Research Of Remote Sensing Image Classification Based On Decision Tree Method

Remote sensing image classification become a very importantcontent in the field of remote sensing research, and it is also an importantpart to extract the surface feature classification information by means ofremote sensing. Remote sensing image classification can be divided intotwo kinds of methods and it includes supervised classification andunsupervised classification，which of them have a variety of classificationmethods. Traditional computer automatic classification that is only basedon the remote sensing image pixel spectral characteristics is affected bymany factors and the classification accuracy can not meet the usersdemand, so the decision tree classification method comes into being. Thedecision tree classification method has been used in research of manyclassification problems, but the research results of applying to remotesensing classification are less common. The decision tree classificationmethod has some characteristics of flexible, intuitive, clear, robust, andoperating efficiency that has shown great advantages in solution ofremote sensing classification problems. The fairly mature decision treeclassification algorithm includes CLS, ID3, C4.5, C5.0series; CART,SLIQ, SPRINT and CHAID, etc. The decision tree algorithm is anon-parametric, nonlinear supervised classification method, as much aspossible to ensure the quality and quantity of training data in this study,and then it properly combines the image spectral information and otherauxiliary information, finally it chooses the best classifier to complete theremote sensing image classification.This study regards Landsat TM images of August1,2010and trackNo.122/30as the basic remote sensing information source, choosing theintersection（E118°22′-118°54′，N43°03′-43°17′）of Bairin Right County,Linxi County, Hexigten County and Wengniute County, the center region of Chifeng City, Inner Mongolia，China as the study area, classifying theremote sensing image based on decision tree method. First, it uses themaximum likelihood classification method to classify remote sensingimages of the study area on The basis of the original spectralcharacteristics.Then,repeatedly modified to improve the training data setin accordance with the classification results, as far as possible to ensurethe precision and sufficient quantities of the training data set, then selecttwenty-one characteristic variables combinations that consist of fivedifferent characteristic variables combinations, and they include sixoriginal bands and the NDVI based on them, the first three principalcomponents (PC1, PC2and PC3), eight texture features (Mean, Variance,Homogeneity, Contrast, Dissimilarity, Entropy and the Second Momentand the Correlation) and three topographical features (DEM, Slope, andAspect). It utilizes the typical decision tree algorithm C5.0, CART andthe traditional maximum likelihood method to classify them. Select theclassification results of three methods based on the all of channels data,then consider to overcome the misclassification problem of classificationresults that select the best classification results in each classificationmodel, and execute the accuracy assessment and comparative analysisseparately. Decision tree classification method of C5.0and CART, whichis low complexity and high operating efficiency, and be easy to realizeand the training processes to need a short time. They don’t have therequirement for normal distribution of the characteristic variables.Especially, when the combinations of characteristic variables areappropriate, they able to effectively use the related auxiliaryinformation,then their final classification results are more satisfactoryand their classification accuracy are higher than the maximum likelihoodclassification accuracy.Typical decision tree algorithm C5.0and CARTcan quickly and accurately find the classification rules from the trainingdata, then effectively integrate the spectral signatures and spatialcharacteristics of the remote sensing image to classify. They are an effective means to improve the classification accuracy of remote sensingimages.