| With the development of our economy and the improvement of people's living,people's purchasing power for fruits has greatly increased.As a large fruit producer,China lacks competitiveness in the international fruit market.The main reason is that the low level of fruit commercialization leads to the uneven quality of fruit.The detection and classification of fruit internal quality is an important part of fruit commercialization,and Vis/NIR spectroscopy has been widely used in the measurement of internal quality parameters of fruit in recent years due to the advantages of fast,nondestructive and easy to realize.However,due to the influence of origin,cultivar and harvest year,the physical and chemical properties of the fruit have changed.This kind of biological variability has a great impact on the modeling and analysis of soluble solids content(SSC)in fruit by Vis/NIR spectroscopy,which makes the model based on data of a certain origin,cultivar or year unable to predict SSC of other origins,cultivars or harvest years accurately.Therefore,based on the intelligent online detection equipment for fruit internal quality of our team,apple was taken as the research object for the general model of online detection of SSC by Vis/NIR diffuse transmission spectroscopy.The main research contents and conclusions were summarized as follows:(1)The influence of origin on the online detection model of apple's SSC was studied.Red Fuji apples that separately from Qixia of Shandong Province,Luochuan of Shaanxi Province and Huining of Gansu Province were used as the research objects.First of all,the calibration model for each of the three origins was established by partial least squares regression(PLSR)respectively,which resulted in the residual prediction deviation(RPD)of3.02,2.62 and 2.29,respectively.It suggested that the internal prediction results of the local origin model were great.Then,taking Qixia origin as an example,10,20,30 and 40 representative samples were selected from Luochuan and Huining samples by K-S algorithm and added to Qixia model.The root mean square error of prediction(RMSEP)of the updated model for Luochuan and Huining samples decreased from 0.82°Brix and 1.24°Brix to0.69°Brix and 0.86°Brix,respectively.It indicated that adding some samples from other origins could improve the prediction ability of the local origin model to other origins samples to a certain extent.Finally,the global origin model was established by mixing the calibration sets of three origins,which resulted in the RMSEP of 0.62°Brix,0.64°Brix and 0.65°Brix,respectively.It suggested that the prediction accuracy of the global origin model was higher than that of the updated local model.Based on the global model,the modeling wavelengths were optimized by uninformative variable elimination method(UVE),which resulted in thatthe RMSEP of the model for Qixia and Luochuan samples except Huining ones decreased to0.50°Brix and 0.63°Brix,respectively.The number of modeling variables and latent variables reduced from 400 and 12 to 58 and 8,respectively.60 test samples were used to assess the performance of the above model,which resulted in the RPD of 2.33.Thus,the impact of origin should be taken into account for the on-line detection of apple SSC.The prediction range of the SSC prediction model among different origins could be expanded by establishing a global general model that could be simplified by suitable wavelength selection methods,which could improve the generality and robustness of the model.(2)The influence of cultivar on the online detection model of apple's SSC was studied.The research objects were Candy Heart apples,Red Fuji apples and Crystal Fuji apples from Shaanxi Province.First of all,the calibration model for each of the three cultivars was established by PLSR respectively,which resulted in the RPD of 2.98,2.80 and 2.10,respectively.It suggested that the internal prediction results of the local cultivar model were great.Then,taking Candy Heart cultivar as an example,10,20,30 and 40 representative samples were selected from Red Fuji and Crystal Fuji samples by K-S algorithm and added to Candy Heart model.The RMSEP of the updated model for Red Fuji and Crystal Fuji samples decreased from 1.25°Brix and 2.73°Brix to 0.98°Brix and 0.80°Brix,respectively.It indicated that adding some samples from other cultivars could improve the prediction ability of the local cultivar model to other cultivars samples to a certain extent.Finally,the global cultivar model was established by mixing the calibration sets of three cultivars,which resulted in the RMSEP of 0.59°Brix,0.64°Brix and 0.63°Brix,respectively.It suggested that the prediction accuracy of the global cultivar model was higher than that of the updated local model.Based on the global model,the modeling wavelengths were optimized by UVE,which resulted in that the RMSEP of the model for Candy Heart and Crystal Fuji samples except Red Fuji ones decreased to 0.54°Brix and 0.61°Brix,respectively.The number of modeling variables and latent variables reduced from 398 and 12 to 144 and 10,respectively.68 test samples were used to assess the performance of the above model,which resulted in the RPD of 1.70.Thus,the impact of cultivar should be taken into account for the on-line detection of SSC.The prediction range of the SSC prediction model among different cultivars could be expanded by establishing a global general model that could be simplified by suitable wavelength selection methods,which could improve the generality and robustness of the model.(3)The maintenance method of the online detection model of apple's SSC in different years was studied.The research objects were Candy Heart apples from Luochuan in2017~2019 as the research object.First of all,the initial calibration model was established for the samples in 2017,which resulted in the RMSEP for the samples in 2017,2018 and 2019 of0.64°Brix,0.94°Brix and 1.38°Brix,respectively.It indicated that the prediction performance of the online SSC prediction model for the new year samples was greatly reduced.Secondly,5,10 and 20 representative samples were selected from the calibration sets in 2018 and 2019 by K-S algorithm and added to the basic model,which resulted in that the RMSEP of the updated model for 2018 and 2019 samples reduced to 0.70°Brix and 0.92°Brix,respectively.Itsuggested that the initial model was maintained by adding some new samples from the harvest year.Finally,5 to 40 samples were randomly selected from the calibration sets in 2018 and2019 as standard sample sets,and the model was maintained by the slope/Bias algorithm(S/B).10 random selections were made for each standard sample number,and the average RMSEP calculated from 10 selections was taken as the final result corresponding to the number of standard samples.For the prediction samples in 2018,the maintenance performance was better when 5 standard samples were selected,which resulted in lower RMSEP of 0.65°Brix.For the prediction samples in 2019,the maintenance performance was better when 10 standard samples were selected,which resulted in lower RMSEP of 0.61°Brix.Thus,S/B method could maintain the online SSC prediction model quickly and conveniently.For the actual production,it could not only save a lot of cost,but also make the process simple to operate without specialized software or professional staff,which could be more practical for applications with low production requirements. |
PDF Full Text Request