We assessed presence/absence prediction of plant species and communities in a southern Mongolian mountain range from geospatial data using a randomized sampling approach. One hundred randomized vegetation samples (3 × 3 m) were collected within the 2 × 2 km summit region of the Dund Saykhan range, which forms part of the core zone of the Gobi Gurvan Saykhan National Park in arid southern Mongolia. Using logistic regression, habitat preference models for all abundant species (n = 52) and communities (n = 5) were constructed; predictors were derived from Landsat 5 imagery and a digital elevation model. Nagelkerkes r 2 was used for an initial data mining, and all significant models were validated by splitting the data and using one half for accuracy assessment based on the AUC (Area Under the receiver operating characteristic Curve)-values. Significant models could be built for half of the species. Altitude proved to be the most important predictor followed by variables derived from Landsat data. The clear altitudinal distribution patterns most definitely reflect precipitation; overall biodiversity in this arid environment is widely controlled by moisture availability. The chosen approach may prove valuable for applied studies wherever spatial data on species distributions are required for conservation efforts.