测绘学报 ›› 2023, Vol. 52 ›› Issue (7): 1074-1089.doi: 10.11947/j.AGCS.2023.20220499

• 高光谱遥感技术专刊 • 上一篇    下一篇

高/多光谱遥感图像超分辨率融合研究进展与展望

张兵1,2, 高连如1,3, 李嘉鑫1,2,3, 洪丹枫1,3, 郑珂1,3   

  1. 1. 中国科学院空天信息创新研究院, 北京 100094;
    2. 中国科学院大学资源与环境学院, 北京 100049;
    3. 中国科学院计算光学成像技术重点实验室, 北京 100094
  • 收稿日期:2022-08-22 修回日期:2023-06-20 发布日期:2023-07-31
  • 作者简介:张兵(1969-),男,博士,研究员,研究方向为高光谱遥感与遥感大数据。E-mail:zb@radi.ac.cn
  • 基金资助:
    国家重点研发计划(2021YFB3900502)

Advances and prospects in hyperspectral and multispectral remote sensing image super-resolution fusion

ZHANG Bing1,2, GAO Lianru1,3, LI Jiaxin1,2,3, HONG Danfeng1,3, ZHENG Ke1,3   

  1. 1. Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing 100094, China;
    2. College of Resources and Environment, University of Chinese Academy of Sciences, Beijing 100049, China;
    3. Key Laboratory of Computational Optical Imaging Technology, Chinese Academy of Sciences, Beijing 100094, China
  • Received:2022-08-22 Revised:2023-06-20 Published:2023-07-31
  • Supported by:
    The National Key Research and Development Program of China (No. 2021YFB3900502)

摘要: 高光谱图像作为多模态遥感数据的重要组成部分,能够捕捉地物精细的光谱特征。由于成像机理的限制,空间细节的损失导致高光谱图像的空间表征能力有所退化,一定程度上限制了数据进一步应用的潜力。数据融合是解决空间/光谱分辨率矛盾的有效手段,近年来相关理论得到了深入发展。本文全面综述了高/多光谱遥感图像超分辨率融合领域的研究进展与展望。首先,将当前方法系统性地分为细节注入、模型优化及深度学习3大类方法,并对不同方法的原理、模型、代表性算法进行了回顾,重点介绍了模型优化中的矩阵分解、张量表示及深度学习中的监督与非监督方法。在此基础上,梳理了该领域技术在像素级分类、目标提取、在轨融合领域的成功应用案例,指出融合产品的潜能在后续遥感应用未被充分挖掘的现状;然后,从退化模型、数据-模型驱动、多任务一体化及应用耦合4个角度对该领域进行展望;最后,就该领域的研究现状与发展趋势进行总结,归纳各类方法优劣势的同时,点明了多类方法协同、外部数据辅助及实际应用驱动等方面的重要性。

关键词: 多模态, 高光谱, 多光谱, 数据融合

Abstract: As an important component of multimodal remote sensing data, hyperspectral images are capable of capturing the fine spectral profiles of objects. However, due to the limitations of the imaging mechanism, the loss of spatial details leads to degradation of the spatial representation capability of hyperspectral images, which to a certain extent limits the potential for further applications. Data fusion is an effective approach to resolve the contradiction in spatial and spectral domains, where related theories have intensively developed in recent years. This paper provides a comprehensive overview of the advances and prospects in hyperspectral and multispectral remote sensing image super-resolution fusion. Firstly, fusion algorithms are systematically introduced and classified into three categories, namely, detail injection-based, model optimization-based, and deep learning-based methods. The principles, models and representative algorithms of different methods are reviewed, with emphasis on matrix decomposition, tensor representation in model optimization-based methods, and supervised and unsupervised methods in deep learning-based methods. On this basis, successful applications of the techniques in the field of pixel-level classification,target extraction,and on-board fusion are given, finding that the potential of fusion products has not been fully exploited in subsequent applications. Then, prospective directions are discussed from four aspects, including degradation models, data-model-driven fusion, multi-task integration fusion and application-coupled fusion. Finally, the current process and prospects of the future development trend in this field are summarized, pointing out the strengths and weaknesses of various approaches while highlighting the importance of multi-approach association, external data assistance, application-driven fusion, et al.

Key words: multimodal, hyperspectral, multispectral, data fusion

中图分类号: