Research On Hyperspectral Image Compression Based On Tensor

The spectrometer of hyperspectral remote sensing imagery can acquire data consisted of hundreds of continuous spectral bands with nano-scale resolution from ultraviolet, visible light to near-infrared and middle infrared. Compared to the traditional remote sensing data, the hyperspectral data have much higher spatial and spectral resolution. They can be used in a lot of applications such as in vegetation survey, ecological environmental survey and geologic studies. With the higher resolution, the problem of increasing data volume becomes an urgent issue. The volumetric data can be difficult to store and transmit. Finding an efficient way to compress the hyperspectral data is inevitable.There are two compression categories for image compression, namely lossless compression and lossy compression. Due to the rigorous demand of hyperspectral imagery compression performance such as the real time and high compression ratio performance, the lossless compression cannot really fulfill the requirements. The lossy compression methods then become the main trend of developing compression schemes. Among the lossy compression, discrete wavelet transform (DWT) based compression scheme has been paid a lot of attention. The DWT can provide high compression ratio and high fidelity, nevertheless its de-correlation performance is not the best. Until now, for the two dimensional cases, principal component analysis (PCA) is the best algorithm to de-correlate, but as the higher dimensional cases, such as hyperspectral data, PCA cannot take good advantage of the data information as an entity.In this thesis, an efficient method for hyperspectral image compression based on tensor decomposition and wavelet packet transform (WPT) is first proposed. The tensor decomposition can reduce the spatial and spectral redundancies and extract the interactive information from different modes of the hyperspectral imagery simultaneously, while wavelet packet transform can provide more flexible and finer decomposition results on the high-frequency components of images. The spectral factor matrix of tensor decomposition is firstly used to de-correlate the spectral dimension in the proposed algorithm; then, with JPEG2000, the more efficient WPT is applied instead of Mallat decomposition to compress the higher-order principal components preserved after spectral de-correlation. To deal with the high computational demand issue of tensor based scheme, we come up with a modified binary search algorithm to simply and efficiently handle the problem. Experimental results on real hyperspectral images demonstrate that the proposed method not only outperforms the conventional 3D wavelet-based algorithms, but also has considerable advantages over the two-dimensional PCA based hyperspectral images compression algorithms on both rate-distortion performance and information preservation.As the uses of hyperspectral images are more common, to specifically deal with the anomalous pixels draws our attention. To maintain these anomalous pixels, we first detect them using anomaly detectors, then remove them for lossless transmission. The original positions of anomalous pixels are interpolated with neighbor pixels so that the original image can be compressed by our tensor based compression scheme.The experimental results of this compression scheme shows great potential on improving the compression performance.With the nonnegativity constraint, the nonnegative tensor decomposition is used to block-compress the whole hyperspectral data. The experiments demonstrate this block based algorithm can reduce the running complexity fiercely, making the proposed algorithm more practical to real applications.