基于协同分割的遥感图像变化检测

发布时间：2018-06-24 17:15

本文选题：变化检测 + 协同分割　；参考：《北京建筑大学》2017年硕士论文

【摘要】：地表覆盖是指地球表面各种物质类型及其自然属性与特征的综合体。航空航天技术的发展,实现了覆盖全球的卫星对地观测。人们可以使用遥感卫星数据来获取大面积甚至全球的地表覆盖信息,及时准确的掌握地表覆盖类型的分布及其变化情况。地表覆盖遥感产品的研制成功,极大地方便了人们在气候变化研究、生态环境监测和可持续发展规划等领域对地表覆盖信息的应用。为了满足人们对地表覆盖信息使用的现势性要求,学者们提出了多种变化检测方法以从不同时相的遥感图像中提取出变化信息来完成对地表覆盖信息的更新。变化检测方法的核心在于如何发现变化,测定变化范围及变化属性。传统的变化检测方法以图像像元作为基本的处理单位,利用像元的光谱特征来构建特征值,将阈值作为判断变化与非变化的标准,在这种情况下,即使同时使用多个特征值组合进行判断,也容易造成错分误差或者漏分误差,影响变化检测的精度。面向对象的变化检测方法考虑了图像的空间信息,将多个像元组成的同质对象作为变化检测的基本单位。如何确定不同地类的最佳分割尺度,如何确定不同时相的图像对象的空间对应性是变化检测开始前需要解决的问题。面向对象的分割方法是“先分割,后检测”的模式,其对变化区域的判断仍然主要依靠阈值来进行。阈值选择的准确性是变化检测精度的决定性因素。计算机视觉中的协同分割算法,能够从同一场景的多视图像中分割出相同或近似的目标。该算法由于利用了图像之间的联系,因此能够挖掘出更多的图像信息。如果把土地覆盖的变化过程看作自然界的运动,则土地覆盖的变化检测问题就可以看作运动图像的协同分割问题。考虑到基于像元和面向对象的变化检测方法将阈值作为判断变化/非变化的标准,本文根据协同分割的思想,构建了适用于变化检测的能量函数。能量函数中包含的变化特征项以两时相图像共同的变化信息为基础,将基于像元和面向对象方法中的阈值条件作为判断是否发生变化的先验知识。同时为了更好的利用图像的空间信息,能量函数中包含的图像特征项能够反映邻域内像元之间的相似性。本文利用基于图割的方法来解决能量函数的最小化问题,将图像映射为图,将能量函数中的各项作为图中不同类的边的权重,使用最小割/最大流方法求得图的最小割,从不同时相的图像中分割出空间对应的变化图斑。本文创新点主要有:(1)根据变化检测方法与计算机视觉中协同分割方法的共同特点,将变化信息特征值的阈值条件作为能量函数中变化特征项的先验知识,改变了基于像元和面向对象的变化检测方法中以阈值条件作为变化和非变化的唯一标准。(2)能量函数中的图像特征项的构建,综合考虑邻域内像元对的光谱和纹理特征,通过对不同时相的图像特征项赋予权重值,构建综合图像特征项。根据不同图像特征参与的分割结果,不仅能够提取出变化位置和区域,也能够对几何属性进行判断。
[Abstract]:Surface coverage refers to a variety of material types and their natural properties and characteristics of the earth's surface. The development of Aeronautics and Astronautics has achieved global satellite observation. People can use remote sensing satellite data to obtain large area and global surface coverage information, and to accurately grasp the distribution of surface cover types. The application of surface coverage information in the fields of climate change research, ecological environment monitoring and sustainable development planning. In order to meet the potential requirements for people to use the surface coverage information, a variety of change detection methods have been put forward by scholars. The core of the change detection method is how to find the change, determine the change range and the change attribute. The traditional change detection method takes the image pixel as the basic processing unit, and uses the spectral characteristics of the pixel to construct the eigenvalues, and the threshold value is used. As a criterion for judging change and non change, in this case, even using multiple eigenvalues at the same time, it is easy to cause error or leakage error and affect the accuracy of change detection. The object oriented change detection method considers the spatial information of the image, and makes the homogeneous objects composed of multiple pixels as the change. The basic unit of detection. How to determine the best segmentation scale of different classes and how to determine the spatial correspondence of the image objects in different phases is a problem that needs to be solved before the beginning of change detection. The object oriented segmentation method is the model of "first segmentation and post detection", and the judgment of the changing region is still mainly based on the threshold. The accuracy of the threshold selection is the decisive factor of the change detection precision. The cooperative segmentation algorithm in computer vision can divide the same or approximate target from the multi view image of the same scene. The algorithm can excavate more image information because of the connection between the images. As the process is regarded as the movement of nature, the problem of change detection of land cover can be regarded as the problem of synergetic segmentation of motion images. Considering the change detection method based on pixel and object oriented change detection method as the criterion for judging change / non change, this paper constructs the energy function suitable for change detection according to the idea of cooperative segmentation. The variation feature contained in the energy function is based on the common change information of the two phase image. The threshold condition based on the pixel and the object oriented method is used as the prior knowledge to judge whether the change is occurring. In order to better use the spatial information of the image, the image feature included in the energy function can reflect the neighborhood. In this paper, a graph cut method is used to solve the minimization of the energy function, and the image is mapped into a graph. The minimum cut of the graph is obtained by using the minimum cut / maximum flow method in the energy function as the weight of the edges of the different classes in the graph. The spatial corresponding change map is segmented from the images of different phases. The main innovations of this paper are as follows: (1) according to the common characteristics of the change detection method and the cooperative segmentation method in the computer vision, the threshold condition of the change of the characteristic value of the information is taken as the prior knowledge of the change feature in the energy function, and the threshold condition is changed and non variable in the variational detection method based on the pixel and the object oriented. The only standard of the transformation. (2) the construction of the image feature item in the energy function, considering the spectral and texture features of the pixel pair in the neighborhood, and building a comprehensive image feature item by giving weight values to the image feature items of different phase. Enough to judge the geometric properties.
【学位授予单位】：北京建筑大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP751

【参考文献】