T91 and 12Cr1MoV are two representative heat-resistant steel widely used in industrial pressure-bearing equipment. During long-term operation, the safety and service life of the equipment will be affected by the change of metallographic structure and mechanical properties of the steel components (i.e. material aging). In this study, the spectral data of eight T91 steel specimens with different aging grades, and seven 12Cr1MoV steel specimens with different grain size grades were obtained by laser induced breakdown spectroscopy (LIES). In order to construct steel aging estimation models, two classification models including logistic regression (LGR) and support vector machine (SVM) were used, and two representative feature selection methods including analysis of variance (ANOVA) and LGR filter were utilized to reduce the high dimensional LIES data (12,281 initial variables) into fewer features for improving the performance of estimation models. Furthermore, a new layered interval wrapper (LIW) feature selection method was proposed for being more targeted toward LIES data. The effects of different feature selection methods on model performance were compared and discussed. The characteristics of the variables selected by different feature selection methods were observed and analyzed. The results showed that the model performance of LGR and SVM can be improved to a certain degree by all three feature selections methods, and LIW showed a greatest improvement for classification prediction. For T91, the prediction accuracy of LGR/SVM coupled with LIW was improved to 0.92/0.94, while the prediction accuracy with full spectral input was 0.76/0.81; for 12Cr1MoV, the prediction accuracy using LIW was improved to 0.87/0.90, while the prediction accuracy with full spectral input was 0.69/0.69. This study demonstrates that LIW is a new effective feature selection method for high dimensional LIES data.