Evaluation of Tree-Based Machine Learning Techniques for Prediction of Stream Reaeration Coefficient

Abstract

Dissolved oxygen (DO) is one of the most important indicators of river water quality and ecological health. Accurate DO estimation is essential for assessing the assimilative capacity of streams, maintaining riverine ecosystems, and designing effective water-quality management and restoration strategies. In many rivers, wastewater discharge from urban areas significantly depletes DO, thereby impairing the water body’s natural self-purification capacity. Under such conditions, the reaeration coefficient ( K 2 ), which governs oxygen transfer across the air–water interface, plays a critical role in replenishing dissolved oxygen and supporting pollutant degradation without adversely affecting river health. For the last few decades, machine learning (ML) techniques have gained significant popularity due to their predictive capabilities. In this study, tree-based ML models, including random forest (RF), gradient boosting (GB), decision tree (DT), and extra tree (ET), are used to predict K 2 for streams. A data set comprising 366 field measurements from multiple river systems reported in the literature was used in this study, covering a broad range of flow velocity ( V ), flow depth ( h ), and bed slope ( S ) conditions. The ML models employed in this study used GridSearchCV for hyperparameter optimization along with k -fold cross-validation. To evaluate the performance of these models, five performance metrics, namely, correlation coefficient (CC), root-mean square error (RMSE), Nash–Sutcliffe efficiency (NSE), Willmott’s index of agreement ( I A ), and percent bias (PBIAS), are used. Results showed that the RF algorithm is the most accurate, with a CC of 0.904. The ML model RF was compared with existing empirical equations, and it was observed that RF outperformed in terms of accuracy and generalization. Shapley additive explanations (SHAP) was used to understand the influence of each input on the model’s output prediction. Sensitivity analysis using SHAP showed that flow velocity ( V ) is the most important feature affecting the reaeration coefficient. The Monte Carlo simulation framework was used to carry out uncertainty analysis in predicting K 2 using the proposed models, and results highlighted the RF model’s superiority, showing the most reliable uncertainty performance, achieving a P -factor of 0.946, with the lowest R -factor (1.667) and d -bar (0.234) compared with the other ML models and empirical equations, securing the first rank.