空—谱联合高光谱数据降维与分类方法研究

发布时间：2018-03-22 19:48

本文选题：高光谱数据　切入点：降维　出处：《西安电子科技大学》2014年硕士论文　论文类型：学位论文

【摘要】：高光谱影像数据分类是近年来遥感数据解译的热点问题之一。随着高光谱成像技术的发展，所获影像数据的空间分辨率得到显著提高，使得在分类中利用空间信息成为可能。本文在挖掘高光谱数据空间信息的基础上，研究了光谱与空间信息联合使用的策略，设计了几种空谱联合的半监督高光谱数据维数约简与分类方法。具体研究内容如下：（1）设计了一种基于空谱稀疏感知图的高光谱数据分类方法。首先，在高光谱数据的空间一致性假设下，挖掘层次化的空域局部结构，联合光谱数据的非负低秩表示构造出空谱结合的稀疏感知图。利用该空谱稀疏感知图形成Laplacian正则，实现半监督学习，，在极少标记样本的条件下，能够获得准确的分类结果。将所提出的方法进行Indian Pines、Salinas和Pavia University等实际高光谱数据的分类，实验结果表明：该算法通过引入空间信息，能够有效提高分类器的识别率，尤其是在小样本情况。（2）设计了一种基于空谱表征学习的高光谱数据降维方法。针对半监督降维中的正则图规模过大的问题，将数据选择归结为稀疏表示问题，利用表征学习获得“代表性”数据。在线性判别分析中构造稀疏样本的光谱Laplacian正则，与层次化空间Laplacian正则，寻找最优投影方向实现半监督降维。在Indian Pines等实际高光谱数据上进行实验仿真，结果显示：与PCA，LDA，LPP，SDLA降维算法相比，能够更好的提取判别性的特征，在采用SVM等基本分类器时，分类正确率有所提高。（3）设计了一种基于空谱张量稀疏表示的高光谱数据分类算法。张量能够将高光谱数据的空谱信息联合表示，有望对高光谱数据实现更高效的处理。在稀疏表示分类器的基础上，设计了空谱张量稀疏表示分类器。在稀疏表示的过程中，考虑高光谱图像的空间一致性，采用张量稀疏表示快速高效的求解稀疏系数，用于预测高光谱数据的标签。在Indian Pines、Salinas和Pavia University等实际高光谱数据上进行实验仿真，实验结果表明：本章提出的方法能得到很高的地物识别率。本文的工作得到了国家重点基础研究发展计划（973计划）： No.2013CB329402，国家自然科学基金61072108，60601029，60971112，61173090)，新世纪优秀人才项目：NCET-10-0668，高等学校学科创新引智计划（111计划）：No. B0704，教育部博士点基金(20120203110005)，武器装备预研基金项目(9*****)，以及华为创新研究计划项目(IRP-2013-01-09)的资助。
[Abstract]:Classification of hyperspectral image data is one of the hot issues in remote sensing data interpretation in recent years. With the development of hyperspectral imaging technology, the spatial resolution of the obtained image data has been significantly improved. It is possible to use spatial information in classification. Based on mining spatial information of hyperspectral data, this paper studies the strategy of combining spectrum and spatial information. Several spatial spectrum joint dimension reduction and classification methods for semi-supervised hyperspectral data are designed. Firstly, under the assumption of spatial consistency of hyperspectral data, a hierarchical local structure of spatial domain is mined. Based on the non-negative low rank representation of the joint spectral data, a sparse perceptual map is constructed, which is used to form Laplacian regularization, which is used to realize semi-supervised learning under the condition of very few labeled samples. The proposed method can be used to classify the actual hyperspectral data such as Indian Pinesco Salinas and Pavia University. The experimental results show that the algorithm can effectively improve the recognition rate of the classifier by introducing spatial information. Especially in the case of small samples. In this paper, a dimensionality reduction method for hyperspectral data based on space-spectrum representation learning is designed. In order to solve the problem of large scale of regular graph in semi-supervised dimensionality reduction, data selection is reduced to sparse representation problem. Representation learning is used to obtain "representative" data. Spectral Laplacian regularity of sparse samples is constructed in linear discriminant analysis (LDA), and Laplacian canonical in hierarchical space. In order to find the optimal projection direction and realize semi-supervised dimensionality reduction, the experimental results on the actual hyperspectral data such as Indian Pines show that the algorithm can extract discriminant features better than the PCA-LDA-LPP SDLA dimensionality reduction algorithm. When using SVM and other basic classifiers, The classification accuracy has been improved. (3) A hyperspectral data classification algorithm based on sparse representation of space-spectrum Zhang Liang is designed. Zhang Liang can combine the spatial spectrum information of hyperspectral data, which is expected to achieve more efficient processing of hyperspectral data. On the basis of sparse representation classifier, A space-spectrum Zhang Liang sparse representation classifier is designed. Considering the spatial consistency of hyperspectral images in the sparse representation process, Zhang Liang sparse representation is used to solve the sparse coefficients quickly and efficiently. The experimental results show that the method proposed in this chapter can obtain a high recognition rate of ground objects by using the Indian Pinesco Salinas and Pavia University data for the prediction of the labels of hyperspectral data, and the experimental results show that the proposed method can achieve a high recognition rate of ground objects. The work of this paper has been obtained from the State key basic Research and Development Plan No. 973 (No. 2013CB329402), the National Natural Science Foundation of China (61072108C) 6060102910 / 6097111261173090, the New Century Outstanding talents Project: NCET-10-0668, the discipline Innovation Program 111 of China's institutions of higher Learning, and the doctoral Program of the Ministry of Education. It is funded by the weapons and equipment Advanced Research Fund (WARF) and the Huawei Innovation Research Program (IRP-2013-01-09).
【学位授予单位】：西安电子科技大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP751

【共引文献】