基于GPU的高光谱图像混合像元分解并行优化研究

发布时间：2019-05-03 18:55

【摘要】：高光谱遥感由于其较高空间分辨率和光谱分辨率的特点,被广泛应用于地球科学的各个领域。在整个高光谱图像处理流程中,混合像元分解技术是其关键环节和研究热点。但现有混合像元分解算法执行效率低,无法满足大数据量遥感图像的实时处理需求,而GPU/CUDA架构能够为算法提供接近计算机集群的高计算能力,利用GPU高并行处理能力和高存储带宽的优势来提高混合像元分解算法的执行效率是一种有效的研究思路。针对上述科学问题,本文分析了高光谱遥感的成像机理与线性光谱混合模型,在研究并行计算发展现状、GPGPU异构编程模型和基于CUDA架构的并行优化模式的基础上,结合GPU/CUDA架构,针对传统高光谱混合像元分解和稀疏性高光谱混合像元分解进行了并行优化处理。首先,分析了传统高光谱端元提取算法的基本原理,结合算法中对不同像元处理的不相关性,设计了基于GPU并行计算的PPI和N-FINDR端元提取算法。将传统PPI算法中的向量投影问题转换为矩阵相乘进行并行优化,在保证精度的同时,取得了最高百倍的加速比；同时,提出了端元集并发替换方法对传统N-FINDR算法进行优化,也取得了显著的加速比。其次,对基于非负矩阵分解的高光谱混合像元分解方法进行了深入研究,针对其中代表性的约束非负矩阵分解算法,通过线程映射、存储器优化等方式设计其并行优化方法,然后分别利用模拟和实际高光谱数据进行实验测试分析,验证了其有效性。最后,研究了基于GPU的稀疏性高光谱图像混合像元分解的并行优化方法。为了满足算法实时性的要求,针对基于L1/2范数的非负矩阵分解高光谱混合像元分解算法(L1/2NMF)中正则化约束高复杂度的问题,采用合理的任务分配,设计CPU+GPU异构并行计算方法,显著提高了算法处理速度。同时针对一种新稀疏性约束的非负矩阵分解高光谱混合像元分解算法(CSNMF),利用大规模线程并行计算技术,结合算法原理进行了优化设计与实现,并在Telsa C2050平台上进行了实验测试,测试结果表明基于GPU的并行优化方法能为高复杂度高精度的稀疏性高光谱图像混合像元分解技术带来极大的效率提升,为此类算法在实时性要求较高的遥感信息处理中应用带来可能。
[Abstract]:Hyperspectral remote sensing is widely used in various fields of earth science because of its high spatial resolution and spectral resolution. In the whole process of hyperspectral image processing, hybrid pixel decomposition is the key link and research focus. However, the existing hybrid pixel decomposition algorithms are inefficient and can not meet the real-time processing requirements of large amounts of remote sensing images. However, GPU/CUDA architecture can provide the algorithm with high computing power close to the cluster of computers. It is an effective research idea to improve the execution efficiency of hybrid pixel decomposition algorithm by using the advantages of high parallel processing ability and high memory bandwidth of GPU. In this paper, the imaging mechanism and linear spectral hybrid model of hyperspectral remote sensing are analyzed. On the basis of studying the development of parallel computing, GPGPU heterogeneous programming model and parallel optimization model based on CUDA architecture, this paper analyzes the imaging mechanism and linear spectral hybrid model of hyperspectral remote sensing. Combined with GPU/CUDA architecture, parallel optimization is carried out for traditional hyperspectral mixed pixel decomposition and sparse hyperspectral mixed pixel decomposition. Firstly, the basic principle of the traditional hyperspectral end element extraction algorithm is analyzed. Combined with the irrelevance of different pixel processing in the algorithm, the PPI and N-FINDR end element extraction algorithms based on GPU parallel computation are designed. The vector projection problem in the traditional PPI algorithm is transformed into matrix multiplication for parallel optimization. The precision is guaranteed and the acceleration ratio is up to 100 times. At the same time, an end-set concurrent replacement method is proposed to optimize the traditional N-FINDR algorithm, and a remarkable acceleration ratio is also obtained. Secondly, the hyperspectral mixed pixel decomposition method based on non-negative matrix decomposition is deeply studied. Aiming at the representative constrained non-negative matrix decomposition algorithm, the parallel optimization method is designed by thread mapping and memory optimization. Then the simulation and actual hyperspectral data are used to test and analyze the experimental results, and the validity of the proposed method is verified. Finally, the parallel optimization method of sparse hyperspectral image hybrid pixel decomposition based on GPU is studied. In order to meet the real-time requirements of the algorithm, a reasonable task allocation method is adopted to solve the problem of high complexity of regularization constraints in the non-negative matrix decomposition hyperspectral mixed pixel decomposition (L1/2NMF) algorithm based on L _ 1 ~ (2) ~ (2) norm. The CPU GPU heterogeneous parallel computing method is designed to improve the processing speed of the algorithm. At the same time, a new non-negative matrix decomposition hyperspectral mixed pixel decomposition algorithm (CSNMF),) with sparsity constraints is designed and implemented by using the massively threading parallel computing technology and combining with the algorithm principle. The experimental results on the Telsa C2050 platform show that the parallel optimization method based on GPU can greatly improve the efficiency of the high complexity and high precision hybrid pixel decomposition technique for sparse hyperspectral images, and the experimental results show that the parallel optimization method based on GPU can greatly improve the efficiency of the hybrid pixel decomposition technique for sparse and sparse hyperspectral images. It is possible for this algorithm to be used in remote sensing information processing with high real-time requirements.
【学位授予单位】：南京理工大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP751

【参考文献】