Gully erosion susceptibility mapping (GESM) is essential to take preventive or control measures, reduce environmental damages and economical costs produced by gully erosion in arid and semiarid area. The purpose of this study is to produce GESM using different data mining models in calcareous soils at Ekbatan Dam drainage basin. Three modeling technique, Generalized Linear Model (GLM), Random Forest (RF) and Multivariate Adaptive Regression Spline (MARS) are used and their results are compared for GESM. The study area is an 18.5 km2 located at Ekbatan Dam drainage basin (Hamadan-Iran), where agriculture activities are limited by intense erosion. By means of extensive field surveys, GPS route of gullies and visual interpretation of satellite images available from Google Earth, we prepared a digital map of the spatial distribution of gullies in the study area. To prepare explanatory variables that affected soil erosion, using the conditioned latin hypercube sampling (cLHS) technique and free survey 130 locations were sampled to create spatial distribution of some soil properties. DEM-derivatives, land use and normalized difference vegetation index (NDVI) maps created by stereo imagery of Spot-6 satellite. Geology map digitized from 1:100,000 scale geological scanned map of Hamadan Province. Road and river networks, derived from 1:25,000 scale topographic maps. The functional relationship between gully erosion occurrence and controlling factor were calculated using the mentioned models (GLM, RF, and MARS). The performance of each models was evaluated using the receiver operating characteristic (ROC), including the area under the curve (AUC), in 10-fold cross-validation. According to the results, RF showed the better performance (with mean AUC94.92%) in compared with MARS (with mean AUC87.31 %) and GLM (with mean AUC82.88%) for calculate the probability to host a gully for each point of study area. Actually these results show that the GLM, RF, and MARS models produce
