Comparative Analysis For Extracting Gobi Types From Images By Applying Different Remote Sensing Classification Methods

Gobi is one of the main landscape types of earth’s surface in the arid region of Northwestof China. At present, it was difficult to do fieldwork due to the execrable natural conditions andthe sparse dweller in Gobi region, which led to the research documents on Gobi were scarce,but the Gobi holds abundant solar power, wind energy and various mineral resources.Exploring the Gobi background is very important for western development and westernecological environment construction in China. With the progress of Remote Sensingtechnology and the upgrade of various image processing software, more and more effective andreliable information can be extracted from Remote Sensing image.The studied region of this paper is located in Ejina County in Inner Mongolia, China. TheRemote Sensing image processing software ENVI4.8was used to extract the Gobi informationby using different classification methods which analyzed the image features from the pixelsand object-oriented perspective based on Landsat TM/ETM image data. The purpose of thispaper is to put forward the suitable Remote Sensing image interpretation method for Gobiinformation extraction based on Landsat TM/ETM images, and provide support for recognitionGobi information from remote sensing images.Six different image classification methods including Minimum Distance, MahalanobisDistance, Maximum Likelihood, Support Vector Machine(containing four different kernelfunctions), Decision Tree(containing four different situations), Object-oriented were adopted toextract Gobi information in this paper and the results were as follows.(1)Comparing the several image classification results, the automatic thresholdclassification of Decision tree was the best result with an overall accuracy at93.3569%; whileKappa coefficient was0.9052; and the mapping accuracy and user accuracy of cumulated Gobiwere93.98%and89.11%respectively; the mapping accuracy and user accuracy of erosionGobi attained84.34%and99.32%respectively. However, this method of decision tree was not suitable for large amount of data. The maximum likelihood method could deal with largeamount of data with a high efficiency and accuracy. In this study, the overall accuracy ofmaximum likelihood method was91.6273%, while the Kappa coefficient is0.8786. Themapping accuracy and user accuracy of stacked Gobi were94.35%and85.37%, and themapping accuracy and user accuracy of erosion Gobi achieved85.39%and98.85%.(2) These methods, except the Minimum Distance method was not suitable for recognitionGobi information, had their own advantages and disadvantages for extraction of the Gobiinformation. Decision Tree method performed well, but not suitable for the large amount ofdata and the calculation speed was too low. Support Vector Machine could be used to processless samples, but it was difficult to decide the kernel function, and it had high requirements onthe amount of data with a slow speed. Maximum likelihood method had a good performanceand high efficiency among the several experimental methods. The Object-oriented approachcould completely extract the surface features with a certain geometry and obvious boundary,but it was not good for extracting Gobi from the TM/ETM images.According to the analysis of the study, it would be a good performance by usingMaximum Likelihood method for large amount of data, while it could obtain a good result byusing automatic threshold patch extraction Gobi information for small amount of data.