Rubin Observatory: six machine learning methods classify stars and galaxies without spectroscopy

Tested on the DP1 and DP2 data releases from the Vera C. Rubin Observatory, the pipeline achieves a macro-F1 of approximately 0.97, taking photometric uncertainties into account.

Before the full launch of its major survey of the sky, the Vera C. Rubin Observatory published data previews (DP1 and DP2). A team used them to automatically distinguish point sources, such as stars, from extended sources, such as galaxies, solely from photometry in six bands and without the use of spectroscopy. Six methods were competed against each other (SVM, random forest, XGBoost, k nearest neighbors, neural network, and voting together) on DP1's deepest field, the Extended Chandra Deep Field South.

The random forest and the set per vote come out on top, tied (macro-F1 of 0.918 on the test data), and the width at mid-height of the sources stands out as the most decisive indicator. Applied without adjustment to other fields, whose extinction and star density vary by a factor of 7 to 10, the pipeline remains above 0.87; on four DP2 fields, it reaches 0.83 to 0.88. Adding the photometric uncertainties per band as variables, the score climbs to about 0.95-0.98 depending on the field.