Use of machine learning and geographical information system to predict nitrate concentration in an unconfined aquifer in Iran

Increased nitrate concentration is one of the main groundwater quality problems today that needs to be measured and monitored. Water quality testing and monitoring are time consuming and costly. Therefore, new modeling methods such as machine learning algorithms can be used as an efficient solution for predicting nitrate concentration. In this study, three machine learning methods including deep neural network (DNN), extreme gradient boosting (EGB), and multiple linear regression (MLR) were used to predict nitrate contamination in groundwater in the north of Iran (Mazandaran plain) and finally the best method was selected for mapping. The mean nitrate concentration in 250 piezometric wells was considered as output variable and the factors affecting groundwater quality (groundwater depth, transmissivity of aquifers, precipitation, evaporation, distance from water resources and Caspian Sea, distance from industries and residential centers, population density, topography, and exploitation from groundwater) as input variables in an alluvial aquifer. The same training and testing data were used in the modeling process of the three machine learning methods. The results of the training and testing stages showed that the EGB method has the highest performance in predicting nitrate concentration due to the lowest error values and highest correlation between the measured and predicted values of nitrate concentration (training R-sqr = 0.98, Nash–Sutcliffe efficiency (NSE) = 0.98, and test R-sqr = 0.86, NSE = 0.84). Further, the results indicate that the factors distance from industries, population density, groundwater depth, and evaporation rates are the most important factors affecting nitrate concentration in groundwater. Finally, the tested EGB model and a geographic information system (GIS) tool were used to prepare a map of groundwater nitrate pollution in the study area. Evaluating the performance of the resulting map by comparing the predicted and measured values indicated a good accuracy (R-sqr = 0.8)