Comparison of Classification Performance of Full-,Dual- and Single-Polarization SAR Images Using SVM
YU Wen-xian
DOI: https://doi.org/10.11834/jrs.20080107
2008-01-01
Journal of Remote Sensing
Abstract:Classification is an important process in interpretation of SAR images.In classification,the information,such as amplitude,phase and texture,is used to arrange all pixels in an image into different classes.A classification map shows directly classes of terrains,which is helpful to understand image.Classification methods of SAR images can be divided into supervised and unsupervised.Support vector machine(SVM) based on statistical learning theory,proposed by Vapnik et al.,is an effective supervised classifier.It is used widely in face recognition,hand-writing identification,and automatic target recognition for good classification performance with small training data sets.It has been a new focus in the field of machine learning.Several researchers have tried to use SVM for classifying polarimetric SAR images,and obtained promising results.As an advanced instrument for remote sensing,polarimetric synthetic aperture radar(SAR) has been applied widely in many fields,such as ecology,environmental monitoring,geological exploration,vegetation investigation,and so on.Compared with single-polarization SAR,to what extent dual-polarization and fully polarimetric SARs can improve in classifieation is important.Classification performance of full polarization versus dual and single polarization is compared qualitatively and quantitatively with SVM taken as the classifier in this paper.For fully polarimetric SAR data,six power values,extracted from the covariance matrix,and three eigenvalues,obtained by eigenvalue analysis technique using the coherency matrix,are contained in an input feature vector.For dual-polarization data,there are only three power values and two eigenvalues.And only one power value is used as an input feature for single-polarization data.In order to equilibrate effect of each element in an input feature vector on classification results,all features are normalized.According to the ground truth or a span image,training samples are selected to train SVM to obtain classifier parameters.Lastly,full-,dual-,and single-polarization SAR images are classified by the trained SVM,and the classification accuracy is calculated if the ground truth is available.In the first experiment,an L-band fully polarimetric image of Flevoland,Netherlands,acquired by the NASA/JPL AIRSAR sensor on August 16,1989,is used to analyze quantitatively the classification accuracy of full-,dual-,and singlepolarization SAR data.The results show that the classification accuracy of fully polarimetric SAR is highest,followed by dual-polarization SAR,and it is lowest for single-polarization SAR.For crop application,the accuracy of HH-VV SAR is greater than other two dual-polarization SARs,comparable with fully polarimetric SAR.If fully polarimetric SAR is unavailable,HH-VV SAR is a proper substitute with acceptable performance.Because of stronger depolarization capability,separability of each terrain in HV data is better than that in the other two cases.Consequently,classification accuracy of HV SAR is better than other two single-polarization SARs.For the two co-polarization SARs,performance of HH SAR is better than another.If the co-polarization transmitter and receiver are used,HH is more proper.In the second experiment,an HH-HV dual-polarization image,obtained in China, is used to analyze qualitatively classification performance of dual-and single-polarization SARs.The experimental results show that scattering power of building is badly confused with that of bank and bare soil due to weak depolarization of building.Thus the classification result of HV SAR is worse than HH SAR.Lastly,using detailed results of the above two experiments,classification performance difference of full-,dual-,and single-polarization SARs is explained from the point of view of scattering characteristics of terrains and operational mechanism of the classifier,SVM.