Modeling groundwater quality by using hybrid intelligent and geostatistical methods

Simulation of groundwater quality is important for managing water resources and mitigating water shortages, especially in arid and semiarid areas. Geostatistical models have been used for spatial prediction and interpolation of groundwater parameters. Recently, hybrid intelligent models have been employed for the simulation of dynamic systems. In this study, hybrid intelligent models, based on a neuro-fuzzy system integrated with fuzzy c-means data clustering (FCM) and grid partition (GP) models as well as artificial neural networks integrated with particle swarm optimization algorithm, were used to predict the spatial distribution of chlorine (Cl), electrical conductivity (EC), and sodium absorption ratio (SAR) parameters of groundwater. Results of the hybrid models were compared with geostatistical methods, including kriging, inverse distance weighting (IDW), and radial basis function (RBF). The latitude and longitude values of observation wells and qualitative parameters in three states of maximum, average, and minimum were introduced as input and output to the models, respectively. To evaluate the models, the root mean squared error (RMSE), the mean absolute error (MAE), and CC statistical criteria were used. Results showed that in the hybrid models, NF-GP with the lowest RMSE and MAE and highest CC was the most suitable model for the prediction of water quality parameters. The RMSE, MAE, and CC values were 107.175 (mg/L), 79.804 (mg/L), and 0.924 in the average state for Cl; were 518.544 (μmho/cm), 444.152 (μmho/cm), and 0.882 for electrical conductivity; and were 1.596, 1.350, and 0.582 for sodium absorption ratio, respectively. Among the geostatistical models, the kriging was found more accurate. Using the coordinates of wells will eventually allow the NF-GP to be used for more sampling and replace the visual techniques that require more time, cost, and facilities.