高光谱遥感图像端元提取算法研究与系统实现
发布时间：2019-01-03 16:35
【摘要】：高光谱遥感是近些年来遥感领域发展的一个重要方向，高光谱遥感是通过航空航天飞行器上携带的光谱成像仪对地面进行拍摄，得到一片地面区域的数百个甚至数千个极其狭窄的波段影像。这样，对于成像范围内的任何一个像元，就可以绘制出一条近乎连续的光谱曲线，这可以用来和已有的光谱曲线数据库进行比对，得到地物的类型或成分。 然而，由于光谱成像仪的空间分辨率有限，实际成像中，一个像元对应地面一个区域，而这一个区域有可能是由多种物质混合而成，因此不能对像元进行直接分析，需要进行解混。解混的目的就是计算出组成该混合像元中不同纯物质的组成比例。 要进行解混，就首先要进行纯物质的提取，在高光谱遥感中，纯物质也被叫做端元。端元在不同的遥感图像中一般是不同的，因此有必要通过一种合理的方法将端元从给定的遥感图像中准确的提取出来。 本文分析了当前主流的几种端元提取算法，例如纯像元指数算法，N-FINDR算法，单形体体积法等，并重点研究了纯像元指数算法。纯像元指数算法的运算时间复杂度很高，这大大限制了在航空航天领域需要实时处理的应用。因此本文在对纯像元指数算法的研究中，提出了矩阵乘法的一种优化结构，，经过优化后的结构，具有了并行性，适合在硬件电路上进行实现。 在硬件系统的实现上，本文采用了Xilinx公司的ZYNQ片上系统平台，在单芯片上进行软硬件协同开发，将纯像元指数算法的核心步骤放在了数字逻辑电路上进行，大大提高了运行速度。在设计的过程中，对存储器结构进行了深入的优化，添加了流水处理机制，并使用高级语言综合工具进行设计，使得处理速度得到了明显的提高。经过实际测试，本文实现的硬件结构的运算速度比PC机上的软件提高了400倍以上。与国际上对该算法在硬件电路上实现的最新结果相比，本文的结果也具有非常显著的优势。
[Abstract]:Hyperspectral remote sensing is an important direction in the field of remote sensing in recent years. Hundreds or even thousands of extremely narrow band images of a ground area are obtained. In this way, for any pixel in the imaging range, a nearly continuous spectral curve can be drawn, which can be used to compare with the existing spectral curve database to obtain the type or composition of the ground object. However, because the spatial resolution of the spectral imager is limited, in actual imaging, a pixel corresponds to a ground area, and this area may be composed of a mixture of a variety of substances, so it is not possible to directly analyze the pixel. It needs to be unmixed. The purpose of unmixing is to calculate the composition ratio of different pure matter in the mixed pixel. In order to demix, the extraction of pure substance is the first step. In hyperspectral remote sensing, pure substance is also called endelement. The endelements are usually different in different remote sensing images, so it is necessary to extract the endelements accurately from the given remote sensing images by a reasonable method. In this paper, we analyze several current algorithms for extracting endelements, such as pure pixel exponent algorithm, N-FINDR algorithm, volume method of single body and so on, and focus on pure pixel exponent algorithm. The computational complexity of pure pixel exponent algorithm is very high, which greatly limits the application of real-time processing in the field of aeronautics and astronautics. Therefore, in the study of pure pixel exponent algorithm, an optimized structure of matrix multiplication is proposed. The optimized structure has parallelism and is suitable to be implemented in hardware circuit. In the realization of the hardware system, this paper adopts the ZYNQ system platform of Xilinx Company, and develops the hardware and software on a single chip. The core steps of the pure pixel exponent algorithm are put on the digital logic circuit. The running speed is greatly improved. In the design process, the memory structure is optimized deeply, the pipeline processing mechanism is added, and the advanced language synthesis tool is used to design the memory structure. The processing speed is improved obviously. After practical test, the calculation speed of the hardware structure is more than 400 times faster than that of the software on PC computer. Compared with the latest results in hardware circuits, the results of this paper also have a very significant advantage.
【学位授予单位】：西安电子科技大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP751
本文编号：2399625
[Abstract]:Hyperspectral remote sensing is an important direction in the field of remote sensing in recent years. Hundreds or even thousands of extremely narrow band images of a ground area are obtained. In this way, for any pixel in the imaging range, a nearly continuous spectral curve can be drawn, which can be used to compare with the existing spectral curve database to obtain the type or composition of the ground object. However, because the spatial resolution of the spectral imager is limited, in actual imaging, a pixel corresponds to a ground area, and this area may be composed of a mixture of a variety of substances, so it is not possible to directly analyze the pixel. It needs to be unmixed. The purpose of unmixing is to calculate the composition ratio of different pure matter in the mixed pixel. In order to demix, the extraction of pure substance is the first step. In hyperspectral remote sensing, pure substance is also called endelement. The endelements are usually different in different remote sensing images, so it is necessary to extract the endelements accurately from the given remote sensing images by a reasonable method. In this paper, we analyze several current algorithms for extracting endelements, such as pure pixel exponent algorithm, N-FINDR algorithm, volume method of single body and so on, and focus on pure pixel exponent algorithm. The computational complexity of pure pixel exponent algorithm is very high, which greatly limits the application of real-time processing in the field of aeronautics and astronautics. Therefore, in the study of pure pixel exponent algorithm, an optimized structure of matrix multiplication is proposed. The optimized structure has parallelism and is suitable to be implemented in hardware circuit. In the realization of the hardware system, this paper adopts the ZYNQ system platform of Xilinx Company, and develops the hardware and software on a single chip. The core steps of the pure pixel exponent algorithm are put on the digital logic circuit. The running speed is greatly improved. In the design process, the memory structure is optimized deeply, the pipeline processing mechanism is added, and the advanced language synthesis tool is used to design the memory structure. The processing speed is improved obviously. After practical test, the calculation speed of the hardware structure is more than 400 times faster than that of the software on PC computer. Compared with the latest results in hardware circuits, the results of this paper also have a very significant advantage.
【学位授予单位】：西安电子科技大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP751
【参考文献】
相关期刊论文 前1条
1 耿修瑞;赵永超;周冠华;;一种利用单形体体积自动提取高光谱图像端元的算法[J];自然科学进展;2006年09期
本文编号：2399625
本文链接：https://www.wllwen.com/guanlilunwen/gongchengguanli/2399625.html
