遥感影像多层次信息变化检测研究
本文选题:变化检测 + 多时相 ; 参考:《武汉大学》2015年博士论文
【摘要】:遥感影像变化检测是出现最早、应用最广泛的遥感数据分析技术之一。遥感变化检测能够提供地球表面大范围、长时间、周期性的地物变化监测信息,为研究地球生态系统变化和人类社会发展提供了丰富的数据,具有重要的研究意义和实用价值。随着遥感技术的发展,目前已经出现了面向不同层次观测任务的多种分辨率遥感数据,包括中低分辨率多光谱遥感影像、高光谱遥感影像和高分辨率遥感影像。不同类型不同分辨率的多时相遥感数据能够为地表变化观测提供变化区域、变化类型以及语义变化情况等多层次变化信息。然而,如何有效地提取多层次变化信息却面临着许多难题。首先,外界成像因素不同所造成的辐射差异使得真实变化信息难以识别;其次,高维光谱特征难以得到充分利用;最后,高分辨率影像所包含的语义层次变化信息难以分析。为了克服以上问题,充分挖掘多种类型遥感数据所包含的多层次信息,本文结合机器学习和模式识别领域的最新研究成果,对遥感影像多层次信息变化检测理论与方法进行了深入的探讨。本文的主要研究工作如下:(1)总结了遥感影像变化检测的经典理论和基本思路,按照预处理、变化检测、阈值分割、精度评价以及应用五个部分对变化检测研究进行了系统的介绍。在预处理部分主要介绍了数据选择、配准和辐射校正三个主要的预处理步骤。在变化检测部分根据不同分辨率数据介绍了传统遥感变化检测方法、高光谱遥感变化检测方法以及高分辨率遥感变化检测方法。接着详细介绍了阂值分割和精度评价中的常用方法,最后对变化检测目前的应用情况进行了概括和总结;(2)发展了针对多时相遥感影像分析的遥感慢特征分析理论,通过提取时间维不变特征,抑制外界成像条件不同造成的辐射差异,突出真实变化信息。根据这一理论提出了非监督、监督和迭代三种变化检测算法,通过未变化训练样本或迭代定权方式来排除真实变化像素对慢特征学习的影响,提高特征提取准确性,有效提高变化检测精度。提出基于迭代慢特征分析的自动相对辐射校正算法,根据迭代慢特征分析算法获取的像素权值自动计算对应波段间的线性关系函数,校正辐射差异,保持光谱一致性。针对地物变化较剧烈的影像,提出监督的自动辐射校正算法,用少量监督信息就可以寻找到正确的线性关系,实现自动相对辐射校正。提出基于慢特征分析的高光谱异常变化检测算法,通过慢特征分析获取最小化背景光谱变化的投影特征,在变化残差中高精度地检测异常地物变化。(3)研究基于子空间理论的遥感变化检测方法,利用子空间的方式来表达地物光谱特征,检测光谱特征变化,提取不同的地物变化类别信息。提出基于子空间距离的高光谱变化检测算法,通过对每一像素构建多时相背景子空间检测对应像素的光谱特征差异,并通过在子空间构建中加入空间信息和光谱信息来减小配准误差和抑制特定地物变化。提出基于独立成分分析的高光谱变化分析算法,利用独立成分分析算法在高光谱差值影像中提取独立地物的变化情况,实现不同地物的变化信息提取。提出了目标变化检测算法,将特定地物变化类型看作由变化前后地物光谱组成的特定类别,在叠加高光谱数据中采用目标探测的方法提取特定地物变化类型的分布情况。(4)首次提出场景变化分析的概念,利用高分辨率遥感影像在场景语义层次检测特定区域内土地利用类型变化情况。在理论上总结并讨论了场景变化分析的定义和研究意义,分析了其应用价值和面临的研究难题。实现了基于视觉词袋模型的遥感影像场景识别算法,通过实验验证了该方法的有效性并对参数设置进行了对比和分析,为场景变化分析研究提供理论和实验基础。提出了基于视觉词袋模型的场景变化分析框架,在字典学习中加入时间信息实现独立字典学习、叠加字典学习和联合字典学习,根据分类变化检测思想提出分类后场景变化分析和联合分类场景变化分析两种方式。实验证明了场景变化分析的有效性和应用潜力。
[Abstract]:Remote sensing image change detection is one of the earliest and most widely used remote sensing data analysis techniques. Remote sensing change detection can provide large range, long and periodic monitoring information of earth's surface change, which provides abundant data for the study of the changes of the earth's ecosystem and the development of human society. It has important research significance and is of great significance. Practical value. With the development of remote sensing technology, a variety of resolution remote sensing data, including medium and low resolution multi spectral remote sensing images, hyperspectral remote sensing images and high resolution remote sensing images, have been developed for different levels of observation tasks. The multi temporal remote sensing data of different types and different resolutions can provide the surface change observation. However, how to effectively extract multi-level change information is faced with many difficult problems. First, the radiation differences caused by different external imaging factors make it difficult to identify the real change information; secondly, the high dimensional spectral characteristics are difficult to be fully utilized; In order to overcome the above problems, the multilevel information contained in many types of remote sensing data is fully excavated. This paper combines the latest research results in the field of machine learning and pattern recognition, and makes an in-depth study of the theory and method of multi-layer information change detection in remote sensing images. The main research work of this paper is as follows: (1) summarize the classic theory and basic idea of remote sensing image change detection, introduce the five parts according to preprocessing, change detection, threshold segmentation, accuracy evaluation and application. In the preprocessing part, it mainly introduces the data selection, registration and radiation. Three main preprocessing steps are corrected. In the change detection section, the traditional remote sensing change detection method, the hyperspectral remote sensing change detection method and the high resolution remote sensing change detection method are introduced in accordance with the different resolution data. Then the common methods in the threshold segmentation and accuracy evaluation are introduced in detail, and the change detection is at present at the end. The application situation is summarized and summarized. (2) the theory of remote sensing slow feature analysis for multi-phase remote sensing image analysis is developed. By extracting the invariant features of time dimension, the radiation differences caused by different external imaging conditions are suppressed and the real change information is highlighted. According to this theory, three kinds of change detection algorithms are proposed, which are unsupervised, supervised and iterated. Method to eliminate the influence of real changing pixels on slow feature learning by unchanging training samples or iterative weights, to improve the accuracy of feature extraction and improve the accuracy of change detection. An automatic relative radiation correction algorithm based on iterative slow feature analysis is proposed, and the weight value of pixels obtained by the iterative slow feature analysis algorithm is automatically calculated. According to the linear relation function between the corresponding bands, the radiation difference is corrected and the spectral consistency is maintained. A supervised automatic radiation correction algorithm is proposed for the image of the more violent ground objects. With a small amount of supervision information, the correct linear relationship can be found and the automatic relative radiation correction is realized. The hyperspectral anomaly based on slow feature analysis is proposed. Detection algorithm, through slow feature analysis, to obtain the projection features of minimal background spectral changes and detect abnormal ground changes in the changing residuals. (3) study the remote sensing change detection method based on subspace theory, use subspace to express spectral characteristics, detect spectral features, and extract different terrain changes. A hyperspectral change detection algorithm based on subspace distance is proposed. By constructing a multi-phase background subspace of each pixel to detect the spectral characteristics of the corresponding pixels, and adding spatial information and spectral information in the subspace construction to reduce registration errors and suppress specific ground changes. The analysis algorithm of hyperspectral change analysis uses the independent component analysis algorithm to extract the change of the independent objects in the high spectral difference image, and realizes the information extraction of the change of different objects. A target change detection algorithm is proposed, which is regarded as a specific category of the spectral composition of the ground objects before and after the change, and the superposition is high. In spectral data, the method of target detection is used to extract the distribution of specific terrain change types. (4) the concept of scene change analysis is proposed for the first time. The change of land use types in specific region is detected by high resolution remote sensing images in the scene semantic level. The definition and research of scene change analysis are also discussed in theory. The significance, analysis of its application value and the problems facing the research. The algorithm of remote sensing image scene recognition based on the visual word bag model is realized. The validity of the method is verified by the experiment and the parameter setting is compared and analyzed. The basis of theory and experiment is provided for the scene change analysis and research. A visual word bag model is proposed. In the framework of scene change analysis, we add time information in dictionary learning to realize independent dictionary learning, superposition dictionary learning and joint dictionary learning. According to classification change detection ideas, two ways are proposed for scene change analysis and joint classification of scene change analysis after classification. The experiment shows the effectiveness and application potential of scene change analysis.
【学位授予单位】:武汉大学
【学位级别】:博士
【学位授予年份】:2015
【分类号】:P237
【参考文献】
相关期刊论文 前10条
1 张良培;;高光谱目标探测的进展与前沿问题[J];武汉大学学报(信息科学版);2014年12期
2 周启鸣;;多时相遥感影像变化检测综述[J];地理信息世界;2011年02期
3 魏立飞;钟燕飞;张良培;李平湘;;遥感影像融合的自适应变化检测[J];遥感学报;2010年06期
4 罗文斐;钟亮;张兵;高连如;;高光谱遥感图像光谱解混的独立成分分析技术[J];光谱学与光谱分析;2010年06期
5 张良培;黄昕;;遥感影像信息处理技术的研究进展(英文)[J];遥感学报;2009年04期
6 徐前祥;盛辉;廖明生;;MNF与MAD变换相结合的城市扩展研究[J];国土资源遥感;2006年04期
7 钟家强;王润生;;基于独立成分分析的多时相遥感图像变化检测[J];电子与信息学报;2006年06期
8 谭炳香;李增元;陈尔学;庞勇;;EO-1 Hyperion高光谱数据的预处理[J];遥感信息;2005年06期
9 李德仁;利用遥感影像进行变化检测[J];武汉大学学报(信息科学版);2003年S1期
10 廖明生,朱攀,龚健雅;基于典型相关分析的多元变化检测[J];遥感学报;2000年03期
相关博士学位论文 前4条
1 汤玉奇;面向对象的高分辨率影像城市多特征变化检测研究[D];武汉大学;2013年
2 杜博;高光谱遥感影像亚像元小目标探测研究[D];武汉大学;2010年
3 黄昕;高分辨率遥感影像多尺度纹理、形状特征提取与面向对象分类研究[D];武汉大学;2009年
4 钟家强;基于多时相遥感图像的变化检测[D];国防科学技术大学;2005年
相关硕士学位论文 前1条
1 戴文渊;基于实例和特征的迁移学习算法研究[D];上海交通大学;2009年
,本文编号:1815294
本文链接:https://www.wllwen.com/kejilunwen/dizhicehuilunwen/1815294.html