Research On Hyper-spectral Remote Sensing Image Compression Based On Three-Dimension Characteristic

In the earth-observing remote sensing systems, hyper-spectral remote sensing images are important data sources, which have widespread military and civil applications. Due to their huge magnitude of data, it is necessary to compress hyper-spectral images in applications. As 3-D images, hyper-spectral images are different from general 2-D images. They have both spatial correlations and spectral correlations and in which the spectral correlations are stronger than spatial correlations. Presently there is no mature or standard compression technique for hyper-spectral images. In this paper, the algorithms for hyper-spectral remote sensing image compression are proposed.Firstly, three kinds of general compression algorithms for hyper-spectral images are presented, which are transform compression, vector quantization and prediction compression algorithms. The comparison of compression ratios of several typical lossless compression algorithms of is given. It shows that although the prediction algorithm is superior to transform algorithm in lossless compression, it is only suitable for 2-D images and has limited capability for 3-D de-correlation. It is necessary to design 3-D prediction algorithm according to the characteristics of hyper-spectral images.Secondly, the algorithms of 3-D optimal combination, 3-D local context and 3-D inter-spectral context are proposed, which extend the prediction algorithm from 2-D images to 3-D images. The 3-D optimal combination algorithm has accurate precision by optimal combinative prediction, without regard to the differences of spatial and spectral correlations. But for the time-consuming calculations of spatial and spectral statistical correlations, it does not meet the real-time requirements. While the 3-D local context algorithm has lower inter-spectral precision and therefore cannot reduce the spatial and spectral correlations effectively, this reduces the predictive precision of the whole image. The inter-spectral context algorithm has better predictive precision than the local context algorithm by exploiting the characteristics that 3-D images have much stronger spectral correlation than spatial correlation. It takes the neighboring spectral pixel value as reference value and modifies the reference value by inspecting and predicting its trend of changes. In the three models of inter-spectral prediction (inter-spectral gradients, inter-spectral gains and inter-spectral LOCO-I), the inter-spectral LOCO-I model has the optimal capability of de-correlations. Finally, the 3-D bit plane transform algorithm is proposed to overcome theshortcoming of 1-D bit plane transform for it can only reduce the correlation when the neighboring pixels have similar values. The algorithm calculates the horizontal, vertical and spectral bit plane transform sequentially. As the spectral bit plane transform, the algorithm can be easily realized by hardware. In addition, because the calculation and encoding of the transform matrix of each bit are independent, the algorithm can be realized by parallel computing model, which can improve the calculation efficiency and save the processing time greatly.