Research On Hyperspectral Remote Sensing Image Compression Technique

Hyperspectral remote sensing data can be acquired by imaging the same area at multiple wavelengths, and normally consist of hundreds of band-images. It contains abundant spectral information, and has been used extensively in enviroment monitoring, mineral investigation and military reconnaissance. However, the corresponding big data volume causes difficulties in informantion transmission and storage, and therefore compression technique is desired.Based on the spatial and spectral correlations of hyperspectral data, this thesis designs efficient compression algorithms by employing prediction, multiscale transform, sparse representation and compressed sensing. The detailed work is conculded as follows.(1) The spatial and spectral statistical properties have been analysed to ensure that the designed algorithms have better pertinency.(2) A searching model has been designed to find the optimal multiple bands to predict the current band, since the hyperspectral data has many bands. The multiband prediction can improve the accuracy, leading to better compression efficiency. Furthermore, the couple grouping prediction and weighted computation have been proposed to reduce the complexity. The analysis shows the presented compression algorithm has satisfying comprehensive performance.(3) The adaptive transform, named DWT-DFB, has been proposed based on the combination of image structural characteristics, DWT and DFB. First, a binary tree is constructed according to the distriution of Fourier transform coefficients. Secondly, the binary tree is employed to direct the decomposition of DWT and DFB, to be specific, the tree height and node-distribution determine the transform levels of DWT and DFB, respectively. Experimental results show that the DWT-DFB transform has better non-linear approximation performance, and thus can reduce the number of reserved coefficients.(4) A compression algorithm for hyperspectral data, based on dictionary learning and sparse representation, has been presented by exploiting the structural similarity of different bands. First, K-SVD is used for training a dictionary from one band-image. Secondly, the joint-sparse-coding is defined to reduce the required additional information. Lastly, quantization and entropy coding are employed to compress the sparse coefficients. Experimental results show that the proposal has better compression and classification performance than that of3D-SPIHT at low bit rates.(5) An inter-band preditor based on MMSE (Minimum Mean Square Error) has been designed to remove the redundancy of CS hyperspectral remote sensing data. The prediction residuals are encoded using arithmetic coder to reduce the data amount, and to improve the transmission efficiency. Moreover, a CS recovery algorithm has been proposed to compromise the complexity and reconstruction quality. Experimental results show that the inter-band prediction can improve the compression performance of CS hyperspectral data, and our CS recovery algorithm can keep the reconstrucion quality with simplified computation.In conclusion, this thesis has presented seveal compression algorithms for hyperspectral data based on prediction, multiscale transform, sparse representation and CS, and the provided experiments have demonstrated their effectiveness.