A Research Of High-resolution PolSAR Image Classification Based On Multiple Features Fusion

PolSAR image classification is the basic problem of PolSAR processing and interpretation.It is also the key step for PolSAR to be actually applied.The researchers pay much attention on this technique for the reason above.The technique tendency of PolSAR is the system's resolution enhancement,which brings about more abundant information along with new challenges.In the case of high-resolution PolSAR,the spatial structure obtains more information,the surface relief is fiercer,the statistical property varies,the data volume increases and the detection scenes become more complicated.These changes lead to the ineffectiveness of lower resolution PolSAR in actual application.This paper studies the image classification algorithms applied to high-resolution PolSAR on the basis of existing image processing methods.The main contents of this paper are as follows:1.We studied the polarimetric features construction method which includes integrated polarimetric information.The surface relief of high-resolution PolSAR is fierce,which makes traditional single polarimetric signature incapable of describing the target characteristic.The polarimetric characteristic extracted from polarimetric algebraic operation or polarimetric target decomposition interprets the PolSAR data in different perspectives.If we collect those polarimetric characteristics and use it to jointly build high dimension polarimetric feature space,then more comprehensive echo scattering information can be obtained.2.According to the advantage of more obvious texture structure information,we brought spatial feature into high-resolution PolSAR classification.To effectively use the spatial feature of PolSAR,we selected three texture feature,namely,Gabor,MP and AP.Additionally,before extracting the texture feature we applied SPAN image as preprocessing procedure in order to restrain the inherent speckle of PolSAR image.The effectiveness of PolSAR image classification using spatial feature is verified by experiments.3.We apply LPDA method which combines the principle of discriminant analysis and manifold learning to high-resolution PolSAR image classification,solving the problem of the increasing data quantity.This method is capable of preserving the within-class manifold structure and maximizing the interclass separability through optimizing the eigenvector structure under a certain ratio criterion.Experimental results show that this dimensional reduction method is able to reduce redundancy and maintain classification accuracy of the original high dimensional features at the same time.4.Facing the enlargement of PolSAR image scene complexity,this paper studied an classification algorithm which takes into accounts both polarimetric information and spatial information.We applied information fusion techniques to make that combination doable.Two fusion strategies are considered,which performs at the feature level and the decision level respectively.In the first strategy,at each pixel two feature vectors are firstly combined,either by feature stacking or composite kernel constructing,and then the classification is performed with one single SVM.In the second strategy,the PolSAR image is firstly classified with polarimetric and spatial features separately,and then the classification results are combined to give the final classification map.Experimental results verify the benefits of using both of polarimetric and spatial information for the classification of high-resolution PolSAR images.