Comparative evaluation of advanced machine learning models for streamflow prediction in a data-scarce mountainous region: a case study of Kohgilueyeh and Boyer-Ahmad, Iran

ABSTRACT Accurate streamflow prediction plays a vital role in managing water resources, especially in regions with climatic variability, complex topography, and limited hydrological data. This study evaluates the performance of four machine learning models – Radial Basis Function Network (RBFN), Support Vector Regression (SVR), Random Forest (RF), and Extreme Gradient Boosting (XGBoost) – for daily streamflow prediction across four hydrometric stations (Batari, Pataveh, Kata, and Tangeh Berim) in Kohgiluyeh and Boyer-Ahmad province, southwestern Iran. The models used meteorological and hydrological inputs including temperature, precipitation, evaporation, sunshine hours, and discharge data from 1999 to 2021. Model performance was assessed using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Nash–Sutcliffe Efficiency (NSE) during both training and testing phases. Results revealed that SVR and XGBoost outperformed the others, achieving NSE values up to 0.99 in training and 0.98 in testing, along with consistently low RMSEs. RBFN demonstrated moderate accuracy, while Random Forest performed poorly, particularly in stations with high streamflow variability. The findings underscore the effectiveness of SVR and XGBoost in modeling nonlinear hydrological processes in data-scarce mountainous environments. These models offer reliable tools for enhancing flood forecasting, drought risk assessment, and sustainable water resource planning in topographically complex regions.