Requirements for anomaly detection in hyperspectral data using spectral unmixing

There is an increasing requirement for the detection of small localized spectral anomalies in hyperspectral data. This requirement is becoming equal in importance to the original use of hyperspectral data as a means to build classification maps of the scene. Initially anomaly detection was considered to be only a military application with the detection of man-made objects in an otherwise natural background an obvious example. Lately, several very interesting applications in civilian remote sensing have developed. The use of hyperspectral sensors in the search for diamonds and the detection of exotic plant species are two applications. These civilian applications and several different military applications have an interest in finding spectral anomalies in the data. A procedure for accomplishing this is to determine certain basis spectra called "endmembers" and then unmix the data cube into fractional abundances of each material. A localized spectral anomaly can be identified as high fractional abundances in few pixels. The determination of the endmembers is often done with an analyst, but several new techniques for automating this procedure have been developed. In this paper the effect of common techniques for reducing the size of the data cube, such as principal component or minimum noise transformations, on the ability to detect local spectral anomalies is explored.