数据与计算发展前沿 ›› 2025, Vol. 7 ›› Issue (4): 143-154.

CSTR: 32002.14.jfdc.CN10-1649/TP.2025.04.012

doi: 10.11871/jfdc.issn.2096-742X.2025.04.012

• 专刊：空间科学大数据智能算法模型与工具 • 上一篇    下一篇

基于PatchCore的科学卫星图像异常检测算法研究

王磊1,2(),马福利1,2,*(),于勤思1,2,魏明月1,2   

  1. 1.中国科学院国家空间科学中心， 空间大数据技术研究室， 北京 100190
    2.国家空间科学数据中心， 北京 100190
  • 收稿日期:2025-04-29 出版日期:2025-08-20 发布日期:2025-08-21
  • 通讯作者: 马福利
  • 作者简介:王磊，中国科学院国家空间科学中心，助理工程师，主要研究方向为计算机视觉、大数据处理技术、大模型应用。
    本文承担工作为：模型算法实现、论文撰写。
    WANG Lei is an assistant engineer at the National Space Science Center, Chinese Academy of Sciences. His research interests include computer vision, big data processing technologies, and the application of large language models.
    In this study, he is primarily responsible for the implementation of the anomaly detection algorithms and paper writing.
    E-mail: wanglei01@nssc.ac.cn|马福利，中国科学院国家空间科学中心，高级工程师，主要研究方向为大数据处理与管理技术、科学卫星地面系统设计。
    本文承担工作为：指导模型设计、模型优化和论文写作。
    MA Fuli is a Senior Engineer at the National Space Science Center, Chinese Academy of Sciences. His research interests include big data processing and management technologies, as well as the design of ground systems for scientific satellites.
    In this study, he is primarily responsible for providing guidance on model design, model optimization, and paper writing.
    E-mail: mafuli007@nssc.ac.cn
  • 基金资助:
    中国科学院战略性先导科技专项(XDA15040200)

Research on Anomaly Detection Algorithms for Scientific Satellite Images Based on PatchCore

WANG Lei1,2(),MA Fuli1,2,*(),YU Qinsi1,2,WEI Mingyue1,2   

  1. 1. National Space Science Center, CAS, Research Laboratory for Spatial Big Data Technology, Beijing 100190, China
    2. National Space Science Data Center, Beijing 100190, China
  • Received:2025-04-29 Online:2025-08-20 Published:2025-08-21
  • Contact: MA Fuli

RichHTML

5

PDF

8

摘要：

【目的】随着空间科学卫星观测数据量的急剧增长，图像异常检测作为保障数据质量与支撑科学研究的重要环节，亟需发展高效、自动化的方法。【应用背景】由于卫星运行初期异常样本稀缺甚至缺失，传统有监督学习方法难以直接应用，因此本文基于无监督学习范式，提出了一套面向空间科学卫星图像的异常检测方法。【方法】以PatchCore算法为核心，本文设计了特征提取、核心集构建、异常评分与图像分类等模块，并结合统计学和聚类分析方法，探索了多种异常分数和阈值设定策略，提升了检测的灵敏度与稳定性。【结果】针对实际太阳观测图像数据集，本文开展了充分的实验验证，并与PaDiM、CS-Flow等主流无监督检测方法进行了对比分析，结果表明本文方法在AUROC、AUPR等指标上均取得了优异性能，分别达到了0.9996和0.9999。系统实现方面，采用ONNX Runtime轻量化部署模型，有效提升了推理速度与部署灵活性，并完成了数据获取、异常检测与预警反馈的全流程闭环。【结论】研究结果表明，该系统可有效提升空间科学任务中的图像质量监控效率，对未来空间观测任务中智能异常检测系统的建设具有重要参考价值。

关键词: 无监督学习, PatchCore, 空间科学卫星, 图像异常检测

Abstract:

[Objective] With the rapid growth of observational data from space science satellites, image anomaly detection has become a critical component for ensuring data quality and supporting scientific research, which requires urgently the development of efficient and automated methods. [Context] Due to the scarcity or even absence of anomaly samples during the early stages of satellite operation, traditional supervised learning approaches are difficult to apply directly. Therefore, this study proposes an anomaly detection framework tailored for space science satellite images based on unsupervised learning. [Methods] Based on the PatchCore algorithm, the proposed method designs modules for feature extraction, core-set construction, anomaly scoring, and image classification. In addition, multiple anomaly scoring and threshold setting strategies based on statistical analysis and clustering methods are explored to enhance detection sensitivity and robustness. [Results] Extensive experiments were conducted on actual solar observation datasets, with comparative analysis against mainstream unsupervised detection methods such as PaDiM and CS-Flow. The results demonstrate that the proposed method achieves outstanding performance, with AUROC and AUPR values reaching to 0.9996 and 0.9999, respectively. In terms of system implementation, ONNX Runtime is adopted for lightweight model deployment, which significantly improves inference speed and deployment flexibility, and establishes a complete closed-loop process covering data acquisition, anomaly detection, and alert feedback. [Conclusions] The study shows that the developed system can effectively enhance the efficiency of image quality monitoring in space science missions, providing valuable references for the construction of intelligent anomaly detection systems in future space observation tasks.

Key words: unsupervised learning, PatchCore, space science satellites, image anomaly detection