JPEG-2000 compression using 3D wavelets and KLT with application to HYDICE data

Abstract

JPEG-2000 is the new image compression standard currently under development by ISOIIEC. Part I of this standard provides a "baseline" compression technology appropriate for grayscale and color imagery. Part II of the standard will provide extensions that allow for more advanced coding options, including the compression of multiple component imagery. Several different multiple component compression techniques are currently being investigated for inclusion in the JPEG-2000 standard. In this paper we apply some of these techniques toward the compression of HYDICE data. Two decorrelation techniques, 3D wavelet and Karhunen-Loeve Transform (KLT), were used along with two quantization techniques, scalar and trellis-coded (TCQ), to encode two HYDICE scenes at five different bit rates (4.0, 2.0, 1 .0, 0.5, 0.25 bits/pixel/band). The chosen decorrelation and quantization techniques span the range from the simplest to the most complex multiple component compression systems being considered for inclusion in JPEG-2000. This paper reports root-mean-square-error (RMSE) and peak signal-to-noise ratio (PSNR) metrics for the compressed data. A companion paper [1] that follows reports on the effects of these compression techniques on exploitation of the HYDICE scenes.

Effects of Lossy Compression of Hyperspectral Imagery

“…This data was collected from an airborne platform at an altitude of 1.5 km; at this altitude, the ground sample distance is approximately 0.8 m. The HYDICE sensor returns data as 12-bit unsigned integers. However, calibration increases the dynamic range and occasionally produces negative values [ 13]; hence, the radiance values are stored as 16-bit signed integers.…”

“…The resulting principal-component images (PC images) are then compressed spatially and independently using the JPEG2000 standard. Other researchers [23,13,24,20] have implemented similar schemes. In this section, we elaborate on the main parts of the scheme, which is diagrammed in Figure 3.…”

“…Also, the energypreserving, energy-packing, and decorrelation properties always hold, and PCA/KLT often provides good compression performance for non-Gaussian data. PCA/KLT forms a crucial part of many HSI compression schemes [13,20,23,24,26], including ours.…”

Feature extraction from hyperspectral images compressed using the JPEG-2000 standard

“…The only studies to date on the exploitability of lossily compressed and reconstructed hyperspectral imagery, however, are quite recent and few in number. The papers by Brower et al [3] and Kasner et al [6] compared different combinations of JPEG-2000 options on hyperspectral imagery in terms of root mean-square error or peak signal to noise ratio. In Shen and Kasner [10], the authors studied anomaly detection and material identification tasks.…”

Hyperspectral Image Coding Using 3D Transform and the Recommendation CCSDS-122-B-1

“…In [8], it is shown that the best option to perform a 3-dimensional DWT is the hybrid transform, which consists of applying first the DWT in the spectral domain followed by a two dimensional squared DWT. In [9,10,11,7], it is demonstrated that, as a spectral decorrelator, PCA provides better performance than DWT (around 5dB). Recent works [12,13] indicate that there exist anomalous pixels, and if they are detected, the performance of the coder can be further improved using the PCA.…”