Influence Of Sample Un-related Factors On Real Detection Of Pear SSC Using Vis/NIR And Model Optimization

With the development of social economy and people living ever better,the demand of fruit grows and more attention is paid to fruit quality.On the other hand,China,as the biggest fruit producer in the world,while lack of competitiveness in the international fruit market.The chief reason is that relative low commercialization treatment ability for domestic fruit leads to uneven quality of fruit even from the same batch.Quality detection and grading of postharvest fruit is an integral part of commoditization processing,which is also an effective way to achieve higher price with better quality.Near-infrared(NIR)spectroscopy with the advantages of rapid,nondestructive and can realize on-line analyzing,has been widely used in agriculture.However,in actual prediction,the changes in environmental conditions,instrument responses and sample related factors often result in the lower prediction accuracy or even render the original model invalid by causing unacceptable deviation in the results.Therefore,this study,based on previous work of our team,selected pear,freshly squeezed juice and fructose solution as the main research objects,using visible/NIR(Vis/NIR)spectroscopy,established different soluble solids content(SSC)prediction models concerned with sample un-related factors.According to the problems faced in actual prediction,the influence of different spectrometers,fruit bags,reference spectrum acquisition modes and informative variable s election methods on real detection of SSC in pear using Vis/NIR were analyzed.The main contents,as well as results and conclusions are summarized as follows:(1)The adaptivity of SSC prediction model of Crown pear was analyzed between different types of fiber optic spectrometer.And model transfer was established by a new combination of mean spectra subtraction correction and direct standardization(MSSC-DS).Firstly,two different fiber optic spectrometers(Ocean Optics,Inc.,USA,model QE65000 and QE65Pro)were applied to collect the transmittance spectra of the same batch of Crown pears.Combining with SSC values,the partial least squares regression(PLSR)models were built,repectively.But both models failed to predict spectra collected by the other spectrometer,which indicated that model built on one exact instrument was usually limited to it own use.Then,QE65000 was selected as the master instrument(QE65Pro was selected as the salve instrument)and four traditional model transfer algorithms of DS,piecewise direct standardization(PDS),slope/bias(S/B)and MSSC were compared.On this basis,MSSC-DS,MSSC-PDS,MSSC-S/B were also developed and compared.SSC prediction model of Crown pear was transferred from QE65000 to QE65Pro through the above algorithms.The results of DS and MSSC-DS were relatively superior with root mean square error of prediction(RMSEP)reduced from 8.482 °Brix to 0.473 °Brix and 0.453 °Brix,respectively,satisfying the requirement of practical production.However,the best predition result on salve instrument after model transfer was still inferior to that predicted using the model of the salve instrument(RMSEP=0.381 °Brix).Therefore,in practice,it needs to balance between cost and grading precision and choose appropriate way of modeling.(2)The influence of two kinds of fruit bag(single layer white bag,SWB and double layer yellow bag,DYB)on Vis/NIR spectra of Cuiguan pear were investigated.In order to reduce the mechanical damage of samples during transportation and grading,users want this process be realized with sample bagged.The single layer white bag used in this study has good transmittance,so the bagged fruit spectra have little differences from non-bagged fruit spectra in absorption,with all absorption peaks reserved.The result of quantitative analysis also showed that the SSC prediction model of Cuiguan pear bagged with SWB(Cross Validation Correlation Coefficient,rcv = 0.7894,Root Mean Square Error of Cross Validation,RMSECV = 0.376 °Brix)was similar to that of non-bagged Cuiguan pear(rcv,0.8282,RMSECV = 0.343 °Brix).And 83.8 percent of the Cuiguan pear bagged with SWB has the prediction error within plus or minus 0.5 °Brix,which could be accepted in practical production.In future,further study that looking for or developing appropriate fruit bag which not only promote fruit quality during its development but also meet the requirement of applying Vis/NIR spectroscopy directly to evaluate its internal quality needs to be carried out.(3)According to previous study of our team,the influence of reference spectrum automatic collection with different frequency on model precision was investigated.Firstly,it was further verified that reference spectrum automatically collected was more stable than manually collected.Besides,in terms of the drift of light source with time going on,the frequency of automatic reference spectrum collection was studied.Reference spectrum measured after every 3,10,20 samples with Teflon ball automatic placement and only one reference spectrum measured at the beginning with Teflon ball automatic placement(marked as A3,A10,A20 and AO in turn)were compared.Model performance from good to bad in turn was obtained under A3,A10,A20 and AO mode.It indicated that as the frequency increased,the model performance improved.But higher frequency will decrease the efficiency of postharvest fruit detecting and grading.Therefore,it needs to select appropriate reference spectrum collection frequency according to actual requirement.(4)Based on informative variable selection,as well as the characteristic wavelengths of fructose solution,SSC prediction models of integrate fruit were built.At first,the predictive ability of SSC models that built after informative variable selection among different years was analyzed.Then,S/B was carried out for further modification.Models that obtained by PLSR coupled with different variable selection algorithms based on Crown pear of 2014 yielded bad results when used to predict samples of 2015 and 2016.But after S/B modification,RMSEP reduced considerably.The best results for Crown pear of 2015 and 2016 were obtained by competitive adaptive reweighted sampling PLSR after S/B modification(CARS-PLSR-S/B)and monte-carlo uninformative variables elimination PLSR after S/B modification(MC-UVE-PLSR-S/B)with RMSEP being 0.575 °Brix and 0.609 °Brix,respectively.Then,full-range PLSR(Full-PLSR),stepwise multi linear regression(SMLR)and interval PLSR(iPLSR)were applied to build the SSC prediction model of freshly squeezed juice and fructose solution.Due to the effect of scattering of particles in freshly squeezed juice,the relationship between absorbance and concentration deviated from the Beer-Lambert Law,which caused bad results of the model.While in fructose solution,fructose as the single solute,yielded ideal absorbance and thus brought desirable model performance.Therefore,SMLR was used to select informative wavelengths of fructose solution spectra in the range of 550-920 nm.Using the selected wavelengths,SSC prediction model of integrate Crown pear was built,whose performance was similar to that of Full-PLSR model obtained by pear spectra in the same wavelength range.The results indicated that it was feasible to use the informative wavelengths of fructose solution to optimize the S SC prediction model of pear.