Automated Solar Activity Detection Based On Deep Models

The Sun has very close relationship with human beings.Various solar activities can lead to changes in space weather,and then have a great impact on magnetic field,ionosphere,climate and space activities of the Earth.Hence,the detection of solar activities plays an important role and has significant practical value for the study of laws and characteristics of solar activities,space weather monitoring,national defense construction and national economic development.With the dramatic increase in the amount of solar observations,there is an increasing need to develop automatic algorithms to deal with the data deluge.Although much progress has already been made,many computer vision,image processing and machine learning techniques have been applied to the recognition of solar activities,the existing methods are almost segmentation-based methods.In this paper,we investigate the problem of automatically identifying and locating active regions(ARs)and coronal holes(CHs)in solar extreme ultraviolet(EUV)images,and we regard it as an object detection task in the field of computer vision.We consider to utilize the emerging deep learning techniques to achieve our goal.The main contributions of the work are listed as follows:1.To solve the problem of recognition and locating of solar activities from a new perspective.We regard it as an object detection task,which tries to find all instances of the category of interest and predict their spatial regions.2.We collect a solar image dataset spanning from 1 January 2012 to 30 June 2012 from the Solar Dynamics Observatory/Atmospheric Imaging Assembly(SDO/AIA)telescope.It includes images from all seven simultaneous EUV wavebands.Then we manually label the ARs and CHs in these images.Thus we construct a dataset to train and test the proposed algorithm.3.We bring the deep learning techniques in the field of solar physics,and apply the leading region-based convolutional neural networks methods in object detection to the detection of ARs and CHs.Besides,we study the effects of various EUV wavebands on the detection of different solar activities.The experiments verify the effectiveness of deep learning in solar activity detection,and show that different wavebands play a different role in identifying different solar activities.4.Taking into account that images of different wavebands provide a different view of the solar atmosphere,we propose a multi-channel model to integrate the complementary and consistent information of all wavebands.The experimental results show the effectiveness of our model.It can achieve more reliable detection.