【作者】 张斌；

【导师】 秦前清； 张志；

【作者基本信息】 武汉大学 ， 信号与信息处理， 2013， 博士

【摘要】 遥感作为一种非接触式的观测技术是获取空间信息的一个非常重要的手段,具有多波段、多时相、重访周期短、信息量丰富、获取效率高等优势;合成孔径雷达(Synthetic Aperture Radar, SAR)是一种微波遥感成像系统,具有全天时、全天候、具有一定的穿透性等特点。可以全天侯、全天时地获取高分辨率、包含极化信息的地面遥感图像数据,同时对光学遥感影像具有较好的互补性。分类是多极化合成孔径雷达(polarimetric synthetic aperture radar, POLSAR)图像解译的一个重要内容。相较于传统SAR影像,POLSAR影像包含了更为丰富的信息。全极化合成孔径雷达不仅可以得到反映目标电磁散射特性的回波幅度、相位、频率等物理量,并能通过电磁波水平或是垂直发射并通过水平或是垂直接收(HH, HV, VH, VV)得到四种不同的极化状态,从而可以得到关于目标的材料或几何形状信息,因此利用全极化SAR (Synthetic Aperture Radar, SAR)图像在进行目标检测和场景分类时有着传统SAR图像无法比拟的优势。POLSAR影像分类算法从某种程度来说,可以分为基于物理散射机制和基于概理统计分类方法,现今主要将这两者结合起来,根据物理散射机制对地物进行初始分类然后由概理统计方法进行迭代聚类。本文结合极化目标分解与马尔科夫随机(Markov Random Field, MRF)方法,利用影像上下文信息,对MRF进行改进,试图达到抑制相干班噪声影响提高分类精度。同时针对地震地质灾害发生后,利用灾后超高分辨率(very high resolution, VHR)SAR影像,并结合灾前高分辨率光学影像,利用SAR模拟仿真技术结合MRF方法,进行灾害变化检测研究。具体来说,本文工作可以概括为：1)介绍了SAR影像的成像特点,MRF模型基本理论,POLSAR影像分割分类算法研究现状,以及高分辨率SAR与光学影像变化检测基本方法。2)基于像素级,提出了将极化目标分解与MRF相结合,最后通过层次聚类算法进行迭代聚类的方法；提出了一个基于分层MRF的POLSAR影像分类方法：将极化目标分解技术与MRF结合起来,先根据物理散射机制对地物进行初始划分,然后由MRF方法进行迭代聚类。将分层MRF方法应用到POLSAR影像分类中,通过先分层,再逐层映射的方法得到最后的分类结果。3)基于区域级,提出了一个基于区域MRF的POLSAR影像分类方法：利用均值漂移算法先进行分割为同质区域,并通过区域增长方法进行区域合并,最后利用分割区域之间的相关性由MRF进行分类得到最后结果。4)针对地质灾害发生后,高分辨率光学图像难以获取或是获取的影像质量较差,利用灾后获取超高分辨率SAR影像,并结合灾前高分辨率光学影像,通过SAR的仿真模拟构建SAR与光学影像变化检测的桥梁,提出了一种基于MRF变化检测的新方法。更多还原

【Abstract】 Remote sensing observation techniques as a non-contact is a very important means to obtain spatial information with a multi-band, multi-temporal, a short revisit cycle, rich information. Synthetic aperture radar (Synthetic Aperture Radar, SAR) is a microwave remote sensing imaging systems, which has the all-day, all-weather, has a certain penetration characteristics.Classification is very important for polarimetric synthetic aperture radar(POLSAR) interpretation. Compared with traditional synthetic aperture radar (SAR), fully POLSAR can acquire more information because it obtains the complete polarization state of electromagnetic waves. Target properties, such as material and geometry, affect the polarization state of the electromagnetic wave; thus, POLSAR has significant advantages in target detection, identification, and classification. POLSAR classification algorithm could be classified to two categrary: based on physical backscattering and statistic method. POLSAR classification algorithm can be divided into two categories, known as the classification method based on physical backscattering, and algorithms based on statistical properties. The current popular classification algorithm is the combination of physical scattering characteristics and statistical properties. Target decomposition combined with Markov Random Filed(MRF) is researched in this paper to reduce the impact caused by speckle noise on classification results. Using very high resolution SAR images after earthquake and very high resolution optical images before earthquake, SAR simulation technique is adopted to change detection.The contribution of this paper could be summaried as:1) SAR imaging system, Markov Random Field, POLSAR classification method, Change detection are introduced.2) Methods are proposed by combining target decomposition and MRF on pixel level, which use agglomerative hierarchical clustering algorithm to decrease class number. A new classification algorithm is proposed for POLSAR images using multi-scale Markov Random Field on pixel level. Mean-Shift algorithm is applied to get the initial classification for the coarsest scale, and then Markov Random Field is introduced to achieve the classification result. Classification result on a coarser scale is employed as the initial classification of the nearest finer scale. Meanwhile, the Wishart distribution is employed to model the observed field, and then the iterative conditional modes algorithm is adopted to implement the maximum a posteriori estimation of pixel labels for each scale.3) An unsupervised classification algorithm is proposed for polarimetric SAR images based on Mean Shift Segmentation and MRF. First, polarimetric features are exacted by target decomposition for Mean Shift segmentation. An initial classification is executed by using target decomposition and agglomerative hierarchical clustering algorithm. This is followed by a classification step based on MRF using the mean coherence matrices obtained for each segment. Under the MRF framework, the smoothness term is defined according to the distance between neighboring areas.4) A damage assessment method is proposed by using pre-event very-high resolution(VHR) optical and post-event synthetic aperture radar(SAR) images to detect buildings damaged in an earthquake. First, by using the ENVI software, the length, width, height and other3-D parameters of a rectangle building are extracted by pre-event VHR optical image. Second, a image-based GPU ray-tracing approach is used to simulate SAR images based on the exacted3-D parameters. Finally, the similarity between the simulated SAR images and post-event actual SAR images is analyzed to determine whether the building is damaged.更多还原