基于核空谱信息挖掘的高光谱图像分类方法研究

发布时间：2018-07-23 09:16

【摘要】：遥感技术发展的总趋势是以更高空间分辨率、更高光谱分辨率、更高时间分辨率对地球进行探测，进而提供地表覆盖环境更加精确、细致的观测信息。自上世纪80年代光谱成像技术被提出以来，高光谱成像已经成为一种重要的遥感探测手段，其本质在于能够同时提供地物分布的空间信息和较高分辨率的光谱信息。因此，高光谱遥感图像数据处理与信息挖掘技术研究具有重要的理论意义和巨大的应用价值，已成为遥感成像探测与信息处理领域的研究热点。本论文以高光谱遥感图像地物分类为背景，以核学习理论和方法为技术框架，针对高光谱遥感图像的特征提取、光谱分类和空间-光谱信息联合分类等问题开展研究，重点研究了基于单核/多核学习理论的高光谱图像光谱信息和空谱联合信息挖掘技术，旨在充分利用高光谱图像所提供的空间-光谱联合信息，提高地物分类性能。本论文研究的主要工作体现在：首先，整体研究核学习理论及其最新进展——多核学习理论及方法，奠定本论文研究内容的理论基础。论文在概要介绍了核学习理论及核方法设计的基础上，研究和分析了多核学习理论所涉及到的多核构造、优化学习方法，在理论上对合成核和多尺度核方法进行了研究。其次，立足于高光谱图像数据自身统计特性，将数据特性同核方法设计有机结合，提出了基于子空间调制核的高光谱图像特征提取方法。论文依据成像光谱探测原理所决定的高光谱图像数据子空间特性，研究了三种子空间划分的度量准则；在此基础上，设计了子空间调制核函数，以使源自成像机理的数据子空间特性融入到核设计及特征提取方法中，进而达到充分利用高光谱成像和数据特性的目的；论文利用地物分类实验验证了所提出的子空间调制核方法的有效性，即在提取特征的同时有效地提高分类性能。再次，以基于光谱信息的地物分类应用为直接导引，重点研究了多尺度多核学习分类模型，提出了多尺度多核最优集成学习方法。针对以支持向量机为代表的传统核方法学习能力受限于单核函数的问题，本文提出了多尺度多核学习模型；进一步，将多尺度多核学习问题分解为多尺度核无监督学习和支持向量机优化两个子问题，并提出了秩“1”约束下的基于非负矩阵分解和核非负矩阵分解的多核最优集成学习方法。相比于传统支持向量机和当前主流的多核学习方法，本文所提出的方法具有更优的性能。最后，为充分挖掘和利用高光谱图像的空-谱信息，构建了多特征多核学习模型，将空间特征和光谱特征有机地融合在多核学习理论框架下，进一步提升了高光谱图像地物分类能力。论文构建了多特征多核学习模型，提出了多特征多核最优集成学习方法，用以实现空谱特征的联合分类，论文针对高光谱图像自身提取的空间-光谱信息联合、高分辨率可见光图像空间信息和高光谱图像光谱信息联合两种情况进行了研究。首先，，针对三类典型的高光谱图像空间特征(局部区域矩特征、Gabor空间纹理特征、多尺度形态学特征)进行了多核分类研究，分析了不同空间特征对于不同数据源的适应性；其次，分别从高空间分辨率的可见光图像和高光谱分辨率的高光谱图像中分别提取空间特征和光谱信息，构建多特征联合分类模型及方法。真实数据的实验结果表明，本文提出的模型及方法有效地提高了空谱特征可利用性和高光谱遥感图像分类性能。
[Abstract]:The general trend of the development of remote sensing technology is to detect the earth with higher spatial resolution , higher spectral resolution and higher temporal resolution , thus providing more accurate and detailed observation information of the surface covering environment . Since the spectral imaging technology of the 1980s has been proposed , hyperspectral imaging has become an important sensing means of remote sensing , which is essential in providing spatial information and high resolution spectral information of the geographical distribution . Therefore , the research of hyperspectral remote sensing image data processing and information mining has important theoretical significance and great application value , and has become a hot spot in the field of remote sensing imaging detection and information processing .

Based on the classification of hyperspectral remote sensing images as background , the paper studies the feature extraction , spectral classification and spatial - spectral information joint classification of hyperspectral remote sensing images by using kernel learning theory and method , and focuses on the high spectral image spectral information and space - spectrum joint information mining technology based on single - core / multi - core learning theory , aiming at fully utilizing the spatial - spectrum joint information provided by hyperspectral images and improving the classification performance of the terrain .

First , the whole study of the theory of nuclear learning and the latest development of multi - core learning theory and method lays a theoretical foundation for the research content of this paper . Based on the summary of the theory of nuclear learning and the design of nuclear method , this paper studies and analyzes the multi - core structure and optimization learning method involved in the multi - core learning theory , and studies the synthetic kernel and multi - scale nuclear method in theory .

Secondly , based on the statistical characteristics of hyperspectral image data , a method for extracting hyperspectral image features based on subspace modulation kernel is proposed . Based on the characteristics of hyperspectral image data subspace determined by imaging spectrum detection principle , the paper studies the measurement criterion of three - seed spatial division .
On this basis , the subspace modulation kernel function is designed so that the data subspace characteristics derived from the imaging mechanism are integrated into the kernel design and feature extraction method , so that the purpose of fully utilizing hyperspectral imaging and data characteristics is achieved ;
In this paper , the validity of the proposed subspace modulation kernel method is verified by means of classification experiments , that is , the classification performance can be improved effectively while extracting features .

Thirdly , the multi - scale multi - core learning classification model is focused on multi - scale multi - core learning classification model based on spectral information , and multi - scale multi - core optimal integrated learning method is proposed .
Further , the multi - scale multi - core learning problem is decomposed into two sub - problems of multi - scale nuclear non - supervised learning and support vector machine optimization , and a multi - core optimal integrated learning method based on non - negative matrix factorization and kernel non - negative matrix decomposition under the constraint of rank " 1 " is proposed . Compared with the traditional support vector machine and the current mainstream multi - core learning method , the method provided by the present invention has better performance .

Finally , the multi - feature multi - core learning model is constructed for the purpose of fully excavating and utilizing the space - spectrum information of hyperspectral image . The multi - feature multi - core learning model is constructed . The multi - feature multi - core learning model is constructed , and the multi - feature multi - core optimal integrated learning method is proposed to realize the joint classification of the spatial - spectral information , the high - resolution visible image space information and the hyperspectral image spectral information .
Secondly , the spatial features and spectral information are extracted from the high spectral images with high spatial resolution and high spectral resolution respectively . The multi - feature joint classification model and method are constructed . The experimental results of real data show that the model and method proposed in this paper effectively improve the classification performance of the spatial spectral features and hyperspectral remote sensing images .
【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;O433

【参考文献】