Rubin Observatory: DeepDISC, first test of a deep learning model on real LSST images

Applied to LSST Data Preview 1 data, DeepDISC detects, separates, and classifies stars and galaxies in a single pass, with galaxy completeness increasing from 40% to 84% after adjustment.

Classical astronomical pipelines detect sources, separate overlapping ones, and then classify them, in distinct steps. DeepDISC does it all at once: it analyzes the entire scene and identifies, cuts, and labels each object as a star or galaxy. Members of the LSST Dark Energy Science collaboration applied it for the first time to real-world data from the Vera C. Rubin Observatory, the Data Preview 1 (DP1) images of two deep fields in the Southern Hemisphere, the ECDFS and the EDFS.

Real-sky simulations

Pre-trained on LSST simulations, the model was readjusted on DP1 images in six bands, then its classification part was refined on a sample of objects identified by spectroscopy, Gaia or photometry. The gain is clear: completeness increases from 40.2% to 84.2% for galaxies and from 13% to 49.1% for stars with strict matching. On never-before-seen sources, its classification reaches 97.4% agreement, compared to 98.5% for a random forest published on the same field; at low magnitudes, it is more likely to describe as stars objects that independent data classify as random forests.

Compared to the deeper imaging of the Hubble telescope (CANDELS), DeepDISC allows 14.9% of unrecognized source mixtures to pass through, compared to 18% for the official DP1 catalogue. The authors see it as a reference base for the next release, Data Preview 2.