An ongoing problem for feature extraction in hyperspectral imagery is that such data consumes large amounts of memory and transmittance bandwidth. In many applications, especially on space based platforms, fast, low power feature extraction algorithms are necessary, but not feasible. To overcome many of the problems due to the large volume of hyperspectral data we have developed a fast, low complexity feature extraction algorithm that is a combination of a fast integer-valued hyperspectral discrete wavelet transform (HSDWT) using a specialized implementation of the Haar basis and an improved implementation of linear spectral unmixing. The Haar wavelet transform implementation involves a simple weighted sum and a weighted difference between pairs of numbers. Features are found by using a small subset of the transform coefficients. More refined spatial and/or spectral identifications can then be made by localized fast inverse Haar transforms using very small numbers of additional coefficients in the spatial or spectral directions. The computational overhead is reduced further since much of the information used for linear spectral unmixing is precomputed and can be stored using a very small amount of additional memory.
[1]
W. Sweldens.
The Lifting Scheme: A Custom - Design Construction of Biorthogonal Wavelets "Industrial Mathematics
,
1996
.
[2]
James F. Scholl,et al.
Higher-dimensional wavelet transforms for hyperspectral data compression and feature recognition
,
2004,
SPIE Optics + Photonics.
[3]
I. Daubechies,et al.
Wavelet Transforms That Map Integers to Integers
,
1998
.
[4]
Michael W. Marcellin,et al.
JPEG2000 - image compression fundamentals, standards and practice
,
2002,
The Kluwer International Series in Engineering and Computer Science.
[5]
Mark S. Schmalz.
Mathematics of Data/Image Coding, Compression, and Encryption Vi, With Applications (Proceedings of Spie, 5208)
,
1998
.