Automatic spectral classification of stellar spectra with low signal-to-noise ratio using artificial neural networks ^⋆

¹ Instituto de Astronomía y Meteorología (IAM), University of Guadalajara Av. Vallarta 2602, Guadalajara, Jal., C.P. 44130, México
e-mail: silvana@astro.iam.udg.mx
² Instituto de Astrofísica de Canarias, 38200 La Laguna, Tenerife, Spain
e-mail: rcorradi@iac.es; amr@iac.es
³ Departamento de Astrofísica, Universidad de La Laguna, 38206 La Laguna, Tenerife, Spain

Received: 31 December 2010
Accepted: 20 October 2011

Abstract

Context. As part of a project aimed at deriving extinction-distances for thirty-five planetary nebulae, spectra of a few thousand stars were analyzed to determine their spectral type and luminosity class.

Aims. We present here the automatic spectral classification process used to classify stellar spectra. This system can be used to classify any other stellar spectra with similar or higher signal-to-noise ratios.

Methods. Spectral classification was performed using a system of artificial neural networks that were trained with a set of line-strength indices selected among the spectral lines most sensitive to temperature and the best luminosity tracers. The training and validation processes of the neural networks are discussed and the results of additional validation probes, designed to ensure the accuracy of the spectral classification, are presented.

Results. Our system permits the classification of stellar spectra of signal-to-noise ratio (S/N) significantly lower than it is generally considered to be needed. For S/N ≥ 20, a precision generally better than two spectral subtypes is obtained. At S/N < 20, classification is still possible but has a lower precision. Its potential to identify peculiar sources, such as emission-line stars, is also recognized.