Parametric projection pursuit for dimensionality reduction of hyperspectral data

Abstract

This paper addresses the feasibility of utilizing parametric projection pursuit (PPP) for dimensionality reduction of hyperspectral data. With hyperspectral data, the large number of spectral bands can be both a blessing and a curse. It is known that for an accurate classification, the number of required training data increases exponentially with the number of spectral bands. When the data consists of many spectral bands, it is often not feasible to collect enough data for classification purposes, so dimensionality reduction methods, such as PPP, have been proposed for preprocessing hyperspectral data. in the past, PPP has been applied to remotely sensed hyperspectral data with the number of spectral bands being limited to /spl sim/200. In that work, it was shown that PPP can reduce the dimensionality form 100's to 10's, while maintaining class separation for target detection purposes. In this paper, the authors will build on previous research of PPP and investigate the feasibility of applying PPP to dat with 1000's of spectral bands. The data will be collected using an ASD spectroradiometer, Also, another dimensionality reduction algorithm, which is called the Best Spectral Band Combination (BSBC) using Receiver Operating Characteristics (ROC) curves, is used to examine the reliability of the PPP algorithm.