Hyperspectral classification using selected contourlet subbands

The contourlet transform is a promising multiscale multidirection image representation technique emerging in recent years. Although it has been adopted in some signal and image processing areas, its application to hyperspectral image analysis has not been adequately studied. In this paper, we explore feature selection in the contourlet domain for hyperspectral classification. We apply a previously developed similarity-based unsupervised band selection approach to all contourlet subbands obtained from a nonsubsampled contourlet decomposition of each spectral image. We then utilize the selected contourlet subbands as features for supervised classification experiments. The results show that, using selected contourlet subbands outperforms using all contourlet subbands, all original spectral bands, selected spectral bands, or principal components. We also examine how the transform parameter selections may impact classification accuracy.