Automated detection of subpixel hyperspectral targets with continuous and discrete wavelet transforms

A major step toward the use of hyperspectral sensors to detect subpixel targets is the ability to detect constituent absorption bands within a pixel's hyperspectral curve. This paper introduces the use of multiresolution analysis, specifically wavelet transforms, for the automated detection of low amplitude and overlapping constituent bands in hyperspectral curves. The wavelet approach is evaluated by incorporating it into an automated statistical classification system, where wavelet coefficients' scalar energies are used as features, linear discriminant analysis is used for feature reduction, and maximum likelihood (ML) decisions are used for classification. The system is tested using the leave-one-out procedure on a database of 1000 HYDICE signals where half contain a subpixel target or additive Gaussian absorption band. Test results show that the continuous and discrete wavelet transforms are extremely powerful tools in the detection of constituent hands, even when the amplitude of the band is only 1% of the amplitude of the background signal.