CP-2008-11
Use of Artificial Neural Networks for Predicting of Groundwater Contamination
Sahoo, Goloka Behari, and Chittaranjan Ray
Groundwater: Modelling, Management and Contamination, eds. L.F. König and J.L. Weiss, 3–14 pp. (2009)
Artificial neural networks are empirical mathematical tools proven to represent complex relationships of hydrological systems. Neural networks are increasingly being applied in subsurface modeling where intricate physical processes and lack of detailed field data prevail. Two types of ANN models: Back propagation neural network (BPNN) and radial basis function neural network (RBFN) are examined to predict the pesticide contamination of domestic wells. Because sample collection, analysis, and re-sampling are expensive, a large dataset is not available for ANN use in this study. This study presents analyzes of raw data for preparation of input subsets for ANN use. Thus, a clustering technique is used to divide the whole dataset into three subsets: training, validating, and testing. The sensitivity analysis was carried out by deleting one or more input variables from the input data set to measure the importance of one variable over the other in terms of ANN prediction performance. It provides a sense of the effect of each parameter on pesticide occurrence in a well. The well depth, depth to aquifer material from land surface, and on-site pesticide storage are found to be important parameters in pesticide occurrence in well.