Monitoring Microbial Quality of Source Waters Using Bayesian Belief Networks

Assessment of microbial quality of drinking water sources and recreational waters is highly dependent on measuring levels of Fecal Indicator Bacteria (FIB). However, there are limitations to FIB as accurate indicators of the presence of pathogens of concern, such as Cryptosporidium, as well as in the time-delay in FIB measurement. The objective of this work was to improve microbial water quality assessments by using data-driven approaches to predict day-to-day FIB, Escherichia coli (E. coli), and Cryptosporidium spp. concentrations based on simple water quality measures and weather data that can be monitored in real-time. To achieve this, we investigated the applicability of Bayesian Belief Networks (BBNs), an Artificial Intelligence-type algorithm that can provide probabilistic predictions using distinct data sources. Compared to other standard prediction methods, BBNs were observed to improve performance significantly and achieved greater than 75% accuracy for predicting FIB and E. coli concentrations. Following success in predicting FIB, predictability of Cryptosporidium, one of the major contributors to worldwide waterborne disease, was then investigated. One of the major challenges of modelling this parasite is dealing with a high number of non-detects and sparse data. Applicability of novel data augmentation and oversampling algorithms, such as the Adaptive Synthetic Sampling Algorithm (ADASYN), was investigated to improve the prediction model's performance. Leveraging the capability of the ADASYN significantly improved prediction performance of Cryptosporidium level presence, with 65% accuracy. The approach presented in this study can address the challenges of limited and time-delayed microbial water quality information and can be used to improve risk-based management and monitoring of water sources.