Evaluating the feasibility of interpretable machine learning for globular cluster detection

Abstract

Extragalactic globular clusters (GCs) are important tracers of galaxy formation and evolution because their properties, luminosity functions, and radial distributions hold valuable information about the assembly history of their host galaxies. Obtaining GC catalogues from photometric data involves several steps which will likely become too time-consuming to perform on the large data volumes that are expected from upcoming wide-field imaging projects such as Euclid. In this work, we explore the feasibility of various machine learning methods to aid the search for GCs in extensive databases. We use archival Hubble Space Telescope data in the F475W and F850LP bands of 141 early-type galaxies in the Fornax and Virgo galaxy clusters. Using existing GC catalogues to label the data, we obtained an extensive data set of 84929 sources containing 18556 GCs and we trained several machine learning methods both on image and tabular data containing physically relevant features extracted from the images. We find that our evaluated machine learning models are capable of producing catalogues of a similar quality as the existing ones which were constructed from mixture modelling and structural fitting. The best performing methods, ensemble-based models such as random forests, and convolutional neural networks recover ∼90−94% of GCs while producing an acceptable amount of false detections (∼6−8%), with some falsely detected sources being identifiable as GCs which have not been labelled as such in the used catalogues. In the magnitude range 22 < m4_g ≤ 24.5 mag, 98−99% of GCs are recovered. We even find such high performance levels when training on Virgo and evaluating on Fornax data (and vice versa), illustrating that the models are transferable to environments with different conditions, such as different distances than in the used training data. Apart from performance metrics, we demonstrate how interpretable methods can be utilised to better understand model predictions, recovering that magnitudes, colours, and sizes are important properties for identifying GCs. Moreover, comparing colour distributions from our detected sources to the reference distributions from input catalogues finds great agreement and the mean colour is recovered even for systems with fewer than 20 GCs. These are encouraging results, indicating that similar methods trained on an informative sub-sample can be applied for creating GC catalogues for a large number of galaxies, with tools being available for increasing the transparency and reliability of said methods.