基于稀疏约束的高光谱解混技术研究

发布时间：2018-12-12 00:43

【摘要】：随着遥感技术和成像光谱仪的发展,高光谱遥感图像在越来越多的领域中得到应用。然而,由于其较低的空间率以及地物分布的复杂性,一个像元往往是由好几种地物构成的,这严重阻碍高光谱图像的实际应用。因而,对高光谱图像进行混合像元分解就特别有意义了。目前在该研究领域,国内外的新方法新思路层出不穷。基于稀疏约束的高光谱解混问题已经成为当今遥感领域的一个热点,它是一个稀疏回归问题,目标是在一个较大的光谱库中寻找能够表征像元的最优光谱子集。但是仍存在着没有利用图像的空间信息等缺点,本论文在总结前人研究成果的基础上,针对基于稀疏约束的高光谱图像混合像元分解作了大量研究,主要的研究内容如下:首先,阐述了高光谱线性和非线性混合模型,介绍了线性解混的基本步骤,对于端元数目估计算法、端元提取算法和丰度反演算法进行了逐一地介绍。然后,详细阐述了基于稀疏约束的高光谱解混模型。它假设混合像元可以被表示成一个已知的光谱库中各光谱曲线的线性组合。这样,解混就相当于在光谱库中寻找能够表征这个像元的最佳的光谱库子集。稀疏解混问题本质上是L0范数的优化问题。通常情况下,将L0范数最小化问题转化为L1范数最小化问题来解。本文采用变量分裂和增广拉格朗日算法来进行稀疏解混,它是一种非常快速的方法,并对其正则化参数和罚参数的取值进行了研究,以便取得一个相对较优的参数。通过模拟实验分析了光谱库的互相关函数值MC对于解混结果的影响,得到了一般性的结论:光谱库矩阵的互相关函数值MC越小,稀疏解混的效果越好。最后,深入研究了加权的L1正则化稀疏解混模型,其相对于L1范数,更加地接近L0范数。针对其仅仅考虑数学意义上的最优化,并没有利用地物的实际分布这一问题,提出了基于修正权值的加权L1正则化稀疏解混算法,在权值的迭代更新过程中引入了空间信息。实验结果了证明改进的加权L1正则化稀疏解混算法在较低信噪比的高光谱图像中能够有效地提升解混精度。
[Abstract]:With the development of remote sensing technology and imaging spectrometer, hyperspectral remote sensing image has been applied in more and more fields. However, due to its low spatial rate and the complexity of the distribution of ground objects, a pixel is usually made up of several ground objects, which seriously hinders the practical application of hyperspectral images. Therefore, it is particularly meaningful to decompose hyperspectral images with mixed pixels. At present, in the field of research, new methods and new ideas emerge endlessly at home and abroad. The problem of hyperspectral descrambling based on sparse constraints has become a hot spot in the field of remote sensing. It is a sparse regression problem. The goal is to find the optimal spectral subset which can represent the pixel in a large spectral database. However, there are still some shortcomings, such as not using spatial information of images. Based on the previous research results, this paper makes a lot of research on mixed pixel decomposition of hyperspectral images based on sparse constraints. The main research contents are as follows: firstly, the hyperspectral linear and nonlinear mixed models are described, and the basic steps of linear unmixing are introduced. The number estimation algorithm, the end component extraction algorithm and the abundance inversion algorithm are introduced one by one. Then, the hyperspectral demultiplexing model based on sparse constraint is described in detail. It assumes that the mixed pixel can be represented as a linear combination of the spectral curves in a known spectral library. In this way, unmixing is equivalent to finding the best subset of the spectrum library that can represent the pixel in the spectral library. The sparse unmixing problem is essentially an optimization problem of L _ 0 norm. In general, the L _ 0 norm minimization problem is transformed into a L1 norm minimization problem. In this paper, variable splitting and augmented Lagrangian algorithm are used for sparse demultiplexing, which is a very fast method, and its regularization parameters and penalty parameters are studied in order to obtain a relatively optimal parameter. The influence of cross-correlation function value (MC) of spectral library on the results of descrambling is analyzed through simulation experiments. The general conclusion is drawn that the smaller the cross-correlation function value MC of spectral library matrix, the better the effect of sparse demultiplexing. Finally, the weighted L 1 regularized sparse demultiplexing model is studied, which is closer to L 0 norm than L 1 norm. Aiming at the problem that it only considers the optimization in mathematical sense and does not make use of the actual distribution of ground objects, a weighted L1 regularized sparse demultiplexing algorithm based on modified weights is proposed, and spatial information is introduced into the iterative updating process of weights. The experimental results show that the improved weighted L1 regularized sparse demultiplexing algorithm can effectively improve the resolution of hyperspectral images with lower SNR.
【学位授予单位】：哈尔滨工程大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP751

【参考文献】