| Radio Frequency Interference（RFI）can have a detrimental impact on the quality of obser-vational data obtained from radio telescopes,leading to a reduction in the signal-to-noise ratio and hindering scientific research.Effective mitigation of RFI has become a pressing issue for the field of radio astronomy,particularly with regards to the construction and operation of the next generation of radio telescopes,which use more antennas to detect a wider range of radio wavelengths.To this end,machine learning techniques are being increasingly explored as a means of automating RFI detection and classification tasks,due to their superior generalization and applicability compared to traditional methods.In this paper,we investigate a machine learning-based approach for RFI identification.We evaluate the performance of four machine learning models and three RFI tagging tools by simulating a batch of SKA-LOW observations contaminated with RFI.Our results demonstrate that Light GBM（Light Gradient Boosting Machine）exhibits the highest overall recognition per-formance,with an F2-Score of 0.9716.Notably,we find that Light GBM outperforms other machine learning methods for weak RFI recognition,and also achieves a two orders of magni-tude improvement in training and prediction speed compared to neural networks in similar recent studies.This points to the potential for real-time machine learning models to enable efficient RFI recognition pipelines in the future.To further demonstrate the effectiveness of Light GBM,we test the model using real ob-servations from Meer KAT,LOFAR and MUSER.Our results indicate that Light GBM performs well across different radio telescopes and science missions,with an overall effectiveness compa-rable to existing tools such as AOFlagger,Tfcrop and Rflag,but with faster recognition speeds and higher practical value.The results also provide a demonstration for producing more accurate real data labels in the future. |
PDF Full Text Request