Mitigation of Radio Frequency Interference in the Solar Radio Spectrum Based on Deep Learning

Radio frequency interference (RFI) may contaminate the signal received by solar radio telescopes. The existence of RFI in the solar radio spectrum affects the accuracy and efficiency of the extraction of burst parameters, which is related to the quality of scientific results and even the authenticity of conclusions. Therefore, it is necessary to carry out research on RFI recognition algorithms for solar radio data. This article aims to compare the recognition performance of six different deep-learning networks (FCN, Deconvnet, Segnet, Unet, Dual-Resunet, and DSC Based Dual-Resunet) on the RFI in solar radio spectra observed by the Chinese Solar Broadband Radio Spectrometer (SBRS). The accuracy and convergence speed in the training process, as well as various performance metrics in the test, indicate that the proposed DSC Based Dual-Resunet is the most suitable neural-network for this task and can achieve both performance and light weight. The RFI recognition accuracy of the DSC Based Dual-Resunet is close to Unet when there is no burst in the spectrum, but in the case of a burst DSC Based Dual-Resunet is obviously better than Unet in terms of RFI recognition. Moreover the model size and number of parameters are approximately 12.5% of those of Unet, and the amount of computation is 38% of that of Unet, which greatly improves the computation efficiency and is of great significance for the realization of the network on mobile hardware. It is promising for the large-scale application of RFI recognition for solar radio telescopes.