【作者】 谭琨；

【导师】 杜培军； RobertS.Chen；

【作者基本信息】 中国矿业大学 ， 摄影测量与遥感， 2010， 博士

【摘要】 随着高光谱遥感数据获取技术的快速发展,高光谱数据处理和分析方法的研究已成为推动其应用发展最为重要的方面。作为统计学习理论最有效的方法,支持向量机(Support Vector Machine, SVM)因其具有小样本学习、高维空间、非线性等优点,在解决高维数据分类、小样本学习、抗噪声影响等方面体现出了明显的优越性,成为高光谱遥感影像分类的一个研究热点。本文从统计学习理论出发,在详细研究SVM理论的基础上,对其从多类分类器设计、核函数构造、组合核函数等方面进行改进,并应用这些改进策略对高光谱遥感数据进行分类。论文主要取得了以下主要研究结论:1)在构建基于SVM的高光谱遥感影像分类技术体系的基础上,分析比较了SVM四种常用核函数的效果,并与现有的几种代表性分类方法进行了比较。结果表明,SVM可以有效地克服样本不足和特征维数高导致的Hughes现象,其训练速度和分类速度均优于径向基神经网络,分类精度高于最小距离分类器、光谱角制图分类器、最大似然分类器和径向基神经网络分类器等常规分类器。在SVM的四种常用核函数中,RBF核函数的分类精度最高。2)通过主成分分析、独立成分分析、最大噪声分离、分组特征提取、导数光谱等进行高光谱遥感数据进行降维与特征提取,作为SVM分类器输入,结果表明SVM分类器的分类精度随着特征维数明显波动,其中主成分分析后提取的特征用于分类通常能够获得最高的精度。基于大量试验,建议对于以分类为目标的高光谱遥感数据降维处理,可以优先选择主成分分析。3)在深入研究多类SVM分类器构建方法的基础上,提出了一个基于分离性测度的二叉树多类SVM分类器,试验结果表明多类SVM的分类精度优于常规分类算法,而相对已有的多分类SVM方法,本文提出的基于分离性测度的二叉树多类SVM的分类精度最高。4)针对SVM分类中核函数构造的重要性,提出了一种再生核Hilbert空间的小波核,构建了小波SVM。小波SVM在应用Coiflet小波核函数的时候能获得最高分类精度,优于光谱角制图分类器、最小距离分类器,及径向基核函数的SVM。5)常用的遥感分类器在组合光谱特征和空间特征进行分类时往往存在局限性,而多核分类器通过对光谱特征和空间特征采用多个核函数进行组合,能够综合空间和结构特征的优点改进分类结果。利用小波变换提取纹理特征,分别对纹理特征和光谱特征采用独立的核函数,构建了多核SVM分类器。试验结果表明,组合光谱和小波纹理信息的多核SVM分类精度高于单核SVM分类器和交叉信息核SVM分类器。将多核SVM与特征提取相结合进行试验,发现当设计多核SVM分类器时,主成分变换后的前四主成分和小波纹理组合作为输入的分类精度最高,是一种快速有效的分类方法。6)基于多核SVM,提出了一种用于高空间分辨率高光谱数据的多核SVM分类模型。该分类方法用数学形态学提取空间结构信息,与光谱信息进行组合实现分类。结果显示融合光谱信息和数学形态学剖面的多核SVM能够得到优于其他方法的分类精度,试验中组合前7主成分分量和数学形态学剖面数据的多核SVM能达到91.0%的分类精度。更多还原

【Abstract】 With the rapid development of the hyperspectral remote sensing data acquisition technology, it has become the most important aspect to process and analyze hyperspectral data with high performance. As the most effective algorithm of the Statistical Learning Theory (SLT), Support Vector Machine (SVM) has the small sample study, high-dimensional space, nonlinear, etc., and becomes the hot issue because of its superiority of the hyperspectral remote sensing image classification. In this paper, SVM theory and its improvement were studied in detail based on the statistical learning concept, and some improvements of SVM algorithms(such as multi-class SVM designed, wavelet Kernel SVM, multiple kernel functions) were successfully applied in hyperspectral remote sensing data classification. The main results are as follows:Four common kernel functions of SVM were analyzed and compared. And we compared SVM with some typical classifiers such as Minimum Distance classifier (MDC), Radial Basis Function Neural Network (RBFNN) classifier, Spectral Angle Mapper(SAM), Maximum Likelihood classifier(MLC). SVM could effectively overcome the Hughes phenomenon with inadequate samples. Overall, it is concluded that the classification speed of SVM are faster than RBFNN, and the accuracy of the SVM classifiers are higher than MDC, RBFNN, SAM, MLC Classifier.By executing the experiments of the feature extraction algorithms including Principal Component Analysis (PCA), Maximum Noise Fraction (MNF), Independent Component Analysis(ICA), feature extraction after correlation coefficient grouping, derivative spectral analysis and so on, it indicates that the SVM model has the fluctuation accuracy with feature dimension and PCA achieves the best accuracy for SVM classification. According to the experiments, it is effective to choose the PCA as the hyperspectral data feature extraction for the classification.According to the SVM theory and the separability measure of hyperspectral data, a novel binary tree multi- class SVM classifier based on separability between classes was put forward. It indicates that the novel binary tree classifier has highest accuracy than the other multi-class SVM classifiers and some traditional classifiers (SAM and MDC).With the study on the SVM theory based on reproducing kernel Hibert Space (RKHS) and wavelet analysis, the wavelet SVM (WSVM) classifier based on wavelet kernel functions was constructed. In the experiments, the WSVM classifier demonstrated more accurate results when it was using Coiflet wavelet Kernel function. Compared with some traditional classifiers (SAM&MDC) and classic kernel (Radial Basis Function kernel) of SVM, it indicates that wavelet kernel SVM classifier is most accurate.Usually, remote sensing image classifiers are limited in terms of the ability to combine spectral features with spatial features. Multiple kernel classifiers, however, are capable to integrate spectral features with spatial or structural features by using multiple kernels for spectral and spatial features and summating them for final outputs. The results show that more accurate classification results can be obtained by integrating the spectral and wavelet texture information, using the multiple kernel SVM classifiers. Moreover, when multiple kernel SVM classifier was being adopted, the highest accuracy can be provided by the combination of the first four principal components from PCA with textural features.Furthermore, a method is proposed for the classification of hyperspectral data with high spatial resolution by using SVM with multiple kernels. In the approach, morphological profile (MP) was used for hyperspectral data classification. The results show that by the integrating the spectral features and MP features, the multiple kernel SVM classifiers obtain more accurate classification results than sole-kernel SVM classifier. Moreover, when the multiple kernel SVM classifier was being used, the combination the first seven principal components derived from Principal Components Analysis (PCA) and MP can provide the highest accuracy (91.0%).更多还原