Adaptive clustering of spectral components for band selection in hyperspectral imagery

A novel unsupervised band selection method is proposed, where adaptive clustering of spectral components is employed. For a given hyperspectral image, its spectral bands are grouped into clusters, based on the similarity measured by histogram-determined mutual information and its normalised version. Then, variable numbers of clusters
can be determined automatically in our approach by selecting the most likely clustering boundaries, thus thresholding of image similarity in grouping bands is avoided. Finally, one representative band is extracted from each cluster by minimising the sum of inter-band difference within the band cluster. Using the well-known 92AV3C dataset, the proposed approach is evaluated in terms of efficiency and effectiveness. Experimental results have demonstrated the great potential of our proposed methodology in automatic band selection for many applications of hyperspectral imagery.