Automatic classification of astronomical objects using Deep Learning 


So far, all searches of the ESA science data archives and all other astronomical archives are based on metadata from the observations (target, sky coordinates, date of observation, instrument, filter, proposer, etc.), while there is an immense amount of information content in the images that has not yet been fully exploited systematically. We present an inference-first approach for identifying and classifying astronomical objects in complete datasets of FITS images, with the objective of simplifying the task of analysing large datasets with different classification models. This work will enable users to do data-driven searches (e.g. show me all the optically resolved galaxies with an X-ray point source counterpart of hardness ratio X and with a Far-infrared Herschel observation available or show me all the co-moving Gaia sources in the star-forming region Y with clean near-IR background and no PSF blending issues) in a relatively near future via the ESASky (add link to portal.


As an initial Proof of Concept, we are applying a current state-of-the-art neural network model (Morpheus to the 2MASS and Skymapper optical image datasets, to generate a Machine Learning catalogue, using ESAC’s GPU grids, Docker and Kubernetes.