Prediction Of Color Change In Nitrite-cured Mutton Using Hyperspectral Imaging Technique During Storage Period

Mutton color is regarded as a visual measure of freshness and quality,which is an important sensory parameter.Nitrite is usually as food additive in the process of meat production,which can develop pink color,inhibit the growth of food spoilage bacteria and contribute to the color.The content of myoglobin in nitrite-cured meat is the key factor which affects the color change of cured meat.Aiming at the problem of rapid and accurate prediction of color of nitrite-cured meat in different storage periods and the unclear status of the change of spectral characteristics,this paper was to assess the potential feasibility of using visible/near-infrared(VIS/NIR)and near-infrared(NIR)hyperspectral imaging for rapid determination of the relative color(L,a,b)and myoglobin contents(DeoMb,MbO2 and MetMb)of nitrite-cured mutton during refrigerated storage.Developing full-wavelength partial least square regression(PLSR)models were first established for predicting mutton color using raw spectra and pre-processed spectra.To reduce the amount and dimension of data,the characteristic wavelengths were extracted by competitive adaptive reweighted sampling(CARS),interval Random Frog(IRF),variable combination population analysis(VCPA)and interval variable iterative space shrinkage approach(iVISSA)respectively.These data were analyzed using evolutionary computing methods,including PLSR and least-squares support vector machines(LSSVM).The primary conclusions are as follows:(1)The partial least squares discriminant analysis(PLSDA)discriminant model of nitrite addition in VIS/NIR and NIR spectra was established,and the modeling effects of VIS/NIR and NIR bands were analyzed and compared.The results showed that the accuracy of the PLSDA discriminant model based on NIR hyperspectral was higher than VIS/NIR hyperspectral.The results indicated that the accuracy of the PLSDA model calibration set was 88%,and the accuracy of the prediction set was 82%,which could well predict the addition of nitrite.(2)VIS/NIR hyperspectral imaging technology was used to predict the color of cured mutton.Different pretreatment methods,feature wavelength extraction algorithms and model development methods were analyzed and compared.The results showed that the simplified iVISSA-CARS-LSSVM model showed the best performance to determine L value with Rc of 0.9017 and Rp of 0.8562.The best model of a value was CARS-LSSVM,with an Rc of 0.9133 and Rp of 0.8983,and VCPA-LSSVM model that yielded peak performance for predicting b value(Rc=0.8380,Rp=0.7605).In addition,the new CARS-PLSR models of MetMb yielded good results,with an was better than predicting DeoMb(Rc=0.9133,Rp=0.8983),and VCPA-PLSR model showed the best performance to determine MbO2(Rc=0.9506*Rp=0.9707),while the best model of MetMb was new iVISSA-CARS-LSSVM model with an Rc of 0.9166 and Rp of 0.9099.(3)NIR hyperspectral imaging technology was used to predict the color of cured mutton.Different pretreatment methods,feature wavelength extraction algorithms and model development methods were analyzed and compared.The results showed that the new CARS-PLSR models of L value and b value yielded good results,with an Rc of 0.7628,and Rp of 0.8003 as well as Rc of 0.8512 and Rp of 0.8606,respectively.The simplified iVISSA-PLSR model showed the best performance to determine a value with Rc of 0.9033 and Rp of 0.8985.Furthermore,the new iVISSA-PLSR models of DeoMb and MetMb yielded better results,with an Rc of 0.8897 and Rp of 0.9017 as well as Rc of 0.9629 and Rp of 0.9661,respectively.