Hyperspectral imagery consists of a large number of spectral bands that is typically modeled in a high dimensional spectral space by
exploitation algorithms. This high dimensional space usually causes no inherent problems with simple classification methods that use
Euclidean distance or spectral angle for a metric of class separability. However, classification methods that use quadratic metrics of
separability, such as Mahalanobis distance, in high dimensional space are often unstable, and often require dimension reduction methods
to be effective. Methods that use supervised neural networks or manifold learning methods can be very slow to train. Implementations of
Adaptive Resonance Theory, such as fuzzy ARTMAP and distributed ARTMAP have been successfully applied to single band imagery,
multispectral imagery, and other various low dimensional data sets. They also appear to converge quickly during training. This effort
investigates the behavior of ARTMAP methods on high dimensional hyperspectral imagery and the potential effects of dimension reduction
methods on performance. Realistic-sized scenes are used and the analysis is supported by ground truth knowledge of the scenes. ARTMAP
methods are compared to a back-propagation neural network, as well as simpler Euclidean distance and spectral angle methods.
Keywords: Supervised classification, neural networks, ARTMAP, Euclidean distance, spectral angle, hyperspectral, segmentation,
pattern
analysis.