多光谱表型成像结合机器学习方法快速无损检测玉米中玉米赤霉烯酮
VideometerLab 4多光谱种子表型成像系统是丹麦理工大学与丹麦Videometer公司开发，是用于种子研究先进的多光谱表型成像设备，典型客户为ISTA国际种子检验协会、ESTA欧洲种子检验协会、John Innes Centre、LGC化学家集团、奥胡斯大学等等，利用该系统发表的文章已经超过300篇。
Videometer种子表型表型成像系统可测量种子如尺寸、颜色、形状等，间接测定种子参数如种子纯度、发芽百分比、发芽率、种子健康度、种子成熟度、中寿命等。种子活力综合种子活力是种子发芽和出苗率、幼苗生长的潜势、植株抗逆能力和生产潜力的总和（发芽和出苗期间的活性水平与行为），是种子品质的重要指标，具体包括吸涨后旺盛的代谢强度、出苗能力、抗逆性、发芽速度及同步性、幼苗发育与产量潜力。种子活力是植物的重要表型特征，传统检测方法包括低温测试、高温加速衰老测试、幼苗生长测定等。
该系统也可以对细菌、虫卵、真菌等进行高通量成像测量，进行病理学、毒理学或其它研究。对于拟南芥等冠层平展的植物，可以进行自动的叶片计数等。
摘要
玉米不可避免地受到玉米赤霉烯酮（zearalenone，ZEN）的污染，对人类造成严重危害。本研究采用多光谱成像（MSI）技术结合不同的机器学习方法检测玉米中的ZEN含量。利用遗传算法和反向传播神经网络（GA-BPNN）可以选择与玉米中ZEN浓度最相关的波长。结果表明，GA-BPNN方法可以检测出ZEN污染水平，准确率为93.33%。此外，GA-BPNN算法是定量预测ZEN污染含量的最佳方法，预测集的相关系数（R p）、均方根误差（RMSEP）、残差预测偏差（RPD）和偏差分别达到0.95、3.66μg/kg、5.39和1.55μg/kg。可以得出结论，多光谱成像与机器学习相结合适用于玉米禅宗含量的快速测量。
关键词：多光谱成像、无损检测、玉米赤霉烯酮、机器学习方法
Application of multispectral imaging combined with machine learning methods for rapid and non-destructive detection ofzearalenone (ZEN) in maize
Abstract
Maize is inevitably contaminated by zearalenone (ZEN) that will cause serious harm to human beings. In this study, multispectral imaging (MSI) technology combined with different machine learning were used to detect ZEN content in maize. The wavelengths that were most related to ZEN concentration in maize could be selected by genetic algorithm with a back-propagation neural network (GA-BPNN). Our results showed that ZEN contamination level could be detected with the accuracy of 93.33% by GA-BPNN method. In addition, for quantitative prediction of ZEN contamination content GA-BPNN algorithm was the best method with the correlation coefficient (R p ), the root means square error (RMSEP), residual predictive deviation (RPD) and bias achieved to 0.95, 3.66 μg/kg, 5.39 and 1.55 μg/kg, respectively in prediction set. It can be concluded that multispectral imaging combined with machine learning was applicable for rapid measurement of ZEN content in maize.
Key words: Multispectral imaging, Non-destructive detection, Zearalenone in maize,Machine learning method
