Robust Unmixing Algorithms for Hyperspectral Imagery

The linear mixture model (LMM) assumes a hyperspectral pixel spectrum to be a linear combination of endmember spectra corrupted by additive noise. This model is widely used for spectral unmixing mainly because of its simplicity. However, the LMM can be inappropriate in presence of nonlinear effects, endmember variability or outliers. This paper presents a comparison between recent robust hyperspectral unmixing algorithms. The mixture models are first introduced followed by the description of their associated unmixing algorithms. The algorithms are then analyzed when considering a real image acquired over the region of Porton Down in England. The results discuss the behavior of each algorithm to unmix these data and compare their ability to detect the natural or man-made outliers in the scene. The obtained results highlight the potential of the studied mixture models to overcome the current limitations of the LMM.