【作者】 王雪；

【导师】 谭琨；

【作者基本信息】 中国矿业大学 ， 摄影测量与遥感， 2019， 博士

【摘要】 高光谱遥感影像具有较好的光学特征和丰富的光谱信息,其图谱合一的特性使得高光谱影像能够为精细分类、遥感反演和目标探测等任务提供新的研究方式。分类是高光谱数据分析中的基本问题之一,虽然高光谱影像中含有丰富的地物信息,但其数据冗余、空间分辨率不足、标记样本缺乏等因素,限制了高光谱影像分类的应用与发展。针对上述问题,引入在计算机视觉任务中的深度生成模型,围绕高光谱影像光谱生成、紧耦合式分类训练和空-谱特征提取等方向开展研究。主要创新工作如下。(1)提出了一种面向光谱生成有效性的条件式变分对抗网络(Conditional Variational and Adversarial Autoencoder,CVA~2E),解决了光谱特征有效性学习的问题。在分析生成模型的训练不稳定、模式坍塌、生成样本质量低等问题形成原因的基础上,在生成对抗网络中引入变分推理过程,提升了模型的学习能力。此外,通过添入两个惩罚项,提高了生成样本的多样性和生成样本的光谱形状特征。三个标准高光谱数据集的实验结果表明,CVA~2E在光谱合成能力上优于其他方法。在使用ROSIS和AVIRIS数据集时,引入全损失惩罚的CVA~2E_SAD_FVA表现最佳,精度依次为96.74%和89.7%,引入了特征角损失的CVA~2E_FVA在HYSPEX数据集上表现最好,精度为98.33%,证明了CVA~2E在挖掘分类特征方面更有效。(2)提出了一种面向分类任务的多组分对抗生成胶囊网络(Capsule Triple Generative Adversarial Network,Caps-TripleGAN),通过引入向量神经元和多组分对抗训练结构进行紧耦合式训练,进一步提高了深度生成框架的判别能力。针对CVA~2E弱耦合训练过程进行优化,引入TripleGAN结构,将生成对抗训练过程,分为分类器、生成器和判别器三组分紧耦合式训练。引入胶囊神经元对光谱特征进行提取,在提取特征时充分考虑特征及其位置和方向信息,提取更加有效的高阶可分特征。实验结果表明,胶囊网络的性能优于对比实验中的深度学习方法,同时TripleGAN可以提高CapsNet在少量训练样本的分类性能。(3)针对高光谱分类过程中空谱信息融合问题,以分离式双通道学习过程和伪标签辅助思想为基础,提出了两种基于深度生成模型的高光谱影像空谱特征学习方法。首先,基于紧耦合对抗机制,将空间信息和光谱信息进行分离式输入,在高阶特征提取层进行合并,提出一种面向空谱特征的双通道多组分生成对抗网络(Dual Triple Generative Adversarial Network,Dual-TripleGAN),完成空谱特征学习和分类。另外,为探索空谱数据块完整性的生成方法,提出一种基于伪标签的多尺度生成辅助分类方法(Multiscale Generative Assistant Capsule Network,MS-GA-CapsNet)。针对空谱数据块的学习过程,使用差异化加权思想,增加中心像素的网络权重,模型考虑了邻域像元类别信息,对网络的光谱特征进行混合式生成。同时,在空谱特征提取过程中,使用多尺度胶囊网络进行分类。实验结果表明,提出的多尺度生成辅助分类方法在四种数据上优于其他算法。该论文有图71幅,表27个,参考文献170篇。更多还原

【Abstract】 Hyperspectral image,which is an important data source in remote sensing area,contains both spectral and spatial information of the scene.Therefore,it is widely utilized in the field of precision classification,quantitative retrieval and target detection etc..Classification is the most primary problem of hyperspectral image analysis.There is a wealth of information in hyperspectral image.However,the issues of large data redundancy,low spatial resolution and small training set are still the barrier of the application and development of hyperspectral image.In view of the above issues,this dissertation introduced the deep generation model which is with remarkable performance in computer vision task,and carried out researches in below aspects:generation of hyperspectral image,stable training process of deep generative model and spatio-spectral feature learning.The contributions of this dissertation are summarized as follows:(1)An innovative generative network named conditional variational and adversarial autoencoder(CVA~2E)has been proposed.Firstly,the causes of training instability,model collapse and low quality of generated samples are analyzed.Then variational inference process is introduced into the generative adversarial network to improve the generative ability of the model.Moreover,to learn the fine-grained spectral characteristics of individual hyperspectral pixels,the spectral angle distance and vectorial angle measurement are introduced in the loss function of CVA~2E.The improved CVA~2E shows a superior performance in the spectral synthesis of different categories.To demonstrate the ability of the generated samples for the classification task,three kind of scenarios,ROSIS,AVIRIS and HYSPEX dataset,were carried out.Experimental results showed that the proposed model achieved the best performance.CVA~2E with spectral angle distance and vectorial angle measurement obtained the best results of 96.74%and 89.7%on the ROSIS dataset and AVIRIS dataset,respectively.Besides,CVA~2E with vectorial angle measurement obtained the best result of 98.33%on the HySpex dataset.(2)An innovative network named capsule triple generative adversarial network(Caps-TripleGAN)has been proposed.Firstly,CVA~2E is rediscussed to explore the reason of its weak-coupling training process.In this regard,TripleGAN is introduced which consists of classifier,generator and discriminator.Moreover,CapsNet is modified to adapt to the hyperspectral imagery classification,which takes the features and their locations,as well as their directions,into account during feature extraction process.To demonstrate the ability of the generated samples for the classification task,three kind of scenarios,ROSIS,AVIRIS and AHSI,are carried out.It has found that the CapsNet outperforms other deep learning approaches and TripleGAN can improve the performance of CapsNet on small training dataset.(3)Spatio-spectral features have been explored and two innovative methods have been proposed based on deep generative model.Firstly,based on the tight-coupling networks proposed in Chapter 4,the spatial information and spectral features are inputted to the network separately.After that,the extracted features are combined in the deep layer to construct the dual channel TripleGAN(Dual-TripleGAN).The other proposed approach is named multiscale generative assistant capsule network(MS-GA-CapsNet).During the learning process of spatio-spectral blocks,pixels in the blocks are reweighted using different weights,and the spectrum of the central pixel is generated in a mixed condition by taking the label information of the neighborhood into account.In the process of spatio-spectral feature extraction,multi-scale CapsNet is utilized.To demonstrate the ability of the proposed methods,three datasets ROSIS,AVIRIS and HYSPEX which from airborne imaging spectrometer and AHSI which from spaceborne imaging spectrometer were utilized.The results showed that the MS-GA-CapsNet achieved the best performance.There are 71 figures,27 tables,and 170 references in this dissertation.更多还原