Research On Sunspot Group Classification Method Based On Deep Learning

Sunspots are solar phenomena that occur in the solar sphere,and are closely related to solar activity,such as flares and coronal mass ejections.These activities can seriously affect the Earth's atmosphere,thereby interfering with short-wave radio communications on the ground and causing "magnetic storms" and other hazards.Related studies have shown that certain types of sunspots are associated with flares,so the type of sunspots can be detected in real time to predict flare outbreaks.The early classification of the sun sunspot group and the automatic classification using image processing technology are not high.The main reason is that the characteristics of the sunspot group are not effectively and fully extracted,resulting in many black subgroups missing or misdetected.Therefore,it is very meaningful to study how to effectively extract the features of the sunspot group to improve the accuracy of the sunspot group classification.The convolutional neural network in deep learning can effectively extract the low-level and high-level fusion features of image targets.This thesis mainly studies how to use the deep learning method to solve the problem of low accuracy of McIntosh classification of sunspots.main tasks as follows:(1)Establish a sunspot group classification sample library McIntosh_SC.The pre-processing of 1840 full disk solar images from 2013 to 2016 is carried out,and the data set is expanded by operations such as sample flipping to compensate for the problem of too few sub-samples of sub-category in some categories.Labelimg is used according to the McIntosh classification standard.The sunspot group is labeled one by one,and more than 8,800 category labels are marked.(2)This thesis studies the classification method of sunspot subgroup based on Faster R-CNN.After researching and comparing the major mainstream convolutional neural network structures,the ZF network is improved.The improved network is used as the main network structure of Faster R-CNN.The sample library is input into the network for migration learning training and fine-tuning parameters to make the characteristics of the black subgroup.Effectively extracted,solved the black subgroup of the multi-class detection frame overlap,the black subgroup multi-class detection is not in the same full disk solar image,the proportion of the black sub-group anchor window is not suitable,etc.The classification of the test shows that the accuracy of the method is 69.94% and the recall rate is 89.03%.(3)The R-FCN based classification method of sunspots is studied.Because the Faster R-CNN model can't meet the speed requirement of real-time detection,and the accuracy is still low,the R-FCN based sunspot sub-group classification method is studied.The residual network ResNet-50 is selected as the R-FCN network structure.The network trained and fine-tuned the parameters,and used the same improved detection method as in(2)to classify the black subgroups on 350 full disk solar images.The statistical analysis showed that the accuracy of the method was 84.99%,and the recall rate was 93.4%.(4)The accuracy,recall,accuracy and mAP were used as evaluation indicators to analyze and compare the results of the two methods.The results show that the accuracy of the R-FCN based method is higher than that of the National Oceanic and Atmospheric Administration NOAA 37%,which is also much higher than the test results of other scholars.