Continuous wavelets for the improved use of spectral libraries and hyperspectral data

Spectral libraries are commonly established as a means to archive representative signatures of natural materials. Such signatures can then be used to train feature extraction and classification algorithms applied to imagery, for comparison with unlabeled spectra. A number of spectral libraries are publicly available and widely used in the community. Disparities in viewing and illumination measurement configurations between libraries generally preclude the direct comparison of spectra for the same materials. Within libraries, measurements may be reported for varying sample properties, such as grain size in the case of powdered minerals or leaf or canopy structure in the case of vegetation. In such instances, use of the library and the selection of representative spectra to identify an unknown material may require a priori knowledge or an educated guess of the physical properties of the unknown material to conduct the comparison.This study demonstrates that continuous wavelet analysis can provide a new and useful representation of spectral libraries and minimize these disparities amongst libraries. In the context of spectral mixture analysis we suggest that the selection of representative endmember spectra from spectral libraries can be more readily defined in the wavelet domain than using reflectance data. In the context of sensing target compositional variability, for example changes in the chemistry of a given mineral, spectral differences due to distinct sample composition are more readily identified using wavelets. The examples provided in this paper are mainly for powdered mineral spectra because there are a number of widely known public spectral libraries of powdered minerals that have been in common use in the hyperspectral community but the principles apply to a range of natural materials including vegetation.