Neural network application in predicting Escherichia coli levels in Indianapolis Pleasant Run waterway in Indiana

Neural network and deep learning techniques are widely used in environmental predictive models. They are especially appropriate to predict time series trends and regress non-linear patterns in hydrological data. However, the application of neural network in assessing surface water quality has been limited. Escherichia coli (E. coli) is an important contaminant in urban streams, commonly correlated with climate and water quality parameters. High levels of E. coli (more than detection limit 24192 MPN/100mL) have been found in a small urban waterway in Indianapolis, and here we examine the effects of weather and water quality variables on E. coli trends using neural network models. Historical weekly E. coli data were available from August 1997 to March 2020. Corresponding weather and water quality data were available from the closest Indianapolis weather station, a United States Geological Survey (USGS) gage station monitoring discharge, and other Indianapolis monitoring sites. These variables included water temperature, air temperature, discharge, daily precipitation, precipitation over the previous 3 days, pH, dissolved oxygen, specific conductance, and total dissolved solids. These data were built as the input layer of a simple neural network and used to predict E. coli levels as the output layer. Predicted E. coli levels showed an increasing trend over the analysis period, mostly affected by the precipitation and stream discharge. The performance of the neural network forecasting of non-linear E. coli trends is fair, with an r2 of 0.24. More features need to be identified and added to the neural network structure and improve model performance. It suggests that neural network is feasible to estimate the water quality changes and forecast climate impacts on E. coli levels.