Comparing artificial neural networks and regression models for predicting faecal coliform concentrations

This paper compares the performance of ordinary least squares (OLS) and binary logistic regression methods, and artificial neural networks (ANNs) for the prediction of surface water faecal coliform concentrations in a 8.2 km(2) mixed land-use watershed. Model inputs consist of precipitation and temperature data, as well as instantaneous measurements of streamflow and conductivity. The ANNs are able to correctly classify 69% and 85% of faecal coliflorm concentrations relative to 20 and 200 cfu/100 mL water quality standards, respectively, results moderately better than those observed for the regression models. The ANN models using only meteorological inputs were able to correctly classify 72% and 81% of the observations relative to the 20 and 200 cfu/100 mL standards, respectively. The ANN models are notably better at predicting when the 200 cfu/100 mL standard is violated. In addition, the ANN models have lower percentages of false negatives, a characteristic desirable for protection of public health.