Study On Terrain Classification Based On MRF Model And Statistical Modeling Of SAR Imagery

As a tool of acquiring information, SAR is of significant strategic sense, the interpretation of SAR imagery is a very important task. Terrain classification plays an important part in the interpretation of SAR imagery. The classification method based on MRF model and statistic modeling, which is able to unify the contextual information from label images and statistical properties of observed images, was widely used and developed in image processing. In this thesis, SAR image classification method based on MRF and statistical modeling is thoroughly studied.Firstly, as the foundation theory, SAR clutter statistical model is studied in depth. According to existing work and the progresses made recently, the relevant techniques of statistical modeling of clutter in SAR image are comprehensively reviewed, and a parameter estimation approach stemming from Mellin transform which is also named"Method of Log-Cumulant"(MoLC) is introduced and analyzed in detail. Based on MoLC, the estimate equations of the statistical distribution models mentioned are induced and comprehensively summarized. To satisfy the requirement in the modeling of high-resolution SAR imagery, a recently proposed method named"Dictionary based stochastic expectation maximization for SAR amplitude probability density function"is studied. Copula, a statistical tool that was designed for constructing joint distributions from marginals with a wide variety of allowable dependence structures, is introduced to model the join distribution of multi-channel space of data.Secondly, the classification of single polarized SAR imagery based on MRF model and statistical modeling is studied. Based on MRF model, the process of image classification is turned to be a problem of energy-minimum by seeking the optimal label Matrix. A classification method based on MRF model and the joint distribution of the data space of gray level and texture information (Contrast Statistics) deduced from Gray Level Co-occurrence Matrix (GLCM) via copulas is proposed.Finally, the classification of multi-polarized SAR imagery based on MRF model and statistical modeling is studied. Statistical modeling approach based on copula theory is studied. After the in depth study of finite mixture copulas model, a dictionary based mixture copulas model is proposed, which has good statistical representation ability. Experiments verified the better performance, compared with the classification process only using single-polarized data.