Abstract: To overcome the technical bottleneck that traditional tobacco leaf grading and blending design relied on subjective experience and were difficult to quantify, a digital discrimination system covering 13 origins in Yunnan and three stalk positions (upper, middle, and lower) for tobacco leaf, as well as style index and membership degree index models, was constructed based on a 70-dimensional chemical fingerprint and pattern recognition. Through one-way ANOVA and principal component analysis of 403 flue-cured tobacco leaf samples, 32 key chemical indicators were screened, and the performance of three algorithms, linear discriminant analysis (LDA), partial least squares discriminant analysis (PLS-DA), and support vector machine (SVM), was compared. The results showed that the SVM-RBF model achieved discrimination accuracies of 99.2% and 95.4% for the 13 origins and three stalk positions on the test set, respectively, while on the external independent validation set, the discrimination accuracies for origins and stalk positions remained at high levels of 98.5% and 93.1%, respectively. Based on the Mahalanobis distance in the optimal discriminant space and Softmax normalization, origin style index and stalk position membership degree index models were proposed, achieving a continuous and vectorized representation of tobacco leaf style characteristics. This system transformed discrete classification results into computable digital indices, providing a high-precision and operable technical tool for intelligent warehousing of tobacco leaf, precise blending substitution, and digital maintenance of brand style, thereby promoting a paradigm shift in tobacco raw material management from experience-driven to data-driven.

Key words: chemical fingerprint, pattern recognition, tobacco leaf, origin, stalk position, discrimination, style index, membership degree index, model, Yunnan