【摘要】 为了实现小麦籽粒蛋白质含量的快速、准确测定,用近红外分析仪对158份小麦进行光谱扫描,采用主成分分析法剔除异常光谱,对剔除异常值后的图谱进行标准正常化及去散射处理,并分别进行一阶和二阶导数处理。并在光谱预处理基础上,建立了预测小麦籽粒蛋白质含量的BP神经网络和偏最小二乘法校正模型。结果表明:经过标准正常化及去散射处理和二阶导数预处理的图谱,运用BP神经网络建立的模型预测小麦籽粒蛋白质含量效果最优,预测的R2和均方根误差分别为0.983和0.067,小麦蛋白质含量的国标测定值与最优条件下的预测值之间的t检验结果为P=0.82>0.05,两种方法测定结果无显著性差异。将其与近红外仪器自带模型相比,预测效果显著提高。采用非线性BP神经网络法建立的定标模型可提高预测小麦蛋白质含量的准确性。更多还原

【Abstract】 In order to achieve a rapid and accurate determination method of protein content in wheat grains,the near-infrared reflectance spectroscopy was used to scan 158 wheat samples.The abnormal spectrums were removed by principal component analysis.After rejecting the abnormal value,the spectrums were processed by standard normalization and scattering,and by first and second derivative processing,respectively.The BP neural network and partial least squares calibration model used to forecast of the protein content of wheat grain were established based on spectral preprocessing.The results showed that:the prediction effect of spectrums processed by standard normalization,scattering and second derivative processing was optimal.The prediction of R2 and RMS error was 0.983 and 0.067 respectively,t-test result between the measured value of national standard method and prediction value of optimum condition was P=0.82>0.05,and there was no significant difference.The prediction effect improved significantly compared with the model installed in near-infrared instrument.The calibration model established by non-linear BP neural network could improve the predicting accuracy of protein content in wheat.更多还原