Classifying spectra of emission-line regions with neural networks

An application to integral field spectroscopic data of M33

¹ Dipartimento di Fisica e Astronomia, Università di Firenze, Via G. Sansone 1, I-50019 Sesto F.no, (Firenze), Italy
² INAF – Osservatorio Astrofisico di Arcetri, Largo E. Fermi 5, I-50125 Florence, Italy
³ INFN – Istituto Nazionale di Fisica Nucleare (INFN), Via Bruno Rossi 1, 50019 Sesto Fiorentino, (FI), Italy
⁴ ICSC – Centro Nazionale di Ricerca in High Performance Computing, Big Data & Quantum Computing, Via Magnanelli 2, 40033 Casalecchio di Reno, (BO), Italy
⁵ University of Trento, Via Sommarive 14, I-38123 Trento, Italy
⁶ Max-Planck-Institut für Extraterrestrische Physik (MPE), Gießenbachstraße 1, 85748 Garching, Germany
⁷ Scuola Normale Superiore, Piazza dei Cavalieri 7, I-56126 Pisa, Italy

^⋆ Corresponding author: caterina.bracci@unifi.it

Received: 20 February 2025
Accepted: 26 March 2025

Abstract

Emission-line regions are key to understanding the properties and evolution of galaxies, as they trace the exchange of matter and energy between stars and the interstellar medium (ISM). In nearby galaxies, individual nebulae can be resolved from the background, and classified through their optical spectral properties. Traditionally, the classification of these emission-line regions in H II regions, planetary nebulae (PNe), supernova remnants (SNRs), and diffuse ionised gas (DIG) relies on criteria based on single or multiple emission-line ratios. However, these methods face limitations due to the rigidity of the classification boundaries, the narrow scope of information they are based upon, and the inability to take line-of-sight superpositions of emission-line regions into account. In this work, we explore the use of artificial neural networks to classify emission-line regions using their optical spectra. Our training set consists of simulated optical spectra, obtained from photoionisation and shock models, and processed to match observations obtained with the MUSE integral field spectrograph at the ESO/VLT. We evaluate the performance of the network on simulated spectra exploring a range of signal-to-noise (S/N) levels, different values for dust extinction, and the superposition of different nebulae along the line of sight. At infinite S/N the network achieves perfect predictive performance, while, as the S/N decreases, the classification accuracy declines, reaching an average of ∼80% at S/N(Hα) = 20. We then apply our model to real spectra from MUSE observations of the Local Group galaxy M33. The network provides a robust classification of individual spaxels, even at low S/N, identifying H II regions and PNe and distinguishing them from SNRs and diffuse ionised gas, while identifying overlapping nebulae. Moreover, we compare the network’s classification with traditional diagnostics, finding a satisfactory level of agreement between the two approaches. We identify the emission lines that are most relevant for our classification tasks using activation maximisation maps. In particular, we find that at high S/N the model mainly relies on weak lines (e.g. auroral lines of metal ions and He recombination lines), while at the S/N level typical of our dataset the model effectively emulates traditional diagnostic methods by leveraging strong nebular lines. We discuss potential future developments focused on deriving segmentation maps for H II regions and other nebulae, and the extension of our method to new datasets and instruments.