FNet II: Spectral Classification of Quasars, Galaxies, Stars, and broad absorption line (BAL) Quasars

TL;DR

This paper introduces FNet II, an enhanced CNN with ResNet architecture, capable of classifying SDSS spectra into quasars, galaxies, stars, and BAL-quasars with high accuracy without needing to identify specific spectral lines.

Contribution

FNet II extends the original FNet by incorporating ResNet architecture and multi-class output, enabling autonomous spectral classification of various celestial objects in SDSS data.

Findings

Achieves over 98.5% completeness across classes.

Comparable performance to QuasarNET in spectral classification.

Does not require explicit line identification for classification.

Abstract

In this work, we enhance the FNet, a 1-dimensional convolutional neural network (CNN) with a residual neural network (ResNet) architecture, to perform spectral classification of quasars, galaxies, stars, and broad absorption line (BAL)-quasars in the SDSS-IV catalog from DR17 of eBOSS. Leveraging its convolutional layers and the ResNet structure with different kernel sizes, FNet autonomously identifies various patterns within the entire sample of spectra. Since FNet does not require the intermediate step of identifying specific lines, a simple modification enabled our current network to classify all SDSS spectra. This modification involves changing the final output layer from a single value (redshift) to multiple values (probabilities of all classes), and accordingly adjusting the loss function from mean squared error (MSE) to cross-entropy. FNet achieves a completeness of 99.00\% $\pm$ 0.20 for galaxies, 98.50\% $\pm$ 0.30 for quasars, 99.00\% $\pm$ 0.18 for BAL-quasars, and 98.80\% $\pm$ 0.20 for stars. These results are comparable to those obtained using QuasarNET, a standard CNN employed in the SDSS routine, comprises convolutional layers without the ResNet structure with equal kernel sizes, and is utilized for redshift measurement and classification by identifying seven emission lines. QuasarNET, in order to overcome the problem of finding a CIV emission line with broad absorption which is slightly more challenging than that of detecting emission lines requires to add BAL CIV line to the list of lines that the network learns to identify. However, this procedure is not necessary in FNet as it learns the features through a self-learning procedure.