Development and testing of improved spectral unmixing techniques

The tradeoff between spatial and spectral resolution gives rise to a finite ground sample size, which produces data with spectrally mixed pixels in a multispectral and hyperspectral systems. Assuming the mixed pixel to be a linear combination of pure spectra known as endmembers, the fractional abundance of the endmembers can be calculated by linear spectral unmixing. Constraints may be placed to force the unmixed fractions to behave realistically. There are also different methods of selecting these endmember spectra, including scene derived and laboratory measured spectra. The stepwise unmixing uses an iterative regression technique to select the optimal endmembers on a per pixel basis. It shows promise of improving the unmixing process by introducing flexibility in endmember selection. However, stepwise procedure has not been rigorously tested, and its parameters are not well characterized. It was the aim of this research to characterize the major parameters of stepwise unmixing, including spectral resolution, library size, F-to-enter/exit, and pixel mixture complexity. In addition to the parametric studies, a comparison between stepwise procedure and hierarchical unmixing was made, as well as applications of unmixing algorithms to non-remote sensing data such as nuclear magnetic resonance (NMR) spectra. The parametric studies were conducted on synthetic data for its advantages of knowing exactly what is in the mixture. The algorithm was also tested with images to confirm the results of the parametric studies. The result showed that the stepwise procedure is capable of results comparable with the best traditional unmixing case