基于实测光谱的恒星大气物理参数估计是探索恒星本质的首要任务。 随着郭守敬望远镜(LAMOST)进入正式巡天阶段, 正以前所未有的速度获取海量的恒星实测光谱数据, 这为星系研究带来了新的机遇和挑战。 由于LAMOST是多目标光纤光谱天文望远镜, 获取的光谱噪音比较大。 光谱前期处理中的波长定标和流量定标精度不高, 导致光谱存在微小畸变, 这些都大大增加了恒星大气物理参数测量的难度。 如何对LAMOST实测光谱的恒星大气物理参数进行自动测量是迫切期待需要研究的一个重要课题, 关键是如何消除噪声, 提高恒星大气物理参数的测量精度和鲁棒性。 提出了一个测量LAMOST恒星光谱大气参数的回归模型(SVM(lasso))。 基本思路是: 首先使用Haar小波对光谱信号进行滤波, 抑制光谱中噪声的不利影响, 最大限度地保留光谱判别信息。 然后采用lasso算法进行特征选择, 选取与恒星大气物理参数相关性强的特征。 最后将选择的光谱特征输入支持向量机回归模型对恒星大气物理参数进行估计, 该模型对光谱畸变和噪音的容忍性比较好, 提高了测量的精确度。 为了验证上述方案的可行性, 在33 963条LAMOST先导巡天恒星光谱库上作了实验研究, 三个恒星大气物理参数的精度分别为log Teff: 0.006 8 dex, log g: 0.155 1 dex, ［Fe/H］: 0.104 0 dex。

It is a key task to estimate the atmospheric parameters from the observed stellar spectra in exploring the nature of stars and universe. With our Large Sky Area Multi-Object Fiber Spectroscopy Telescope (LAMOST) which begun its formal Sky Survey in September 2012, we are obtaining a mass of stellar spectra in an unprecedented speed. It has brought a new opportunity and a challenge for the research of galaxies. Due to the complexity of the observing system, the noise in the spectrum is relatively large. At the same time, the preprocessing procedures of spectrum are also not ideal, such as the wavelength calibration and the flow calibration. Therefore, there is a slight distortion of the spectrum. They result in the high difficulty of estimating the atmospheric parameters for the measured stellar spectra. It is one of the important issues to estimate the atmospheric parameters for the massive stellar spectra of LAMOST. The key of this study is how to eliminate noise and improve the accuracy and robustness of estimating the atmospheric parameters for the measured stellar spectra. We propose a regression model for estimating the atmospheric parameters of LAMOST stellar(SVM(lasso)). The basic idea of this model is: First, we use the Haar wavelet to filter spectrum, suppress the adverse effects of the spectral noise and retain the most discrimination information of spectrum. Secondly, We use the lasso algorithm for feature selection and extract the features of strongly correlating with the atmospheric parameters. Finally, the features are input to the support vector regression model for estimating the parameters. Because the model has better tolerance to the slight distortion and the noise of the spectrum, the accuracy of the measurement is improved. To evaluate the feasibility of the above scheme, we conduct experiments extensively on the 33 963 pilot surveys spectrums by LAMOST. The accuracy of three atmospheric parameters is log Teff: 0.006 8 dex, log g: 0.155 1 dex, ［Fe/H］: 0.104 0 dex.