【作者】 王磊；

【导师】 徐新；

【作者基本信息】 武汉大学 ， 信号与信息处理， 2020， 博士

【摘要】 合成孔径雷达(Synthetic Aperture Radar,SAR)是一种主动微波雷达成像系统,具有全天时全天候成像、一定的穿透性、高精度和大幅面获取地表散射信息等特点,在军用及民用领域发挥着重要作用。极化SAR(Polarimetric SAR,PolSAR)可以发射和接收不同组合方式的极化波,为SAR影像解译提供更丰富的地物目标极化信息。极化SAR影像信息表达具有多样性、复杂性,极化散射机理、极化信息的有效提取和表达均有待进一步的研究,这为极化SAR影像的可视化、场景分类等解译工作提出了更高的挑战。深度学习在自然图像处理等领域发挥着革命性的作用,在极化SAR影像解译中也已逐步展现出优异的性能以及巨大的潜力,为极化SAR影像的精细解译提供新的解决方案。基于深度学习的极化SAR影像解译仍存在一些问题,首先,极化SAR影像具有特殊的微波成像机理,目视解译较为困难。极化SAR影像地物信息由极化相干矩阵等表达,具有多种散射信息表达形式。从目视效果和数据表达上,极化SAR影像均与光学影像存在明显差异。其次,在极化SAR影像分类的数据集上,获取大量准确的标注样本非常困难,深度学习模型难以学习极化SAR影像的信息表达,整景极化SAR影像的场景分类速度有待提高。因此,需要研究适用于极化SAR影像的具有较强泛化能力深度学习模型,学习极化SAR影像有效的信息表达,从而进一步提高极化SAR影像解译的性能。针对上述问题,在高分专项和信息融合项目的资助下,围绕极化SAR影像场景分类和可视化解译要求,在深度学习框架下,开展了面向极化SAR影像信息表达与精细分类的研究,以及极化SAR影像信息可视化表达的研究。目前基于深度学习的监督方式的PolSAR影像分类方法,通常将邻域窗数据分成一个类别,未考虑邻域窗中的像素可能由多种地物类型构成的情况。本文首次将卷积神经网络输出二维分类图的思想引入到PolSAR影像分类,提出了一种特征尺寸不变的卷积神经网络多像素同时分类方法,充分利用邻域像素间不同地物类型的相互关系信息表达,实现对邻域窗中的所有像素同时分类。在高分3号等多种常用PolSAR平台影像整景数据分类实验中,具有更快的速度、更强的泛化性能和更高的分类精度。在此基础上,本文进一步引入PolSAR影像旋转域的特征,使用卷积长短时记忆网络学习旋转域极化相干矩阵序列的隐藏信息表达,提升PolSAR影像监督分类性能。PolSAR影像难以获取大量准确有效的标注信息,无监督的PolSAR影像分类对于PolSAR影像解译非常重要,常规机器学习的无监督分类方法具有冗长的处理流程。本文提出了一种基于深度互信息表达的PolSAR影像无监督分类方法,以像素的邻域窗数据及其随机几何变换数据组成邻域窗样本对作为卷积神经网络的输入,学习邻域窗样本对的深度互信息表达,指导网络保持邻域窗数据的共有信息,去除特殊特性。本文方法在真实的PolSAR影像无监督分类中展示了优异的性能,超过了目前先进的无监督分类方法。本文方法无需繁琐的方法流程以及额外预处理和后处理,实现了端到端的操作。此外,基于深度互信息表达学习的网络模型,其深度特征表达可将半监督分类提升至优异水平。由于SAR特殊的侧视主动成像机制以及严重的相干斑,非专业人员很难通过极化SAR影像进行目视观察与解译。本文提出一种基于监督循环一致生成对抗网络的大尺寸极化SAR影像转换为类光学遥感影像的方法,为极化SAR影像信息可视化表达提供了新的方式。在光学遥感影像因天气影响而信息缺失时,转换的类光学遥感影像可以辅助非专业人员的目视观察。为促进深度学习在遥感影像解译的健康发展,本文公开了基于哨兵1号与哨兵2号影像的配对数据集。在此数据集上验证了本文的极化SAR到光学遥感影像转换模型,可以同时保持地物目标的类型以及结构信息,超过了一些著名的影像转换模型。转换的类光学遥感影像可以进一步用于光学遥感影像去云,补足光学遥感影像的信息缺失。本文在深度学习框架下,开展了极化SAR影像的特征学习与信息表达研究,发展了基于卷积神经网络的PolSAR影像全监督以及无监督分类方法,基于生成对抗网络的极化SAR影像到类光学遥感影像转换方法。所有方法均在我国高分3号极化SAR影像数据上取得了优异的结果。其中,极化SAR到类光学遥感影像转换方法已应用于“通导遥综合应用与多源数据融合系统”项目。更多还原

【Abstract】 Synthetic aperture radar(SAR)is an active microwave imaging system and it is one of the most important methods of Earth observation.SAR has the advantages of working under all weather conditions,large scope and certain penetration capacity,and it plays an important role in military and civilian fields.Modern SAR systems can provide polarimetric SAR(PolSAR)images by emitting and receiving fully polarized radar waves.The information representation of the PolSAR image has the characteristics of diversity and complexity.The extraction of scattering mechanism and polarimetric information requires further research.PolSAR image interpretation,such as visualization and classification,is a great challenge。Deep learning plays an important role in natural image processing and has shown excellent performance and great potential in PolSAR image interpretation.It provides new solutions for PolSAR image interpretation.Deep learning based PolSAR image interpretation methods still have some problems.First,PolSAR images have special imaging mechanism and the visual interpretation is difficult.The land cover information of PolSAR images is expressed by polarization-coherence matrix,and has various forms of scattering information representation.The visual appearance and data formation of PolSAR images are very different from optical images.Second,it is difficult to obtain a large number of accurately labeled samples for PolSAR image classification.Deep learning is difficult to learn the information representation of PolSAR images.The classification speed of a whole PolSAR image needs to be improved.Therefore,it is necessary to study deep learning models which have strong generalization ability and can learn the information representation of PolSAR images.With the help of the deep learning models,the PolSAR image interpretation performance can be improved.To solve the above problems,with the support of the sub-project of high resolution Earth observation system and the information fusion project,we conduct the researches on PolSAR image information representation and classification,and PolSAR image information visualization based on deep learning.Most deep learning based supervised PolSAR image classification methods can only classify a neighborhood window data to one class.The pixels in a neighborhood window may have different land cover types.The algorithm of outputting a 2-D classification map with Convolutional Neural Networks(CNN)is introduced to the PolSAR image classification.We use a Fixed-Feature-Size CNN(FFS-CNN),which can use the interrelation information representation of land cover types,to classify all pixels in a neighborhood window simultaneously.Experiments on a variety of PolSAR images,such as Gaofen-3,show that proposed method has faster classification speed,stronger generalization ability and higher classification accuracy.Furthermore,in order to make full use of hidden features of PolSAR data in rotation domain,Convolutional Long Short Term Memory networks(Conv LSTM)is used to learn the hidden information representation of the polarization coherent matrix sequence and the PolSAR image classification performance is further improved.It is difficult to obtain a large number of accurately labeled samples of PolSAR images,so unsupervised image classification is important for PolSAR image interpretation.Traditional machine learning based unsupervised classification methods have cumbersome pipelines.A method that uses the deep mutual information representation for unsupervised PolSAR image classification is proposed.The neighborhood window data of a pixel in a PolSAR image and its geometry transformed version are used to generate a paired sample as the input of a CNN.The deep features are used to compute deep mutual information representation,which can guide the CNN model to learn common information,and discard instance-specific details.Experiments on three real PolSAR datasets show that proposed method can achieve good unsupervised classification results and is superior to state-of-art methods.Proposed method is trained in an end-to-end manner and do not require extra preprocessing and post-processing.The deep features extracted by the CNN model also can be used to improve the performance of semi-supervise classification.Due to the imaging mechanism of SAR and the speckle noise,untrained people are difficult to recognize the land cover types visually from SAR images.A Supervised Cycle-Consistent Adversarial Network(S-Cycle GAN)is proposed to translate SAR images to the form of optical remote images and improve the SAR image visualization.When the optical RS data are unavailable or partly unavailable,the generated optical images can be alternative data that aid in land cover visual recognition for untrained people.To further promote the development of deep learning on remote sensing(RS),a paired SAR-optical image dataset based on Sentinel-1 and Sentinel-2 images is published and then used to evaluate proposed method.Proposed method can keep both the land cover and structure information well,and its performance is superior to some famous image-to-image translation models.The generated image can be used for optical RS image cloud removal and make up the optical RS image.In summary,the deep features and information representation of SAR images are studied based on deep learning.The fully supervised and unsupervised PolSAR image classification methods based on CNNs are developed.The SAR-to-optical image translation method based on S-Cycle GAN is also developed.All the methods achieve excellent performance on Gaofen-3 PolSAR images and the SAR-to-optical image translation method has already applied to the information fusion project.更多还原