A comparative study of ensemble and non-ensemble machine learning methods for predicting river pollution index

Accurate prediction of river water quality is fundamental to environmental sustainability and public health, particularly amid increasing freshwater scarcity. This study develops a robust Machine Learning (ML) framework to forecast the River Pollution Index (RPI) using a comprehensive 36-year national dataset from Taiwan’s Environmental Protection Administration, covering over 500 monitoring stations. We conducted a systematic comparison of ensemble methods (CatBoost, XGBoost, NGBoost) and non-ensemble benchmarks (SVM, ElasticNet, and 1D CNN). Hyperparameters were optimized via Bayesian optimization, and statistical significance was ensured by evaluating model stabilityusing a suite of complementary indicators (RMSE, MAE, R 2 , A10 index) across 30 independent experimental runs. The results demonstrated the consistent superiority of ensemble models over non-ensemble counterparts. Among them, CatBoost achieved the highest accuracy and stability (RMSE ≈ 0.85, MAE ≈ 0.61, R 2 = 0.78), reducing prediction error by approximately 20% relative to SVM and ElasticNet. These findings highlight the capacity of ensemble learning techniques to capture complex, non-linear interactions inherent in water quality data. The study makes two principal contributions: (1) the systematic implementation, optimization, and comparison of ensemble and non-ensemble ML models for river pollution prediction on a long-term national dataset; and (2) the identification of ensemble-based methods, particularly CatBoost, as robust and data-driven tools to enhance RPI forecasting and to support informed decision-making in sustainable water resource management. • Evaluated ML models on a 36-year nationwide river pollution dataset from Taiwan. • Compared ensemble (CatBoost/XGBoost/NGBoost) vs non-ensemble methods (SVM/CNN). • CatBoost achieved superior stability and accuracy (RMSE ≈ 0.85, R 2 = 0.78). • Ensemble models reduced prediction error by ∼ 20% versus non-ensemble benchmarks. • Bayesian optimization and 30 runs validated model robustness and reliability.