Comparing the Best Input Combining Artificial Neural Networks and Decision Tree Method to Identify Factors that Influence the Phenomenon of Dust Storm (Case Study: Yazd Province)

One of the major natural disasters that caused tremendous damages each year in many areas of the country is desert، especially in Yazd، Iran. Strong winds and the formation of dust storms will take place several times each year. In this study، data from meteorological stations in Yazd (thunder storms، wind magnitude (size، amount)، wind duration، visibility، fastest wind speed، average wind speed، prevailing winds and dust storms) in the period 1953-2005 were used on a monthly basis. In order to determine the most appropriate combination of neural network and input parameters (inputs) that influence the phenomenon of dust storms from Variable reduction technique of factor analysis (maximum likelihood، principal component)، principal component analysis، stepwise progressive and gamma test were used. Each of the methods presented، each with a different combination of these compounds neural network feed Forward back propagation with the algorithm of Levenberg-Marquardt have been used. The results showed that the stepwise progressive R² = 0. 87 and RMSE = 0. 04 provides the most suitable combination for a neural network. Comparison of simulated dust storm phenomenon in seasons and months in different years showed that the phenomenon of dust storm in summer and spring seasons and months of April، May، June، July، August and September are different. In comparing the neural network feed forward back propagation models with algorithm of Levenberg-Marquardt and decision tree with algorithm CART، Neural networks model with a correlation coefficient of 0. 87 and the root mean square error of 0. 04، the decision tree method with a correlation coefficient of 0. 86 and the root mean square error of 0. 06 has more carefully in order to simulate dust storm.