Hyperspectral band selection using the N-dimensional Spectral Solid Angle method for the improved discrimination of spectrally similar targets

Abstract

Selecting a subset of bands from hyperspectral data can improve the discrimination of ground targets because the most distinguishing spectral features are utilized. Targets with similar spectra are particularly challenging for band selection. A band selection method using the N-dimensional Solid Spectral Angle (NSSA) was recently proposed by Tian et al. (2016) to select the most dissimilar spectral regions amongst targets, but no case studies have been conducted using data from natural targets and there are currently no guidelines for the parameter selection in the NSSA band selection method. This study uses two spectral datasets of geologic relevance (clay minerals and ultramafic rocks), each with spectrally similar materials, to establish guidelines for the selection of two parameters (k and threshold) that will enable the use of the method for practical applications. K defines the band interval (relates to feature width) from which NSSA is calculated, and the threshold defines the number of bands selected from a profile of NSSA as a function of wavelength. The first guideline consists in constraining the maximum k value based on the spectral dimensionality of the widest significant spectral feature for the materials under study. The second guideline is to use a profile of the NSSA value as a function of wavelength for each permissible k value to capture the primary wavelength regions of high NSSA values. Finally, the threshold parameter for each k is estimated from a graph of the NSSA value as a function of the number of bands. The guidelines on the parameter definition allow non-expert users to select a subset of bands while capturing both narrow and broad discriminating features.Results show that bands selected from the two datasets are in good agreement with known spectral features. Of significance is that the bands encompass a range of distinguishing and often subtle spectral characteristics that include absorption feature position, and shape (asymmetry) and depth, the same that are recognized by experts, and thus can be used to assist experts in identifying key features. Moreover, the bands selected with NSSA show improved class separation as illustrated for datasets of spectrally similar materials. When tested for the discrimination of clay minerals, a competitive method named Variable Number Variable Band Selection (VNVBS) did not provide adequate information for the selection of bands. There are several valuable band selection methods reported in the literature, but few can be applied to datasets encompassing a relatively small number of spectra and to select bands that enable the discrimination of spectrally similar materials. As demonstrated in this study, the NSSA method should be of great value to studies that require feature identification from spectral libraries either resulting from the collection of field spectra or the extraction of endmembers from imagery.