【作者】 陈一祥；

【导师】 秦昆；

【作者基本信息】 武汉大学 ， 地图制图学与地理信息工程， 2013， 博士

【摘要】 近十年来高分辨率对地观测技术得到了迅猛发展,已成为当前世界高科技发展的前沿领域。高分辨率商业卫星的空间分辨率已达到米级或亚米级的水平,能够获取更详细的地面信息,如何实现高分影像解译与地物信息提取已成为当前重要的研究课题。空间结构特征是高分影像最显著的特征之一,高分影像上的各种地物类型都表现出不同程度的结构特性,这些结构既包括纹理的、几何的也包括空间关系的。有效利用高分影像的空间结构信息对于影像的解译与地物信息的提取将具有重要的作用。本文以高分影像的空间结构特征建模与信息提取为主题,系统研究了基于空间统计和数据场的高分影像空间结构特征描述、建模与提取方法,并将获取的空间结构特征与结构模型用于改善高分影像的分类或实现居民区的检测。本文的主要研究工作如下：(1)对高分影像空间结构特征的概念、内涵和特点进行了探讨。借鉴系统科学对结构的认识,本文从影像的组成“元素”以及这些元素之间的“空间关系”来认识影像的空间结构。这里的“元素”是结构分析的基本单元,它既与影像自身的空间分辨率有关,也与人们对事物的认知模式密切相关。考虑到空间结构的多尺度特性,本文提出了高分影像空间结构特征的一个多层次分析框架,该框架将高分影像的空间结构划分为像元结构、目标结构和场景结构三个不同的层次。然后在该框架下,重点了研究高分影像在前两个层次的空间结构特征,并分别研究了基于空间自相关统计、空间半变差函数和数据场的空间结构特征建模方法。(2)研究了基于空间自相关统计的高分影像空间结构描述方法。首先分析了空间自相关统计量对高分影像中不同地物空间依赖的响应。实验发现：Moran’sⅠ和Getis统计量对影像中的同质性区域响应较好,并且后者还能区分是高亮同质区还是低暗同质区；Geary’s C统计量对影像中的异质性区域或边缘区域响应较好。进一步,根据影像中不同区域空间关联强度的差异,使用局部空间统计量和3D阈值分割法实现了高分影像中同质区和边缘结构的提取。接下来,本文将六个空间自相关统计量(包括三个全局统计量和三个局部空间统计量)用于高分影像局部空间结构的描述,并将提取的空间结构特征与光谱特征融合用于高分影像的分类,通过比较分类结果对各统计量的局部空间结构的描述能力进行了比较分析。实验发现：Moran’s I和Getis统计量在改善影像中同质性地物的分类效果较好,同时Getis统计量还具有一定的平滑滤波的作用；Geary’s C统计量能够显著改善影像中的异质性较大的地物类(如建筑区和植被)的分类结果；在相同的邻域窗口条件下,全局统计量能够比局部统计量获取更多的空间结构信息,但是其计算量会显著增大。最后,针对传统的空间自相关统计量描述高分影像局部空间结构信息的不足,本文提出了基于方向空间关联性的特征描述方法。该方法利用了方向射线,能够同时考虑空间关联的强度信息和长度信息,与通常的基于窗口的描述方法相比,它不仅能够提取高分影像局部的空间变异或表面纹理信息,还能够利用地物类的形状和结构信息,同时它还能够减少计算代价。与传统的基于窗口的Moran’s I、Geary’s C、灰度共生矩阵和小波变换方法进行比较,本文提出的方法都表现出了较优越的性能。(3)研究了基于空间半变差函数的高分影像空间结构特征建模与居民区提取方法。本文将半变差函数的形态分为“平衡型”、“不平衡型”和“周期型”三种类型,并且给出了描述半变差函数曲线形态的多个参数,这些参数可用于描述高分影像的空间结构特征。进一步,本文提出了一种基于半变差的高分影像居民区检测方法。该方法利用半变差函数来建模居民区的空间结构特征,实验发现,高分影像中居民区的半变差函数符合具有一定周期的孔穴效应模型。基于该模型本文提取了表征居民区结构的特征参数,并利用该特征通过对高分影像进行基于分块的分类实现居民区的检测。(4)研究了基于数据场的高分影像空间结构建模与居民区提取方法。首先比较了数据场方法与空间统计学中局部空间统计量之间的关系,本文获得了以下结论：通过选择合适的空间权重,局部空间统计量可以统一到数据场势函数体系,或者说局部空间统计量可以作为数据场描述数据对象相互作用的一种势函数。然后借鉴局部空间统计量描述空间关联性的方法,将数据场方法用于高分影像局部空间结构特征的提取,提取的新特征改善了高分影像的分类。最后,将数据场方法用于居民区空间结构的描述与建模,提出了一种基于数据场的高分影像居民区检测方法。该方法将居民区中的建筑物或其部分特征(如特征点、特征线)看成是质点,居民区内部的空间结构通过场来进行描述和建模；进一步,根据居民区与非居民区在势值上的差异,通过阈值分割的方式实现居民区的分割,并通过“噪声”去除、孔洞填充等方法完善了居民区的提取结果。更多还原

【Abstract】 High-resolution earth observation technology has obtained rapid development over the past decade, and at present, it has become one of the frontiers of high-tech development in the world. With a spatial resolution of meter or sub-meter level, high-resolution images can provide more detailed ground information and how to achieve the image interpretation and information extraction has become an important research topic. Spatial structural feature is one of the most significant spatia features of high resolution images, in which most of the object types exhibit different degrees of structural characteristics such as textural structure, geometrc structure and spatial relationships. It is well known that effectively utilizing spatial structural information can play an important role in impoving the image classfication result and increasing the object detection accuracy. In this paper, we studied the methods of spatial structural feature description, modeling and extraction based on spatial statistics and data field. The structural features or structural model obtained by the proposed methods were then applied to improving the image classification or achieving the detection of residential area class.The main work of this paper are as follows:(1) Discussed the concept, connotation and characteristics of spatial structural features in high resolution images. Drawing on the theory of system science which includes the cognition and understanding of structure, we recognized the spatial structure of an image from two aspects:one is the composed "element" of the image, and the other is the "spatial structural relationship" among these elements. Here the "element" is the basic unit of structure analysis, which is closely related to the spatial resolution of the image itself as well as human cognitive pattern of things. Taking into account the multi-scale characteristics of spatial structure, this paper proposed a multi-level analysis framework of spatial structure, which divided the spatial structure of high resolution images into three different levels including pixel structure, object structure and scene structure. In this framework, this paper focused on the first two levels of spatial structure, and studied the methods of spatial structural feature modeling based on spatial autocorrelation statistics, semi-variance function and data field, respectively. (2)Studied the methods of spatial structural feature description based on spatial autocorrelation statistics. First, we tested the response characteristics of spatial autocorrelation statistics to spatial dependance of different object classes in high resolution images. It was found that, Moran’s I and Getis statistics can make a good response to homogeneous regions in an image, and the latter can distinguish between homogeneous regions of high value and low value; Geary’s C statistic can make a strong response to heterogeneous regions or edges in the image. Further, according to intensity differences of spatial correlation in different regions in the image, this paper proposed a novel method of extracting homogeneous regions and edge structure from images using local spatial statistics. Second, the well-known six spatial autocorrelation statistics (including three global statistics and three local spatial statistics) were applied to the local spatial structural feature description and extraction of high resolution image, and the extracted spatial features were integrated with the spectral features for image classification. By comparing the classification results of using each statistic, the performance of each statistic was evaluated. Experimental results indicate that Moran’s I and Getis statistics can better improve the classification of homogeneous object classes, while Geary’s C statistic can significantly improve the classification of heterogeneous object classes (such as built-up areas and vegetation). It was also found that within the same neighborhood window, global statistics can get more spatial structural information but also need more computation cost than local statistics. Finally, to overcome some shortcomings of spatial autocorrelation statistics in the local spatial structure description, this paper proposed a novel method of modeling spatial structural features using directional spatial correlation(DSC). This method used eight directional half-lines instead of a window to measure spatial correlation in neighborhoods of pixels, which can take into account both the intensity and the length of spatial correlation. Compared with the usual window-ased methods, the proposed method can not only extract the information of spatial variability or surface texture but also can use the geometric information about shape and structure of object classes. In particular, it can also reduce the computation cost. By comparison, it was found the proposed method outperform some existing algorithms including Moran’s I, Geary’s C, GLCM, and showed competitive performance against wavelet transform based methods in accuracy or computation time.(3) Studied the methods of spatial structural feature modeling and residential area extraction based on based on semi-variance function. This paper first divided the common semi-variograms into three different types:"balanced","unbalanced" and "periodic", and then studied the spatial structural feature extraction method based on parameters characterizing semi-variogram curve shape. Then, a novel method of residential areas detection based on semi-variance function was proposed. This method used semi-variance function to modeling the spatial structure of residential areas, and it was found by experiments that the semi-variogram of residential areas is subject to the hole effect model with a certain period. Based on this model, this paper extracted feature parameters characterizing the spatial structure of residential areas, and further the extracted features were applied to the detection of residential areas.(4) Studied the methods of spatial structural feature modeling and residential area extraction based data field. First, this paper studied the relationship between the local spatial statistics in spatial statistics and the data field, and found that by selecting appropriate spatial weights, local spatial statistics can be unified into the potential function system of data field. In other words, local spatial statistics can be used as a potential function of data field, describing the interaction between data objects. Second, drawing on the method of local spatial statistics describing the spatial correlation, the data field method was applied to the spatial structural feature extraction of high resolution images, and the extracted features can be used to improve image classification. Finally, the data field method was used to describe the spatial structure of residential areas, based on which this paper proposed a novel method of residential area detection form high resolution images. This method regarded buildings or their partial structures (such as feature points, feature lines) in residential areas as mass points, and used potential functions of data field to describe the internal structure of residential areas; Then, according to potential value differences between the residential and non-residential areas, residential areas can be extraced from backgrounds by threshold segmentation, and the detected residential areas can be further improved by some post-processings such as "noise" removal, hole filling.更多还原