2024MNRAS.527.1163W

Emission-line galaxy classification plays an important role in comprehending the formation and evolution of galaxies. The widely used optical spectral classification method for galaxies is the BPT diagram, which classifies emission-line galaxies on the basis of precise spectral line measurements. Various classical machine learning methods have been utilized to classify galaxy spectra. Deep learning (DL) is more feasible for a huge amount of data, as it can learn patterns autonomously from the original data. This study aims to explore the possibility of applying DL to classify galaxy spectra and improve classification efficiency. A one-dimensional convolutional neural network model called GalSpecNet was constructed to classify emission-line galaxy spectra, which recognizes star-forming, composite, active galactic nucleus (AGN), and normal galaxies with an accuracy of over 93 per cent. This study employs the Gradient-weighted Class Activation Mapping to elucidate the decision-making process of the model by inspecting spectral features that the model prioritizes for each type of galaxy. The findings suggest that the model considers features highly consistent with the conventional BPT method. Subsequently, we applied the model to the cross-matched galaxies of Sloan Digital Sky Survey Data Release 16 (DR16) and Large Sky Area Multi-Object Fiber Spectroscopic Telescope DR8 and present a catalogue comprising of 41 699 star-forming candidates and 55 103 AGN candidates. The catalogue is publicly available.