The main goal of the study was to estimate the efficiency and predictive ability of a series of tree-based ensemble methods in groundwater spring potential mapping. The study introduces a novel hybrid integration approach based on J48 Decision Trees (J48), AdaBoost (AB), Bagging (Bag), RandomSubSpace (RS), Dagging (Dag) and Rotation Forest (RF) algorithms for constructing a groundwater spring potential map. The performance of the ensemble models were evaluated at the Wuqi County in Shaanxi Province, China. At first, a groundwater spring inventory map with 235 groundwater springs was constructed. The groundwater spring database was randomly divided into a training (70% of the total number of groundwater springs) and validating subset (remaining 30%). Secondly, sixteen groundwater spring related variables were selected and analyzed, such as elevation, slope, aspect, planform and profile curvature, curvature, topographic wetness index, stream transport and stream power indexes, distance from river network, lithology, soil and landuse cover, normalized difference vegetation index, mean annual rainfall and distance from road network. The contribution of each groundwater spring related variable so as to identify the most predictive was based on the correlation attribute evaluation method. The predictive performance of each model was evaluated by estimating the area under the receiver operating characteristic (ROC) curve (AUC) and several statistical indexes (accuracy, sensitivity, specificity). The outcomes revealed that all models had a good predictive performance, having AUC values greater than 0.74. Specifically, RF-J48 model appeared with the highest prediction capability (AUC = 0.797), followed by the AB-J48 model (0.793), the Bag-J48 (0.768), the RS-J48 (0.766) and the Dag-J48 (0.748), respectively. The study highlights that the proposed ensemble methodology is efficient and highly accurate appropriate for groundwater spring potential mapping.
