Study And GPU Parallel Implementation Of The Hyperspectral Image Compression Method Based On Interband And Calibration Correlations

Hyperspectral remote sensing is a remote sensing system that collects spectra, which is divided into a number of very narrow and continuous bands from ultraviolet to mid-infrared band range. And it plays an important role in military reconnaissance and national economy. Hyperspectral remote sensing has hundreds of spectral channels. Increasing spectral and spatial resolutions make the data volume rapid expansion, which puts great pressure on transmission and storage of data. So there needs to compress hyperspectral images using effective compression technologies. At the same time, in order to meet the real-time of the proposed algorithm, parallel computation is achieved on CPU/GPU hardware platform, which provides support for theoretical research of the hyperspectral image compression method and efficient coding of the compression model.Aimed at the spatial, spectral and correction correlations of hyperspectral images, this paper explores effective compression algorithms such as interband gradient adaptive prediction, three-stage prediction, recursive least squares filter with constant coefficients and the GPU parallel implementation of the three-stage hyperspectral image compression algorithm. The main contests are summarized as follows.(1) The spatial, spectral and correction statistical characteristics in hyperspectral images are analyzed, which establishes the basis of designing an effective compression algorithm of hyperspectral images.(2) An interband gradient adaptive prediction(IGAP) algorithm is proposed combined with the interband structure characteristics of hyperspectral images, the linear prediction(LP) algorithm and the gradient adaptive prediction(GAP) algorithm. Firstly, the horizontal and vertical local gradients between two neighboring prediction bands are estimated. Then using the gradient value and related thresholds to verify that the local boundary and corresponding degree of strength exists, and prediction is carried out by dynamically adjusting functions according to the computation results. Experimental results show that the proposed IGAP algorithm has better comprehensive performance.(3) On the basis of third-order predictor and backward pixel search technology(IP3-BPS), a compression third-order predictor algorithm using three-stage prediction with adaptive predictor reordering was proposed to overcome the calibration-induced data correlation of hyperspectral images. Firstly, hyperspectral images are divided into groups adaptively according to the correlation factor between adjacent bands. Then using the calibration-induced data correlation and the band scaling factor, a recursive bidirectional pixel search method and an adaptive band grouping method are proposed, respectively, for these groups with spectral correlation factor more than 0.9. The proposed algorithm takes the recursive bidirectional pixel search and the backward pixel search as the last two predictors, and adjusts adaptively their orders to achieve better prediction values. Finally, in order to further optimize computation complexity, a computing method of adaptive search threshold is proposed using the original image data and its one-stage prediction value, and the backward pixel search is made by substituting the adaptive search threshold for the band scaling factor. Experimental results show that the IP3-PS2-APR not only decreases the computation complexity, but also has better compression effect.(4) A compression method of hyperspectral imagery via recursive least squares filter with constant coefficients(RLS-CC) is proposed according to the converge property of K coefficient in recursive least squares filter. Firstly, an increasing series with step length 50 is adopted. On the basis of this, the optimize number of prediction bands can be determined using statistical techniques. Then, using the K coefficient character, which converges to zero as pixels get more, constant coefficients are given, and the recursive least squares filter compression is performed using these coefficients. Experimental results show that compared to the RLS, the RLS-CC obviously improves compression effect and efficiency.(5) Aimed at high computation complexity but parallel properties of the three-stage compression algorithm of hyperspectral image(IP3-PS2-APR), an IP3-BS2-APR algorithm based on GPU parallel computing is proposed. By analyze the computation complexity of each prediction stage, the computation bottlenecks such as the three order interband prediction and the adaptive search threshold backward pixel search are found. The three order interband prediction mainly includes matrix inversion and matrix multiplication, so different kernel functions are specified to decrease the computation complexity of this part. For the adaptive search threshold backward pixel search, the main idea is that each pixel is searched separately according to the rules of the BPS-AST algorithm. Experimental results show that when the GPU with single-core is adopted to compress the hyperspectral image, the parallel IP3-PS2-APR can achieve about 35× speedup compared with the corresponding serial code, while the parallel IP3-PS2-APR runs 64 times faster on GPU with dual-core than CPU.