高光谱图像解混方法的GPU并行设计研究

发布时间：2018-01-23 22:58

本文关键词： 高光谱图像端元提取丰度估计 LSE OSP OVP ATGP UOVP　出处：《大连海事大学》2017年硕士论文　论文类型：学位论文

【摘要】：高光谱遥感数据由于具有空间和光谱的双重信息,愈加广泛地应用在军事、医学、农业以及公共安全等领域。但由于自然界中地物的复杂性以及高光谱图像空间分辨率的限制使得每个像元包含了较多的物质信息,导致大量混合像元的存在,从而增加了数据分析的难度,光谱解混技术可以定量地对地物属性进行描述。端元提取和丰度估计是高光谱解混技术中最重要的两个主题。端元代表图像中纯粹的光谱特征,而丰度可以精确地分析混合像素的比重。在端元提取中,ATGP算法是提取端元的代表算法之一;在丰度估计中,LSE和OSP是最常用的两种方法。但是传统的算法,如ATGP、LSE和OSP的设计思路,通常具有过多的矩阵求逆和乘法运算,使得它们在软件实现时速度慢,在硬件上难以实现。因此,这些算法不能满足许多应用的实时需求,应该寻找一种适合快速处理具有大量数据的遥感图像的算法。丰度估计OVP算法和端元提取UOVP算法,其通过Gram-Schmidt正交化的思想进行解混,不涉及任何矩阵求逆操作,更适合于并行计算。论文对上述几种监督式端丰度估计算法和非监督式端元提取算法进行研究,给出了基于GPU端设计方案,详细工作如下:通过深入研究丰度估计的三种算法(LSE、OSP和OVP)和非监督式端元提取算法(ATGP和UOVP)的设计思想,分别完成了基于GPU并行平台的设计和CPU串行平台的LSE、OSP、OVP、ATGP和UOVP算法的设计,其中OVP算法分为CUDA架构以及OpenMP+CUDA混合架构两种设计模式,并对各种算法的并行效果进行比较和分析。分别在模拟高光谱图像和真实高光谱图像上进行实验,纵向比较了各个算法在GPU并行情况下和CPU串行情况下的时间性能;横向比较了丰度估计三种算法和端元提取两种算法在GPU平台下的时间性能。从而验证OVP-GPU和UOVP-GPU并行设计的有效性,提高了高光谱解混的实时性。理论分析和实验结果表明,GPU并行设计可以很大幅度提高算法的运行速度,能够更好地满足系统对实时性的要求,且丰度估计算法OVP和端元提取UOVP更适合GPU并行设计。
[Abstract]:Hyperspectral remote sensing data have been widely used in military and medical science because of the dual information of space and spectrum. However, due to the complexity of ground objects in nature and the limitation of spatial resolution of hyperspectral images, each pixel contains more material information, resulting in the existence of a large number of mixed pixels. Thus increasing the difficulty of data analysis. Spectral demultiplexing technique can be used to describe the properties of ground objects quantitatively. End-component extraction and abundance estimation are the two most important topics in hyperspectral demultiplexing technology. Endelements represent pure spectral features in images. The abundance can accurately analyze the specific gravity of the mixed pixels. The ATGP algorithm is one of the representative algorithms for the extraction of the end elements. LSE and OSP are the two most commonly used methods in abundance estimation, but the traditional algorithms, such as ATGP LSE and OSP, usually have too many matrix inverse and multiplication operations. These algorithms can not meet the real-time requirements of many applications because they are slow in software implementation and difficult to implement in hardware. We should find a suitable algorithm for fast processing remote sensing images with a large amount of data. The OVP algorithm of abundance estimation and the UOVP algorithm of End-element extraction should be found. It is unmixed by the idea of Gram-Schmidt orthogonalization, and does not involve any matrix inverse operation. It is more suitable for parallel computing. In this paper, several supervised end abundance estimation algorithms and unsupervised end component extraction algorithms are studied, and the design scheme based on GPU is given. The detailed work is as follows: the design ideas of LSEOSP and OVPs and unsupervised End-element extraction algorithms (ATGP and UOVPP) are studied in detail. The design of parallel platform based on GPU and the design of UOVP algorithm based on CPU serial platform are completed respectively. The OVP algorithm is divided into two design patterns: CUDA architecture and OpenMP CUDA hybrid architecture. The parallel effects of various algorithms are compared and analyzed. Experiments are carried out on simulated hyperspectral images and real hyperspectral images. The time performance of each algorithm in GPU parallel case and CPU serial case is compared longitudinally. The time performance of the three algorithms of abundance estimation and End-component extraction on GPU platform is compared in order to verify the effectiveness of OVP-GPU and UOVP-GPU parallel design. The theoretical analysis and experimental results show that the parallel design of GPU can greatly improve the speed of the algorithm, and can better meet the real-time requirements of the system. The abundance estimation algorithm OVP and End-component extraction UOVP are more suitable for GPU concurrent design.
【学位授予单位】：大连海事大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP751

【参考文献】