Nondestructive Detection Of Wheat Flour Quality Based On Hyperspectral Imaging Technology

Wheat is one of the most widely grown and important commodity food crops in the world.Flour(wheat flour)is a major component of many foods,and the quality of flour is generally described by its moisture,protein,starch,ash,gluten content,and pasting properties.By understanding these parameters,the processing suitability of flour can be assessed,and the correct flour can be selected for specific food formulations.Deoxynivalenol(DON),a common mycotoxin produced by the fungus Fusarium in wheat and its derivatives,seriously endangers human and animal life safety.Moreover,DON is difficult to completely destroy during food processing.Therefore,effective monitoring of the quality of high-contamination batches of flour before entering the food chain is the key to ensuring the safety of flour quality.Hyperspectral imaging(HSI)is a promising new nondestructive testing technology,which can replace the traditional testing methods and provide an efficient,high-precision,nondestructive and rapid testing method for flour.HSI combines the advantages of near-infrared spectral technology and digital imaging technology,which can provide information on the physical and chemical properties and their spatial distribution of the sample,and achieve non-destructive analysis of sample physical and chemical properties.This research aims to explore the feasibility of using HSI combined with traditional chemical methods to determine the protein,starch,and moisture content in flour,using a rapid viscometer to measure pasting properties,and predicting the possibility of deoxynivalenol contamination in flour.A series of complete methods for testing the quality of flour is established,and the corresponding theoretical basis is provided.The main contents of this study are as follows:1.Based on near-infrared hyperspectral imaging technology,predictive models for protein,starch,and moisture content in 77 different flours are constructed.Firstly,partial least squares regression(PLSR),principal component regression(PCR),support vector machine regression(SVMR),and multiple linear regression(MLR)models are constructed based on raw spectral data and nine different preprocessing algorithms.It is found that the PCR and MLR exhibit better predictive performance for protein and starch content prediction models without preprocessing,and the moisture content prediction model combined with PLSR shows better performance after SNV preprocessing.Then,four regression models are established based on the original spectral information,preprocessing,and eight feature wavelength extraction algorithms to determine the relationship between spectra and detection indices.Different feature wavelength extraction algorithms are used to optimize the models.The optimal predictive models for protein,starch,and moisture content are IVISSA-IRIV-PCR,IVISSA-IRIV-MLR,and SNV-MASS-PLSR,respectively,with prediction set determination coefficients(R_P~2)and root mean square errors(RMSEP)of 0.9859,1.1580 g/100g,0.9243,0.2068 g/100g,and 0.8646,2.1669 g/100g.The visualization of protein,starch,and moisture content is realized using optimized models.2.Detection of deoxynivalenol-contaminated flour using hyperspectral imaging technology was performed by preparing 240 samples of contaminated flour with different concentration gradients.The samples were divided into 180 calibration sets and 60prediction sets using the KS algorithm.Linear discriminant analysis(LDA),principal component analysis-linear discriminant analysis(PCA-LDA),partial least squares-discriminant analysis(PLS-DA),soft independent modeling of class analogy(SIMCA),and K-nearest neighbor(KNN)models were constructed based on raw spectral data and five different preprocessing algorithms.The results showed that modeling analysis based on raw spectral data exhibited better predictive performance.Discriminant effects of different classification models were evaluated based on spectral features,texture features,and spectral-texture fusion features.The BOSS-SIMCA model based on spectral features was the optimal model,with a prediction set recognition accuracy of 92%;the optimal model based on texture features was LDA,with a prediction set recognition accuracy of 77%;and the optimal model based on spectral and texture features fusion was LDA,with a prediction set discrimination accuracy of 92%.3.Hyperspectral imaging technology was used to detect the gelatinization characteristics of flour using peak viscosity as a characteristic indicator.Three models,namely,principal component regression(PCR),partial least squares regression(PLSR),and support vector machine regression(SVMR),were constructed based on raw spectral data and nine different preprocessing methods.SNV preprocessing method was found to have better predictive effect.Considering the calculation speed,robustness,and stability of the model,the optimal prediction model for flour peak viscosity was SNV-IVISSA-IRIV-SVMR,with R_P~2=0.8955 and RMSEP=115.4859 c P.This model can effectively predict the gelatinization characteristics of flour,providing theoretical support for flour processing and manufacturing.The results showed that near-infrared hyperspectral imaging technology can effectively and accurately predict the protein,starch,moisture content,and gelatinization characteristics of flour.Moreover,it is feasible to detect the level of deoxynivalenol contamination in flour by combining the texture features extracted from the gray-level co-occurrence matrix(GLCM)with spectral features.This study provides a new set of methods for flour quality testing and a reference for non-destructive and rapid detection of grains and their chemical compositions.