Comparative Analysis of Data Augmentation Techniques for Imbalanced Water Quality Classification: Prioritizing Public Safety through Enhanced Minority Class Detection

Abstract

Water quality monitoring encounters a significant challenge when dealing with imbalanced datasets where contamination events are rare but critical. In our study utilizing a dataset of 7,999 samples with 21 physicochemical parameters, unsafe water samples constitute merely 11% of the data, reflecting the natural imbalance found in real-world monitoring scenarios. To address this critical imbalance, this study offers a comprehensive evaluation of eight data augmentation techniques, encompassing traditional oversampling methods (SMOTE, ADASYN, Borderline-SMOTE), hybrid approaches (SMOTE-ENN), interpolation-based methods (MixUp), and contemporary generative approaches (CTGAN, TVAE), applied to water quality classification. We evaluate each technique across five classification algorithms: LightGBM, XGBoost, CatBoost, Random Forest, and Multi-Layer Perceptron. Through this extensive analysis, we demonstrate that the primary evaluation metric must be recall for the minority class (unsafe water), as false negatives pose severe health risks. Specifically, our experiments indicate that SMOTE-ENN achieves the highest recall of 97.77 % for unsafe water detection (using LightGBM and CatBoost), significantly surpassing the baseline approach's best performance of 85.47% (LightGBM), although this comes at the cost of a lower overall F1-score. Furthermore, we introduce a novel preprocessing pipeline incorporating outlier detection, feature engineering, and adaptive transformations that enhances model robustness. Consequently, our findings suggest that for safety-critical applications such as water quality monitoring, optimizing for minority class recall should take precedence over balanced accuracy metrics, thereby establishing SMOTE-ENN as the optimal choice for operational deployment.