VEXAS: VISTA EXtension to Auxiliary Surveys - Data Release 2. Machine-learning based classification of sources in the Southern Hemisphere

We present the second public data release (DR) of the VISTA EXtension to Auxiliary Surveys (VEXAS), where we classify objects into stars, galaxies and quasars based on an ensemble of machine learning algorithms. The aim of VEXAS is to build the widest multi-wavelength catalogue, providing reference magnitudes, colours and morphological information for a large number of scientific uses. We apply an ensemble of 32 different machine learning models, based on three different algorithms and on different magnitude sets, training samples and classification problems on the three VEXAS DR1 optical+infrared (IR) tables. The tables were created in DR1 cross-matching VISTA near-IR data with WISE far-IR data and with optical magnitudes from the Dark Energy Survey (VEXAS-DESW), the Sky Mapper Survey (VEXAS-SMW), and the PanSTARRS (VEXAS-PSW). We assemble a large table of spectroscopically confirmed objects (415 628 unique objects), based on the combination of 6 different spectroscopic surveys that we use for training. We develop feature imputation to classify also objects for which magnitudes in one or more bands are missing. We classify in total ~90 million objects in the Southern Hemisphere. Among these, ~62.9M (~52.6M) are classified as 'high confidence' ('secure') stars, ~920k (~750k) as 'high confidence' ('secure') quasars and ~34.8M (~34.1M) as 'high confidence' ('secure') galaxies, with probabilities pclass≥0.7 (pclass≥0.9). The density of high-confidence extragalactic objects varies strongly with the survey depth: at pclass≥0.7, there are 111/deg2 quasars in the VEXAS-DESW footprint and 103/deg2 in the VEXAS-PSW footprint, while only 10.7/deg2 in the VEXAS-SM footprint. Improved depth in the mid-IR and coverage in the optical and near-IR are needed for the SM footprint that is not already covered by DESW and PSW.