SDSS-DR12 bulk stellar spectral classification: Artificial neural networks approach

Abstract

This paper explores the application of Probabilistic Neural Network (PNN), Support Vector Machine (SVM) and K-means clustering as tools for automated classification of massive stellar spectra. The data set consists of a set of stellar spectra associated with the Sloan Digital Sky Survey (SDSS) SEGUE-1 and SEGUE-2, which consists of 400,000 data from 3850 to 8900 Å with 3646 data points each. We investigate the application of principal components analysis (PCA) to reducing the dimensionality of data set to 280, 400 and 700 components. We show that PNN can give fairly accurate spectral type classifications $\sigma _{\mathit{RMS}}=1.752$, $\sigma_{\mathit{RMS}}=1.538$ and $\sigma_{\mathit{RMS}}=1.391$ and K-means can classify these spectra with an accuracy of $\sigma_{\mathit{RMS}}=1.812$, $\sigma_{\mathit{RMS}}=1.731$ and $\sigma_{\mathit{RMS}}=1.654$ and SVM with the accuracy of $\sigma_{\mathit{RMS}}=1.795$, $\sigma_{\mathit{RMS}}=1.674$ and $\sigma_{\mathit{RMS}}=1.529$ across the 280, 400 and 700 components, respectively. By using K-means the classification of the spectra renders 38 major classes. Furthermore, by comparing the results we noticed that PNN is more successful than K-means and SVM in automated classification.