[4Xin1] Poster session 2

[4Xin1-30] Development of an automatic solar filament detection method using deep learning

〇Tomoya Sato¹, Yusuke Iida¹, Shuichi Ando¹, Akira Sasaki¹ (1.Niigata University)

Keywords:Space Weather, Deep Learning, Object Detection

The eruption of solar filament, which consists of gas masses, can cause the serious problems such as satellite and GPS operations when they reach the Earth's upper atmosphere. This is an important issue from the perspective of space weather. On the other hand, the automatic detection methods based on deep learning have been developed, but a fragmentation problem, a single filament is divided into multiple filaments, is reported. In this study, we aimed to solve this fragmentation problem by the hyperparameters tuning and the data modification. Hyperparameters were tuned to allow detection according to filament geometry, and data modification increased the percentage of large filaments prone to eruptions and excluded those that were improperly annotated. As a result, the fragmentation problem is partially solved for some cases and the the evaluation score, Average Precision, also increases from 0.322 to 0.480 with the improvements in this study.