Neural network assessment of groundwater contamination of US Mid-continent

Artificial neural networks (ANNs) were applied to data taken from 1,302 domestic and rural hydraulic wells in the Mid-continent of the USA including Illinois, Iowa, and other 12 States to predict the contamination of the groundwater with pesticides. Preliminary hydrogeological and geostatistical analyses were carried out to assess groundwater vulnerability and data variability and weight, where data attributes were pre-processed and grouped into three main categories: hydrologic, human interaction, and climatic groups. ANNs are computer parallel-based systems that are characterized by their topologies, transfer functions, and learning algorithms. The backpropagation network (BP-NN) learning algorithm, used here, involves incremental adjustment of a set of parameters to minimize the error between the desired output and the actual output. Sensitivity analysis of the main BP-NN attributes was conducted to improve the BP-NN performance. Results of several trials demonstrated that the BP-NN have predicted the contaminated wells within each minor group in high precision. Sensitivity analysis revealed that BP-NN topologies and transfer functions were the main factors that affected its performance. It is evident that BP-NN is a powerful tool to predict the groundwater contamination over a wide area with limited data availability, which can provide an alternative cheap and effective tool to assess groundwater contamination worldwide.