改进的经验模态分解算法及其在高光谱图像分类中的应用
发布时间：2018-09-04 09:08
【摘要】：源于20世纪80年代的高光谱,是遥感技术的一大飞跃。与传统可见光或多光谱图像相比,高光谱图像能提供更为丰富的地表覆盖信息和地物光谱信息,在航天和军事等领域具有很大潜力,因而受到国内外学者的极大关注,成为目前的一个研究热点。然而,如何遵循高光谱数据“非线性”、“非平稳”的本质,充分利用采集到的图像信息,以此提高地物分类识别的能力,成为高光谱遥感技术推广和应用道路上亟待解决的难题。虽然近年来发展起来的经验模态分解算法(Empirical Mode Decomposition,EMD)对处理复杂的“非线性”、“非平稳”高光谱数据具有先天优势,但迄今为止尚无一套完整的、公认的理论基础。因此,如何从理论上对EMD进行改进,尤其是端点效应的抑制和包络求取时的“上冲”或“下冲”现象的消除等,仍然是摆在研究者们面前的重大挑战。本文围绕EMD算法的理论改进及其在高光谱图像分类中的应用展开深入细致的研究。一方面,基于灰色模型和最优化理论,从端点效应抑制和包络求取等方面对EMD进行改进,并将改进的EMD用于高光谱地物谱线和空间特征提取;另一方面,基于稀疏表示分类器(Sparse RepresentationClassifier,SRC)和超像素图像分割算法,研究谱-空间特征相结合的高光谱图像分类方法。本文的主要研究内容和创新成果如下:研究基于单变量灰色模型的一维EMD(Gray Mode-based EMD,GM-EMD)端点效应抑制方法。端点效应是EMD算法中的公开理论问题,它是由于信号两端附近极值点无法被准确确定而产生的。本文先证明极值点和包络对一维EMD至关重要,然后在微分方程离散化过程中对现有单变量灰色模型进行了改进,提出GM-EMD方法,在不改变原始信号特性的前提下,充分发挥单变量灰色模型适合少量数据预测、计算量小和短期预测精度高等优势,对一维EMD迭代过程中的信号向外延拓2个点,有效抑制一维EMD的端点效应。研究基于交替方向乘子法(Alternating Direction Method of Multipliers,ADMM)的一维EMD包络求取方法。由于样条插值过程中未对非极值点位置上的信号进行约束,容易造成“上冲”或“下冲”现象。于是,构建上、下包络应满足的严格数学不等式,用ADMM来求解这些最优化问题,与样条插值相比,能有效消除“上冲”或“下冲”现象,得到更准确的分解结果。研究基于多变量灰色模型的二维EMD(Gray Mode-based Bi-dimensionalEMD,GM-BEMD)端点效应抑制算法。由于图像边界处极值点无法被准确求取,造成了二维EMD(Bi-dimensional EMD,BEMD)的端点效应问题。于是,提出GM-BEMD方法,该方法由基于复化Simpson公式的多变量灰色模型将图像向四周延拓,再用传统的BEMD方法对延拓后的图像进行分解,得到延拓的二维本征模态函数(Bi-dimensional Intrinsic Mode Function,BIMF)和残差,然后截取出与原始图像区域相对应的分解结果,从而抑制BEMD端点效应。研究基于核稀疏多任务学习(Kernel-based Sparse Multitask Learning Classi-fier,KSMTLC)的谱-空间特征分类方法。在SRC的框架下,提出用KSMTLC来综合利用高光谱谱-空间特征,所涉及的最优化问题由加速近端梯度法(Accelerated Proximal Gradient,APG)求解。与单一的光谱或空间特征相比,KSMTLC的分类效果更好。研究基于支持向量机(Support Vector Machine,SVM)和超像素图像分割的谱-空间特征分类方法。先利用光谱特征得到高光谱的初步分类结果,再通过后处理的方式利用空间特征对初步分类结果进行修正,从而得到最终的谱-空间特征分类结果。其中,初步分类结果由SVM对光谱特征进行分类得到,而空间后处理采用超像素图像分割方法。这种方法极为简单但却是被首次提出,加入超像素图像分割空间后处理,能改善高光谱图像分类效果。
[Abstract]:Hyperspectral imagery, which originated in the 1980s, is a great leap forward in remote sensing technology. Compared with traditional visible and multispectral imagery, hyperspectral imagery can provide more abundant information of surface coverage and surface features, and has great potential in space and military fields. Therefore, it has attracted great attention of scholars at home and abroad and has become a current one. However, how to follow the "non-linear" and "non-stationary" nature of hyperspectral data and make full use of the collected image information to improve the ability of classification and recognition of ground objects has become an urgent problem in the promotion and application of hyperspectral remote sensing technology. However, there is no complete and accepted theoretical foundation so far. Therefore, how to improve the EMD theoretically, especially the "up" or "down" phenomena in the suppression of the endpoint effect and the envelopment calculation, is an important issue. Elimination is still a major challenge for researchers. This paper focuses on the theoretical improvement of EMD algorithm and its application in hyperspectral image classification. On the one hand, based on the grey model and optimization theory, EMD is improved from the aspects of end-effect suppression and envelope extraction, and the improved EMD is proposed. On the other hand, hyperspectral image classification based on sparse Representation Classifier (SRC) and super-pixel image segmentation algorithm is studied. The main research contents and innovations of this paper are as follows: 1. A Gray Mode-based EMD (GM-EMD) endpoint effect suppression method is proposed. The endpoint effect is an open theoretical problem in EMD algorithm. It is caused by the fact that the extreme points near the two ends of the signal can not be accurately determined. This paper first proves that the extreme points and envelopes are very important to one-dimensional EMD, and then it is important to the existing single-dimensional EMD in the discretization process of differential equations. The variable grey model is improved and the GM-EMD method is proposed. On the premise of not changing the original signal characteristics, the single variable grey model is fully utilized for small amount of data prediction, small amount of calculation and high precision of short-term prediction. The signal in one-dimensional EMD iteration process is extended outward by two points, and the end-point effect of one-dimensional EMD is effectively suppressed. One-dimensional EMD envelope extraction method based on Alternating Direction Method of Multipliers (ADMM) is proposed. Since the signal at the position of non-extreme points is not constrained in the spline interpolation process, it is easy to cause "up" or "down" phenomena. Compared with spline interpolation, it can effectively eliminate the "up" or "down" phenomena and get more accurate decomposition results. A two-dimensional EMD (Gray Mode-based Bi-dimensional EMD, GM-BEMD) endpoint effect suppression algorithm based on multivariable grey model is studied. The problem of end effect in two-dimensional EMD (Bi-dimensional EMD, BEMD) is caused. Therefore, a GM-BEMD method is proposed, which extends the image to its surroundings by a multivariate grey model based on the complex Simpson formula, and then decomposes the extended image with the traditional BEMD method to obtain the extended two-dimensional intrinsic modal function (Bi-dimensional Intrins). IC Mode Function (BIMF) and Residual, and then intercept the decomposition results corresponding to the original image region to suppress the BEMD endpoint effect. The spectral-spatial feature classification method based on Kernel-based Sparse Multitask Learning Classi-fier (KSMTLC) is studied. Hyperspectral spectral spectral-spatial features are solved by Accelerated Proximal Gradient (APG). Compared with single spectral or spatial features, KSMTLC has better classification performance. Spectral-spatial feature classification based on support vector machine (SVM) and superpixel image segmentation is studied. The initial classification result of hyperspectral feature is obtained by spectral feature, and then the initial classification result is modified by spatial feature through post-processing. The final classification result of spectral-spatial feature is obtained by SVM, and the superpixel image segmentation is used in spatial post-processing. This method is very simple, but it is the first time proposed that the hyperspectral image classification effect can be improved by adding super-pixel image segmentation space post-processing.
【学位授予单位】：哈尔滨工业大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：TP751
，
本文编号：2221632
[Abstract]:Hyperspectral imagery, which originated in the 1980s, is a great leap forward in remote sensing technology. Compared with traditional visible and multispectral imagery, hyperspectral imagery can provide more abundant information of surface coverage and surface features, and has great potential in space and military fields. Therefore, it has attracted great attention of scholars at home and abroad and has become a current one. However, how to follow the "non-linear" and "non-stationary" nature of hyperspectral data and make full use of the collected image information to improve the ability of classification and recognition of ground objects has become an urgent problem in the promotion and application of hyperspectral remote sensing technology. However, there is no complete and accepted theoretical foundation so far. Therefore, how to improve the EMD theoretically, especially the "up" or "down" phenomena in the suppression of the endpoint effect and the envelopment calculation, is an important issue. Elimination is still a major challenge for researchers. This paper focuses on the theoretical improvement of EMD algorithm and its application in hyperspectral image classification. On the one hand, based on the grey model and optimization theory, EMD is improved from the aspects of end-effect suppression and envelope extraction, and the improved EMD is proposed. On the other hand, hyperspectral image classification based on sparse Representation Classifier (SRC) and super-pixel image segmentation algorithm is studied. The main research contents and innovations of this paper are as follows: 1. A Gray Mode-based EMD (GM-EMD) endpoint effect suppression method is proposed. The endpoint effect is an open theoretical problem in EMD algorithm. It is caused by the fact that the extreme points near the two ends of the signal can not be accurately determined. This paper first proves that the extreme points and envelopes are very important to one-dimensional EMD, and then it is important to the existing single-dimensional EMD in the discretization process of differential equations. The variable grey model is improved and the GM-EMD method is proposed. On the premise of not changing the original signal characteristics, the single variable grey model is fully utilized for small amount of data prediction, small amount of calculation and high precision of short-term prediction. The signal in one-dimensional EMD iteration process is extended outward by two points, and the end-point effect of one-dimensional EMD is effectively suppressed. One-dimensional EMD envelope extraction method based on Alternating Direction Method of Multipliers (ADMM) is proposed. Since the signal at the position of non-extreme points is not constrained in the spline interpolation process, it is easy to cause "up" or "down" phenomena. Compared with spline interpolation, it can effectively eliminate the "up" or "down" phenomena and get more accurate decomposition results. A two-dimensional EMD (Gray Mode-based Bi-dimensional EMD, GM-BEMD) endpoint effect suppression algorithm based on multivariable grey model is studied. The problem of end effect in two-dimensional EMD (Bi-dimensional EMD, BEMD) is caused. Therefore, a GM-BEMD method is proposed, which extends the image to its surroundings by a multivariate grey model based on the complex Simpson formula, and then decomposes the extended image with the traditional BEMD method to obtain the extended two-dimensional intrinsic modal function (Bi-dimensional Intrins). IC Mode Function (BIMF) and Residual, and then intercept the decomposition results corresponding to the original image region to suppress the BEMD endpoint effect. The spectral-spatial feature classification method based on Kernel-based Sparse Multitask Learning Classi-fier (KSMTLC) is studied. Hyperspectral spectral spectral-spatial features are solved by Accelerated Proximal Gradient (APG). Compared with single spectral or spatial features, KSMTLC has better classification performance. Spectral-spatial feature classification based on support vector machine (SVM) and superpixel image segmentation is studied. The initial classification result of hyperspectral feature is obtained by spectral feature, and then the initial classification result is modified by spatial feature through post-processing. The final classification result of spectral-spatial feature is obtained by SVM, and the superpixel image segmentation is used in spatial post-processing. This method is very simple, but it is the first time proposed that the hyperspectral image classification effect can be improved by adding super-pixel image segmentation space post-processing.
【学位授予单位】：哈尔滨工业大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：TP751
，
本文编号：2221632
本文链接：https://www.wllwen.com/guanlilunwen/gongchengguanli/2221632.html
