摘要
【目的】遥感细粒度目标识别是对地观测及计算机视觉领域中的一项核心且极具挑战的任务,其聚焦高空间分辨率图像中对象的定位及精细分类。当前遥感细粒度目标识别算法突破了对像素级、对象级与邻域级不同层次的对象语义特征、纹理像素特征及空间邻域特征等多源多尺度特征的协同综合,但仍然无法直接利用场景构成、实体内涵、特征描述与时序变化等细粒度的目标识别相关特征,其原因在于其缺乏形式化的知识组织与表达方法。【方法】本文提出了一种面向遥感细粒度目标识别的多层级知识图谱组织与表达方法:通过设计场景、实体、特征与变化的四层知识表达框架,采用时空参照、空间形态、关联关系等特征对遥感目标进行多粒度动态化描述,实现了场景约束、实体约束、特征约束、时序约束下的多层级遥感细粒度目标识别知识的组织与表达。【结果】本文提出的多层级知识图谱方法能够有效组织和表达场景、实体及时序等知识,并有效助力细粒度目标识别性能的提升。其中,在基准模型STD上引入知识图谱后,整体mAP提升了约3.82%,recall提升了约3.92%;并通过在多种典型神经网络(Oriented R-CNN、Oriented RepPoints、LSKNet、STD)模型上均实现稳定的性能提升,验证了方法的普适性与鲁棒性。【结论】多层级知识图谱不仅提升了遥感细粒度目标识别的精度,还增强了模型的可解释性和动态适应性,能够为遥感智能解译提供从特征感知向知识推理转变的有效途径,在地理信息分析、军事情报分析等领域具有重要应用价值。
[Objectives]Fine-grained object recognition in remote sensing is a fundamental yet highly challenging task within both Earth observation and computer vision.It involves the accurate localization and detailed classification of objects in High-Spatial-Resolution(HSR)imagery,which often features highly complex backgrounds,inter-class similarities,and intra-class variations.In recent years,notable progress has been driven by algorithms that jointly exploit pixel-level,object-level,and neighborhood-level information.These approaches combine semantic features,texture characteristics,and spatial contextual relationships to form multi-source and multi-scale feature representations.Despite these advances,existing methods remain inadequate for directly utilizing higher-level fine-grained knowledge such as scene composition,entity semantics,attribute descriptions,and temporal dynamics.The core limitation lies in the absence of a formalized knowledge organization and representation paradigm capable of systematically bridging low-level visual perception and higher-order semantic reasoning.[Methods]To address these limitations,this study proposes a multi-level knowledge graph-based organization and representation framework specifically designed for fine-grained remote sensing object recognition.The framework adopts a four-layer hierarchical structure encompassing scene,entity,feature,and change dimensions,enabling dynamic and semantically rich descriptions of remote sensing targets.In this structure,scene nodes provide contextual constraints,entity nodes capture essential connotations of objects,feature nodes encode visual and semantic attributes,and change nodes represent temporal evolution.[Results]By incorporating spatiotemporal references,spatial morphology,and inter-object relationships,the proposed approach enables knowledge organization under multiple constraints,including scene,entity,feature,and temporal conditions.In doing so,it moves beyond purely data-driven perception and establishes a mechanism for knowledge-driven reasoning in remote sensing interpretation.Extensive experiments were conducted to validate the effectiveness of the proposed framework.When integrated into the baseline model STD,the knowledge graph yielded an improvement of approximately 3.82%in mean Average Precision(mAP)and 3.92%in recall,demonstrating its ability to enhance detection accuracy.Beyond this single case,the universality and robustness of the framework were confirmed by consistent performance improvements across several representative neural networks,including Oriented R-CNN,Oriented RepPoints,LSKNet,and STD.These results indicate that the proposed method not only improves recognition performance but also enhances interpretability and adaptability across heterogeneous architectures and datasets.[Conclusions]Overall,this study demonstrates that a multi-level knowledge graph provides an effective pathway for advancing fine-grained object recognition in remote sensing,transitioning from feature perception to knowledge reasoning.The method not only increases recognition accuracy but also enhances semantic interpretability and dynamic adaptability,offering a scalable solution for intelligent remote sensing analysis.Importantly,it provides new theoretical and practical insights for applications in geospatial information extraction,environmental and urban monitoring,disaster assessment,and military intelligence analysis.By systematically integrating structured knowledge with data-driven models,the proposed framework enriches the semantic depth of remote sensing interpretation and demonstrates strong potential for future developments in intelligent Earth observation systems.
作者
颜秋宇
王曙
华一新
张江水
YAN Qiuyu;WANG Shu;HUA Yixin;ZHANG Jiangshui(Information Engineering University,Institute of Geospatial Information,Zhengzhou 450002,China;State Key Laboratory of Geographic Information Science and Technology,Institute of Geographic Sciences and Natural Resources Research,Chinese Academy of Sciences,Beijing 100101,China;University of Chinese Academy of Sciences,Beijing 100049,China;Jiangsu Center for Collaborative Innovation in Geographical Information Resource Development and Application,Nanjing 210023,China)
出处
《地球信息科学学报》
北大核心
2025年第12期2833-2849,共17页
Journal of Geo-information Science
基金
国家重点研发计划项目(2021YFB3900900、2022YFB3904201)
国家自然科学基金项目(42471503)
中国科学院基础与交叉前沿科研先导专项项目(XDB0740200-01)。
关键词
遥感知识图谱
细粒度目标识别
遥感场景
时空实体
时空知识图谱
多粒度时空对象
remote sensing knowledge graph
fine-grained target recognition
remote sensing scene
spatiotemporal entity
spatiotemporal knowledge graph
multi-granularity spatiotemporal object
