An information theoretic comparison of projection pursuit and principal component features for classification of Landsat TM imagery of central Colorado

Projection pursuit (PP) and principal component analysis (PCA) projections derived from Landsat Thematic Mapper (TM) imagery of central Colorado were compared. While PCA is a simple subset of the general class of PP algorithms, it cannot distinguish Gaussian from non-Gaussian distributions, since it maximizes projected variance. PP algorithms, which maximize higher-order statistics, can be used to find skew or multi-modal projections in order to reveal underlying class structure. These data projections have greater fidelity to underlying land-cover distributions. On sequestered test data, PP projections improved separation of individual categories from a few percent to as much as 24%. PP performance exceeded that of PCA for all but one of the 14 land-cover categories.