Detecting the edges of galaxies with deep learning

¹ GIR GCME. Departamento de Informática, Universidad de Valladolid, 47011 Valladolid, Spain
e-mail: jesus.fernandez.iglesias@estudiantes.uva.es
² Departamento de Física Teórica, Atómica y Óptica, Universidad de Valladolid, 47011 Valladolid, Spain
³ Instituto de Astrofísica e Ciências do Espaço, Universidade de Lisboa, OAL, Tapada da Ajuda, 1349-018 Lisbon, Portugal

Received: 16 June 2023
Accepted: 17 November 2023

Abstract

Galaxy edges or truncations are low-surface-brightness (LSB) features located in the galaxy outskirts that delimit the distance up to where the gas density enables efficient star formation. As such, they could be interpreted as a non-arbitrary means to determine the galaxy size and this is also reinforced by the smaller scatter in the galaxy mass-size relation when comparing them with other size proxies. However, there are several problems attached to this novel metric, namely, the access to deep imaging and the need to contrast the surface brightness, color, and mass profiles to derive the edge position. While the first hurdle is already overcome by new ultra-deep galaxy observations, we hereby propose the use of machine learning (ML) algorithms to determine the position of these features for very large datasets. We compare the semantic segmentation by our deep learning (DL) models with the results obtained by humans for HST observations of a sample of 1052 massive (M_stellar > 10¹⁰ M_⊙) galaxies at z < 1. In addition, the concept of astronomic augmentations is introduced to endow the inputs of the networks with a physical meaning. Our findings suggest that similar performances than humans could be routinely achieved, although in the majority of cases, the best results are obtained by combining (with a pixel-by-pixel democratic vote) the output of several neural networks using ensemble learning. Additionally, we find that using edge-aware loss functions allows for the networks to focus their optimization on the galaxy boundaries and, therefore, to provide estimates that are much more sensitive to the presence of neighboring bodies that may affect the shape of the truncation. The experiments reveal a great similarity between the semantic segmentation performed by the AI compared to the human model. For the best model, an average dice of 0.8969 is achieved, while an average dice of 0.9104 is reached by the best ensemble, where the dice coefficient represents the harmonic mean between the precision and the recall. This methodology will be profusely used in future datasets, such as that of Euclid, to derive scaling relations that are expected to closely follow the galaxy mass assembly. We also offer to the community our DL algorithms in the author's github repository.