光谱学与光谱分析, 2020, 40 (8): 2415, 网络出版: 2020-12-02  

流形学习方法及近红外透射光谱的新疆冰糖心红富士水心鉴别

Watercore Identification of Xinjiang Fuji Apple Based on Manifold Learning Algorithm and Near Infrared Transmission Spectroscopy
作者单位
1 新疆农业大学机电工程学院, 新疆 乌鲁木齐 830052
2 江苏大学食品与生物工程学院, 江苏 镇江 212013
3 新疆农业大学数理学院, 新疆 乌鲁木齐 830052
摘要
苹果水心病在众多苹果主产区都有发生, 现阶段没有合适的方法实现快速鉴别和分类。 为了探索苹果水心鉴别新方法, 采用近红外透射光谱与化学计量学方法结合非线性流形学习数据降维技术, 逐个采集好果与疑似水心病果样本590~1 250 nm的近红外透射光谱, 将经光谱校正后的原始光谱做多元散射校正(multivariate scattering correction, MSC)、 标准正态变量变换(standard normal variate transformation, SNVT)、 二阶求导(2nd derivative)、 一阶求导(1st derivative)、 归一化(normalization)、 卷积平滑法(savitzky-golay smoothing, SG)、 均值中心化(mean centering, MC)、 移动平均平滑(moving average, MA)、 直接差分二阶求导(direct differential second derivative, DDSD)以及直接差分一阶求导(direct differential first derivative, DDFD)等10余种光谱预处理; 先对预处理后的光谱数据建立全波长模式识别模型从而找出多元散射校正是最优预处理方法, 而后再分别用多维尺度分析(multidimensional scaling, MDS)、 分布邻域嵌入(stochastic neighbor embedding, SNE)、 对称分布邻域嵌入(symmetric stochastic neighbor embedding, SymSNE)、 t分布邻域嵌入(t-distributed stochastic neighbor embedding, t-SNE)、 拉普拉斯映射(laplacian eigenmaps, LE)、 等距特征映射(isomap)、 地标等距映射(landmark isomap)、 局部线性嵌入(locally linear embedding, LLE)、 扩散映射(diffusion maps, DM)等多种流形学习方法对经多元散射校正预处理后的光谱数据做降维处理, 并结合马氏距离判别(mahalanobis distance discrimination, MD)、 二次判别分析(quadratic discriminant analysis, QDA)、 贝叶斯判别(Bayesian discrimination, BD)、 K最近邻法(K nearest neighbor, KNN)识别其水心存在与否。 结果表明, 提取前12主成分, 采用多元散射校正-地标等距映射-K最近邻法(MSC-landmark isomap-KNN)模型识别效果最优, 校正集和预测集识别率分别为97.5%和96.3%。 故, 流形学习方法结合近红外透射光谱可成功、 高效地实现苹果水心鉴别, 为进一步研发水心鉴别设备提供新的理论指导。
Abstract
Apple watercore occurs in many major apple producing areas, while there is no suitable way to sort apple type with watercore quickly. This research applies near infrared transmission spectroscopy, chemometric methods and manifold learning algorithm, selecting Xinjiang Red Fuji apple and watercore disease ones as samples, collecting near infrared transmission spectrum within 590 to 1 250 nm, spectroscopically corrected spectrum is used to do ten more speciesof spectral pretreatment. Firstly, full-wavelength pattern recognition is performed on the pre-processed spectral data to find out that multivariate scattering correction is the best pretreatment method. Then dataset preprocessed by multivariate scattering correction is used to make dimension reduction by using many other manifold learning algorithms such as Multidimensional Scaling, Stochastic Neighbor Embedding, Symmetric Stochastic Neighbor Embedding, t-Distributed Stochastic Neighbor Embedding, Laplacian Eigenmaps, Isomap, Landmark Isomap, Locally Linear Embedding, Diffusion Maps, combining Mahalanobis distance discrimination, quadratic discriminant analysis, K-nearest neighbor method to identify if watercore exist or not. Results indicate that an optimal identification model is obtained by using MSC-Landmark Isomap-KNN when principal components equal to twelve, and the identification rates for the calibration set and prediction set are 97.5% and 96.3% respectively. Hence, manifold learning algorithm and near infrared transmission spectroscopy technology can successfully realize the watercore identification of Xinjiang Red Fuji apple, which provides a theoretical basis for developing identification device in further research.

郭俊先, 马永杰, 郭志明, 黄华, 史勇, 周军. 流形学习方法及近红外透射光谱的新疆冰糖心红富士水心鉴别[J]. 光谱学与光谱分析, 2020, 40(8): 2415. GUO Jun-xian, MA Yong-jie, GUO Zhi-ming, HUANG Hua, SHI Yong, ZHOU Jun. Watercore Identification of Xinjiang Fuji Apple Based on Manifold Learning Algorithm and Near Infrared Transmission Spectroscopy[J]. Spectroscopy and Spectral Analysis, 2020, 40(8): 2415.

关于本站 Cookie 的使用提示

中国光学期刊网使用基于 cookie 的技术来更好地为您提供各项服务,点击此处了解我们的隐私策略。 如您需继续使用本网站,请您授权我们使用本地 cookie 来保存部分信息。
全站搜索
您最值得信赖的光电行业旗舰网络服务平台!