Recent developments, based on lattice auto-associative memories, have been proposed as novel and alternative
techniques for endmember determination in hyperspectral imagery. The present paper discusses and compares
three such methods using, as a case study, the generation of vegetation abundance maps by constrained linear
unmixing. The first method uses the canonical min and max autoassociative memories as detectors for lattice
independence between pixel spectra; the second technique scans the image by blocks and selects candidate
spectra that satisfies the strong lattice independence criteria within each block. Both methods give endmembers
which correspond to pixel spectra, are computationally intensive, and the number of final endmembers are
parameter dependent. The third method, based on the columns of the matrices that define the scaled min and
max autoassociative memories, gives an approximation to endmembers that do not always correspond to pixel
spectra; however, these endmembers form a high-dimensional simplex that encloses all pixel spectra. It requires
less computations and always gives a fixed number of endmembers, from which final endmembers can be selected.
Besides a quantification of computational performance, each method is applied to discriminate vegetation in the
Jasper Ridge Biological Preserve geographical area.