Research on Mount Wilson Magnetic Classification Based on Deep Learning

The Mount Wilson magnetic classification of sunspot groups is thought to be meaningful forecast flares’ eruptions. In this paper, we adopt a deep learning method, CornerNet-Saccade, perform the groups. It includes three stages, generating object locations, detecting objects, and merging detections. key technologies consist backbone as Hourglass-54, attention mechanism, points’ mechanism including top-left corners bottom-right by corner pooling layers. These improve efficiency objects without sacrificing accuracy. A dataset built total 2486 composited images which are with continuum corresponding magnetograms from HMI MDI. After training network, in solar full image detected classified 3 seconds on average. test results show that method has good performance, accuracy, precision, recall, mAP 0.94, 0.93, 0.90, respectively. Moreover, flare productivities different types 2011 2020 calculated. As I tot id="M2"> ≥ 1, id="M3"> α , β γ δ , id="M4"> 0.14, 0.28, 0.61, 0.71, 0.87, id="M5"> id="M6"> 10, 0.02, 0.07, 0.27, 0.45, 0.65, means id="M7"> id="M8"> indeed very closely related eruption flares, especially id="M9"> type. Based reliability classified, detailed data shared website (https://61.166.157.71/MWMCSG.html).