Random projection as dimensionality reduction and its effect on classical target recognition and anomaly detection techniques

There is usually a large amount of redundancy in hyperspectral pixels as they are acquired in hundreds of narrow and continuous spectral bands. Numerous techniques have been proposed to reduce the dimensionality of hyperspectral data in order to improve both computational and memory efficiency. In this paper, we explore the effect of random projection as a dimensionality reduction method on the performance of classical target detection techniques for hyper-spectral images. Specifically, each spectral pixel is projected onto a measurement space with a much smaller dimensionality by a linear transformation represented by a matrix whose entries are randomly generated. The detectors are then applied to the measurement vectors to detect the targets of interests. The detection performances are compared to those obtained from the entire spectrum by the receiver operating characteristics curves. Experimental results demonstrate that only a small number of measurements are necessary to achieve detection performance comparable to that obtained by exploiting the full-spectrum pixels.