Multi-Method Machine Learning Classification of Astronomical Sources with Rubin Data Previews: a Spectroscopy-Free Approach with DP1 and DP2
We present a multi-method machine-learning study for the binary photometric classification of point-like and extended sources in the Rubin Observatory Data Preview 1 (DP1) and Data Preview 2 (DP2). The workflow runs on the Rubin Science Platform using DP1 and DP2 ugrizy photometry, the DP1 ancillary photometric-redshift catalog, and the Monster astrometric and photometric reference catalog, without relying on external spectroscopy. Six methods (SVM, Random Forest, XGBoost, k-Nearest Neighbors, an Artificial Neural Network, and a soft Voting ensemble) are benchmarked in the deepest DP1 field, the Extended Chandra Deep Field South. These methods reach 10-fold cross-validation macro-F1 scores of 0.95-0.97 on the balanced training subset and test macro-F1 scores of 0.87-0.92 on the imbalanced held-out set, with RF and the SV ensemble tied at the top (0.918); the differences among the leading methods are within the statistical uncertainty. Permutation importance identifies FWHM as the dominant feature, and the six methods agree on 79.5% of the full catalog. The same pipeline is then applied without retuning the settings to the two other DP1 fields with full ugrizy coverage and photometric-redshift support, Euclid Deep Field South and the Low Galactic Latitude Field. Together, the three fields span a factor of approximately 7 in E(B-V) and approximately 10 in stellar density, yet headline macro-F1 remains above 0.87 in every field. We then apply the same pipeline to four DP2 fields (ECDFS, COSMOS, EDFS, and ELAIS-S1). Using colors and morphology, the best method reaches a test macro-F1 of 0.83 to 0.88 across these fields, and the six methods agree on 77 to 86% of each catalog. Adding the per-band photometric uncertainties as features raises the best macro-F1 to about 0.95 in ECDFS DP1 and 0.97 in ECDFS DP2 (and to approximately 0.97-0.98 in the other fields).