Comparative Analysis of Machine Learning Models for Predicting River Water Quality: A Case Study of the Zayandeh Rood River

Given the key role of rivers in supplying drinking water, supporting industry, agriculture, and ecosystems, water quality assessment and pollution quantification are essential for sustainable use. This study evaluated five machine learning models, i.e., Lasso Regression, Random Forest (RF), Gradient Boosting (GB), XGBoost, and Support Vector Machine (SVM) for predicting four water quality parameters—EC (Electrical Conductivity), TDS (Total Dissolved Solids), Sodium Adsorption Ratio (SAR), and TH (Total Hardness)—using data collected over a 31-year period from eight monitoring stations along the Zayandeh Rood River, a vital water source for drinking, agriculture, and industry in the arid region of central Iran. The models were evaluated based on five statistical criteria: R², RMSE, RRMSE, r, and MAE. Two dimensionality reduction techniques—PCA and correlation matrix-based feature reduction—were implemented to enhance model efficiency and mitigate multicollinearity. The findings indicate that the best-performing model for a given parameter varied across stations. However, the differences in evaluation metrics between the best models were quite low in most stations. The GB and SVM models outperformed other models in predicting EC, and TDS (0.80<R²<0.99). However, in predicting SAR, the GB and XGBoost models (0.955<R ²<0.999), and in predicting TH, the Lasso and SVM models achieved higher accuracy (0.830<R²<0.996). The Lasso regression model proved to be the most effective for predicting TH at half of the monitoring stations.