测绘通报 ›› 2026, Vol. 0 ›› Issue (10): 34-41,60.doi: 10.13474/j.cnki.11-2246.2026.1005

• 学术研究 • 上一篇    

基于DeepSeek的遥感影像目标实体关系抽取方法

颜秋宇1, 王曙2,3,4, 华一新1, 张江水1, 曹一冰1,5   

  1. 1. 信息工程大学地理空间信息学院, 河南 郑州 450002;
    2. 中国科学院地理科学与资源研究所地理信息科学与技术全国重点实验室, 北京 100101;
    3. 中国科学院大学, 北京 100101;
    4. 江苏省地理信息资源开发与利用协同创新中心, 江苏 南京 210023;
    5. 航空电子综合与体系集成全国重点实验室, 上海 200241
  • 收稿日期:2026-04-14 发布日期:2026-11-06
  • 作者简介:颜秋宇(1995—),男,博士生,主要研究方向为地理人工智能与遥感目标智能解译。E-mail:yanqiuyu1@gmail.com
  • 基金资助:
    国家自然科学基金(42471503);广西重点研发计划(桂科FN2504240012);国家重点研发计划(2022YFB3904201;2022YFF0711601);中国科学院基础与交叉前沿科研先导专项(XDB0740200);航空电子综合与体系集成全国重点实验室开放基金(2025AIASS0601)

The method of target entity relationship extraction from remote sensing images based on DeepSeek

Yan Qiuyu1, Wang Shu2,3,4, Hua Yixin1, Zhang Jiangshui1, Cao Yibing1,5   

  1. 1. Institute of Geospatial Information, Information Engineering University, Zhengzhou 450002, China;
    2. State Key Laboratory of Resources and Environmental Information System, Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences, Beijing 100101, China;
    3. University of Chinese Academy of Sciences, Beijing 100101, China;
    4. Jiangsu Center for Collaborative Innovation in Geographical Information Resource Development and Application, Nanjing 210023, China;
    5. National Key Laboratory of Avionics Integration and Aviation System-of-Systems Synthesis, Shanghai 200241, China
  • Received:2026-04-14 Published:2026-11-06

摘要: [目的] 遥感影像目标识别和知识图谱技术已成为实景三维数据更新与应用的重要技术方法之一,本文旨在解决遥感影像目标实体间关联关系的抽取过程中存在的自动化水平低、数据标注量大,以及关系模型和分类体系不统一等问题。[方法] 本文提出基于DeepSeek大模型的遥感影像目标实体关系抽取技术框架,通过关系模型、类别体系构建及基于提示工程的遥感影像实体关系数据生产,完成遥感影像实体知识图谱的构建。为检验方法的有效性,在3392张高分辨率遥感影像数据集上开展试验验证,围绕典型复杂地表目标实体关系抽取任务进行对比分析。[结果] 结果显示,本文方法在主流大模型上整体关系抽取F1值均达到0.85以上。[结论] 本文方法能够快速适配不同遥感场景,操作灵活便捷,并为遥感识别、地理问答等任务提供结构化知识支撑,为遥感影像目标关系数据生产与知识图谱构建提供结构化支撑。

关键词: 遥感目标识别, 实体关系, 知识图谱, DeepSeek, 关系抽取

Abstract: [Purposes] Remote sensing image target recognition and knowledge graph technology have become one of the important technical methods for the update and application of real-scene 3D data.However,there are still problems such as low automation level,large amount of data annotation,and inconsistent relationship models and classification systems in the extraction of the association relationships between remote sensing image target entities. [Methods] This paper proposes a technical framework for extracting the relationship between remote sensing image target entities based on the DeepSeek large model.Through the construction of relationship models and category systems and the production of remote sensing image entity relationship data based on prompt engineering,the construction of remote sensing image entity knowledge graphs is completed.To verify the effectiveness of the proposed method,experiments were conducted on a dataset of 3392 high-resolution remote sensing images,focusing on relation extraction for typical complex surface targets. [Findings] The results show that the overall F1 value of the relationship extraction of this method on mainstream large models is above 0.85. [Conclusions] This method can quickly adapt to different remote sensing scenarios,is flexible and convenient to operate,and provides structured knowledge support for tasks such as remote sensing recognition and geographic question answering,promoting the transformation of remote sensing intelligent interpretation from the “feature perception” paradigm to the “knowledge reasoning”paradigm.

Key words: remote sensing target recognition, entity relationship, knowledge graph, DeepSeek, relationship extraction

中图分类号: