【作者】 王海江；

【导师】 皮亦鸣；

【作者基本信息】 电子科技大学 ， 信号与信息处理， 2008， 博士

【摘要】 合成孔径雷达(SAR)图像分类是SAR图像处理中的重要内容，也是SAR图像解译的关键技术之一，它是将解译系统中的前端部分单独提取出来作为具体应用的一个典型实例。快速、准确的SAR图像分类是实现各种实际应用的前提。国外(如美国、加拿大、德国、日本等)在SAR图像分类方面处于领先水平，我国的SAR信息处理技术基本上还停留在SAR图像解译的初级阶段，在SAR图像分类算法研究上落后于国外。但不管是国内还是国外，SAR图像分类的准确率、精细度以及分类算法的效率和稳定性都需要改善，SAR图像分类方法的研究仍然是一个热点和难点。本文则重点开展了极化SAR(PolSAR)图像分类方法的研究工作。极化SAR是用来测量目标散射信号极化特征的新型成像雷达，它具有能够获得多通道极化图像的优越性，有利于确定和理解散射机理，提高目标检测、辨别和分类能力，有利于抑制杂波，提高抗干扰能力。极化SAR的出现，扩大了SAR系统的应用范围，在采集地表或地面覆盖物的物理和电磁结构信息的应用中起着越来越重要的作用。全极化SAR数据不仅提供了各个极化通道的幅度和相位信息，还提供了各个极化通道间的相对信息。因此，极化SAR图像包含了更加丰富的地物信息。利用这些信息，可以带来更好的SAR图像处理效果。而现有极化SAR图像分类方法对极化信息发掘还不够，充分发掘和利用这些极化信息，可以更好的确定和理解目标散射机理，从而带来更好的SAR图像分类结果。另外，数据处理手段不断发展，近些年来出现了一些新的数据处理技术，如独立分量分析(ICA)，将这些技术用于极化SAR图像分类，也可能会改善SAR图像分类效果。本文在充分利用极化信息，引入独立分量分析技术的同时，还利用子孔径分析这一新的SAR图像处理技术来对目标极化散射行为进行统计分析，并把分析结果用于分类，改善了分类效果。本文的主要工作和贡献如下：(1)介绍了雷达极化的基本理论，包括极化表示以及极化SAR成像中的散射坐标系和极化散射描述。并对Jones矢量复标量表示下的一些极化问题作了推导。(2)对现有的一些极化SAR图像分类方法进行了研究和实验，如Pauli分解、Krogager分解、H／α分类、H／A／α分类、Wishart分类、以及模糊c均值分类等。对给出垂直水平线极化基下散射功率矩阵的条件下进行Krogager分解的问题作了推导。(3)对独立分量分析的基本理论、模型、约束条件和优化算法进行了介绍。(4)研究了基于独立分量分析的极化SAR图像分类方法，提出了结合ICA相干斑抑制的极化SAR图像分类和直接基于ICA的SAR图像分类。在结合ICA相干斑抑制的极化SAR图像分类中，提出了在Pauli分解后的颜色通道上进行相干斑抑制的方法，克服了在原极化通道上进行相干斑抑制造成的相对相位丢失问题，改善了分类效果。在直接基于ICA的极化SAR图像分类中，使用ICA同时分离各个有用独立分量和相干斑噪声，使相干斑抑制与图像分类成为一个统一的过程，并提高了分类准确度。(5)研究了极化SAR图像子孔径分析。首先研究了基于子孔径分析的非平稳目标检测，提出用不同的方法检测两类非平稳目标；接着将非平稳检测结果与H／α分类结合起来，提出了平稳性-H／α分类方法，增加了分类精细度；然后研究了基于子孔径数据进行非监督Wishart分类的方法，提出了子孔径-H／α／Wishart分类方法，改善了非监督Wishart分类的效果。(6)对全文进行了总结，对有待进一步研究的问题作了展望。更多还原

【Abstract】 SAR image classification is an essential integral part of SAR image processing and also a key technique for SAR image interpretation. Fast and accurate SAR image classification is a critical pre-processing step in various practical applications. Nowadays countries like USA, Canada, Germany and Japan are playing the leading role in SAR image classification, while China, owing to its late start, is still in the elementary phase of development in this field. But for all of them, leading or following, the performance of the SAR image classification need be improved and the efficiency and robustness of the classification algorithms be enhanced. By now, the field of SAR image classification methods is still a hotspot for intensive and original research.. This dissertation is devoted to the exploration and development of new methods of classification based on Polarimetric SAR (PolSAR) images.Polarimetric synthetic aperture radar is a new type of imagery radar which is used to measure the polarization feature of the scattering signal from the targets. It has the advantage of getting SAR images of multiple polarization channels and it is helpful in understanding and defining scattering mechanism, and improving radar performance in target detection, discrimination and classification. Furthermore, it is also superior in clutter suppression and anti-jamming to those which do not acquire polarimetric information. PolSAR expands the application area of the SAR system and becomes more and more important in applications to collecting information of physical and electromagnetic architecture of superstratum on the ground.Full polarization SAR data give not only the amplitude and phase information of each polarization channel, but also the relative phase information between each two channels. So, PolSAR data provide more target information on the ground than ordinary SARs do. With the additional polarimetric information, better SAR image processing results can be obtained. The existing PolSAR image classification methods have not dug the polarimetric information in depth. In fact, Extracting and utilizing the polarimetric information adequately would enable us to define and interpret scattering mechanisms better, and then better classification results can be obtained. Furthermore, data processing techniques are in fast progress and some new ones such as independent component analysis (ICA) are emerging continuously in recent years. The classification results may be improved further if these techniques are integrated into the applications of SAR image classification.In addition to utilizing the polarimetric information adequately and introducing ICA into PolSAR image classification, this dissertation applies subaperture analysis, which is a new SAR image processing technique, to statistical analysis of polarimetric scattering. Making use of the analysis results in PolSAR image classification, the classification effects are improved.The main work and contributions accomplished in this dissertation are as follows:(1) The fundamental theory of electromagnetic wave polarimetry in radar, including the polarization representation of EM wave, the scattering coordinate system in polarimetric SAR imaging and the polarimetric scattering description, is presented. And some polarization problems under Jones vector’s complex scalar representation are derived.(2) The existing classification algorithms for PolSAR images, such as Pauli decomposition, Krogager docomposition, H/αclassification, H/A/αclassification, Wishart classification and fuzzy c -means classification are analytically and experimentally studied. The Krogager decomposition under scattering power matrix in HV linear polarization basis is derived.(3) The basic theory, model, constraints and optimized algorithms of ICA are introduced.(4) ICA based PolSAR image classification methods are studied in this dissertation. The classification method associated with ICA based speckle reduction and the classification method directly using ICA are presented. In classification associated with ICA based speckle reduction, the speckle suppression process is carried out on the color channels after decomposition rather than the polarization channels. This solves the problem of relative phase missing and improves the classification results. In classification based on ICA, the desirable independent components and the speckle are separated simultaneously, so speckle reduction and classification become a unified process. And by using this method, the classification accuracy is enhanced.(5) The subaperture analysis of PolSAR images is studied. Firstly the nonstationary target detection based on subaperture analysis is investigated and two different methods are proposed to detect different kinds of nonstationary targets. Secondly the nonstationary target detection results are combined with H/αclassification and the stationarity- H/αclassification method is presented. With this method, the classification fineness is enhanced. Then the unsupervised Wishart classification algorithm based on subaperture data is studied and the subaperture-H/α/Wishart classification method is proposed. This method improves the classification effects of unsupervised Wishart classification.(6) Finally, several conclusions of the dissertation are drawn and the research subjects in the future are proposed.更多还原