Unmixing Methods For Hyperspectral Images Based On GPU Parallel Design

Hyperspectral remote sensing data is widely used in military,medical,agricultural and public security fields because of the dual information of space and spectrum.However,because of the complexity of the terrain and the limitation of the spatial resolution of hyperspectral images,each pixel point contains more material information,resulting in the existence of a large number of mixed pixels,thus increasing the difficulty of data analysis.Spectral unmixing technology can be quantified ground property description.Endmember extraction and abundance estimation are the two most important themes in hyperspectral technology.The endmember represents the pure spectral features in the image,and the abundance can accurately analyze the specific gravity of the mixed pixels.In the endmember extraction,ATGP algorithm is one of the representative algorithms of extracting end element;In the abundance estimation,LSE and OSP are the most commonly used two methods.But traditional algorithms such as ATGP,LSE,and OSP,their design ideas usually have too many matrix inversion and multiplication,making them slow on the software,hard to achieve in the hardware.Therefore,these algorithms cannot meet the real-time requirements of many applications,and should find an algorithm for fast processing remote sensing images with large amounts of data.Abundance estimation OVP algorithm and end-extraction UOVP algorithm,which through Gram-Schmidt orthogonalization of the idea,and does not involve any matrix inversion operation,more suitable for parallel computing.In this paper,several supervised abundance estimation algorithm and non-supervised endmember extraction algorithm are studied,and the design scheme based on GPU is given.The detailed work is as follows:through the in-depth study of supervised abundance estimation of three algorithms(LSE,OSP and OVP)and non-supervised endmember extraction algorithm(ATGP and UOVP)design ideas,respectively,with CUDA C completed based on GPU parallel platform and C to complete the CPU serial platform LSE,OSP,OVP and UOVP algorithm design.The OVP algorithm is divided into two types of CUDA architecture and OpenMP+CUDA architecture.And The parallel effects of various algorithms are compared and analyzed.Experiments on simulated hyperspectral image and real hyperspectral image are carried out separately,comparing the various algorithms in the GPU parallel case and CPU serial time trend longitudinally;And conmparing the three algorithms of abundance estimation and the two algorithms of extracting endmember in the GPU platform horizontally.Thereby verifying the performance of OVP-GPU and UOVP-GPU.The real-time performance of hyperspectral unmixing is improved.Theoretical analysis and experimental results show that GPU parallel design can greatly improve the operation speed of the algorithm,and can better meet the requirements of real-time system,the proposed algorithm for abundance estimation OVP and end extraction UOVP are more suitable for parallel design.