Multiple data compression
To answer the needs related to the transmission, processing and archiving of multispectral images for the next generation of SPOT satellites, this project aims at elaborating efficient lossless and quasi-lossless data compression algorithms, optimally exploiting the redundancy between spectral bands.
A comparative litterature review to identify a reference reversible compression technique for monospectral data has shown that:
- optimal predictors combined with an arithmetic coder with contextual modelling yield the highest mean compression ratios (2.74% for SPOT-2/3 images, 2.11% for simulated SPOT-5 data);
- the MAP predictor combined with a Rice coder leads to mean compression ratios of 2.38% for SPOT-2/3 images and 2% for simulated SPOT-5 data. These ratios comply with SPOT data payload constraints;
- wavelet-based schemes are applicable only if their hierarchical structure is required by some block in the image processing chain.
Modeling the intrinsic multispectral properties of SPOT data at the level of compression schemes in lossless and quasi-lossless contexts has then resulted in elaborating a genuinely 3D approach in which three 3D predictors, allowing to account for spectral correlations, are combined with a Rice coder. Performance assessment over a SPOT 2/3 database has shown a significant increase of compression ratios w.r.t. the reference monospectral technique (+18% for SPOT 2/3 images, +13% for simulated SPOT-5 data). When introducing the spectral information contained in the Panchromatic band, these ratios are increased by 10%.
This technique has been extended to quincunx grids, yielding a minimal increase of +16% of compression ratios. Finally, it has been further extended to a quasi-lossless context for which a performance assessment study based on subjective image quality metrics has been carried out.
Keywords: Compression, Modeling