欧亚国际

来自中国农业大学草业科学与技术学院的科学家利用VideometerLab 4多光谱成像系统，在草业领域知名期刊草业学报发表了题为“基于多光谱成像技术快速无损检测紫花苜蓿人工老化种子”的文章。这是该院科学家利用该系统发表的第三篇文章。

图14组种子材料的多光谱RGB成像

草业学报, 2022, 31(7): 197-208 DOI:10.11686/cyxb2021198

研究论文基于多光谱成像技术快速无损检测紫花苜蓿人工老化种子

王雪萌，何欣，张涵,宋瑞,毛培胜,贾善刚

中国农业大学草业科学与技术学院，草业科学北京市重点实验室，北京 100193

Non-destructive identification of artificially aged alfalfa seeds using multispectral imaging analysis

WANG Xue-meng，HE Xin,, ZHANG Han, SONG Rui, MAO Pei-sheng, JIA Shan-gang,

College of Grassland Science and Technology，China Agricultural University，Key Laboratory of Pratacultural Science，Beijing Municipality，Beijing 100193，China

摘要

种子老化是影响种子生产和储藏的重要因素，给农业生产带来了严重的经济损失，已成为影响种子活力中最具威胁性的因素之一。老化种子的检测以及种子老化后发芽情况的鉴定对种子生产具有重要意义，但目前常用的检测手段都是一次性、破坏性的。因此，一种快速、无损的种子老化和发芽检测方法不仅是研究的需要，也是种子行业进行种子检测分选所急需的。利用多光谱成像技术，采集紫花苜蓿种子的形态和光谱特征数据，利用LDA（线性判别分析）、SVM（支持向量机）和nCDA（归一化标准判别分析）3种多元分析方法，对不同老化程度苜蓿种子及其发芽情况分别进行分类和预测。结果表明，不同老化程度种子平均光谱反射率在470~660 nm处出现了明显的区别。LDA可以区分老化种子和未老化种子（准确度93.0%~97.7%），也可以区分不同老化程度的种子（准确度75.3%~91.7%），且均高于SVM的分类结果（准确度分别为92.4%~94.9%和68.7%~78.8%）；nCDA对老化种子进行区分的准确度高达88%~98%。同时，LDA可以准确预测发芽种子和不发芽种子，准确度可达98.7%，高于SVM的92.1%；nCDA预测老化种子发芽准确度达到了90%~99%。本研究证明了多光谱成像与分析技术不仅可以区分老化种子，也可以预测种子的发芽。上述结果证实多光谱成像技术结合多元分析为高效无损检测苜蓿种子活力提供了新途径，具有良好的应用前景。

关键词：老化种子;多光谱成像;多元分析;紫花苜蓿;无损检测

Abstract

Aging during the storage of seeds reduces seed vitality and causes serious economic losses to the agricultural industry， and has become one of the biggest factors involved in decreased seed vigor. Distinction between aged and viable seeds is of high importance in alfalfa seed planting and production， but the existing methods are time-consuming or destructive. Therefore， a rapid and non-destructive screening method to distinguish aged and viable seeds is not only very necessary in seed testing and the alfalfa seed industry， but also potentially useful in alfalfa seed research. In this study， we collected data of both morphological features and spectral traits of alfalfa seeds using multispectral imaging （MSI） technology. Then， we evalsuated three multivariate analysis methods： linear discriminant analysis （LDA）， support vector machines （SVM） and normalized canonical discriminant analysis （nCDA）， to classify seeds artificially aged for 0， 3， 6 and 14 days， and predict viable seeds which could germinate. It was found that the mean light reflectance at 470-660 nm differed significantly between non-aged and aged seeds. The LDA model based on a “hold-out method” provided accuracies of 93.0%-97.7% in distinguishing aged seeds from nonaged seeds， and 75.3%-91.7% in distinguishing the different groups of aged seeds. Corresponding values for the SVM model were a little lower， being 92.4%-94.9% and 68.7%-78.8%， respectively. The nCDA model also exhibited achieved aged seed discrimination with an accuracy of 88.0%-98.0%. Finally， viable seeds could be distinguished from dead seeds in all the categories of aged seeds， with accuracies of 98.7% and 92.1% in LDA and SVM analysis， respectively， while the accuracy of nCDA in predicting the germination of aged seeds ranged from 90% to 99%. This study showed that MSI could successfully distinguish aged seeds， and also predict germination of seeds. In summary， we demonstrated a nondestructive， rapid and high-throughput approach to screen both aged and viable seeds in alfalfa， and showed that MSI together with multivariate analysis is promising as a new tool for application in seed testing and field planting of alfalfa seeds.

Keywords：aged seeds;multispectral imaging;multivariate analysis;alfalfa;non-destructive identification