关键词: 烤烟, 近红外光谱, 香型, 分类模型, 随机森林

Abstract: In order to quickly and non⁃destructively discriminate the aroma type of domestic flue⁃cured tobacco and guide the utilization of raw materials for cigarette formulation，near infrared spectra were collected from 1 383 tobacco samples，the aroma was classified based on the results of eight aroma zones，and the samples from each stable aroma zone were selected to build the aroma classification model.The results showed that the accuracy of the fragrance classification model constructed by using random forest based on the NIR raw spectral data was only 48.64%；the model accuracy was improved by 29.54 percentage points after the spectral data were pre⁃processed by SG filtering first or second derivatives and multiple scattering correction；then the model accuracy was improved to 85.91% by factor analysis and dimensionality reduction（45 factors）；finally，the model key parameters were optimized，the highest model accuracy of 90.45% was achieved when the number of base evaluators was 500 and the random seed was 9.This classification model was used to predict the prediction set，the recall rates of fresh⁃sweetness type，burnt⁃sweetness⁃burnt type，fresh⁃honey⁃sweetness type and woody⁃honey⁃sweetness type all reached more than 90.00%.The lowest recall rate was for honey⁃sweetness type and burnt⁃honey⁃sweetness type（66.67%）.The results show that the use of NIR spectroscopy can effectively identify the eight aroma types of flue⁃cured tobacco，which provides new ideas and technical support for the rapid identification of flue⁃cured tobacco aroma types and digital evaluation of tobacco styles.

Key words: Flue?cured tobacco, Near infrared spectroscopy, Aroma type, Classification model, Random forest