Parallel Data Compression for Hyperspectral Imagery

Abstract

The high dimensionality of hyperspectral imagery challenges image processing and analysis. It has been shown that hyperspectral compression can be achieved by principal component analysis (PCA) for spectral decorrelation followed by the JPEG2000-based coding. This approach, referred to as PCA+JPEG2000, provides superior ratedistortion performance and can preserve useful data information. However, its main disadvantage is high computational complexity in the PCA process which entails the calculation of the data covariance matrix and its eigenvectors. Parallel processing is an appropriate approach to relieve the computation burden of such a PCA-based compression. In this paper, several parallel PCA implementations are proposed and their processing speed and resulting compression performance are investigated.

A Parallel PCA Neural Network Approach for Feature Extraction

“…Several parallel implementations of traditional PCA have been introduced by Yang et al [6], who have investigated such proposals considering time speed and resulting compression performance. They have showed that parallel implementations using an eigenspace merging approach have lower speed performance than other based on covariance matrix merging.…”

Segmented Principal Component Analysis for Parallel Compression of Hyperspectral Imagery

“…In this letter, we investigate this issue of transform-matrix overhead in the context of parallel compression in which we partition a hyperspectral image spatially or spectrally into multiple subscenes to be distributed to and compressed by multiple independent processing elements (PEs) (e.g., [5]). In this parallel-compression paradigm, the hyperspectral image as a whole is likely to have a very large spatial extent (hence, the motivation for parallel compression in the first place); however, depending on how the data set is partitioned, the individual subscene compressed by a given PE may have a spatial size that is small with respect to the number of spectral bands and, thus, may be detrimentally affected by the transform-matrix overhead.…”

“…Along these lines, we focus on the compression of reflectance data since reflectance-domain compression is preferable to radiance-domain compression for applications requiring atmospheric correction prior to data analysis [7]. That said, the strategies we employ in this letter to Our focus in this section is on parallel implementations of compression in which a data set is partitioned spectrally or spatially in order to distribute data to multiple parallel processing units (e.g., [5]). However, cropping data sets to smaller spatial or spectral sizes can occur in other situations as well-for example, when end users are interested in only a certain part of a larger scene or in only several bands out of a much larger hyperspectral volume, i.e., a region of interest.…”