基于高光谱影像的抗噪模型及解混算法研究
发布时间：2019-04-03 20:32
【摘要】：为提高遥感影像在使用过程中的精度,高光谱图像的混合像元解混是研究者们亟待解决的问题。传统的混合像元解混算法主要基于线性解混模型、双线性模型或者非线性解混模型,这些模型均以最小化加性噪声为目标,实现端元提取并获得端元丰度系数。本论文在分析传统解混模型基础上,构建抗噪模型,进而提出基于抗噪模型的混合像元解混算法,最终由实验验证算法在高光谱图像解混中的有效性。本文研究内容包括如下几个方面： 1混合像元解混的基础知识：研究高光谱图像成像原理,分析高光谱图像成像过程中产生的噪声,总结混合像元解混之前常见的预处理算法存在的问题。深入研究目前广泛使用的三种混合像元解混模型与其对应的混合像元解混的典型算法。并介绍常见的混合像元解混精度评价机制。 2针对传统的线性模型及非线性模型均以加性噪声作为随机误差的产生原因,忽略了高光谱图像成像过程中可能产生的其他噪声。本文分析噪声对高光谱数据解混结果的稀疏性及稳定性的影响,推导传统解混模型误差产生机制及乘性噪声、混合噪声的危害,提出基于Itakura-Saito (IS)距离的抗噪模型。并通过理论推导证明抗噪模型能够克服独立于高光谱数据存在的加性噪声及乘性噪声、混合噪声对混合像元解混的影响。 3为了证明提出抗噪模型的必要性,作者改进基于传统模型的混合像元解混算法,并由理论及实验说明传统模型的固有局限性。本论文以线性模型为例,推导基于线性模型的解混结果精确度,分析影响线性模型混合像元解混精度的因素,在基于对传统线性模型的误差推导基础上,提出基于传统解混模型改进的混合像元解混算法。在新的解混算法里,全面考虑影响结果的随机测量误差及其他因素,此外,通过添加阻尼项,提高解的稳定性。由共轭梯度算法加速新的解混算法的收敛速度,并由实验验证其有效性。 4研究基于抗噪模型的混合像元解混算法：在抗噪模型基础上,研究对应的抗噪的混合像元解混算法。针对混合像元解混问题是不适定问题,混合像元解混结果对初始值极其敏感,本文提出构建端元字典,并改进传统的正交基匹配追踪(OMP)算法,根据改进的正交基匹配追踪算法具有基正交性及非负性,在没有约束参数的情况下,求解Lo范数约束问题,提高解混结果的稳定性及稀疏性。采用基于IS距离迭代的非负矩阵分解算法,克服混合像元解混过程中的噪声,提高解混算法的解混能力及抗噪性能。
[Abstract]:In order to improve the precision of remote sensing images, it is an urgent problem for researchers to decompose the mixed pixels of hyperspectral images. The traditional mixed pixel unmixing algorithms are mainly based on linear unmixing model bilinear model or nonlinear unmixing model. These models are aimed at minimizing additive noise to achieve end-element extraction and obtain the end-element abundance coefficient. In this paper, based on the analysis of the traditional deconvolution model, the anti-noise model is constructed, and then the hybrid pixel unmixing algorithm based on the anti-noise model is proposed. Finally, the effectiveness of the algorithm in the hyperspectral image unmixing is verified by experiments. The main contents of this paper are as follows: (1) the basic knowledge of mixed pixel unmixing: studying the principle of hyperspectral image imaging and analyzing the noise produced in the process of hyperspectral image imaging. This paper summarizes the problems existing in the pre-processing algorithms before the mixing of mixed pixels. In-depth study of three widely used mixed pixel demultiplexing models and their corresponding mixed pixel de-mixing typical algorithms. At the same time, the common evaluation mechanism of mixed pixel unmixing accuracy is introduced. (2) in the traditional linear model and nonlinear model, the additive noise is used as the cause of random error, and the other noises which may be produced in the process of hyperspectral image imaging are ignored. This paper analyzes the influence of noise on the sparsity and stability of the results of hyperspectral data unmixing, deduces the mechanism of error generation and multiplicative noise of traditional unmixing model and the harm of mixed noise, and puts forward an anti-noise model based on Itakura-Saito (IS) distance. It is proved by theoretical derivation that the anti-noise model can overcome the additive noise and multiplicative noise independent of hyperspectral data, and the influence of mixed noise on the unmixing of mixed pixels. 3 in order to prove the necessity of putting forward the anti-noise model, the author improves the mixed pixel deconvolution algorithm based on the traditional model, and explains the inherent limitation of the traditional model by theoretical and experimental results. Taking the linear model as an example, this paper deduces the precision of the unmixing result based on the linear model, analyzes the factors that influence the precision of the mixed pixel of the linear model, and on the basis of the error derivation of the traditional linear model, An improved hybrid pixel deconvolution algorithm based on traditional unmixing model is proposed. In the new unmixing algorithm, the random measurement error and other factors affecting the results are considered comprehensively. In addition, the stability of the solution is improved by adding damping term. The conjugate gradient algorithm is used to accelerate the convergence rate of the new deconvolution algorithm, and the effectiveness of the algorithm is verified by experiments. 4. The hybrid pixel unmixing algorithm based on the anti-noise model is studied. On the basis of the anti-noise model, the corresponding anti-noise mixed pixel deconvolution algorithm is studied. In view of the ill-posed problem of mixed pixel unmixing, the results of mixed pixel deconstructing are very sensitive to the initial value. In this paper, an end-element dictionary is proposed and the traditional orthogonal basis matching tracking (OMP) algorithm is improved. According to the basis orthogonality and non-negativity of the improved orthogonal basis matching tracking algorithm, the Lo norm constrained problem is solved without the constraint parameters, and the stability and sparsity of the unmixed results are improved. The non-negative matrix decomposition algorithm based on IS distance iteration is used to overcome the noise in the process of mixed pixel de-mixing and to improve the unmixing ability and anti-noise performance of the unmixing algorithm.
【学位授予单位】：华东师范大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：TP751
本文编号：2453554
[Abstract]:In order to improve the precision of remote sensing images, it is an urgent problem for researchers to decompose the mixed pixels of hyperspectral images. The traditional mixed pixel unmixing algorithms are mainly based on linear unmixing model bilinear model or nonlinear unmixing model. These models are aimed at minimizing additive noise to achieve end-element extraction and obtain the end-element abundance coefficient. In this paper, based on the analysis of the traditional deconvolution model, the anti-noise model is constructed, and then the hybrid pixel unmixing algorithm based on the anti-noise model is proposed. Finally, the effectiveness of the algorithm in the hyperspectral image unmixing is verified by experiments. The main contents of this paper are as follows: (1) the basic knowledge of mixed pixel unmixing: studying the principle of hyperspectral image imaging and analyzing the noise produced in the process of hyperspectral image imaging. This paper summarizes the problems existing in the pre-processing algorithms before the mixing of mixed pixels. In-depth study of three widely used mixed pixel demultiplexing models and their corresponding mixed pixel de-mixing typical algorithms. At the same time, the common evaluation mechanism of mixed pixel unmixing accuracy is introduced. (2) in the traditional linear model and nonlinear model, the additive noise is used as the cause of random error, and the other noises which may be produced in the process of hyperspectral image imaging are ignored. This paper analyzes the influence of noise on the sparsity and stability of the results of hyperspectral data unmixing, deduces the mechanism of error generation and multiplicative noise of traditional unmixing model and the harm of mixed noise, and puts forward an anti-noise model based on Itakura-Saito (IS) distance. It is proved by theoretical derivation that the anti-noise model can overcome the additive noise and multiplicative noise independent of hyperspectral data, and the influence of mixed noise on the unmixing of mixed pixels. 3 in order to prove the necessity of putting forward the anti-noise model, the author improves the mixed pixel deconvolution algorithm based on the traditional model, and explains the inherent limitation of the traditional model by theoretical and experimental results. Taking the linear model as an example, this paper deduces the precision of the unmixing result based on the linear model, analyzes the factors that influence the precision of the mixed pixel of the linear model, and on the basis of the error derivation of the traditional linear model, An improved hybrid pixel deconvolution algorithm based on traditional unmixing model is proposed. In the new unmixing algorithm, the random measurement error and other factors affecting the results are considered comprehensively. In addition, the stability of the solution is improved by adding damping term. The conjugate gradient algorithm is used to accelerate the convergence rate of the new deconvolution algorithm, and the effectiveness of the algorithm is verified by experiments. 4. The hybrid pixel unmixing algorithm based on the anti-noise model is studied. On the basis of the anti-noise model, the corresponding anti-noise mixed pixel deconvolution algorithm is studied. In view of the ill-posed problem of mixed pixel unmixing, the results of mixed pixel deconstructing are very sensitive to the initial value. In this paper, an end-element dictionary is proposed and the traditional orthogonal basis matching tracking (OMP) algorithm is improved. According to the basis orthogonality and non-negativity of the improved orthogonal basis matching tracking algorithm, the Lo norm constrained problem is solved without the constraint parameters, and the stability and sparsity of the unmixed results are improved. The non-negative matrix decomposition algorithm based on IS distance iteration is used to overcome the noise in the process of mixed pixel de-mixing and to improve the unmixing ability and anti-noise performance of the unmixing algorithm.
【学位授予单位】：华东师范大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：TP751
【参考文献】
相关期刊论文 前6条
1 张伟;杜培军;张华鹏;;基于神经网络的高光谱混合像元分解方法研究[J];测绘通报;2007年07期
2 黎贵发;;基于改进BP神经网络的遥感影像混合像元分解[J];地理空间信息;2009年03期
3 李乐;章毓晋;;非负矩阵分解算法综述[J];电子学报;2008年04期
4 王立国;邓禄群;张晶;;基于线性最小二乘支持向量机的光谱端元选择算法[J];光谱学与光谱分析;2010年03期
5 王强;束炯;尹球;;高光谱图像光谱域噪声检测与去除的DSGF方法[J];红外与毫米波学报;2006年01期
6 刘雪松;王斌;张立明;;基于非负矩阵分解的高光谱遥感图像混合像元分解[J];红外与毫米波学报;2011年01期
相关博士学位论文 前1条
1 李二森;高光谱遥感图像混合像元分解的理论与算法研究[D];解放军信息工程大学;2011年
，本文编号：2453554
本文链接：https://www.wllwen.com/guanlilunwen/gongchengguanli/2453554.html
