Relative performance of artificial neural networks and regression models in predicting missing water quality data

Groundwater quality data are essential in providing valuable insight about the magnitude and source of contamination, as well as spatial and temporal variations. Under many circumstances, due to missing observations, forecasting, and backfilling of the groundwater quality data becomes mandatory. This study is aimed to investigate the potential of the artificial neural networks and the regression models for forecasting and backfilling the groundwater quality data. Sulfate, chemical oxygen demand, sodium, potassium, and phosphorus were chosen as dependent (output) variables. Chemical and the hydrometeorological data collected over a 2-year time period in an industrial area in India were used for developing these models. Artificial neural networks were trained using the backpropagation algorithm on four different feed-forward architectures as well as the radial basis function. Relative strength effect was used to examine the usefulness of the input variables. Model comparison statistics indicate that neural network techniques based on backpropagation algorithm training are better than the regression models and can be the effective modeling tool for predicting and backfilling the water quality data.