| Aksu Fuji Apple has become a specialty of Aksu region due to its sweet and refreshing taste and unique "sugar heart".With the increasing influence of Aksu Fuji Apple brand,the phenomenon of selling counterfeit products has gradually appeared in the market.In order to enhance the brand effect of Aksu Fuji Apple and make Aksu Fuji Apple famous all over the country,this article is based on the machine learning method to study the identification of Aksu Fuji Apple,using hyperspectral imaging technology to provide a rapid identification method for apple varieties,in order to improve the accuracy of the classification model,using chemometrics The method of collecting data is fused with the spectral data model,reducing the number of samples required to build the model.Because this experiment is limited by time and research level,there are still many shortcomings.It is hoped that in future research,we can learn Raman spectroscopy technology to establish a more efficient fusion model to deal with practical problems.The research content and results of this article are as follows:(1)This paper uses 9 preprocessing methods(first derivative,standard normal transformation,multivariate scattering correction,principal component analysis,principal component analysis based on kernel function,linear discriminant analysis,local linear embedding,factor analysis,Multi-dimensional scale analysis),the first-order derivative and multivariate scattering processing spectral noise classification method has a higher accuracy rate,and the principle component analysis and factor analysis select the characteristic variable classification method with higher accuracy rate.(2)Collect apple physical and chemical content data by chemometric method,including soluble solids,hardness,p H,moisture,weight and volume.Through continuous control of data variables during the modeling process,it is finally determined that there is a greater correlation with apple classification The physical and chemical components are: soluble solids,hardness,p H value.According to the above nine methods to construct the classifier,the experimental results are that the random forest,K-nearest neighbor and naive Bayes method are better,and the classification accuracy is above 0.77,and the multi-layer perceptron is only 0.547 poor.(3)Use PCA,FCA,MDS and other algorithms to extract spectral features from the denoised spectral data,and simplify the model input variables.Among them,PCA reduces the dimensionality to the sum of the correlation coefficients of 5 feature vectors accounting for 97%.A total of 162 apple classification models were established by using 9 methods including KNN,Naive Bayes algorithm,and quadratic discriminant analysis.The experimental results showed that the QDA-D1-PCA model has the best classification effect,and its accuracy rate is 0.862.(4)The classification model of apple spectrum data and physical and chemical data was successfully established.In order to improve its accuracy,the model fusion method was used to fuse the apple hyperspectral data model and the physical and chemical content data model to improve the average accuracy of the model.In this paper,the fusion model adopts the voting method,and selects 5high-accuracy hyperspectral data models and physical and chemical content data models for fusion.Through continuous exploration,finally adopts a 5:2 fusion method to increase the classification accuracy of the fusion model to 90.5%.,The identification rate of Aksu Fuji Apple is 93.8%. |
