A CW-SSIM distance measure-based affinity propagation for hyperspectral band selection

Dimensionality reduction is often adopted for hyperspectral imagery in advance to improve the efficiency of following processing, such as classification and identification. Affinity propagation (AP), which is a clustering algorithm, has shown the ability to automatically pick out the representative bands from the hyperspectral imagery. Several distance measures have been proposed to construct the similarity matrix, which is an important issue for AP, but the spatial structural information of the image is not considered. In this paper, a structural approach used to evaluate the image quality, called complex wavelet structural similarity (CW-SSIM) index, is developed to build the similarity between band images. The CW-SSIM index could capture the spatial structural information of compared images. Experiments on the real Kennedy Space Center (KSC) hyperspectral data set has demonstrated the efficacy of the proposed distance criterion for AP.