Research On Denoising And Burst Detection Methods Of Solar Radio Dynamic Spectrogram

With the continuous development of observation technology,astronomy has entered the era of big data.In the face of massive observation data,traditional image processing methods are inefficient,and a large number of observation data meet the needs of deep learning.Deep learning is an important research direction in the field of machine learning and is widely used in the field of image processing.In this thesis,based on the actual research project,the denoising and burst detection of solar radio spectrogram are studied,and the denoising method and burst event detection method based on deep learning are proposed.This thesis analyses the sources and effects of noise in the solar radio observation system in Chashan,Weihai City,Shandong Province.It is believed that the noise mainly comes from the intrinsic noise inside the equipment and the complex interference in the external environment.The presence of noise not only reduces the observation quality of the solar radio spectrogram,but also affects the subsequent detection and determination of solar radio burst events.After experimenting with the traditional noise reduction methods,it is considered that the denoising method based on the neighborhood pixel values can only reduce the noise of a certain type,but it has no generalization for other types of noise and can not meet the denoising requirements of the solar radio spectrogram.Therefore,the deep learning technology is considered to construct the denoising method.In this thesis,two denoising methods based on deep learning for solar radio spectrogram are constructed and compared.First,two kinds of deep learning networks FFDNet and DnCNN are set up.Compared with DnCNN,FFDNet add noise estimation subnetworks and resample images by up-sampling and down-sampling,thus increasing the network's perception of noise levels.Secondly,in the process of building the dataset,the images from three sources,namely,the Chashan observatory,e-CALLSITO observatory and Nancay observatory,are filtered,and the problem of insufficient contrast image is solved by noise enhancement.Finally,the denoising methods of solar radio spectrogram based on two network models are tested,and the results show that the denoising methods based on FFDNet network are better than DnCNN in noise reduction ability.In this thesis,during the process of building object detection method for solar radio burst events,the advantages and disadvantages of different types of target detection networks are introduced,and the RetinaNet network is chosen to build the burst detection method according to the actual needs.Feature Pyramid Network(FPN)is used as the backbone network in the RetinaNet network,and ResNet is used for feature extraction.FPN combines features of different scales and semantics of different levels through simple connection in structure,which improves the performance of object detection without affecting detection speed.Secondly,detailed statistics are made on the distribution attributes of solar radio burst in the dataset while building the dataset,and the solar radio type ? burst are taken as an example for experiments.Finally,in order to further improve the detection ability of multiple bursts,this thesis puts forward an improved method to classify and recombine batch data in network training and carries out experiments again.The results show that the improved network improves the detection accuracy of multiple bursts.