首页 >  , Vol. , Issue () : -

摘要

全文摘要次数: 154 全文下载次数: 80
引用本文:

DOI:

10.11834/jrs.20243547

收稿日期:

2023-12-26

修改日期:

2024-04-10

PDF Free   EndNote   BibTeX
知识与数据驱动的遥感图像智能解译:进展与展望
孟瑜, 陈静波, 张正, 刘志强, 赵智韬, 霍连志, 史科理, 刘帝佑, 邓毓弸, 唐娉
中国科学院空天信息创新研究院 国家遥感应用工程技术研究中心
摘要:

知识与数据是贯穿遥感图像解译数十年发展历程的两大要素。随着传感器平台的不断丰富,以及深度学习、大数据、多模态、长时序解译方法的快速突破,数据驱动的智能解译成为了近年来的热点研究方向。然而在不断深入扩展的研究与应用中,数据驱动方法迁移复用难、样本依赖强、可解释性弱等局限开始显露。在长期解译实践中积累的各类知识具有客观实在性、确定性、场景适应性、解释推理性等特点,可以与数据驱动的方法互为补充,知识与数据双驱动正逐渐成为遥感图像解译的新方向。本文首先回顾了遥感图像解译发展的几个主要阶段以及知识和数据在各个阶段分别发挥的作用,继而总结了十四类遥感图像解译涉及的主要知识类型。知识与深度学习的融合是实现知识与数据双驱动的重要路径,本文梳理了五大类十五小类知识与深度神经网络的融合方法并例举了相关案例。以知识类型为主要脉络,本文进一步对现有知识与数据联合的遥感解译应用进行了综述,通过典型案例分析了效益能力增量。最后本文对知识与数据联合驱动的遥感图像智能解译框架及关键技术进行了展望。

Knowledge and Data Driven Remote Sensing Image Interpretation: Recent Developments and Prospects
Abstract:

Knowledge and data are the two main elements that have characterized the development of remote sensing image interpretation for decades. With the continuous enrichment of sensor platforms and rapid breakthroughs in deep learning, big data, multi-modal, and long time-series methodologies, data-driven intelligent remote sensing image interpretation has become a hot research direction in recent years. However, in the deepening and expanding research and applications, the limitations of data-driven methods such as difficult reuse between different scenarios, strong training sample dependence, and weak interpretability are beginning to emerge. Various types of knowledge accumulated in the long-term remote sensing image interpretation practice have the characteristics of objective reality, certainty, scene adaptability, interpretability, etc., which can be complemented with data-driven approaches, and the dual-driven of knowledge and data is becoming a new direction of remote sensing image interpretation. This paper first reviews the major stages in the development of remote sensing image interpretation and the respective roles of knowledge and data in each of these stages. Then the main types of knowledge involved in remote sensing image interpretation are summarized and categorized into fourteen types. The fusion of knowledge and deep learning is an important path to achieve the dual-drive of knowledge and data, and this paper summarizes five categories and fifteen subcategories of knowledge and deep neural network fusion methods with relevant cases. From the perspective of knowledge types, this paper further provides an overview of existing applications of remote sensing interpretation with joint knowledge and data. The effectiveness and capability increment of fusing knowledge and data is demonstrated by the analyses of typical examples. Lastly, this paper gives a systematic prospect on the framework and key techniques for knowledge and data compound driven remote sensing image interpretation.

本文暂时没有被引用!

欢迎关注学报微信

遥感学报交流群