Modeling Fecal Indicator Bacteria in Urban Waterways Using Artificial Neural Networks

Abstract

Fecal indicator bacteria (FIB) are used as proxies to measure the microbial water quality of aquatic ecosystems. Methods of modeling FIB have evolved in order to provide accurate and timely prediction to inform decisions by governing authorities to prevent risks to public health. A predictive model to forecast the FIB concentrations of an urban waterway, the Chicago River, utilizing the artificial neural network (ANN) method was developed. To address tuning of hyperparameters of the ANN model, an exhaustive testing was performed to select optimal hyperparameters. The root-mean-square propagation (RMSprop) optimizer performed better than the stochastic gradient descent (SGD) and adaptive moment estimation (Adam) optimizers in this study. Eight input variables were eventually selected from 10 initially proposed variables: water temperature; turbidity; daily, 2-day, and 7-day cumulative rainfall; river flow discharge; distance from the upstream water reclamation plant; and number of upstream combined sewer outfalls. Water reclamation plants and combined sewer overflows were found to be critical contributors of microbial pollution in this urban waterway and should be considered in the ANN model. The developed model has an accuracy of 86.5% to predict whether fecal coliform concentration is above or below a regulatory threshold.