Directly Deriving Parameters from SDSS Photometric Images

Abstract

Stellar atmospheric parameters (effective temperature, surface gravity, and metallicity) are fundamental for understanding the formation and evolution of stars and galaxies. Photometric data can provide a low-cost way to estimate these parameters, but traditional methods based on photometric magnitudes have many limitations. In this paper, we propose a novel model called Bayesian Convit, which combines an approximate Bayesian framework with a deep-learning method, namely Convit, to derive stellar atmospheric parameters from Sloan Digital Sky Survey images of stars and effectively provide corresponding confidence levels for all the predictions. We achieve high accuracy for T _eff and [Fe/H], with σ ( T _eff ) = 172.37 K and σ ([Fe/H]) = 0.23 dex. For $\mathrm{log}g$ , which is more challenging to estimate from image data, we propose a two-stage approach: (1) classify stars into two categories based on their $\mathrm{log}g$ values (>4 dex or <4 dex) and (2) regress separately these two subsets. We improve the estimation accuracy of stars with $\mathrm{log}g\gt 4$ dex significantly to $\sigma (\mathrm{log}g\gt 4)=0.052$ dex, which are comparable to those based on spectral data. The final joint result is $\sigma (\mathrm{log}g)=0.41$ dex. Our method can be applied to large photometric surveys like Chinese Space Station Telescope and Large Synoptic Survey Telescope.

Keywords

• Astronomy image processing
• Stellar photometry
• Stellar abundances
• Neural networks
• Bayes’ Theorem