Finding AGN remnant candidates based on radio morphology with machine learning

Abstract

Context. Remnant radio galaxies represent the dying phase of radio-loud active galactic nuclei (AGN). Large samples of remnant radio galaxies are important for quantifying the radio-galaxy life cycle. The remnants of radio-loud AGN can be identified in radio sky surveys based on their spectral index, and identifications can be confirmed through visual inspection based on their radio morphology. However, this latter confirmation process is extremely time-consuming when applied to the new large and sensitive radio surveys. Aims. Here, we aim to reduce the amount of visual inspection required to find AGN remnants based on their morphology using supervised machine learning trained on an existing sample of remnant candidates. Methods. For a dataset of 4107 radio sources with angular sizes of larger than 60 arcsec from the LOw Frequency ARray (LOFAR) Two-Metre Sky Survey second data release (LoTSS-DR2), we started with 151 radio sources that were visually classified as ‘AGN remnant candidate’. We derived a wide range of morphological features for all radio sources from their corresponding Stokes-I images: from simple source-catalogue-derived properties to clustered Haralick-features and self-organising-map(SOM)-derived morphological features. We trained a random forest classifier to separate the AGN remnant candidates from the yet-to-be inspected sources. Results. The SOM-derived features and the total-to-peak flux ratio of a source are shown to have the greatest influence on the classifier. For each source, our classifier outputs a positive prediction, if it believes the source to be a likely AGN remnant candidate, or a negative prediction. The positive predictions of our model include all initially inspected AGN remnant candidates, plus a number of yet-to-be inspected sources. We estimate that 31 ± 5% of sources with positive predictions from our classifier will be labelled AGN remnant candidates upon visual inspection, while we estimate the upper bound of the 95% confidence interval for AGN remnant candidates in the negative predictions to be 8%. Visual inspection of just the positive predictions reduces the number of radio sources requiring visual inspection by 73%. Conclusions. This work shows the usefulness of SOM-derived morphological features and source-catalogue-derived properties in capturing the morphology of AGN remnant candidates. The dataset and method outlined in this work bring us closer to the automatic identification of AGN remnant candidates based on radio morphology alone and the method can be used in similar projects that require automatic morphology-based classification in conjunction with small labelled sample sizes.