Classification of Spectra of Emission-line Stars Using Feature Extraction Based on Wavelet Transform

Our goal is to automatically identify spectra of emission (Be) stars in large archives and classify their types based on a typical shape of the H-alpha emission line. Due to the length of spectra, classification of the original data is very time-consuming. In order to lower computational requirements and enhance the separability of the classes, we have to find a reduced representation of spectral features, however conserving most of the original information content. As the Be stars show a number of different shapes of emission lines, it is not easy to construct simple criteria (like e.g. Gaussian fits) to distinguish the emission lines in an automatic manner. We proposed to perform the wavelet transform of the spectra, calculate statistical metrics from the wavelet coefficients, and use them as feature vectors for classification. In this paper, we compare different wavelet transforms, different wavelets, and different statistical metrics in an attempt to identify the best method.