基于高分影像的滑坡提取关键技术研究

发布时间：2018-06-23 04:58

本文选题：高分影像 + 变化检测　；参考：《中国地质大学》2017年博士论文

【摘要】：快速准确地从滑坡数据中寻找到灾害发生区域,并标示出受灾范围和程度等信息对于开展灾后救援具有重要意义。遥感技术可以不进入现场即可获得灾情信息,在滑坡灾后救援方面得到了重要应用。然而,由于灾损地区往往地表覆盖复杂多变,故在缺少辅助数据的情况下难以仅利用灾后影像通过分类的方式提取灾情信息。随着遥感技术的发展,目前基于双时相影像的变化检测技术得到了快速发展。利用灾前灾后影像结合变化检测方法可以更为直观地显示灾损情况,因此发展基于变化检测原理的滑坡信息提取方法具有十分重要的意义。近年来高分辨率影像得到了迅速的发展,并被广泛应用于资源调查、数据库更新、灾后救援等领域。早期的变化检测算法研究都是针对中低分辨率影像开展的,现有针对高分影像的算法还鲜有应用于滑坡研究的文献报道,且方法一般会针对某种类型的变化,故当将其应用于滑坡信息提取时会存在诸多方面的问题,造成提取结果精度低下。因此需要开展基于高分辨率影像的变化检测方法在滑坡提取方面的研究,以满足灾后救援对滑坡信息快速、高精度提取的迫切要求。由于高分影像变化检测面临数据量庞大和影像分割尺度难以精准确定问题,且在提取变化信息时阈值也难以最佳获取,现有研究一定程度上可以解决上述问题,但方法的适应性等有待深入探讨。基于此,本文开展了基于变化检测原理的高分辨率影像滑坡信息提取研究,完善了基于高分影像在像素和对象不同层面的滑坡信息提取及阈值自动选取理论。本文主要研究内容及创新点如下:1)以像素为处理单位,将ICA原理引入到滑坡提取研究,提出了基于ICA/MNF原理的提取方法。由于地表覆盖差异、噪声等问题,尤其是灾后数据中地表杂乱无章,影像往往不会服从高斯分布,这与PCA、CCA等常用方法以服从高斯分布为基础的基本假设不相符,则运用此类方法提取相关信息难以取得最好的提取效果。而ICA方法以信号的非高斯性为基础,故而用其提取影像的独立成分是十分有利的。本文针对遥感影像波段之间存在信息冗余且高分影像数据量庞大的问题,提出了基于ICA/MNF的变化信息提取方法。算法以单一时相的影像为基础,对双时相影像分别运用基于负熵最大化的Fast-ICA算法分离出相互正交的独立成分,并构建对应独立成分的差异成分。由于变化信息分布于多个成分中,为利用少量的成分获取到最大的变化信息,本文使用MNF算法对差异成分进行了变化信息的集中,通过设置滑动阈值获取了差异影像的阈值,最终提取了滑坡信息。通过对两组数据进行实验证实了方法的可行性,灾害提取结果的整体精度和Kappa精度分别达到了75%和0.47以上;同时与基于像素的差异主成分法、波段差异法两种方法进行了实验对比,验证了方法的优越性。2)开展了面向对象基于RF原理的全特征、优选特征影像分类研究,通过对比灾前灾后影像的对象类别提取了滑坡信息,同时与SVM、KNN算法进行了对比分析,得出了部分有益的结论。由于以像素为单位进行滑坡提取存在严重的“椒盐”效应,且在提取的连续变化区域中存在“空洞”现象,从而影响最终结果的精度。本文深入分析了影像分割理论,对叠加的灾前灾后影像经过多次尝试获取了较优分割尺度并进行了分割,提取了对象的光谱、纹理、形状、语义等41个特征;为使用尽量少的特征实现影像的高精度分类,本文使用VarSelRF程序包进行特征优选方面的分析,统计分析了优选特征重要性排序、被选中次数的关系,获取了不同影像的优选特征;使用对象的全特征与优选特征分别进行了RF分类研究,并与SVM、KNN算法进行了对比分析,结果表明RF分类效果明显优于SVM、KNN算法,SVM结果最差,且整体上RF全特征分类精度稍优于优选特征,单一影像分类的整体精度和Kappa精度可分别达到94%和0.89以上;最后结合分类后比较思想,利用优选特征的RF分类结果提取了滑坡信息,结果的整体精度和Kappa精度分别提高到了77%和0.59,变化类的生产者精度较基于像素的ICA/MNF方法得到了提升,一定程度上抑制了漏检滑坡的现象。获得如下结论:I.在特征优选方面:a)被选特征与地表覆盖有重要关系,而不同时相、不同种类影像之间有一定的差异性;b)被选中次数与特征重要性具有一定关系,重要性排名前五的特征在整体上的被选中次数表现为高频,反之亦然;c)高分影像对象中,纹理和形状特征一般是分类中不可或缺的,但当分辨率较低时,纹理和形状信息则不突出,一般表现为极低的被选概率;II.在分类方面:a)全特征的分类精度整体上稍优于优选特征,但二者差别不大;b)将影像分为若干个精确地类时的分类精度明显优于二分类问题,提高二分类精度的可能解决办法是添加其它辅助数据或是加大样本的种类和各个类别样本的数量;III.在应用分类后比较法提取滑坡时,受单一类别分类精度影响,连续变化区域一般可以取得较好的结果,而在不连续区域效果极差。3)开展了影像分割尺度研究,提出了多序列影像对象的概念,以子对象为处理单位进行了变化检测和滑坡提取研究。针对面向对象方法中最优尺度难以获取但其对结果至关重要的问题,本文深入分析了面向对象影像分割中难以确定最优分割尺度问题,提出了多序列影像对象的概念,将等差数列引入到影像分割领域用于影像分割尺度参数的生成,进行了面向对象的滑坡信息提取研究。算法以单一时相的影像为基础分别对其进行多个单一尺度的有序分割,以便观察对象在空间的变化规律,并以双时相影像的最小分割尺度为基准分裂获取子对象,以子对象为单位搜索其在双时相影像各个分割尺度层中的关联对象,通过构建变化特征向量并获取阈值,根据变化特征向量的大小确定子对象是否发生了变化,并最终获得变化信息。结合对两组数据进行实验,滑坡提取结果的整体精度和Kappa精度分别达到了85%和0.68,表明提出方法是切实可行的;通过与以像素为单位的变化向量方法和面向对象方法对比,验证了提出方法的优越性,以及与前述基于ICA/MNF方法和面向对象RF分类后比较法的结果对比分析,发现整体提取精度不仅是最高的,且滑坡类的用户精度提升到了76%,有效抑制了提取结果中的误检现象。4)开展了影像阈值选取研究,改进了蜂群算法并用于最优阈值自动获取。基于前文的分析和实验结果,阈值对于滑坡信息具有至关重要的影响,选择不当会极大的损害算法最终结果的精度。人工蜂群算法具有控制参数少、计算简便、全局搜索能力强等优势。为快速高效地进行图像分割,针对人工蜂群算法存在的收敛速度慢、易陷入局部最优解等问题,提出了一种基于改进人工蜂群算法分割二维OTSU图像的新方法。算法通过对蜜源更新过程中向当前最优蜜源方向进行引导,加快了算法的收敛速度;为避免算法陷入局部最优并加快收敛速度,在对当前最优解附近局部搜索过程中动态缩减了搜索范围,加大了更优解被发现的概率;针对较大梯度值无意义的问题,限定了蜜源范围,提高了算法的效率。以灰度-梯度二维直方图中背景类和目标类的方差-协方差矩阵的迹为测度函数,结合具有不同直方图分布的图像进行了实验,统计不同算法在各个影像获得最优解的用时和迭代次数等信息量,结果表明改进算法具有稳健、高效、快速的特性;同时发现算法对标准测试图像和滑坡提取均具有较好的分割效果,且在含有噪声情况下算法对滑坡提取比测试图像具有相对较优的结果;通过与改进算法但未限制蜜源生成范围、经典ABC算法且不限制蜜源范围两种方法的对比,实验显示改进算法在获得最优解时的迭代次数、整体运行时间以及获得最优解时的用时三个方面均明显优于对比方法,证明了改进算法的优越性。
[Abstract]:It is of great significance to quickly and accurately find the area of the disaster from the landslide data and to indicate the extent and degree of the disaster. The remote sensing technology can obtain the disaster information without entering the site, and it is important to be used after the landslide disaster relief. With the development of the remote sensing technology, the technology of change detection based on the dual phase image has been developed rapidly. Using the pre disaster post disaster image combined with the change detection method can show the damage situation more intuitively, with the development of remote sensing technology. Therefore, it is very important to develop the method of landslide information extraction based on the principle of change detection. In recent years, the high resolution image has been developed rapidly, and is widely used in the fields of resource investigation, database updating, disaster relief and so on. The research of early change detection algorithms is carried out for low and medium resolution images. There are few literature reports applied to the study of landslide research, and the method generally aims at some kind of change, so when it is applied to the extraction of landslide information, there will be many problems, resulting in the low precision of the extraction results. Therefore, a change detection method based on high resolution image is needed to carry out the landslide extraction. In order to meet the urgent requirements of rapid and high precision extraction of landslide information after disaster relief, it is difficult to accurately determine the high resolution image change detection because of the huge amount of data and the image segmentation scale, and the threshold is difficult to obtain when the change information is extracted. The existing research can solve the above problems to some extent. However, the adaptability of the method needs to be discussed in depth. Based on this, this paper carries out the study of high resolution image landslide information extraction based on the principle of change detection, and improves the theory of landslide information extraction and threshold automatic selection based on high resolution images at different levels of pixels and objects. The main research content and innovation points are as follows: 1) pixels For the processing unit, the ICA principle is introduced to the study of landslide extraction, and an extraction method based on the principle of ICA/MNF is put forward. Due to the difference of surface coverage and noise, especially in the post disaster data, the image often does not obey the Gauss distribution, which is not based on the basic assumption that the common methods such as PCA and CCA are based on the distribution of Gauss. It is difficult to obtain the best extraction effect by using this method, and the ICA method is based on the non Gauss character of the signal, so it is very beneficial to extract the independent components of the image. This paper proposes a ICA/MN based on the problem of the information redundancy and the large amount of image data between the remote sensing image bands. The method of extracting the change information of F. Based on the image of a single phase, the Fast-ICA algorithm based on the maximum of negative entropy is used to separate the independent components of each other, and the difference components corresponding to the independent components are constructed. In this paper, the MNF algorithm is used to focus the variation information on the difference components. The threshold of the difference image is obtained by setting the sliding threshold, and the landslide information is extracted. The feasibility of the method is verified by the experiment of two groups of data. The overall accuracy and the Kappa precision of the disaster extraction results are 75% and 0.47 respectively. At the same time, compared with two methods based on pixel differential principal component and band difference method, the superiority of the method is verified by the two methods. The whole feature based on the object oriented RF principle is carried out, the feature image classification is optimized, and the landslide information is extracted by comparing the object categories before the disaster, and the SVM, KNN algorithm is also obtained. A comparative analysis is carried out and some useful conclusions are drawn. Because of the serious "salt and pepper" effect in the extraction of landslides in pixels, there is a "hollow" phenomenon in the continuous changing region of the extraction, thus affecting the accuracy of the final result. After many times, we try to get the better segmentation scale and divide it, extract the 41 features of the spectral, texture, shape, and semantics of the object. In order to realize the high precision classification of the image by using as few features as possible, this paper uses the VarSelRF package to analyze the feature selection, and statistics and analyze the priority order of the selected features, and the selected time is selected. The relationship between the number and the optimal feature of different images is obtained. The RF classification is carried out with the full features of the object and the preferred feature, and the comparison analysis is carried out with the SVM and KNN algorithm. The results show that the RF classification effect is obviously better than the SVM, KNN algorithm and the worst result of SVM, and the overall classification accuracy of RF is slightly better than the optimal feature, and the single image is a single image. The overall accuracy and Kappa precision of the classification can reach 94% and more than 0.89 respectively. Finally, the landslide information is extracted with the RF classification results of the selected features. The overall accuracy and the Kappa accuracy of the results are increased to 77% and 0.59 respectively. The producer precision of the change class is improved compared with the pixel based ICA/MNF method. To a certain degree, the phenomenon of landslide is suppressed to a certain extent. The following conclusions are obtained: I. in feature selection: a) the selected feature is closely related to the surface coverage, but the difference between different types of images is different; b) has a certain relationship with the feature importance, the first five characteristics of the importance ranking are selected as a whole. In the high score image objects, the texture and shape features are generally indispensable in the classification, but when the resolution is low, the texture and shape information is not prominent and generally shows a very low probability of selection; the classification accuracy of the II. in the classification aspect: a) is a little better than the preferred feature, but the two are two. The difference is not significant; b) the classification accuracy of the image classification is obviously superior to the two classification problem, and the possible solution to improve the two classification accuracy is to add other auxiliary data or to increase the type of samples and the number of samples in each category; III. is subject to a single classification precision when the landslide is extracted after the application of classification. As a result, better results can be obtained in the continuous changing region, while the image segmentation scale is studied in the discontinuous region of.3). The concept of multi sequence image objects is proposed, and the research of change detection and landslide extraction is carried out with the sub object as the processing unit. The problem is very important. This paper deeply analyzes the problem that is difficult to determine the optimal segmentation scale in object image segmentation. The concept of multi sequence image object is put forward, and the arithmetic sequence is introduced into the image segmentation field for the generation of image segmentation scale parameters, and the landslide information extraction is studied. The algorithm is single. The image of the time phase is based on the sequential segmentation of multiple single scales, so as to observe the variation of the object in the space, and take the minimum segmentation scale of the dual phase image as the base division to obtain the sub objects, and search the related objects in each cut scale layer of the dual phase image by the sub object, and build the change through the construction. According to the size of the change feature vector, it determines whether the subobjects change, and finally obtains the change information. The overall accuracy and Kappa accuracy of the landslide extraction results are 85% and 0.68 respectively. The results show that the proposed method is feasible, and the method is based on pixels. The comparison between the change vector method and the object oriented method verifies the superiority of the proposed method, and the comparison analysis of the results based on the ICA/MNF method and the object oriented RF classification method. It is found that the overall extraction precision is not only the highest, but the user accuracy of the landslide class is raised to 76%, and the error detection in the extraction results is effectively suppressed. Phenomenon.4) carry out the study of image threshold selection, improve the colony algorithm and use the optimal threshold automatically. Based on the previous analysis and experimental results, the threshold value has a crucial influence on the landslide information, and the accuracy of the final result of the damage algorithm is very large. The artificial bee colony algorithm has less control parameters and simple calculation. In order to segment the image quickly and efficiently, in order to solve the problem of slow convergence and local optimal solution of artificial bee colony algorithm, a new method based on improved artificial bee colony algorithm to segment two-dimensional OTSU images is proposed. The algorithm is directed to the current optimal nectar source in the process of honeysource update. In order to avoid the local optimization and speed up the convergence speed of the algorithm, the search range is reduced dynamically in the local search process near the current optimal solution, and the probability of finding the better solution is increased in the process of local search near the current optimal solution; the nectar source range is limited to the larger gradient value and the efficiency of the algorithm is improved. The trace of the variance covariance matrix of the background class and the target class in the gray-scale gradient two-dimensional histogram is the measure function. The experiments are carried out with the images of different histogram distribution, and the information amount of the optimal solution and the number of iterations of the different algorithms are obtained. The results show that the improved algorithm is robust, efficient and fast. At the same time, it is found that the algorithm has better segmentation effect on the standard test image and landslide extraction, and the algorithm has a relatively superior result on the landslide extraction compared with the test image under the condition of noise, and the comparison of the two methods of the classic ABC algorithm and the nectar source range without limiting the range of the nectar source generation with the improved algorithm is compared with the improved algorithm. It is proved that the improved algorithm is superior to the contrast method in three aspects, the number of iterations, the overall running time and the time of obtaining the optimal solution.
【学位授予单位】：中国地质大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：P237;P642.22

【相似文献】