基于GPU的高光谱遥感主成分分析并行优化

主成分分析（principal component analysis, PCA）是高光谱遥感图像特征提取的重要方法。为了在保证精度的同时，提高高光谱遥感PCA算法的计算效率，文章提出一种基于图形处理器（graphic processing unit，GPU）＋中央处理器（central processing unit，CPU）异构系统的PCA并行优化方法。该方法利用GPU的并行计算能力实现PCA中复杂的协方差矩阵计算与维数缩减过程，优化了像元去均值的计算流程；解决了GPU内核计算像元累加和非合并访问问题；利用共享内存机制，提高了访存效率。此外，该方法采用改进的Jacobi快速迭代法在CPU中进行特征分解，保证了算法的精度。实验结果表明，该方法在保证精度的同时能够有效提高计算效率，在Quadro600平台上的加速比达到141倍，满足了高光谱遥感图像实时应用的需求。

Parallel Optimization for Principal Component Analysis of Hyperspectral Remote Sensing Based on GPU

Principal component analysis （PCA） is an important method for feature extraction of hyperspectral remote sensing image. In order to improve the computational efficiency of PCA, a novel parallel optimization method of PCA is proposed based on GPU＋CPU heterogeneous platform. It takes the advantage of GPU’s parallel computing ability to implement the complex calculation of covariance matrix and dimensionality reduction process of PCA. And it also optimizes the decentralized flow of image data, solves the non-consolidated access of summation on GPU and uses the shared memory mechanisms to improve the efficiency of memory access. Furthermore, the modified Jacobi iterative solution is proposed for eigen-decomposition on CPU to ensure the accuracy of the algorithm. Experimental results show the efficiency of proposed method is achieved, and the maximum speedup is up to 141X at Quadro 600 platform, which can meet the requirement of real-time hyper spectral remote sensing image applications.