Efficient wavelet-based predictive Slepian-Wolf coding for hyperspectral imagery

• Authors: Ngai-Man Cheung;Caimu Tang;Antonio Ortega;Cauligi S. Raghavendra
• Venue: Signal Processing - Special section: Distributed source coding
• Year: 2006

Hyperspectral imagery is usually highly correlated, in some cases within each spectral band, but in particular across neighboring frequency bands. In this paper, we propose to use distributed source coding (DSC) to exploit this correlation with an eye to a more efficient hardware implementation. The theoretical underpinnings of DSC are laid out in the pioneering work of Slepian and Wolf, and Wyner and Ziv, which provide bounds on the achievable compression when encoding correlated sources with side information available at the decoder. We apply DSC principles to hyperspectral images by encoding individual images (each image representing a spectral band) under the assumption that these bands are correlated. Using DSC tools allows us to operate in "open loop" at the encoder, so that encoding a band does not require having access to decoded versions of (spectrally) neighboring bands. We first compute the parameters of a linear predictor to estimate the current spectral band from a neighboring one, and estimate the correlation between these two bands (after prediction). Then a wavelet transform is applied and a bit-plane representation is used for the resulting wavelet coefficients. We observe that in typical hyperspectral images, bit-planes of same frequency and significance located in neighboring spectral bands are correlated. We exploit this correlation by using low-density parity-check (LDPC)-based Slepian-Wolf codes. The code rates are chosen based on the estimated correlation. We demonstrate that set partitioning of wavelet coefficients, such as that introduced in the popular SPIHT algorithm, can be combined with our proposed DSC techniques so that coefficient significance information is sent independently for all spectral bands, while sign and refinement bits can be coded using DSC. Our proposed scheme is appealing for hardware implementation as it is easy to parallelize and has modest memory requirements. In addition to these implementation advantages, our scheme can achieve competitive coding performance. Our results for high-correlation spectral bands from the NASA AVIRIS dataset show, at medium to high reconstructed qualities, gains of up to 5dB as compared to encoding the spectral bands independently using SPIHT. Our proposed techniques are also competitive compared to 3D wavelet coding methods, where filtering is applied spatially within each spectral band, as well as across spectral bands.