基于高光谱成像技术的鲜烟叶叶位识别方法  

Identification Method of Green Tobacco Leaf Positions Based on Hyperspectral Imaging Technology

作　　者：李粉粉 王爱霞 赵晨 白涛[3] 毛岚[3] 张豹林 李生栋 宋朝鹏[1] 王涛[1,3] LI Fenfen;WANG Aixia;ZHAO Chen;BAI Tao;MAO Lan;ZHANG Baoin;LI Shengdong;SONG Zhaopeng;WANG Tao(College of Tobacco,Henan Agricultural University,Zhengzhou 450002,China;China Tobacco Henan Industry Co.,Ltd.,Zhengzhou 450016,China;Qujing Tobacco Company of Yunnan Province,Qujing 655000,China)

机构地区：[1]河南农业大学烟草学院,河南郑州450002 [2]河南中烟工业有限责任公司,河南郑州450016 [3]云南省烟草公司曲靖市公司,云南曲靖655000

出　　处：《河南农业科学》2024年第2期144-151,共8页Journal of Henan Agricultural Sciences

基　　金：中国烟草总公司科技重点研发项目(110202102007);中国烟草总公司云南省公司资助项目(2021530000241036)。

摘　　要：为实现鲜烟叶叶位的快速无损识别,以不同着生部位烟叶为研究对象,应用高光谱成像技术,构建基于特征光谱的鲜烟叶叶位判别模型。首先,利用标准正态变换(SNV)、二阶导数(2ND)、SavitzkyGolay卷积平滑(SG)和多元散射校正(MSC)4种光谱预处理方法对烟叶原始高光谱数据进行处理,然后采用预处理后的全波段光谱数据和特征波段光谱数据,构建基于支持向量机(SVM)、偏最小二乘判别分析(PLS-DA)和反向传播神经网络(BPNN)的鲜烟叶叶位识别模型。结果表明:采用SG滤波预处理和BPNN所构建的模型识别效果最好,训练集和预测集的预测准确率分别为91.15%和90.63%。此外,利用竞争性自适应重加权算法(CARS)所筛选的特征波长所建立的BPNN模型最优,训练集和预测集的预测准确率达到了93.23%和92.19%。表明利用高光谱成像技术判别鲜烟叶所属部位是可行的,可以实现鲜烟叶所属部位快速、无损检测。s:In order to realize the rapid non‐destructive identification of fresh tobacco leaf position,this study took tobacco leaves with different leaf positions as the research object,applied hyperspectral imaging technology to construct a fresh tobacco leaf position discrimination model based on characteristic spectrum.We processed the original hyperspectral data of tobacco leaves by using SNV(standard normal variate),2ND(2nd derivative),SG(Savitzky‐Golay smoothing filter)and MSC(multiplicative scatter correction)four spectral preprocessing methods,and used the preprocessed full‐band spectral data and characteristic band spectral data to construct fresh tobacco leaf position recognition models based on SVM(support vector machines),PLS‐DA(partial least squares‐discriminant analysis,PLS‐DA)and BPNN(back propagation neural network).The results showed that the model constructed by SG filter preprocessing and BPNN had the best recognition effect,and the discrimination results of the training set and prediction set were 91.15%and 90.63%,respectively.In addition,the BPNN model established by using the characteristic wavelengths screened by CARS was the best,and the prediction accuracy of the training set and prediction set reached 93.23%and 92.19%.This study shows that it is feasible to use hyperspectral imaging technology to identify the parts of fresh tobacco leaves,which can realize rapid and nondestructive detection of the parts of fresh tobacco leaves.

关 键 词：烟叶 叶位 无损识别 高光谱成像 光谱特征 

分 类 号：S126[农业科学—农业基础科学] S572