Research On Compression Technology Of Hyperspectral Images

Hyperspectral remote sensing images are acquired by observing the same object(area or target) in multiple narrow wavelength slices at the same time and reveal thereflection, transmission, or radiation features of the observed object in multiple spectralbands. With more information about the observed object contained, hyperspectralremote sensing images have wide potential applications, such as aerospace, mineralexploration, environment monitoring, lunar exploration, etc. Because huge amount ofdata, there are heavy difficulties in storing and transmitting the data of hyperspectralremote sensing images. Therefore, compression of hyperspectral images has becomeone of the most important research tasks in the discipline of signal and informationprocessing. In this thesis, the technologies for compression of hyperspectral images aresystemically and deeply investigated, which are based on Wavelet Transform, NeuralNetwork, Multiscale Geometric Analysis, Independent Component Analysis and Coding,The main work and results are outlined as follows:Based on the analysis of the wavelet transform performance and the charactensticsof hyperspectral images, a new algorithm of lossless compression of hyperspectralimagery is proposed, which combines integer wavelet transform and linear predictiontechnique. The spatial redundancy in the images is removed by the integer-to-interwavelet transform and the inter-band redundancy is removed by the linear predictionaccording to the correlation between two bands. This algorithm is simple in calculationand suitable for parallel processing and hardware realization. The experimental resultsfrom 64 bands hyperspectral images have shown that this algorithm can get highercompression ratio and is of lower computing complexity than the improved predictiontree.The codebook design is one of the key techniques in vector quantization. Based onthe analysis and introduction of vector quantization technique and Self-OrganizingFeature Mapping(SOFM), the learning algorithm of SOFM is improved in order toovercome the insufficiencies of traditional LBG and SOFM learning algorithms. A newalgorithm of lossless compression of hyperspectral imagery is proposed, which is basedon vector quantization and classification prediction technique. At first, the algorithmmakes the vector quantization for spectral dimensions of hyperspectral images. Then,the differences of band vectors and code vectors are used to remove the spatialredundancy and spectral redundancy is removed by classification inter-band prediction. At last, entropy coding is done. The experimental results show that the performance ofthe proposed algorithm outperforms that of the algorithm with LBG codebook.It is very important to preserve detail information for sequential analysis. Afteranalyzing Contourlet transform (CT) and wavelet transform coefficient-basedContourlet transform(WBCT), a new transform named wavelet-based uniformdirectional filter banks(WUDFB) is proposed, which not only overcomes thedisadvantage of 4/3 redundancy with CT, but also reduce the noises of WBCT in thecontour and smooth area. A novel WUDFB-based compression algorithm ofhyperspectral imagery is proposed, in which intra-band WUDFB and inter-bandprediction are used to remove the spatial and spectral redundancy separately, no-list-setpartitioned embedded block coder(NLS) makes output embedded stream. Theexperiment results show that this algorithm is effective and restored images are good inpreserving detail information of original images.Independent component analysis(ICA) is applied to reduce dimensionality ofhyperspectral imagery according to the characteristic of ICA. As the appearing sequenceof IC images produced by ICA is random, a method for selection of IC images ispresented, in which high order statistic is used as measuring rule. A compressionalgorithm of hyperspectral images based on ICA for Hyperspectral image analysis isproposed. At first, hyperspectral features are extracted using ICA and dimensionalityreduction is made in the same time. Then, IC images are compressed by the predictivecode and adaptive arithmetic code. In order to identify the sequential analysis capabilityof the algorithm, the experiments about classification accuracy and preservinginformation for small targets are desiged and accomplished. The experimental resultsshow that the proposed algorithm achieves higher compression ratio, more stronganalysis capability and lower peak signal-to-noise ratio than the dimensionalityreduction based on principal components analysis.The theoretical analysis and algorithms proposed in this thesis have been verifiedwith the 64-band hyperspeetral data provided by 863-308 office and America AVIRIS220-band hyperspectral data. It is also shown that the results have wide potentialapplications to defense engineering.