IEEE PIC 2017 keynote speech (1): Reconstruction and analysis of hyperspectral imagery using random

The ever-growing spatial and spectral resolution of hyperspectral imagery has increased interest in dimensionality reduction that takes place on-board the data-acquisition platform. However, traditional dimensionality-reduction algorithms are data dependent, computationally expensive, and, consequently, prohibitive for many resource-constrained airborne or satellite-borne sensor platforms. Random projections offer a means for accomplishing dimensionality reduction simultaneously with data acquisition, such that the sensor projects onto a lower-dimensional subspace chosen at random. The problem of reconstruction from random projections is addressed, considering both compressed sensing as well as an alternative based on principal components. The effect of random projections on the performance of hyperspectral analysis is also investigated, with particular focus on anomaly detection and classification. It is observed that strongly anomalous vectors are likely to be identifiable in the domain of the random projections even at low dimension, while widely separated classes are likely to remain so. Finally, the ability of applying such anomaly or class analysis in the random-projection domain is exploited to improve subsequent reconstruction of the hyperspectral dataset.