Projection pursuits for dimensionality reduction of hyperspectral signals in target recognition applications

Three dimensionality reduction methods, all based on parametric projection pursuits (PPP), are investigated for hyperspectral signatures consisting of 1000's of bands. These methods are the parallel parametric projection pursuits (PPPP), projection pursuits best band selection (PPBBS), and sequential parametric projection pursuits (SPPP). Two performance metrics are investigated for maximization during the design phase of the PPP-based methods; these include the Bhattacharyya distance (BD) and receiver operating characteristics (ROC) curves. The three PPP-based methods and the two performance metrics are compared and tested on hyperspectral signatures consisting of approximately 2000 bands. In each case, the PPP-based methods are followed by feature extraction (selecting an optimum subset of projected spectral bands) and classification. The signatures are from an precision agricultural application, where the goal is to distinguish two weeds (sicklepod and cocklebur) commonly found in soybean and cotton crops. The classification accuracies of both the PPPP and PPBBS methods varied significantly depending on the type of classifier utilized. And the PPBBS method produced results with no improvement over classification that had no PPP-based preprocessing. The SPPP method was optimum, producing accuracies >95%, with the ROC metric producing marginally better results than the BD.