高光谱图像结合深度学习的油菜菌核病早期识别  

Early Recognition of Sclerotinia Stem Rot on Oilseed Rape byHyperspectral Imaging Combined With Deep Learning

作　　者：梁万杰[1] 冯辉[2] 江东[3] 张文宇[1,4] 曹静 曹宏鑫[1] LIANG Wan-jie;FENG Hui;JIANG Dong;ZHANG Wen-yu;CAO Jing;CAO Hong-xin(Institute of Agricultural Information,Jiangsu Academy of Agricultural Sciences,Nanjing 210014,China;Institute of Plant Protection,Jiangsu Academy of Agricultural Sciences,Nanjing 210014,China;Institute of Geographical Sciences and Natural Resources Research,Chinese Academy of Sciences,Beijing 100101,China;School of Agricultural Engineering,Jiangsu University,Zhenjiang 212013,China)

机构地区：[1]江苏省农业科学院农业信息研究所,江苏南京210014 [2]江苏省农业科学院植物保护研究所,江苏南京210014 [3]中国科学院地理科学与资源研究所,北京100101 [4]江苏大学农业工程学院,江苏镇江212013

出　　处：《光谱学与光谱分析》2023年第7期2220-2225,共6页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金面上项目(31871522);江苏省农业科技自主创新资金项目(CX(20)3073)资助。

摘　　要：油菜菌核病为土传病害,发病早期叶片无可见症状,从植株表面很难发现。用叶片的普通光谱图像或RGB图像无法对其进行识别。采用高光谱图像作为监测技术,结合深度学习模型构建油菜菌核病发病早期识别模型,并取得了较好的识别效果。以油菜菌核病为研究对象,采用菌丝块接种法,在油菜根部诱发病害。分别于发病后第2、5、7、9天采集发病油菜植株和健康植株光谱图像。对高光谱图像去除背景、S-G光谱曲线平滑处理、剪切、分割等处理后构建模型训练测试数据集。以Resnet50深度学习模型为基础,通过增加特征图数量,减小第1层卷积核大小来提高模型对油菜菌核病发病早期的识别能力。通过交叉验证、模型结构改进前后识别能力对比、模型泛化能力测试等,验证了改进模型的识别能力和泛化能力。Resnet50模型结构改进前后,对油菜菌核病发病早期的识别正确率分别是66.79%、83.78%和88.66%,改进后模型的识别正确率分别提高了16.99%和4.88%,模型的识别精度和召回率也得到很大提高。所提出的识别模型平均识别正确率为88.66%,精度和召回率达到83%以上,只有对发病第7天的召回率为79.04%。把构建的多分类模型设定为是否受病害胁迫的二分类模型,则模型的正确率97.97%,精度99.19%,召回率98.02%,同时,模型对第9天测试集的识别正确率达到91.25%。改进后的Resnet50模型可有效保留数据的光谱特征和局部特征,使模型对油菜菌核病发病早期的识别能力显著提高。该模型对发病1周内的油菜菌核病严重程度具有较好的识别能力。对是否发病的识别能力更高,模型识别正确率、精度和召回率均达到97.97%以上。模型对油菜菌核病发病早期识别具有很好识别能力和泛化能力。因此,该模型可综合利用高光谱图像的光谱和图像特征,解决油菜菌核病发病早期无症状、识别困难的问题;�The sclerotinia stem rot on oilseed rapeis soil-borne disease.There are no visible symptoms in the leaves in the early onset stage,so it is not easy to monitor from the plant surface.It cannot be recognized by ordinary spectral images or RGB images of oilseed rape leaves.In this study,hyperspectral imaging is used as monitoring technology,combined with deep learning to build an early identification model of sclerotinia stem rot on oilseed rape to solve the problem of early identification of sclerotinia stem rot on oilseed rape.In this study,the stem rot on oilseed rape was used as the research object,and the mycelium inoculation method was used to induce the disease in the root of oilseed rape.The hyperspectral images of diseased rape plants and healthy plants were collected on the 2nd,5th,7th and 9th day after onset.After removing the background,S-G smoothing of the spectral curve,cutting and segmentation,the model training and testing dataset was constructed.Based on the resnet50,the number of feature images was improved,and the first layer’s convolution kernelsize was reduced to improve the model’s recognition ability.The model’s recognition performance and generalization ability were verified based on cross validation.The accuracy of the three models with different structures was 66.79%,83.78%and 88.66%respectively.The accuracy of the improved model was increased by 16.99%and 4.88%respectively,and the precision and recall rate were improved too.The average accuracy of the improved resnet50 model was 88.66%,the precision and recall rate was more than 83%,and only the recall rate on the seventh day of onset was 79.04%.If the model is binary whether the rape is under disease stress,the accuracy of the model is 97.97%,the precision is 99.19%,and the recall rate is 98.02%.At the same time,the accuracy of the model for the test dataset reached 91.25%.The results of cross-validation showed that the improved model had a good recognition ability for sclerotinia stem rot on oilseed rape within one week and could

关 键 词：深度卷积神经网络 高光谱图像 油菜菌核病 早期诊断 Resnet50 

分 类 号：S127[农业科学—农业基础科学]