Comparative analysis of ensemble learning algorithms in water quality prediction

Water is essential for all life forms but is increasingly at risk of contamination. Monitoring water quality is crucial to protect ecosystems and public health. This study evaluates ensemble learning techniques - AdaBoost, Gradient Boost, XGBoost, CatBoost, and LightGBM - for predicting key water quality parameters in the Bara River Basin, Pakistan. Initially, a random forest model identified optimal input-target parameter combinations. Machine learning models were then developed and evaluated using R2, MSE, and MAE, with the best models selected via compromise programming. Results show XGBoost and Gradient Boost outperformed other methods. XGBoost achieved near-perfect R2 values for bicarbonate (HCO3), carbonate (CO3), and magnesium (Mg), while Gradient Boost excelled with parameters like electrical conductivity (EC), sulfate (SO4), temperature, and calcium (Ca). XGBoost demonstrated high training R2 values (0.999) but slightly lower testing R2 (e.g., 0.8636 for HCO3). Gradient Boost exhibited greater stability, maintaining high accuracy in both phases (e.g., Ca testing R2 = 0.9433). AdaBoost and CatBoost showed moderate performance for parameters like chloride (Cl) and pH, while CatBoost and LightGBM performed well for pH and dissolved solids but varied across other indicators. These findings underscore the potential of ensemble methods for accurate water quality prediction, aiding future management and environmental protection efforts.