GPU Parallel Research Of Hyperspectral Image Endmember Extraction With Discrete Particle Swarm Algorithm

Hyperspectral remote sensing is an important means of observing the earth.However,due to the limitation of sensor accuracy,the data detected by the detector often contains multiple objects information.The mixed information forms a mixed image element.The existence of mixed pixels restricts the improvement of classification accuracy and the accuracy of target detection in hyperspectral image processing.One of the most important steps in hyperspectral image unmixing based on a linear spectral hybrid model is endmember extraction.Traditional endmember extraction algorithms such as N-FINDR and VCA cannot adapt to the hyper-spectral images which has lots noise with large amounts of data,and lack information feedback mechanisms.The discrete particle swarm endmember extraction algorithm can solve the problems above,but the discrete particle swarm endmember extraction algorithm still has the disadvantage of low computational efficiency.In recent years,GPUs have been applied to hyperspectral image processing which have shown good performance.Therefore,the GPU-based parallel design could be applied to the discrete particle swarm endmember extraction algorithms.In this paper,a parallel discrete particle swarm optimization for hyperspectral endmember extraction algorithm based on the GPU/CUDA architecture is ddopted.For the DPSO algorithm contains a large number of matrix operations,such as matrix multiplication,matrix inversion,matrix summation and other time-consuming part with computing by CPU,we can use the CULA library,CUBLAS library and write the kernel functions to parallel optimization.Comparing the experimental results of serial version with parallel version on the personal PC,the acceleration ratio can be achieved 2.At the same time,the factors that affect the number of iterative convergence of the algorithm are further studied.Through the actual experimental research,D-PSO with the new parameter optimization strategy that can reach the time requirements and can maintain good extraction accuracy.The G-DPSO experiment based on GPU/CUDA architecture using the Tesla K20 platform shows that the GPU-based parallel optimization method can greatly improve the efficiency of high-spectral image end-member extraction which has high complexity and large amount of data,and can achieve quasi-real-time level.