Classification Of Polarimetric SAR Imagery Based On Statistical Model TMF

Polarimetric synthetic aperture radar(Pol SAR) is a multi-channel imaging radar system.Compared with single-polarization SAR intensity image, full-polarization SAR image contains more information about the ground truth and full polarimetry information.Therefore, Pol SAR image interpretation has become a researching focus in the field of SAR image processing. Up to today, Pol SAR has been broadly and deeply used in all walks of life. In terms of military, Pol SAR can be used to detect and discriminate military targets. In terms of civil use, Pol SAR can be used to monitor vegetation growth, evaluate natural disaster, research the ocean, monitor glacier cover and so on. Based on the triplet Markov field(TMF), this paper introduces advanced statistical models of Pol SAR data into the likelihood energy term of TMF to study Pol SAR image supervised classification. The main work can be divided into two parts.Based upon the TMF model, Pol SAR image supervised classification by the original Wishart TMF integrates a Gaussian complex Wishart distribution for the Pol SAR data statistics conditioned to each image cluster and a spatial contextual model. For nonstationary images, TMF can obtain more promising classification results than MRF(Markov Random Field). However, it is difficult to interpret the physical meaning of the auxiliary field derived by TMF. To solve this problem, we integrate a normalized smoothness characteristic into the energy of the proposed TMF to supervise the auxiliary field to be classified as smoothness stationary and nonsmoothness stationary. In addition,during constructing the initial labeling field of TMF, the original Wishart TMF merely performs an ML(Maximum Likelihood) classification based on the Gaussian complex Wishart distribution, and abundant polarimetry information isn't fully used. To solve this problem, we extract polarization signatures by polarimetric decomposition, which are combined with a supervised SVM(Support Vector Machine) classifier to construct the initial labeling field of TMF. Experiments on three real Pol SAR data demonstrate the effectiveness of the above two improvements.A supervised classification method for Pol SAR image based on feature extraction by polarimetric decomposition and TMF under the non-Gaussian K-Wishart distribution is proposed. Gaussian complex Wishart distribution is strictly justified only for homogeneous regions of the image, but non-Gaussian K-Wishart distribution is able to account for potential textural differences in the classes, implying that processing algorithms based on non-Gaussian K-Wishart distribution should improve classification performance. We introduce the non-Gaussian K-Wishart distribution into the likelihood energy term of TMF.The shape parameter a in the K-Wishart probability density function(PDF) is estimated with the method of matrix log-cumulants(Mo MLCs) based on complex matrix-variate Mellin kind statistics. During constructing the initial labeling field of TMF, we continue to extract polarization signatures by polarimetric decomposition, combined with a supervised SVM(Support Vector Machine) classifier. Experiments on two real Pol SAR data demonstrate that compared with Gaussian complex Wishart distribution, non-Gaussian K-Wishart distribution is more suitable for modeling non-homogeneous regions of the image and can achieve higher classification accuracy.