病害胁迫下玉米LAI遥感反演研究  

Analysis of Effect of Disease Stress on Maize LAI Remote Sensing Estimation

作　　者：刘帅兵 金秀良 冯海宽[4] 聂臣巍 白怡 程明瀚 LIU Shuaibing;JIN Xiuliang;FENG Haikuan;NIE Chenwei;BAI Yi;CHENG Minghan(School of Remote Sensing and Information Engineering,Wuhan University,Wuhan 430079,China;Institute of Crop Science,Chinese Academy of Agricultural Sciences,Beijing 100081,China;National Nanfan Research Institute(Sanya),Chinese Academy of Agricultural Sciences,Sanya 572024,China;Information Technology Research Center,Beijing Academy of Agriculture and Forest Sciences,Beijing 100097,China)

机构地区：[1]武汉大学遥感信息工程学院,武汉430079 [2]中国农业科学院作物科学研究所,北京100081 [3]中国农业科学院国家南繁研究院,三亚572024 [4]北京市农林科学院信息技术研究中心,北京100097

出　　处：《农业机械学报》2023年第3期246-258,共13页Transactions of the Chinese Society for Agricultural Machinery

基　　金：国家重点研发计划项目(2016YFD0300602);国家自然科学基金项目(42071426、51922072、51779161、51009101);海南省崖州湾种子实验室项目(JBGS+B21HJ0221);中国农业科学院南繁研究院南繁专项(YJTCY01、YBXM01)。

摘　　要：为了在病害发生条件下进行玉米LAI的遥感估算,针对41个不同抗性的玉米自交系品种,通过人工接种方法,获得了不同病害严重程度(1~9级)的LAI数据,同时采集了地面高光谱和无人机多光谱数据,构建了K近邻算法、支持向量机、梯度提升分类树和决策分类树分类模型对病害进行分类,对玉米种质资源抗病性进行了划分。基于不同玉米病害胁迫程度分类结果,采用随机森林回归、梯度提升回归树、极端梯度增强算法、轻量梯度提升机4种机器学习模型对玉米LAI进行反演,讨论了不同模型在病害胁迫下的鲁棒性。研究结果表明,对不同生育期玉米病害程度进行划分,基于地面高光谱识别精度分别为84.72%(梯度提升分类树)、47.67%(支持向量机)、55.05%(K近邻算法)、83.02%(决策分类树)。基于病害分类结果,本文利用无人机多光谱数据估算了不同病情等级胁迫下的玉米LAI。构建了4种集成学习模型对不同病情等级的LAI进行估算,4个LAI反演模型的总体反演精度(rRMSE)分别为:19.11%(梯度提升回归树)、15.94%(轻量梯度提升机)、14.51%(随机森林回归)和15.45%(极端梯度增强算法)。其中极端梯度增强算法对病害胁迫的普适性最好,不同病害等级下的反演精度rRMSE为15.19%(轻微)、17.46%(中等)、9.12%(严重)和9.63%(不抗病)。LAI反演模型普遍在病害早期和中期(病情等级1~7)对玉米LAI估算精度相差不大。但是对病情极其严重的玉米样本,其玉米LAI估算结果精度差异较大,田间不同病情等级胁迫会影响玉米LAI的准确估算。Leaf area index(LAI)is a significant phenotypic parameter to characterize maize growth information.Accurate estimation of LAI under disease stress using UAV with multispectral camera is important for phenotypic research and breeding engineering.Maize disease is a common problem in the process of germplasm resources identification and breeding.The remote sensing estimation of maize LAI under the condition of disease occurrence needs to be considered.Firstly,the LAI data of different disease severities(grade 1~9)were obtained from 41 maize inbred lines with different disease resistances by artificial inoculation.The leaf-scale ASD FieldSpec PRO 4 hyperspectral data of maize were collected to classify the disease resistance of different maize germplasm resources.The hyperspectral indexes related to leaf nitrogen,chlorophyll and specific leaf weight were constructed,and the recognition models of different maize disease grades were developed.To identify different maize disease grades,the hyperspectral indexes relating to leaf nitrogen,chlorophyll,and specific leaf weight were developed and used in the recognition models.Four integrated learning models,K-nearest neighbours(KNN),support vector machine(SVM),gradient boosting decision tree(GBDT)and decision tree(DT)were constructed to classify the disease resistance of different maize germplasm resources.Multi-spectral images were obtained from the Rededge MX5 multispectral camera carried on DJI Matrice M600 Pro(UAV)at different stages of maize growth.Based on red edge and near infrared bands,a variety of vegetation indices were constructed to estimate the maize LAI.According to maize disease recognition model,different disease stress levels were classified and identified.The robustness of LAI estimation model under different disease stress was discussed by using the four integrated machine learning models of random forest regression(RFR),extreme gradient boosting(XGBoost),light gradient boosting machine(LightGBM),and GBDT.The results showed that based on the ground ASD h

关 键 词：玉米病害 叶面积指数 无人机 高光谱 多光谱 

分 类 号：TP75[自动化与计算机技术—检测技术与自动化装置]