Estimation of Potential of the Ground Water Arsenic Contamination in Sanandaj Area Using Artificial Neural Network Model

  Identification of ground waters contaminated by arsenic using surface soil parameters and modeling this relationship in two models including artificial neural network and multiple linear regression can be useful in managing the water resources of the region.
Material & The purpose of the present study was to estimate the potential of arsenic pollution in the Sanandaj ground waters using multiple linear regression (MLR) and artificial neural
network (ANN) models. In this regards, 35 number of wells were selected among the permissible wells with considering watershed area, appropriate distribution, and different geological structure. The water samples stored in polyethylene bottles and kept at 4°C until transferred to the laboratory. For consideration of the relationship between the soils characteristics around the wells and ground water, the soil samples were collected from 0-20 cm of topsoil with composite sampling technique. The soil samples were air-dried and prepared for analysis. For long term storage of water samples nitric acid were added and the concentration of arsenic in water samples were measured by graphite furnace atomic absorption analyzer. Physical and chemical characteristics of the soil samples including: arsenic, arsenate, arsenite, phosphate, nitrate, total iron, amorphous iron, total manganese, amorphous manganese, clay, sand, silt, organic matter, pH and CEC were measured. Then all water and soil data were normalized and finally, accuracy of the MLP and ANN models was assessed to investigate the relationship between arsenic of water and soil parameters. Results showed that the arsenic concentration of ground waters were lower than the standard level in the study area. This can be due to high concentration of arsenate in the study area soils
compared arsenite and increasing the cationic exchange capacity of soil under the influence of clay particles, organic matter and free iron oxides. Compression of models accuracy result showed that ANN model with R=0.835, RMSE=0.156 and MAE =0.118 in the training phase and R =0.816, RMSE=0.177 and MAE=0.158 in the testing phase has higher accuracy and lower errors in the estimation of ground waters arsenic contamination than MLP model