Lossy Hyperspectral Image Compression Using Improved Classified DCT

Hyperspectral images can obtain spectral data of hundreds of different wavelengths on the same scene. The abundant information in these hyperspectral images can remarkably improve credibility and reliability of data analysis. Therefore, the analysis of hyperspectral image data can effectively helps us identify types of objects on the surface of the earth from the pixel level to sub pixel level. Now, they have beenwidely used in earth resources survey and geological exploration, remote sensing of urban planning and management, disaster and environment pollution monitoring, mapping and other archaeological studies. With the increasing application of hyperspectral image data, how to collect and transfer it efficiently has attracted more and more attention of scholars at home and abroad.Many hyperspectral image compression methods have already been proposed.Because hyperspectral image date is very large, we usually compressed it with a lossy compression method,which can get a high compression ratio, before we transfer it. As hyperspectral image is three-dimensional data, we commonly use a two-dimensional orthogonal wavelet transform in spatial domain, and then in the spectral domain with KLT, DCT or the one dimensional DWT. The strong correlation between bands of hyperspectral image determined that we must consider more about the spectral decorrelation. Classical transform methods have some problems. For example, the complexity of KLT is too high, even though it is an optimal algorithm. DCT and DWT have good complexity performance, but the reconstruction performance is not very well.This paper proposes a three steps clustering(TSC) methods, performing on spectral vectors, used in a transform scheme for hyperspectral images(HSIs). The spectral vectors come from 9/7 wavelet transform in spatial domain. The clustering algorithm classifies these vectors based on their statistics with 3 steps to ensure the classifying efficiency. Then a 1D-DCT is adopted to deal with the residual of each class whose performance can be as good as KLT. Thus, both the spatial and spectral data are decorrelated. Since the transformed coefficients are not symmetrical, the asymmetrical 3D-SPIHT is introduced followed by an adaptive entropy coding. The performance is compared with other state-of-the-art transform-based coding algorithms on Airborne Visible/Infrared imaging Spectrometer(AVIRIS) images. Results show that the TSC lossy compression has a better performance.