A new approach to band clustering and selection for hyperspectral imagery

In this paper a new approach for band selection is introduced based on statistical and geometrical characteristics of band images, where the basic idea is to measure the spread of image data in each band and then clustering the bands in such a way to keep intracluster variance minimum and intercluster variance maximum. For optimal number of bands to be selected, recently developed concept of virtual dimensionality (VD) is used. For endmember extraction, vertex component analysis (VCA) is used. A comparative study is conducted to show the effectiveness of the new approach with other unsupervised band selection methods.