Research And Implementation Of Solar Radio Classification Based On Deep Learning

Solar radio bursts usually occur during intense solar activity,and they carry important information about the physical environment and radiation conditions in the burst area.Therefore,the study of solar radio bursts can learn the characteristics of magnetic field structure and particle movement in the explosion area,which has high practical significance.Generally speaking,solar radio requires astronomers with expertise to classify it manually,which was time-consuming and laborious.In recent years,researchers have tried to use traditional machine learning methods to achieve automatic classification of solar radio.The method of machine learning usually needs to design features manually to describe pictures,while the information in solar radio pictures is very complicated,so the design features are usually very complex,and it is difficult to accurately describe the deep information of pictures,which will have a great impact on the accuracy of classification.With the rapid development of deep learning technology,it has also made great breakthroughs in the field of image classification.In this thesis,we propose a classification method of solar radio based on deep learning.Firstly,we use principal component analysis to reduce the dimension of radio spectrograms,and then according to the characteristics of the spectrograms,we propose a rectangular convolution kernel to extract image features.In addition,we propose a two-stage strategy to solve the problem of uneven distribution of samples in solar radio data sets.In the first stage,we use the method of generating new samples and resampling original samples to supplement categories with fewer samples.In the second stage,we propose a cost-sensitive multi-classification loss function to make the network pay more attention to the category with fewer samples in the training process.Generally speaking,we first supplement the sample size with the help of sample enhancement strategy for later training,then we use the proposed loss function to train the solar radio classifier.We conduct experiments on the solar radio spectrum database and compare them with other methods,the experimental results show that our method can achieve good classification effect.