基于大米特征矿物元素产地鉴别建模方法比较研究

Comparative Study on Modeling Methods of Rice Origin Identification Based on Characteristic Mineral Elements

基于前期实验筛选到的与产地和母质土壤直接相关的23种特征矿物元素作为产地判别特征指标,比较Fisher判别模型和前馈神经网络预测模型的适用性。以连续3年随机采集的五常、查哈阳和建三江地理保护区274份样本作为建模对象,以模型的判别率为指标,采用电感耦合等离子体质谱仪(ICP-MS)和X射线荧光光谱仪测定样本中特征矿物元素含量(Mg、Ca、Cr、Mn、Zn、As、Rb、Sr、Ag、Cd、Sb、Te、Ba、La、Nd、Sm、Gd、Dy、Ho、Er、Yb、Pb、U),结合线性(一般线性判别分析)和非线性(前馈神经网络训练方法)模型构建方法用于产地鉴别。结果表明,Fisher判别分析模型对训练集判别率为81.5%,交叉检验判别率为78.8%,测试集总体判别率为87.5%。前馈神经网络预测模型对大米产地识别结果平均相对误差值为17.14%,产地的整体识别准确率为100%。筛选到的特征矿物元素携带了足够多的产地信息,通过前馈神经网络法建立的判别模型具有更优的判别能力,能解决小距离相似自然环境产地样本难以识别的问题。

Twenty-three mineral elements directly related to the producing area and parent soil were selected from the previous experiments as the discriminant index of producing area.The applicability of Fisher discriminant model and feed forward neural network prediction model was compared.A total of 274 samples from Wuchang,Cha-hayang and Jiansanjiang geographical protected areas were selected as the modeling objects.The discriminant rate of the model was used as an indicator.The contents of characteristic mineral elements(Mg,Ca,Cr,Mn,Zn,As,Rb,Sr,Ag,Cd,Sb,Te,Ba,La,Nd,Sm,Gd,Dy,Ho,Er,Yb,Pb,U)in samples were determined by inductively coupled plasma mass spectrometry(ICP-MS)and X-ray fluorescence spectrometer.Linear(general linear discriminant analysis)and nonlinear(feed forward neural network training method)model building methods were combined for origin identification.The results were as follows,the discriminant rate of Fisher discriminant analysis model to the training set was 81.5%.The discriminant rate of cross test was 78.8%.The overall discriminant rate of the test set was 87.5%.The average relative error of the feed forward neural network prediction model for rice origin identification was 17.14%.The overall identification accuracy of origin was 100%.The selected characteristic mineral elements carried enough information of origin.The discriminant model established by feed forward neural network had abetter discriminant ability.This could solve the problem that it was difficult to identify samples from similar habitats in small distances.