【作者】 马丽；

【导师】 田金文； Melba M. Crawford；

【作者基本信息】 华中科技大学 ， 控制科学与工程， 2010， 博士

【摘要】 高光谱遥感数据的光谱分辨率很高,能够获取地球表面丰富的光谱和空间信息,反映地物本质的物理化学特性,使得在宽波段遥感中不可识别的地物,在高光谱遥感中能够被识别。地物分类和异常检测是高光谱遥感的重要应用方向,对理解地物分布规律以及探测感兴趣目标具有重要作用,是论文研究的主要内容。高光谱图像具有很高的维数,存在数据冗余,噪声,和维数灾难问题,而且不是每个特征对所分析的问题都具有作用,降维是解决这个问题的有效方法。降维可以发现高维数据中所隐藏的低维结构,减少后续处理的负担,而且可能提高数据分析的质量。流形学习是一类重要的非线性降维方法,它假设高维数据位于一个低维流形中,该低维流形能够表示数据的本征结构和非线性特性。由于高光谱数据存在固有的非线性特性,线性降维可能丢失数据某些重要的非线性信息,因此,论文研究基于流形学习非线性降维的高光谱数据分析。流形学习分为全局算法和局部算法,论文重点研究局部流形学习算法在高光谱图像分类和异常检测中的应用,主要从以下几个方面开展研究工作:(1)流形学习算法存在的一个问题是无法对新数据进行泛化,论文采用Bengio提出的基于核的泛化算法框架,实现对多种流形学习算法的泛化。泛化算法的关键是给出流形学习算法对应的核函数,对于局部切空间排列算法,论文推导出其核函数,实现其对新数据的泛化。(2)对于高光谱图像分类,论文对多种流形学习算法进行比较研究,采用k近邻分类器对各种降维算法的分类性能进行评价,以更好理解流形学习算法的性能,以及高光谱图像在流形域的数据特性。通过大量实验得到多个有意义的结论:流形学习算法是一种有价值和前景的高光谱数据预处理方法,其最大优势是在两类分类问题中,可以提高难以区分地物的分类效果;另外,有监督局部流形学习算法具有最好表现,能够较大幅度提高分类性能。(3)基于流形学习算法和k近邻分类器结合的研究,提出一种新的基于有监督局部流形学习算法的加权k近邻分类器,应用于高光谱图像分类。权值由流形学习算法的核函数计算,可以描述数据局部几何结构,有效评价各近邻点对测试点分类的作用,提高k近邻分类器的性能。该分类器计算简单,只需要计算近邻点权值,因而适用于大数据量情况,还可以有效缓解不均衡样本对k近邻分类器的影响。(4)针对高光谱图像异常检测存在的问题,采用鲁棒的流形学习算法,以避免异常信息对背景特性的影响,建立更准确的背景流形,提高异常检测性能。首先对鲁棒的局部线性嵌入算法,通过将图像分割成多个子块的方式降低其计算量,但是无法得到全局降维结果。然后提出基于背景训练点选择的鲁棒流形学习算法,其中背景训练点由递归多层图像分割方法获得,可以建立全局数据流形,并避免异常对背景流形的影响。在局部流形学习算法中,鲁棒的局部切空间排列算法具有最好表现。更多还原

【Abstract】 The hyperspectral data with high spectral resolution are able to obtain abundant spaial and spectral information, characterize the inherent physical and chemical properties of land cover types, and provide superior capability for discriminating materials than multispectral data. Classification and anomaly detection in hyperspectral data are the focus of this research, which can facilitate understanding the land cover distribution and detecting the interesting targets.The high number of spectral bands, interband spectral redundancy, and the ever presented noise present challenging problems for analysis of hyperspectral data. Moreover, not all the bands are important for understanding the phenomena. Dimensionality reduction is an important preprocessing step for many approaches to analysis of hyperspectral data, which is capable of exploring the inherent low dimensional structure, reducing the computational complexity, and improving the performance of data analysis. Manifold learning is proposed for nonlinear dimensionality reduction. It assumes that the original high dimensional data lie on a low dimensional manifold that can characterize the structure and nonlinear properties of the original data. Since hyperspectral data exhibit intrinsic nonlinearities, the commonly used linear feature extraction methods may lose some important nonlinear properties of hyperspectral data, motivating the research of manifold learning nonlinear dimensionality reduction for hyperspectral data analysis. Manifold learning methods are categorized as global and local techniques. This study focuses on the local manifold learning methods for hyperspectral image classification and anomaly detection. The main work is as follows:(1) The traditional manifold learning methods are restrictedly implemented on training data and lack generalization to new data, and therefore the kernel-based out-of-sample extension method proposed by Bengio is employed. The key point of this approach is to find the kernel function of the specific manifold learning method. Our contribution is to derive the kernel function of local tangent space alignment (LTSA) algorithm and achieve its generalization to new data.(2) For hyperspectral image classification, the paper compares multiple manifold learning methods via the classification using k nearest neighbor (kNN) classifier, with the goal of better understanding the capability of manifold learning for classification and the characteristics of hyperspectral data in the manifold domain. Valuable conclusions are achieved using the experiments implemented on several space-based and airborne hyperspectral data sets. The experimental results demonstrate that the nonlinear manifold learning is promising as dimensionality reduction methods. Its greatest advantage is to discriminate difficult classes in two-category classification problems. Moreover, the supervised local manifold learning methods obtain the best performance and can largely improve the classification.(3) Based on the research of manifold learning in conjunction with the kNN classifier, a new supervised local manifold learning weighted kNN classifier (SLML-WkNN) is proposed and applied to hyperpectral image classification. The weight that is calculated by the kernel function of the specific manifold learning method can capture the geometric properties of each neighborhood and provide a meaningful measure of the contributions of neighbors. The new classifier does not involve dimensionality reduction, and thereby is suitable for large data sets. Further, it can mitigate the influence of imbalanced data sets on kNN classifier.(4) Anomaly detection in hyperspectral images has a problem that the background characteristics may be contaminated by anomalies. Therefore, the robust manifold learning that mitigates the influence of anomalies on background manifold is employed to improve the detection performance. For the robust locally linear embedding (LLE) algorithm, the image is divided into sub-images to reduce the computional complexity. However, this approach cannot obtain the global dimensionality reduced data. Therefore, a background training data selection based robust manifold learning method is then proposed, where the background training data is obtained by the recursive hierachical segmentation. This method can both achieve global data manifold and reduce the complexity. Among the local manifold learning methods, robust LTSA has the best performance.更多还原