17:15 〜 18:45
[PEM11-P03] Comparison of simple U-Net model and U-Net models with Attention mechanism in detecting solar coronal hole regions from magnetic field images.
In this study, the Encorder-Decorder model is used to detect coronal hole regions from magnetic field images only, and the detection results provide insight into the formation mechanism of the magnetic field structure of coronal holes. The formation of coronal holes is thought to be greatly affected by variations in the magnetic field on the solar surface, but the detailed formation mechanism has not been clarified because the characteristics of the magnetic field structure are too fine to be determined visually. Therefore, this study aims to elucidate the formation mechanism by constructing a predictive detection model based only on magnetic field images using deep learning. As for the detection of coronal holes themselves, models focusing on ultraviolet images that clearly capture coronal holes have been developed and reported. However, there is no precedent for the development of a detection model based only on magnetic field images, which is planned in this study, other than the master's thesis by Takebe et al. In their master's thesis, Takebe et al. achieved TSS=0.587 using U-Net, which has a simple structure. In this study, we aim to improve the detection accuracy by applying a model with an attention mechanism, which has recently been pointed out for its usefulness in image data. So far, image generation models with Attention mechanism such as Attention U-Net and Trans U-Net, and even hyper-parameter adjustment of the U-Net model proposed by Takebe et al. have been performed and statistically evaluated. As a result, the Attention mechanism did not function as effectively as expected, and the U-Net model with adjusted hyperparameters achieved the highest detection accuracy of TSS=0.618±0.054. As for future research, first, since there are still many over-detections in the detection results, the hyper-parameters will be adjusted for a while longer with a view to suppressing over-detections in the U-Net model, which had the highest detection accuracy. Next, we will attempt to elucidate the coronal hole magnetic field structure from the detection results. In particular, by using XAI techniques such as Grad-CAM, we will find features that the model focuses on that are not captured visually.