高光谱影像混合像元分解及亚象元定位

发布时间：2018-05-16 05:16

本文选题：高光谱图像 + 混合像元分解　；参考：《长安大学》2013年硕士论文

【摘要】：随着高分辨率遥感卫星的发射，海量遥感影像获得，但是实时分析处理能力欠缺，其中最欠缺的是各种有效的算法。对于高光谱遥感影像，急需要解决的一个问题便是混合像元分解，其直接制约了影像的实际应用。但仅仅解决混合像元问题是不够的，其只能获得端元丰度图，不能确定亚象元的空间位置分布，因此还需要解决亚象元定位问题。至此，高光谱影像才能真正得到普遍应用。本文阐述了高光谱遥感的基本概念；研究了高光谱图像的特点；总结了现有的混合像元分解技术，并着重分析研究了几种常见的端元提取算法；同时也总结了现有的亚象元定位技术，并用程序实现了一种亚象元定位算法。最后，通过总结研究现有的高光谱混合像元分解技术，提出了基于顶点成分分析的端元优化算法。顶点成分分析算法（VCA）的本质是一种纯数学方法，具有良好的理论基础，取得了良好效果。但是VCA算法具有三方面缺陷：没有考虑图像空间信息，对于噪声较大的高光谱图像其有效性可能会降低；算法需要预先确定端元数目，但是预先确定正确的端元数目很困难；VCA算法多次运行结果不稳定。针对以上问题，本文提出改进VCA的算法(Improve-VCA)，其指定候选端元数，用候选端元区间的迭代计算、结合图像空间信息以及病态矩阵规避的优化机制，实现了VCA算法的改进。为定量评价算法，充分印证本文算法思想的正确性与有效性，模拟生成了高光谱数据，对常用的端元提取算法(N-FINDR、SGA、VCA、ACEEHIIU)及本文算法(Improve-VCA)进行同条件对比实验与检验，并进行严格的定量分析和说明。定量研究指标采用平均光谱角mSAD、平均光谱信息散度mSID、组分平均夹角mAAD以及丰度反演得到的组分总体均方根误差mARMSE进行综合评价和分析。通过对比分析可知，本文算法能够自动确定端元数目，准确提取端元，在很多方面可以和常用的端元提取算法相媲美，甚至在某些方面更优于常用的端元提取算法。最后，本文用代码实现了一种基于正则MAP模型的高光谱影像亚象元定位算法，对亚象元定位进行了初探，为以后的研究奠定了基础。
[Abstract]:With the launch of high-resolution remote sensing satellite, massive remote sensing images are obtained, but the ability of real-time analysis and processing is lacking, among which various effective algorithms are the most deficient. For hyperspectral remote sensing images, a problem that needs to be solved urgently is mixed pixel decomposition, which directly restricts the practical application of images. However, it is not enough to solve the mixed pixel problem. It can only obtain the endmember abundance graph and can not determine the spatial distribution of the sub-pixel, so it is necessary to solve the sub-pixel localization problem. At this point, hyperspectral images can really be widely used. This paper describes the basic concept of hyperspectral remote sensing, studies the characteristics of hyperspectral images, summarizes the existing mixed pixel decomposition techniques, and focuses on the analysis of several common End-element extraction algorithms. At the same time, the existing sub-pixel localization technology is summarized, and a sub-pixel localization algorithm is implemented by program. Finally, by summarizing and studying the existing hyperspectral mixed pixel decomposition techniques, an end element optimization algorithm based on vertex component analysis is proposed. The essence of Vertex component Analysis (VCA) is a pure mathematical method with good theoretical foundation and good results. However, the VCA algorithm has three defects: it does not consider the spatial information of the image, and its validity may be reduced for the noisy hyperspectral image, and the algorithm needs to determine the number of endpoints in advance. However, it is difficult to determine the correct number of endpoints in advance. In order to solve the above problems, this paper proposes an improved VCA algorithm, which specifies the number of candidate endpoints, and implements the improvement of VCA algorithm by the iterative calculation of candidate endmember interval, the combination of image spatial information and the optimization mechanism of ill-conditioned matrix evasion. In order to evaluate the algorithm quantitatively and fully verify the correctness and validity of the algorithm in this paper, the hyperspectral data are generated by simulation. Experiments and tests are carried out on the same conditions for the common endmember extraction algorithm (N-FINDRN SGASGAACEEEHIIUU) and the improved prove-VCAA algorithm in this paper. And carries on the strict quantitative analysis and the explanation. The quantitative study indexes were evaluated and analyzed by means of average spectral angle mSAD, average spectral information divergence mSID, component mean inclusion angle mAAD and mARMSE of root mean square error obtained by abundance inversion. Through comparison and analysis, we can see that the algorithm can automatically determine the number of end elements, accurately extract the end elements, in many ways can be compared with the commonly used end element extraction algorithm, and even better than the common end element extraction algorithm in some aspects. Finally, this paper implements a hyperspectral image sub-pixel localization algorithm based on canonical MAP model, and makes a preliminary study of sub-pixel localization, which lays a foundation for future research.
【学位授予单位】：长安大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP751;P237

【参考文献】