A multiple endmember mixing model to handle spectral variability in hyperspectral unmixing

This paper proposes a novel mixing model that incorporates spectral variability. The proposed approach relies on the following two ingredients: i) a mixed spectrum is modeled as a combination of a few endmember signatures which belong to some endmember bundles (referred to as classes), ii) sparsity is promoted for the selection of both endmember classes and endmember spectra within a given class. This leads to an adaptive and hierarchical description of the endmember spectra. A proximal alternating linearized minimization algorithm is derived to minimize the objective function associated with this model, providing estimates of the bundling coefficients and abundances. Results showed that the proposed method outperformed the existing methods in terms of promoting sparsity and selecting endmember classes within each pixel.