[P-EM11] Space Weather and Space Climate

[PEM11-P14] Automatic Detection of Sigmoidal Structures in Solar X-ray Images Using Machine Learning Models

*Hibiki Iwanaga¹, Satoshi Nakahira², Yusuke Iida³ (1.Osaka Metropolitan University, 2.ISAS/JAXA, 3.Niigata University)

Keywords:Space Weather, Coronal Mass Ejection, Machine-Learning, X-ray

A coronal mass ejection is a massive release of plasma in the solar corona, which can cause various problems such as communication failures, satellite malfunctions, and even radiation exposure to astronauts. To minimize these effects, it is extremely important to predict coronal mass ejections in advance.In many cases, large-scale coronal mass ejections in active solar regions are accompanied by a precursor phenomenon called a sigmoid structure, which is an S-shaped twisted magnetic field structure.Therefore, it is very important to detect the sigmoid structure. However, the detection process currently requires manual analysis, which is problematic in terms of real-time performance.In this study, we developed a machine learning model to automatically detect sigmoid structures from X-ray image data acquired by the solar observing satellite "Hinode".The model is based on the Vision Transformer, which is an application of the Transformer, which has been effective in the field of natural language processing, to image processing. The Vision Transformer divides an image into patches, transforms each patch linearly, and embeds the transformed image as input. It is reported that this improves computational efficiency and enables pre-training with a huge amount of data, resulting in better performance than conventional machine learning models.To create the training dataset for the model construction, 3000 images that captured the sigmoid structure and 3000 images that did not capture the sigmoid structure were extracted by matching Hinode's X-ray telescope observation data from 2016 to 2019 with the Heliophysics Events Knowledgebase.The test data were also extracted in the same way, 450 images each, and applied to the constructed model. As a result, the recall was 0.813 and the precision was 0.723 against the test data, demonstrating the potential of machine learning for automatic sigmoid structure detection.At this conference, we would also like to discuss specific policies for visualization of gaze region using the Attention mechanism.