Automatic Classification Of Sunspots Using Convolutional Neural Networks

Space weather is a phenomenon in solar-terrestrial space that affect the short time scale variations of the earth's magnetosphere,ionosphere,and thermosphere.Space weather arises as a result of bursts of solar activity.Solar activity is the source of catastrophic space weather.An important task of space weather forecasting is to predict the occurrence and evolution of explosive solar activity in a timely and effective manner,especially solar flares and coronal mass ejections.The parameters of sunspots and active regions composed of sunspot groups are important indicators for the prediction of solar outburst activity,and the study of the identification and classification of sunspot groups is the basis for space weather forecasting.Using the high-resolution totality images provided by the helioseismic and magnetic imager and atmospheric imaging array carried by the Solar Dynamics Observatory,the sunspot data were collected for the time period from 2011 to 2018 to establish a sunspot database,the characteristic parameters of sunspots were measured and extracted,and the sunspot categories were statistically analyzed.And based on the machine learning method,a suitable convolutional neural network was selected to realize the automatic sunspot categorization,and a reasonable correct rate was achieved.Based on the classification of linguistic descriptions of sunspot magnetic fields,the sunspot magnetic classification method was chosen to collect data on seven different types of sunspots.A sunspot database was established,and the parameters related to the sunspot magnetic field and geometric features were extracted and their distributions were statistically analyzed.According to the characteristic distributions of sunspots based on their magnetic field strength,length and width,longitude and latitude span,area and latitude distribution,it was found that complex sunspots have a larger latitude and longitude span and a larger inclination.The statistical analysis of the magnetic field strength of sunspots verifies that the magnetic field strength of complex sunspots is stronger.The significant asymmetry in the number of sunspots at northern and southern latitudes can be seen based on the latitudinal distribution characteristics of sunspots,which verifies the asymmetry of solar activity from north to south.In order to make an accurate classification of sunspot groups on the solar surface more efficiently,machine learning methods were introduced.A suitable convolutional neural network was selected for construction,the sunspot database was expanded by means of data augmentation,the network was trained using the augmented data set,and the network parameters,including dropout value,learning rate and number of iterations,were adjusted to achieve the purpose of automatic sunspot categorization.In addition,the recognition accuracy of convolutional neural networks was tested and evaluated.The experiments provided a new method for machine recognition and classification of sunspots,and laid the foundation for automatic sunspot classification in the future.