Remove First Detect Later: a counter-intuitive approach for detecting radio frequency interference in radio sky imagery

Abstract

ABSTRACT The modern era has witnessed a rapid uptake of technological use – from air travel to mobile cellphones. Technological advancement has however come at the cost of radio spectrum crowding and as such the efficient detection of radio frequency interference (RFI) from radio sky images has become more paramount. Detecting RFI is a complex task that blends semantic segmentation and anomaly detection, further complicated by the limited availability of public data sets with accurate ground truth labels. Recent studies show that deep learning models improve RFI detection compared to current state-of-the-art tools. However, many astronomers are hesitant to adopt these models, possibly due to the dependence of these models on noisy labels from existing tools when accurate ground truth labels are largely unavailable in the public domain. This study argues that utilizing large weakly labelled training data sets yields lower performance than appropriately employing a modest set of expertly annotated samples. Further, Remove First Detect Later (RFDL), an augmented deep learning framework, is proposed. First, counter-intuitively, removing RFI with inpainting, RFDL feeds the difference between the original and inpainted images into existing detection models. RFDL’s performance is benchmarked against current state-of-the-art deep learning methods and the prevalent AOFlagger pipeline, using AUROC, AUPRC, and F1 score metrics. It is shown that RFDL significantly outperforms the state-of-the-art while only necessitating the use of 20 expertly labelled images.