Accurate prediction of river flow remains a major challenge in hydrological systems due to nonlinear relationships and temporal variability governing streamflow behavior. Machine learning algorithms are increasingly used for this purpose because of their strong capability in modeling complex nonlinear interactions among hydrological variables. However, when a large number of input predictors are used, model complexity, computational cost, and the risk of overfitting increase, highlighting the importance of applying effective feature selection methods.

This study evaluates the impact of different feature selection strategies on the performance of machine learning models for monthly streamflow prediction of the Kashkan–Poldokhtar River. Two feature selection approaches SelectFromModel (SFM), based on model-derived feature importance, and recursive feature elimination with cross-validation (RFECV) were implemented alongside the inherent feature-handling mechanisms embedded within certain algorithms. Four machine learning models, including Random Forest (RF), Extreme Gradient Boosting (XGBoost), Least Absolute Shrinkage and Selection Operator (LASSO), and Support Vector Regression (SVR), were trained and tested. Their performances were evaluated using R², RMSE, MAE, and KGE metrics and compared with a climatology baseline. Results show that feature selection significantly influences the predictive capability of machine learning models. The RF–SFM model achieved the best testing performance (R² = 0.926, KGE = 0.934), demonstrating the effectiveness of SFM in identifying key predictors and improving forecast accuracy. The RF-RFECV model produced a simpler and more stable structure by iteratively removing less important features, making it suitable for interpretation and scientific analysis. XGBoost showed a tendency toward overfitting, while SVR exhibited the weakest performance on this dataset. Comparison with the climatology baseline confirmed the superiority of machine learning methods in representing complex hydrological dynamics.

Overall, selecting appropriate feature selection strategies is essential for balancing predictive accuracy and model simplicity and can improve operational streamflow forecasting and support water resources management