Automated Source Detection in Radio Astronomy Images Using Deep Neural Networks

Abstract Radio astronomical observations inherently capture both spatial and spectral emission, creating high-dimensional data cubes. As next-generation facilities like the Square Kilometre Array (SKA) come online, the volume and complexity of these data present significant challenges for automated source detection. Faint sources are often obscured by complex noise and artifacts, while their morphology evolves across the spectral range, limiting the efficacy of traditional threshold-based algorithms. To address these challenges, we present a volumetric detection methodology based on a 3D U-Net++ architecture designed to identify intricate structures within the 3D domain. We incorporate Monte Carlo Dropout to perform Bayesian inference, providing robust per-detection uncertainty estimates. Validated against simulated data from the SKAO Science Data Challenge 2 (SDC2), our method demonstrates a substantial performance improvement over the standard SoFiA pipeline, achieving a total score nearly 1000 points higher under the official metric. This gain is driven by superior sensitivity: our approach increases recall by approximately 2 % and recovers 22.2 % more confirmed sources than SoFiA, albeit with a 6 % decrease in precision. Furthermore, the uncertainty estimates derived from our Bayesian framework provide critical insights for candidate vetting and the prioritization of follow-up observations.