基于近红外光谱技术和人工神经网络的玉米品种鉴别方法研究

提出了一种采用近红外光谱技术结合人工神经网络对玉米品种进行鉴别的方法。在3 800～10 000 cm-1(波长1 000～2 632 nm)范围内采集四种玉米单粒完整籽粒的近红外漫反射光谱，经Savitky-Golay平滑和多重散 射校正预处理后，对数据进行主成分分析，再结合人工神经网络技术进行品种鉴别。主成分分析表明，前8个主成分的累积贡献率达到99.602%。以前8个主成分作为网络输入，品种类型作为输出，建立三层LMBP神经网络模型。每个品种 各取30粒共120个样本用于建模，10粒共40个样本用于预测。模型对建模集120个样本鉴别率为100%，对预测集40个样本的鉴别率为95%。实验结果说明该方法能快速无损地鉴别玉米品种，为玉米的品种鉴别提供了一种新方法。

Study on Discrimination of Corn Seed Based on Near-Infrared Spectra and Artificial Neural Network Model

A new non-destructive and rapid method was developed to discriminate varieties of corn seeds.The method is based on near-infrared reflectance spectroscopy(NIRS) and artificial neural network(ANN).The corn seeds used for this study involved four varieties: Gaoyou115,Nongda368,Nongda108 and Nongda4 967.After collecting the near-infrared reflectance spectrum of each single seed in the range between 1 000 and 2 632 nm,the principal component analysis(PCA) was used to compress the NIR spectra,which had been preprocessed with Savitky-Golay and multiplicative scatter correction(MSC).The analysis results showed that the cumulate reliabilities of PC1 to PC8(the first eight principal components) were 99.602%.A three-layer back-propagation neural network(BPNN) was developed for classification,which was trained by the Levenberg-Marquard algorithm to improve the network training speed and efficiency.The LMBP was activated by the sigmoid function,and normalization of targets was used to get the best discrimination result of network.The first eight principal components of the samples were applied as LMBPNN inputs,and the values of the type of corn seeds were applied as the outputs.In this model,120 kernels were used as the training data set and 40 kernels were used as the test data set.Calculation results showed that the distinguishing rate of the four corn seed varieties was 95%.This model is reliable and practicable.The results demonstrated that this identification method was rapid and non-destructive,and could be used for classification.