Parallel Optimization Of Hyperspectral Image Unmixing Algorithm Based On GPU

Hyperspectral remote sensing is a technique for continuous remote sensing of ground objects using narrow and continuous spectral channels.From visible light to short wave infrared band,the spectral resolution is as high as the order of nanometer,and the number of spectral bands is as many as dozens or even hundreds of bands.The existence of a large number of mixed pixels in hyperspectral remote sensing images reduces the accuracy of data analysis.The hyperspectral deconvolution technique decomposes the mixed pixels in the images to obtain the corresponding endelement matrix and abundance matrix.The quantitative analysis of remote sensing data can be realized.Due to the large amount of data in hyperspectral remote sensing images,the efficiency of unmixing algorithm is improved and the fast processing of unmixing algorithm is always a hot topic in the field of unmixing.In recent years,with the rapid development of graphics processor in the field of high-performance computing,how to achieve the parallel optimization of GPU-based hyperspectral remote sensing image deconvolution and improve the efficiency of the algorithm,The realization of hyperspectral remote sensing image processing is very important.In this paper,aiming at the endmember extraction and abundance estimation in hyperspectral unmixing algorithm,the typical ATGP algorithm and OSP algorithm are studied at first.On one hand,the ATGP algorithm in endelement extraction,an improved ATGP algorithm is based on recursive(RATGP).On the other hand,an improved ROSP algorithm is studied for the OSP algorithm in abundance estimation Then,the parallel optimization problem of the algorithm is studied from the characteristics of the algorithm itself and the GPU architecture system The factors are considered such as storage,access,data interaction and task partitioning of the unmixing algorithm.The parallel optimization of RATGP algorithm based on GPU and parallel optimization of ROSP algorithm based on GPU are proposed respectively.The experiment is carried out on two hyperspectral images.First,the time performance of RATGP algorithm and ROSP algorithm under CPU platform and GPU platform is compared.Then,the effects of GPU with different computing power and images with different data sizes on the performance of the algorithm are compared.Finally,the RATGP-GPU algorithm,the ROSP-GPU algorithm,the classical ATGP-GPU algorithm and the OSP-GPU algorithm,which are more suitable for parallel optimization The experimental results on the hyperspectral images show that the proposed optimization method improves the efficiency of the algorithm while keeping the precision of the unmixing results.