高光谱估测稻叶SPAD值预处理方法与模型比较  

Comparison of Pre-processing Methods and Models for Estimating SPAD of Rice Leaves by Hyperspectrum

作　　者：严志雁 王芳东 郭熙[4] 丁建 YAN Zhi-yan;WANG Fang-dong;GUO Xi;DING Jian(Institute of Agricultural Economics and Information,Jiangxi Academy of Agricultural Sciences,Nanchang 330200,China;Jiangxi Province Engineering Research Center of Information Technology in Agriculture,Nanchang 330200,China;Base Management Center,Jiangxi Academy of Agricultural Sciences,Nanchang 330200;Key Laboratory of Poyang Lake Watershed Agricultural Resources and Ecology of Jiangxi Province,Jiangxi Agricultural University,Nanchang 330045,China)

机构地区：[1]江西省农业科学院农业经济与信息研究所,江西南昌330200 [2]江西省农业信息化工程技术研究中心,江西南昌330200 [3]江西省农业科学院基地管理中心,江西南昌330200 [4]江西农业大学江西省鄱阳湖流域农业资源与生态重点实验室,江西南昌330045

出　　处：《江西农业大学学报》2020年第6期1130-1138,共9页Acta Agriculturae Universitatis Jiangxiensis

基　　金：国家重点研发计划项目(2017YFD0301603);江西省现代农业产业技术体系建设专项资金资助(JXARS-21-农机信息化应用)。

摘　　要：【目的】比较筛选高光谱数据预处理方法和水稻叶片SPAD值估测模型,为利用高光谱技术测定水稻叶片叶绿素含量提供依据。【方法】研究对象为96份不同施肥处理下的水稻叶片样本,利用ASD FildSpec 4测定叶片光谱,采用叶绿素计SPAD-502测定SPAD值。采用7种光谱预处理方法处理350~2500 nm光谱,结合3种回归模型(偏最小二乘、支持向量机和随机森林算法),建立了高光谱反射率与水稻叶片SPAD值的映射关系,比较了模型的预测精度。【结果】(1)BC、SG、SG+BC、SG+SNV预处理提高了PLSR模型验证集建模精度;SG、SG+BC预处理提高了SVR模型验证集建模精度;SG、SG+BC、SG+MSC、SG+SNV提高了RFR模型验证集建模精度;(2)SG+BC预处理能提高PLSR、SVR、RFR模型建模精度,说明采用消除信号不稳定造成的噪声、背景细小噪声和低频信号干扰对于提高水稻叶片的高光谱反演精度有重要的作用。(3)数据预处理后RFR模型精度最佳,验证集的平均决定系数R2为0.84,RMSE为13.70,RPD为2.59。【结论】SG及其复合预处理方法与随机森林回归模型结合使用,可作为高光谱估测水稻叶片SPAD值的参考方法。[Objective]The aim of this study was to compare and screen hyperspectral data preprocessing methods and SPAD estimation models of rice leaves,to provide a basis for the determination of chlorophyll content in rice leaves by hyperspectral technology.[Method]96 samples of rice leaves under different fertilization treatments were collected as the research object.The leaf spectrum was measured by ASD FildSpec 4 and the SPAD value was measured by chlorophyll meter SPAD-502.7 spectral preprocessing methods were used to process the 350-2500 nm spectrum,combined with 3 regression models(partial least squares,support vector machine and random forest algorithm)to establish the mapping relationship between the hyperspectral reflectance and the SPAD value of rice leaves,and the prediction accuracy of the models was compared.[Result]The results showed that,(1)BC,SG,SG+BC,SG+SNV improved the accuracy of the PLSR model;SG,SG+BC improved the accuracy of the SVR model;SG,SG+BC,SG+MSC,SG+SNV improved the accuracy of RFR model;(2)SG+BC Pre-processing could improve the modeling accuracy of PLSR,SVR and RFR models,which showed that eliminating the noise caused by signal instability,background fine noise and lowfrequency signal interference played an important role in improving the hyperspectral inversion accuracy of rice leaves.(3)The accuracy of RFR model was the best after data preprocessing.The average determination coefficient R2 of validation set is 0.84,RMSE was 13.70 and RPD is 2.59.[Conclusion]The combination of SG preprocessing and random forest regression can be used as a reference method for hyperspectral estimation of rice leaf SPAD value.

关 键 词：水稻叶片 SPAD 高光谱 预处理 

分 类 号：S511[农业科学—作物学] T179[一般工业技术]