基于可见/近红外反射光谱的稻米品种与真伪鉴别

利用可见/近红外光谱技术对市场上5种稻米进行了鉴别.以ASD FieldSpec3地物光谱仪采集了5种稻米的光谱数据,各获取35个样本,随机分成训练集(150份)和检验集(25份),并分别采取全波段与特征波段(400~500nm、910~1400nm与1940~2300nm)两种方法建立模型进行分析.光谱经S.Golay平滑和标准归一化(SNV)处理后,以主成分分析法(PCA)降维.将降维所得的前9个主成分数据作为BP人工神经网络(BP-ANN)的输入变量,稻米品种作为输出变量,建立3层BP-ANN鉴别模型.利用25个未知样对模型进行检验,结果表明两类模型预测准确率均高达100%,其中特征波段模型比全波段模型具有更高的预测精度,说明利用可见/近红外技术结合PCA-BP神经网络分析法进行稻米品种与真伪的快速、无损鉴别是可行的,且提取特征波段是优化模型的有效方法之一.

DISCRIMINATION OF VARIETY AND AUTHENTICITY FOR RICE BASED ON VISUAL/NEAR INFRARED REFLECTION SPECTRA

Five different varieties of rice bought from the supermarket were identified with visual/near infrared spectroscopy (NIRS) technology. The spectra were collected by using ASD FieldSpec . 3 spectrometer, and 35 samples were obtained for each variety of rice. All the samples were divided randomly into two groups, one group with 150 ones used as calibrated set, and the other with 25 ones as validated set. Then the samples were analyzed with the whole wave band and the charac- teristic wave band(400～500nm, 910～1400nm and 1940～2300nm) models, respectively. The samples data were pre- treated by the methods of S.Golay smoothing and standard normal variable (SNV), and then the pretreated spectra data were analyzed with principal component analysis (PCA). The anterior 9 principal components computed by PCA were used as the input variables of back-propagation artificial neural network (BP-ANN) which included one hidden layer, while the values of rice varieties were used as the output variables of BP-ANN, and then the three layers BP-ANN discrimination model was built. The 25 unknown samples in the validated set were predicted by ANN-BP model. The results show that the discriminating rates are 100% in both models, and the accuracy of characteristic wave band model is higher than that of the whole wave band model. The results indicate that it is feasible to discriminate the variety and authenticity of rice by visible/ near infrared reflectance spectra as a rapid and non-contact way, and selecting characteristic wave band is one of the valida- ted methods to improve the precision of the discrimination model.