高光谱图像中的异常成分检测

发布时间：2018-04-03 20:50

本文选题：高光谱图像　切入点：遥感　出处：《南京航空航天大学》2017年硕士论文

【摘要】：高光谱遥感技术的快速发展,提高了拍摄场景的信息丰富度,为遥感图像中的异常(小目标)检测开辟了新的途径,赋予了其更重要的实际意义。为了构建一个完整的高光谱遥感图像异常检测系统,实现异常成分的自动定位,需要对系统中涉及的波段选择、特征提取、异常检测等方法进行研究。本文的主要工作如下:首先,探索了一种基于子空间中主成分最优线性预测的波段选择方法。采用改进相关性度量的谱聚类方法将高光谱波段划分为不同的子空间,并对各子空间中的波段进行主成分分析(Principal Component Analysis,PCA),选择主要分量作为重构目标;以子空间追踪法为搜索策略,从各子空间中选择数个波段对其重构目标进行联合最优线性预测;合并各子空间中的所选波段得到最佳波段子集。实验结果表明,该方法选择的波段子集可以较完整地重构原始数据,与原始数据以及自适应波段选择(Auto Band Selection,ABS)方法、线性预测(Linear Prediction,LP)方法、最大方差主成分分析(Maximum-variance Principal Component Analysis,MVPCA)方法、自相关矩阵波段选择(Auto Correlation Matrix-based Band Selection,ACMBS)方法得到的波段子集相比,其波段子集具有更好的异常检测性能。然后,讨论了一种基于模糊监督与流形结构保持的特征提取方法。采用RX方法粗略计算像元隶属于异常的可能性,并将其作为模糊监督信息;在模糊监督信息的基础上,分别建立同类局部、非同类非局部以及全局的流形结构近邻图,并结合这3种近邻图构建低维映射的目标函数;为了使该方法能够处理新像元并适应非线性结构数据,给出了其线性化与核化方法,以获得线性与非线性投影矢量。实验结果表明,与主成分分析方法、核主成分分析(Kernel Principal Component Analysis,KPCA)方法、局部线性嵌入(Locally Linear Embedding,LLE)方法、局部保持投影(Locality Preserving Projection,LPP)方法相比,该方法的特征数据拥有更高的异常像元显著性,能够获得更佳的异常检测结果。其次,研究了一种基于背景聚类与加权迭代RX的异常检测方法。采用改进密度峰值快速搜索算法将高光谱像元划分为不同的背景类群,在此基础上,结合3种方法对RX检测窗中的外窗像元进行加权,以获得更准确的背景统计模型:依据外窗像元与各背景类群的马氏距离加权,以减小外窗中异常像元的权重;根据各背景类群对待检测像元的贡献度加权,以降低外窗中相异类群背景像元的影响;利用初次检测结果加权后再迭代检测,达到进一步纯化外窗中背景的目的。实验结果表明,与常规RX方法、分块自适应异常点计算(Blocked Adaptive Computationally Efficient Outlier Nominators,BACON)方法,异常加权RX(Weighted Anomaly RX,WARX)方法、概率异常检测(Probabilistic Anomaly Detector,PAD)方法相比,该方法不仅可以明显提高RX方法对检测窗口尺寸的鲁棒性,而且能够获得更高的检测精度。再次,提出了一种基于自适应参数支持向量机(Support Vector Machine,SVM)的异常检测方法。通过无监督检测方法对异常像元进行快速、粗略定位,并将该定位结果作为后验信息输入到支持向量机中;依据后验信息与核空间散度准则自适应确定支持向量机中核函数的参数,并使用该支持向量机在核空间中寻找分离异常和背景的最佳超平面;利用该超平面将像元重新分类为背景和异常,并且迭代上述操作,得到稳定的异常检测结果。实验结果表明,与常规RX方法、核RX(Kernel RX,KRX)方法、支持向量数据描述(Support Vector Data Description,SVDD)方法相比,该方法可以更有效、精确地检测出高光谱遥感图像中的异常成分。最后,提出了一种基于蜂群优化投影寻踪(Projection Pursuit,PP)与加权K最近邻(Weighted K-nearest Neighbor,WKNN)的异常检测方法。结合邻域像元联合定义的峰度与偏度为投影指标,以MABC为寻优方法,使用投影寻踪从高光谱图像中逐次获取投影图像,再根据其直方图提取异常像元;在初检结果的基础上,提取包含像元判别信息与主要结构的特征,结合WKNN方法对初检结果进行提纯。实验结果表明,与RX方法、独立分量分析(Independent Component Analysis,ICA)方法以及混沌粒子群优化(Chaotic Particle Swarm Optimization,CPSO)投影寻踪方法相比,该方法不但可以获得虚警率更低的异常检测结果,而且具有更快的运算速度。
[Abstract]:The rapid development of hyperspectral remote sensing technology, improve the shooting scene information richness, abnormal in remote sensing image (target) provides a new way to detect, given its important practical significance. In order to build a complete hyperspectral remote sensing image anomaly detection system, automatic positioning and implementation of abnormal components. To extract the features of the selection, involving the band system, studied the anomaly detection method. The main work of this paper is as follows: firstly, to explore a method to select the principal component subspace optimal linear prediction based on the improved correlation metric band. The spectral clustering method of high spectral band is divided into different sub spaces. In the space of the bands in the principal component analysis (Principal Component, Analysis, PCA), select the main component as the target for reconstruction; search strategy search for subspace tracking method, from the air In the choose a number of bands are combined to reconstruct the optimal linear prediction; with each subspace in the selected band get the best band subsets. The experimental results show that this method can select band subset of a complete reconstruction of the original data with the original data and adaptive band selection (Auto Band Selection, ABS) method of linear prediction (Linear Prediction LP) method, principal component analysis with varimax (Maximum-variance Principal Component Analysis, MVPCA) method, choose the autocorrelation matrix band (Auto Correlation Matrix-based Band Selection, ACMBS) compared with the band subset obtained, the subset of the performance anomaly detection has better. Then, is discussed. A fuzzy feature extraction method based on manifold structure and maintain supervision. By the method of RX pixels belonging to rough calculation the possibility of abnormal, and as a fuzzy Based on fuzzy information supervision; supervision information, establish similar local manifold structure, nearest neighbor graph of non similar non local and global, and combined with the objective function of the 3 nearest neighbor graph based low dimensional mapping; in order to make the method can handle the new pixel and adapt to the nonlinear structure of the data, given its linearization and nuclear method to obtain the linear and nonlinear projection vector. The experimental results show that the analysis method and the principal component, kernel principal component analysis (Kernel Principal Component Analysis, KPCA), locally linear embedding (Locally Linear Embedding, LLE) method, locality preserving projection (Locality Preserving, Projection, LPP) compared with other methods, the method of data characteristics have a higher significant pixel anomaly, anomaly detection can obtain better results. Secondly, the anomaly detection method is studied based on the background of clustering and weighted iterative RX mining. With the improvement of the peak density of fast search algorithm for hyperspectral pixel is divided into groups of different background, on the basis of this, 3 methods are weighted window RX pixels in the detection window, to obtain a statistical background model is more accurate: Based on Mahalanobis distance weighted window pixel and the background of the group, to reduce weight the abnormal pixels outside the window; according to the weighted degree for each background pixel groups to be detected with, in order to reduce the influence of external window in different groups of background pixels; using the initial test results after weighted iterative detection, to achieve further purification in the background window. The experimental results show that with the conventional RX method, block adaptive anomaly (Blocked Adaptive Computationally Efficient calculation of Outlier Nominators, BACON) method, weighted RX (Weighted Anomaly RX anomaly, WARX) method, the probability of anomaly detection (Probabilistic Anomaly Detector, P AD) compared with other methods, this method not only can improve the robustness of RX method for detecting the size of the window, and can obtain higher detection accuracy. Thirdly, we propose a support vector machine based on adaptive parameters (Support Vector, Machine, SVM) anomaly detection method. The detection method of abnormal fast unsupervised pixel rough, positioning, and the positioning results as a posteriori information input to the support vector machine; based on a posteriori information and spatial divergence criterion to adaptively determine the nuclear parameters of kernel function for support vector machine, and use the support vector machine to find the best hyperplane separating anomaly and background in the kernel space; the super plane the pixel is reclassified as background and anomaly, and the iterative operation, get the results of anomaly detection stability. Experimental results show that with the conventional RX method, RX (Kernel RX, KRX) method, support vector Data description (Support Vector Data Description, SVDD) compared with other methods, this method can effectively and accurately detect abnormal components of hyperspectral remote sensing images. Finally, proposes a bee colony optimization based on projection pursuit (Projection Pursuit, PP) and the weighted K nearest neighbor (Weighted K-nearest, Neighbor, WKNN) anomaly detection methods. The combination of kurtosis and skewness of neighborhood pixels defined as joint projection index, with MABC as the optimization method, using projection pursuit from hyperspectral image to obtain successive projection images according to the histogram extraction abnormal pixel; based on initial inspection results, including extraction characteristics of pixels discriminant information and main structure, combined with WKNN method for purification of the initial inspection results. The experimental results show that, with the RX method, independent component analysis (Independent Component, Analysis, ICA) method and chaotic particle swarm optimization (Chaotic Particle Compared with the Swarm Optimization, CPSO) projection pursuit method, this method can not only obtain the abnormal detection results with lower false alarm rate, but also have faster operation speed.

【学位授予单位】：南京航空航天大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP751

【参考文献】