Abstract: By using the Wrapper feature extraction algorithm in data mining methods, the key indicators were extracted, which significantly impacted the aroma quality of flue-cured tobacco. The method was compared with the simple correlation coefficient method. The results showed that there were 20 kinds of aroma components which mainly impacted the aroma type, such as geranylaceione, pentanoic acid, total acid, megastigmatrienone, solanone, while 25 types of aroma components which mainly impacted the volume of aroma, such as geranylacetone, 2-acetyl pyrrole, tetramethylpyrazine, benzaldehyde, solanone, isovaleric acid, β-dama-scenone. There were 18 types of aroma compontents which mainly impacted the quality of aroma, such as megastigmatrienone, isovaleric acid, solanone, total neutral, neophytadiene, geranylacetone, farnesylacetone, 3-methylpentanoic acid, tetramethylpyrazine, β-damascenone, pyridine. The performance of the model was improved greatly by using the extracted features, which were built between aroma components and aroma quality. Compared with piecewise linear regression model, the RBF neural network model has better accuracy in prediction.