【作者】 孙盛；

【导师】 田金文；

【作者基本信息】 华中科技大学 ， 控制科学与工程， 2013， 博士

【摘要】 极化合成孔径雷达成像是一种基于微波的遥感技术，它能够提供对观测对象的大范围的二维高空间分辨率图像。随着各种机载、星载合成孔径雷达系统的出现，相关的数据集变得越来越丰富，这使得合成孔径雷达系统数据处理技术的研究进入了一个快速发展的时期。近年来，关于合成孔径雷达测量及数据处理的技术得到了较快发展，其中的部分技术和应用已经较为成熟。但是，该领域的相关技术仍然有待进一步深入研究。合成孔径雷达成像系统由单极化、单频率成像发展到多极化、多频率成像。多极化、多频率的极化合成孔径雷达图像又被称为多通道的雷达图像，使用多通道雷达对地物进行观测，可以获得比单通道雷达更加丰富的信息。本文围绕单基站、单频率全极化合成孔径雷达图像处理中的相干斑滤波、极化目标特征分解参数的误差校正、非监督分类这三个方面的问题进行了研究。首先，对电磁波极化状态的描述方法进行了介绍。对极化合成孔径雷达成像的几何模型和基本原理进行了陈述。对欧洲空间局的开放式处理平台及常见的机载平台数据源进行了介绍。其次，对极化合成孔径雷达图像的滤波原则进行了简单分析，归纳出相干斑滤波的目标不仅仅是降低噪声水平，更重要的是，滤波之后的数据所包含的极化属性不能受到破坏。滤波算法必须保持相干矩阵或者协方差矩阵在滤波之后仍然为半正定的Hermitian矩阵。对非线性各向异性扩散方程的扩散行为和能力进行了分析，然后对多维非线性各向异性扩散方程进行了扩展，使之能够适用于相干斑滤波。采用极化合成孔径雷达图像的总功率分量来构建扩散率函数，使用边缘增强方案来设计扩散张量，最后给出了基于非线性各向异性扩散的相干斑滤波算法。实验结果表明，本文提出的算法能够降低噪声水平，同时也可以实现滤波之后的协方差矩阵保持原有极化属性。之后，在极化目标特征分解的特征参数计算过程中，研究了由空域平均代替统计平均计算极化目标的二阶统计量所产生的误差。对极化目标特征分解参数的估计量的统计特性进行了分析。利用开放式处理平台生成了若干仿真数据，仿真数据是一种相对理想化的数据。对仿真数据实施渐进式的统计分析，得到了平均阿尔法角和熵参数的统计特性。通过对这两个参数的统计特性的分析，可以得知平均阿尔法角对于采样窗口尺寸的影响几乎可以忽略不计，而熵参数对于采样窗口的尺寸比较敏感。在对仿真数据中的多个随机点进行平均处理后，得出熵与视数具有近似线性变化的关系。通过这种线性变化关系，设计了熵参数的校正算法。实验结果表明，经过校正后的熵参数对于基于特征分解的非监督分类精度的提高有较大帮助。最后，极化合成孔径雷达图像的分类是该领域中非常重要的应用，基于监督分类的核心思想将基于散射模型的分解与监督分类结合起来。将基于四分量散射模型分解的结果作为非监督分类的初始分类数据，然后应用威沙特分类器进行分类，并利用离差平方和方法对分类的结果进行聚类。实验结果表明，与传统的三分量散射模型非监督分类相比，新的非监督分类方法的精确度更高。与基于特征分解的非监督分类相比，新的方法能够避免特征参数上下界无法确定的问题。更多还原

【Abstract】 Polarimetric synthetic aperture radar imaging is a technique based on microwaveremote sensing. Polarimetric synthetic aperture radar can provide large-scaletwo-dimensional high spatial resolution images of the observed objects. In addition, withthe wide availability of PolSAR data from both space borne and airborne SAR systems,the research on processing of PolSAR data is developing fastly. For the past recent years,significant advances in polarimetric synthetic aperture radar instruments and informationextraction techniques have flourished, and related research and applications have reacheda certain degree of maturity. However, part of the techniques in this field deserves athorough investigation. Synthetic aperture radar systems come along from singlepolarization and monochromatic imaging to multi-polarization and multi-frequencyimaging. The later is also called the multi-channel radar imagery. Using multi-channelradar system for observing the earth can acquire much more information than singlechannel radar system. In this thesis, three problems, including speckle filtering, biascorrection for eigen-decomposition parameters and unsupervised classification, in theprocessing of monostatic and monochromatic PolSAR data will be studied.At first, the description method for polarization state of electromagnetic wave ispresented. Then the geometrical configuration and principles of imaging system inpolarimetric synthetic aperture radar is put forward. The open-source platform by theEuropean space agency and mainstream airborne datasets are listed.Thereafter a simple analysis is put forward at first about the principles in specklefiltering. One of the most important clauses is preserving the polarimetric properties andreducing the speckle level in the data. The coherency matrix and covariance matrix afterspeckle filtering must be semi-definite positive. A study on the behavior and performanceof the nonlinear anisotropic diffusion equation is accomplished. Whereafter an extensionto the multi-dimensional nonlinear anisotropic diffusion is done. The total scattered poweris employed to construct the diffusivity function and the edge-enhancing diffusion schemeis adopted to design the diffusion tensor. At last the algorithm of speckle filtering usingnonlinear anisotropic diffusion is listed. It can be easily observed from the experimentalresults that the new proposed method can preserve the polarimetric properties and reducethe speckle level.Whereafter in the calculation of eigen-parameters for polarimetric target decomposition, the ensemble averaging is substituted by the spatial averaging. Thus theestimation for coherency matrix or covariance matrix will contain bias, which will bestudied in this thesis. Consequently the statistics of these parameters are analyzed at first.The simulated data generated by the PolSARpro platform is used for the reason that it hasdodged the influence of the artifact and system noise. The simulated data can be viewedas a realized data. An asymptotically statistical analysis is performed on simulated dataand thereafter the statistical behavior of averaged alpha angle and entropy are acquired.Based on the statistical behavior, it can be concluded that the influence caused by thewindow size on averaged alpha angle can be ignored. Nevertheless the entropy issensitive to the windows size. After performing an averaging of many random points insimulated data, it can be seen that approximate linear relation between entropy andwindow size is valid. Bias correction algorithm is constructed based on the aboveapproximate linear relation. It can be concluded from the experimental results that theentropy which has been corrected will enhance the accuracy of unsupervisedclassification using eigen-decomposition.At last, polarimetric classification is one of the most important applications inPolSAR imaging. The kernel idea of supervised classification is employed in this thesis.The scattering model based decomposition is combined with the supervised classification.The four-component scattering model based decomposition is used as the initial input ofclassification. The Wishart classifier is used iteratively. It is shown from the experimentalresults that the new unsupervised classification method possesses higher accuracycompared with the classic three-component method. Compare with the classic methodbased on eigen-decomposition, the new method stay away from the problem of uncertainupper-bound and lower-bound.更多还原