极地研究

极地研究 ›› 2025, Vol. 37 ›› Issue (3): 427-436.DOI: 10.13679/j.jdyj.20240016

• 研究论文 • 上一篇    下一篇

基于Swin-Transformer的太空台风识别模型

乔枫1,3, 张清和1,2, 邢赞扬1, 王勇1, 马羽璋1, 陆盛1, 张红波4, 王飞飞4   

  1. 1山东大学空间科学研究院, 山东 威海 264209;
    2中国科学院国家空间科学中心, 空间天气学国家重点实验室, 北京 100190; 
    3山东大学信息科学与工程学院, 山东 青岛 266237; 
    4中国电波传播研究所, 电波环境特性及模化技术重点实验室, 山东 青岛 266107

  • 收稿日期:2024-02-01 修回日期:2024-04-24 出版日期:2025-09-30 发布日期:2025-09-25
  • 通讯作者: 张清和
  • 基金资助:
    国家自然科学基金、山东省自然科学基金、小米青年学者项目和北极黄河地球系统国家野外科学观测研究站开放研究基金资助

A Swin Transformer-based space hurricane identification model

QIAO Feng1,3, ZHANG Qinghe1,2, XING Zanyang1, WANG Yong1, MA Yuzhang1, LU Sheng1, ZHANG Hongbo4, WANG Feifei4   

  1. 1Institute of Space Sciences, Shandong University, Weihai 264209, China;
    2State Key Laboratory of Space Weather, National Space Science Center, Chinese Academy of Sciences, Beijing 100190, China; 
    3School of Information Science and Engineering, Shandong University, Qingdao 266237, China;
    4National Key Laboratory of Electromagnetic Environment, China Research Institute of Radiowave Propagation, Qingdao 266107, China
  • Received:2024-02-01 Revised:2024-04-24 Online:2025-09-30 Published:2025-09-25

PDF (PC)

52

摘要： 太空台风是发生于地磁平静期极盖区内的一种涡旋状极光亮斑结构, 可由局地堪比磁暴的太阳风能量注入引起。为解决从海量星载极光数据中准确而高效地识别出太空台风事件, 进而研究太阳风能量注入过程等关键问题, 本文基于Swin Transformer构建了可用于美国国防气象卫星星载紫外光谱成像仪极光图像的太空台风识别模型。该模型通过拆分窗口加速计算, 并使用移动窗口多头自注意力方法搭建窗口间信息传输通道, 从而实现太空台风的自动识别。研究表明, 该模型通过南北半球极盖区的太空台风图像数据集进行训练, 实现了对太空台风事件更加精准的识别, 其准确率高达95.94%。

关键词: Swin-Transformer, 太空台风, 识别模型, 深度学习

Abstract: A space hurricane is a vortex-like auroral bright-spot structure that occurs in the polar cap region during quiet geomagnetic periods and is caused by the local injection of a large amount of solar wind energy into the polar ionosphere, comparable to a magnetic storm. Achieving accurate and effective identification of space hurricane events from a large amount of auroral data is essential for studies of solar wind energy injection. In this paper, a Swin Transformer model, which is used to identify space hurricanes from Defense Meteorological Satellite Program/Special Sensor Uleraviolet Spectrographic Imager (DMSP/SSUSI) images, is constructed. This model improves computation time using splitting windows and establishes inter-window information transfer channels using the Shifted Window Multi-Head Self-Attention (SW-MSA) method, achieving automatic identification of space hurricanes. The study demonstrates that the model trained using a dataset consisting of space hurricanes in the northern and southern hemispheres identifies space hurricane events more accurately. The accuracy of the Swin Transformer-based space hurricane identification model is 95.94%.

Key words: Swin Transformer, space hurricane, identification model, deep Learning