【作者】 吴永辉；

【导师】 郁文贤； 粟毅；

【作者基本信息】 国防科学技术大学 ， 信息与通信工程， 2007， 博士

【摘要】 极化SAR是一种先进的遥感信息获取手段。与单极化SAR相比,它通过测量每个分辨单元在不同收发极化组合下的散射特性,更完整地记录了目标后向散射信息,为详尽分析目标散射特性提供了良好的数据支持。极化SAR图像分类是图像解译的重要步骤,分类图既可作为中间结果为边缘提取、目标检测、识别等提供辅助信息,也可作为最终结果直接输出给用户。开展极化SAR图像分类研究对于探索目标散射特性、提高极化SAR系统的应用水平具有重要的理论意义和实用价值。论文以揭示目标散射机理和提高分类精度为主要目的,系统地研究了极化SAR图像分类方法。揭示目标散射机理是极化SAR数据分析的重要目的之一,现有这方面研究一般只针对全极化数据。双极化是极化SAR常用的工作模式,为研究双极化SAR对目标散射机理的识别性能,论文对H-α方法进行了修正。推导了双极化H-α平面有效区域的边界,从理论角度研究了双极化SAR对各向同性表面、偶极子和各向同性二面角三种基本散射的识别能力,通过实验分析了双极化SAR对H-α平面有效区域内八种基本散射机理的识别性能,并给出了HH-VV双极化H-α平面的一种可行划分方式。现有极化SAR图像分类算法大多以像素为基本分类单元。为提高分类精度,论文从改进已有算法和引入发展于其它领域的算法两个角度,对基于像素的分类方法进行了研究。H-α方法是最为著名的极化SAR图像非监督分类方法之一,但其分类图存在地物类别模糊问题。针对该问题,通过结合C-均值算法,提出H-α-CM算法,从而将H-α方法的散射机理分类转化为地物分类。为自动确定迭代次数,定义了分类图像熵,以熵最大作为H-α-CM算法的迭代终止准则,并通过实验分析了该准则的合理性。SVM是一种较新的分类和回归算法,论文研究了它在极化SAR图像分类中的应用。为避免依据经验选择特征导致分类性能不稳定,提出以支持向量个数作为评估指标的NSVFS特征选择算法,并将其用作SVM分类的预处理,从而构成完整的利用SVM进行特征选择和分类的NSSVM算法。实验结果表明该算法对SVM参数的敏感性较低,具有较强的自适应性。最后,以SVM为分类器,定性、定量地比较了全极化、双极化和单极化SAR的分类性能,并从目标散射特性和分类器工作原理的角度阐明了三者性能差异的成因。基于像素的方法能保持分类图的地物细节,但其性能易受相干斑影响。基于区域的方法在分类过程中考虑像素空间相关性,可有效减弱相干斑的不良影响。MRF是一种常用的描述像素空间相关性的模型,在将MRF引入极化SAR图像分类的过程中,为充分利用数据的统计先验知识和避免拆分协方差矩阵导致信息损失,将MRF与协方差矩阵的Wishart分布结合,提出WMICM算法。随后,针对watershed算法过分割时存在锯齿效应的问题,采用MRF进行过分割,提出MOS-ML算法。WMICM和MOS-ML在分类的不同阶段考虑像素的空间相关性,前者由初始分类和ICM调整两个步骤构成,在第二步中引入相邻像素相关性:后者包括仞始过分割和ML分类,在构造过分割得到的大量子区域的过程中利用像素相关性。由于利用了极化数据的统计先验知识和相邻像素的空间相关性,因此两种算法均具有较高的分类精度,并能获得清晰平滑的分类图。更多还原

【Abstract】 Polarimetric synthetic aperture radar (SAR) is an advanced instrument for remote sensing. It obtains scattering characteristics of each resolution cell under different combinations of receiving and transmitting polarization, and records back scattering information of targets more completely than single-polarization SAR. It is helpful for analyzing target scattering characteristics. Classification of polarimetric SAR images is an important procedure of image interpretation. The classification map can be used as the middle result for edge extraction, target detection and recognition, etc., and also as the final result output directly to users. Investigation of classification of polarimetric SAR images is of much theoretical and applicable significance in the exploitation of target scattering characteristics, as well as the improvement of the application efficiency of polarimetric SAR systems.To reveal target scattering mechanisms and improving classification accuracy, classification methods of polarimetric SAR images are investigated in this thesis.It is very important to reveal target scattering mechanisms in analysis of polarimetric SAR data, nevertheless it is done generally for fully polarimetric data. Dual polarization is a frequently used operational mode of polarimetric SAR systems. In order to investigate performance of scattering mechanism identification of dual-polarization SARs, H-αdecomposition is modified. The boundary of the feasible region in H-αplane for dual-polarization cases is derived. Performance of dual-polarization SARs to distinguish three basic scattering mechanisms from an isotropic surface, a dipole, and an isotropic dihedral is studied theoretically. Performance of dual-polarization SARs to identify the eight scattering mechanisms inside the feasible region in H-αplane is analyzed, and a feasible division of H-αplane for HH-VV dual-polarization SAR is obtained in an experimental manner.Individual pixels are taken as elements in most classification methods for polarimetric SAR images so far. In order to improve classification accuracy, pixel-based classification methods are investigated by improving an existing method and introducing an algorithm developed in other fields. H-αdecomposition is one of the most famous unsupervised methods for classifying polarimetric SAR images. However, terrain classes are confused in the classification map of H-αdecomposition. Therefore the H-α-CM is proposed by integrating the C-mean algorithm, thus classification of scattering mechanisms is transformed into terrain classification. In this algorithm, to determine the number of iteration automatically, the entropy of a classification map is defined, and maximizing the entropy is taken as the termination criterion, which is demonstrated to be reasonable by experimental results. SVM is a new algorithm for classification and regression. It is used in classification of polarimetric SAR images herein. To avoid instability of classification performance induced by selecting features using experience, the NSVFS is proposed for feature selection, in which the number of support vectors is taken as the estimation index. Then it is used as a preprocess step in SVM classification, thus the NSSVM is constructed to select features and classify images using SVM. It is demonstrated by experiments that this algorithm is not very sensitive to SVM parameters, and has better self-adaptability. Finally, classification performance of full polarization versus dual and single polarization is compared qualitatively and quantitatively using SVM. Causes resulting in performance difference are illuminated by target scattering characteristics and operational mechanism of the classifier.Terrain details can be preserved in classification maps obtained by pixel-based methods. But performance of these methods is affected by the speckle. It is different for region-based methods. The spatial relation of neighboring pixels is considered in these methods, thus speckle effect is reduced effectively. MRF is a frequently used model for describing the spatial correlation between adjacent pixels. While introducing MRF into classification of polarimetric SAR images, in order to use completely the statistical a priori knowledge of the data and avoid information loss induced by separating the covariance matrix, the WMICM is proposed by integrating MRF and Wishart distribution of the covariance matrix. Then, aiming at the sawtooth effect of the watershed algorithm in over-segmentation, the MOS-ML is proposed using MRF for over-segmentation. The spatial relation of neighboring pixels is considered in different phase of the two algorithms. It is introduced in the second step of the WMICM, which contains initial classification and ICM adjustment, while introduced in the first step of the MOS-ML, containing initial over-segmentation and ML classification. Clear and smooth classification maps and high accuracy are observed using the two algorithms, due to full consideration of the statistical characteristic of polarimetric data and the spatial relation between adjacent pixels.更多还原