Unsupervised Hyperspectral Band Selection Based on Spectral Rhythm Analysis

Remote sensing image classification aims to automatically categorize a monitored area in land cover classes. Hyperspectral images, which provide plenty of spectral information per pixel, allow achieving good accuracy results in classification problems. However, the vast amount of information also can compromise the efficiency due to noisy bands, redundancy, and high-dimensionality. Some dimensionality reduction techniques have been proposed in order to better use the available information. One approach is to perform a band selection, which aims to select the best bands for the classification in order to decrease the dimensionality without degradation of information, i.e., keeping the physical properties acquired by the sensors. This paper introduces a new unsupervised band selection method based on dissimilarity between bands, which are represented by a spectral rhythm, using a bipartite graph matching approach. We carried out experiments in three well known real hyperspectral images datasets. The accuracy results with few bands can achieve levels comparable with the classification made with all data. Our approach can also yield better results in some cases, which is only observed with using supervised approaches in the literature.