A Robust RFI Identification Method for Radio Interferometry Based on LightGBM

Abstract

Abstract Radio frequency interference is an essential factor affecting the observation data quality of radio telescopes. In constructing the Square Kilometer Array (SKA) radio interferometer, dealing with radio frequency interference (RFI) effectively is always a hot issue in data processing. Traditional identification methods have poor precision or recall, and existing machine-learning-based methods have complicated models and low processing efficiency. We propose a LightGBM identification method based on previous machine-learning research to identify RFI. Based on the data of SKA1-LOW simulation observations, we construct five visibility function data sets, one for modeling and the rest for validation. The experimental results show that the F 2-score reaches 0.9583, and the training and prediction speed are much more efficient than those of convolutional neural networks in a similar recent study. Then, we further investigate the effectiveness of this model in identifying RFI from actual MeerKAT observations. The results show that the overall effectiveness is comparable to tools such as Tfcrop and Rflag, improving over existing methods in identification speed.