Atmospheric parameter measurement of Low-S/N stellar spectra based on deep learning

Deriving accurate atmospheric parameters from stellar spectra is of fundamental importance for stellar research. At present, machine learning, such as multiple linear regression, artificial neural networks (ANN), and support vector machines, have been widely used to derive atmospheric parameters. However, these methods are generally only applicable to estimate the atmospheric parameters of high signal-to-noise ratio (high-S/N) stellar spectra. For low signal-to-noise ratio (low-S/N) stellar spectra, these methods tend to perform poorly. In order to address the problem, we propose a one-dimensional convolutional neural network StarNet. The proposed method includes the following three steps: firstly, select the spectra with S/N< = 15 as the input spectra; secondly, extract representative features from stellar spectra through two convolutional layers and a max pooling layer; thirdly, learn the mapping function from spectra to atmospheric parameters through two fully connected layers. We evaluate the proposed method on both synthetic spectra calculated from Kurucz models and observed spectra from the Large Sky Area Multi-Object Fiber Spectroscopic Telescope (LAMOST), as well as compare it with other commonly used methods including Lick + OLS and Wavelet + ANN. Experiments show that the proposed method is efficient in estimating the atmospheric parameters of low-S/N stellar spectra and more accurate than other methods.