Hyperspectral band selection based on the incremental n-dimensional solid spectral angle

An incremental n-dimensional solid spectral angle based band selection method is proposed in this paper, denoted as the INSSA-BS method, for hyperspectral dimension reduction. Inspired by the commonly used spectral angle mapper (SAM) method which measures the difference between two spectra through calculating the cosine value of the angle, we present a n-dimensional solid spectral angle (NSSA) method. The proposed NSSA method is not constrained by the number of input spectra vectors as well as the number of bands. It measures the high dimensional virtual solid spectral angle constituted by a set of interested spectral vectors quantitatively. A higher NSSA value indicates a more dramatic spectral profile difference. According to this character, this paper aims at searching out a suboptimal band set which constitutes the NSSA value as large as possible. Thus, an incrementation coefficient is derived to appraise the contribution of the current band to the total NSSA value, forwardly. The bands which enlarge the NSSA value are selected. Experiment results prove that the proposed INSSA-BS method has the capacity of retaining spectra profile features and enhancing the separability between either inter and intra spectra. Furthermore, as opposite to full bands unmixing, the unmixing errors using bands selected by INSSA-BS method decrease significantly.