基于非负矩阵分解的高光谱图像解混研究

发布时间：2018-06-19 18:20

本文选题：高光谱图像 + 混合像元分解　；参考：《中国科学院大学(中国科学院遥感与数字地球研究所)》2017年硕士论文

【摘要】：高光谱遥感技术同时探测获得地物的二维几何空间和一维光谱信息,因获取高光谱分辨率的遥感图像,其相邻波段间隔窄且连续而被广泛应用。但由于成像光谱仪的限制且地物复杂多样,高光谱遥感数据的空间分辨率相对较低,混合像元普遍存在,但高光谱遥感所获得的高光谱分辨率使得混合像元分解在高光谱图像中成为可能,而混合像元问题也是高光谱图像应用的主要障碍,因此高光谱图像解混问题成为高光谱遥感图像数据处理分析中一个重要的研究方向。高光谱图像解混是研究混合像元的组成端元以及各端元参与混合的程度的过程。非负矩阵分解作为一种盲源分离的方法,其矩阵分解的数学模型与线性解混模型相似,且分解前的矩阵和分解所得矩阵都具有非负性,能够满足高光谱图像解混问题中的非负性要求,因此被运用到解混问题中。但基于非负矩阵分解的解混问题是欠定问题,容易获得局部最小值,受初始值和约束项影响较大,针对这个问题,本文在考虑高光谱图像本身的理化特性的基础上,挖掘高光谱图像的空间光谱信息,设计基于非负矩阵分解的解混算法的优化方法。本文主要内容以及创新点如下:(1)总结分析高光谱解混问题和非负矩阵分解相关的理论基础,包括线性解混模型、端元数量确认、端元提取、丰度反演以及精度评定的方法,非负矩阵分解以及目前已有的基于约束非负矩阵分解的解混算法。(2)提出基于约束非负矩阵分解的高光谱解混算法的空间光谱预处理流程。在利用非负矩阵分解做高光谱图像的解混问题的算法中,很容易出现的局部最小问题,通过图像预处理,综合考虑高光谱图像数据的空间信息和光谱特征,获得预选端元,为非负矩阵分解带来更好的初始值。通过空间预处理(SPP)和空间光谱预处理(SSPP)两种预处理方法对约束非负矩阵分解解混算法(MVCNMF和GNMF)进行预处理,分别使用模拟数据和Cuprite矿区真实实验数据进行有效验证。(3)提出基于丰度稀疏性和光谱平滑性的约束非负矩阵分解的高光谱图像解混算法。针对高光谱遥感数据的本身理化特性而对基于非负矩阵分解的高光谱混合像元分解算法的优化改进,主要从高光谱传感器获取地物光谱信息的方式在遥感图像上体现出的平滑性以及地物分布的固有特征在遥感图像上的体现出的丰度稀疏性两方面考虑,介绍稀疏非负矩阵分解(SNMF)和局部光滑约束非负矩阵分解(PSNMF),提出对丰度稀疏约束和对端元平滑约束的稀疏平滑非负矩阵分解(SPSNMF)解混算法,然后通过模拟数据和Washington,DC地区、Cuprite矿区两组真实实验数据分别进行了有效验证。
[Abstract]:Hyperspectral remote sensing technology simultaneously detects the two-dimensional geometric space and one-dimensional spectral information of ground objects. Because of the high spectral resolution of remote sensing images, its adjacent bands are widely used because of their narrow interval and continuity. However, because of the limitation of imaging spectrometer and the complexity of ground objects, the spatial resolution of hyperspectral remote sensing data is relatively low. However, the hyperspectral resolution obtained by hyperspectral remote sensing makes it possible to decompose mixed pixels in hyperspectral images, and the mixed pixel problem is also the main obstacle to the application of hyperspectral images. Therefore, the problem of hyperspectral image de-mixing has become an important research direction in hyperspectral remote sensing image data processing and analysis. Hyperspectral image demultiplexing is a process of studying the composition of the mixed pixel and the degree to which each end element is involved in the mixing. As a blind source separation method, the nonnegative matrix factorization has similar mathematical model to the linear unmixing model, and the matrix before decomposition and the matrix obtained from the decomposition are nonnegative. It can satisfy the requirement of nonnegativity in the problem of hyperspectral image unmixing, so it is applied to the problem of descrambling. However, the unmixing problem based on nonnegative matrix factorization is an underdetermined problem, which is easy to obtain the local minimum and is greatly affected by initial values and constraints. In view of this problem, the physical and chemical properties of hyperspectral images are considered in this paper. The spatial spectral information of hyperspectral images is mined and an optimization method based on nonnegative matrix decomposition is designed. The main contents and innovations of this paper are as follows: 1) summarizing and analyzing the theoretical basis of hyperspectral unmixing problem and nonnegative matrix factorization, including linear demultiplexing model, endmember quantity confirmation, endcomponent extraction, abundance inversion and accuracy evaluation. Non-negative matrix factorization and the existing unmixing algorithm based on constrained nonnegative matrix factorization. 2) the spatial spectral pretreatment process of hyperspectral descrambling algorithm based on constrained nonnegative matrix decomposition is proposed. In the algorithm of decomposing hyperspectral image by using non-negative matrix decomposition, the local minimum problem is easy to appear. By image preprocessing, the spatial information and spectral characteristics of hyperspectral image data are considered synthetically, and the pre-selected endpoints are obtained. It brings better initial value for nonnegative matrix decomposition. Spatial preprocessing (SPP) and spatial spectral preprocessing (SSPP) are used to preprocess the constrained nonnegative matrix factorization unmixing algorithms (MVCNMF and GNMFs). Simulation data and real experimental data of Cuprite mining area are used to verify the hyperspectral image unmixing algorithm based on constrained nonnegative matrix factorization based on abundance sparsity and spectral smoothness respectively. According to the physical and chemical properties of hyperspectral remote sensing data, the hyperspectral mixed pixel decomposition algorithm based on non-negative matrix decomposition is optimized and improved. Mainly from the hyperspectral sensor to obtain ground object spectral information on the remote sensing image reflected in the smoothness and the distribution of the inherent characteristics of the remote sensing image reflected in the abundance of two aspects of the sparsity. This paper introduces sparse nonnegative matrix factorization (SNMF) and local smooth constraint nonnegative matrix factorization (PSN MFF), and proposes a sparse smooth nonnegative matrix factorization (SNMF) unmixing algorithm for abundance sparse constraint and end element smoothing constraint. Then the simulation data and two groups of real experimental data of Cuprite mining area in Washington DC area are validated.
【学位授予单位】：中国科学院大学(中国科学院遥感与数字地球研究所)
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP751

【相似文献】