When laser induced breakdown spectroscopy (LIBS) is used for detection, due to the many and complex spectral lines, there are much redundant information, which will affect the quantitative analysis. Therefore, extracting effective feature variables is of great significance in the quantitative analysis of LIBS. In this paper, the method of selecting the spectral characteristics of the Ca element in the CaCl₂ solution was analyzed, and the accuracy and stability of the univariate model, partial least square regression and CART regression tree calibration model were compared. In view of the large volatility of the surface of the water body, the poor spectral stability, and the fact that the spectrum is affected by the matrix effect and the self-absorption effect, the fitting coefficient (R²) obtained by the univariate model is only 0.933 2, and the training root mean square error (RMSEC), prediction root mean square error (RMSEP) and average relative error (ARE) are 0.019 2 Wt%, 0.017 7 Wt% and 11.604% respectively. After partial least squares regression optimization, the model R² is increased to 0.975 3, and RMSEC, RMSEP and ARE are reduced to 0.010 8 Wt%, 0.013 Wt% and 7.49%, respectively. Although the model’s accuracy has been improved, it is still difficult to meet the analysis requirements. In order to further improve the accuracy of quantitative analysis, a CART regression tree calibration model was established. When constructing the tree model, this method uses the square error minimization criterion to select the optimal combination of characteristic variables from the complex spectral information to make classification decisions, thereby establishing the calibration curve of the Ca element. Through the variable selection of the CART regression tree, the number of characteristic variables is reduced from 100 to 6, and the compression rate of variables reaches 94%, which significantly reduces the interference of irrelevant spectral lines. The correlation coefficients of the regression tree model are R², RMSEC, RMSEP and ARE is 0.997 5, 0.003 5 Wt%, 0.006 1 Wt% and 2.500%, respectively. Compared with the traditional univariate and partial least square regression, the CART regression tree model has higher accuracy and lower error. Through effective screening of characteristic variables, this paper eliminates the interference of irrelevant signals, significantly reduces the influence of matrix effect and self-absorption effect on LIBS quantitative analysis, and improves the accuracy and stability of quantitative analysis.