New insights into machine learning prediction techniques for real-time sanitary risk assessment in karst drinking water sources affected by faecal contamination

Abstract

Safe drinking water supply from karst aquifers faces several challenges due to their high vulnerability to contamination, which can result in abrupt water quality variations in matter of hours. The analysis of faecal bacterial activity requires time-consuming cultures and expensive reagents in the laboratory, which can delay the detection of an imminent contamination event. An innovative methodology is proposed in this research to overcome these limitations and provide real-time insights about water quality in drinking sources. Fieldwork activities included the continuous monitoring of water parameters (spring discharge, electrical conductivity, turbidity and Tryptophan-like fluorescence) and groundwater sampling for Escherichia coli determination in two springs draining a binary karst aquifer in S Spain during three hydrological years (2020/21 to 2022/23). Ten supervised Machine Learning models were then tested to infer five sanitary risk levels (based on E. coli activity) from continuous measurements at the springs. The combination of two water parameters was the most effective predictor at the two drinking water sources, which showed different optimal configuration of proxy parameters depending on their hydrogeological features and contaminant transport mechanisms. Gaussian Processes, Neural Networks, Naïve Bayes and Quadratic Discriminant Analysis provided the highest probability of correctly discriminating between sanitary risk levels. This methodology holds significant potential to be integrated as an early-warning protection tool for real-time sanitary risk assessment and, thus, safeguarding drinking water supplies worldwide against microbial threats.