Hyper-DEMIX: Blind Source Separation of Hyperspectral Images Using Local ML Estimates

We propose a new method to unmix hyperspectral images. Our method exploits the structure of the material abundance maps by assuming that in some regions of the spatial dimension, only one material is present. Such regions provide a local estimate of the endmember spectrum of the corresponding material. Our main contribution is a new clustering algorithm called Hyper-DEMIX to estimate the endmember spectrum of each material based on such local estimates. The abundance map of each material is then recovered with a binary masking technique. Experimental results over noisy hyperspectral images show the effectiveness of the proposed approach.