AbstractEnhancing Pure-Pixel Identification Performance via Preconditioning In this talk, we discuss hyperspectral unmixing in the presence of pure pixels. This problem is equivalent to separable nonnegative matrix factorization (NMF), a recently introduced subclass of NMF problems with applications for example in document classification. This problem can be solved using pure-pixel search algorithms (such as N-FINDR or VCA). However, in noisy settings, the performance of a pure-pixel search algorithm usually depends on the conditioning of the endmember matrix. We describe and analyze three data preconditioning methods for mitigating the aforementioned issue: (1) an approach based on semidefinite programming, (2) pre-whitening, and (3) an approach based on a pure-pixel search algorithm itself. The developments are based on a pure-pixel search algorithm called the successive projection algorithm (SPA). Simulations based on synthetic data sets show that preconditioning makes SPA much more robust against noise. This is joint work with S. Vavasis (U. of Waterloo) and Wing-Kin Ma (The Chinese U. of Hong Kong). |