Lossy compression of ultraspectral images: integrating preprocessing and compression stages

Abstract

In this paper we focus on lossy compression of Atmospheric Infrared Sounder images that include around 40 MB of data distributed over more than two thousand bands. We present a novel architecture that integrates both preprocessing and compression stages providing efficient lossy compression. As part of preprocessing the spectral bands are normalized and reordered such that the bands of the transformed cube are spatially segmented and scanned to generate a unidimensional signal. This signal is then modeled as an autoregressive process and subjected to linear prediction (LP) for which a valid filter order is obtained by analyzing the prediction gain of the filter. The outcomes of this procedure are LP coefficients and an error signal that, when encoded, results in a compressed version of the original image. Performance of this novel architecture is mathematically justified by means of rate-distortion analysis and compared against other well-known compression techniques.