Unsupervised Hyperspectral Band Selection By Sequentially Clustering A Mahalanobis-Based Dissimilarity Of Spectrally Variable Endmembers

This paper introduces an unsupervised hyperspectral band selection approach based on sequential clustering of a dissimilarity measure between spectrally variable endmembers. The proposed method combines two main techniques. It begins by automatically extracting spectrally variable endmembers, from the considered hyperspectral image, by means of a linear spectral unmixing technique that takes into account the spectral variability phenomenon. Then, dissimilarity values vector, based on Mahalanobis distances that are calculated between samples of extracted spectrally variable endmembers, is segmented into a fixed number of clusters by using a sequential clustering technique. For each obtained cluster, a single spectral band is chosen. This band corresponds to the wavelength at which a maximum dissimilarity value is found. The proposed method is applied to a real hyperspectral image, and its effectiveness is evaluated by means of a supervised pixel-based classification in comparison with three literature band selection methods. Experimental results show that the proposed approach outperforms tested literature techniques.