Nondestructive Testing Of "Yun He" Pear Based On Fusion Strategy And Visible-near Infrared Spectroscopy

With the development of social economy,consumers have increasingly strict requirements for fruit's quality.The total production of Chinese fruit ranks in the world's leading position,but the annual export volume accounts for only about 3%.The main reason is that fruit in China is not effectively graded and the level of commercialization is also low.So,the quality is uneven.Meanwhile,the market value and international competitiveness are unable to effectively increase.This paper takes"Yun he"pear of Zhejiang province as the object of research,using visible-near infrared spectroscopy(Vis-NIRS)and chemometric methods to detect its soluble solids content(SSC).SSC is an important indicator for evaluating the internal quality of pears.The main content and research conclusions are as follows:(1)Collect the"Yun he"pear's diffuse transmission spectrum and determine the SSC through six pre-processing methods(mean centralization,multiple scattering correction,standard normal transform,first-order derivative,second-order derivative,smoothing)optimize the spectrum and establish a partial least squares(PLS)model for the“Yun he”pear's SSC.The results show that the mean centralization is the optimal preprocessing scheme.(2)After optimization of the pre-processing,a total of three strategies are adopted:Uninformative variable elimination(UVE),Competitive Adaptive Reweighting Sampling(CARS)and its combination(UVE-CARS)Screen the effective band of the spectrum,and establish the PLS prediction model of"Yun he"pear's SSC.Comparing the prediction performance of the full-variable model and the three characteristic variable models,the UVE-CARS-PLS model achieved the best detection effect.(3)Two fusion strategies of stack fusion and error fusion are used to model the fusion of the two member models.One is an optimized PLS model determined by the first 12 LVs that minimized cross-validated root mean square error(RMSECV)during the calibration model phase,and the other is a multiple linear regression(MLR)model developed using the last 8 LVs.Compared with the general PLS model,the optimized PLS model,the MLR model based on the remaining information,the deviation model has a better predictive ability,its R_Pis 0.9026,and the RMSEP is0.5945 Brix.Compared with the best PLS member model,the evaluation indicators are improved 0.26%and 7.5%.(4)Comparing the prediction results of the UVE-CARS-PLS model based on variable selection and the error fusion model,the prediction performance of the error fusion model is better than any single model and stacked fusion model.Overall,the effective spectral information of the two member models is effectively used to improve the model prediction performance.(5)In order to explore the effectiveness and generality of error fusion strategy in the quality detection of other products,this paper established an error fusion model for predicting satsuma's SSC.The good prediction performance further proves that the fusion strategy has its specific advantages.