Hyperspectral image compression using sparse representations and wavelet transform based spectral decorrelation

Being a spectral imaging technique, hyperspectral imagery enables acquisition of electromagnetic spectrum data in hundreds of narrow bands. The chemical composition of objects within the viewing range of hyperspectral sensors may be analyzed with the help hyperspectral imaging techniques. Providing a huge amount of data comes with a heavy cost of size. To that extent, hyperspectral compression methods become essential for transmission and storage purposes. There are two fundamental ways for data compression, namely, lossless and lossy. Lossy methods aims at compressing the data as much as possible while keeping the reconstructed data quality as high as possible. On the other hand, lossless compression techniques, as the name suggests, are targeted towards compressing the data in a lossless manner with the largest possible compression ratio value defined as the ration of uncompressed data size to the size of the compressed data. Lossless compression schemes typically achieve much higher compression ratios as compared with lossy techniques. For hyperspectral compression purposes, data-driven approaches, such as dictionary learning based lossy compression methods, yield better compression performance compared with other state-of-the-art methods. Regarding the lossless compression methods for hyperspectral imagery, integer wavelet transform based techniques are reported to perform better in terms of compression ratio. In this thesis, a novel hyperspectral image compression approach that blends a fully data-driven technique based on sparse representations and an integer wavelet transform based algorithm is proposed. The hyperspectral data is spectrally decorrelated using various integer wavelet filters at different levels. The decorrelated data is then processed according to its frequency content in such a way that highly correlated, sparse high-band block is compressed in a lossless fashion with JPEG2000. On the other hand, a sparse representation is obtained for the low-band block making it possible to represent the lower subband data block using a few number of coefficients. Experimental results indicate that, the proposed hybrid method perfroms better than the existing hyperspectral image compression techniques.