Qin Fang-pu, Zhang Ai-wu, Wang Shu-min, Meng Xian-gang, Hu Shao-xing, Sun Wei-dong
Guang Pu Xue Yu Guang Pu Fen Xi. 2015 May;35(5):1357-64.
With the development of remote sensing technology and imaging spectrometer, the resolution of hyperspectral remote sensing image has been continually improved, its vast amount of data not only improves the ability of the remote sensing detection but also brings great difficulties for analyzing and processing at the same time. Band selection of hyperspectral imagery can effectively reduce data redundancy and improve classification accuracy and efficiency. So how to select the optimum band combination from hundreds of bands of hyperspectral images is a key issue. In order to solve these problems, we use spectral clustering algorithm based on graph theory. Firstly, taking of the original hyperspectral image bands as data points to be clustered , mutual information between every two bands is calculated to generate the similarity matrix. Then according to the graph partition theory, spectral decomposition of the non-normalized Laplacian matrix generated by the similarity matrix is used to get the clusters, which the similarity between is small and the similarity within is large. In order to achieve the purpose of dimensionality reduction, the inter-class separability factor of feature types on each band is calculated, which is as the reference index to choose the representative bands in the clusters furthermore. Finally, the support vector machine and minimum distance classification methods are employed to classify the hyperspectral image after band selection. The method in this paper is different from the traditional unsupervised clustering method, we employ spectral clustering algorithm based on graph theory and compute the interclass separability factor based on a priori knowledge to select bands. Comparing with traditional adaptive band selection algorithm and band index based on automatically subspace divided algorithm, the two sets of experiments results show that the overall accuracy of SVM is about 94. 08% and 94. 24% and the overall accuracy of MDC is about 87. 98% and 89. 09%, when the band selection achieves a relatively optimal number of clusters using the method propoesd in this paper. It effectively remains spectral information and improves the classification accuracy.
随着遥感技术和成像光谱仪的发展，高光谱遥感图像的分辨率不断提高，其海量数据在提高遥感探测能力的同时，也给分析处理带来了极大困难。高光谱图像的波段选择能够有效减少数据冗余，提高分类精度和效率。因此，如何从高光谱图像的数百个波段中选择最优波段组合是一个关键问题。为了解决这些问题，我们采用基于图论的光谱聚类算法。首先，将原始高光谱图像波段作为待聚类的数据点，计算每两个波段之间的互信息以生成相似性矩阵。然后根据图划分理论，对相似性矩阵生成的非规范化拉普拉斯矩阵进行光谱分解以得到聚类，类间相似性小而类内相似性大。为了达到降维的目的，计算每个波段上特征类型的类间可分性因子，将其作为进一步在聚类中选择代表性波段的参考指标。最后，采用支持向量机和最小距离分类方法对波段选择后的高光谱图像进行分类。本文方法不同于传统的无监督聚类方法，我们采用基于图论的光谱聚类算法，并基于先验知识计算类间可分性因子来选择波段。与传统自适应波段选择算法和基于自动子空间划分算法的波段指数相比，两组实验结果表明，当采用本文提出的方法使波段选择达到相对最优的聚类数时，支持向量机的总体精度约为94.08%和94.24%，最小距离分类的总体精度约为87.98%和89.09%。它有效地保留了光谱信息，提高了分类精度。
