Optimization of DRASTIC Model by Support Vector Machine and Artificial Neural Network for Evaluating of Intrinsic Vulnerability of Ardabil Plain Aquifer

With respect to population growth and agricultural development in Ardabil Plain, vulnerability assessment of the plain aquifer is necessary for management of groundwater resources and the prevention of groundwater contamination. In this study, vulnerability of Ardabil plain aquifer to pollution was evaluated by DRASTIC method. DRASTIC model was prepared by seven effective parameters on vulnerability, including groundwater depth, net recharge, aquifer media, soil media, topography, impact of vadose zone, and hydraulic conductivity as seven raster layers at 1:30000 scales. Then DRASTIC index was calculated after ranking and weighting that it was obtained 82 to 151 for Ardabil plain. The support vector machine (SVM), feedforward network (FFN) and recurrent neural network (RNN) models were adapted for optimizing the DRASTIC model to obtain the most accurate results of vulnerability evaluation. For this purpose, the DRASTIC parameters and the vulnerability index were defined as inputs data and output data respectively for models, and nitrate concentration data were divided in two categories for training and testing.DRASTIC index in training step was corrected by the related nitrate concentration, and after model training, the output of model in test step was verified by the nitrate concentration. The results show that 3 models of artificial intelligence are able to assessment of aquifer vulnerability, but the Support vector machine (SVM) with the least value of RMSE for all Eastern, Western and Southern parts of the plain is 6.74, 3.93 and 3.78, respectively and the highest value of R2 is 0.73, 0.79 and 0.72, respectively had the best results in the test step.According to this model, the northern and western parts of the plain are classified as high pollution potential areas and should be more protection of these areas.