【作者】 成宝芝；

【导师】 赵春晖；

【作者基本信息】 哈尔滨工程大学 ， 信号与信息处理， 2012， 博士

【摘要】 高光谱图像是一种新型的具有图谱合一特性的遥感图像，与多光谱等图像相比，其对地表物质间的细微差异通过连续的光谱曲线有更好的表达，使得高光谱遥感图像在地表物质的分类、解混、目标探测和异常检测等方面得到广泛应用。在高光谱遥感图像的诸多应用中，异常目标检测是近年来的一个热点，因为其不需要先验光谱信息，仅通过高光谱遥感图像自身特性就可以把需要检测的目标从背景信息中提取出来，具有较强的实用性。许多研究者针对异常目标检测进行了深入研究，并提出了相应的异常检测算法。但是，由于高光谱图像自身存在的高数据维，严重的背景干扰，同物异谱，波段间非线性相关性和图像中含有混合像元等问题，使得现有的异常检测算法在一定程度上都存在着或多或少的不足。现实的需要要求研究者开发一些新的算法解决目前异常目标检测存在的问题。本文依据真实的高光谱遥感图像，在分析其光谱分辨率特性、波段间相关性和背景模型等特性的基础上，以高光谱遥感图像的光谱特性为主线，利用高光谱降维技术、特征提取技术和非线性核机器学习等方法，提出了三种有创新性的异常目标检测算法。论文主要的工作如下：首先，论文对高光谱遥感图像的成像机理进行了简单介绍，详细分析了高光谱遥感图像的光谱分辨率特性、波段间相关性和背景模型等特性。并重点介绍了核函数涉及的相关内容，特别是对后续研究中用的最多的高斯径向基核函数进行了详细阐述。对这些光谱特性和核函数等问题的分析，为后文的研究工作开展打下基础。其次，针对高光谱图像的高数据维、数据间的冗余性和背景噪声干扰问题，论文提出了利用光谱维变换和空域滤波的异常检测算法。该算法首先利用最大噪声分量变换技术对原始高光谱图像进行变换，按照变换后的各波段的信噪比，设定合适的阈值，提取一定数量的波段对应的最大噪声分量变换矩阵，构造正交子空间投影算子。将变换后的高光谱图像数据投影到这个正交子空间上，得到残差数据；再利用主成分分析方法对此残差数据进行特征提取，使异常目标的能量集中在前面几个主成分中，并设定合适的阈值，确定需选择的主成分数目；然后，把含有丰富异常目标信息的主成分分量构成的高光谱图像数据集通过空域滤波器滤波，最大可能消除数据中含有的各类噪声。最后，把滤波后的波段子集数据输入RX异常检测器进行异常目标检测。该算法很好地解决了已有算法在处理高光谱图像高数据维、数据间的冗余性和背景噪声干扰等方面的不足。再次，针对高光谱图像波段间数据冗余性和背景信息干扰导致经典核RX算法检测性能不高等问题，提出了基于四阶累积量的波段子集非线性异常检测算法。该算法先依据高光谱图像波段具有分块相关性的特点，计算相邻波段的相关系数，将原始图像划分为维数不同的波段子空间；然后，利用主成分分析构造的正交子空间对各波段子集进行背景抑制，得到图像误差数据；在此基础上，再次利用主成分分析提取各波段子集的特征信息，使异常目标信息集中于前面几个波段；最后，利用高光谱数据含有异常目标时背景呈现奇异性不符合高斯分布的特性，提取各子集主成分中含有最大四阶累积量值的波段，构成最优波段子集，并与核RX算法结合进行异常检测。该算法充分利用了高光谱图像的光谱特性，克服了高光谱图像的高维性、波段间的冗余性和背景信息干扰对经典核RX算法的影响。最后，针对含有混合像元情况的异常目标检测问题，提出了一种新的结合光谱解混技术的支持向量数据描述的高光谱图像异常检测算法。该算法将光谱解混技术引入到异常检测问题中，利用新的约束非负矩阵分解的解混算法实现高光谱图像复杂背景信息和目标信息的分离，使解混后的误差数据含有丰富的目标信息，抑制了背景干扰信息；然后将解混误差数据利用非线性的支持向量数据描述方法映射到高维特征空间，在充分利用高光谱图像波段间的非线性统计特性的基础上，完成异常目标的检测。该算法很好的解决了背景信息和异常目标混合问题，最大程度抑制了背景信息对异常目标的干扰，突出了异常目标，提高了高光谱图像异常目标检测的精度。更多还原

【Abstract】 Hyperspectral image is a new class of remote sensing image with the property of imageand spectrum, it has very good expression to the slight differences in surface substance bycontinuous spectrum curve by comparing with multispectral, and make hyperspectral image tohave widely application on spectral classification, spectral unmixing, target detection andanomaly detection etc. In recent years, the anomaly target detection is a hot issue in theapplication field of hyperspectral remote sensing image because of no prior knowledge of thetarget spectral signature is utilized or assumed. That is very useful in reality because thehyperspectral remote sensing image can put the need to detection target from backgroundinformation through spectral features. The existing algorithms are boundedness due to thehigh dimensionality, background interference, nonlinear correlation between bands, andmixed pixels in hyperspectral imagery. Based on the real hyperspectral image, the thesisanalysizes spectral resolution characteristics, band correlation and the background model.Then, based on the spectral characteristics, it is proposed several innovative anomaly targetdetection algorithm by using hyperspectral dimension reduction techniques, feature extractiontechnology and nonlinear kernel machine learning method. The thesis mainly work asfollows:Firstly, the thesis introduces the imaging mechanism of hyperspectral remote sensingimage in brief, and analysizes spectral resolution characteristics, band correlation andbackground model in detail. And the thesis mainly describes the kernel function methods,specially, it describes the gaussian RBF kernel function which is a foundation for the laterresearch work.Secondly, a novel anomaly detection algorithm is proposed for hyperspectral images toresolve the high dimensionality, the nonlinear statistical property and background interference,which is the extended RX algorithm based on spectral dimension transformation and spatialfilter(STSF-RX). Firstly, the maximum noise fraction transform is performed on the originalhyperspectral images, and it obtains MNF transform matrices by setting a SNR threshold, thatthe SNR of their corresponding bands are larger than the threshold. Then, for suppressingbackground interferences, The orthogonal subspace projection is estabished by MNFtransform matrices, and project the hyperspectral data of MNF transform to the orthogonal subspace, and obtain the error data of hyperspectral. In order to concentrate the energy ofdetection targets a few components for the first, the method of principal components analysisis performed on the processing hyperspectral image. Finally, we obtain bands for PCAtransform based on the eigenvalue threshold, the eigenvalue of the bands are larger thanthreshold, and the bands are input to the RX detector. the proposed STSF-RX algorithm areeffectively resolved the high dimensionnality and background interferences.Thirdly, The Kernel RX is the classical anomaly detection algorithm of hyperspectralimages, which exploits nonlinear statistical property between bands. But, the algorithm hasn’tthink about the complexity of spectral and spatial speciality. In addition, the high dimensi-onality and redundancy between bands affect the performance of the KRX algorithm.Therefore, the detection performance of Kernel RX is low. Aiming at the problem, the paperproposes the algorithm of band subsets anomaly detection of hyperspectral image based onfourth order cumulant(KBS-KRX). Firstly, the algorithm divides the original hyperspectralimage to subset of low dimensions according to the correlation coefficient between spectralbands. Then, the error image data is achieved via subset bands were suppressed backgroundinterferences for detection by using orthogonal subspace projection, which is producd by theprincipal component analysis. Based on the data, the feature Information of all band subsetswere extracted by using the principal component analysis, which make the information ofanomaly target concentration on previous bands. At last, the optimal band subsets wereachieved by fourth order cumulant in principal component. In band subsets, the anomalydetection is carried by the kernel RX. KBS-KRX algorithm makes full use of blockingcharacteristics of band correlation in hyperspectral images.Finally, the thesis proposes a novel algorithm for spectral unmixing support vector datadescription(SU-SVDD), it resolves the mixed pixels anomaly detection problem, the proposedalgorithm introduces hyperspectral unmixing into the problem of anomaly detection toseparate target information from complicated background clutter. After spectral unmixing, theerror datum includes abundant target information, and at the same time suppressedbackground interference; then the error datum is mapped into a high dimensional featurespace by a nonlinear SVDD. By exploiting nonlinear information between the spectral bandsof hyperspectral imagery, the anomaly targets can be detected. The proposed algorithm canhighlight anomaly target and highly detect precision.更多还原