Using Nir Spectra In Non-invasive Measurements Of Apple

China has become the biggest producer country of apples in the world. Due to the lowcommercialization processed capacity of apple after harvest, exports of fresh apple was lessthan4%of the national total output; near-infrared spectroscopy technique has advantagessuch as no pre-treatment, no pollution, good repeatability, speediness, and non-destructive; soit is widely used in food, agriculture, medicine and other fields. Using near-infraredtechnology to detection quality and grade apples could promote the standardization of appleindustry and enhance the commercialization processed of apple in China.With the apples which were produced in Shaanxi province as materials, the research usednear-infrared spectral to detect the quality （chemical quality and texture quality）, varieties andbruised of apples, and the chemometrics methods were used to analyze and establish themodels of quality, varieties and bruised for apple. It can provide theoretical basis tonon-destructive detection techniques and online classification for apples.Specific results were as follows:（1）The near-infrared models which were used to distinguish unbruised from bruisedapples were built. After spectra of unbruised and bruised Gala apples were preprocessed,models were built using the principal component combined with Multilayer Perceptron neuralnetwork （MLP-ANN）, Radial Basis Function neural network （RBF-ANN）, Fisherdiscriminant analysis （Fisher-DA）, BP neural network （BP-ANN） and partial least squarediscrimination analysis （PLS-DA）. Results were as follows: after spectra were preprocessedby wavelet analysis, models of MLP-ANN, RBF-ANN, Fisher-DA and BP-ANN coincidedwith the first principal component of12000cm^-1～4000cm^-1were established, the correctdiscriminant rates were97.8%,87.2%,84.8%and95.7%respectively; correct discriminantrates of PLS-DA model were both100%for calibration set and testing set, so PLS-DA modelwas superior to others.After smoothing the spectra of unbruised and bruised of Starking apples, the MLP-ANN,RBF-ANN, Fisher-DA, BP-ANN, and PLS-DA were used to build models. The correctdiscriminant rates of models established by MLP-ANN, BP-ANN, and PLS-DA were better than RBF-ANN and Fisher-DA, the correct discriminant rates were97%、90%、95%、82.6%and85%separately.Results indicated that it was practical to build models using PLS-DA, MLP-ANN andBP-ANN together with principal component analysis to distinguish bruised from unbruisedapples; it would be study using the RBF-ANN and Fisher-DA to distinguish bruised fromunbruised apples in future.（2） The chemical quality of Fuji and Pick Lady apples were detected using FT-NIRduring storage time, and the models （TA、pH values、SSC and ratio of SSC to TA） based ontwo kinds of variety were built. The coefficients of correlation （R²c） were0.9019,0.9033,0.9021and0.9008separately; root mean square error of cross-validation （RMSECV） were0.366%、0.0384%、0.0896and23.9separately; the coefficient of prediction（R²p） were0.9391,0.9027,0.9012, and0.9038separately，root mean square errors of prediction （RMSEP）were0.386%、0.0359%、0.0834�'�25.1separately. Results indicated that the mathematical modelsof the chemical quality were acceptable and the robustness of the model was increased;applicability of the models was further extended, it provided the basis to establish a modelwhich was suitable for all species of apples.（3） The texture quality of Pink Lady apples was checked by FT-NIR, and texture qualitymodels were built using chemometrics. Firstly, partial least squares method was used toestablish the models of flesh firmness and crispness, R²cwere0.9029and0.9009, RMSECVwere63.7g and52.6g; R²pwere0.9071and0.9060, RMSEP were62.9g and69.9g. Then,using BP neural network combined with principal component analysis to establish models ofadhesiveness, chewiness, springiness, cohesiveness, resilience; R²cwere0.7040,0.6408,0.7369,0.7928and0.8256respectively, RMSECV were1.8868g·sec,16.2133,0.0282,0.0149and0.0075respectively; R²pwere0.6260,0.7486,0.6182,0.7853and0.7896respectively, RMSEP were0.617g·sec,23.0192,0.0836,0.0164and0.0062, respectively.（5） Texture of Pink Lady apples for different storage period was analysis using TPA. Theresults showed that: adhesiveness of flesh had negative correlation with hardness, crispness,cohesiveness and resilience, respectively. Flesh Cohesiveness was positively correlated withhardness, chewiness and springiness （R were0.583,0.886and0.928, respectively）;springiness had poor related to other parameters, while the resilience had better correlationswith hardness, cohesiveness and chewiness （R were0.6,0.928and0.824, respectively）.（5） The near-infrared models distinguished varieties of apples were built. MLP-ANN,RBF-ANN, Fisher-DA combined with principal component which were extracted from8000cm^-14500cm^-1and12000cm^-14000cm^-1respectively were used. Results indicated that:models of the whole rage12000cm^-14000cm^-1was superior to8000cm^-14500cm^-1, and models of MLP-ANN were better than RBF-ANN and Fisher-DA, but effects of the threekinds of models were poor, the correlate rates would be improved in future. PLS-DA wasfirstly applied to determine the varieties of apples, when choosing the eliminate constant topreprocess the map, the wavelength range of7502cm^-16098cm^-1, the principal componentwas5, discriminant model of PLS-DA was better, correct classification rates of the trainingset and validation set were both100%, so the PLS-DA model was superior to others models.