基于烟株冠层高光谱的烟草病毒病分类模型构建  

Construction of a tobacco virus disease classification model based on canopy hyperspectra of tobacco plants

作　　者：姜雪妍 辛星龙 薛文鑫 李文杰 汪季涛[3] 江厚龙 王树林 王新伟[1] JIANG Xueyan;XIN Xinglong;XUE Wenxin;LI Wenjie;WANG Jitao;JIANG Houlong;WANG Shulin;WANG Xinwei(Key Laboratory of Monitoring,Early Warning and Integrated Control of Pests and Diseases in the Tobacco Industry/Tobacco Research Institute,Chinese Academy of Agricultural Sciences,Qingdao,Shandong 266101,China;Qingdao Agricultural University,Qingdao,Shandong 266109,China;Anhui Branch of China Tobacco Corporation,Hefei 230081,China;Chongqing Branch of China Tobacco Corporation,Chongqing 404100,China;Sichuan Branch of China Tobacco Corporation,Chengdu 610000,China)

机构地区：[1]中国农业科学院烟草研究所/烟草行业病虫害监测预警与综合防治重点实验室,山东青岛266101 [2]青岛农业大学,山东青岛266109 [3]安徽中烟工业有限责任公司,安徽合肥230081 [4]中国烟草总公司重庆市公司烟叶分公司,重庆404100 [5]四川中烟工业有限责任公司,四川成都610000

出　　处：《生物灾害科学》2024年第4期561-573,共13页Biological Disaster Science

基　　金：中国烟草总公司科技项目(110202102027、110202201051(SJ-1))。

摘　　要：【目的】烟草病毒病影响烟草生长,甚至导致烟叶减产、烟株死亡,严重影响烟叶的品质。为实现快速高效烟草病毒病害监测。【方法】采用芬兰SPECIM IQ高光谱相机,通过无损检测烟草病毒病感染程度的技术,建立了高光谱相机在烟田中采集健康和不同级别病毒病感染程度的烟草植株图像,预处理后提取波长范围为397~1004 nm的204个波段光谱特征,使用四种波长选择方法与四种机器学习算法构建了16个田间烟草病毒病等级分类模型。【结果】模型检测结果表明一半以上模型精度超过0.70。其中,鲸鱼优化算法结合随机森林算法的模型识别准确率达到0.83,表现最佳。波段选择与变量贡献分析表明,近红外光谱区在区分健康与感染病毒病烟草叶片中具有重要参考价值。【结论】研究证明了基于高光谱成像技术结合机器学习和波长选择算法,能够有效监测烟草病毒病,为大范围田间烟草病毒病害监测提供了基础与理论支持。[Objective]Tobacco virus diseases affect the growth of tobacco,and even lead to the reduction of tobacco yield and death of tobacco plants,which seriously affects the quality of tobacco.This study was conducted to achieve rapid and efficient tobacco virus disease monitoring.[Method]Using the Finnish SPECIM IQ hyperspectral camera,through the non-destructive detection of the infection degrees of tobacco virus diseases,a hyperspectral camera was used in the tobacco field to collect images of healthy and different levels of virus disease infection degrees of the tobacco plants,pre-processing to extract the wavelength range of 397-1004 nm of 204 bands of spectral features,and using four wavelength selection methods with four machine learning algorithm,and so sixteen field tobacco virus disease class classification models were constructed.[Result]Model testing results showed as follows:more than half of the models had an accuracy of more than 0.70.Among them,the model identification accuracy of the whale optimization algorithm combined with the random forest algorithm reached 0.83,which was the best performance.The analysis of band selection and variable contribution showed that the near-infrared spectral region had important reference values in distinguishing between healthy and infected tobacco leaves with viral diseases.[Conclusion]This study has proved that based on hyperspectral imaging technology combined with machine learning and wavelength selection algorithms,tobacco virus diseases can be effectively monitored,which provides the foundation and theoretical support for the monitoring of tobacco virus diseases on a wide range of fields.

关 键 词：烟草 病毒病 高光谱成像 特征波段 机器学习 

分 类 号：S435.72[农业科学—农业昆虫与害虫防治]