note 3.1AISRP workshop '97 presentation,    E. Merényi, U of Arizona, LPL

The following examples illustrate the relative discrimination capabilities of an ANN and Maximum Likelihood (MLH) classifier, and linear mixture modeling (currently one of the most advanced anaylsis tools for hyperspectral imagery).

(Click on image for full size.) Upper left: Color composite of the Lunar Crater Volcanic Field (LCVF), Nevada, from a 1994 AVIRIS image. The labels indicate distinct surface units where training pixels were extracted. Cinder cones (red features, A), young (F) and old (I) basalt flows , iron oxide rich soils (L, W), a small rhyolite outcrop (B), and a narrow strip of tuff exposed along a scarp (G) are among the tertiary volcanic products in the area. The dry Lunar Lake playa contains two large compositionally distiguishable parts (D and E), and several small units (Q, R, S, T) at its northern outflow area. Sediment transport is indicated in the long, downsloping outflow channels with varied water and clay contents (N, O, P, U). The basalt cobbles strewn around the edge of the playa (K) have a characteristic, albeit subtle distinction in their spectra (which is a mix of playa and basalt signature), which allows mapping of this unit.

Upper right: MLH classification map; Lower right: mixture model with playa, cinder, and rhyolite endmembers; Lower left: ANN classification map.

Note that the MLH classifier could not work with the full AVIRIS resolution, because the largest number of training pixels that could be identified reliably was 32. Therefore for the MLH, the image had to be subsampled to 32 bands (from 194, which was the input to the ANN and mixture model after removing bands with atmospheric contamination). The subsampling was based on strategic selection of bands that best characterize the minerals in the scene, but even so the MLH misses several classes completely, and makes severe overestimation for others. Note also that for several of the classes (most notably the rhyolite, B) that are of geologic interest it would be impossible to collect 195 training pixels as they do not contain that many.

Details on classifier evaluations are in Merényi, E. Minor, T. B., Taranik, J. V., and Farrand, W.H., 1996, Quantitative Comparison of Neural Network and Conventional Classifiers for Hyperspectral Imagery, subitted to Remote Sensing of Environment.