利用高光谱反射率预测枸杞叶片氮素含量

Predicting Leaf Nitrogen Content in Wolfberry Trees Using Hyperspectral Reflectance

为实现枸杞氮素含量的快速无损监测,以“宁杞7号”为研究对象,同步测定枸杞叶片光谱反射率与叶片氮素含量,选用快速傅里叶变换(FFT)对测定的枸杞叶片光谱进行平滑滤波处理,并获取原始光谱(OS)。采用一阶微分、二阶微分及连续统去除法对原始光谱进行变换处理,获取枸杞叶片一阶微分光谱(FDS)、二阶微分光谱(SDS)及连续统去除光谱(CRS),将原始光谱及3种变式光谱分别与叶片氮素含量进行相关性分析,进而筛选出敏感波长,并构建预测枸杞叶片氮素含量的随机森林回归模型(RFRM)和多元线性回归模型(MLRM)。结果表明,基于3种变换光谱构建的RFRM和MLRM的预测精度均优于基于OS构建的模型;其中:基于FDS构建的模型预测效果最优,其次为基于SDS和基于CRS构建的模型,基于OS构建的模型预测精度最差。同时表明,RFRM的拟合度均优于MLRM,基于原始光谱OS、一阶微分光谱FDS、二阶微分光谱SDS和连续统去除光谱CRS构建的RFRM,同MLRM相比,其模型拟合度分别提高0.258、0.259、0.275和0.291,RMSE分别降低0.044、0.054、0.059、0.076,MAE分别降低0.045、0.043、0.066、0.059。基于FFT+FDS组合方法下筛选的敏感波长构建的RFRM,其建模数据集的拟合度、RMSE和MAE分别为0.897、0.071和0.058,检验数据集的决定系数、RMSE和MAE分别为0.689、0.129和0.102,模型具有良好的精度和稳定性,可作为枸杞叶片氮素的高光谱估测方法。

To realize rapid and non-destructive monitoring of nitrogen content in wolfberry tree,"Ningqi No.7"wolfberry was selected as the research object to synchronously measure the spectral reflectance and nitrogen content of wolfberry leaves.Fast fourier transform(FFT)was performed to smooth and filter the measured spectra to obtain the original spectra(OS).Three types of mathematical transformations including first-derivative(FD),second-derivative(SD)and continuum removal(CR)transformations were performed on the original spectra,and the corresponding spectral datasets including first-derivative spectra(FDS),second-derivative spectra(SDS)and continuum removal spectra(CRS)were obtained.The correlation analysis between the spectra including OS,FDS,SDS and CRS and nitrogen of wolfberry leaves were performed to select sensitive wavelengths based on the value of correlation coefficients.Random forest regression models(RFRM)and multiple linear regression models(MLRM)were constructed using selected sensitive wavelengths to predict the nitrogen content of wolfberry leaves.The study indicated that the prediction accuracy of RFRM and MLRM constructed using the three types of transformation spectra were better than those constructed using OS.Among them,the models constructed using FDS had the best prediction performance,followed by the models constructed using SDS and the models constructed using CRS,the models constructed using OS had the worst prediction accuracy.Meanwhile,it is shown that the prediction accuracy of RFRM were superior to MLRM.Compared to MLRM,the fitting degree of RFRM constructed using OS,FDS,SDS and CRS increased by 0.258,0.259,0.275 and 0.291,the root mean square error(RMSE)decreased by 0.044,0.054,0.059 and 0.076,and the mean absolute error(MAE)decreased by 0.045,0.043,0.066 and 0.059,respectively.The RFRM model constructed using the sensitive wavelengths selected from FDS had best accuracy and stability,with the determination coefficient,RMSE and MAE of calibration set of 0.897,0.071 and 0.058,respectively,with the determination coefficient,RMSE and MAE of validation set of 0.689,0.129 and 0.102,respectively.It can be used as a hyperspectral estimation method for leaf nitrogen content.