عنوان مقاله [English]
This study was carried out in order to evaluate the ability of artificial neural network in the preparation of predictive distribution map of plant species habitat and recognizing the strengths and weaknesses of this method. For this purpose, vegetation and soil sampling was conducted after determination of the homogenous unit using digital elevation model and geological map with a scale of 1: 25,000. Then environmental variable maps were obtained using GIS and Geostatistic. Input variables were selected based on the results of logistic regression. In order to develop of the artificial neural network model the best structure of the neural network was determined using Mean Square Error (MSE) after data normalizing and randomly data portioning into train, test and validation sets. Then simulation of presence / absences probability species was done with optimal network and continuous probability maps of presence or absence species in each habitats were prepared using Arc GIS software and presence optimal threshold was determined. Then compliance of the predictive and actual maps was assessed through the calculation of the kappa coefficient. The results showed which prediction maps of Pteropyrum olivieri-Stipa barbata habitat has excellent correspondence with actual vegetation map (kappa= 90%) and Amygdalus scoparia, Artemisia aucheri-Astragalus glaucacanthus, Scariola orientalis- Stipa barbata habitats have very good correspondence with the ground truth maps. These results indicate that the use of logistic regression preprocessing simplify the network architecture and has increased learning speed and accuracy of the simulation results. Therefore in case of performing appropriate preprocessing of the data and selection of the appropriate input variables, neural networks can be an appropriate approach to estimation of the distribution of plant species habitats and selection of suitable species for rehabilitation activities in the rangelands.