced high streamflow forecasting

Accurate forecasting of high streamflow remains a significant challenge and is essential for sustainable water resource management and disaster mitigation, particularly due to the data imbalance often present during model development. This study proposes novel hybrid models through a comprehensive investigation of resampling techniques and machine learning algorithms. Four ensemble methods—Random Forest (RF), Extremely Randomized Trees (ET), Adaptive Boosting (ADA), and Extreme Gradient Boosting (XGB)—along with traditional methods such as Support Vector Regression (SVR) and K-Nearest Neighbors (KNN), were employed and compared for daily streamflow forecasting in the Thale Sap Songkhla Basin, southern Thailand. The key finding indicated that the recursive method consistently outperformed the direct method across all models. Additionally, combining original and resampled data enhanced forecast accuracy for various models. Even models such as RF, ET, ADA, and XGB, which typically show limited responsiveness to resampling, benefited to some extent from this approach. SVR demonstrated the highest sensitivity to resampling adjustments, particularly when paired with SVMSMOTE and Org-Resampling methods. KNN also exhibited notable improvements under several Org-Resampling strategies. These results present a promising framework for high streamflow prediction that can be adapted and applied to other river basins. • Accurate high streamflow forecasting achieved using hybrid machine learning models • Recursive forecasting method consistently outperforms the direct approach • Resampling techniques effectively address data imbalance in streamflow prediction • Combining original and resampled data (Org-Resampling) enhances forecast accuracy • SVR shows the highest sensitivity to resampling techniques