Statistical Modeling And Classification Of SAR Image Based On The Alpha-stable Distribution

SAR is a full-time, all-weather active microwave imaging system with some penetrating ability. It is irreplaceable especially to the area which the traditional optical sensor is difficult to image. SAR has been broadly applied in various field of military and civilian because of its capability of providing unique information about the earth. Under the advancement of technology, SAR has been developed from single frequency, single polarimetric and single mode at the beginning to multiple frequencies, full polarimetrc and various mode nowadays. With the rapid growth of available SAR data, the development of SAR image understanding techniques was desired. SAR image is generally subject to Speckle phenomenon which disturbs the interpretation of SAR image. Because of the existence of Speckle, information contained in SAR image is mainly depended on the statistical property of grayscale instead of grayscale itself. The traditional digital image processing and analysis techniques based on grayscale that work successfully on nature images often do not perform as well on SAR image. SAR image classification is one of the basic research directions in SAR image understanding and has been the hot topic and difficult problem in remote sensing field recently.Statistical modeling is critical in SAR image classification. It is the primary problem in traditional Bayesian method of pattern recognition. Recently, the research on statistical modeling of SAR data has also been one hot direction in SAR image processing. This dissertation focuses on the research of statistical modeling and classification of SAR image based on its statistical property.SAR image always presents a type of statistical property of heavy tail and impulsive peak, usual statistical distributions (such as Gamma, Weibull, Log-Nornal, K, G0, etc.) can't model well all types of SAR image. Alpha-stable distributin is a family of heavy-tailed with slow discrepancy, whch agrees with the particular statistical property of SAR image. Concerning this kind of statistical property, this dissertation introduces the Alpha-stable distribution for SAR image statistical modeling after studying the methods of parameter estimaton and probability density estimation. The fitting performances on TerraSAR-X samples and the comparison with other statistical models validate the effectiveness. Finally, a Bayesian classification method is proposed by combining Alpha-stable distribution with MRF which introduce the prior information by building a labeling restraint between the current pixel and its neighborhood. The classification performances on TerraSAR-X image of both Wuhan and Foshan are satisfying.With the improvement of resolution, the increasing complexity and randomness of SAR image scenes, multimodal statistical property is more and more prominent in SAR image histogram. In this case, any unimodal distribution is impossible to model well. The mixture of Alpha-stable distributions is proposed to model the multimodal statistical property of high resolution SAR image. Experimental results show that the modeling performances on various types of SAR samples are improved when the number of components of the mixture model is increased. Concerning the difficulty of maximum likelihood estimation and moment based estimation over Alpha-stable distribution due to lack of an analytical expression for its probability density function, a novel estimator called PSA MCMC is proposed for mixture of Alpha-stable distributions based on MCMC. In PSA MCMC estimator, a simulated annealing strategy is introduced which aims to overcome the shortcoming of MCMC that is easily trapped in local optimal for long period of time especially when the target distribution is multimodal. PSA MCMC also updates parameters of the candidate distribution, which makes the candidate samples close to the true posterior distribution. Experimental results on simulated data show that estimated parameters by PSA MCMC are close to the true ones. At last, a Bayesian classification method is proposed by combining the mixture of Alpha-stable distribution and MRF model. Classification accuracies of the proposed classifier on TerraSAR-X image of both Wuhan and Foshan are higher than those of other compared classifiers.Usual used multichannel statistical models for SAR image mainly include complex Gaussian, complex Wishart, Kp, Gp0. This dissertation studies the multichannel SAR image statistical modeling method based on elliptical Alpha-stale distribution, as well as the method of combining Copula function and the Alpha-stable distribution. The fitting performances on RadarSAT-2data illustrate their effectiveness. Finally, a Bayesian classification algorithm is proposed by combining these two methods and MRF. The classification performances on RadarSAT-2data are satisfying, and the classification accuracies are higher than those of complex Gaussian and complex Wishart distribution.The research results of this dissertation are applied in the project of National Natural Science Foundation of China.