基于正则化张量表达模型的时变高光谱遥感图像变化检测研究

As a set of high-dimensional data, multitemporal hyperspectral remote sensing imagery (HSI) consists of one spectral dimension, one temporal dimension and two spatial dimensions, providing rich spectral-spatial information and multiple temporal change details. The abundance of information is essential to change detection, while also making the task challenging. To deal with major problems in HSI change detection such as curse of dimensionality, corruption of noise and lack of prior knowledge, we will make use of tensor, which is naturally advantageous in representing and analyzing high-dimensional data, and propose a framework based on regularized tensor representation model for HSI change detection. The scheme of our research is laid out as follows. (1) We study the relation between tensor model and multitemporal HSI, design suitable regularization forms, and use them to innovate regularized tensor representation models, which are expected to accurately depict the inherent data structures of change features and noise. This study is to address the issues of dimensionality reduction and denoising. (2) With the proposed models, we design effective tensor analysis methods for simultaneous feature extraction in multiple dimensions, so as to increase change detection accuracy. (3) Within our framework of tensor analysis, we conduct adaptive pattern recognition for multiple change classes, which exhibits decreased dependency on prior knowledge. The findings of this project are expected to offer novel ideas and systematic methods for change detection in multitemporal HSI, as well as promoting the values of multitemporal HSI in applied areas.

时变高光谱遥感图像是一组高维数据，具有一维光谱、一维时间、二维空间共四个数据维度，含有丰富的空谱信息和时间变化信息，为变化检测提供了基本条件，但同时也带来了挑战。为了解决维数灾难、噪声干扰、先验知识稀缺性等问题，本项目利用张量在表达和分析高维数据上的天然优势，提出基于正则化张量表达模型的时变高光谱遥感图像变化检测框架。主要研究内容如下：（1）根据时变高光谱遥感图像与张量模型的契合点，设计合理的约束项，建立新型的正则化张量表达模型，以期精确表达变化特征和噪声的内在结构，为降维和降噪的奠定基础；（2）利用所提出的模型，针对多维联合变化特征的提取，设计有效的张量分析方法，提高变化检测精度；（3）在张量分析的框架下，实现多变化种类的自适应识别，降低对先验知识的依赖。研究成果有望给时变高光谱遥感图像的变化检测提供新型的研究思路和系统的实现方案，并对时变高光谱遥感图像应用的发展起到推动作用。

本项目在张量分析框架下探究时变遥感图像变化检测理论和方法，提出多种正则化张量特征表达和判别模型，在多维变化特征表达，自监督张量网络及其在地物变化分类的应用，深层张量网络的应用推广等方面取得良好进展。项目贡献主要体现在以下两点：1）研究张量分析理论与方法，提出新型的高维数据特征表达模型，有效学习时变遥感图像的“时—空—谱”联合特征，提升变化检测精度；2）建立张量分析与神经网络的数学联系，将张量分解模型和卷积神经网络充分结合，提出新型深层张量网络和学习框架，推进机器学习在时变遥感图像分析的应用研究。结果表明，提出的正则化张量模型在多时相遥感图像地物变化检测、时变光谱云图热带气旋定强等应用上表现出优越的性能和实用价值，验证正则化张量模型在时变遥感图像分析的可行性和有效性。在研期间，负责人以一作或通讯作者共发表16篇论文（其中4篇为SCI二区及以上论文），被授权1项国家发明专利，获得上海市教委“晨光计划”资助（编号：18CG38），培养硕士研究生10余名。