The aim of this study was to evaluate the efficiency of artificial neural networks in supplying predictive maps of plant species habitats distribution in Qum province rangelands, Iran. For this purpose, soil and vegetation sampling was done after determination of homogenous unit by combining of slope, aspect and elevation maps and environmental variables maps were prepared using geostatistics and GIS. To prepare the artificial neural network models, the best network structure, was determined following required data preprocessing (normalizing data and partitioning of data into three sets, training, test and validation), improving adjustable parameters (such as transfer function, learning rule, the number of hidden layers, number of hidden layer neurons) and using statistical parameters calculated in the test phase (MSE). After selecting the optimal network, simulations were performed to estimate the probability of the presence or absence of the species and continuous probability maps of the presence or absence was prepared at each species habitat using Arc GIS. Then the optimal threshold was determined using equal sensitivity and specificity method and the compliance between predicted and actual maps were examined by calculating kappa. Based on the results, the most accurate prediction models were obtained for all habitats using sigmoid transfer function and the Levenberg Marquardt algorithm. The results also showed that habitat predictive and actual maps of Artemisia sieberi, Halocnemum strobilaceum, Tamarix passerinoides, Seidlitzia rosmarinus and Artemisia sieberi have excellent, very good, good, fair and poor compliance, respectively. These results indicate that the multilayer perceptron has acceptable accuracy in the modeling and estimation of the geographical distribution of the studied species habitat and if the network input variable be chosen properly, the network can simulate presence or absence of plant species with high accuracy.