Synthetic Aperture Radar Imagery Statistical Modeling And Filtering

Synthetic aperture radar (SAR), which is the most crucial developmentof radar technology from target one-dimension detection (e.g., range, azimuth,and height) to target two-dimension imaging, has the capability of all-whetherand day-and-night operations. SAR, attributable to its merits, graduallybecomes one of the most important equipment of microwave remote sensing,and simultaneously, SAR images have increasingly extensive applications inthe field of remote sensing. The tremendous divergence which exists betweenSAR images and optical images of general sense displays in many aspects,such as the imaging mechanism, noise pattern, and distribution of pixel gray.To transform SAR images from "data" to "information" and then obtain theeventual "knowledge" which meets the demand for remote sensingapplications, understanding the basic property and enhancing the visualreadability of SAR imagery are fundamental tasks.This thesis advances the research of SAR speckle in view of statisticsand then reviews probability distribution models of SAR imagery;furthermore, the estimation of model parameters and the evaluated criterion ofmodeling precision are developed and improved; lastly, the filtering methods,based on the Contourlet transform, are proposed to reduce SAR speckle.In the study of basic properties of SAR imagery, the fundamental speckleof SAR imagery is introduced, including the imaging mechanism, statisticalmodels, and simulated methods. Provided that scattering components of atarget obey the property of random walk, the first-order statistical models ofcomplex data, amplitude, intensity, and phrase of speckle are obtained andthen validated with real SAR images. Furthermore, the formation andstatistical models of multi-look SAR imagery are also considered. Finally, thesimulated methods of speckle noise are developed based on speckle statistics.In the study of statistical modeling of SAR imagery, this thesis reviewsall statistical models, whose parameter estimation and evaluation of modelingprecision are also provided. According to the origin of each model, all statistical models, which are amount of eleven kinds, can be divided into twocategories: the statistical models of prior assumptions and the empiricaldistribution models. In addition, we propose a modified generalized GaussianRayleigh (GGR) model based on the GGR model. Meanwhile, we provideparameter estimates of statistical models under the following three methods:the method of moments (MoM), the maximum likelihood (ML), and themethod of logarithmic cumulants (MoLC). The evaluated criterions ofmodeling precision, including Pearson’s2test, Kolmogorov-Smirnov (K-S)distance, and Kullback-Leibler (K-L) distance, are also developed. Finally, weuse real SAR images to exam statistical models and evaluated criterions, andconclude that the inverse Gaussian, Fisher, and generalized Gammadistributions show good capabilities of statistical modeling of SAR imagery.In the study of speckle filtering of SAR imagery, the filtering methods ofSAR speckle are proposed in the basis of Contourlet transform. TheContourlet transform, which characterizes image multi-scale decomposition,including scale decomposition of Laplacian pyramid (LP) and directionaldecomposition of directional filter bank (DFB), is introduced. Sequentially,the Contourlet Bayes soft-threshold and the hidden Markov tree (HMT)model of Contourlet coefficients, which are under the same philosophy as theWavelet, are raised to reduce SAR speckle. Furthermore, the merits ofproposed methods are demonstrated through simulated speckle noise and realSAR images. The result indicates that the de-speckling image of Contourlettransform reduces obviously "artifacts" and enhances the visual readability incomparison with the de-speckling image of Wavelet.