中文摘要：

目的随着成像光谱仪的发展，高光谱遥感图像的空间分辨率和光谱分辨率越来越大，这给高光谱遥感图像解译处理及应用带来挑战。本文提出一种基于MapReduce模式的分布式混合并行处理模型来加速高光谱解混处理。方法为降低算法计算复杂度，对原串行算法进行并行化设计，并采用行列式分块计算法对原算法进行化简计算；最后在分布式集群环境下，采用Jama和JCuda组件来加速算法执行过程中的矩阵运算操作。结果针对224波段，400×400像素空间分辨率的高光谱图像，采用分布式混合计算模型进行解混处理比原始的处理方法在速度上有近十倍的提高，且算法计算量越大，加速效果越明显。结论本文提出了一种基于MapReduce模式的分布式混合并行处理方法来加速最大单形体体积端元提取算法，加速效果明显；采用分块法求解行列式可以大大降低算法复杂度。该方法对计算任务可并行划分、主机与节点间数据交换量少且计算复杂类算法加速效果明显。

英文摘要：

Objective With the development of imaging spectrometer, the spectral resolution and space resolution have been enhanced dramatically which makes a challenge to hyperspectral unmixing processing. So a new distributedhybrid parallel model has been proposed to accelerate hyperspectral unmixing processing. Method In order to reduce the computational complexity of endmember extraction algorithm, the original serial method has been redesigned for parallel computation and a fast implementation of improved method has been proposed based on partitioned determinant operations. At the same time, the Jama and JCuda components have been used to accelerate the computation in distributed cluster environment. Result The proposed distributed hybrid parallel method plays a large role in accelerating hyperspectral unmixing based on maximum simplex volume algorithm. The improved MapReduce model method is near ten times more rapid than the original method for the hyperspectral image which size is 400 × 400 × 224. And the more computational load, the more speed up. Conclusion In this paper, the proposed distributedhybrid parallel method can increase the hyperspectral unmixing processing speed dramatically. At the same time, the partitioned determinant solving method can reduce the complexity of MSVA algorithm. The experimental resuhs indicate that the proposed method can achieve great speedup to the algorithms which have characters of parallel executive tasks, lower data transmission between main node and sub nodes and massive calculations.