Research On Lossless Compression Of Hyperspectral Data And CUDA Parallel Acceleration

The hyperspectral data studied in this dissertation specifically refers to hyperspectral remote sensing images and auroral spectral data.Hyperspectral remote sensing images are widely used in agriculture and forestry,military reconnaissance,urban planning,pollution monitoring and other fields.The study of aurora is of great significance to the study of solar activities and their impact on the earth.But the data volume of hyperspectral remote sensing image and auroral spectrum data is very large.For example,the size of a hyperspectral remote sensing image collected by AVIRIS exceeds 148 MB,and with the continuous improvement of spatial resolution and inter-spectral resolution,the size of the hyperspectral remote sensing image is still increasing.And the total amount of aurora spectral data collected each day can reach several GB,which is because that the aurora is dynamically changing and requires the detection equipment to work continuously.The huge amount of data brings great challenges to storage,transmission and processing,so data compression techniques must be used to reduce the amount of data.In addition,because of the huge amount of data,compression is also very time-consuming.For hyperspectral remote sensing images,the huge time overhead introduced by compression is unacceptable in some application scenarios that require real-time performance;for auroral spectral data,the huge time overhead introduced by compression is also unacceptable to realize real-time transmission of aurora spectral data from the polar scientific research stations to the domestic.Therefore,the compression of hyperspectral remote sensing images and auroral spectral data must be accelerated.This dissertation uses GPU to accelerate the compression of hyperspectral remote sensing images and auroral spectral data in parallel.GPU has powerful parallel computing capabilities,which can greatly increase the compression speed without affecting the compression performance.The main research content and characteristics of this dissertation are as follows:1.The C-DPCM algorithm is used for lossless compression of hyperspectral remote sensing images.Its compression performance is tested and compared with several other excellent lossless compression algorithms.The experiments show that the C-DPCM algorithm has excellent compression performance,but the serial compression time of a hyperspectral remote sensing image on the CPU is about 17 to 38 minutes.To improve the compression speed,this dissertation uses GPU to accelerate the C-DPCM algorithm in parallel,and uses shared memory and registers,multi-CUDA stream technique and multiGPU technique to optimize the CUDA implementation of the C-DPCM algorithm.Finally,the compression time of a hyperspectral remote sensing image is shortened to about 2seconds,and the acceleration ratio is close to 700 times.2.The online linear regression prediction algorithm is applied to the lossless compression of auroral spectral data.Improvements are proposed for the prediction coefficients calculation and the residual coding to improve the compression performance of the online linear regression prediction algorithm.In addition,there exist a large amount of repeated calculations in the calculation of prediction coefficients,which not only waste computing resources,but also seriously affect the compression speed.To reduce the amount of calculations and improve the calculation speed,a decomposition computing method for prediction coefficients is proposed.The experiments show that the compression time of an auroral spectral image is about 11 seconds when serially compressed on the CPU.To further improve the compression speed,this dissertation uses GPU to accelerate the compression of aurora spectral images in parallel,and uses shared memory and registers,multi-CUDA stream technique and multi-GPU technique to optimize the parallel CUDA programs,respectively.In the end,the compression time of an auroral spectral image is shortened to about 620 milliseconds,and the speedup ratio is about 18 times.Although the speedup is not high,the parallel compression can still save a lot of time compared to the serial compression due to the large amount of aurora spectral data.3.The deep neural network is applied to aurora spectral data compression,and a lossless compression framework for aurora spectral data based on LSTM is designed.When compressing,the predicted value for each pixel of an aurora spectral image is calculated by the trained neural network firstly,then the predicted image is subtracted from the original auroral spectral image to get the residual image,and finally the residual image is encoded.Experiments show that the LSTM neural network has good compression performance,and the average compression rate is 0.11 bpp lower than the online linear regression prediction algorithm.