不同核函数支持向量机和可见-近红外光谱的多种植被叶片生化组分估算
|
摘要
氮、磷、钾元素是植物有机质的重要生化组分,准确估算其含量对监测管理植被的新陈代谢和健康状况具有重要意义。可见-近红外光谱结合多种建模方法已被用于植被生化参数的监测,其中支持向量机回归方法被证明能够较好拟合反射光谱和植被生化参数之间的非线性关系,而选取适当的核函数是其成功的关键。以宜兴地区水稻、玉米、芝麻、大豆、茶叶、草地、乔木和灌木等八种植被叶片样本为研究对象,分析比较基于径向基核函数、多项式核函数和S形核函数的支持向量回归模型估算叶片氮、磷、钾元素含量的能力。利用一阶微分变换、标准正态变量变换和反对数变换对叶片可见-近红外光谱进行预处理,运用bootstrapping法生成1 000组校正集和验证集,分别建立基于三种核函数的支持向量回归估算模型,以决定系数(R2)和相对分析误差(RPD)的均值作为评价指标。结果显示,结合一阶微分和反对数变换光谱,采用径向基核函数模型对氮、钾元素估算精度最高(氮:平均R2=0.64,平均RPD=1.67;钾:平均R2=0.56,平均RPD=1.48),结合一阶微分变换光谱,采用径向基核函数模型对磷元素估算精度最高(磷:平均R2=0.68,平均RPD=1.73)。研究表明,结合不同预处理的可见-近红外光谱,基于径向基核函数的支持向量回归模型具有较好的估算多种植被叶片生化组分含量的潜力。
Nitrogen(N),phosphorus(P)and potassium(K)are important biochemical components of plant organic matters,and estimating their contents are useful for monitoring plant metabolism processes and health.Visible and near-infrared(VNIR)spectroscopy has been applied to monitor plant biochemical parameters with many modeling methods,in which support vector machine(SVM)has been proved to be a potential approach for modeling the nonlinear relationships between the reflectance spectra and biochemical parameters of plant organic matters,and the successful application of SVM relies on the proper selection of kernels.This study aimed to compare the performances of radial basis function(RBF),polynomial and sigmoid kernels based support vector machine regression(SVR)models in estimating the contents of nitrogen(cN),phosphorus(cP)and potassium(cK)of diverse plant leaves using laboratory-based VNIR spectroscopy.The cN,cP,cKand VNIR reflectance of leaf samples in eight plant species(rice,corn,sesame,soybean,tea,grass,shrub and arbor)were measured in laboratory.Three transformation methods,namely the first derivative(FD),standard normal variate(SNV)and logrithmic reciprocal transformation(Log(1/R))were used for spectral transformation.The SVR models using three aforementioned kernels were calibrated and validated with 1 000 bootstrap sample datasets.The average determination coefficients(R2)as well as ratio of performance to standard deviate(RPD)were calculated to compare the performances of three different kernels.The results showed that,the RBF kernel based SVR model with FD and absorbance transformation obtained the best accuracy for cNand cKestimations(cN:mean R2=0.64,mean RPD=1.67;cK:mean R2=0.56,mean RPD=1.48),and the RBF kernel based SVR model with FD transformation obtained the best accuracy for cPestimations(cP:mean R2=0.68,mean RPD=1.73).The study indicated that RBF kernel based SVR model has great potential in estimating biochemical component contents of diverse plant leaves with VNIR spectroscopy.
Nitrogen(N),phosphorus(P)and potassium(K)are important biochemical components of plant organic matters,and estimating their contents are useful for monitoring plant metabolism processes and health.Visible and near-infrared(VNIR)spectroscopy has been applied to monitor plant biochemical parameters with many modeling methods,in which support vector machine(SVM)has been proved to be a potential approach for modeling the nonlinear relationships between the reflectance spectra and biochemical parameters of plant organic matters,and the successful application of SVM relies on the proper selection of kernels.This study aimed to compare the performances of radial basis function(RBF),polynomial and sigmoid kernels based support vector machine regression(SVR)models in estimating the contents of nitrogen(cN),phosphorus(cP)and potassium(cK)of diverse plant leaves using laboratory-based VNIR spectroscopy.The cN,cP,cKand VNIR reflectance of leaf samples in eight plant species(rice,corn,sesame,soybean,tea,grass,shrub and arbor)were measured in laboratory.Three transformation methods,namely the first derivative(FD),standard normal variate(SNV)and logrithmic reciprocal transformation(Log(1/R))were used for spectral transformation.The SVR models using three aforementioned kernels were calibrated and validated with 1 000 bootstrap sample datasets.The average determination coefficients(R2)as well as ratio of performance to standard deviate(RPD)were calculated to compare the performances of three different kernels.The results showed that,the RBF kernel based SVR model with FD and absorbance transformation obtained the best accuracy for cNand cKestimations(cN:mean R2=0.64,mean RPD=1.67;cK:mean R2=0.56,mean RPD=1.48),and the RBF kernel based SVR model with FD transformation obtained the best accuracy for cPestimations(cP:mean R2=0.68,mean RPD=1.73).The study indicated that RBF kernel based SVR model has great potential in estimating biochemical component contents of diverse plant leaves with VNIR spectroscopy.
引文
[1] Mahajan G,Sahoo R,Pandey R,et al.Precision Agriculture,2014,15(5):499.
[2] Tang R,Chen X,Li C.Applied Spectroscopy,2018,72(5):740.
[3] Cui L,Fei T,Qi Q,et al.International Journal of Remote Sensing,2013,34(5):1866.
[4] Sanches I,Tuohy M,Hedley M,et al.International Journal of Remote Sensing,2013,34(1):276.
[5] Vapnik V N.The Nature of Statistics Learning Theory.2nd ed.Springer-Verlag New York Inc.,2000.
[6] Wang L,Zhou X,Zhu X,et al.Computers and Electronics in Agriculture,2017,140:327.
[7] Yao X,Huang Y,Shang G,et al.Remote Sensing,2015,7(11):14939.
[8] Zhai Y,Cui L,Zhou X,et al.International Journal of Remote Sensing,2013,34(7):2502.
[9] LIU Wen-jie,LI Wei-jun,LI Hao-guang,et al(刘文杰,李卫军,李浩光,等).Spectroscopy and Spectral Analysis(光谱学与光谱分析),2017,37(7):2024.
[10] Yang X,Huang J,Wu Y,et al.Science China Life Sciences,2011,54(3):272.
[11] Smola A,Sch9lkopf B.Statistics and Computing,2004,14(3):199.
[12] Chang C C,Lin C J.ACM Transactions on Intelligent Systems and Technology(TIST),2011,2(3):Article 27(doi:10.1145/1961189.1961199).
[13] Zhou X,Ma Y.Communications in Statistics-Simulation and Computation,2013,42(10):2175.
[14] Axelsson C,Skidmore A,Schlerf M,et al.International Journal of Remote Sensing,2013,34(5):1724.
[15] Xu S,Zhao Y,Wang M,et al.Geoderma,2018,310:29.
[2] Tang R,Chen X,Li C.Applied Spectroscopy,2018,72(5):740.
[3] Cui L,Fei T,Qi Q,et al.International Journal of Remote Sensing,2013,34(5):1866.
[4] Sanches I,Tuohy M,Hedley M,et al.International Journal of Remote Sensing,2013,34(1):276.
[5] Vapnik V N.The Nature of Statistics Learning Theory.2nd ed.Springer-Verlag New York Inc.,2000.
[6] Wang L,Zhou X,Zhu X,et al.Computers and Electronics in Agriculture,2017,140:327.
[7] Yao X,Huang Y,Shang G,et al.Remote Sensing,2015,7(11):14939.
[8] Zhai Y,Cui L,Zhou X,et al.International Journal of Remote Sensing,2013,34(7):2502.
[9] LIU Wen-jie,LI Wei-jun,LI Hao-guang,et al(刘文杰,李卫军,李浩光,等).Spectroscopy and Spectral Analysis(光谱学与光谱分析),2017,37(7):2024.
[10] Yang X,Huang J,Wu Y,et al.Science China Life Sciences,2011,54(3):272.
[11] Smola A,Sch9lkopf B.Statistics and Computing,2004,14(3):199.
[12] Chang C C,Lin C J.ACM Transactions on Intelligent Systems and Technology(TIST),2011,2(3):Article 27(doi:10.1145/1961189.1961199).
[13] Zhou X,Ma Y.Communications in Statistics-Simulation and Computation,2013,42(10):2175.
[14] Axelsson C,Skidmore A,Schlerf M,et al.International Journal of Remote Sensing,2013,34(5):1724.
[15] Xu S,Zhao Y,Wang M,et al.Geoderma,2018,310:29.