基于红边参数和人工神经网络的苹果叶片叶绿素含量估算

Estimation of chlorophyll content in apple leaves based on red edge parameters and artificial neural network

【目的】应用人工神经网络来解决苹果叶片叶绿素含量与高光谱数据之间的非线性关系,建立估测苹果叶片叶绿素含量的最佳模型,为苹果叶片叶绿素含量的快速无损监测提供参考。【方法】以位于陕西扶风杏林镇的树龄为4～5年的15棵苹果树为研究对象,2015年分别于果树花期(4月27日)、幼果期(5月30日)、果实膨大期(7月6日)、果实着色期(8月5日)、果实成熟期(9月11日)采集叶片样本,利用SVC HR-1024i型高光谱仪和SPAD-502叶绿素仪同步获取苹果叶片光谱值和对应的叶绿素含量,对原始光谱和一阶导数光谱与叶绿素含量之间的关系进行分析,从一阶导数光谱中提取苹果叶片光谱的5个红边参数,从5个红边参数中筛选出相关性好的红边参数,使用传统单变量回归算法、反向传播(back propagation,BP)神经网络和径向基函数(radial basis function,RBF)神经网络,建立叶绿素含量估测模型,用决定系数(R2)、均方根误差(RMSE)和相对误差(RE)来验证模型的准确性。【结果】与原始光谱相比,一阶导数光谱与苹果叶片叶绿素含量的相关性更高。5个红边参数中,红边位置、峰度系数、偏度系数与叶绿素含量的相关系数均较高,且均达极显著水平。建立的传统单变量回归模型中,基于红边位置、峰度系数和偏度系数的R2均大于0.7,其中基于偏度系数建立的多项式模型模拟精度最高,验证结果R2为0.872,RMSE为4.631,RE为8.81%。选取红边位置、峰度系数和偏度系数为人工神经网络的输入变量,分别优化BP神经网络的隐含层节点数和网络类型以及RBF神经网络的扩展系数(SPREAD值)来提高预测精度,结果发现,建立隐含层节点数为4的双隐含层BP神经网络最优模型R2为0.891,RMSE为3.844,RE为7.55%;当SPREAD值为0.6时,建立的RBF神经网络最优模型的R2为0.955,RMSE和RE分别为2.517和3.69%。【结论】估算苹果叶片叶绿素含量时,人工神经网络模型比传统单变量模型精度高,其中RBF神经网络模型学习速度快、精度高,拟合结果更加可靠。

Objective】The nonlinear relationship between hyperspectral data and chlorophyll content of apple leaf was solved using artificial neural network(ANN)and the best model for estimating chlorophyll content of apple leaves was established to provide basis for rapid and non destructive monitoring of chlorophyll content in apple leaves.【Method】A total of 15 apple trees at the age of 4-5 years in Xinglin,Fufeng,Shaanxi were selected and leaf samples were collected in flowering(April 27),young fruit(May 30),young fruit expansion(July 6),coloring(August 5),and fruit maturation(September 11)stages in 2015.Hyperspectral reflectance was collected by spectrometer(SVC HR 1024i),and relative chlorophyll content was measured by chlorophyll meter SPAD 502.The corrections between original spectrum,first derivative spectrum and chlorophyll content were analyzed.Five red edge parameters were calculated as the first derivative.The traditional single variable empirical regression algorithms,back propagation(BP)and radial basis function(RBF)based on the well related red edge parameters were applied to establish chlorophyll content estimated models.The accuracy of models was determined based on determination coefficient(R^2),root mean square error(RMSE)and relative error(RE).【Result】Compared with the original spectrum,the first derivative spectra had higher relevance with chlorophyll content of apple leaves.The red edge position,kurtosis and skewness had higher correlation among the 5 red edge parameters(P<0.01).The correlation coefficients of models based on red edge position,kurtosis and skewness reached over 0.7.The skewness based polynomial model was the best with R^2,RMSE and RE of 0.872,4.631 and 8.81%,respectively.Red edge position,kurtosis and skewness were selected as input variables of ANN to build model for estimating chlorophyll content.The node of hidden layer and network type in BP ANN and the SPREAD value in ANN RBF were optimized to improve precision.The R^2 based on the BP neural network optimal model with 4 node of double hidden layer was 0.891,and RMSE and RE were 3.844 and 7.55%.When SPREAD was 0.6,R^2,RMSE and RE of the optimized RBF-ANN were 0.955,2.517 and 3.69%,respectively.【Conclusion】Compared to the models with single variable,ANN can improve estimation accuracy of chlorophyll content prediction.RBF ANN had the advantages of rapid,accurate,and reliable.