Groundwater contamination risk mapping constitutes an important component of groundwater management and quality
control. In the present study, we describe a method for such mapping that is more suitable for arid regions than other methods developed in previous work. Specifically, we integrate machine learning tools, interpolation and process-based models with a modified version of DRASTIC-AHP to evaluate groundwater vulnerability to nitrate contamination, and to map this contamination in Jiroft plain, Iran. The DRASTIC model provides a tool for evaluating aquifer vulnerability by using seven parameters related to the hydrogeological setting (depth to water, net recharge, aquifer media, soil media, topography, impact of vadose zone and hydraulic conductivity), while the criteria ratings and weights of these parameters are evaluated by means of an analytic hierarchy process (AHP). However, to obtain the risk map, the model predictions related to groundwater vulnerability are combined here with a contamination hazard map, which we estimate by applying ensemble modeling. This modeling builds on the occurrence probability predicted by means of a modeling framework that is based on generalized
linear modeling (GLM), flexible discriminant analysis (FDA) and support vector machine (SVM). We find that the application
of our ensemble modeling to predicting groundwater contamination in Jiroft plain leads to better results (AUC = 0.916,
Kappa = 0.89, MSE = 0.18 and RMSE = 0.11) compared to the separated employment of the various machine learning (ML)
methods, i.e., either SVM (AUC = 0.847, Kappa = 0.86, MSE = 0.19 and RMSE = 0.29), GLM (AUC = 0.829, Kappa = 0.81,
MSE = 0.23 and RMSE = 0.37) or FDA (AUC = 0.816, Kappa = 0.8, MSE = 0.26 and RMSE = 0.42). Our integrated modeling framework provides an assessment of both regional patterns of groundwater contamination and an estimate of contamination impacts based on socio-environmental variables, being particularly suitable for applications in which the amount of available data is scarce. The groundwater contamination risk map obtained from our case study shows that the central and southern regions of the Jiroft plain display high and very high contamination risk, respectively. This result is associated with the high production rate of urban waste in residential lands and an overuse of nitrogen fertilizers in agricultural lands throughout the study area. Therefore, while the present work introduces a new model which is applicable to arid regions in situations of scarce data availability, our results both provide insights for the future assessment of groundwater contamination in Jiroft plain and have potential impacts for the management and control of water resources in arid and semiarid environments.