基于PLS-DA和LS-SVM的可见/短波近红外光谱鉴定港种四九、十月红和九月鲜菜心种子的可行性研究  

Feasibility Study on Identification of Seeds of Hong Kong Seeds 49,October Red and September Fresh Cabbage Based on Visible/Shortwave Near-Infrared Spectroscopy of Partial Least Squares Discriminant(PLS-DA)and Least Squares Support Vector Machine(LS-SVM)

作　　者：章海亮 聂训 廖少敏 詹白勺 罗微 刘书玲 刘雪梅[2] 谢潮勇 ZHANG Hai-liang;NIE Xun;LIAO Shao-min;ZHAN Bai-shao;LUO Wei;LIU Shu-ling;LIU Xue-mei;XIE Chao-yong(School of Electrical and Automation Engineering,East China Jiaotong University,Nanchang 330013,China;School of Civil Engineering and Architecture,East China Jiaotong University,Nanchang 330013,China;Jiangxi Institute of Science and Technology Information,Nanchang 330046,China)

机构地区：[1]华东交通大学电气与自动化工程学院,江西南昌330013 [2]华东交通大学土木建筑学院,江西南昌330013 [3]江西省科学技术信息研究所,江西南昌330046

出　　处：《光谱学与光谱分析》2024年第6期1718-1723,共6页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金项目(61565005,41867020);江西省03专项及5G项目(20212ABC03A17)资助。

摘　　要：目前市面上菜心的品种复杂,不同菜心种子的品质与发芽率不同,但菜心种子单从外观上差别不大,因此区分菜心种子的类别成为了一大难题。为了实现菜心种子类别的快速区分,探究了基于可见/短波近红外光谱分析菜心种子类别的可行性。从南昌市种子交易场所购买了港种四九、十月红和九月鲜三个品种的菜心种子,从中挑选出品相较好且大小适中的子粒,将每种菜心种子均匀分为30份,按照2∶1划分为建模集和预测集,所有样本共计90份。通过近红外光谱仪获取采样间隔为1 nm的菜心种子的光谱反射率,波长覆盖范围325~1075 nm,将原始光谱数据采用多元散射校正(MSC)、卷积平滑(S-G)和标准正态变换(SNV)三种预处理方法进行预处理,预处理后的光谱变量建立偏最小二乘回归(PLSR)模型,确定了SNV是最佳预处理方法。采用主成分分析(PCA)对菜心种子进行了聚类分析,从前三个主成分因子(PCs)得分图可知三种菜心种子存在光谱特征差异。将原始光谱变量、前三个PCs(累计贡献97.15%)和基于随机蛙跳(RF)算法挑选的13个特征波长作为偏最小二乘判别(PLS-DA)和最小二乘支持向量机(LS-SVM)模型的输入变量,从模型结果可知:三种输入变量中,采用RF筛选特征波长作为模型输入变量时,模型预测效果最好,PCs建立的模型最差,相比于PCA分析,采用RF筛选出的特征波长更能够反映原始光谱信息。比较不同模型预测效果,LS-SVM模型比PLS-DA模型得到的预测精度更好,其中RF-LS-SVM模型是所有模型中最佳的预测模型,建模集和预测集均为100%。采用可见/短波近红外光谱研究菜心种子的类别可行,并且能够获得很好地预测效果,为菜心种子的快速区分提供了理论依据。At present,the varieties of cabbage on the market are complex;the quality and germination rate of different cabbage seeds are different,but the appearance of cabbage seeds is not very different,so it has become a big problem to distinguish the types of cabbage seeds.This paper explores the feasibility of analyzing cabbage seed categories based on visible/short-wave near-infrared spectroscopy to achieve rapid differentiation of cabbage seed categories.The experiment purchased three varieties of cabbage seeds of Hong Kong species Sijiu,October red and September fresh from the Nanchang Seed Trading Place.The seeds with good appearance and moderate size were selected,and each kind of cabbage seed was evenly divided into three categories.30 copies,divided into modeling and prediction sets according to 2∶1,totalling 90 copies of all samples.The near-infrared spectrometer was used to obtain the spectral reflectance of cabbage seeds with a sampling interval of 1 nm,and the wavelength coverage was 325~1075 nm.The original spectral data were corrected by multivariate scattering(MSC),convolution smoothing(S-G)and standard normal transformation(SNV).)Three preprocessing methods were used for preprocessing.A partial least squares regression(PLSR)model was established for the spectral variables after preprocessing,and SNV was determined to be the best preprocessing method.In addition,principal component analysis(PCA)was used to conduct cluster analysis on cabbage seeds.The scores of the first three principal component factors(PCs)show that the three kinds of cabbage seeds have differences in spectral characteristics.Finally,the original spectral variables,the first three PCs(with a cumulative contribution of 97.15%)and 13 characteristic wavelengths selected based on the random frog(RF)algorithm were used as partial least squares discriminant(PLS-DA)and least squares support vector machines(The input variables of the LS-SVM)model,from the model results,we can see that among the three input variables when the RF screening char

关 键 词：菜心种子 主成分分析 随机青蛙 偏最小二乘判别 最小二乘支持向量机 

分 类 号：S339.31[农业科学—作物遗传育种]