Automated Classification of Auroral Images with Deep Neural Networks
Zhiyuan Shang,
Zhonghua Yao,
Jian Liu,
Linli Xu,
Yan Xu,
Binzheng Zhang,
Ruilong Guo,
Yong Wei
Affiliations
Zhiyuan Shang
Key Laboratory of Earth and Planetary Physics, Institute of Geology and Geophysics, Chinese Academy of Sciences, Beijing 100000, China
Zhonghua Yao
Key Laboratory of Earth and Planetary Physics, Institute of Geology and Geophysics, Chinese Academy of Sciences, Beijing 100000, China
Jian Liu
School of Nuclear Science and Technology, University of Science and Technology of China, Hefei 230026, China
Linli Xu
Anhui Province Key Laboratory of Big Data Analysis and Application, School of Computer Science and Technology, University of Science and Technology of China, Hefei 230026, China
Yan Xu
Key Laboratory of Earth and Planetary Physics, Institute of Geology and Geophysics, Chinese Academy of Sciences, Beijing 100000, China
Binzheng Zhang
Department of Earth Sciences, The University of Hong Kong, Hong Kong SAR, China
Ruilong Guo
Laboratory of Optical Astronomy and Solar-Terrestrial Environment, Institute of Space Sciences, School of Space Science and Physics, Shandong University, Weihai 264200, China
Yong Wei
Key Laboratory of Earth and Planetary Physics, Institute of Geology and Geophysics, Chinese Academy of Sciences, Beijing 100000, China
Terrestrial auroras are highly structured that visualize the perturbations of energetic particles and electromagnetic fields in Earth’s space environments. However, the identification of auroral morphologies is often subjective, which results in confusion in the community. Automated tools are highly valuable in the classification of auroral structures. Both CNNs (convolutional neural networks) and transformer models based on the self-attention mechanism in deep learning are capable of extracting features from images. In this study, we applied multiple algorithms in the classification of auroral structures and performed a comparison on their performances. Trans-former and ConvNeXt models were firstly used in the analysis of auroras in this study. The results show that the ConvNeXt model can have the highest accuracy of 98.5% among all of the applied algorithms. This study provides a direct comparison of deep learning tools on the application of classifying auroral structures and shows promising capability, clearly demonstrating that auto-mated tools can help to minimize the bias in future auroral studies.
