Stream automatic detection with convolutional neural networks

¹ Departamento de Astronomía, Universidad de La Serena, Raúl Bitrán N° 1305, La Serena, Chile
² Departamento de Física, Universidad Técnica Federico Santa María, Avenida España 1600, 2390123 Valparaíso, Chile
³ Max-Planck-Institut für Astrophysik, Karl-Schwarzschild-Str. 1, 85748, Garching, Germany
⁴ Cardiff Hub for Astrophysics Research and Technology, School of Physics and Astronomy, Cardiff University, Queen’s Buildings, Cardiff CF24 3AA, UK
⁵ Center for Theoretical Astrophysics and Cosmology, Department of Astrophysics, University of Zurich, Switzerland
⁶ Astrophysics Research Institute, Liverpool John Moores University, 146 Brownlow Hill, Liverpool L3 5RF, UK
⁷ Department of Physics and Astronomy “Augusto Righi,” University of Bologna, via Gobetti 93/2, 40129 Bologna, Italy
⁸ INAF, Astrophysics and Space Science Observatory Bologna, Via P. Gobetti 93/3, 40129 Bologna, Italy

^★ Corresponding author: alex.vera@userena.cl

Received: 21 March 2025
Accepted: 15 September 2025

Abstract

Context. Galactic halos host faint substructures such as stellar streams and shells, which provide insights into the hierarchical assembly history of galaxies. To date, such features have been identified in external galaxies by visual inspection. However, with the advent of larger and deeper surveys and the associated increase in data volume, this methodology is becoming impractical.

Aims. Here we aim to develop an automated method to detect low surface brightness features in galactic stellar halos. Moreover, we seek to quantify the performance of this method when considering progressively more complex datasets, including different stellar disc orientations and redshifts.

Methods. We have developed the stream automatic detection with convolutional neural networks (SAD-CNNs) models. This tool was trained on mock surface brightness maps obtained from simulations of the Auriga Project. The model incorporates transfer learning, data augmentation, and balanced datasets to optimise its detection capabilities at surface brightness limiting magnitudes ranging from 27 to 31 mag arcsec⁻².

Results. The iterative training approach, coupled with transfer learning, allowed the model to adapt to increasingly challenging datasets, achieving precision and recall metrics above 80% in all considered scenarios. The use of a well-balanced training dataset is critical for mitigating biases and ensuring that the CNN accurately distinguishes between galaxies with and without streams.

Conclusions. SAD-CNN is a reliable and scalable tool for automating the detection of faint substructures in galactic halos. Its adaptability makes it well suited to future applications that would include the analysis of data from upcoming large astronomical surveys such as LSST and JWT.