Measurement Of A-type Stellar Atmospheric Parameters Based On Ensemble Tree Model

At present,as the astronomical telescope with the highest celestial spectra acquisition rate in the world,the LAMOST is a powerful tool for large field-view and large sample astronomical research.Up to now,the latest data set released by LAMOST named as LAMOST-DR6 includes more than 10 million spectra.And it is the first telescope device in the world which has achieved more than ten million orders of magnitude spectra data.In this thesis,we mainly focus on the research about the automatic algorithms to process the massive celestial spectral data obtained by LAMOST,especially focus on the automatic estimation algorithms of the stellar atmospheric parameters.About the feature information of a stellar spectrum,its line indices are able to be used to represent the stellar physical features.So,applying the relations between the line indices and stellar atmospheric parameters to estimate the three parameters is effective,at the same time,we can also solve the problem of high computational complexity of the high-dimensional spectral data.Therefore,in this thesis,according to the above ideas and based on the spectral data of the LAMOST,the two integrated tree models called as Extremely Randomized Trees(Extra-Trees)and Extreme Gradient Boosting(XGBoost)are applied in the automatic estimation of Atype stellar atmospheric physical parameters.This thesis mainly includes the following contents:(1)An algorithm to estimate the stellar atmospheric parameters based on ExtraTrees model is designed and implemented.And a series of experiments are performed on the A-type stellar spectra data set from LAMOST-DR6.The correlation analysis between the 26 line indices and the three stellar parameters: effective temperature(Teff),surface gravity(log g)and metallicity([Fe/H])are performed in detail.The analysis results show that different parameters have different dependencies on the 26 line indices,so,in this thesis,we apply the line indices with higher correlation as the input features of Extra-Trees model to estimate the three parameters separately.On the other hand,the interaction between three physical parameters and the peak wavelength of the continuum is analyzed in this thesis,and the result shows that the peak wavelength can only affect the estimated accuracy of the Teff and [Fe/H].The final experimental results show that the accuracy of parameters estimation is better than the other two algorithms: Random Forest(RF)and Support Vector Machines Regression(SVR).(2)An algorithm to estimate the stellar atmospheric parameters based on the XGBoost model is designed and implemented in this thesis.The XGBoost is an improved algorithm of the traditional Gradient Boosting Decision Tree(GBDT)algorithm.And it can be used to deal with high dimensional sparse features in a distributed way,which has the advantages of fast running speed,high accuracy and avoiding over-fit.Firstly,in this thesis,the grid search algorithm is applied to optimize the main parameters of XGBoost model.Secondly,we also consider the interrelations among the three output atmospheric parameters.Based on our experimental results,we design an input-output orders of each parameter and the other two parameters,and this design is able to increase the input feature information of the estimation algorithm.Finally,a series of experiments are designed and preformed,compared with ExtraTrees,the traditional GBDT and SVR algorithms,the experimental results of XGBoost show that the estimated accuracy of the three parameters is better than the other three algorithms.