Abstract: Aiming at the parallel processing of remote sensing image classification, the existing researches are generally carried out through computer cluster and workstation. These ways have the disadvantage of high cost and are difficult to establish. Only a few researches which are based on GPU mainly intend to demonstrate?the availability of this parallel architecture from the perspective of workflow and pay little attention to the significant storage optimization strategies. Directed?against?the shortages of the existing studies, taking the high performance computing of hyperspectral image classification using the method of spectrum waveform matching and spectral angle mapping based on CUDA for example, this paper places emphasis on researching the optimization strategies of GPU storage and their improvement method. The experimental results show that, the optimization strategies of GPU storage and their improvements are effective, besides, for a variety of images of different sizes and data volume, the efficiency of algorithm has been promoted remarkably compared with the situation before these strategies are applied. At the same time, The hyperspectral image classification?based on CUDA acquires accurate computing?results.

Key words: CUDA, hyperspectral remote sensing image, spectral angle mapping, constant memory, shared memory, merged accessing of memory

摘要： 针对高光谱遥感影像分类的并行化处理，现有研究一般是通过集群和工作站来开展，成本较高，部署困难。少数基于GPU方式的研究主要是从流程的角度来论证该并行架构对提高算法效率的有效性，对于算法关键的存储器优化策略等研究相对较少或不详细。针对现有研究的不足，以CUDA架构下高光谱遥感影像的光谱波形匹配法和光谱角填图法分类的高性能计算为例，对算法存储优化策略进行重点研究，深入探讨了一系列存储优化及其改进方法。通过实验论证分析表明：存储优化策略及其改进方法有效，并且对于多种不同尺寸与数据量的影像，CUDA架构下算法的运行效率都有了较为显著的提升。同时，基于CUDA的高光谱影像分类维护了计算结果的准确性。

关键词: CUDA架构, 高光谱遥感影像, 光谱角填图, 常量存储器, 共享存储器, 存储器合并访问