Nondestructive Measurement Of Properties Of Chestnut Based On Near Infrared Spectroscopy

Chestnut (Castanea mollisima) is one of the most important fruit in China which has the greatest area of chestnut cultivation and products more than 800,000 tons fruit every year. However, due to the highly variable properties and the less optimized method of store and processing, the rate of post harvest lost is more than 30% and the profits is less than it should be. Therefore, to discriminate the moldy fruit rapidly and to measure the content of main components accurately play the important roles in the chestnut industry.Near Infrared Spectroscopy (NIR) is a spectroscopic method, which has employed to perform the quality and quantity analysis of agricultural product, with the advantages of fast and easy to operate, nondestructive measurement and minimized preparation of sample. Howevery, the application of NIR in the target with ligneous peel is rare reported. In this study, the NIR spectroscopy technology was applied in assessing the properties of chestnuts nondestructively, including collection of moldy chestnut samples, acquirement of the NIR spectra, discrimination of the inner mildew and correlating of the spectra feature with the main components content. In the qualitative analysis, both the unsupervised and the supervised pattern recognition methods, including HCA (Hierarchical Cluster Analysis) and PCA-HCA (Principle Component Analysis-Hierarchical Cluster Analysis), PLSDA (Partial Least Squares Discriminant Analysis), SVMDA (Support Vector Machine Discriminant Analysis) and SVM-PLS (Support Vector Machine-Partial Least Squares), were executed to detect the inner mildew of the intact chestnut, combined with the common preprocessing methods, including derivation, MSC (multiplication scatter correction) and SNV (standard normalized variate). In the quantitative analysis, the PCR (Principle Component Regression) and PLSR (Partial Least Squares Regression) were used to establish the correlativity between the spectra of intact or peeled sample and the reference value of the moisture, sugar, starch and protein content. The effects to models of different subset partition algorithms, named Concentration Gradient, Kennard-stone Algorithm and Sample Set Partitioning based on Jiont X-Y Distances Algorithm were compared, as well as diverse preprocessing methods, including derivation, MSC and SNV. Additionally, OSC (Orthogonal Signal Correction), iPLS (Interval Partial Least Squares), VIP (variable importance on the projection), GA (genetic algorithm) UVE (uninformative variable elimination) and SPA (successive projection algorithm) were applied to optimize and simplify the predictive models. The main conclusions are listed in following:1. The feasibility of discriminating the interior mildew in intact chestnuts using NIR spectroscopy was firstly proved by experiment. The detection of moldy extent was exploratory studied. The method for discriminating the moldy chestnut was optimized by comparing the effects of different preprocessing methods, definitions of distance and pattern recognition algorithms in analyzing the spectra of moldy and sound sample. The results shown that the reflectance NIR spectrum of intact chestnut sample contained the information of whether mildew occurred interiorly even the exterior appearance was normal and the NIR spectroscopy can be used in detection of the inner mildew in intact chestnuts. The supervised pattern recognition achieved better results than the unsupervised pattern recognition. Additinally, the accuracy rates of distinguishing mildew were higher than that of detecting the mildew extent. When differentiating the spectra into mildew sample and sound sample, the PLSDA model established from spectra preprocessed by MSC obtained 99% as the accuracy rate and 0.94 as the sensibility and the specificity in cross validation; the best result of the PCA-HCA models was 89%. When classifying the spectra into slight mildew sample, severe mildew sample and sound sample, the PLSDA model established from spectra preprocessed by MSC had the highest accuracy rate at 97%, with 0.91 as the sensibility and 0.92 as the specificity in cross validation. While the model calculated by PCA-HCA combined with the Ward's Algorithm got the accuracy rate at 76%.2. This work validated the feasibility of applying NIR spectroscopy in quantitative analysis of interior properties of intact and peeled chestnuts nondestructively. By comparing the function of different sample subset partitioning methods, preprocessing methods and regression algorithms in establishment of the correlation ship between the reference value of the target components and the feature of spectra acquired from intact and peeled sample, this work figured out the best subset partitioning method, preprocessing method and regression algorithm for the quantity analysis of moisture, sugar starch and protein in intact and peeled chestnuts baed on NIR spectroscopy. The results shown that, firstly, both the PCA models and PLS models based on the subsets chosen by SPXY algorithm got better results than those based on the subsets from other algorithm. Secondly, the PLSR performed better than PCR in analyzing the spectra of both peeled and intact samples. Thirdly, the models based on spectra of peeled sample performed better than the model of intact spectra of sample established by same preprocessing and regression method. Furthermore, the derivation preprocessing method was more effective than the other methods in purifying the information of the target component. Additionaly, the precisions of the models established from the spectra of peeled sample were positive linear correlated to the concentration of the target component, while the precision of models calculated from the spectra of intact sample did not appear in this way.3. For the analysis of the spectra of peeled samples, the PIS models based on 1st pretreated spectra acchived the best results for each component. The model for moisture content obtained 0.9359 and 0.8437 as the correlation coefficient of calibration subset and validation subset separately, with 1.44 as the RMSEC and 1.83 as the RMSEP. The model for sugar content had 0.9072 and 0.8649 as the correlation coefficient of calibration subset and validation subset separately, with 0.76 as the RMSEC and 0.74 as the RMSEP. The one for starch content achieved 0.9168?0.8376?1.07 and 1.26 as these parameters separately, while the model for protein content got 0.9044?0.8029?0.29 and 0.40 separately. For the analysis of the spectra of intact samples, the PIS models based on 1st pretreated spectra acchived the best results for moisure, sugar and starch content, while the PIS models based on 2nd pretreated spectra performed outstanding for protein content. In these models, the correlation coefficient of calibration and validation subset, the RMSEC and the RMSEV of the model for moisture content were 0.8270?0.7655?2.27 and 2.35 separately. These parameters of the model for sugar content were 0.8379?0.5891?1.116 and 1.320, while the model for starch content had 0.9137?0.8432?1.09 and 1.24 separately. The model for protein content got 0.8748 and 0.7324 as the correlation coefficient of calibration and validation subset, with 0.35 as the RMSEC and 0.38 as RMSEV.4. The predictive models were optimized by different algorithms namely OSC, iPLS, VIP, GA, UVE and SPA. The performance of model for moisture content was improved most prominently. The precision of model for protein content could be enhanced by OSC,VIP and SPA. The model for sugar content can achieve better parameters combined with OSC, while it only can be simplified by the variable selection methods. The model for starch content can be optimized by OSC, but the variable selection methods brown down the predictive function.5. For the effectiveness of variable selection methods, iPLS can establish a model using 50% variable to perform similar as the model based on all variable. VIP can keep the performances of the models when removed more than 65% variable. However, its function on enhancing the predictive capable was limited, only the model for protein content was improved. GA selected less than 30% of variable as the informative variable, but the models performed worse than the model based on all variable. Although the numbers of informative variable selected by GA in each calculation were significant different, the variable with high selected frequency were kept consistently. More than 90% variable were labeled as non-informative variable and removed by UVE, while SPA selected less than 3% variable as the informative variable. The model for moisture content based on 29 variable selected by SPA performed best in all the models for moisture. However, SPA did not optimize the model for sugar and starch content distinctly.