极地研究

极地研究 ›› 2026, Vol. 38 ›› Issue (2): 265-274.DOI: 10.13679/j.jdyj.20250075

• 研究论文 • 上一篇    下一篇

基于FA-GAN的极光卵强度建模

韩冰1, 赵玮雄1, 胡泽骏2,3,4   

  1. 1西安电子科技大学电子工程学院, 陕西 西安 710071;
    2中国极地研究中心(中国极地研究所), 极区空间物理与天文研究所, 上海 201209;
    3北极黄河地球系统国家野外科学观测研究站, 中国极地研究中心(中国极地研究所), 上海 201209;
    4浙江大学海洋研究院, 浙江 舟山 316021
  • 收稿日期:2025-11-11 修回日期:2026-01-30 出版日期:2026-06-30 发布日期:2026-07-13
  • 通讯作者: 胡泽骏
  • 基金资助:
    国家重点研发项目资助

Auroral oval intensity modeling based on FA-GAN

HAN Bing1, ZHAO Weixiong1, HU Zejun2,3,4   

  1. 1School of Electronic Engineering, Xidian University, Xi’an 710071, China;
    2Polar Atmosphere and Space Physics Laboratory, Polar Research Institute of China, Shanghai 201209, China；
    3Arctic Yellow River Earth System National Observation and Research Station, Polar Research Institute of China, Shanghai 201209, China;
    4Ocean academy, Zhejiang University, Zhoushan 316021, China
  • Received:2025-11-11 Revised:2026-01-30 Online:2026-06-30 Published:2026-07-13

PDF (PC)

156

摘要：

极光卵的极光强度能够直观反映极光粒子沉降的分布特征, 是研究磁层动力学以及空间环境的重要物理参量, 对理解磁层、电离层、热层及其耦合关系具有重要意义。研究发现, 部分空间物理参数与极光强度密切相关, 但目前并没有明确的数学关系, 而神经网络可以通过映射关系建模两者的关系。因此, 本文基于Polar卫星的紫外极光图像数据, 利用NASA OMNI数据库中与极光卵强度相关的6个空间物理参数, 构建1996年12月—1997年1月的极光卵强度数据集。利用基于文本到图像的特征感知生成式对抗网络(Feature-Aware Generative Adversarial Network, FA-GAN)构建物理参数到极光卵强度的预测模型。利用4个评价指标(KL散度、结构相似度(structure similarity, SSIM)、Inception Score(SIS)和Frechet Inception Distance(DFID))对极光强度预测模型进行性能评估。结果显示, 基于FA-GAN方法的极光强度预测模型预测结果对应的SSIM、SIS以及DFID明显优于现有的极光强度预测模型。


关键词: 紫外极光, 极光卵强度建模, 生成对抗式网络, 空间物理参数

Abstract:

The auroral intensity of the auroral oval directly reflects the distribution characteristics of auroral particle precipitation. It serves as a critical physical parameter for investigating magnetospheric dynamics and the space environment, playing a vital role in understanding the coupling mechanisms between the magnetosphere, ionosphere, and thermosphere. Recent studies have demonstrated that specific space physical parameters are closely correlated with auroral intensity; however, their explicit mathematical relationships remain poorly defined. Neural networks offer a robust alternative by modeling these complex non-linear associations through learned mapping relationships. In this study, an auroral oval intensity dataset spanning from December 1996 to January 1997 was constructed using ultraviolet auroral imagery from the Polar satellite, integrated with six key space physical parameters retrieved from the NASA OMNI database. A prediction model mapping physical parameters to auroral oval intensity was developed based on the Feature-Aware Generative Adversarial Network (FA-GAN), a text-to-image synthesis framework. The performance of the proposed model was evaluated using four metrics: Kullback-Leibler (KL) divergence, Structural Similarity (SSIM), Structural Inception Score (SIS), and Denoising Feature Inception Distance (DFID). The experimental results indicate that the FA-GAN-based prediction model significantly outperforms existing auroral intensity models in terms of SSIM, SIS, and DFID.

Key words:

ultraviolet aurora, auroral oval intensity modeling, generative adversarial network, space physics parameters