Examination of the use of artificial neural networks to model fecal indicator organism concentrations in urbanizing watersheds
Pathogen contamination is a primary source of surface water quality impairment in the United States and modeling tools to predict concentrations of indicator organisms have utility for those involved in watershed management and ecosystem restoration, as well as public health and source and recreational water protection. This research evaluated the use of artificial neural networks (ANNs) for simulating concentrations of fecal indicator organisms in the surface waters of the Gates Brook and lower Charles River watersheds in Massachusetts.
ANN model performance was assessed in terms of both the ability of the ANN models to accurately match observed indicator organism concentrations, as well as the ability of the models to correctly predict when a relevant water quality standard is violated due to exceedance of an indicator organism concentration. In addition to the fundamental question of ANN model performance, several other issues related to the development and implementation of the ANN models were explored, including the effects of different methods Of input selection and logarithmic transformation of data, the temporal transferability of the ANN models, and the potential for use of ANN models in ungaged watersheds. A comparison of ANN and ordinary least squares (OLS) regression models for fecal coliform prediction was also performed.
In both watersheds, for all combinations of input parameters considered, values of average absolute percent error (AAPE) and root mean square error (RMSE) are high. When the model performance assessment is based on the ability to identify violations of a water quality standard, the impression of model performance is quite different. The best performing ANN models in each watershed are able to predict approximately 60% to 90% of the violations of the 200 CFU/100 mL standard, which is the fecal coliform standard for primary contact recreation in Massachusetts. The ANN models developed for Gates Brook show performance that is comparable to and, in some situations, slightly better than other ANN and regression models developed for indicator organisms and pathogens.