Combining laser induced breakdown spectroscopy(LIBS)with machine learning methods to do quantitative analysis,variable selection is the key.There are many problems in the existing variable selection methods,such as requirement for prior knowledge and large amount of computation,so a two-stage variable selection method is proposed.In the first stage,Pearson correlation coefficient,as a sort criterion,quickly excludes variables independent of the target elements concentration,the set of retained variables is recorded as S1.In the second stage,the Approximate Markov Blanket(AMB)excludes redundant variables in S1 as a search criterion,and the set of reserved variables is S2.To test the effectiveness of the method,the variable set S2 and variable set S3,which is got by the PLS-VIP,were used to establish quantitative analysis models of strontium in soil by combining three types of machine learning methods respectively,and then the quality of S2 can be tested by comparing the prediction ability of the two groups of models.The result shows that all the calibration models obtained by S2 with the R2 of more than 0.99,RE of lower than 5 %,RMSE of lower than 22 ppm and RSD of lower than 20 %,significantly better than the calibration models obtained by S3.As a result,this two-stage variable selection method can not only work effectively,but also owns certain universality when combining with different machine learning methods to do a better LIBS quantitative analysis.