In this paper, the laser-induced breakdown spectroscopy and random forest are combined to study the GCr15 steel samples with different hardness. Through the linear relationship between the spectral line intensity ratio (I_{Fe Ⅱ 316.786}/I_{Fe Ⅰ 375.745} and I_{Cr Ⅱ 482.413}/I_{Cr Ⅰ 302.067}) and the sample hardness, the experimental results show that the spectral line intensity ratio of the matrix element to the alloy element and the hardness of the sample present a linear correlation, in which the linear correlation between I_{Fe Ⅱ 316.786}/I_{Fe Ⅰ 375.745} and hardness is higher, indicating that the method of the spectral line intensity ratio has a certain dependence on the selection of spectral lines. The LIBS-RF method is proposed to estimate the hardness of the sample. First, PCA is used to reduce the dimensionality of the original data, and subsequently a random forest model is established. It is found that the hardness of the sample can not be effectively predicted. Then, the random forest is used to select feature spectra based on the importance of the variables to establish a random forest model. The results show that the prediction accuracy of the LIBS-RF model based on the full-spectrum data is lower than that based on the partial characteristic spectrum data. This is because the full-spectrum data also contains a lot of noise and other redundant information, which also participates in the training of the model and results in a decrease in model accuracy. In addition, it is found that as the number of decision trees and the number of random features increase, the prediction accuracy of the model increases to a certain value and remains relatively stable. Based on this, the parameters of the model can be adjusted reasonably while satisfying the accuracy requirements, the complexity of the algorithm is reduced, and the efficiency of the algorithm is improved. Above all, as a novel hardness measurement technology, LIBS-RF has the advantages of simpler and faster than the traditional LIBS hardness measurement, and the results here provide a theoretical basis for engineering practice applications.