Low complexity dimensionality reduction for hyperspectral images

Abstract

In hyperspectral imaging systems, principal component analysis (PCA), also known as the Karhunen Loeve Transform (KLT) is the conventional way of spectral dimensionality reduction which compacts the image energy into relatively few coefficients to enable compression. The computational burden of the data dependent PCA/KLT often exceeds the capacity of resource constrained hyperspectral sensing platform considering the large size of the hyperspectral image. We propose to use a spectral dimensionality reduction method based on the relationship between KLT and the Discrete Prolate Spheroidal Sequences (DPSS). DPSSs construct a highly efficient basis that captures most of the signal energy while in signal processing the KLT is used to find the filter that maximizes the concentration of the output energy for a given spectrum of the input signal. On the other hand, spatial dimensionality reduction can provide significant amount of reduction as well. The reduction in the spatial domain can be implemented by subsampling. We demonstrate our method's performance on the AVIRIS Hyperspectral data.