Hyperspectral sensors produce high-resolution spectral images of a scene simultaneously, which are referred to as hyperspectral imagery or a datacube. Hyperspectral imagery possess much-richer spectral information than multispectral imagery because the number of spectral bands in hyperspectral imagery is in the hundreds instead of a dozen or less. Hyperspectral image analysis has become one of the most-active research areas in remote sensing. The large data volume produced by hyperspectral sensors presents a challenge to traditional data processing techniques. For example, conventional classification methods may not be used without dimensionality reduction (DR) as a preprocessing step, as shown in Fig. 13.1. A number of methods have been developed to mitigate the effects of dimensionality on information extraction from hyperspectral data, such as dimensionality reduction using principal component analysis (PCA), minimum noise fraction (MNF), wavelet dimensionality reduction,and independent component analysis (ICA) methods. Other authors have improved the locally linear embedding (LLE) method by introducing a spatial neighborhood window for hyperspectral dimensionality reduction. The same authors further proposed a nonlinear hyperspectral dimensionality reduction method by combining LLE with Laplacian eigenmaps.
In this chapter, three popular DR methods—PCA, wavelet, and
MNF—and a band-selection method are reviewed and then evaluated and compared in order to demonstrate which method is more robust for a specific application. Experiments are conducted using endmember (EM) extraction as an example of applications. The EMs were extracted from an AVIRIS datacube using the N-FINDR algorithm, which is an EM-extraction method based on the geometry of convex sets to find a unique set of purest pixels in a datacube. Experiments are also performed using mineral detection and the classification of mineral and forest as examples of applications.
Online access to SPIE eBooks is limited to subscribing institutions.