Astronomical Data Analysis Based On Machine Learning Method

Machine learning has an incredible ability to classification and regression.They are able to identify the complicated nonlinear behavior of multi-dimensional feature space,providing accurate solutions for some problems in astronomy.In this dissertation,we use machine learning algorithms research star/stars/QSO classification and spectral type classification and estimate the stellar surface temperature.The specific work content is as follows:This dissertation utilizes 6)-nearest neighbors algorithm(6)NN),decision tree(DT),random forest(RF)and support vector machine(SVM)for star/galaxy/QSO classification on the light source of LAMOST DR5.Select u,g,r,i,z,J,H as the algorithm input features.The results show that the above four algorithm models can identify the star source type with ～95% accuracy.After screening out the stars,this paper continues to use the above four mainstream machine learning algorithms to identify the spectral types of stars(G,K,M,F,A).The results show that the areas under the ROC(receiver operating characteristic)curves of the four algorithm models are all close to 1,and all four models can identify the spectral types of stars to a certain extent.The recognition accuracy of RF and SVM is higher than that of 6)NN and DT.Random forest can identify stellar spectral types with an accuracy of ～73%.The reason is that G stars and F stars are seriously crossed and misclassified,and the prediction effect of the two stars is not as good as that of other three stars.This dissertation uses random forests to estimate the effective surface temperature of stars.The stars are mainly derived from APOGEE and LAMOST DR6 of SDSS DR16.In addition,this dissertation also compares and analyzes the prediction effects of the two input feature sets.When using NUV-u,u-g,g-r,r-i,i-J,JH,H-K,K-WISE＿4＿5 as the algorithm input features,the accuracy of random forest on APOGEE stars is as high as 94.91%,and the estimated temperature is the same as the standard given temperature.The deviation of 93.89K;the accuracy of random forest on LAMOST stars is as high as 90.46%,and the deviation of its estimated temperature from the standard given temperature is 113.10 K.When NUV-J,J-H,H-K,K-WISE＿4＿5 are used as the input features of the algorithm,the accuracy of random forest on APOGEE stars is as high as 94.37%,and the deviation between the estimated temperature and the temperature given by the standard is 96.59K;The accuracy is as high as 88.89% on the LAMOST stars,and its estimated temperature deviates from the standard given temperature by 119.92 K.Compared with the former,the accuracy will not decrease too much,and NUV,J,H,K,and WISE＿4＿5 are all derived from all-sky observations,which can be more widely used in estimating the effective temperature of stars.Moreover,J-H is the most important feature for estimating effective temperature,followed by NUV-J.NUV is a non-negligible band.