极地研究 ›› 2018, Vol. 30 ›› Issue (2): 123-131.DOI: 10.13679/j.jdyj.20170038

• 研究论文 •    下一篇

基于卷积神经网络的极光图像分类

王菲, 杨秋菊   

  1. 陕西师范大学
  • 收稿日期:2017-09-29 修回日期:2017-12-06 出版日期:2018-06-30 发布日期:2018-06-30
  • 通讯作者: 杨秋菊
  • 基金资助:

    国家自然科学基金(41504122)、陕西省自然科学青年人才项目(2016JQ4001)、陕西省高校科协青年人才托举计划
    (20160211)资助

Classification of auroral images based on convolutional neural network

Wang Fei, Yang Qiuju   

  • Received:2017-09-29 Revised:2017-12-06 Online:2018-06-30 Published:2018-06-30
  • Supported by:

    National Natural Science Foundation of China

摘要:

极光是由带电粒子经磁层—电离层碰撞大气而产生的。面对形态各异、演变过程复杂的极光图像, 对
其合理分类为进一步探究日地电磁活动和能量耦合等空间物理问题奠定了基础。针对该问题, 引入深度学
习的方法, 通过卷积神经网络模型自主表征极光特征并实现极光图像分类。该方法对2003 年北极黄河站越
冬观测的38 044 幅和8 001 幅典型极光图像分类正确率达93.17%和91.5%; 自动识别2004—2009 年观测数
据的极光形态, 4 类极光时间分布规律与三波段激发谱能量分布基本一致。实验结果表明, 基于卷积神经网
络的极光表征方法, 能有效实现极光图像的自动分类。

关键词: 极光, 卷积神经网络, 分类

Abstract:

Auroral light is the result of charged particles interacting with the magnetosphere and ionosphere.
Proper classification of complex morphological all-sky auroral images is meaningful for studying the relationship
between electromagnetic activity and energy coupling. To address these issues, a method of deep
learning based on a convolutional neural network was proposed to explore the feature space of auroral data
and to achieve automatic auroral recognition. The representation method was used in automatic recognition
of four primary categories of aurora observed in 2003 at the Yellow River Station. The supervised classification
accuracy rates on labeled data between dataset1 and dataset2 were 93.17% and 91.5%, respectively. The
occurrence distributions of the four categories obtained through automatic classification of data from
2004–2009, were consistent with the spectral energy distribution excited by three bands. The experimental
results showed that the presented representation method is effective for automatic auroral image recognition.

Key words: aurora, convolutional neural network, classification