烟叶霉变的快速识别——基于近红外光谱与随机森林算法

为建立烟叶霉变快速识别模型，以复烤片烟为研究对象，在高温高湿条件下进行霉变实验，获得不同霉变程度的烟叶样本。应用近红外光谱技术在4000~12000 cm-1范围内对烟叶的近红外光谱进行采集，获得烟叶样本的基础光谱数据。采用小波分解法对基础光谱数据进行解析，选择中间频率小波系数[cd4, cd5]为光谱变量，利用随机森林算法建立了不同霉变烟叶的识别模型。模型对训练集预测准确率达到93.82%，独立测试集判别准确率达到94.84%，对未霉变样品、临近霉变样品和霉变样品的判别均取得了令人满意的结果。 

Rapid identification of tobacco mildew based on near infrared spectroscopy and random forest algorithm

In order to establish a rapid identification model of tobacco mildew,the tobacco mildew experiment was carried out under high temperature and high humidity to obtain tobacco leaf samples with different degrees of mildew.Near-infrared spectroscopy (NIRS) was applied for direct and rapid collection of characteristic spectra from flue-cured tobacco strips.In the range of 4000~12000 cm^-1,the NIR spectra were collected to obtain the basic spectra data of tobacco samples,and their NIR spectra were pretreated by the wavelet transform (WT) method.A training set of tobacco objects was modeled based on wavelet coefficients by using the random forest (RF).The classification accuracy for training dataset was 93.82%,while that for independent dataset was 94.84%.As for the prediction model,10-fold cross-validation was used to optimize parameters,and different wavelet coefficients were also compared and optimized to improve prediction ability of the model.Satisfactory identification of tobacco mildew degree can be achieved,which showed that the mildew degree of tobacco samples can be distinguished accurately based on NIRS combined with suitable chemometrics approaches.