Study of Spatial Distribution of Groundwater Quality Using LS-SVM, MLP, and Geostatistical Models

Abstract

Groundwater quality control is of great importance in (semi-)arid zones due to the water deficit in these regions. Geostatistical models are techniques commonly developed for the interpolation and spatial prediction of groundwater quality parameters. In this study, IDW, Kriging, and CoKriging methods were used in the geostatistical, LS-SVM, and MLP models to predict the spatial distribution of groundwater EC. The models were then compared in terms of their efficiency. For the purposes of this study, data were collected from 120 wells in the Mashhad plain. Variograms were then drawn after normalizing the data for application in the geostatistical models. In the next stage, the lowest RSS value was used for selecting the one model that was suitable for fitting the experimental variogram while cross-validation and RMSE were used to select the best method for interpolation. Comparison of the three models in question was accomplished by using 25% of the observation data and the statistical parameters of RMSE, R2, and MAE were determined. Results showed that the CoKriging method outperformed its Kriging counterpart in the geostatistic model for interpolating groundwater quality. Finally, the most accurate values for the quality parameters (i.e., R2=0.932, RMSE=367.9, MAE=265.78() were obtained with the MLP model.

Keywords

• Variogram
• Multilayered perceptron (MLP)
• Least Squares Support Vector Machine (LS-SVM)