基于级联分类的复烤片烟产地预测方法研究  

Study on prediction method of the geographical origin of flue-cured tobacco strips based on cascade classification

作　　者：帖金鑫[1] 何文苗[1] 李石头[1] 郝贤伟[1] 李永生[1] 张立立[1] 钟永健[1] 毕一鸣[1] TIE Jinxin;HE Wenmiao;LI Shitou;HAO Xianwei;LI Yongsheng;ZHANG Lili;ZHONG Yongjian;BI Yiming(Technology center,China Tobacco Zhejiang Industrial Co.,Ltd.,Hangzhou 310008,China)

机构地区：[1]浙江中烟工业有限责任公司技术中心,浙江杭州310008

出　　处：《轻工学报》2022年第4期94-99,共6页Journal of Light Industry

基　　金：中国烟草总公司项目(110202103003);浙江中烟有限责任公司科技项目(ZJZY2021A001)。

摘　　要：为提高在有限样本条件下近红外预测复烤片烟产地的准确率,提出了一种基于级联分类的复烤片烟产地预测方法,该方法首先通过近红外光谱判断样本的香型属性并构建香型模型作为中间层,再使用LDA或PLS方法在单一香型框架下构建产地模型进行产地预测。以全国主要烟叶产区的复烤片烟为对象进行验证,结果表明:通过引入香型模型作为中间层,基于LDA的分类模型产地预测准确率由83.33%提升至94.44%;基于PLS的分类模型的准确率则由72.22%提升至86.11%。在有限样本数据和不引入新的模型参数的条件下,该方法有效降低了复烤片烟的产地误判比例。In order to improve the accuracy of near-infrared prediction under limited sample conditions,a prediction method of the geographical origin of flue-cured tobacco strips based on cascade classification was proposed.In this method,the flavor attributes of samples were judged by near-infrared spectroscopy,and the flavor model was constructed as the middle layer,and then the origin model was constructed by LDA or PLS method under a single flavor framework to predict the geographical origin.The redried tobacco in the main producing area in China was taken as the objects in the experiment.The results showed that by introducing the flavor model as the middle layer,the accuracy of the classification model based on LDA was improved from 83.33%to 94.44%;the accuracy of PLS based classification model increased from 72.22%to 86.11%.Under the condition of limited sample data and without introducing new model parameters,this method could effectively reduce the proportion of misjudgment of the origin of redried tobacco.

关 键 词：级联分类 香型 预测产地 准确率 

分 类 号：TS424[农业科学—烟草工业] X795[环境科学与工程—环境工程]