In this paper, 3 914 near infrared spectrums of flue-cured tobacco samples was tested. These tobacco samples were collected in 17 provincial origins, including the NONG (Luzhou) flavor 865 cartons, Intermediate flavor 1 646 cartons and QING flavor (Fen) 1 646 cartons. We used near-Infrared spectroscopy and multiple algorithms fusion to analyze flue-cured tobacco flavor style features. Based on the preliminary classification of tobacco flavor according to different origins, accepting transitional and atypical flavor, tobacco flavor classification models of PPF (Projection of Basing on Principal Component and Fisher Criterion), DPLS (Partial least squares discriminant) and SVM (Support vector machine) were established respectively; and 1st and 2nd discriminant results of each algorithm can be known. Using PPF-DPLS-SVM fusion and discriminant results (1st and 2nd) of each algorithm, prediction results can be refined into typical, transitional and atypical flavor. The numbers of three flavors were 493, 392 and 115, respectively. The discriminant accuracy rate of typical flavor was improved to 927%. And it was improved 302%, 154% and 166% to compared with those achieved using PPF, DPLS and SVM, respectively. The tested samples were collected in main origins of China, which were abundant with great representativeness, therefore, the analysis result had practical application. The analysis method presented greatly improved the discriminant accuracy rate of flue-cured tobacco flavor, which was better than that of the classification according to objective data. The method refining flue-cured tobacco into typical, transitional and atypical flavor, provided guidance to the scientific application and module industrial processing of raw flue-cured tobacco.