A comparison of unmixing algorithms for hyperspectral imagery

In this paper, we present an experimental comparison of unmixing using the constrained positive matrix factorization (cPMF) with SMACC and MaxD unmixing algorithms that retrieve endmembers from the image pixels. The comparison was made using hyperspectral images collected over Vieques Island in Puerto Rico using the AISA sensor. Based on field work, six information classes were identified in the area of interest and the algorithms are evaluated in their capability to retrieve information about the classes of interest. The cPMF was the only approach capable of identifying all six informational classes with one or more spectral classes assigned to them. SMACC and MaxD were unable to extract one of the classes. The abundance maps from cPMF describe the spatial distribution of the information classes.