Machine learning Applied to Star-Galaxy-QSO Classification and Stellar Effective Temperature Regression

Abstract

In modern astrophysics, the machine learning has increasingly gained more popularity with its incredibly powerful ability to make predictions or calculated suggestions for large amounts of data. We describe an application of the supervised machine-learning algorithm, random forests (RF), to the star/galaxy/QSO classification and the stellar effective temperature regression based on the combination of LAMOST and SDSS spectroscopic data. This combination enable us to obtain reliable predictions with one of the largest training sample ever used. The training samples are built with nine-color data set of about three million objects for the classification and seven-color data set of over one million stars for the regression. The performance of the classification and regression is examined with the validation and the blind tests on the objects in the RAVE, 6dFGS, UVQS and APOGEE surveys. We demonstrate that the RF is an effective algorithm with the classification accuracies higher than 99\% for the stars and the galaxies, and higher than 94\% for the QSOs. These accuracies are higher than the machine-learning results in the former studies. The total standard deviations of the regression are smaller than 200 K that is similar to those of some spectrum-based methods. The machine-learning algorithm with the broad-band photometry provides us a more efficient approach to deal with massive amounts of astrophysical data than traditional color-cuts and SED fit.