Real-Time Progressive Band Processing for Linear Spectral Unmixing and Endmember Extraction

Hyperspectral imaging systems generate large volumes of data that contain the subtle yet critical knowledge their users need, if they are processed intelligently. Band selection algorithms attempt to separate image bands with high-quality information from those with irrelevant noise. They are a pre-processing filter applied before several other classes of algorithms to avoid the Hughes phenomenon, of diminishing (and even negative) returns caused by supplying those algorithms with too much data. Effective band selection enables real-time and band-progressive image processing algorithms. Real-time algorithms are ones that process image data at an equal or faster rate than it is acquired, analyzing one pixel before the next one arrives from the imaging device. Band-progressive algorithms process entire image subsets at once, then use the results to process a slightly modified subset with small numbers of bands added or removed. This dissertation develops three major technologies. Progressive Band Selection (PBS) is a method to select the most valuable bands of a hyperspectral image according to a changeable, application-specific criterion, and compose a reduced image subset. Real-time linear unmixing algorithms use bands from this subset to determine the abundance of physical materials in each pixel, where the materials are represented by their spectroscopic profile (spectra). This implies that material spectra data is separate from the image data, and the spectra are available before the image is acquired. However, band-progressive endmember extraction algorithms are designed to detect image pixels that represent pure material spectra, negating the need for a separate library. Each technology is described both from a theoretical standpoint and in terms of their practical implementation, and experiments are performed to demonstrate their value relative to the state of the art