Detection of Radio Frequency Interference Using an Improved Generative Adversarial Network

Radio Frequency Interference (RFI) is a type of inevitable noise in the radio astronomy data collection process. It can corrupt weak cosmic signals and potentially lead to misleading results. The proper identification of RFI during data processing thus is critical to obtain clean and high-quality data for analysis. This need will become even more urgent when next generation of large radio telescopes, such as the Five-hundred-meter Aperture Spherical radio Telescope (FAST) comes into service and generates an increasing amount and complexity of radio signal data. Among RFI identification methods, detection using artificial intelligence (AI) has particularly demonstrated advantages in superior efficiency, accuracy and less human intervention. We thus propose a RFI detection model based on Pix2Pix, an image-to-image translation solver using Generative Adversarial Network (GAN): RFI-GAN. This model transforms the RFI detection to an image translation problem and trains two deep neural networks that contest each other to output a binary RFI mask image to eliminate RFI noises. We also optimize the network structures of the generator and discriminator used in the Pix2Pix model for better quality of RFI detection, making it suitable for processing data from single antenna radio telescope. The model is designed to serve the upcoming FAST data and has been evaluated using a standard simulation data set generated by the HI Data Emulator (HIDE). Experimental results have shown that our model can achieve higher scores (99%) on accuracy, recall and F1-score than other RFI detection methods