Galaxy Spectra neural Network (GaSNet). II. Using deep learning for spectral classification and redshift predictions

The size and complexity reached by the large sky spectroscopic surveys require efficient, accurate, and flexible automated tools for data analysis and science exploitation. We present the Galaxy Spectra Network/GaSNet-II, a supervised multinetwork deep learning tool for spectra classification and redshift prediction. GaSNet-II can be trained to identify a customized number of classes and optimize the redshift predictions. Redshift errors are determined via an ensemble/pseudo-Monte Carlo test obtained by randomizing the weights of the network-of-networks structure. As a demonstration of the capability of GaSNet-II, we use 260k Sloan Digital Sky Survey spectra from Data Release 16, separated into 13 classes including 140k galactic, and 120k extragalactic objects. GaSNet-II achieves 92.4 per cent average classification accuracy over the 13 classes and mean redshift errors of approximately 0.23 per cent for galaxies and 2.1 per cent for quasars. We further train/test the pipeline on a sample of 200k 4MOST (4-metre Multi-Object Spectroscopic Telescope) mock spectra and 21k publicly released DESI (Dark Energy Spectroscopic Instrument) spectra. On 4MOST mock data, we reach 93.4 per cent accuracy in 10-class classification and mean redshift error of 0.55 per cent for galaxies and 0.3 per cent for active galactic nuclei. On DESI data, we reach 96 per cent accuracy in (star/galaxy/quasar only) classification and mean redshift error of 2.8 per cent for galaxies and 4.8 per cent for quasars, despite the small sample size available. GaSNet-II can process ∼40k spectra in less than one minute, on a normal Desktop GPU. This makes the pipeline particularly suitable for real-time analyses and feedback loops for optimization of Stage-IV survey observations.