基于高光谱成像技术的青花椒产地识别研究

Research on origin identification of Zanthoxylum schinifolium based on hyperspectral imaging technology

中文关键词:  高光谱成像技术  青花椒  产地识别  机器学习  兴趣区域

英文关键词:hyperspectral imaging technology  Zanthoxylum schinifolium  origin identification  machine learning  region of interest

中文摘要:

      目的 基于高光谱成像技术结合机器学习建立一种青花椒产地的快速识别方法, 实现四川、贵州、云南、重庆等10个青花椒主要产地样品的快速无损鉴别。方法 本研究利用全平皿法、五点平均法和中心点法3种不同的兴趣区域(region of interest, ROI)提取方式, 获得平行光谱数据, 分别采用5种预处理方法消除数据噪声提升模型性能, 并比较了偏最小二乘判别分析(partial least square-discriminant analysis, PLS-DA)、随机森林(random forests, RF)和支持向量机(support vector machine, SVM) 3种模型的产地识别效果。结果 采用全平皿法提取兴趣区域, 通过二阶导(second derivative, D2)预处理后建立的RF模型分类效果最佳, 训练集和测试集的准确率均可达到100%。进一步采用连续投影算法(successive projections algorithm, SPA)选择27个特征波长建模, 结果表明多元散射校正(multiplicative scatter correction, MSC)-RF模型判别效果最优, 训练集准确率为98.8%, 测试集准确率达到98.3%。结论 本研究建立的方法可实现不同青花椒主要产地样品的快速无损鉴别, 为高光谱成像技术在食品和药品领域的推广应用及专属小型化仪器装备系统的开发提供了理论依据。

英文摘要:

      Objective To establish a rapid identification method of the origin of Zanthoxylum schinifolium based on hyperspectral imaging technology co mbined with machine learning, and realize rapid and non-destructive identification of 10 major origins of Sichuan, Guizhou, Yunnan and Chongqing. Methods In this study, 3 kinds of different region of interest (ROI) extraction methods, including full plate, five-point average and center point, were used to obtain parallel spectral data, and 5 kinds of pretreatment methods were respectively used to eliminate data noise and improve model performance. The effects of origin identification of 3 kinds of models: Partial least square-discriminant analysis (PLS-DA), random forests (RF) and support vector machine (SVM) models were compared. Results The full plate was used to extract regions of interest, and the RF model established after second derivative (D2) preprocessing had the best classification effect, the accuracy of both training set and test set could reach 100%. The successive projections algorithm (SPA) was further used to select 27 characteristic wavelengths for modeling. The results showed that the multiplicative scatter correction (MSC)-RF model had the best discriminating effect, the accuracy of training set was 98.8%, and the accuracy of test set was 98.3%. Conclusion The method established in this study can demonstrate the rapid and non-destructive identification of samples from different major producing areas of Zanthoxylum schinifolium, which provides a theoretical basis for the popularization and application of hyperspectral imaging technology in food and medicine fields and the development of specialized miniaturized instrument and equipment system.