Automated classification of stellar spectra

Abstract

We have analyzed the performance of a PCA-based automated classifier of stellar spectra into the MK system, using as benchmark the dataset of optical spectra listed in the SDSS-DR15. We have found that it is possible to account for 99% of the total variance arising from good-quality A, F, G, K, and M stellar spectra with only three principal components, though we have ended up using four to further increase the discriminant power of our methodology. The projections of a subset of 50; 000 good-quality spectra on this 4D space have been used to determine the most probable spectral type of test stars in samples of spectra of increasing quality, with which we have evaluated the goodness of our classification procedure. Within a general scenario of excellent results, we found that a Gaussian Kernel performs somewhat better than a Top-Hat Kernel when calculating membership probabilities, that the efficiency of our classification method improves with the S=N of the spectra, and that the classification of G-type stars is the less reliable and that of F-type stars the most incomplete.