ML-based monthly runoff forecasting using gridded hydroclimatic products

This study investigated the temporal variation of hydroclimatic variables and their influence on monthly runoff forecasting using ensemble and kernel-based machine learning models. The Alafiya River basin situated across the River Alafiya in the state of Florida, United States, was selected for study. Three widely used machine learning models: Random Forest (RF), Extreme Gradient Boosting (XGB) and Support Vector Regression (SVR) were employed. This study utilizes a high-resolution Terra Climate (∼4 km) hydroclimatic dataset in monthly time steps (January 1958–December 2024), with input as well as target variables including rainfall (R), soil moisture (SM), actual evapotranspiration (AET) and runoff (Q). The input scenario including the combination of all hydroclimatic products (R + SM + AET) achieved the highest predictive performance (NSE > 0.80) in all employed models. SVR exhibited a higher percentage bias (−20 < Pbias < +5), indicating a consistent tendency to under- or overestimate runoff compared to RF and XGB. RF and XGB outperformed SVR by consistently achieving higher NSE and minimum error (RMSE and Pbias). These findings underscore the advantage of ensemble and kernel-based approaches for robust monthly runoff forecasting in data-rich hydroclimatic environments.