【作者】 张腊梅；

【导师】 张晔； 邹斌；

【作者基本信息】 哈尔滨工业大学 ， 信息与通信工程， 2010， 博士

【摘要】 极化合成孔径雷达(PolSAR)可以利用不同极化通道的SAR复图像区分物体的细致结构、目标指向、几何形状以及物质组成等参数,在遥感领域具有广阔的应用前景。利用极化信息提取技术对SAR图像中的典型目标进行特征提取和检测是PolSAR图像解译和应用的热点课题,具有重要的理论意义和实用价值。论文立足于SAR极化信息的提取,以建筑物目标检测为目的,重点围绕极化目标分解、PolSAR图像分类和PolSAR目标检测以及PolInSAR目标检测等内容进行系统深入的研究。首先,本文对目标的极化特性和目标分解方法进行深入研究,包括相干目标分解、基于特征值的非相干目标分解和基于散射模型的非相干目标分解方法。在深入研究已有极化目标分解方法和其应用范围的基础上,针对建筑物的特殊结构和特有散射特性,提出基于多成分散射模型(MCSM)的极化目标分解方法,综合考虑了奇次散射、偶次散射、体散射、螺旋散射和线散射五种基本散射机理。利用E-SAR和EMISAR的PolSAR数据进行实验,验证了基于MCSM目标分解方法的有效性,分解得到的各散射成分将作为主要特征用于后续的PolSAR图像分类和PolSAR目标检测。其次,在PolSAR图像分类研究中,利用支持向量机(SVM)在小样本情况下良好的学习能力和结构风险最小化的特性,在MCSM目标分解的基础上,本文提出联合MCSM和SVM的PolSAR图像分类方法。将MCSM提取的目标散射特征与纹理特征相结合,考虑目标的自身散射特性及其空间纹理,运用SVM分类器进行PolSAR图像分类。基于该方法对EMISAR数据进行了分类实验和性能评估,并且与基于Freeman分解和SVM的分类实验比较,结果表明联合MCSM和SVM的分类方法能够获得良好的分类效果和较高分类精度。针对PolSAR图像目标检测,本文提出基于商空间粒度计算的PolSAR图像目标检测算法,将基于MCSM的目标分解结果、极化相似性参数和极化白化滤波结果作为粗粒度空间分别进行目标检测,再利用商空间粒度合成将三个检测结果进行加权融合得到细粒度空间,获得最优的检测结果。该方法可综合三种方法的优点,充分考虑目标的散射特性、与典型目标的相似性以及对比度,并将其优化组合实现目标的高精度检测。利用EMISAR数据分别进行了MCSM目标分解、极化相似性参数、极化白化滤波以及基于商空间粒度合成的目标检测实验。对比各种方法的检测结果表明,基于商空间粒度合成的目标检测方法能够获得较好的检测效果。将基于商空间粒度合成的检测结果和人工标定的建筑物进行匹配,结果表明基于商空间粒度合成的检测方法可有效用于PolSAR的目标检测。最后,由于建筑物在SAR图像中大多为分布式目标,本文将极化相似性参数的定义范围进行拓展,提出基于Stokes矩阵的极化相似性参数。又由于建筑物在PolInSAR图像中具有较高的相干性,提出极化干涉的广义特征值相似性参数,并与极化干涉相干矩阵特征值联合用于PolInSAR的目标检测。利用E-SAR的PolInSAR图像进行检测实验,检测结果证明了该方法的有效性。更多还原

【Abstract】 Polarimetric Synthetic Aperture Radar (PolSAR) identifies the fine configuration, orientation, geometric shape and composition of target using the SAR complex images in different polarimetric channels, and PolSAR represents wide applications in remote sensing. Feature extraction and target detection in SAR images using polarimetric information extraction technology are hot issues of PolSAR image interpretation and application with much theoretical and applicable significance. Based on the extraction of polarimetric information in SAR images, in order to improve the capability of image analysis and building target detection in PolSAR images, the polarimetric target decomposition, PolSAR image classification, target detection using PolSAR and PolInSAR images are studied systematically and detailedly in this dissertation.Firstly, the polarimetric characteristics of target and polarimetric target decomposition are deeply studied, including the coherence target decomposition, the incoherence decomposition based on eigenvalues and the incoherence target decomposition based on scattering model. Based on the theories and applications of existing decomposition methods, an extended Multiple-Component Scattering Model (MCSM) is proposed for PolSAR image decomposition, which considers single-bounce, double-bounce, volume, helix and wire scattering as elementary scattering mechanisms in the analysis of PolSAR images. The proposed MCSM is demonstrated with L-band full polarized images of DLR E-SAR of Oberpfaffenhofen test site in Germany and Danish EMISAR of Foulum test site in Denmark. The results validate that MCSM is effective for analysis of buildings in urban areas. Furthermore, the decomposition results can be used for further PolSAR classification and target detection.Secondly, PolSAR image classification is researched. Support Vector Machines (SVM) have good learning ability in case of small samples and structure risk minimization. The classification of polarimetric SAR image based on MCSM and SVM is presented in this paper. In order to take the scattering characteristics of itself and the spatial distribution into consideration, the decomposition results of MCSM and the texture features are combined in the SVM classifier. The validation experiment and performance evaluation are implemented using EMISAR PolSAR images. Compared with the classification result using Freeman and SVM, it can be found that the proposed classification method based on MCSM and SVM can obtain a good classification result and high precision. Subsequently, target detection based on granularity computing of quotient space theory using PolSAR images is proposed in this dissertation. The detection results of MCSM decomposition, polarimetric white filter, and polarimetric similarity parameter are considered as coarse granularity spaces. Then these three coarse granularity spaces are combined to construct the fine granularity space by using granularity synthesis algorithm based on quotient space theory. The fine granularity space is namely the optimal detection result. This method comprehensively utilizes the scattering characteristics, contrast and the similarity with typical target, and optimally combines the advantages of these three methods to realize high-precision detection. The target detection experiments based on MCSM, polarimetric similarity parameter, polarimetric white filter and their combination are also implemented using EMISAR data. The detection results demonstrate that the detection method based on granularity computing is better than a single detection algorithm. Compared the results based on granularity computing with the manual marks of buildings, it is found that the proposed detection method based on granularity computing is an effective target detection method in PolSAR images.Generally, buildings are distributed target in SAR image, and thus polarimetric similarity parameter based on Stokes matrix is presented, which is an extension of polarimetric similarity parameter based on Scattering matrix. Because targets present high coherence in PolInSAR image, polarimetric interferometric generalized eigenvalue similarity parameter is proposed. The target detection based on eigenvalues of PolInSAR coherence matrix and generalized eigenvalue similarity parameter is applied to E-SAR L band PolInSAR data and the results verify its effectiveness.更多还原