Predicting soil–water distribution coefficients of heavy metals using artificial neural networks

Subject Areas : Farm water management with the aim of improving irrigation management indicators

Amin Falamaki ¹ , Mahnaz Eskandari ²

1 - Department of Civil Engineering; Payame Noor University, PO Box 19395-3697, Tehran, Iran
2 - Department of Soil Science, Science and research Branch, Islamic Azad University, Tehran, Iran

Received: 2012-04-07 Accepted : 2012-09-20 Published : 2012-12-10

Keywords: Artificial Neural Networks, Cadmium, chromium, soil–water distribution coeffi,

Abstract :

Contamination of soil and water resources is a major concern for optimal use of these resources worldwide. The so-called distribution coefficient (Kd) is an applied parameter not only for modeling contaminant transport in soil but also for risk analysis of soil and water resources contamination. The most common quantitative model for estimating Kd is parametric method. The correlation coefficient of this model is usually low, however, the predicted Kd values may cause significant inaccuracy in predicting the impacts of contaminant migration or siteremediation options. The objective of this study was to investigate application of artificial neural networks (ANN) for improving Kd prediction of heavy metals. Consequently, three ANN types including multi layer perceptron (MLP), redial basis function (RBF) and hierarchical networks (HN) and two heavy metals of Chromium (VI) and cadmium were used for modeling purposes. The collected data were first divided into two training and test groups. The first group was used to train ANN and the second to evaluate generalized ANN models. The most suitable geometry of networks were obtained with trial and error procedure. The results of modeling Kd(Cr) revealed that both MLP and RBF networks are reasonable tools, but MLP was more accurate than RBF. Although the applied input data for training networks were not so much (at least 9 and the maximum of 16), but they were sufficient for modeling Kd(Cr). This finding is a promising result because direct measurement of Kd is expensive and time consuming. Further, usually limited numbers of available data are existing in each case. The results of predicting Kd(Cd) approved the preferences of MLP for modeling purposes. The ANN model can significantly enhance the correlation coefficient between predicted and measured data form 0.37 of parametric method to 0.63.