基于核岭回归方法的恒星大气物理参数的自动测量

我国大科学工程项目LAMOST巡天计划每观测夜能获取多达数万条天体光谱数据,天文学家通过对天体光谱的分析观察可以获取有效的天文信息用于天文学或天体物理学的研究。而针对海量数据,寻找自动方法分析天体光谱并进行天体各种物理参数的测量就具有重要研究意义和价值。这一课题也吸引了许多学者进行研究,但目前所尝试的算法和相应结果仍然需要进一步改进,针对这一需求深入研究了核岭回归(KRR)方法在恒星大气物理参数(包括有效温度、表面重力和金属丰度)自动测量方面的应用,特别是在我国大科学工程项目LAMOST所释放光谱数据上的应用。核岭回归是岭回归算法的进一步发展,而岭回归是最小二乘方法的一种变形,其具有解决高维多重共线性问题的能力。所以KRR方法适合于处理高维的天体光谱信息,从LAMOST的第五期释放数据中随机选择了2万条被识别为恒星的光谱数据用于实验测试,该数据既包含低信噪比数据,也包含高信噪比数据(g,r,i波段平均信噪比最低至6.7,最高到793)。首先,本文对光谱进行预处理,包括三个步骤:(1)利用小波变换对光谱数据进行去噪处理;(2)因为LAMOST采用的是后期修正的流量定标设计,所以还通过流量归一化来避免部分光谱流量值不准确的问题;(3)由于每条光谱维数高达数千维,利用主成分分析方法(PCA)对光谱进行了降维。然后,利用KRR方法建立了光谱数据和标准化后的三大参数值之间的回归模型。最后,通过设计进行不同的组合实验对KRR算法模型进行了测试分析,并与经典算法支持向量回归(SVR)进行了对比。综合所有实验结果显示KRR方法对应的有效温度、表面重力和金属丰度的测试平均绝对误差分别为82.9897 K,0.1858 dex和0.1211 dex,优于SVR的144.2308 K,0.1886 dex和0.1246 dex。特别是KRR在温度测试结果上有较大优势,由此表明KRR方法能够有效地应用于天体光谱特别是恒星光谱参数的自动测量处理中。

Automatic Measurement of Stellar Atmospheric Physical Parameters Based on Kernel Ridge Regression Method

Through observing and analyzing these celestial spectra,astronomers can obtain effective astronomical information to research the theory of astronomy and astrophysics.And a large scientific engineering project in China,the LAMOST survey program can obtain a large number of celestial spectra every observation night.For the massive data,it is very important to find an automatic method to analyze the celestial spectrum and measure various physical parameters of the celestial body.In fact,many scholars had been attracted to research this topic before,however,the current algorithms and corresponding results that were presented by them cannot meet the accuracy of manual measurement,and it means that we should find more appropriate algorithms to improve the effect of automatic processing.In this paper,we study deeply the applications of the Kernel Ridge Regression(hereinafter referred to as KRR)method in the automatic measurement of the stellar atmosphere physical parameters(including temperature,gravity and chemical abundance),especially the applications in the spectral data released by LAMOST.The KRR is a further development of the Ridge Regression algorithm,and the Ridge Regression is a variant of the Least Squares Method with a regularization term,it has an ability to solve high dimensional multi-collinearity problems.Therefore,the KRR method is suitable for processing high-dimensional celestial spectral information.In this paper,20000 stellar spectral data identified as stars are randomly selected from the release data of LAMOST for experimental testing.The data set contain spectra with low SNR and high SNR(the average SNR in g,r,i-band are from 6.7 up to 793).In this paper,we preprocess all the spectra firstly,including three steps:one is the de-noise phase based on the wavelet transform;In order to avoid some inaccuracies in the spectral flux value,the second step is spectral flux normalization;Because the dimension of each spectrum is up to several thousand dimensions,the principal component analysis method(PCA)is used to reduce the spectral dimensions as the third step.Then,we establish a regression model between the spectral data and the normalized three stellar parameters based on the KRR method.Finally,we design many different combinations of experiments to test and analyze the KRR model,and compare its results with the results of the classical algorithm SVR.All the experimental results using KRR method show that the average absolute error of temperature,gravity and chemical abundance is 82.9897 K,0.1858 dex and 0.1211 dex,respectively,it is better than the result of the SVR which is 144.2308 K,0.1886 dex and 0.1246 dex,respectively.In particular,the KRR method has a large advantage in temperature test results,which indicates that the KRR method can be effectively applied to the automatic measurement of the stellar spectral parameters.