Study On Origin Identification Of Bingtang Sweet Orange Based On Spectral Data Fusion

As a high-value agricultural product with geographical characteristics,the identification of Bingtang sweet orange from different origins has important reference value for producers and consumers.In this thesis,Inductively Coupled Plasma-Atomic Emission Spectroscopy,Fourier Transform Infrared Spectroscopy and High Performance Liquid Chromatography were used to collect the mineral elements,infrared spectra and chromatographic data of Bingtang sweet orange from four different places of origin.Partial Least Squares Discriminant Analysis（PLS-DA）,Support Vector Machine（SVM）and Extreme Learning Machine（ELM）models were established by using three types of data with Principal Component Analysis（PCA）,Backward interval PLS（Bipls）,Synergy interval PLS（Sipls）,Variable importance in projection（VIP）and other feature variable extraction methods,and improve the performance of the model through multi-source and homologous data fusion strategies at three levels to achieve the origin identification of Bingtang sweet orange.The main work is as follows:1.PCA,PLS-DA and SVM models were established by Fourier transform mid-infrared spectra and 15 mineral elements content information of Bingtang sweet orange’s peel and pulp combined with multiple feature variable extraction methods to identify the origin of Bingtang sweet orange.Infrared spectrum showed that the main components of Bingtang sweet orange were sugar and protein.The infrared spectral feature variables of peel and pulp were extracted by 4Sipls＿VIP were mainly concentrated in 1150～1000cm^-1,which further showed that there are sugar differences in Bingtang sweet orange from different regions.The Ca,Cr,K,Mg and Se elements in the peel and the Fe,K,Mg,P and Se elements in the pulp of Bingtang sweet orange showed significant differences among the four producing areas.PCA models based on mineral elements content information could identify the origin of Bingtang sweet orange,but some Bingtang sweet orange samples from Yunnan and Hunan provinces still overlap.4Sipls and 4Sipls＿VIP and other methods could eliminate the useless information of the infrared spectrum,reduce the calculation amount of the model,improve the accuracy.In addition,the PLS-DA model improved by 4Sipls and the SVM model improved by 4Sipls＿VIP achieved more than 96.67%accuracy in both infrared data sets.The PLS-DA models based on the content information of mineral elements could not accurately identify the origin of Bingtang sweet orange,the accuracy of the models in test set were less than 73.08%.The SVM model based on PCA extraction of mineral elements feature variables achieved 100%accuracy in the mineral elements data set of peel and pulp.The results shows that the mid-infrared spectral information and mineral elements information of peel and pulp could reflect the differences between different regions of Bingtang sweet orange.2.The HPLC data of Bingtang sweet orange pulp were collected and the content of citric acid,malic acid and tartaric acid in Bingtang sweet orange pulp were determined based on the external standard method.In addition,PCA,PLS-DA and SVM models were establish with multiple characteristic variable extraction methods to identify the origin of Bingtang sweet orange samples.The identification accuracy of PLS-DA model and SVM model based on 4Sipls were more than 95.83%.The results show that compared with the contents of three organic acids,the HPLC information could more comprehensively and accurately reflect the differences between different origins of Bingtang sweet orange.3.Three levels multi-source data fusion strategies were carried out for the mineral elements,infrared spectrum and chromatographic data of the pulp of Bingtang sweet orange.The infrared spectra and mineral elements information of the peel and pulp were fused by three levels homologous data fusion strategies respectively.Based on the fused data,PLS-DA,SVM and ELM models were established to determine the feasibility of multi-source and homologous data fusion strategies in the study of origin identification of Bingtang sweet orange.In the origin identification of Bingtang sweet orange based on multi-source data fusion,the multi-source data fusion strategy could improve the performance of the identification model compared with the model established using a single technology.The accuracy range of the identification model established based on infrared spectra and mineral elements fusion information,mineral element and chromatographic fusion information,infrared spectrum and chromatographic fusion information and three types of data fusion information were 100%～100%,93.33%～100%,91.67%～100%and 80.95%～100%respectively.Among PLS-DA,SVM and ELM origin identification models based on mid level multi-source data fusion,SVM model had the best performance.The mid level multi-source data fusion performance of infrared spectrum and mineral elements information were the best,and the identification accuracy of the models based on infrared spectrum and mineral elements mid level fusion information were 100%.In the origin identification of Bingtang sweet orange based on homologous data fusion,the homologous data fusion strategy could improve the performance of the identification model compared with the model established using single information.The models built based on the infrared spectrum mid level homologous fusion information of peel and pulp only needed 399 variables and the accuracy of identification were 100%,and the SVM and ELM models built using mid level homologous fusion information of peel and pulp mineral elements only needed20 variables and the accuracy of identification were 100%.The research results show that the combination of infrared spectra,mineral elements and HPLC information,as well as the combination of peel and pulp spectral information play a synergetic role in the origin identification of Bingtang sweet orange.The mid level multi-source and homologous data fusion strategy could effectively fuse the characteristic information to develop the origin identification model of Bingtang sweet orange with low computational complexity and high accuracy.The work of the thesis shows that the infrared spectra,mineral elements and HPLC information can reflect the differences between different origins of Bingtang sweet oranges.The mid level multi-source and homologous data fusion strategies can effectively fuse the characteristic information,achieve more accurate and scientific identification of Bingtang sweet orange origins while reducing the amount of data,which can provide a reference for the quality control of Bingtang sweet orange and other fruits.