关键词: 烟叶, 化学指标, 感官质量, 分类模型, 江西省

Abstract: In order to explore the quality characteristics of tobacco leaves in Jiangxi Province, statistical methods such as analysis of variance and interval estimation were used to analyze the differences and coordination of 11 chemical indicators in the upper, middle, and lower parts of tobacco leaves with different sensory quality grades （Class A, Class B, Class C） in Jiangxi Province. On this basis, this study constructed support vector machine （SVM） models and random forest （RF） models for predicting the sensory quality classification of tobacco leaves. The results showed that Class C tobacco leaves in the upper, middle, and lower parts had 7, 6, and 6 chemical indicators higher than Class A and Class B, respectively. Class B tobacco leaves in the upper, middle, and lower parts had 3, 3, and 2 chemical indicators higher than Class A and Class C, respectively, while Class A tobacco leaves in the upper, middle, and lower parts had 0, 2, and 3 chemical indicators higher than Class B and Class C, respectively; from the perspective of interval length, class C tobacco leaves in the upper, middle, and lower parts had 9, 9, and 7 chemical indicators higher than Class A and Class B, respectively. Class B tobacco leaves in the upper, middle, and lower parts had 1, 0, and 3 chemical indicators higher than Class A and Class C, respectively, while Class A tobacco leaves in the upper, middle, and lower parts had 1, 2, and 1 chemical indicators higher than Class B and Class C, respectively. The coordination of chemical components in Class C tobacco leaves was much worse than that in Class A and Class B, which might be an important reason for the deterioration of sensory quality. The weighted average of accuracy, recall, and F1 score for both SVM and RF models exceeded 84%, and the SVM model had slightly higher three indicators than the RF model. There were significant differences in the chemical index characteristics of Class C tobacco leaves compared to Class A and Class B, while the differences between Class A and Class B were relatively small; the SVM model had better classification performance for Class A and B tobacco samples than the RF model, while the RF model had better recognition performance for Class C than the SVM model.

Key words: tobacco leaves, chemical indicators, sensory quality, classification model, Jiangxi Province