【作者】 汤守鹏；

【导师】 朱艳；

【作者基本信息】 南京农业大学 ， 作物栽培学与耕作学， 2009， 硕士

【摘要】 作物氮素营养的快速、无损估测对提高产量和改善品质具有重要意义。近红外光谱为作物无损监测和信息的准确获取提供了有效手段。本研究的目的是以不同年份、不同品种、不同施氮水平的小麦田间试验为基础,基于傅立叶近红外光谱仪获取小麦主要生育时期鲜叶、干叶及其成熟期籽粒的光谱信息,运用偏最小二乘法(PLS)、BP神经网络(BPNN)和小波神经网络(WNN),分析并量化其与叶片全氮含量、糖氮比、籽粒蛋白质含量的关系,以实现小麦生长和品质信息的近红外快速预测模型。首先运用多种光谱预处理技术,对鲜叶在1155～1803 nm和2118～2500 nm光谱区间,干叶在2046～2155 nm、2297-2339 nm和2343～2378 nm光谱区间,用PLS、BPNN和WNN方法分别确立了小麦叶片氮含量的定量监测模型。结果显示,鲜叶和干叶的最优光谱预处理方法分别为MSC+Savitzky-Golay二阶导数和MSC+Norris一阶导数。对于鲜叶样品,PLS模型的预测均方根误差(RMSEP)和决定系数(R2)分别为0.216%和0.841；BPNN模型的RMSEP和R2分别为0.175%和0.894；WNN模型的RMSEP和R2分别为0.169%和0.901。对于粉末状干叶样品,PLS模型的RMSEP和R2分别为0.147%和0.910；BPNN模型的RMSEP和R2分别为0.101%和0.960；WNN模型的RMSEP和R2分别为0.094%和0.978。从模型精度和稳健性来看,神经网络法相对优于PLS法；粉末状干叶模型优于鲜叶模型。采用上述类似方法建立了小麦鲜叶和干叶糖氮比预测模型。结果显示,鲜叶光谱模型预测性能不佳；干叶在1655～2378 nm谱区范围内采用MSC+Norris一阶导数光谱预处理方法构建模型,表现较好；基于PLS、BPNN和WNN方法构建的糖氮比估算模型,其预测均方根误差(RMSEP)分别为0.332%、0.292%和0.288%,决定系数(R2)分别为0.853、0.865和0.870。进一步基于小麦籽粒近红外漫反射光谱构建了整粒小麦蛋白质含量预测模型。结果表明,对光谱进行多元散射校正并结合Norris一阶导数在1242～2230 nm谱区构建模型,表现较好；检验结果显示,PLS模型的RMSEP和R2分别为0.848%和0.794,BPNN模型的RMSEP和R2为0.770%和0.814,WNN模型的RMSEP和R2为0.761%和0.816；神经网络回归效果好于偏最小二乘法。最后探讨了同时估测小麦鲜叶和干叶可溶性总糖和全氮含量的可行性。结果表明,PLS、BPNN和WNN三种方法均不能准确的同时测定小麦鲜叶的全氮和总糖含量,但所构建的干叶WNN模型预测效果较好,其预测均方根误差(RMSEP)分别为0.101%和0.089%,决定系数(R2)分别为0.957和0.941；并且,在收敛速度和预测精度上,WNN模型均明显优于BPNN和PLS模型,其光谱预处理方法为：鲜叶为MSC+Savitzky-Golay二阶导数；干叶为MSC+Norris一阶导数；建模谱区为1100～2500nm。更多还原

【Abstract】 Quick and non-destructive monitoring of crop nitrogen status is useful for the managements to improve crop grain yield and quality. Near infrared (NIR) spectroscopy technology has provided an effective tool for non-destructive monitoring and crop information acquisition. In this study, a series of field experiments using different wheat varieties under various nitrogen levels were carried out in different years. Time-course near infrared spectrum (1100～2500 nm) were taken by Fourier transform near infrared spectrometer from fresh and dry wheat leaves, and mature grain. The purposes of this study were to develop prediction models for leaf nitrogen, the ratio of sugar to nitrogen and grain protein of wheat using different methods of chemometrics including partial least squares (PLS), back-propagation neural network (BPNN) and wavelet neural network (WNN), to realize quantitative analyzing wheat growth and quality information with NIR technique.Different spectra preprocessing ways combined with PLS, BPNN and WNN were used respectively to develop the models, in which 1155～1803 nm and 2118～2500 nm for fresh leaf nitrogen,2046～2155 nm,2297～2339 nm and 2343～2378 nm for dry leaf nitrogen. The result showed that MSC combined Savitzky-Golay second derivative, MSC combined Norris first derivative could be the suitable methods to develop nitrogen monitoring models for fresh and dry leaf, respectively. For fresh leaves, the root mean square errors of prediction (RMSEP) and coefficients of determination (R) by PLS model were 0.216% and 0.841, that by BPNN model were 0.175% and 0.894%, while that by WNN model were 0.169% and 0.901, respectively. For dry leaves, RMSEP and R2 by PLS model were 0.147% and 0.910, that by BPNN model were 0.101% and 0.960, while that by WNN model were 0.094% and 0.978, respectively. In term of prediction precision and stability of model, dry leaf was superior to fresh leaf, and neural network was superior to PLS relatively.Then, quantitative models for soluble sugar to nitrogen ratio in wheat leaves were established with the same way for leaf nitrogen. The evaluation results showed that, the models precision for fresh leaves was not satisfactory, but RMSEP by PLS, BPNN and WNN models for dry leaves based on 1655-2378nm were 0.332%,0.292% and 0.288%, while the coefficients of determination (R2) were 0.853,0.865, and 0.870, respectively. It seemed that artificial neural network models were superior to PLS model on prediction performance.Thirdly, the models for grain protein content of wheat were constructed with near infrared diffuse reflectance based on 1242～2230 nm after spectra data were pre-processed with MSC combined Norris first derivative method, which was optimized from MSC, Savitzky-Golay smoothing and derivative, Norris derivative methods. The evaluation results showed that, RMSEP and R2 by PLS model were 0.848% and 0.794, that by BPNN model 0.770% and 0.814, while that by WNN model were 0.761% and 0.816, respectively, it was clear that neural network models were superior to PLS model.Finally, the feasibility of simultaneous determination of soluble sugar and total nitrogen contents in fresh and dry wheat leaves was examined. The results showed that models based on PLS, BPNN and WNN could not be applied to estimating soluble sugar and total nitrogen contents in fresh wheat leaves simultaneously, but prediction precision of WNN model for dry leaves was satisfactory, RMSEP were 0.101% and 0.089%, with R2 of 0.957 and 0.941, respectively. In addition, WNN model was superior to BPNN and PLS models on convergence speed and prediction precision obviously. Spectra preprocessing methods was that, MSC combined Savitzky-Golay second derivative for fresh leaves, MSC combined Norris first derivative for dry leaves, and modeling spectrum range was 1100～2500 nm.更多还原