Coronal Loop Detection Using Multiscale Convolutional Neural Networks

Abstract

Solar magnetic fields play an important role in many solar activities, such as the solar wind, coronal mass ejections, and coronal oscillation. Coronal loops are curvilinear structures in the solar atmosphere and are closely related to coronal magnetic fields, so the study of their structure is very important. However, it is difficult to identify coronal loops accurately because of the complexity of their features. Therefore, we propose a two-stage detection method, using multiscale convolutional neural networks, to identify coronal loops. The regions including initial coronal loops are first marked by a improved Res-UNet model. The loop structures in the region are then detected using a improved dense extreme inception network for edge detection model. We selected the coronal images observed by the Transition and Coronal Explorer and the Atmospheric Imaging Assembly of the Solar Dynamics Observatory in the 171 Å channel to illustrate the detection processing. Meanwhile, we also compared the accuracy of our method to others. The results demonstrate that our proposed method has a high recognition rate and good robustness over previous identification methods and can be used to study the physical characteristics of coronal loops.