A comparison of machine learning models for the mapping of groundwater spring potential

Abstract

Groundwater resources are vitally important in arid and semi-arid areas meaning that spatial planning tools are required for their exploration and mapping. Accordingly, this research compared the predictive powers of five machine learning models for groundwater potential spatial mapping in Wadi az-Zarqa watershed in Jordan. The five models were random forest (RF), boosted regression tree (BRT), support vector machine (SVM), mixture discriminant analysis (MDA), and multivariate adaptive regression spline (MARS). These algorithms explored spatial distributions of 12 hydrological-geological-physiographical (HGP) conditioning factors (slope, altitude, profile curvature, plan curvature, slope aspect, slope length (SL), lithology, soil texture, average annual rainfall, topographic wetness index (TWI), distance to drainage network, and distance to faults) that determine where groundwater springs are located. The area under the curve (AUC) of the receiver operating characteristic (ROC) curve was employed to evaluate the prediction accuracies of the five individual models. Here the results were ranked in descending order as MDA (83.2%), RF (80.6%), SVM (80.2%), BRT (78.0%), and MARS (75.5%).The results show good potential for further use of machine learning techniques for mapping groundwater spring potential in other places where the use and management of groundwater resources is essential for sustaining rural or urban life.