RFI-HWT: Hybrid deep-learning algorithm for advanced RFI mitigation in radio astronomy

¹ School of Physics and Electronic Science, Guizhou Normal University, Guiyang 550025, China
² Guizhou Provincicial Key Laboratory of Radio Astronomy and Data Processing, Guizhou Normal University, Guiyang 550025, China
³ College of Big Data and Information Engineering, Guizhou University, Guiyang 550025, China
⁴ CAS Key Laboratory of FAST, National Astronomical Observatories, Chinese Academy of Sciences, Beijing 100101, China
⁵ Key Laboratory of Information and Computing Guizhou Province, Guizhou Normal University, Guiyang 550001, China
⁶ National Space Science Center, Chinese Academy of Sciences, Beijing 100000, China
⁷ Guizhou Software Engineering Research Center, Guiyang 550000, China

^★ Corresponding authors: xiaoshuo@gznu.edu.cn; lizhang.science@gmail.com

Received: 7 March 2025
Accepted: 16 July 2025

Abstract

Context. High-sensitivity radio telescopes face increasing challenges of radio frequency interference (RFI) from various sources, including communication base stations, TV broadcasts, satellite signals, and so on. This RFI degrades the quality of observational data, posing a significant obstacle to the detection of pulsars and fast radio bursts (FRBs). Consequently, the development of effective RFI mitigation techniques is crucial.

Aims. To address this issue, this study aims to develop an advanced RFI mitigation algorithm, RFI-HWT, to overcome the limitations of existing machine-learning-based methods by integrating multi-scale and multidirectional signal decomposition with deep-learning-based denoising.

Methods. RFI-HWT integrates the multilevel two-dimensional wavelet transform (2D WT) with a deep-learning denoising model, namely DnCNN, to accurately identify and eliminate RFI via multi-scale and multidirectional signal decomposition. Furthermore, our model employs a self-training semi-supervised learning strategy, effectively utilizing both limited labeled and abundant unlabeled data during training to enhance its generalization and adaptability.

Results. Preliminary experiments on FAST and Parkes datasets demonstrate RFI-HWT’s superior performance: it achieves a 15% average signal-to-noise ratio (Avg.S/N) enhancement and 14% average structural similarity index measure (ASSIM) improvement over PRESTO’s “rfifind” method, outperforms the original 2D WT (11% Avg.S/N, 12% ASSIM gains), and surpasses the representative deep-learning methods RFI-Net (3.26% Avg.S/N, 4.46% ASSIM) and RFDL (2.12% Avg.S/N, 2.71% ASSIM). Furthermore, higher precision, recall, and F1-score values across both datasets confirm its strong generalization capability. CUDA parallelization ensures efficient processing while maintaining excellent performance.

Conclusions. These findings demonstrate that RFI-HWT is a feasible solution for mitigating multiple types of RFI sources and is capable of improving the data quality and efficiency of pulsar and FRB searches with high-sensitivity telescopes.