Research On The Confirmation Technology Of Geographical Indication Rice Origin Based On Machine Learning Method

Study the feasibility of the application of machine learning method in the confirmation of the geographical origin of rice,and establish the confirmation model of origin in adjacent areas,which can provide the theoretical basis for constructing the geographical protection rice protection system.In this study,166 samples of rice were collected from Meihekou City and its adjacent areas in Jilin Province.The contents of 10 kinds of mineral elements(Cu,Zn,Fe,Mn,K,Ca,Na,Mg,Pb,Cd)in rice samples were measured by atomic spectrophotometer.The data obtained by the instrument analysis are sampled by sampling packet,the data is divided into training set and test set in the ratio of 7:3,which is for model establishment and prediction.The model was established by random forest(RF)and SVM in the machine learning method,and compared with the multivariate statistical discriminant model established by the linear discriminant analysis(LDA)method.The main conclusions are as follows:(1)Random forest and support vector machine,these two machine learning methods can be applied to the confirmation of geographical origin of rice,the establishment of adjacent areas of origin confirmation model has a high accuracy.The prediction accuracy is 96% and 94% respectively.(2)In the RF model,the eight parameters are selected from 10 elements as feature subsets by feature selection and parameter optimization.The original parameter mtry=3 ntree=500 is optimized as mtry=1 ntree=600.After optimization,building an RF model requires only eight elements.The accuracy rate of the external test set is increased from 94% to 96%,and the generalization ability is improved.(3)The support vector machine(SVM)is based on four kernel functions(linear kernel,Gaussian kernel,polynomial kernel,Sigmoid kernel).The accuracy of the model based on four kernel functions is improved by parameter optimization.The kernel function model has the highest precision,the number of support vector and optimization parameters are the least.Therefore,the feature of the external test set is improved from 91.67% to 94% based on the linear kernel function.(4)The LDA model is optimized by feature selection and the accuracy rate of the external test set is 92%.Compared with the model established by machine learning method(RF,SVM),there is no hypothesis constraint on the initial data,the generalization ability is stronger,the prediction of unknown data is more accurate,and they are superior to LDA model.(5)By comparing the three methods in the model accuracy and generalization ability,over-fitting degree and model construction cost comparison.The results show that the RF model is optimal,and its prediction accuracy is high,the generalization ability is strong,the degree of over-fitting is low,and the model construction cost is small.