Data driven water quality assessment using machine learning and synthetic data generation

Providing fresh and clean water to all is the objective of Sustainable Development Goal 6. Consuming clean water improves health for all living beings worldwide. Therefore, accurately estimating the water quality and classifying the type of water (i.e. whether it is used for potable purposes or not) based on physicochemical parameters is considered a challenging issue due to the dynamic nature of water quality data, the selection of physical or data-driven models used to estimate the water quality and the lack of a large water dataset for training the data-driven models. The primary objective of this research is to generate synthetic dataset(s) and propose state-of-the-art machine learning models for predicting water quality. In this study, synthetic data generation via oversampling is used to balance the water quality dataset and train models on a balanced dataset for water quality classification and prediction. The experimental results indicate that the quality of the synthetic Drinking Water Final dataset produced using SMOTE is satisfactory, as evidenced by a Maximum Mean Discrepancy (MMD) score of 0.0067. Furthermore, the test accuracy of the GB and XGB machine learning models is notably high at 99.47% on the synthetic Drinking Water Final dataset. The performance of the ML model on the synthetic Drinking Water Final dataset generated by a GAN is outstanding; however, the MMD score indicates that the quality of the synthetic data is subpar. Likewise, the MMD score quality of the synthetic Water Quality Analysis dataset produced using SMOTE (MMD = 0.0006) and GAN (MMD = 0.0016) is noteworthy. The comprehensive evaluation findings demonstrate that the synthetic dataset produced through oversampling techniques will enhance the models' predictive accuracy.