Research On Hyperspectral Imagery Compression Based On Online Learning Dictionary

Hyperspectral remote sensing technology has been progress in recent years. Spectral channels of imaging spectrometer have increased. The spatial and spectral resolution have become more and more high, which lead to bigger data size. A number of factors results in a large increase in hyperspectral image data. Although the progress of data processing capacity by computer raises efficiency of hyperspectral image transmission and storage, the mass data of hyperspectral image brings pressure to the transmission and storage, which restricts the development of hyperspectral remote sensing technology in some degree. So hyperspectral image compression technology plays a crucial role for application of hyperspectral remote sensing data.First, hyperspectral image was introduced briefly. The necessity and feasibility of hyperspectral image compression are analyzed by the description of hyperspectral data characteristics. After analyzing the hyperspectral data from multiple perspective, we could make a conclusion that hyperspectral image has spectral correlation which is different from common natural image and the spectral correlation is stronger than the spatial correlation.Second, decorrelation method of hyperspectral image data was studied. The traditional transform methods were applied in 3D transform of hyperspectral data cube to compare their ability of decorrelation. After that, a method that combines spectral curve clustering and principal component analysis was proposed. The experiments proved that the method could remove the data correlation betterThird, a new algorithm was designed for spectral curve clustering problem to get better clustering result. A new similarity function of spectral curve was given to fit the hyperspectral data. The new compression method that consisted of this new clustering algorithm and principal component analysis was designed and its better performance was validated by experiments on AVIRIS data.Last, spectral dictionary that learned in sparse coding mode could be used to represent the corresponding material. Because a better compression performance could be reached when a dictionary learned in specific data set was used to reconstructing data. From the perspective of sparse coding, learning a sparse dictionary could achieve a better result of data decorrelation. In order to compress the hyperspectral data, an online learning sparse coding dictionary which could describe the characteristic of spectral curve was created to represent and reconstruct hyperspectral data. This method outperforms other traditional lossy compression.