Karhunen-Loève Transform based Lossless Hyperspectral Image Compression for Space Applications

The research presented in this thesis is concerned with lossless hyperspectral image compression of satellite imagery using the Integer Karhunen-Loève Transform (KLT). The Integer KLT is addressed because it shows superior performance in decorrelating the spectral component in hyperspectral images compared to other algorithms, such as, Discrete Cosine Transform (DCT), Discrete Wavelet Transform (DWT) as well as the Lossless Multispectral and Hyperspectral Image Compression algorithm proposed by the Consultative Committee for Space Data Systems (CCSDS-MHC). The aim of the research is to develop a reliable low complexity implementation of the computationally intensive Integer KLT, which is suitable for use on board remote sensing satellites. The performance of the algorithm in terms of compression ratio (CR) and execution time was investigated for different levels of clustering and tiling of hyperspectral images using airborne and spaceborne test datasets. It was established that the clustering technique could improve the CR, which is a completely new finding. To speed up the algorithm the Integer KLT was parallelised based on the clustering concept and was implemented using a multi-processor environment. The core part of the Integer KLT algorithm, i.e. the PLUS factorisation, has proven to be the most vulnerable part to single-bit errors that could cause a large loss to the encoded image. An error detection algorithm was proposed which was incorporated in the Integer KLT to overcome that. Based on extensive testing it was shown that it is capable of detecting errors with a sufficiently low error tolerance threshold of 1e-11 featuring a low execution time depending on the extent of clustering and tiling. A new fixed sampling method for the covariance matrix calculation was proposed, which could avoid variation in the data volume of the encoded image that would be beneficial for remote debugging. Analysis of the overhead information generated by the Integer KLT was carried out for the first time and a compaction method which is crucial to clustering and tiling was also suggested. The full range of the proposed enhanced Integer KLT schemes was implemented and evaluated on a desktop computer and two DSP platforms, OMAP-L137 EVM and TMDSEVM6678L EVM in terms of execution time and average power consumption. A new method for estimating the best clustering level, at which the compression ratio is maximised for each tiling level involved, was also proposed. The estimation method could achieve 87.1% accuracy in determining the best clustering level based on a test set of 62 different hyperspectral images. The best average compression ratio, recorded for Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) and Hyperion (spaceborne) images is 3.31 and 2.39, respectively. The fully optimised KLT system, achieving a maximum CR, could compress an AVIRIS image in 3.6 to 8.5 seconds, depending on the tiling level, while a Hyperion image - in less than 1 second on a desktop computer. On the multi-core DSP, an AVIRIS image could be compressed in 18.7 seconds to 1.3 minutes, depending on the tiling level, whereas a Hyperion image - in around 3.4 seconds. On the low power DSP platform OMAP-L137 the compression of an AVIRIS image takes 5.4 minutes and of a Hyperion image - 44 seconds to 2.1 minutes, depending on the tiling level.