End-member extraction using cone non-negativity constraints

This paper presents a new factorization approach for hyperspectral data based on non-negativity constraints. The method does not assume a one to one correspondence between the pseudo-rank of the data matrix and the number of unique components present. Rather it assumes that the number of unique components is related to the number of extreme points of the cone formed by the data matrix. The cone is represented by singular vectors and a set of linear homogeneous inequality constraints. The extraction of extremes is based on the identification of non-redundant inequalities. The approach is illustrated in an application to an AVIRIS spectral image of the Cuprite mining site.