Title : Spatially Adaptive Constrained Non Negative Matrix Factorization for Hyperspectral Unmixing
Place : Celis 010, UPRM
Time : 11:00 AM on June 19, 2014
The use of a single endmember spectra, to represent an endmember class does not take into account the variability of spectral signatures caused by natural factors. For instance, in a forest scene, the spectral signature of a particular tree species may vary due to minerals in the soil or by water content. A single spectral signature can, by itself, provide suitable accuracies in some relatively homogeneous environments. This research proposes an approach to perform unsupervised unmixing where endmember classes, composed of multiple spectral signatures, are used to describe the “endmember” across a large scene. Local endmember spectral signatures are extracted from spectrally homogeneous regions. Spectrally homogeneous regions are identified using a quadtree region partitioning method. Once the regions are identified, local endmembers can be extracted using any of the available endmember extraction methods. Endmember classes are built by clustering local spectral endmember signatures to build up a more precise description of the landscape under study. The cNMF is used at individual image tiles to perform endmember extraction. We call the proposed approach the spectrally adaptive constrained NMF or sacNMF.