What question did this study set out to answer?

March 26, 2026

A multi-task deep learning approach for automatic classification and redshift measurement of astronomical spectra

Key Points

To enhance the automatic classification and redshift measurement of astronomical spectra using a novel deep learning model.
Developed SpecCZ-Net, a multi-module deep learning model based on 1D-CNN.
Integrated Gabor filters and SE attention mechanism for feature extraction.
Constructed two datasets: simulated and empirical from SDSS DR17.
Evaluated model performance against baseline models.
Achieved 95.40% average classification accuracy on the empirical dataset.
F1-score for QSO_BROADLINE reached 99.40%.
Redshift measurement with R^2 up to 99.83% and NMAE as low as 0.0015.
Model confirmed excellent stability and reproducibility over five independent experiments.

Abstract

Abstract To address the demand for processing massive astronomical spectral data generated by large-scale sky survey projects, as well as the issues of existing methods—such as ambiguous classification boundaries in feature-overlapping scenarios and insufficient measurement accuracy under low SNR or high redshift conditions—this study proposes SpecCZ-Net, a deep learning model integrating multi-module collaborative feature extraction. Based on a 1D-CNN, the model incorporates Gabor filters and the SE attention mechanism and realizes the simultaneous output of celestial object types and redshift values through a main classification branch and a multi-sub-branch parallel structure. To verify the model performance, two datasets were constructed: a simulated dataset generated based on galaxy templates (covering four types: S0, Sa, Sb, Sc) and an empirical dataset from the SDSS DR17 (including five categories: Star, GalaxyAGN, GalaxySTARBURST, GalaxySTARFORMING, and QSOBROADLINE). The results demonstrate that on the empirical dataset, SpecCZ-Net achieves an average classification accuracy of 95. 40%—with the F1-score of the QSOBROADLINE category reaching as high as 99. 40%, and those of GalaxyAGN and GalaxySTARBURST being 94. 28% and 94. 01%, respectively. In terms of redshift measurement, R² can reach up to 99. 83%, the NMAE is as low as 0. 0015, and the GF remains stable within the range of 93% to 95%. Specifically, for QSOBROADLINE, the NMAE is 0. 0169 and the GF is 93. 03%. Overall, its performance is comprehensively superior to that of the other five comparative baseline models. Five independent repeated experiments confirm the model’s excellent stability and reproducibility. Finally, model visualization analysis verifies that its attention weight distribution is highly consistent with the positions of characteristic spectral lines (e. g. , O ii, O iii, Mg ii) provided by SDSS for the same spectrum. SpecCZ-Net provides a new approach for the automated and high-precision processing of massive astronomical spectral data.

Bookmark

A multi-task deep learning approach for automatic classification and redshift measurement of astronomical spectra

Key Points

Abstract

Cite This Study