A Standardized Framework for Radio Frequency Interference Detection and Benchmarking
Radio Frequency Interference (RFI) contamination poses a significant challenge to modern radio astronomy observations, requiring robust detection, mitigation and flagging strategies. Traditional statistical thresholding methods such as TFCROP, RFLAG, and AOFlagger have been de rigeur, but the emergence of machine learning techniques necessitates systematic performance comparison across methodologies. We present RFI Toolbox, a comprehensive framework designed to establish standardized benchmarks for RFI detection algorithms through controlled synthetic datasets and uniform evaluation metrics.
Our approach generates realistic synthetic RFI patterns across multiple interference categories including broadband, narrowband, time-bursty, and complex sweeping signals. The framework supports both traditional convolutional neural networks and advanced segmentation models, with direct integration to the SAM-RFI project for Segment Anything Model adaptation to radio astronomy data. We demonstrate the toolbox's capabilities through systematic comparisons between SAM-RFI models and established statistical methods (TFCROP, RFLAG, AOFlagger) using standardized datasets and evaluation protocols.
The framework establishes comprehensive evaluation metrics including Dice coefficient, Intersection over Union, precision, recall, and F1-score to enable direct quantitative comparison across methodologies. Integration capabilities support seamless data loading from Measurement Sets and export to machine learning platforms, facilitating reproducible research workflows. Through open-source code, publicly available training datasets, and open model repositories, this work aims to enhance scientific trust and transparency in RFI detection research. By providing standardized benchmarks and accessible tools, we invite community contributions to advance RFI mitigation capabilities for next-generation radio telescopes. The open science approach ensures reproducible results and enables collaborative development of improved detection algorithms across the radio astronomy community.
This standardized evaluation framework addresses the critical need for reproducible RFI detection benchmarks, ultimately advancing our ability to identify interference in an increasingly congested radio frequency environment.