Research On Statistical Modeling Of SAR Images And Its Application Based On Generalized Gamma Distribution

The study of interpretation algorithms of Synthetic Aperture Radar(SAR)images is a hot advanced topic in the field of SAR information processing in recent decades.The statistical modeling of SAR images is a technology that applies mathematical model to describe the statistical property of SAR images precisely.The research on application algorithms based on the statistical modeling of SAR images is a main way for interpreting SAR images.Recently,with the improvement of SAR imaging technologies,SAR imaging modes have been increasingly enriched,such as the multi-frequency,multi-polarization and interference.Besides,the spatial resolution of SAR images has been improved seriously too.On the one hand,much more information of terrains and targets in such images is carried.On the other hand,however,the statistical properties of such SAR images have become more complicated,making the precisely description of their statistical properties a new challenge.To improve the level of automatically interpreting SAR images,it is an urgent issue to carry out the research on the statistical modeling of such SAR images as well as its application algorithms.In view of the high flexibility of generalized gamma distribution(G?D),the paper focuses on this distribution for modeling SAR images and its applications.We perform theoretical researches on the statistical modeling of both single-channel and polarimetric SAR images.Besides,the studies of some of its hot typical application algorithms are also carried out,including the target detection,segmentation and terrain classification.These contents constitute a relatively complete system of the statistical modeling and application of SAR images.The main work and innovative contributions of this paper on the theoretical researches and application algorithms are as follows.1.Hierarchical researches on the theories of statistical modeling of SAR image are performed,of which the main work and innovative contributions are as follows.(1)Existing researches on the statistical modeling of single-channel and polarimetric SAR images are reviewed systematically,on both constructing models and parameter estimation.Firstly,the work of statistical modeling of single-channel SAR images is stated.Then that of polariemtric SAR(PolSAR)images is summarized.(2)An algorithm of parameter estimator for G?D is proposed by parameter decoupling.In this algorithm,a Scale Independent Shape Estimation(SISE)equation is derived.This method is superior to the existing feasible Method of Log-Cumulants(MoLC)on both stability and effectiveness.(3)A novel statistical model called G?-Wishart distribution for PolSAR images is proposed,of which a parameter estimator based on Matrix Method of Log-Cumulants(MoMLC)is developed too.By introducing the G?D to model the texture component of PolSAR images,the G?-Wishart distribution is derived,which has generalized many classical PolSAR distributions,including the complex Wishart,?p and Gp0 distributions.The experimental results on real PolSAR images show that this distribution is more flexible than the previous classical models,even superior to the recently developed KummerU distribution.(4)A method of simulating spatial correlated single-channel and polarimetric SAR clutter images is developed.Firstly,under the framework of non-linear transform method,the scheme of simulating spatial correlated SAR clutter images following the G?D is theoretically derived.Besides,the methods of simulating the PolSAR images following several typical polarimetric distributions are also proposed.The simulated SAR images are with great benefit to the evaluation of parameter estimation algorithms,the design and evaluation of statistical algorithms and so on.2.Based on our previous work on the statistical modeling of SAR images,the paper then carries out the researches on the SAR image interpretation algorithms,focusing on three typical issues including the constant false alarm rate(CFAR)algorithm,SAR image segmentation technology and terrain classification method.The main work and innovative contributions of these aspects can be summarized as follows.(1)A CFAR algorithm for SAR images based on the G?D is proposed.With the G?D describing the background clutter of SAR images,the expression of the CFAR detection threshold is analytically derived.Compared to the state-of-the-art CFAR algorithms for SAR image,the proposed method is superior for maintaining the false alarm rate and is more efficient.(2)A hierarchical segmentation algorithm of SAR images based on the G?D is developed.By using the G?D to model SAR images,the segmentation algorithm is designed with two steps.Firstly,the SAR image regions are merged by a minimum log-likelihood loss criterion.Then,the region edges are evolved with a local Bayesian criterion.Compared to the general hierarchical segmentation algorithms of SAR images,the proposed method is capable to improve the quality of SAR image segmentation.(3)A terrain classification method of SAR images based on the G?D is proposed.With the G?D as the theoretical model of SAR images,the analytical expression of the Kullback-Leibler(KL)distance between G?D is derived.Then,a method of terrain classification of SAR images is proposed based on a criterion of minimizing the KL distance between regions.The proposed method is superior to the classical pixel-based and region-based Bayesian classifiers with respect to both the classification accuracy and the robustness on biasd statistical modeling.