A comparative analysis of classic Geometrical methods and sparse regression methods for linearly unmixing hyperspectral image data
This research report presents an across-the-board comparative analysis on algorithms for linearly unmixing hyperspectral image data cubes. Convex geometry based endmember extraction algorithms (EEAs) such as the pixel purity index (PPI) algorithm and N-FINDR have been commonly used to derive the material spectral signatures called endmembers from the hyperspectral images. The estimation of their corresponding fractional abundances is done by solving the related inverse problem in a least squares sense. Semi-supervised sparse regression algorithms such as orthogonal matching pursuit (OMP) and sparse unmixing algorithm via variable splitting and augmented Lagrangian (SUnSAL) bypass the endmember extraction process by employing widely available spectral libraries a priori, automatically returning the fractional abundances and sparsity estimates. The main contribution of this work is to serve as a rich resource on hyperspectral image unmixing, providing end-to-end evaluation of a wide variety of algorithms using di erent arti cial data sets.
A research report submitted to the Faculty of Science, University of Witwatersrand, Johannesburg, in the ful lment of the requirements for the degree of Masters of Science by Coursework and Research Report, 2019