基于多源信息的烤烟香型分类

摘要

为我国烤烟典型香型的分类提供参考,以2014年全国浓香型、中间香型、清香型烟叶代表性产区18个州(市)45个县(区)的520个C3F初烤烟叶样品为研究对象,通过对常规化学成分、外观质量、物理特性进行测定,基于多源信息应用随机森林算法进行分类模型构建.结果表明:利用随机森林算法基于多源信息构建的烤烟香型分类模型较好,分类正确率为91.54％,AUC(ROC曲线下的面积)为0.972 6.其中,对浓香型烤烟分类效果最好,没有误判,中间香型误判率为7.57％,清香型烤烟误判率为18.87％;影响烤烟香型分类的重要变量为氯离子、总糖、宽度、厚度和钾离子,影响浓香型和中间香型烤烟分类最重要的变量为氯离子,影响清香型烤烟分类最重要的变量为总糖.基于烤烟多源信息利用随机森林算法进行香型分类,在浓香型和中间香型分类效果较好,清香型分类效果稍差.%The conventional chemical components,appearance quality and physical property of 520 C3F samples from 45 counties of 18 prefectures (cities) with the representative producing areas of strong,intermediate and clear aroma flue-cured tobacco were determined and then the classification model was established by the random forest algorithm based on multiple source information to provide a reference for classification of flue-cured tobacco aroma type in China.Results:The classification accuracy rate and AUC of the established classification model of flue-cured tobacco aroma type reach 91.54％ and 0.972 6 respectively.The application effect of the established classification model is the best in strong aroma fluecured tobacco.The misjudgment rate of the established classification model in intermediate and clear aroma flue-cured tobacco is 7.57％ and 18.87％ respectively.The important variables to influence classification of flue-cured tobacco aroma type are chlorine ion,total sugar,leaf width,leaf thickness and potassium ion.The important variable to influence classification of flue-cured tobacco with strong and intermediate aroma type is chlorine ion.The important variable to influence classification of flue-cured tobacco with clear aroma type is total sugar.In conclusion,the classification effect of the established classification model of flue-cured tobacco aroma type based on multiple source information is good in strong and intermediate aroma flue-cured tobacco and poor in clear aroma flue-cured tobacco.