基于高光谱成像和卷积神经网络的‘库尔勒’香梨 黑斑病潜育期诊断研究  

Diagnosis of Korla pear black spot in incubation period based on hyperspectral imaging and convolutional neural network

作　　者：胡泽轩 王文秀 张凡 赵丹阳 马倩云 孙剑锋[1] HU Zexuan;WANG Wenxiu;ZHANG Fan;ZHAO Danyang;MA Qianyun;SUN Jianfeng(College of food science and technology,Hebei Agricultural University,Baoding,071000,China;Agricultural Engineering Key Laboratory,Tarim University,Alar 843300,China)

机构地区：[1]河北农业大学食品科技学院,河北保定071000 [2]塔里木大学现代农业工程重点实验室,新疆阿尔罕843300

出　　处：《河北农业大学学报》2022年第5期86-92,F0002,共8页Journal of Hebei Agricultural University

基　　金：河北省重点研发计划项目(20327111D);河北省省属学校基本科研业务费研究项目(KY202002).

摘　　要：黑斑病是‘库尔勒’香梨贮藏期的易染病害之一,在潜育期外观无明显变化,很难直接通过肉眼进行准确识别。本研究结合高光谱成像和卷积神经网络(CNN),实现了‘库尔勒’香梨黑斑病潜育期的识别。获取健康和不同病害程度香梨样品的高光谱图像,提取感兴趣区域内光谱后,利用不同预处理方法对其进行处理,分别基于常规算法(最小二乘-支持向量机、K最邻近法、随机森林)和CNN建立病害识别模型。结果表明,与常规算法建模结果相比,CNN模型的识别效果最优。当卷积层数为3,全连接层数为3,学习率为0.0005时,CNN模型的识别效果最佳,对样品的总体识别准确率为99.70%,对潜育期样品的识别准确率为99.76%,分别较常规算法提高了12和14个百分点。该结果证实CNN模型能够显著提高对‘库尔勒’香梨黑斑病潜育期识别的准确率,为‘库尔勒’香梨黑斑病的早期诊断防治提供了1种新的方法。Black spot is one of the infectious diseases of‘Korla’Fragrant Pear during storage.There is no obvious change in the appearance during the incubation period,so it is difficult to identify it accurately by the eyes.In this study,hyperspectral imaging and convolutional neural network(CNN)were combined to identify the black spots of‘Korla pear’in the incubation period.The hyperspectral images of healthy and fragrant pear samples with different disease degrees were obtained.After extracting the spectra in the region of interest,different preprocessing methods were employed to pretreat the spectra.Then the disease identification models were established based on conventional algorithms(least square-support vector machine,K nearest neighbor method,and random forest)and CNN,respectively.The results showed that CNN model achieved the best recognition effects compared with the results obtained by conventional algorithm modeling.The recognition accuracy of CNN model was the best when the convolution layer number was 3,the full connection layer number was 3,and the learning rate was 0.0005.The overall recognition accuracy was 99.70%,and the recognition accuracy for the incubation period sample was 99.76%,which was 12%and 14%higher than that using the conventional algorithms model.The results showed that CNN model can significantly improve the identification accuracy of black spot of‘Korla’fragrant pear during incubation period,and provide a new method for the early diagnosis and control of‘Korla’fragrant pear black spot.

关 键 词：高光谱成像 卷积神经网络 ‘库尔勒’香梨 黑斑病 潜育期 

分 类 号：S379.9[农业科学—农产品加工]