中文摘要：

活体植物叶片叶绿素含量SPAD值易受叶片厚度、水分等影响,提出了基于多参数神经网络建模的叶绿素含量精细反演方法。通过测量叶片在中心波长分别为650,940和1 450nm光照射下的透过率,获得叶片的SPAD值和水分指数WI（water index）,同时用数字螺旋测微仪测量相应的叶片厚度并用分光光度法测得其叶绿素含量。利用建模集样本分别建立SPAD值与实测叶绿素含量之间的单参数模型和基于BP神经网络的WI、厚度及SPAD值与实测叶绿素含量之间的非线性模型。利用这两种模型分别计算获得验证集样本的叶绿素含量预测值,对预测值和实测值进行了相关分析和相对误差的分析。实验以340个三种不同植物叶片为样本,用以上方法进行了分析。结果表明,利用BP神经网络建模后,每种植物样本的叶绿素含量预测精度都有不同程度的提高,尤其对于叶片厚度值较大的样本,效果更为明显。数据显示所有混合样本平均相对误差绝对值由单参数模型的7.55%降低到5.22%,实测值与预测值的拟合决定系数由0.83提高到0.93。验证了利用多参数BP神经网络模型可以有效地提高活体植物叶绿素含量预测精度的可行性。

英文摘要：

Aiming at SPAD values of living plant leaf chlorophyll content affected easily by the blade thickness,water content, etc,a fine retrieval method of chlorophyll content based on multiple parameters of neural network model is presented.The SPAD values and water index（WI）of leaves were obtained by the leaf transmittance under the irradiation of light central wavelength in 650nm,940nm,1450nm respectively.Meanwhile,the corresponding blade thickness is got by micrometer and the chlorophyll content is measured by spectrophotometric method.To modeling samples,the single parameter model between SPAD values and chlorophyll content was built and the nonlinear model between WI,thickness,SPAD values and chlorophyll content was estab-lished based on BP neural network.The predicted value of chlorophyll content of test samples were calculated separately by two models,and the correlation and relative errors were analyzed between predicted values and actual values.340 samples of three different plant leaves were tested by the method described above in experiment.The results showed that compared with single parameter model,the prediction accuracy of three different plant samples were improved in different degrees,the average abso-lute relative error of chlorophyll content of all pooled samples predicted by BP neural network model reduced from 7.55% to 5.22%.the fitting determination coefficient is increased from 0.83 to 0.93.The feasibility were verified in this paper that the prediction accuracy of living plant chlorophyll content can improved effectively using multiple parameter BP neural network model.