Classification for Unrecognized Spectra in LAMOST DR6 Using Generalization of Convolutional Neural Networks

Abstract

Commonly used template classification for celestial spectra always fails dealing with low signal-to-noise ratio (S/N) spectra, which are very numerous in spectroscopic surveys. In the sixth data release of Large sky Area Multi-Object Fiber Spectroscopic Telescope (LAMOST DR6 V1), more than 0.7 million bad quality data were refused to classify by LAMOST pipeline and archived as “UNKNOWN.” To recognize as many objects with low S/N spectra as possible in the “UNKNOWN” data set, one-dimensional convolutional neural network (CNN) based classifier was adapted from the widely used two-dimensional CNN. In this work, two CNN based classifier were applied, a classifier for distinguishing galaxy, QSO and star, and a classifier for discriminating subtypes of stars. To solve the problem caused by imbalanced training samples among different classes for the stellar classifier, a semi supervised learning algorithm by two CNNs and Spectral Generative Adversarial Network (SGAN) was introduced to produce artificial spectra for the minority O type. The SGAN solution is better than over-sampling in solving overfitting caused by imbalanced training set. The trained CNN classifiers were applied to classify “UNKNOWN” spectra into candidates of galaxies, QSOs, and stars. and further classify star candidates into spectral subclasses of O to M. Each spectra can be recognized to a spectral type with a probability by CNN algorithm, and 101,082 stellar spectra were remained with the probability larger than 99%, making up a supplemental star catalog of LAMOST DR6, which includes 294 O, 2 850 B, 269 A, 6 431 F, 626 G, 60 527 K, and 30 085 M types. To verify the catalog, the distances to corresponding templates from recognized spectra in each class were also checked comparing with known spectra. In addition, 200 O type stars were manually confirmed from 294 automatically identified O type stars in the catalog, because O type spectra have weak features and easily to be confused with no signal spectra. The classification result as a part of this work are available at http://paperdata.china-vo.org/Classification_SGAN/result.zip.