论文标题
从低SNR的LAMOST低分辨率光谱中估算大气参数
Estimating Atmospheric Parameters from LAMOST Low-Resolution Spectra with Low SNR
论文作者
论文摘要
大天空区多对象光纤光谱望远镜（Lamost）获得了数千万的低分辨率恒星光谱。大量光谱导致探索自动大气参数估计方法的紧迫性。有很多具有低信噪比（SNR）的杆光谱，导致其估计准确性急剧下降。因此，有必要探索低SNR光谱的更好估计方法。本文提出了一种基于神经网络的方案，以提供大气参数，即lasso-mlpnet。首先，我们采用一种多项式拟合方法来获取伪符号并将其删除。然后，使用最少的绝对收缩和选择算子（LASSO）检测到高噪声中的某些参数敏感的特征。最后，lasso-mlpnet使用多层感知器网络（MLPNET）来估计大气参数$ t _ {\ mathrm {eff}} $，log $ g $和[fe/h]。在Apogee（Apache Point天文台银河系进化实验）和Lamost之间的某些lamost恒星光谱上评估了Lasso-Mlpnet的有效性。结果表明，估计精度在恒星光谱上以$ 10 <\ mathrm {snr} \ leq80 $显着提高。尤其是，套索 - Mlpnet从（144.59 k，0.236 dex，0.108 dex）（lasp）（90.29 k，0.152 dex dex dex dex dex dex dex dex，log $ g $ and [fe/h]减少$ t _ {\ mathrm {eff}} $估计的平均绝对错误（MAE）至（90.29 k，0.152 dex，lass），last dex，0.152 dex，last）。售价为$ 10 <\ mathrm {snr} \ leq20 $。为了促进参考，我们以$ 10 <\ Mathrm {snr} \ leq80 $和3500 <snr $ g $ $ \ $ \ $ 6500的估计来释放超过482万个恒星光谱的Lasso-Mlpnet估计值。
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.
