Radio source-component association for the LOFAR Two-metre Sky Survey with region-based convolutional neural networks

Radio loud active galactic nuclei (RLAGNs) are often morphologically complex objects that can consist of multiple, spatially separated, components. Astronomers often rely on visual inspection to resolve radio component association. However, applying visual inspection to all the hundreds of thousands of well-resolved RLAGNs that appear in the images from the Low Frequency Array (LOFAR) Two-metre Sky Survey (LoTSS) at $144$ MHz, is a daunting, time-consuming process, even with extensive manpower. Using a machine learning approach, we aim to automate the radio component association of large ($> 15$ arcsec) radio components. We turned the association problem into a classification problem and trained an adapted Fast region-based convolutional neural network to mimic the expert annotations from the first LoTSS data release. We implemented a rotation data augmentation to reduce overfitting and simplify the component association by removing unresolved radio sources that are likely unrelated to the large and bright radio components that we consider using predictions from an existing gradient boosting classifier. For large ($> 15$ arcsec) and bright ($> 10$ mJy) radio components in the LoTSS first data release, our model provides the same associations for $85.3\%\pm0.6$ of the cases as those derived when astronomers perform the association manually. When the association is done through public crowd-sourced efforts, a result similar to that of our model is attained. Our method is able to efficiently carry out manual radio-component association for huge radio surveys and can serve as a basis for either automated radio morphology classification or automated optical host identification. This opens up an avenue to study the completeness and reliability of samples of radio sources with extended, complex morphologies.