Modular neural networks to predict the nitrate distribution in ground water using the on-ground nitrogen loading and recharge data

Artificial neural networks have proven to be an attractive mathematical tool to represent complex relationships in many branches of hydrology. Due to this attractive feature, neural networks are increasingly being applied in subsurface modeling where intricate physical processes and lack of detailed field data prevail. In this paper, a methodology using modular neural networks (MNN) is proposed to simulate the nitrate concentrations in an agriculture-dominated aquifer. The methodology relies on geographic information system (GIS) tools in the preparation and processing of the MNN input-output data. The basic premise followed in developing the MNN input-output response patterns is to designate the optimal radius of a specified circular-buffered zone centered by the nitrate receptor so that the input parameters at the upgradient areas correlate with nitrate concentrations in ground water. A three-step approach that integrates the on-ground nitrogen loadings, soil nitrogen dynamics, and fate and transport in ground water is described and the critical parameters to predict nitrate concentration using MNN are selected. The sensitivity of MNN performance to different MNN architecture is assessed. The applicability of MNN is considered for the Sumas-Blaine aquifer of Washington State using two scenarios corresponding to current land use practices and a proposed protection alternative. The results of MNN are further analyzed and compared to those obtained from a physically-based fate and transport model to evaluate the overall applicability of MNN.