Hyperspectral Data Processing: Algorithm Design and Analysis

31.7 Conclusion

When hyperspectral imagery (HSI) was available for data processing in early 1990s, a common approach is to extend existing multispecral imaging techniques in a straightforward manner for processing HSI with a general understanding that hyperspectral imagery is an extension of multispectral imagery by including more spectral bands with better spectral resolutions. Unfortunately, this is generally not true. One main reason is that issues to be resolved in HSI such as subpxiels, mixed pixels, and endmembers are quite different from those in multispectral imagery (MSI) such as land cover/use classification, geographical information system (GIS), etc. The work in Chang (2003a) was developed to design statistical signal processing algorithms for subpixel detection and mixed pixel classification from a viewpoint of HSI. Some of them such as orthogonal subspace projection (OSP) and constrained energy minimization (CEM) have been shown to be promising in LSMA. However, as noted in Ren and Chang (2000a) such hyperspectral imagery-based developed techniques suffer from an issue of intrinsic dimensionality constraint, which does not necessarily guarantee that the same success can also be applied to multispectral imagery. This chapter investigates this issue and further develops two approaches to nonlinear dimensionality expansion to MSI. One is band dimensionality expansion (BDE) which creates new spectral band images resulting from implementing nonlinear functions on the original ...