An Investigation of Spectral Metrics in Hyperspectral Image Preprocessing for Classification

The paper investigates the efficiency of spectral metrics when used in spectral screening of hyperspectral imagery. Spectral screening is the technique of selecting from the data a subset of representative spectra that can be used in further processing. The subset is formed such that any two spectra in it are dissimilar and, for any spectrum in the original image cube, there is a similar spectrum in the subset. The similarity can be described through various spectral distances and can be controlled by a threshold value. The spectral screening can be improved by associating to each spectrum in the subset a weighing factor proportional to the number of spectra in the original image that are similar to it. Following its generation, the screened subset is used in further computations instead of the full data. The resulting processing mappings are then applied to the data. The investigation has focused on the comparison between Spectral Angle (SA), Spectral Correlation Angle (SCA), Spectral Information Divergence (SID), and spectral gradient angle (SGA) in terms of accuracy of the results and speedup obtained. Spectral screening is performed prior to Principal Component Analysis. The PCA result is next extended to the full data. To quantify the accuracy we rely on unsupervised classification of the resulting processed data. Results from experiments on AVIRIS data show that no significant classification accuracy is recorded while the main processing was done on a subset representing only a very small fraction of the original data size.