A Machine Learning Approach to Mitigating Radio Interference in Deep Wide-field Observations of the Lockman Hole

Abstract

A study of the Lockman Hole at 1.5GHz with the e-MERLIN radio telescope is performed, along with an experiment in using machine learning techniques to identify and remove interference in astronomical radio observations, and an investigation of a faint sub-kpc scale binary black hole candidate. It is found that manually flagged observations used as the basis for a training set does not produce a trained network that performs better than the current state of the art statistical tool, at least when using a GAN architecture. Though, it remains possible that refinements to the GAN model, the pre-processing method, and the training set could produce better results. Machine learning techniques remain as one of the most promising methods of dealing with interference in an increasingly difficult RFI environment populated by massive satellite networks that will impede the operation of next generation radio observatories, such as the SKA. In the investigation of the sub-kpc scale binary AGN candidate, two radio sources are detected that overlap with two previously identified stellar cores, both with a separation of � 1". Specifically, the local epoch separation between the two radio sources is measured as 0.95�0.29 kpc. However, follow up radio observations with greater resolving power did not produce a second detection. The source is very faint, so it is likely these observations were not sensitive enough to provide a second detection; further observation is recommended to reduce the noise level. Finally, a source catalogue of 78 sources in the Lockman Hole is produced from a wide field image with a mean sensitivity of 12�Jy/beam. Radio measurements from other studies at 140MHz and 3GHz are included for � 50% of sources, along with estimates of redshift for 35 sources. Five sources from this new catalogue are studied in detail and the differential source counts are calculated and compared with similar work. The source counts are consistent with an observed faint radio population with a suppressed SFG component at . 1mJy, which is expected at VLBI scales. The overall trend agrees with source counts calculated using a catalogue of VLA sources. The e-MERLIN image is an important step towards producing a deeper radio image of the Lockman Hole that can be used to classify sources within the field, and investigate the nature of the faint radio population.