Hyperspectral imaging sensors provide high-spectral resolution images about natural phenomena in hundreds of bands. High storage and transmission requirements, computational complexity, and statistical modeling problems motivate the idea of dimension reduction using band selection. The optimal band-selection problem can be formulated as a combinatorial optimization problem where p-bands from a set of n-bands are selected such that some measure of information content is maximized. Potential applications for automated band selection include classifier feature extraction, and band location in sensor design and in programming of reconfigurable sensors. The computational requirements for standard search algorithms to solve the optimal band selection problem are prohibitive. In this paper, we present the use of singular value and rank revealing QR matrix factorizations for band selection. These matrix factorizations can be use to determine the most independent columns of a matrix. The selected columns represent the most independent bands that contain most of the spatial information. It can be shown that under certain circumstances, the bands selected using these matrix factorizations are good approximations to the principal components explaining most of the image spatial variability. The advantage of matrix factorizations over the combinatorial optimization approach is that they take polynomial time and robust and proven numerical routines for their computation and are readily available from many sources. In the paper, we will present results comparing the performance of the algorithms using AVIRIS and LANDSAT imagery. Algorithms are compared in their computational requirements, their capacity to approximate the principal components, and their performance as an automated feature extraction processor in a classification algorithm. Preliminary results show that under certain circumstances selected bands can have over 90% correlation with principal components and classifiers using these algorithms in feature extraction can outperform spectral angle classifiers.

Date Published: 23 August 2000
PDF: 10 pages
Proc. SPIE 4049, Algorithms for Multispectral, Hyperspectral, and Ultraspectral Imagery VI, (23 August 2000); doi: 10.1117/12.410351

Published in SPIE Proceedings Vol. 4049:
Algorithms for Multispectral, Hyperspectral, and Ultraspectral Imagery VI
Sylvia S. Shen; Michael R. Descour, Editor(s)