A Comparative Analysis of Spectral Band Selection Techniques
ROCHESTER INST OF TECH NY
Pagination or Media Count:
The ability to determine optimal spectral band sets for the exploitation of multispectral and hyperspectral imagery is of great concern due to data transfer, storage, and computational constraints. This study compares the performance of three band selection techniques across a range of scenarios and image exploitation algorithms. Thresholded Divergence, a technique based on Gaussian Maximum Likelihood classification, Forward Sequential Band Selection, an iterative method based on target identification algorithms, and Spectral Basis Functions, a method independent of end-exploitation, were selected for evaluation. Each of these band selection techniques was applied to two M7 multispectral images and two HYDICE hyperspectral images. Each selected optimal spectral band set for each image was classified and assessed for classification accuracy. Comparisons between band selection techniques were made based on spectral band subset size, image exploitation algorithm, image and scene type, and input parameter set.