Research On Unsupervised Classification Of Polarimetric SAR Image Based On Target Decomposition

Synthetic Aperture Radar(SAR),which is an advanced microwave imaging system,can actively emit electromagnetic waves and receive scattered echoes from targets to detect targets with its advantages of all-weather,all-time.The polarimetric SAR(Pol SAR)which can obtain relatively complete polarimetric scattering information through four different channels(HH,HV,VH and VV)is referred to as a fully Pol SAR,and theory and practice show the differences in polarization signatures help distinguish different types of ground objects.Pol SAR image classification uses the polarization informations of the target to classify.It is important for exploring the target's polarization scattering characteristics and improving the interpretation of polarized SAR images.Because it is difficult to obtain prior knowledge such as training samples and large-scale real features,this thesis focuses on the unsupervised classification method of transforming SAR images based on target decomposition.The main research contents are as follows:(1)An improved hybrid Freeman/Eigenvalue decomposition method is studied.This thesis proposes two improvements for the original hybrid Freeman/Eigenvalue decomposition method in order to solve the problems of negative power value.The first is to remove the azimuth and phase randomness of the target in accordance with the rotation invariance of the matrix under the dominant scattering mechanism.The second is to select the corresponding volume scattering model under different dominant scattering mechanisms.When the scattering mechanism is dominant,a rotated dihedral angle scattering model suitable for the artificial area is selected;when the secondary scattering mechanism is dominant,a generalized volume scattering model suitable for the natural area is selected.The effectiveness and practicability of the method is verified by experiments on several sets of Pol SAR data.(2)An unsupervised classification method of Pol SAR images based on target decomposition and class adaptive recognition is studied.Firstly,the polarization scattering parameters are extracted through improved hybrid Freeman/Eigenvalue decomposition,which is applied to achieve the initial classification.Then the image is classified into more classes under the co-polarization ratio(? and ? parameters),followed by the Visual Assessment of Cluster Tendency and Dark Block Extraction algorithms determine the number of the class.At last,the complex Wishart iteration is performed on the classification results to further improve the classification accuracy.The experimental results are analyzed using polarimetric Pol SAR data of real ground targets,which verifies the effectiveness of the classification performance of the method in this thesis.(3)An unsupervised classification method of Pol SAR images based on target decomposition and large-scale spectral clustering is studied in chapter 5 of this thesis.Aiming at the shortcomings of high complexity and poor performance of pixel-level classification algorithms,we propose a large-scale spectral clustering algorithm based on region-level "superpixels".Firstly,the polarization scattering parameters are extracted through improved hybrid Freeman/Eigenvalue decomposition,which gives rise to more accurate feature vector estimate.Then this involves a beginning sub-step of superpixels generation via the Simple Linear Iterative Clustering(SLIC)algorithm.Finally,a sub-step of representative points selection bipartite graph fromation,followed by the spectral clustering algorithm to complete the classification task.Compared with several existing Pol SAR n SAR classification methods,this method has better classification accuracy.