面向对象的高分辨率影像城市多特征变化检测研究

发布时间：2018-09-07 13:33

【摘要】：遥感影像空间分辨率的提高为城市发展的监测与规划提供了大量地面细节信息,使得城市遥感变化检测研究成为当前遥感研究领域的热点之一。然而,数据信息量的急剧增加,也为遥感影像变化检测技术的发展提出了新的问题与挑战：首先,丰富的地物细节信息使得单一地物对象由多个空间相邻的像元组成,单个像元的光谱变化不足以反映其所在地物的变化情况；其次,影像空间分辨率提高、光谱分辨率受限,导致同类地物光谱差异变大、不同地物光谱相互重叠,“同物异谱、异物同谱”的现象普遍存在；再次,多时相高分辨率影像成像条件的差异,导致同一地物在不同时相的影像中呈现出光谱、空间特征的差异,仅仅通过影像预处理很难彻底消除这些差异,例如：针对海拔高度较高的地物,多时相成像角度的差异将严重影响变化检测效果；最后,成倍增长的多时相影像数据量使得对算法自动性的要求更高。本文在现有变化检测技术的基础上,针对高分辨率遥感影像中的地物对象变化情况,提出了几种新的面向对象的多特征变化检测模型。它们分别着眼于改善面向对象的变化检测方法中的对象“匀质性”问题,提高自动搜索全局最优的变化检测结果的能力,解决多源影像光谱分辨率差异、多时相影像复合分割误差的问题,改进多时相影像房屋变化显著性度量方式,以及提高针对多时相成像角度差异带来的房屋“伪变化”的容错性变化检测能力等等。通过利用QuickBird、 IKONOS等高分辨率遥感卫星影像数据进行实验,验证了各种检测模型的有效性。为了引入本文提出的变化检测模型,我们首先总结了传统遥感影像变化检测方法的基本思路,详细介绍了影像预处理、变化信息提取、阈值选择与精度评价这四项关键技术。其中,在变化信息提取的部分,详细介绍了现有的基于像元光谱信息的变化分析方法中的代数运算法、影像变换法,以及顾及影像空间信息的面向对象的方法和基于神经网络的方法,并运用一组共用多时相QuickBird影像数据对这些方法进行了实验验证与分析。结果证明：传统的基于像元光谱信息的变化检测算法由于没有考虑影像空间上下文信息,已无法满足高分辨率影像变化分析的需要；而现有的顾及影像空间信息的变化检测算法虽然实现了对高分辨率影像空间信息的利用,但仍存在一些问题,包括面向对象的方法中的对象“匀质性”问题、复合影像分割失真的问题以及多时相影像成像角度差异对变化检测结果的影响间题等等。由此引出了本文针对这些问题的具体解决方案：首先,针对高分辨率遥感影像面向对象的变化分析的两大关键问题——阈值选择对自动获取全局最优解的影响以及面向对象方法的对象“均值化”问题,本文提出了两种新的检测模型,在面向对象的思想下,分别利用遗传算法(Genetic Algorithm,GA)自动搜索全局最优解的机制,以及针对多时相影像对象内部像元光谱特征的K-S (Kolmogorov-Smirnov)统计检验,有效解决了上述问题。根据两组多时相QuickBird影像数据的实验证明,基于GA的方法能够通过循环迭代中的遗传操作自动搜索全局最优的地物对象变化检测结果,避免了阈值选择对算法自动性和最优解选择效果的影响：而基于K-S检验的方法有效保留并考察了多时相影像对象内部的像元光谱统计差异,解决了传统面向对象的方法中的对象“均值化”问题,从不同角度提高了高分辨率遥感影像面向对象的变化检测的有效性。其次,通过分析总结多源高分辨率影像变化分析与同源影像的区别与联系,总结了当前多源影像变化分析的难点,即多源影像的光谱分辨率差异问题。针对这一问题,我们提出了一种新的解决方案,根据变化区域与其所在对象的空间关系,定义多时相影像对象的相似性特征,提取相似性较小的影像对象并将其视为变化的影像区域。同时,该方法通过多时相影像分割映射的方式解决了影像复合分割误差的问题,并针对不同基准影像与不同空间尺度下的多类变化检测结果进行了影像区域融合的后处理。根据两组获取自QuickBird与IKONOS卫星传感器的多源多时相影像数据的实验验证与分析,证明了该方法能有效检测多源高分辨率影像的变化区域。最后,为了针对性地监测反应城市发展的房屋目标变化情况,我们总结了高分辨率影像房屋变化分析现存的两大问题：变化显著性度量方式与多时相成像角度差异的影响。首先,针对变化显著性度量方式的问题,提出了基于脉冲耦合神经网络(Pulse-Couplec Neural Network, PCNN)的房屋变化检测方法,通过神经网络的构建,充分考虑各时相房屋特征影像的对象空间上下文信息,并使用多种相关性度量方式,全面考察房屋对象的变化显著性程度,并据此判断房屋对象的变化情况。通过两组多时相QuickBird影像的实验验证,证明了该方法能有效提取高分辨率影像的房屋变化区域。其次,为了尽可能地降低多时相成像角度差异对房屋变化检测结果的影响,房屋容错性变化检测方法通过对多时相房屋特征点的局部影像匹配,容错性地识别出不同时相中空间几何分布特征存在差异的同一房屋对象,并将其从真实变化的房屋区域中剔除。通过多组QuickBird或IKONOS影像实验证明,该方法能有效降低多时相成像角度的差异导致的对房屋“伪变化”的误检,明显提高了房屋变化检测的精度。
[Abstract]:The improvement of remote sensing image spatial resolution provides a lot of ground detail information for urban development monitoring and planning, which makes the research of urban remote sensing change detection become one of the hot topics in the field of remote sensing. Firstly, the abundant detail information makes a single object consist of several adjacent pixels, and the spectral change of a single pixel is not enough to reflect the change of the object. Secondly, the spatial resolution of the image is improved, and the spectral resolution is limited, which results in the spectral difference of the same object becomes larger, and the spectra of different objects overlap each other. Thirdly, the different imaging conditions of multi-temporal and high-resolution images lead to the differences of spectral and spatial characteristics of the same object in different temporal images. It is difficult to completely eliminate these differences only through image pre-processing, such as: for high-altitude objects, more. The difference of temporal imaging angle will seriously affect the effect of change detection. Finally, the multiplication of multi-temporal image data makes the algorithm more automatic.
Based on the existing change detection techniques, several new object-oriented multi-feature change detection models are proposed for the change of objects in high-resolution remote sensing images. These models focus on improving the homogeneity of objects in object-oriented change detection methods to improve the global optimization of automatic search. The ability of change detection, the resolution difference of multi-source image, the error of multi-temporal image composite segmentation, the improvement of multi-temporal image house change saliency measurement method, and the improvement of the ability of fault-tolerant change detection for multi-temporal image angle difference caused by house "pseudo-change" and so on. Experiments on IKONOS and other high resolution remote sensing satellite images verify the effectiveness of various detection models.
In order to introduce the change detection model proposed in this paper, we first summarized the basic ideas of traditional remote sensing image change detection methods, and introduced four key technologies in detail: image preprocessing, change information extraction, threshold selection and accuracy evaluation. Algebraic operation, image transformation, object-oriented method considering image spatial information and neural network-based method are used in information change analysis. A group of common multi-temporal QuickBird image data are used to validate and analyze these methods. The results show that the traditional method based on pixel spectral information is effective. Change detection algorithms can not meet the needs of high-resolution image change analysis because they do not take into account the spatial context information of the image; while the existing change detection algorithms which take into account the spatial information of the image realize the use of high-resolution image spatial information, there are still some problems, including the object-oriented method of the pairing. Such as "homogeneity" problem, composite image segmentation distortion problem and multi-temporal image angle differences on the impact of change detection results and so on.
Firstly, aiming at the two key problems of object-oriented change analysis of high-resolution remote sensing images, the influence of threshold selection on automatic global optimal solution and the problem of object-oriented "mean" of object, two new detection models are proposed in this paper. Under the object-oriented idea, genetic algorithm (Genetic A) is used respectively. The mechanism of lgorithm (GA) searching global optimal solution automatically and the K-S (Kolmogorov-Smirnov) statistical test for the spectral characteristics of pixels in multi-temporal image objects have effectively solved the above problems. Experiments on two sets of multi-temporal QuickBird image data show that the GA-based method can be automatically operated by the genetic operation in the iterative cycle. Searching for globally optimal change detection results avoids the influence of threshold selection on algorithm automaticity and optimal solution selection effect. The K-S test-based method effectively preserves and inspects the statistical difference of pixel spectrum within multi-temporal image objects, and solves the problem of object "mean" in traditional object-oriented methods. It improves the effectiveness of object oriented change detection of high resolution remote sensing images from different perspectives.
Secondly, by analyzing and summarizing the differences and relations between multi-source high-resolution image change analysis and homologous image, the difficulty of multi-source image change analysis is summarized, that is, the spectral resolution difference of multi-source image. Meanwhile, this method solves the problem of image composite segmentation error by means of multi-temporal image segmentation and mapping, and detects multi-class changes for different reference images and different spatial scales. According to the experimental verification and analysis of two groups of multi-source and multi-temporal image data obtained from QuickBird and IKONOS satellite sensors, it is proved that this method can effectively detect the change area of multi-source and high-resolution images.
Finally, in order to monitor and reflect the changes of housing targets in urban development, we summarize two existing problems in high-resolution image housing change analysis: the impact of change saliency measurement and multi-temporal imaging angle differences. Firstly, to solve the problem of change saliency measurement, a new method based on pulse coupling God is proposed. Through the method of house change detection based on the Pulse-Couplec Neural Network (PCNN), and through the construction of the neural network, the object space context information of the house feature images at different time phases is fully considered, and a variety of correlation measures are used to comprehensively inspect the change significance of the house object and judge the change of the house object. The experimental results of two sets of multi-temporal QuickBird images show that the proposed method can effectively extract the change area of houses from high-resolution images. Secondly, in order to minimize the influence of the difference of multi-temporal imaging angles on the detection results of house changes, the method of house fault-tolerant change detection is based on the local shadows of multi-temporal house feature points. By image matching, the same building object with different spatial geometric distribution characteristics in different time phases can be identified faultlessly and removed from the real changing housing area. The accuracy of house change detection is obviously improved.
【学位授予单位】：武汉大学
【学位级别】：博士
【学位授予年份】：2013
【分类号】：TP751;P237

【参考文献】