Effectiveness of groundwater heavy metal pollution indices studies by deep-learning

Abstract

Globally, groundwater heavy metal (HM) pollution is a serious concern, threatening drinking water safety as well as human and animal health. Therefore, evaluation of groundwater HM pollution is essential to prevent accompanying hazardous ecological impacts. In this aspect, the effectiveness of various groundwater HM pollution evaluation approaches should be examined for their level of trustworthiness. In this study, 226 groundwater samples from Arang of Chhattisgarh state, India, were collected and analyzed. Measured concentration for various HMs were further used to calculate six groundwater pollution indices, such as the HM pollution index (HPI), HM evaluation index (HEI), contamination index (CI), entropy-weight based HM contamination index (EHCI), Heavy metal index (HMI), and principal component analysis–based metal index (PMI). Groundwater in the study area was mainly contaminated by elevated Cd, Fe, and Pb concentrations due to natural and anthropogenic pollution. Moreover, this study explored the performance of deep learning (DL)-based predictive models via comparative study. Two hidden layers with 26 and 19 neurons in the first and second hidden layers, respectively, were optimised along with rectified linear unit activation function. A mini-batch gradient descent was also applied to ensure smooth convergence of the training dataset into the model. Results demonstrated that the DL-PMI scored lowest errors, 0.022 for mean square error (MSE), 0.140 for mean absolute error (MAE), and 0.148 for root mean square error (RMSE), in the model validation than the other DL-based groundwater HM pollution model. Prediction performances of all pollution indices were also verified using artificial neural network (ANN)-based models, which also highlighted the lowest validation error for ANN-PMI (MSE = 3.93, MAE = 1.38, and RMSE = 1.98). Furthermore, the prediction accuracies of PMI using both ANN and DL models scored the highest R2 value of 0.95 and 0.99, respectively. Therefore it is suggested that groundwater HM pollution using PMI as the best indexing approach in the present study area. Moreover, compared to benchmark, ANN, the DL performed better; hence, it could be concluded that the proposed DL model may be suitable approach in the field of computational chemistry by handling overfitting problems.