Feature Extraction And Classification Of PolSAR Image

Polarimetric synthetic aperture radar (PolSAR) alternately transmits and receivesradar signals in different polarization ways, and can obtain abundant information abouttargets scattering. Up to now, PolSAR has been being one of the most important toolsfor earth observation. Successful applications of PolSAR depend on imageinterpretation technology (IIT). IIT can reveal the nature of PolSAR image, and is thebasis of automatic target recognition (ATR) for PolSAR system. However PolSAR IIThas not met the requirement of military and civil affairs yet, since the start of PolSARIIT is late and its study is not mature. For PolSAR IIT to be well developed, we commitourselves to studying its two key issues: feature extraction and classification.The body of this dissertation consists of two modules: feature extraction andclassification of PolSAR image. First, we study feature extraction of PolSAR imageusing polarimetric target decomposition (PTD) and non-negative eigenvalue theory(NNET). Second, based on polarization features and statistical distributions of PolSARdata, we study PolSAR image classification using kernel function, support vectormachines (SVM) and triplet Markov field (TMF). The main content of this dissertationconsists of five parts as follows:In the first part, we construct a knowledge system of NNET which is based on VanZyl’s work. Van Zyl’s work concludes two points:1) he proves that covariance matrices,which can represent targets scattering mechanisms, are positive semi-definite;2) heproposes the model of non-negative eigenvalue decomposition (NNED), and gives thesolution to NNED in reflection symmetry case of PolSAR data. However Van Zyl’swork needs to be matured in theory. Van Zyl’s work and the complementary work in thisdissertation are called by a joint name: NNET. The complementary work consists ofthree aspects.1) The term that polarimetric matrices should satisfy positivesemi-definiteness is called non-negative eigenvalue constraint (NNEC); we analyze therelationship between NNEC and NNED, and introduce the function of NNEC on PTD.2) We interpret the physical meaning of the NNED model; propose and prove severalimportant properties of NNED.3) We propose a fast solution to NNED in non-reflectionsymmetry case of PolSAR data. NNET enriches the essential theory of PolSAR, and cansupport the related techniques of PolSAR information processing. In addition, we applyNNET to determine the threshold of subspace decomposition filter. Real PolSAR data experiments demonstrate that the proposed filter can enhance the speckles suppressionand retain the information of edges and point targets very well.In the second part, NNED backward strategy (NNED-BS) of NNET is proposed,and the failure of Freeman decomposition (FD) is overcome. This part contributes tofeature extraction of PolSAR image in three aspects:1) we prove that the validity of FDis equivalent to that the remainder of FD satisfies NNEC. And we analyze that theexisting methods, NNED forward strategy (NNED-FS) and its derivation strategies,cannot ensure that the remainder satisfies NNEC.2) NNED-BS based on NNET isproposed. And we analyze that NNED-BS can ensure that the remainder satisfies NNEC,which demonstrates that NNED-BS can deal with the failure of FD.3) We applyNNED-BS to FD and its improved methods, and propose the feature extraction methodcombining FD and NNET. Real PolSAR data experiments show that compared with theexisting FDs, the proposed feature extraction method can greatly enhancedouble-bounce scattering powers and suppress the over-estimation of volume scatteringpowers.In the third part, we propose NNED hierarchy backward strategy (NNED-HBS) ofNNET, and use NNED-BS and NNED-HBS to overcome the failure of Yamaguchidecomposition (YD), respectively. This part contributes to feature extraction of PolSARimage in three aspects:1) we analyze and prove that YD has the failure, and point outthat the failure can be dealt with if the remainder satisfies NNEC.2) Based on NNET,we propose the improved NNED-BS, viz. NNED-HBS. It can overcome the failureproblem of YD and further reduce the power of the remainder.3) NNED-BS andNNED-HBS are applied to YD and its improved methods, respectively. And then thefeature extraction method combining YD and NNET is proposed. Real PolSAR dataexperiments show that compared with the existing YDs, the proposed feature extractionmethod can greatly enhance double-bounce scattering powers and suppress theover-estimation of volume scattering powers, and in addition the remainder power ofNNED-HBS is much lower than that of NNED-BS.The fourth part consists of two aspects. First, we use SVM to evaluate the effect ofscattering powers on PolSAR image classification. These scattering powers areextracted by NNED-BS or NNED-HBS. Real PolSAR data experiments demonstratethat these scattering powers are helpful to improve the accuracy of PolSAR imageclassification. Second, we use SVM and weight composite kernel (WCK) to propose aPolSAR image classification method. The proposed method can improve theperformance of PolSAR image classification based on SVM. The key of constructing WCK is to determine the weight coefficients, and these weight coefficients aredetermined by the distances between training samples in feature space. Real PolSARdata experiments show that WCK can efficiently fuse features, and enhance theperformance of PolSAR image classification based on SVM.In the fifth part, we apply TMF and its improved model to PolSAR imageclassification, and improve the performance of PolSAR image classification based onMRF. This part contributes to PolSAR image classification in three aspects:1) weanalyze the nonstationary statistical property of PolSAR image, and point out that TMFis suitable for dealing with PolSAR image classification.2) We propose the model andalgorithm of TMF for PolSAR image classification. Real PolSAR data experimentsdemonstrate that TMF is superior to MRF for PolSAR image classification.3) Weanalyze the limitations of the auxiliary field of TMF. For the limitations to be overcome,we define a smoothness feature of PolSAR image, and add this feature into the TMFmodel to propose a Wishart TMF with a specific auxiliary field (TMF-SAF). In addition,we analyze that TMF-SAF can overcome the limitations in principle. Real PolSAR dataexperiments demonstrate that TMF-SAF is superior to both MRF and TMF for PolSARimage classification.