Neural network modelling for nitrate concentration in groundwater of Kadava River basin, Nashik, Maharashtra, India

Abstract

An attempt has been made to develop Artificial Neural Network (ANN) model for prediction of nitrate concentration in groundwater of Kadava River basin, Nashik District, Maharashtra. The study area lies between longitude 73°55:74°15’E and latitude 19°55’:20°25’N. River Kadava is one of the tributary of Godavari which originates in Sahyadri hills and flows towards NW to SE direction. Forty (40) representative groundwater samples were collected from dug/bore wells and analyzed for major cations and anions during pre and post monsoon seasons of 2012. Physicochemical results confirm that, 67.50% and 75% of groundwater samples having NO3 concentration beyond the permissible limit of the BIS (>45mg/L) in both the seasons. The spatiotemporal analysis inferred that, nitrate prone areas located in North and Central part of the study area, may be due to intense agriculture and overuse of nitrogen rich fertilizers and natural processes viz., dissolution, percolation and leaching. The consumption of high nitrate containing water is harmful to human health; consequently, it reduces the oxygen carrying capacity of the blood and in infant causes methemoglobinemia. In view of this the assessment and prediction of groundwater quality is an essential to predict the nitrate content to avoid future consequences, therefore there is need to develop a consistent, precise and resilient predictive model. The present study utilizes the algorithms viz., Levenberg - Marquardt three layer back propagation, traditional back propagation algorithm, resilient back propagation with and without weight algorithm, smallest absolute derivative (sag) and smallest learning rate (slr) for nitrate prediction. The Levenberg - Marquardt three layer back-propagation algorithm was found effective with 7 and 8 input neurons for pre and post monsoon season; 6 hidden neurons and nitrate content as a output variable. The efficiency of model is validated through coefficient of determination (R2), Residual Mean Square Error (RMSE) and Mean Absolute Relative Error (MARE) values. The present model gives satisfactory results and confirms consistent acceptable performance in both the seasons. The proposed ANN model may be helpful for similar studies and will be helpful to local public health bodies and policy makers to develop the management strategies.