From low-cost measurements to rapid water quality monitoring: a hybrid machine learning-optimization framework for predicting laboratory-intensive parameters.

Abstract

Reliable estimation of laboratory-intensive water-quality variables remains challenging in spatially heterogeneous river systems. This study developed a machine-learning framework for predicting BOD, COD, DO, NH₄, NO₂, NO₃, and TA across major rivers of Khuzestan Province, Iran, using a 12-year monitoring dataset. After temporal matching and quality-control screening, 1427 observations from 26 stations were retained for the entire-network scenario. Eight predefined spatial scenarios were evaluated using feature construction, MIC-based feature selection, nine machine-learning algorithms, and multiple hyperparameter-optimization strategies. Model generalization was assessed using expanding-window temporal and station-grouped cross-validation. The best temporal median KGE values ranged from 0.699 to 0.844 for six targets, while NH₄ showed lower temporal performance (0.467); station-grouped KGE values ranged from 0.667 to 0.866. RF, XGBoost, and SVM were most frequently represented among the best-performing models, while GWO, HHO, PSO, BO, and SSA emerged among the leading optimization strategies. No single model, optimizer, or spatial scenario was universally superior. Overall, the results highlight the importance of target-specific modelling and structured validation and demonstrate the potential of spatially informed machine learning to support water-quality monitoring and decision-making in heterogeneous river networks.