Research On Correction Of Low-frequency SKA Bandwidth Smearing Effect Based On Deep Learnin

As the observation and exploration of the Universe deepens,mankind has started to build the next generation of super radio telescope SKA to address one of the ambitious science goals based on the unknown of the Universe.The SKA1-low,currently under construction,will soon begin its survey and enter the first science goal of cosmic dawn and reionization detection.However,the bandwidth smearing effect brought by low-frequency SKA imaging distorts the observed signals during the reionization period,making it impossible to complete more accurate signal separation and other subsequent work.And traditional astronomical broadband interferometer imaging algorithms cannot complete accurate batch effect correction in the face of massive data.In recent years,the rapid development of deep learning in the field of image processing provides a new idea for the correction of bandwidth smearing effect in low-frequency SKA imaging.In this paper,we explore the bandwidth smearing effect correction method for low-frequency SKA imaging from a practical research project.Through the theoretical analysis of the bandwidth smearing effect,we introduce the bandwidth smearing effect in the OSAKR simulation and obtain the simulated images of the bandwidth smearing effect observation.These images show that the bandwidth smearing phenomenon increases with the distance from the center of the field of view,which greatly limits the observation range of the field of view and affects the separation of the subsequent reionization signal from the foreground interference.The paper perform multi-frequency synthetic imaging correction experiments with different channel numbers on the bandwidth smearing images.It is concluded that the traditional broadband imaging algorithm cannot strike a balance between fast correction and high quality imaging,and cannot meet the requirements for efficient correction of low-frequency SKA survey imaging.Therefore,it is necessary to apply deep learning method in bandwidth smearing effect correction.In order to address the shortcomings of traditional astronomical imaging algorithms,three deep learning network-based methods are constructed and compared for the correction of low-frequency SKA bandwidth smearing effect.Firstly,a CNN model is built,and a dataset based on extragalactic point source,radio halos and a perfect evaluation system are constructed.Through the training and testing results,it is found that the CNN network can achieve a high level of correction speed,imaging effect and evaluation index in the correction test of extragalactic point source,which improves the deficiency of efficient correction that cannot be achieved in traditional imaging.The innovation of deep learning method applied in bandwidth smearing effect correction is realized.However,the evaluation index of the CNN network is not satisfactory for the partial correction results of the extragalactic point source,and the partial correction results in the radio halos also produce weak artifacts.Based on the above problems,this paper also proposes two improved convolutional neural networks based on SE and CBAM attention mechanisms named SE＿CNN and CBAM＿CNN respectively,and applies them to the correction of extragalactic point source and radio halos.This paper achieves the innovation and improvement of the algorithm.The improved networks were trained and tested with the same data set to ensure the homogeneity of data processing.And the experiments demonstrated that SE＿CNN further improved the evaluation metrics for extragalactic point source correction without significantly increasing the training computational resources.However,CBAM＿CNN better improved the problem of weak artifacts of radio corona and improved the evaluation metrics of two kinds source correction results.The overall experimental results show that SE＿CNN and CBAM＿CNN have their own advantages and disadvantages in the correction tests of radio halo and extragalactic point source.In particular signal correction,at least one of these networks is superior to traditional imaging algorithms in both correction efficiency and imaging evaluation index.This paper demonstrates the innovation and feasibility of the deep learning-based correction method for low-frequency SKA bandwidth smearing effect through multiple rounds of experiments.