Comparative evaluation of machine learning models for water quality classification using physicochemical parameters and SMOTE in tropical rivers

ABSTRACT Figure showing an interpretable framework for water quality parameters in tropical rivers Machine learning offers a promising approach for water quality assessment by modelling complex nonlinear relationships among physicochemical parameters. This study compares four machine learning classifiers. Artificial Neural Network (ANN), Decision Tree (DT), Support Vector Machine (SVM), and Random Forest (RF), for multi-class river water quality classification using five physicochemical parameters (pH, dissolved oxygen, biochemical oxygen demand, total suspended solids, and ammonium nitrogen) collected from three tropical rivers in Penang, Malaysia, between July 2023 and August 2024. We defined water quality classes according to the Malaysian Inland Water Quality Standards (INWQS), using a conservative maximum-parameter approach for overall classification. The dataset was divided into training (80%) and testing (20%) sets using stratified sampling, while Synthetic Minority Over-sampling Technique (SMOTE) addressed class imbalance. All models used identical preprocessing and were evaluated using accuracy, macro-averaged precision, recall, and F1-score. Random Forest achieved the highest predictive performance (95.0% accuracy; macro F1-score = 0.951), followed by Decision Tree (93.4%), SVM (90.1%), and ANN (89.3%). The deterministic INWQS rule achieved 100% accuracy, providing a baseline for contextualising machine learning performance. Overall, tree-based models demonstrated superior performance for structured tabular data, highlighting the value of SMOTE and rigorous comparative evaluation for tropical river water quality classification.