Study On Nondestructive Detection Of Fruit Quality Based On Hyperspectral Technology

With the development of agricultural technology and the improvement of people's living standards,the varieties of fruits on the market are becoming more and more abundant,and consumers' requirements for fruit quality are getting higher and higher.However,people cannot intuitively judge the variety and internal quality of fruits.Proper detection methods are needed to help people make judgments.Due to the drawbacks of the traditional methods in detecting fruit quality,such as destroying fruit integrity,time consuming and difficulty in batch processing,etc,it is necessary to develop a novel,non-destructive and efficient detection method to identify the quality of fruits.Hyperspectral technology can reflect the differences of physical structure and chemical composition of the sample under investigation by measuring its spectral changes.At the same time,it has the advantages of wide band range and high spectral resolution,no damage to sample,simple operation and low analysis cost,etc.The research objects of this study are jujube and cherry tomato,which are rich in nutrients and popular among consumers.This thesis focuses on two aspects: to identify the variety of jujubes and to predict the soluble solid content of cherry tomatoes by using hyperspectral technology.The main research contents and results are as follows.(1)Based on the hyperspectral technology,the varieties of Xinjiang jujubes,including Jinsi-jujube,Jun-jujube and Tan-jujube were identified.Firstly,the hyperspectral data of three varieties of jujubes were collected and the calibration set and prediction set were divided by sample set partitioning methods of random sampling(RS),kennard-stone(KS)and sample set partitioning based on joint X-Y distance(SPXY).The identification model of jujube varieties based on support vector machine(SVM)was established.The results showed that the prediction set based on the SPXY partition method was more representative.Secondly,the effects of various pre-processing methods on modeling were studied.The original spectra were preprocessed by using various data preprocessing methods,such as multiplicative scatter correction(MSC),standard normal variate transformation(SNV),first derivative(FD),Savitzky-Golay(SG)smoothing and so on.After the calibration set and the prediction set were divided by the SPXY sample partitioning method,the identification model of jujube varieties was established based on the pre-processed spectra using linear discriminant analysis(LDA),k-nearest neighbor(KNN),SVM and other algorithms.The results showed that the first-derivative method had the best effect in various pre-processing methods with the accuracy of 76.32% in LDA model and 100% in KNN and SVM models.we utilized the combination of the two pretreatment methods of SNV and FD in the LDA model,its accuracy was improved to 84.21%.Next,the characteristic bands of the full-band spectra were extracted by principal component analysis(PCA),successive projections algorithm(SPA)and competitive adaptive reweighted sampling(CARS)with the characteristic bands of 10,13,and 275,respectively.Then the jujube variety identification model was established based on the characteristic bands and compared with the model built by the full-band spectra.It was found that the accuracy of the model based on the full-band spectra is higher than that based on the characteristic bands.The CARS method can achieve better results in several characteristic extraction methods with the accuracy rates of 89.91%,98.68%,98.25% in the LDA model,KNN model and SVM model,respectively.Finally,taking the SVM model as an example,the modeling runtime was compared.The time required to build the SVM model based on the full-band spectra was 1.497 s,which was much longer than the time based on the characteristic wavelength modeling(running time for PCA,SPA and CARS were 0.026 s,0.032 s and 0.167 s,respectively).(2)The content of soluble solids in cherry tomato was detected based on hyperspectral technology.The hyperspectral data of the cherry tomato were collected and the corresponding soluble solid content were measured.The calibration set and prediction set were divided by the three sample partitioning methods of RS,KS and SPXY.According to the analysis of the results of partial least squares regression(PLSR)model,the SPXY partitioning method was finally used to divide the sample set.Secondly,the influence of various preprocessing methods on modeling was studied.The original spectra was preprocessed by using various data preprocessing methods,such as MSC,SNV,FD,SG smoothing and so on.The SPXY sample partitioning method was used to divide the sample set.The prediction model of soluble solid content of the fruit was established based on the pre-processed spectra using principal component regression(PCR),PLSR and support vector regression(SVR)and other algorithms.The results showed that the MSC had good effects in the PCR and PLSR models.The effect of SNV was the best in the SVR model.Finally,the characteristic bands were extracted from the full-band spectra by PCA,SPA and CARS,and the prediction model of the soluble solid content of the cherry tomato was established based on the characteristic wavelength.By comparing with the models built by the full-band spectra,the results showed that the prediction model established by the characteristic wavelength extracted by PCA had the best performance in the PCR model,while in the PLSR model and SVR model the prediction models based on the characteristic wavelength extracted by CARS performed the best.