Resampling Techniques for Imbalanced Water Quality Classification

Early detection of water quality status is crucial for preventing water pollution that could negatively impact community health. However, the infrequency of pollution events in water quality data leads to a class imbalance issue, where one class has significantly more observations than the other. Training machine learning models on imbalanced data can result in overfitting, reducing sensitivity and impairing predictive performance. Therefore, this study uses various oversampling techniques, including Synthetic Minority Oversampling Technique (SMOTE), Random Oversampling (ROS), Rapidly Converging Gibbs Sampler (RACOG) and Adaptive Synthetic Oversampling (ADASYN), and under-sampling techniques such as Random Under-Sampling (RUS), to balance the data before fitting into the machine learning. Secondary data on multiple water quality parameters, which are from the Department of Environment Malaysia, were utilized. The dataset consisted of a binary target variable, which is the water quality classification (WQC) and nine physicochemical parameters. The performance of artificial neural networks (ANN), support vector machines (SVM), and gradient boosting (GB) classifiers trained on the balanced dataset was assessed using balanced accuracy, sensitivity, f-measure, and Area Under the Curve (AUC). The results showed that the optimal performance for gradient boosting was achieved with ROS samples, yielding a balanced accuracy of 89.65%, sensitivity of 82.50%, f-measure of 84.62%, and AUC of 98.35%. In contrast, the best performance for ANN was achieved with RACOG samples, while for SVM, ADASYN samples produced the best results. Different sampling techniques showed the best results for different models because each machine learning algorithm has unique ways of learning patterns from data, and different resampling techniques address class imbalance in different ways.