Spatial mapping of water spring potential using four data mining models

Abstract

Abstract Population growth and overexploitation of water resources pose ongoing pressure on groundwater resources. This study compares the capability of four data mining methods, namely, boosted regression tree (BRT), random forest (RF), multivariate adaptive regression spline (MARS), and support vector machine (SVM), for water spring potential mapping (WSPM) in Al Kark Governorate, east of the Dead Sea, Jordan. Overall, 200 spring locations and 13 predictor variables were considered for model building and validation. The four models were calibrated and trained on 70% of the spring locations (i.e., 140 locations) and their predictive accuracy was evaluated on the remaining 30% of the locations (i.e., 60 locations). The area under the receiver operating characteristic curve (AUROCC) was employed as the performance measure for the evaluation of the accuracy of the constructed models. Results of model accuracy assessment based on the AUROCC revealed that the performance of the RF model (AUROCC = 0.748) was better than that of any other model (AUROCC SVM = 0.732, AUROCC MARS = 0.727, and AUROCC BRT = 0.689).