Abstract: Leaf tobacco grading involves multiple leaf characters and a great number of computation, the training model becomes very complex, a grading method on the basis of principal component analysis(PCA), genetic algorithm (GA) and support vector machine (SVM) was proposed. PCA was used to remove cross redundant information via dimension reduction, the resulted 15 leaf characters were input into SVM. The penalty parameter C and kernel function parameter g in the SVM model were optimized by GA. In combination with the actual demands of leaf tobacco grading, and comparing the identification rate with running time, PC_best and corresponding C_best and g_best were determined, and PC_best was taken as the principal component standard after dimension reduction. C_best and g_best were used as the parameter training models of SVM model, and the test set samples were tested by the trained model. The results showed that comparing with SVM model and GA-SVM model, the identification rate and grading efficiency of PCA-GA-SVM model promoted by 24.86% and 35.64%, respectively. This method provides technical supports for promoting leaf tobacco grading efficiency and accuracy.