Polarimetric SAR Image Classification

Polarimetric synthetic aperture radar (SAR) has become one of the most advanced remote sensors in recent years. As one of the main tasks for understanding polarimetric SAR images, polarimetric SAR image classification has been playing an important role in many fields of both civil and military applications. To improve classification accuracy and reveal target scattering mechanisms, some key techniques concerned with polarimetric SAR image classification are investigated in this dissertation.1) Deep analysis of some fundamental theories of radar polarimetry. To further clear some misunderstandings and inconsistencies of some concepts related to the basic polarimetric equations, an exact derivation of basis transformation of the Sinclair scattering matrix is given using the directional Stokes vector and the time reversal operator, which are introduced by Graves and Luneburg, respectively. The consistency in essence of the Mueller matrix and the Kennaugh matrix is explained in theory by introducing the concepts of positive and opposite propagation spaces. In addition, partially polarized waves are also deep analyzed.2) Statistical modeling of polarimetric SAR data. Based on the multiplicative speckle model, five new statistical distributions ( KP distribution, GP 0 distribution, GP H distribution, GP 1 distribution and GP 2 distribution) for the scattering vector and two new statistical distributions ( GP 1 distribution and GP 2 distribution) for the polarimetric covariance matrix are proposed. The GP 2 distribution is most appropriate to model homogeneous, heterogeneous and extremely heterogeneous clutter. The estimators using moments for the roughness parameters of the GP 1 and GP 2 distributions are given. To obtain more accurate and robust estimation, an optimization method is proposed.3) Supervised statistical classification of polarimetric SAR images. An iterative classification method of polarimetric SAR images (GMMAP method), based on the maximum a posteriori (MAP) criterion, the GP 2 distribution and the Markov random field (MRF), is proposed. The method can achieve least classification error in theory. And because of the introduction of new samples through iterations, the method solves the problem that class statistics are probably not estimated accurately with a limited training sample set.4) Scattering randomness measurement of polarimetric SAR targets. In order to reflect the variation of target scattering randomness with the polarization sates of incident waves, a novel measure of target scattering randomness, which is called Degree of Randomness (DoR), is proposed. The concept of DoR Signature is introduced for the visualized description of the DoR. The mean and standard deviation of the DoR is defined. The mean of the DoR for"horizontal-vertical"linear,"45°-135°"linear and"left-right"circular polarization waves is analyzed. The variation of this parameter is almost the same as that of the scattering entropy and there exists an approximation relationship between them. And the computation of the parameter is simpler and faster than that of the scattering entropy.5) Unsupervised scattering classification of polarimetric SAR images. First, based on the dominant scattering mechanism and the scattering randomness measure, a scattering classification frame is constructed. Then, to solve the problems existing in the H /αclassification, a new method (EKE method), based on the eigen decomposition, Krogager decomposition and scattering entropy, is proposed. To directly extract and discriminate the dominant scattering mechanism in the incoherent case, another new method (FDR method), based on the Freeman decomposition and the mean DoR, is proposed. Finally, to further improve the classification performance, combining the EKE and FDR methods with the Wishart distance measure, two iterative classification methods (EKE-Wishart and FDR-Wishart) are developed.