Liang Liang, Yang Min-Hua, Liu Zhi-Xiao, Xu Hai-Wei, Liu Fu-Hui, He Qi-Zhuang, Luo Yun-Fei
School of Info-Physics and Geomatics Engineering, Central South University, Changsha 410083, China.
Guang Pu Xue Yu Guang Pu Fen Xi. 2009 Nov;29(11):2962-5.
A rapid and non-invasive method was put forward to measure the purity of hybrid rice seed by visible-near infrared reflectance spectra. Ninety hybrid rice seed samples (Yixiang 725) with the purity of 90%-99% were collected using a FieldSpec 3 visible-near infrared spectometer. All samples were divided randomly into two groups, one group with 75 samples used as calibrated set, and the other with 15 samples used as validated set. Based on the spectra in the range of 380-2 400 nm, the regression model was established using the PLS (partial least square), and different spectra pretreatment methods were compared. The study showed that spectra information can be extracted thoroughly by the pretreatment method of first derivative combined with standard normal variate, with the SEC (standard error of calibration) of 0.002 5, SEP (standard error of prediction) of 0.006 6, and determination coefficients of 0.988 4 (calibration set) and 0.922 7 (validation set) respectively. The spectra, which were pretreated with the method of first derivative combined standard normal variate, were analyzed by principal component analysis (PCA). The top 20 principal components, which were computed by PCA and accounted for 86.09% variation of the original spectral information, were used to build BP-ANN model for measuring the purity of hybrid rice seed as the new variables. The study showed that the SEC and SEP of BP-ANN model were 0.001 7 and 0.006 1, and the determination coefficients of that were 0.995 2 (calibration set) and 0.936 9 (validation set) respectively. Therefore, the predictive power of BP-ANN model was better than that of PLS model. Results indicated that it was feasible to measure the purity of the hybrid rice seed by visible-near reflectance spectra as a rapid and non-contact way, and PCA combined with BP-ANN was a preferred method.
提出了一种利用可见-近红外反射光谱快速、无损检测杂交水稻种子纯度的方法。使用FieldSpec 3可见-近红外光谱仪采集了90个纯度在90%-99%之间的杂交水稻种子样本（宜香725）。所有样本随机分为两组，一组75个样本作为定标集，另一组15个样本作为验证集。基于380-2400nm范围内的光谱，采用偏最小二乘法（PLS）建立回归模型，并比较了不同的光谱预处理方法。研究表明，一阶导数结合标准正态变量的预处理方法能充分提取光谱信息，定标标准误差（SEC）为0.0025，预测标准误差（SEP）为0.0066，定标集决定系数为0.9884，验证集决定系数为0.9227。对经一阶导数结合标准正态变量预处理的光谱进行主成分分析（PCA）。通过PCA计算得到的前20个主成分占原始光谱信息变异的86.09%，将其作为新变量建立用于测量杂交水稻种子纯度的BP-人工神经网络（BP-ANN）模型。研究表明，BP-ANN模型的SEC和SEP分别为0.0017和0.0061，定标集决定系数为0.9952，验证集决定系数为0.9369。因此，BP-ANN模型的预测能力优于PLS模型。结果表明，利用可见-近反射光谱快速、非接触地测量杂交水稻种子纯度是可行的，PCA结合BP-ANN是一种优选方法。
