JpGU-AGU Joint Meeting 2020

Presentation information

[E] Poster

P (Space and Planetary Sciences ) » P-EM Solar-Terrestrial Sciences, Space Electromagnetism & Space Environment

[P-EM13] Dynamics of Magnetosphere and Ionosphere

convener:Aoi Nakamizo(Applied Electromagnetic Research Institute, National Institute of Information and Communications Technology), Mitsunori Ozaki(Faculty of Electrical and Computer Engineering, Institute of Science and Engineering, Kanazawa University), Akiko Fujimoto(Kyushu Institute of Technology), Yuka Sato(Nippon Institute of Technology)

[PEM13-P14] Automatic Classification of Auroral videos based on 3D-CNN

*Toru Mikuriya1, Akiko Fujimoto1, Yoshizumi Miyoshi2, Yasunobu Ogawa3, Keisuke Hosokawa4 (1.Kyushu Institute of Technology, 2.Nagoya University, 3.National Institute of Polar Research, 4.The University of Electro-Communications )

Keywords:Machine Learning, 3D-CNN, auroral movies

There are several approaches for classifying all-sky images of aurora with handcrafted methods. The recent enhanced on the fields of computer vison and machine learning research resulted from growing the environment of information processing allows to easily develop the automatic classification of aurora images into some labeled groups. The previous researchers have proposed handcrafted classification models of snapshot aurora images with aurora shapes, color features and aurora velocity (motional differences between two aurora images), etc. They work well for the type of discrete aurora, since they have fine structure of aurora shape. Now we attempt the non-handcrafted and automatic distinction of diffuse aurora such as pulsating aurora with the supervised 3D-Convolutional Neural Network (3D-CNN) technique. We use 605 aurora all-sky imager videos (each movie has 60 frames with 1 second time resolution) to extract the feature of quasi-periodic pulsation aurora motion which is the main characteristic of pulsating aurora. The supervised learning data has two categories, 1) pulsating aurora and 2) no pulsating aurora, and we divide all videos into 8:2 for learning /evaluation processes and test process, respectively. The k-Cross Validation (k=5) resulted in higher accuracy of our model (larger than 90 %). We found the classification accuracy depends on the number of 3D-CNN filter in time direction, and our model works well on the lower number than the typical frequency of pulsating aurora.