Research Of Nondestructive Detection For Assessing Internal Quality Attributes Of Pear Based On Spectroscopy Technology And Model Optimization

With the improvement of people’s living standards,the consumption of fruits is increasing year by year in my country,and the requirements for fruit quality are getting higher and higher.Among them,soluble solids contents and hardness are the two internal qualities that people are most concerned about.Besides,our country is a large fruit producer in the world,but its export volume is very small.It also lacks competitiveness in the international market.The main reason is that our country’s fruit post-harvest testing methods are backward,and the quality of the produced fruits is difficult to meet the national import as well as export standards.Therefore,enhancing the non-destructive testing of fruit quality can not only meet the requirements `of consumers for fruit quality,but also promote the development of fruit industry in China.In order to solve the above-mentioned problems in China’s fruit industry,this paper takes "Yuluxiang" pears,"Xuehua" pears and crisp pears produced by Taigu District of Shanxi Province as the research objects,and uses hyperspectral imaging technology as well as visible/near infrared spectroscopy to detect soluble solids contents and hardness.It provides the algorithm support and theoretical foundation for the fast and accurate detection of the pear internal quality.The main contents and research conclusions of the paper are as follows:(1)The distribution of SSC in different parts of “Yuluxiang” pear was analyzed.The back propagation neural network and partial least square regression model were constructed based on hyperspectral data.Genetic algorithm was used to optimize the back propagation neural network.The impact of sample size of a single location and a mixed location on model performance was compared and analyzed.The SSC and hyper-spectral data of the top,equator and bottom of "Yuluxiang" pear was collected,and the abnormal samples were eliminated by Monte Carlo partial least square method.The full-band hyper-spectral data was de-noised by median filter,Savitzky-Golay,baseline correction,standard normal variate,de-trending,multiplicative scatter correction,as well as other preprocessing methods.Building a PLSR model,the best preprocessing method was determined according to the performance of the model.The pre-processed hyper-spectral data were used as independent variables and the SSC values as dependent variables.The BP as well as GA-BP detection models were established and compared with the traditional PLSR model.Finally,different locations of samples were mixed to train the GA-BP neural network with different numbers of samples.According to the results,there are some differences in the spectral data and SSC of the top,equator and bottom of "Yuluxiang" pear,and the SSC model of the equatorial position is better.After MF pretreatment,the denoising effect of spectral data is the best.At the same training sample,the GA-BP model has the best performance.The coefficient of determination for calibration set was 0.9788,and the root mean square error of calibration set was 0.1865;The coefficient of determination for prediction set was 0.8565,and the root mean square error of prediction set was 0.4311.The residual predictive deviation was 2.445.Building the GA-BP model with 300 samples,Rc2 was 0.9986 and RMSEC was 0.2185;Rp2was 0.9805 as well as RMSEP was 0.2031,and the ratio of the coefficient of determination for prediction set to the coefficient of determination for calibration set was 0.9819.The over-fitting phenomenon of the model was significantly improved,and the predictive accuracy of the model was further enhanced.A theoretical basis for the hyper-spectral technology nondestructive testing of "Yuluxiang" pear SSC is provided in this research.(2)Based on high spectroscopic imaging technology,the influence of single and mixed pears on the stability of the SSC detection model was studied.The SSC and hyper-spectral data from the equatorial parts of "Yuluxiang" Pear,"Xuehua" Pear,and crisp Pear was collected.The full-band hyper-spectral data was de-noised by moving average,median filter,Savitzky-Golay,baseline correction,standard normal variate,de-trending,multiplicative scatter correction and MSC-SNV,as well as other preprocessing methods.The PLSR detection model is established based on the raw and preprocessed spectral data.The best preprocessing method is selected according to the model performance.Finally,the partial least squares regression and least squares support vector regression detection models were established using a single variety and a mixed sample of 2 and 3 varieties,respectively.The results show that the spectral data pretreated by MSC-SNV method is best.The performance of the PLSR model established by a single variety sample is better.With the mixture of varieties,the prediction performance of the PLSR model was gradually decreased,and the prediction performance of the LSSVR model was gradually improved.When mixing the 3 varieties samples,the model with highest prediction accuracy was the LSSVR.The coefficient of determination for prediction set was 0.8706;the Root Mean Square Error of prediction set was 0.4960;the residual predictive deviation was 2.2333.By mixing the samples,on the one hand,the SSC detection model of fruit was optimized;on the other hand,the generalization ability of the mode based on varieties is enhanced,so that a better prediction effect on the SSC of a single variety of pears was obtained.This research provides a reference for internal quality universal model established by different varieties of pears.(3)Based on visible/near-infrared spectroscopy technology,a one dimensional convolutional neural network model was established to detect the SSC and hardness of different varieties of pears.And the Bayesian optimization algorithm was used to optimize the hyper-parameters of the model.Taking "Yuluxiang" Pear,"Xuehua" Pear,and crisp Pear as the research objects,the spectral information of the equator of the pears was collected.The detection indicators includes SSC and hardness and the spectrum data is preprocessed.Competitive adaptive reweighted sampling is used to extract 20,30,and 40 characteristic wavelengths respectively.The PLSR detection model of SSC and hardness was established for the full wavelength as well as characteristic wavelength spectral data.The performance of the model is evaluated by the relevant indexes.And then the optimal number of wavelengths was confirmed.According to the one-dimensional characteristics of near infrared spectrum data,a new 8-layer one-dimensional convolutional neural network model is designed in this paper,which can reduce the preprocessing requirements of spectral data.Besides,the prediction accuracy of the model is improved by selecting the hyper-parameters of the model with Bayesian optimization algorithm.However,because of the long training time of convolution neural network model,it is necessary to use GPU to accelerate the training of the model.The spectral data corresponding to the characteristic band were used as input variables,the SSC and hardness as the output variables.The PLSR and BO-1D CNN detection models on SSC and hardness of different varieties of pear were established.The optimal model was determined by comparing the coefficient of determination for prediction set,the root mean square error and the residual prediction deviation.Finally,the effects of different numbers of samples on the performance of the model were compared.The results show that the SSC and hardness of pears can be detected by using visible/near infrared spectroscopy data.The model performance based on the original spectral data was better than the pre-processed spectral data,and and 40 was the optimal number of characteristic wavelengths.When a small number of samples are participating in the modeling,the prediction performance of the BO-1D CNN model is slightly higher than that of PLSR.After increasing the number of samples,the prediction performance of the BO-1D CNN model will be much higher than that of the PLSR.In the hardness testing,the Rp2 of the best BO-1D CNN detection model is 0.9868.The RMSEP is 0.0559,and the RPD is 8.1719;When testing the SSC,the Rp2 of the best BO-1D CNN detection model is 0.9909.The RMSEP is 0.1757,and the RPD is 9.6939.Therefore,this research provides algorithmic support for the non-destructive testing SSC and hardness of different varieties of pear which is based on convolutional neural network as well as near-infrared spectroscopy data.To sum up,the non-destructive testing on Pear SSC and hardness was based on the spectral information in this paper,which is to enhance the detection technology of the internal quality of fruits in our country.By combining hyper-spectral and visible / near infrared spectroscopy data with neural network,SSC and hardness detection models of pear are established.The model has a fast convergence speed and good stability.It provides theoretical support for the detection of other internal qualities of pears,such as titratable acidity and vitamin C content,etc.Further,it can promote the development of the pear industry in China.