稀疏与低秩先验下的高光谱分类与检测方法

发布时间：2018-08-14 11:13

【摘要】：高光谱遥感是以遥感影像与光谱的合一为特征的新型遥感技术,其光谱分辨率可高达纳米级,是20世纪80年代以来地球观测技术所取得的重大技术突破之一,在现代军事、矿物勘测、精确农业以及环境监控等领域有着广泛的应用。因此,研究高光谱数据的高效处理与解译具有重要的理论意义和实际应用价值。高光谱分类与目标检测是高光谱数据处理的主要内容。高光谱数据提供了丰富的光谱信息,为人们研究地表物体的特性、进行地物识别创造了条件。但是,这些海量的信息和特殊的数据结构又给人们在图像处理、信息分析、分类和检测等方面提出了严峻的挑战,也就要求人们从多个方面去理解和揭示其物理特征及其变化。本论文在总结高光谱分类与目标检测现状的基础上,通过对高光谱数据自身特性的深入分析,挖掘高光谱数据的稀疏与低秩先验知识,研究了联合空谱信息的高光谱分类与目标检测技术,并设计了相应的高效算法。论文的主要工作和研究成果如下:(1)针对高光谱分类问题,提出了一种基于自适应上下文信息的联合稀疏表示分类方法。高光谱图像不仅具有丰富的光谱信息,而且图像中的每个像元还含有其自身的空间结构。为了充分利用这种空间结构信息,本文利用高阶steering核函数来刻画各个像元的局部空间结构,并将所得到的空间结构与联合稀疏表示模型融合得到最终的分类结果。结果表明,基于自适应上下文信息的联合稀疏表示分类方法可以有效描述地物的空间上下文信息,提高分类精度。(2)通过充分挖掘高光谱图像的全局和局部空间信息,提出了一种基于低秩矩阵分解的高光谱图像分类方法。首先采用一种基于图的分割算法,将高光谱图像分成多个匀质区域。由同一匀质区域内的像元组成的矩阵(矩阵的列向量对应像元的光谱向量)具有很强的列相关性,因此可将该二维矩阵分解成低秩矩阵和稀疏矩阵之和。然后以低秩矩阵中的列向量作为相应像元的特征,利用概率支持向量机(Probabilistic support vector machine,PSVM)进行分类。同时为了精细化分类结果,在PSVM中引入马尔科夫随机场正则化,保证高光谱图像中的全局空间信息、局部空间信息和光谱信息有效结合。实验结果表明,该方法的分类结果优于其他主流分类方法。(3)针对高光谱图像异常检测问题,提出了一种基于低秩和稀疏表示的高光谱异常检测方法。该方法将高光谱图像分解为背景和异常两部分进行分别建模,对于异常部分,根据异常像元在整个图像中只占极少部分,本文将异常部分建模为一个列稀疏矩阵,并采用l2,1范数来刻画。对于背景部分,由于高光谱图像中的背景像元可认为是位于多个子空间,采用低秩表示模型进行建模,在背景字典下搜寻数据的最低秩表示,从而有效刻画背景的全局结构。为了能够准确表示每个像元光谱,刻画像元的局部光谱信息,对其表示系数进行稀疏约束。同时,由于背景字典中的原子须要覆盖所有背景地物种类且不能为异常像元,本文提出了一种新的背景字典构造方法。针对所提模型,设计一个有效算法求解。在模拟和真实高光谱图像上的实验表明,本文所提出的异常检测方法能够抑制背景像元突出异常像元,达到了较高的检测准确率。(4)针对高光谱视频序列气体检测问题,提出了一种基于时空TV(Total Variation)正则化的目标检测方法。根据高光谱视频序列中的气体在空间维和时间维都具有连续性,在RPCA(Robust principal component analysis)模型中引入时空TV正则项刻画气体的空间连续性和时间连续性。同时为了充分利用高光谱视频的光谱信息,采用主成分分析方法抽取每帧图像中的主要特征,设计了一种多特征融合的气体检测方法。实验结果表明本文所提出的时空TV正则化模型能够有效刻画气体的结构,提高检测精度。
[Abstract]:Hyperspectral remote sensing is a new remote sensing technology characterized by the combination of remote sensing images and spectra. Its spectral resolution can reach nanometer level. It is one of the important breakthroughs in earth observation technology since the 1980s. It has been widely used in modern military, mineral exploration, precision agriculture and environmental monitoring. Hyperspectral classification and target detection are the main contents of hyperspectral data processing. Hyperspectral data provide abundant spectral information for people to study the characteristics of surface objects and identify them. Information and special data structure have posed a serious challenge to people in image processing, information analysis, classification and detection, which requires people to understand and reveal its physical characteristics and changes from many aspects. In-depth analysis of characteristics, mining sparse and low-rank prior knowledge of hyperspectral data, hyperspectral classification and target detection technology based on joint spatial spectrum information are studied, and corresponding efficient algorithms are designed. The hyperspectral image not only has abundant spectral information, but also each pixel in the image contains its own spatial structure. In order to make full use of this spatial structure information, this paper uses high-order steering kernel function to characterize the local spatial structure of each pixel, and the resulting spatial structure and structure. The results show that the combined sparse representation and classification method based on adaptive context information can effectively describe the spatial context information of objects and improve the classification accuracy. (2) A low rank moment based method is proposed by fully mining the global and local spatial information of hyperspectral images. A method of hyperspectral image classification based on matrix decomposition is presented. Firstly, a graph-based segmentation algorithm is used to divide hyperspectral images into homogeneous regions. Then, the column vectors of the low rank matrix are used as the feature of the corresponding pixels, and the probabilistic support vector machine (PSVM) is used to classify them. Experimental results show that the proposed method outperforms other mainstream classification methods. (3) To solve the problem of hyperspectral image anomaly detection, a hyperspectral anomaly detection method based on low rank and sparse representation is proposed. For the abnormal part, the abnormal part is modeled as a column sparse matrix and characterized by l2,1 norm. In order to represent the local spectral information of each pixel accurately, the representation coefficients of each pixel are sparsely constrained. At the same time, the atoms in the background dictionary must cover all species in the background and can not be abnormal images. An efficient algorithm is designed to solve the proposed model. Experiments on simulated and real hyperspectral images show that the proposed anomaly detection method can restrain background pixels from highlighting abnormal pixels and achieve a high detection accuracy. (4) For hyperspectral video sequences, the proposed method is effective. In this paper, a target detection method based on the regularization of space-time TV (Total Variation) is proposed for column gas detection. According to the continuity of gas in both space and time dimensions in hyperspectral video sequences, the space-time TV regularization term is introduced into the RPCA (Robust principal component analysis) model to characterize the space continuity and time continuity of gas. At the same time, in order to make full use of the spectral information of hyperspectral video, the principal component analysis (PCA) method is used to extract the main features of each image, and a multi-feature fusion gas detection method is designed.
【学位授予单位】：南京理工大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP751

【相似文献】