Using self-organizing maps for anomaly detection in hyperspectral imagery

Hyperspectral imagery datasets contain tremendous amounts of information. Unfortunately, due to the homogeneity of the Earth´s surface many pixels in such imagery function as background and serve to obscure or hide a desired target. In many cases, the primary reason for collecting the hyperspectral imagery is to find a pixel or two representing statistical scene anomalies. Few commercial applications focus on this essential aspect of hyperspectral image processing. More often, the primary focus of commercial image processing packages is the more abundant pixels to the exclusion of the few anomalous pixels. Toward this end we have developed a system based on Self-Organizing Maps (SOM) that cluster the data then evaluates the relationship of the data to the cluster centers. Those pixels located farthest from the cluster centers are found on the outer surface of the convex hull enclosing the hyperspectral dataset. In addition, these outlying pixels represent anomalies within the dataset and their location in proximity to an individual cluster center may merely be serendipitous. This procedure for locating anomalous pixel is demonstrated in a 1998 AVIRIS scene of the Copper Flat porphyry copper deposit. The approach is applicable to other domains besides geology and mineral exploration.