基于机器学习方法的地理标志大米产地确证技术研究

发布时间：2018-08-21 11:11

【摘要】：探讨机器学习方法在地理标志大米产地确证技术上应用的可行性,建立相邻区域的产地确证模型,可为构建地理标志大米保护体系提供理论依据。本研究采集来自吉林省梅河口市及其相邻区域的水稻样本166份,通过原子分光光度计检测大米样本中10种矿物质元素(Cu、Zn、Fe、Mn、K、Ca、Na、Mg、Pb、Cd)含量。将仪器分析得到的数据通过sampling包实现数据分层抽样,以7:3的比例划分训练集和测试集。分别利用机器学习方法中的随机森林(Random Forest,RF)和支持向量机(Support Vector Machine,SVM)建立模型,并与线性判别分析(Linear Discriminant Analysis,LDA)方法建立的多元统计判别模型进行了比较。主要结论如下:(1)随机森林和支持向量机这两种机器学习方法能够应用于地理标志大米产地确证中,建立的相邻区域产地确证模型具有较高的判别精度,其预测的准确率分别为96%、94%。(2)RF模型通过特征选择和参数优化,从10个元素中选取8个元素为特征子集,将原始参数mtry=3 ntree=500优化为mtry=1 ntree=600,优化后仅通过8个元素即可构建RF模型,外部测试集准确率由94%提高至96%,泛化能力提高。(3)支持向量机分别以四个核函数(线性核、高斯核、多项式核、Sigmoid核)为基础建立了确证模型,经参数优化后基于四个核函数的模型精度均有提高,其中线性核函数模型精度最高,所用支持向量和优化参数最少,因此基于线性核函数进一步进行特征选择,优化后模型的外部测试集准确率由91.67%提高至94%。(4)LDA模型经特征选择优化后外部测试集准确率为92%。机器学习方法(RF,SVM)建立的模型与之相比较,对初始数据没有假设前提的约束,泛化能力更强,对未知数据的预测更加准确,均优于LDA模型。(5)三种方法在模型精度与泛化能力、过拟合程度以及模型构建代价的比较结果说明,RF模型最优,其预测精度高,泛化能力强,过拟合程度低,模型构建代价小。
[Abstract]:This paper discusses the feasibility of applying machine learning method to the identification technology of geographical indication rice producing area, and establishes the model of producing area confirmation in adjacent regions, which can provide the theoretical basis for the construction of geographical indication rice protection system. In this study, 166 rice samples from Meiehekou City, Jilin Province, and their adjacent regions were collected. The contents of 10 mineral elements (Cu ~ (2 +) Zn ~ (Zn) Fe ~ (2 +) K _ (+) K _ (+) (Na) mg _ (Pb) (CD) in rice samples were determined by atomic spectrophotometer (AAS). The data obtained by instrument analysis are stratified by sampling package, and the training set and test set are divided according to 7:3 scale. The models of Random forest (RF) and support vector machine (Support Vector machine) were established and compared with the multivariate statistical discriminant model established by the linear discriminant analysis (Linear Discriminant) method. The main conclusions are as follows: (1) the two machine learning methods, random forest and support vector machine, can be applied to the identification of rice geographical indication. The prediction accuracy of RF model is 96 / 94 respectively. (2) the RF model selects 8 elements from 10 elements as feature subsets and optimizes the original parameter mtry=3 ntree=500 to mtry=1 ntree600. After optimization, the RF model can be constructed by only 8 elements. The accuracy of external test set is improved from 94% to 96%, and the generalization ability is improved. (3) support vector machine is based on four kernel functions (linear kernel, Gao Si kernel, polynomial kernel Sigmoid kernel). After parameter optimization, the accuracy of the model based on four kernel functions is improved, among which the linear kernel function model has the highest accuracy, the support vector and the optimized parameters are the least, so the feature selection is further based on the linear kernel function. The accuracy of the external test set of the optimized model is improved from 91.67% to 94. (4) the accuracy of the external test set of the LDA model is 922 after the feature selection and optimization. Compared with the model established by the machine learning method (RFS-SVM), it has no presupposition constraints on the initial data, has stronger generalization ability, and is more accurate in predicting unknown data than the LDA model. (5) the accuracy and generalization ability of the three methods are better than that of the LDA model. The comparison of the degree of overfitting and the cost of constructing the model shows that the RF model is optimal, with high prediction accuracy, strong generalization ability, low degree of overfitting and low cost of model construction.
【学位授予单位】：吉林农业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：S511;TP181

【参考文献】