Redshifts of radio sources in the Million Quasars Catalogue from machine learning

TL;DR

This study explores machine learning methods to estimate redshifts of radio sources using optical and near-infrared data, achieving reliable predictions for a subset of sources and expanding redshift estimates for a larger sample, with implications for future surveys.

Contribution

The paper demonstrates the effectiveness of machine learning in predicting radio source redshifts using optical and infrared data, and highlights the potential for future surveys like SkyMapper and SKA.

Findings

Reliable redshift predictions for 12,503 sources using optical/infrared data.

Expanded redshift estimates to 32,698 sources by imputing missing magnitudes.

Model performance improves with larger training samples, but radio spectra alone are insufficient.

Abstract

With the aim of using machine learning techniques to obtain photometric redshifts based upon a source's radio spectrum alone, we have extracted the radio sources from the Million Quasars Catalogue. Of these, 44,119 have a spectroscopic redshift, required for model validation, and for which photometry could be obtained. Using the radio spectral properties as features, we fail to find a model which can reliably predict the redshifts, although there is the suggestion that the models improve with the size of the training sample. Using the near-infrared--optical--ultraviolet bands magnitudes, we obtain reliable predictions based on the 12,503 radio sources which have all of the required photometry. From the 80:20 training--validation split, this gives only 2501 validation sources, although training the sample upon our previous SDSS model gives comparable results for all 12,503 sources. This makes us confident that SkyMapper, which will survey southern sky in the u, v, g, r, i, z bands, can be used to predict the redshifts of radio sources detected with the Square Kilometre Array. By using machine learning to impute the magnitudes missing from much of the sample, we can predict the redshifts for 32,698 sources, an increase from 28% to 74% of the sample, at the cost of increasing the outlier fraction by a factor of 1.4. While the "optical" band data prove successful, at this stage we cannot rule out the possibility of a radio photometric redshift, given sufficient data which may be necessary to overcome the relatively featureless radio spectra.