Classifying Stars, Galaxies and AGN in CLAUDS+HSC-SSP Using Gradient Boosted Decision Trees

Abstract: Classifying catalog objects as stars, galaxies, or AGN is a crucial part of any statistical study of galaxies. We describe our pipeline for binary (star/galaxy) and multiclass (star/galaxy/Type I AGN/Type II AGN) classification developed for the very deep CLAUDS+HSC-SSP $u^*grizy$ dataset. Our method uses the XGBoost implementation of Gradient Boosted Trees (GBT) to train ensembles of models which take photometry, colours, maximum surface brightnesses, and effective radii from all available bands as input, and output the probability that an object belongs to each of the classes under consideration. At $i_{AB}<25$ our binary star/galaxy model has AUC=0.9974 and at the threshold that maximizes our sample's weighted F1 score, selects a sample of galaxies with 99.7% purity and 99.8% completeness. We test the model's ability to generalize to objects fainter than those seen during training and find that extrapolation of ~1-2 magnitudes is reasonable for most applications provided that the galaxies in the training sample are representative of the range of redshifts and colours of the galaxies in the target sample. We also perform an exploratory analysis of the method's ability to identify AGN using a small x-ray selected sample and find that it holds promise for classifying type I AGN, although it performs less well for type II AGN. Our results demonstrate that GBTs provide a flexible, robust and efficient method for performing classification of catalog objects in large astronomical imaging surveys.