The Research Of SVM On Remote Sensing Image Classification Application

rural- urban fringe is an active area for human economic society,concentration zone construction and development of unban and rural,new "cambia" for city expand,and leader of rural urbanization.With the fast development of urbanization,the land changing in rural- urban fringe is obvious,land use survey is caused more and more attention.Remote sensing technology(RS) is the main method for regional land use survey and ration collection on use information.Thereinto,remote sensing image classification is the first step of remote sensing data of land resource analysis and application,the key to the question of remote sensing image research is how to classify many categories image exactly,it is important,and with good basic for next research.Support vector machine(SVM) is a new machine learning arithmetic,it has good generalization performance,many special advantage in solving small sample,non-linear and high dimension mode recognition.This text aim at character of multispectral data,take Landsat ETM+/TM image of 2000 and 2009 as main information source,make the research on using multispectral image classification, the main working as below:1. Collect spectral and texture information for TM image,select the best band combination as classification feature,the selection is:image in 2000 year,select wave 2,6,8,15,17 and 16;image in 2009 year,select wave3,6,8,10,11 and 12.2. Do land use classification by combining spectral signature and texture information, the result show spectral signature take leading effect.Well extract vegetation and waters after combining vegetation index, waters index and texture information,increase classify precision.3. On sorting technique, use SVM for image classification, compare with other methods, SVM has special advantage and high precision in solving small sample,non-linear and high dimension mode recognition.4. Combined with the advantages and disadvantages of traditional support vector,and do the classification testing and precision comparison on computer.Confirm new type SVM classification method has reality operability on remote sensing image classification. Thereinto, LS-SVM classification that selected by bsyes parameter is best,but the parameter training cost of time is too large;C-SVM classification is satisfied, and take least time,have best reality operability;GEPSVM has advantage on mixed pixels classification,but is sensitive on unbalancedness of 2 kinds data.5. Make pilot study on special questions,such as multiclass classification, training samples selection,model parameter selection,kernel function selection etc. for each SVM used in multispectral remote sensing image classification, get below conclusion:(a) kernel function selection has large effect on SVM classification accuracy,the selected RBF kernel function in this training has good adaptability and can apply to various kinds classification,but stability of accuracy is not high, generalization ability is low.(b)RBF kernel function whether can reach the best classification speed and accuracy,is main decided by setting the parameter.This article seek the best parameter for fitting to research area by using grid-search method.