Research On Lossless Predictive Compression Technique Of Hyperspectral Images

Hyperspectral images can be acquired by imaging the same landmark at multiple wavelengths, and normally consist of hundreds of bands in the range from visible to infrared. Hyperspectral images contain abundant spectral information due to their nanometer-grade spectral resolution, and they can provide precise details of landmarks. Hyperspectral images have extensive use in areas such as environment monitoring, reconnaissance, resource management, mineral exploration and vegetation analysis. However, the growing scientific and technological demands in spatial and spectral resolutions have drastically increased the data volume of hyperspectral images, which make storing and transferring information much more difficult. Thus, compression technique for hyperspectral images is desired.Compression technique for hyperspectral images can be divided into three categories: lossless compression, near-lossless compression and lossy compression. Hyperspectral images are primarily intended for precise analysis and calculation where the majority of the data is automatically analyzed by computers. It is crucial that any distortion caused by lossy compression is not acceptable. The hyperspectral images must by compressed by a lossless form of compression, for only then can both the data reduction and the original quality of the data be guaranteed. Thus, lossless compression algorithm for hyperspectral images has becoming an active research topic in recent years.Nowadays, there is not a unified standard of lossless compression for hyperspectral images. Thus, new lossless compression algorithms for hyperspectral images are needed to promoted the development of hyperspectral remote sensing for our country. We take the lossless compression technique for hyperspectral images as the principal line in this dissertation, and mainly work on these topics: lossless compression algorithm based on recursive least-squares, lossless compression algorithm based on spectra clustering and band reordering, and lossless compression algorithm based on approximate search.First, the properties of the hyperspectral images have been analyzed by calculating the spatial correlation coefficient and the spectral correlation coefficient. We find that hyperspectral images contain two different kinds of correlations: spatial correlations and spectral correlations, and spectral correlations are much greater than spatial correlations. The analysis ensures that the designed algorithms have better pertinence.Secondly, a new lossless compression algorithm using recursive least-squares predictor is proposed. The algorithm derives the spectral domain predictor by using recursive least-squares filtering, and using different number of prediction bands to predict the current band according to its different spectral correlations. An adaptive prediction bands selection module is proposed to reduce the operation time for determining the optimal prediction bands. Experimental results show that the proposed algorithm obtains very close compression results to the state-of-the-art C-DPCM-APL. At the same time, however, the computation complexity of the proposed algorithm is far less than C-DPCM-APL.Thirdly, aiming at the structural characteristics of hyperspectral images, a new lossless compression algorithm using spectra clustering and band reordering is proposed. On one hand, the proposed algorithm uses K-means clustering method to cluster the spectra into clusters to avoid using unbefitting pixels to predict the current pixel. The K-means clustering method removes part of the spatial redundancy. On the other hand, the proposed algorithm uses optimal band reordering method to improve the spectral correlations of the neighboring bands. The band reordering method removes part of the spectral redundancy. Experimental results show that the spectra clustering method and band reordering method are efficient. The two kinds of preprocessing help the predictive algorithm obtains better compression performance.Finally, we introduce the calibration procedure of the Airborne Visible/Infrared Imaging Spectrometer(AVIRIS) and find that the AVIRIS 1997 hyperspectral images have undergone a radiometric calibration procedure that introduces artificial regularities into the data. To exploit the calibration artifacts, a new lossless compression algorithm using approximate search method is proposed. The proposed algorithm uses a classical predictor for the initial prediction and obtains the final prediction value by selecting the existing pixel values in the current pixel line. Experimental results show that the approximate search method which includes the use of equivalent coefficient is efficient for the AVIRIS 1997 hyperspectral images. The proposed algorithm obtains satisfying comprehensive compression performance and it provides a solution to improve the situation that classical predictive compressors no longer work well as calibrated images for uncalibrated images.