Sunspot Groups Detection and Classification on SDO/HMI Images using Deep Learning Techniques

Sunspot groups detection, classification and tracking is a key component in many space weather forecasting/nowcasting applications. Reliable knowledge on current active regions is a crucial factor that can influence a trustworthy space weather prediction. In this paper, we propose a method where a pipeline of state-of-the-art object detection and classification deep neural networks (DNNs) is used to detect and classify sunspots groups in real time. We use the McIntosh system for sunspot groups classification. Our method classifies a full-disk image under the cadence of 45s, the time of acquisition of the Helioseismic and Magnetic Imager (HMI) onboard the Solar Dynamics Observatory (SDO). Additionally, this may allows scientists to automatically analyze the most dynamic phases of sunspot groups evolution, like the first stages of a group emergence, when it grows and change group class very quickly. The dataset used is produced efficiently, combining solar data from SDO/HMI images and from National Oceanic and Atmospheric Administration (NOAA/SWPC) Solar Region Summary (SRS) tables for annotations. The generation process is carried out on the overall HMI data available on the timespan that goes from 2010 to 2021 with a sampling time of 1 day, according to the availability of SRS Tables annotations. The main challenge of this project was found in dealing with a great imbalance in class distribution; the three categories approach made this problem feasible. Finally, our paper also adresses the extent to which the decision-making processes of the model can be explained.