基于非负矩阵分解的高光谱图像解混技术研究

发布时间：2018-03-21 04:01

本文选题：非负矩阵分解　切入点：高光谱图像解混　出处：《西安电子科技大学》2014年硕士论文　论文类型：学位论文

【摘要】：高光谱遥感成像是一种新型的对地观测技术。成像光谱仪能够记录数十至数百个波段的光谱信息，使得对各种类型地物的准确分类与识别成为可能。然而，因为二维平面的分辨率的约束作用，高光谱图像中的一个像元可能含有几种不同类型的地物，形成了所谓的“混合像元”，直接影响了后续地物的检测与识别。光谱解混技术能够将“混合像元”分解为几种基本类型的地物光谱向量(即端元)与其对应的混合比例(即丰度)的乘积，从而获得亚像元信息，进而提高后续数据应用的效果，最近在遥感领域得到了普遍关注。非负矩阵分解（Nonnegative Matrix Factorization, NMF）将一个非负矩阵分解为两个非负矩阵的乘积，其分解模型与混合像元解混模型相似，非常适于解决光谱解混问题。然而，基于基本NMF算法的高光谱图像解混算法是一个欠定问题。为了得到唯一确定的解混结果，需要加入正则约束消除其欠定性，针对该问题，本论文在研究流形学习的基础上，挖掘高光谱数据的先验信息，设计了多种流形正则下的非负矩阵分解解混技术。具体工作如下： (1)针对现有NMF解混算法仅利用了高光谱数据的光谱信息，忽略了高光谱数据空间信息的缺陷，设计了一种基于空-谱流形正则的NMF高光谱数据解混方法。构造局部窗挖掘高光谱数据的空间信息，设计基于空谱流形正则的半监督NMF解混算法。在人工合成高光谱数据和实际高光谱数据上进行试验，对于S3NMF的各个正则参数的选择、算法的收敛性、对于解混出的端元和丰度进行对比分析、算法的鲁棒性分析和在含有不同端元数的高光谱数据解混结果分析结果表明：本方法在数值指标方面、视觉效果方面和算法的鲁棒性均优于CNMF、GLNMF等方法。 (2)针对高光谱图像特性中的稀疏先验，设计了一种基于稀疏多流形正则的非负矩阵分解高光谱数据解混方法。高光谱图像中的混合像元仅由有限多种地物的光谱向量混合产生，基于这种稀疏性，以及稀疏编码系数相近的像元具有相近丰度向量的假设，构造稀疏流形正则，设计基于稀疏多流形正则的非负矩阵分解高光谱数据解混算法（MMSNMF）。实验中将与S3NMF等方法的高光谱解混结果进行对比，在人工合成高光谱数据和实际高光谱数据上进行试验，结果表明：本方法相比前面的方法解混的在数值和视觉方面的效果有了一定的提高。 (3)针对现有NMF解混算法未充分挖掘图像空间信息的局限，设计了一种基于相似性流形正则的非负矩阵分解高光谱数据线性解混方法。采用邻接权值计算权值的方法比用在K-近邻中用热核函数获得更加准确的像元空间关系信息，针对高光谱图像中光谱之间相似性的特性，引入相似度函数对光谱信息的相似度特性进行描述。构造相似性正则，设计基于相似性流形正则的非负矩阵分解高光谱数据线性解混方法（SMNMF）。实验中将本方法与MMSNMF等方法的高光谱解混结果进行对比，在人工合成高光谱数据和实际高光谱数据上进行试验，结果表明：本方法相比前面方法，在解混的数值和视觉结果方面均有了一定的提高。本文的工作得到了国家重点基础研究发展计划（973计划）： No.2013CB329402，国家自然科学基金61072108，60601029，60971112，61173090)，，新世纪优秀人才项目：NCET-10-0668，高等学校学科创新引智计划（111计划）：No. B0704，教育部博士点基金(20120203110005)，武器装备预研基金项目(9*****)，以及华为创新研究计划项目(IRP-2013-01-09)的资助。
[Abstract]:Hyperspectral imaging is a new technology of earth observation. The spectral information of imaging spectrometer to record tens to hundreds of bands, making the accurate classification and identification of various types of objects possible. However, because of the confinement effect of two-dimensional resolution, high optical spectrum of a pixel in the image may contain several different the type of the object, the formation of the so-called "mixed pixel", directly affects the detection and recognition of objects. The spectral unmixing technique can be "mixed pixel" is divided into several basic types of spectral vector (endmembers) corresponding to the mixing ratio (i.e., abundance) product, so as to obtain sub-pixel information then, to improve the application effect of follow-up data, recently in the field of remote sensing has been widely concerned.
Non negative matrix factorization (Nonnegative Matrix, Factorization, NMF) be a non negative matrix factorization is the product of two non negative matrices, the decomposition model and unmixing model similarity, is very suitable for solving spectral unmixing problem. However, based on the basic NMF algorithm for hyperspectral image unmixing algorithm is an underdetermined in order to get mixed results. The only solution to determine the need to add regular constraints to eliminate the underdetermined, aiming at this problem, this thesis research on manifold learning in hyperspectral data mining, prior information, design a variety of manifold is the non negative matrix factorization unmixing technology. The specific work is as follows:
(1) the existing NMF unmixing algorithm using only the spectral information of hyperspectral data, ignoring the defects of high spectral data of spatial information, design a solution of mixed NMF hyperspectral data in space and spectrum. Based on Manifold Regularization to construct a local window mining spatial information of hyperspectral data, the design of air manifold regularization spectrum semi supervised NMF algorithm based on mixed solution. Experiments were carried out in synthetic hyperspectral data and real hyperspectral data, for each of the regularization parameter selection of S3NMF, the convergence of the algorithm, the endmember and abundance of the mixed solution were analyzed, the robustness of the algorithm and Analysis on hyperspectral data with different end the number of element unmixing results analysis results show that this method in the numerical index, visual effects and robustness of the algorithm are better than those of CNMF, GLNMF and other methods.
(2) according to the sparsity characteristics of hyperspectral image in the design of a non negative matrix sparse Manifold Regularization Based on decomposition of high spectral data unmixing method. Mixed pixels in hyperspectral images is only produced by the spectral vector mixed finite variety of features, which based on sparse, and sparse pixel encoding similar coefficient with similar abundance vector hypothesis, construct sparse Manifold Regularization, design based on non negative matrix sparse Manifold Regularization decomposition of high spectral data unmixing algorithm (MMSNMF). The S3NMF with high spectral unmixing for comparison test in synthetic hyperspectral data and real hyperspectral data. The results show that the method of this method compared to previous mixing solutions in numerical and visual effect are improved.
(3) the existing NMF unmixing algorithm does not fully tap the image spatial information limitations, the design of a non negative matrix similarity canonical manifold decomposition based on linear unmixing method for hyperspectral data. The pixel spatial information method using adjacency weight calculation weights is more accurate than that used in the K- function was used in thermonuclear neighbor the in between spectral hyperspectral image similarity features, the similarity characteristics of the spectral information into the similarity function described. Structural similarity design based on regular, non negative matrix similarity canonical manifold decomposition of hyperspectral data linear unmixing method (SMNMF). In this experiment, high spectral method and MMSNMF method the unmixing results were compared in the experiment, synthetic hyperspectral data and real hyperspectral data. The results showed that this method compared to the previous method, the numerical and visual mixing solutions There is a certain improvement in the fruit.
This work was supported by the national basic research program (973 Program):No.2013CB329402, the National Natural Science Foundation 61072108606010296097111261173090), New Century Talents Project: NCET-10-0668, higher school subject innovation engineering plan (111 plan): No. B0704, Doctoral Fund of Ministry of Education (20120203110005), Armament Research Foundation (9*****). And HUAWEI innovation research project (IRP-2013-01-09) funding.

【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】