高光谱图像空谱联合监督分类算法及软件系统

发布时间：2018-05-17 17:31

本文选题：高光谱图像 + 监督分类　；参考：《南京理工大学》2017年硕士论文

【摘要】：高光谱遥感图像作为一种新型细分光谱成像遥感技术,已经广泛应用于军事勘测、环境调查、灾害评估、矿物探测等各个领域。高光谱图像监督分类是高光谱图像处理中的重要组成部分。传统监督分类方法仅仅挖掘了光谱维度的信息,但是忽视空间维度中包含的丰富信息,导致它们的分类效果不够理想。此外,高光谱图像的数据量较大,普通方法的算法复杂度比较高,造成分类的效率较低。因此,迫切需要开发一种能够充分挖掘空谱信息的兼具高效和高精度的高光谱图像分类方法。极限学习机(ELM)是一种新型的单隐层前馈神经网络算法,它能随机生成单隐层参数,因此具有快速学习、泛化能力强的特点。本文详细分析了极限学习机在高光谱图像分类中的优势和问题,接着从特征提取和核函数两个方面对极限学习机方法进行改进,针对像素级ELM监督分类错误率高的问题重点研究了一个空谱联合信息的ELM高光谱图像分类方法,集成了主成分分析、独立成分分析等几种特征提取方法和支持向量机、核支持向量机、极限学习机、核极限学习机等几种监督分类方法设计实现了一个"特征提取-监督分类-精度评价"一体化的高光谱图像监督分类算法原型系统。本文主要工作为:(1)设计了各类特征提取方法与常规ELM结合的高光谱图像监督分类两步方法。实验分析表明:不同特征提取方法(如主成分分析PCA、独立成分分析ICA、局部保持投影LPP、核主成分分析KPCA等)都能提高常规ELM对高光谱图像监督分类的精度。相比而言,KPCA与ELM的结合性能更佳。(2)比较研究了核化ELM框架中不同核函数对高光谱图像分类的影响,设计了结合复合核(Composite Kernel)和核化ELM(KELM)的高光谱图像监督分类方法,称为KELM-CK。该方法综合了空间-光谱维度信息,提升了小样本情况下的监督分类精度并克服了噪声影响。各类数据集上验证了该方法的有效性。(3)提出一种基于核范2DPCA(N-2DPCA)和KELM的多视学习监督分类方法。该方法利用N-2DPCA抽取高光谱图像不同视角下的空谱特征,通过多个KELM分类器集成学习和多数投票,可取得很高的监督分类精度。实验表明,该方法优于SVM、SVM-CK、ELM、MH_KELM等主流方法。(4)设计并实现一个高光谱图像分类原型系统。该系统集成多种特征提取方法和监督分类算法,其能够实现"特征提取-监督分类-精度评价"的功能。
[Abstract]:As a new subdivision spectral imaging remote sensing technology, hyperspectral remote sensing image has been widely used in military survey, environmental investigation, disaster assessment, mineral exploration and other fields. Hyperspectral image supervised classification is an important part of hyperspectral image processing. The traditional supervised classification method only excavates the spectral dimension information, but neglects the rich information contained in the spatial dimension, resulting in their classification effect is not ideal. In addition, the amount of data in hyperspectral images is large, and the algorithm complexity of common methods is high, which results in low efficiency of classification. Therefore, there is an urgent need to develop a highly efficient and accurate hyperspectral image classification method which can fully mine space-spectrum information. ELM) is a new single hidden layer feedforward neural network algorithm, which can generate single hidden layer parameters at random, so it has the characteristics of fast learning and strong generalization ability. In this paper, the advantages and problems of extreme learning machine in hyperspectral image classification are analyzed in detail, and then the method of extreme learning machine is improved from two aspects: feature extraction and kernel function. Aiming at the problem of high error rate of pixel level ELM supervised classification, this paper focuses on a ELM hyperspectral image classification method based on space-spectrum joint information, which integrates several feature extraction methods such as principal component analysis (PCA) and independent component analysis (ICA) and support vector machine (SVM). Several supervised classification methods such as kernel support vector machine (KSVM), extreme learning machine (LLM) and kernel extreme learning machine (KLM) are designed and implemented to realize a hyper spectral image supervised classification algorithm prototype system which integrates feature extraction, supervised classification and accuracy evaluation. The main work of this paper is to design a two-step method of hyperspectral image supervised classification, which combines all kinds of feature extraction methods with conventional ELM. The experimental results show that different feature extraction methods (such as PCA, ICA, LPP, KPCA, etc.) can improve the accuracy of supervised classification of hyperspectral images by conventional ELM. In this paper, the effects of different kernel functions on hyperspectral image classification in the framework of kernelized ELM are compared and studied. A hyperspectral image supervised classification method combined with Composite Kernel and Kernel Kernel Kernel is designed, which is called KELM-CK. This method integrates spatial and spectral dimension information, improves the accuracy of supervised classification in the case of small samples and overcomes the influence of noise. The validity of this method is verified on various datasets.) A multi-view learning supervised classification method based on kernel norm 2DPCACN-2DPCAA and KELM is proposed. This method utilizes N-2DPCA to extract space-spectrum features from different angles of view of hyperspectral images, and achieves high supervised classification accuracy by integrating learning and majority voting with multiple KELM classifiers. The experimental results show that this method is superior to SVMN SVM-CKM ELMHKELM and other mainstream methods in designing and implementing a hyperspectral image classification prototype system. The system integrates various feature extraction methods and supervised classification algorithms, and it can realize the function of feature extraction, supervised classification and accuracy evaluation.
【学位授予单位】：南京理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP751

【相似文献】