Estimating Atmospheric Parameters from LAMOST Low-Resolution Spectra with Low SNR
Large Sky Area Multi-Object Fiber Spectroscopic Telescope (LAMOST) acquired tens of millions of low-resolution stellar spectra. The large amount of the spectra result in the urgency to explore automatic atmospheric parameter estimation methods. There are lots of LAMOST spectra with low signal-to-noise ratios (SNR), which result in a sharp degradation on the accuracy of their estimations. Therefore, it is necessary to explore better estimation methods for low-SNR spectra. This paper proposed a neural network-based scheme to deliver atmospheric parameters, LASSO-MLPNet. Firstly, we adopt a polynomial fitting method to obtain pseudo-continuum and remove it. Then, some parameter-sensitive features in the existence of high noises were detected using Least Absolute Shrinkage and Selection Operator (LASSO). Finally, LASSO-MLPNet used a Multilayer Perceptron network (MLPNet) to estimate atmospheric parameters $T_{\mathrm{eff}}$, log $g$ and [Fe/H]. The effectiveness of the LASSO-MLPNet was evaluated on some LAMOST stellar spectra of the common star between APOGEE (The Apache Point Observatory Galactic Evolution Experiment) and LAMOST. it is shown that the estimation accuracy is significantly improved on the stellar spectra with $10<\mathrm{SNR}\leq80$. Especially, LASSO-MLPNet reduces the mean absolute error (MAE) of the estimation of $T_{\mathrm{eff}}$, log $g$ and [Fe/H] from (144.59 K, 0.236 dex, 0.108 dex) (LASP) to (90.29 K, 0.152 dex, 0.064 dex) (LASSO-MLPNet) on the stellar spectra with $10<\mathrm{SNR}\leq20$. To facilitate reference, we release the estimates of the LASSO-MLPNet from more than 4.82 million stellar spectra with $10<\mathrm{SNR}\leq80$ and 3500 < SNR$g$ $\leq$ 6500 as a value-added output.