基于波段聚类的高光谱图像波段选择

Band Selection Based on Band Clustering for Hyperspectral Imagery

为使无监督的波段选择能够更好地保留高光谱图像的信息,提出一种基于波段聚类的高光谱图像无监督波段选择方法.首先,计算高光谱图像各波段间的互信息,以此衡量各波段间的相关程度;然后,根据各波段间的互信息,对波段集合进行聚类;通过迭代使得各波段分组自动地聚集在信息量较大且具有代表性的波段周围,直到各聚类中心不再变化,则聚类结束.通过波段聚类过程保证了冗余波段的去除和有用信息的保留,最后,以各聚类中心波段作为所选的波段组合.实验结果证明,与传统方法相比,使用文中的方法选择波段,能够更有效地保留光谱信息,得到更高的分类精度.

In order to preserve the information of hyperspectral imagery more effectively, an unsupervised band selection algorithm for hyperspectral imagery based on band clustering is proposed in this paper. Firstly, mutual information between every two bands is calculated to measure the degree of correlation. Then, clustering within bands is realized by calculating the mutual information. After iterative computation, the bands within the same class are clustered around the most important bands automatically, and the clustering operation stops until the clustering center does not change. As a consequence of clustering, the redundancy bands are removed while the useful information is retained. Finally, the selected band subsets are determined by the clustering centers. Experimental results show that compared with traditional methods, the band set obtained by the proposed method can preserve more spectral information effectively and acquire higher classification accuracy.