Cd modeling of soils around Garmsar salt mines based on artifical neural network (MLP) model

During the past two decades, computer aid models for simulation of heavy metals have been remarkably developed. Prediction of soil pollution plays an important role in pollution control and land management. But in large areas, collecting data in a direct way is challenging in terms of cost and time. In recent years, the use of indirect methods such as artificial neural network (ANN) and other similar models to estimate heavy metals has been considered. There are 27 salt mines in Garmsar city. Of these, 16 mines are active. Salt extracted from these mines are used as one of the food spices. On the other hand, due to mining activities, the soils of this region may be contaminated with heavy metals. Therefore,in this study, the effectiveness of terrain and spectral indices for predicting total soil Cadmium (Cd) around the soils of Garmsar salt mines was evaluated by ANN – multilayer perceptron (MLP) model.

For this research, 49 soil samples were collected from the 0-20 cm Physicochemical properties of soil samples such as percentage of clay, sand, silt, soil acidity (pH), electrical conductivity (EC) and lime percentage were determined. Total Cd concentration was measured by atomic absorption spectroscopy (AAS) (Varian, Spectra 220). All terrain attributes used in this study were derived from a digital elevation map (DEM) and to calculate the spectral indices, Landsat-8 OLI/TIRS bands image with a resolution of 30 meters were used. Twenty-five auxiliary data variables derived from a DEM and Landsat-8 were used to predict total soil Cd in the study area. Based on the auxiliary data obtained and the correlation coefficients between these data and the predicted total Cd value, 2 models were evaluated. The collected data were randomly divided into categories training and validation and were used to evaluate the MLP model.

The results of this study show that the auxiliary data extracted from landsat-8 bands (with the highest accuracy and lowest error rate) were the effective parameters in predicting soil contamination with Cd. Based on the results obtained from the evaluation of ANN performance in estimating total Cd, the value of root mean square error (RMSE) and coefficient of explanation (R2) were 0.05 and 0.95 for the first model and 0.10 and 0.80 for the second model. In model 1, saturation index (Sat I), grain size index (GSI), carbonate index (CrI), soil color index (color I) and gypsum index (GI) were important and main parameters in total Cd modeling. The results of the present study showed the high efficiency of the ANN model in predicting total soil Cd.

Due to the development of machine learning models in the field of environmental engineering especially in simulation of heavy metals, having a turning point for their advancement is very important. The results of this research show that the MLP model is suitable for total soil Cd prediction and this method can save the cost of soil sampling and analysis. Therefore, it is recommended to validate the method applied in this study to prepare total soil Cd map in similar areas.