【作者】 呙维；

【导师】 龚健雅； 廖明生；

【作者基本信息】 武汉大学 ， 摄影测量与遥感， 2010， 博士

【摘要】 近几年来,随着传感器技术的不断发展,高分辨雷达遥感卫星不断成功发射,标志着原来仅可用于军事用途的高分辨率雷达数据成为一种可方便获取的新数据来源,这不仅扩大了遥感应用的范围,也增加了对地观测的尺度,并提高了数据更新的速度。研究者可通过解译高分辨率SAR遥感影像,实现在较小的空间尺度和短时间基线上观测城市地表的细节变化,用于城市生态环境评价、灾害应急与快速响应、城市建设与规划、地籍调查等方面的研究。目前常用的雷达图像解译方法主要是通过提取像素级特征来实现地物分类或提取,通过本文分析,这种像素级影像分析方法对高分辨率SAR图像效果较差,为了得到理想的结果,需要引入更多的对象级特征。由于密集城区高层建筑较多,多次折射现象导致SAR成像机制非常复杂,故本文重点针对非密集城区,通过应用面向对象的影像分析方法对高分辨率SAR图像进行特征分析和解译。重点围绕对象级上下文特征表达、适合雷达成像特点的多尺度分割算法和解译方法进行了研究。主要研究成果及创新有如下几点：(1)针对高分辨率SAR图像上的叠掩具有高亮,易检测的特点,本文提出了两种分别针对大型建筑物和密集居民区的对象级上下文特征,分别是叠掩邻接强度和亮点散布密度。通过散点图对这两种特征的空间聚集性分析,说明两种特征能作为高分辨率SAR图像中的建筑区域提取的主要参考特征。(2)我们提出了具有可并行、多尺度和参数自适应等特点,并对雷达噪声和弱边缘较好适应能力的eFNEA分割算法。具体而言,本文在分析光学多尺度图像分割(FNEA)算法用于雷达图像易产生边缘漂移问题的基础上,提出将分割过程分解成两个步骤,并通过在小尺度下融入边缘拓扑约束来减少边缘漂移现象出现和提高分割精度；针对图像分割占用内存资源多,无法处理具有海量数据特点的遥感图像的问题,本文根据eFNEA算法特点,设计了基于多级网格的并行任务分解与合并策略；由于eFNEA全局固定权重参数对城区和郊区无法同时适应的问题,本文提出了广义异质性规则,通过利用eFNEA多尺度特点来进行颜色和形状权重的自适应调整。(3)本文针对城市地区的我国机载、E-SAR机载和TerraSAR-X星载三种SAR图像特点,综合利用eFNEA和基于叠掩特点的对象级特征,采用三种不同的面向对象解译方法与流程,实现了对以建筑物为主的地物提取与分类。实验结果说明面向对象的解译方法能得到大约86%左右的精度,明显优于像素上下文特征方法,具有应用潜力。(4)最后,本文针对SAR遥感图像处理的计算模式进行了研究,探讨了云计算环境中的遥感影像处理系统构成和实现方法,以一个具体的原型系统(OpenRS-Cloud)展现了遥感云计算模式的用户界面、技术手段与运行流程,并且将本文提出的并行eFNEA算法在OpenRS-Cloud平台上进行了实现与试验,结果证明云计算是一种能有效提升算法运行和开发效率的新型计算环境。更多还原

【Abstract】 In recent years, with the development of sensor technology, high-resolution (HR) radar remote sensing satellites were successfully launched continuously, marking the high-resolution radar data became a easily obtained datasource which only can be used for military purposes originally. This datasource not only expands the scope of remote sensing application, and also increases the observation scale and the speed of data updating. Researchers can study urban surface details by interpreting HR SAR image on more refined spatial scale and shorter time baseline for changes observed for the urban ecological environment assessment, disaster rapid response, urban planning and cadastral survey. At present, common interpreting method for radar image is based on feature classification and recognition through extracting feature characteristic pixel by pixel. According to our study, pixel-level image analysis method is less effective for high-resolution SAR images. In order to get the more desired results, object-level features need be considered.Because there are many high-height buildings at dense urban areas, multi-refraction cause SAR imaging mechanism is very complicated, so this paper studis the object-oriented image analysis (OBIA) method for HR SAR images on non-dense urban areas, focusing on object-level context features, multi-scale segmentation algorithm and interpretating methods while taking the sar imaging characteristics into account. Main research content and innovations are the followings:(1) Benefiting from the highlight layerover can be easily detected on HR SAR images, this paper proposes two kinds of object-level context features named LAI (Layerover Adjacent Intensity) and SDD (Shining point Distribute Density) respectively for large buildings and dense residential areas. The scatter diagram analysis shows that these two features bring better spatial clustering ability than pixel-level features (such as color and texturue). Therefore, we regard these two object-level context features as the main features for urban auto-interpreting in HR SAR images.(2) We propose a segmentation algorithm named eFNEA (Edge restricted Fractal Net Evolution Approach) which not only can adapt to the Speckle noise and weak edge on HR SAR images, and also has parallel, multi-scale and adaptive-parameters characteristics. Firstly, we analyze edge-shift problem which exposed when the famous optical multi-scale image segmentation algorithm (FNEA) processes radar image, and then, propose splitting the original segmentation into two steps to improve the affection and accuracy by merging the edge information to constraint the region topology while small scale level region merge. On the other hand, aiming at the massive memory resource demand of HR SAR image processing and restricted capability of processing huge image data with single compute, we analysis the eFNEA characteristics and propose a parallel task decomposition and merging strategy based on multi-level grid. In term of parameters setting, we propose the generalized heterogeneity rule, and use multi-scale features to adapt weight of color and shape, because eFNEA global weight parameters can not adapt to urban and suburban issues at the same time.(3) Focusing on the there characteristics of China airborne image, E-SAR airborne image and TerraSAR-X satellite images, this paper propose three different object-oriented interpretation methods to extract and classify buildings based on eFENA segmentation algorithm and object-level features. Experiment result confirms the applicable potential of object-oriented method which can get 86% around accuracy, and is significantly better than the pixel analysis method in these cases.(4) Finally, this paper discusses SAR remote sensing image processing mode on the emerging cloud computing environment and proposes a specific prototype system (OpenRS-Cloud) to show GUI, computing mode, technical method and running flow. We realize the above eFNEA algorithm on OpenRS-Cloud platform, and the experiment result shows that cloud computing as the new computing environment is an effective method to improve the efficiency of algorithm processing and algorithm development in remote sensing area.更多还原