Rapid Detection Research Of Bruising Lingwu Jujube Based On Vis/NIR Hyperspectral Imaging

Lingwu long jujube is the dominant characteristic jujube fruit in Ningxia.Long jujube contains a variety of minerals and vitamins,known as the "fruit treasure",its market share is increasing year by year.However,in the process of picking,grading,processing,packaging and transportation of lingwu long jujubes,mechanical damage(internal bruising)by collision,extrusion,vibration and other reasons causes were cause economic losses.Therefore,some simple,sensitive,economical,reliable and real-time evaluation methods are urgently needed for the rapid detection of bruising lingwu long jujube.This paper takes lingwu long jujube as the research object.The visible near infrared(Vis/NIR)hyperspectral imaging was used to distinguish the bruising grade and time of lingwu long jujube in bruised state,which provided a theoretical basis for online quality detection of lingwu long jujube,and could effectively guide the scientific storage of lingwu long jujube after harvest.The main research results were as follows:(1)Study on the classification discrimination of bruising lingwu long jujube.Hyperspectral imaging system was used to obtain images of complete Lingwu long jujube and bruising lingwu long jujube(Grade ?,?,?,? and ? bruised lingwu long jujubes).The original spectral data were calculated by ENVI software.PLS-DA and SVM classification discrimination models were established based on the original spectra.The results showed that the accuracy of the calibration set of the PLS-D A model was 72.70%and the accuracy of the prediction set was 86.67%,The original spectra were preprocessed by various algorithms and the models of PLS-DA,SVM and LDA were established.The results showed that the PLS-DA model was established after MSC preprocessing has the best effect.The MSC-PLS-DA model,the accuracy of calibration set was 76.19%and the accuracy of the prediction set was 86.67%.Multiple algorithms were used to select characteristic wavelengths of preprocessed spectral data,and PLS-DA,SVM and LDA models were established based on characteristic wavelengths.The results showed that the model SG-iVISSA-PLS-DA has the best effect,and the accuracy of calibration set was 70.79%,and the accuracy of prediction set was 90.48%.(2)Study on the detection of bruising time of lingwu long jujube.Hyperspectral imaging system was used to collect the spectral images of complete lingwu long jujube and 5 time periods after grade I bruising(2 h,4 h,8 h,12 h and 24 h after mechanical impact).Then the spectral data of the lingwu long jujube was obtained by extracting the region of interest.The original spectral data were preprocessed by different algorithms and the PLS-DA model was established.The results showed that the DT-PLS-DA model has the best efect.The accuracies of cailbration set and prediction set were 85.56%and 92.22%,respectively.The characteristic wavelengths of DT spectral data were extracted by different algorithms,and the classification and discrimination models of PLS-DA and SVM were established based on the characteristic wavelengths.The results showed that the DT-iRF-UVE-PLS-DA model has the best effect.The accuracies of calibration set and prediction set were 85.93%and 92.22%,respectively.And the number of modeling variables used in this model was 50,accounting for 40%of the total wavelengths.At the same time,it was found that the model could quickly identify and detect lingwu long jujubes 8 h after bruising,and it could be used for online rapid detection of early bruising of lingwu long jujubes.(3)Study on the detection of bruise grade of lingwu long jujube by Vis/NIR combined with original pectin.The images of complete jujubes and bruising jujubes(Grade ?,?,?,? and ? bruised)were collected by hyperspectral imaging system.The regions of interest were selected and the spectral average value was calculated.In PLSR modeling,the results showed that SNV-PLSR model has the best modeling effect by using different algorithms to preprocess spectral data.Rc,Rcv and Rp were 0.7449,0.4719 and 0.6937,RMSEC,RMSECV and RMSEP were 0.0524,0.0727 and 0.1023,respectively.Different algorithms were used to select the characteristic wavelength of SNV spectral data,and a model based on the characteristic wavelength was established.The results showed that the model with the characteristic wavelength selected by iRF and CARS algorithm has the best effect.The Rc,Rcv and Rp of SNV-iRF-CARS-PLSR were 0.7167,0.5871 and 0.7512,respectively,and the RMSEC,RMSECV and RMSEP were 0.0548,0.0646 and 0.0948,respectively.Among the PLS-DA classification and discrimination models,the accuracy of modeling after SG-1st pretreatment was the highest,and the accuracies of calibration set and prediction set were 68.65%and 86.05%,respectively.Among the models based on characteristic wavelengths,SG-1st-iRF-PLS-DA model has the highest classification accuracy,and the accuracy of calibration set and prediction set are 67.46%and 84.88%,respectively.