基于深度学习与可见-近红外光谱的患腥黑穗病小麦籽粒分类研究  

Research on Classification of Common Bunt of Wheat Kernels Based on Visible-Near Infrared Spectroscopy Combined with Deep Learning Algorithms

作　　者：宋金鹏 梁琨[2,4] 张驰 梅秀明 蒋鹏飞 袁锐[2] SONG Jin-peng;LIANG Kun;ZHANG Chi;MEI Xiu-ming;JIANG Peng-fei;YUAN Rui(College of Engineering,Nanjing Agricultural University,Nanjing 210031,China;College of Artificial Intelligence,Nanjing Agricultural University,Nanjing 210031,China;Key Laboratory of Biotoxin Analysis&Assessment for State Market Regulation,Nanjing Institute of Product Quality Inspection&Testing,Nanjing 210019,China;Jiangsu Province Engineering Lab for Modern Facility Agriculture Technology and Equipment,Nanjing Agricultural University,Nanjing 210031,China)

机构地区：[1]南京农业大学工学院,江苏南京210031 [2]南京农业大学人工智能学院,江苏南京210031 [3]南京市产品质量监督检验院国家市场监管重点实验室(生物毒素分析与评价),江苏南京210019 [4]南京农业大学江苏省智能化农业装备重点实验室,江苏南京210031

出　　处：《分析测试学报》2023年第7期784-793,共10页Journal of Instrumental Analysis

摘　　要：针对小麦腥黑穗病快速无损的检测需求,该文将可见-近红外光谱与深度学习算法结合建立了小麦腥黑穗病籽粒的分类模型。采用多元散射校正算法(MSC)和标准正态变换算法(SNV)对光谱进行预处理,消除光谱噪音的影响,分别使用竞争性自适应重加权算法(CARS)和随机蛙跳算法(RF)对预处理后的光谱进行特征波长提取。结果显示,特征提取算法可去除大量冗余信息,波段减少比率为93.7%~94.2%,有效降低了模型运行成本,并可防止模型过拟合。结果显示:MSC+CARS+VGG16模型训练集的准确率为96.39%,测试集准确率为91.67%,取得了较好的分类结果。最终建立的VGG16深度学习模型实现了健康、轻度患病和重度患病3类小麦籽粒的分类。对比传统机器学习模型,VGG16模型能够充分提取光谱特征信息,更好地区分健康与轻度患病籽粒。该研究表明深度学习结合可见-近红外光谱方法,能够实现对不同患病程度腥黑穗病小麦籽粒的有效分类,为腥黑穗病小麦籽粒的快速无损检测提供了一种新方法。In view of the need for rapid and non-destructive detection of common bunt,a visiblenear infrared spectroscopy combined with deep learning algorithms was proposed to construct a classi⁃fication model for identifying common bunt of wheat kernels in this paper.The spectral data were pre⁃processed by multiple scattering correction(MSC)and standard normal variate(SNV)algorithms to eliminate noise.Meanwhile,feature wavelength extraction was performed on the preprocessed spec⁃tra using the competitive adaptive reweighted sampling(CARS)and random frog(RF)algorithms.The results indicated that the feature extraction algorithms effectively removed redundant information,resulting in a reduction of 93.7%to 94.2%in the number of spectral bands.This significantly re⁃duced the model’s computational cost and prevented overfitting.The achieved results demonstrated the effectiveness of the MSC+CARS+VGG16 model with an accuracy of 96.39%for the training set and 91.67%for the test set,providing satisfactory classification outcomes.The final VGG16 deep learning model successfully categorized wheat kernels into three classes:healthy,light suscept,and heavy suscept.Compared to traditional machine learning models,the VGG16 model demon⁃strates a superior capability in extracting spectral features,and could more accurately distinguish healthy wheats and light suscept wheats.Results showed that visible-near infrared spectroscopy com⁃bined with deep learning could realize the effective classification on the wheat kernels with different degrees of common bunt,providing a novel approach for the rapid and non-destructive detection on common bunt wheat kernels.

关 键 词：可见-近红外光谱 深度学习 腥黑穗病 小麦 

分 类 号：O657.3[理学—分析化学] S512.1[理学—化学]