Evaluation of artificial neural network algorithm in desertification assessment (Case Study: South of Neishabour township)

Desertification، after climate changes and deficiency of freshwater، is the third important global challenge during 21th century. Investigating and assessment of the effective factors on desertification trend are very important issues for better land management. The study area is located at south of Neishabour township. This region has been faced to increasing rate of desertification because of drought، vegetation removal، change of rangelands to dry farming lands، water and wind erosion and lack of proper land management over previous years. In this research، data of field study، analysis of aerial photo and satellite image were collected and analyzed in geographic information system. Then FAO-UNEP desertification method (1984) was used in this study. According to this method، vegetation condition، rangeland condition، water and wind erosion and salinity were defined and categorized as the factors defining the status of desertification. Finally، desertification map was prepared. Based on the result، desertification in this region is severe in the northern part of the region mostly due to reduction of canopy، vegetation removal and water erosion. Meanwhile، 57% and 30% of the studied region could be categorized in moderate and slight desertification conditions، respectively. To compare the results and effects of input parameters and their weights; artificial neural networks have been used. After data collection، classification and categorization of and input parameters in four levels including slight، moderate، severe and very severe، desertification map of FAO-UNEP method was used in GIS based on the artificial neural networks approach for predicting desertification condition. Final optimized perception model with one input layer composed of 4 neurons (vegetation condition، water and wind erosion and salinity) and one hidden layer including 39 neurons; and one output layer (desertification condition) based on momentum training algorithm for hidden and output layers was accepted to reach the Mean Square Error (MSE) of 0. 07975. Comparison of the outputs of artificial neural network with FAO-UNEP method، shows artificial neural network capability to explain desertification condition. It was also revealed that، definition of structure of networks، training algorithm and cycles in addition to input neurons may effectively influence on the performance of various structures of artificial neural networks.