Predicting the concentration range of trace organic contaminants in recycled water using supervised classification

Trace Organic Contaminants (TrOCs) have evidence for many health and environmental issues. Frequent monitoring of TrOC concentration is a time-consuming and costly process, which cannot be achieved easily. Identifying surrogate markers for these contaminants is a practical solution to monitor and ensure water quality. However, this topic is seldom explored in previous literature. This study aimed to find surrogate markers to predict concentration class (i.e., three classes: low concentration, medium concentration, high concentration) for a set of widely used pharmaceutical and personal care product TrOCs (e.g., Fluoxetine, Primidone, Saccharin, Sucralose to name a few) in recycled water from Melbourne Eastern Treatment Plant (ETP), Melbourne, Australia. For this purpose, three popular supervised learning classification algorithms namely Naïve Bayes, Random Forest and Support Vector Machines were utilized. Physicochemical parameters colour, Chemical Oxygen Demand (COD) and Total Organic Carbon (TOC) were found to be the top three predictive features for the majority of the investigated TrOCs. UV Transmittance (UVT) and the total amount of suspended solids (TSS) were the next frequent features. The Random Forest model resulted in the highest classification accuracy (≥73 %) for the majority of compounds. This paper presents evidence that with the acquired intelligence of supervised machine learning, the concentration range of hard to measure TrOCs in water can be predicted from a handful of low-cost and easy-to-measure physicochemical parameters.