Support vector machine and artificial neural network to model soil pollution: a case study in Semnan Province, Iran

Abstract

To study the extent of soil pollution in Shahrood and Damghan located in Semnan Province, Iran, 229 soil samples were taken and the levels of 12 heavy metals (Ag, Co, Pb, Tl, Be, Ni, Cd, Ba, Cu, V, Zn and Cr) were analyzed. Elevated values of some heavy metals such as Cr, Ni and V were detected in the study area. In order to predict soil pollution index (SPI) with respect to the concentration levels of 12 detected heavy metals, support vector machines (SVMs) with different kernels (linear, RBF and polynomial) and artificial neural networks (ANNs) were utilized. The database was repeatedly randomly split into training and testing data sets, and both SVMs and ANNs were trained and tested for each split. The testing results of the support vector regression (SVR) model with combinations of parameter sets were compared to optimize the parameters of SVMs with different kernels. The out-of-sample generalization ability of different kernels was roughly high and the same. Therefore, RBF kernel was selected for comparison with ANNs with early stopping. The correlation coefficients between the predicted and observed SPI for the RBF kernel and ANN with early stopping were 0.997 and 0.995, implying the same performance of these two methods. The results indicated that because of some problems associated with ANNs (such as local minima), for cases in which there are quite comparable results for ANNs and SVMs, the usage of SVMs is preferable.