基于集成学习的水稻氮素营养及籽粒蛋白含量监测  

Monitoring Nitrogen Nutrition and Grain Protein Content of Rice Based on Ensemble Learning

作　　者：张杰[1,2] 徐波 冯海宽[1] 竞霞[2] 王娇娇[1] 明世康 傅友强 宋晓宇[1] ZHANG Jie;XU Bo;FENG Hai-kuan;JING Xia;WANG Jiao-jiao;MING Shi-kang;FU You-qiang;SONG Xiao-yu(Information Technology Research Center,Beijing Academy of Agriculture and Forestry Sciences,Beijing 100094,China;School of Surveying and Mapping Science and Technology,Xi’an University of Science and Technology,Xi’an 710054,China;Rice Research Institute,Guangdong Academy of Agricultural Sciences,Guangzhou 510640,China)

机构地区：[1]北京市农林科学院信息技术研究中心,北京100094 [2]西安科技大学测绘科学与技术学院,陕西西安710054 [3]广东省农业科学院水稻研究所,广东广州510640

出　　处：《光谱学与光谱分析》2022年第6期1956-1964,共9页Spectroscopy and Spectral Analysis

基　　金：广东省重点领域研发计划项目(2019B020214002);国家自然科学基金项目(42171394)资助。

摘　　要：利用高光谱遥感技术在水稻收获前对籽粒品质相关的蛋白质含量进行监测,一方面可以及时调整栽培管理方式,指导合理追肥,另一方面,有助于提前掌握籽粒品质信息,明确市场定位。该研究以广东省典型优质籼稻为研究目标,基于2019年和2020年两年氮肥梯度实验,以水稻分化期和抽穗期冠层尺度高光谱数据、水稻氮素参数,包括叶片氮素含量(LNC)、叶片氮素积累量(LNA)、植株氮素含量(PNC)、植株氮素积累量(PNA)及籽粒蛋白含量数据为基础,利用四种个体机器学习算法partial least square regression(PLSR)、K-nearest neighbor(KNN)、Bayesian ridge regression(BRR)、support vector regression(SVR),三种集成学习算法random forest(RF)、adaboost、bagging,针对水稻不同生育期氮素状况进行监测建模,在此基础上构建基于水稻冠层光谱信息、光谱信息结合水稻农学氮素参数的籽粒蛋白含量的监测模型,并对模型进行精度对比。研究结果表明,在水稻氮素营养监测方面,利用水稻冠层454~950 nm波段信息,采用RF及Adaboost算法,在水稻分化期、抽穗期及全生育期LNC、LNA、PNC及PNA模型R^(2)均达到0.90以上,同时也具有较低的RMSE和MAE。在水稻籽粒蛋白品质监测方面,采用全波段光谱信息进行籽粒蛋白含量监测时,RF具有最高的精确度与稳定性,两生育期的RF模型对籽粒蛋白含量的监测结果R^(2)分别为0.935和0.941,RMSE分别为0.235和0.226,MAE分别为0.189和0.152;两生育期以全波段光谱信息结合长势参数进行籽粒蛋白监测时,Adaboost模型具有最高的精确度和稳定性,其中分化期全波段光谱信息结合PNA作为输入参数,Adaboost模型R^(2)为0.960,RMSE为0.175,MAE为0.150,以抽穗期全波段光谱信息结合PNC作为输入参数,R^(2)为0.963,RMSE为0.170,MAE为0.137。研究结果表明,与PLSR,KNN,BRR和SVR几种个体学习器算法相比,集成算法RF,Adaboost和Bagging具备良好的处理多重共线�The use of hyperspectral remote sensing technology to monitor the protein content related to grain quality before rice matures is important.It can promptly adjust cultivation management methods and guide reasonable fertilization and help to grasp rice grain quality information in advance and clarify market positioning.This study took typical high-quality Indica rice in Guangdong Province as the research goal.Two-year nitrogen gradient experiments were carried on in 2019 and 2020.The canopy level hyperspectral data and rice nitrogen parameters,including leaf nitrogen content(LNC),leaf nitrogen accumulation(LNA),plant nitrogen content(PNC),and plant nitrogen accumulation(PNA),were collected at the rice panicle initiation stage and heading stage.Then,four individual machine learning algorithms,Partial Least Square Regression(PLSR),K-Nearest Neighbor(KNN),Bayesian Ridge Regression(BRR),Support Vector Regression(SVR),and three ensemble learning algorithms,Random forest(RF),Adaboost,Bagging were used for monitoring and modeling the nitrogen status of rice at different growth stages.After that,the rice grain protein content estimation models based on rice canopy spectral information,and spectral information combined with rice nitrogen parameterswere constructed by different machine learning algorithms.The rice nitrogen and grain protein content estimation models’accuracy were evaluated and compared.The study results showed that for rice nitrogen nutrition monitoring,using the rice canopy spectral information from 454~950 nm,the R^(2)of LNC,LNA,PNC and PNA estimation models based on RF and Adaboost algorithms achieved above 0.90 at the rice,heading stage,with low RMSE and MAE.Panicle initiation stage When using full-band spectral information to estimaterice grain protein content,RF had the highest accuracy and stability,with R^(2)of 0.935 and 0.941,RMSE of 0.235 and 0.226,and MAE of 0.189 and 0.152 at rice panicle initiation and heading stage,respectively.Adaboost model has the highest accuracy and stability for seed p

关 键 词：高光谱遥感 水稻品质 机器学习 集成算法 ADABOOST Random forest 

分 类 号：TP79[自动化与计算机技术—检测技术与自动化装置]