Research On The Estimation Model Of Main Biochemical Parameters Of Flue-cured Tobacco Leaves Based On Hyperspectral Technolog

Tobacco,as one of the most important leaf cash crops in China,the quality of tobacco leaves after roasting is the lifeblood of the entire tobacco industry,which directly affects the quality of cigarettes and the national economy.The tobacco roasting process has a significant impact on the quality of the tobacco leaves after roasting,so studying the changes of different biochemical parameters in the baking process of tobacco leaves is not only of great significance to the baking process but also has a guiding and judging effect on the quality of the tobacco leaves after roasting.The use of hyperspectral technology to quickly monitor the changes in biochemical parameters during tobacco roasting can guide the tobacco roasting process to obtain high-quality flue-cured tobacco,which is of great significance for tobacco scientific research,production and processing.In this paper,the lower part of the conventional flue-cured tobacco variety Yunyan 87in Pingba District,Anshun City,Guizhou Province is used as the research object.First,the hyperspectral image information acquisition system built by ourselves is used to capture the hyperspectral images of tobacco leaves of different maturity and different curing temperatures.The chlorophyll content and nicotine content of cured tobacco leaves were measured,the spectral characteristics of flue-cured tobacco at different maturity and different curing temperatures were analyzed,and the relationship between hyperspectral characteristics of tobacco leaves and chlorophyll content and nicotine content was discussed.The reflectance values corresponding to different wavelengths generated under the spectrum were obtained.Finally,the estimation models of chlorophyll content and nicotine content of cured tobacco leaves based on hyperspectral technology were established respectively.The main research contents and results are as follows:(1)The rapid prediction of chlorophyll a content of cured tobacco leaves was realized.First,five preprocessing methods were used to preprocess the spectral data of cured tobacco leaves,and then the competitive adaptive algorithm and the continuous projection algorithm were used to predict the chlorophyll a content of 176.The characteristic sensitive bands were screened in the dimensional spectral data,and finally,5 kinds of CARS-BP and 6 kinds of SPA-BP content prediction models were established for chlorophyll a by using the BP neural network model.By comparing the model output results with the detection index determination coefficient and root mean square error,it is concluded that the MSC-CARS-BP neural network model with the highest model accuracy is selected from five CARS-BP models for chlorophyll a,and its test set determination coefficient R~2=0.951,root mean square error RMSECV=1.492,at the same time,the SNV-SPA-BP neural network model with the highest model accuracy was selected from 6SPA-BP models,the test set determination coefficient R~2=0.949,root mean square error RMSECV=1.118.Finally,by comparing the test results of the MSC-CARS-BP model and the SNV-SPA-BP model on the prediction set of chlorophyll a content,it is shown that the fitting effect of the measured value and the predicted value of the prediction set of the SNV-SPA-BP model is better than that of the MSC-CARS-BP.Finally,the SNV-SPA-BP model was used to predict the chlorophyll a content of cured tobacco leaves.(2)The rapid prediction of chlorophyll b content in cured tobacco leaves was realized.First,five preprocessing methods were used to preprocess the spectral data of cured tobacco leaves,and then the competitive adaptive algorithm and the continuous projection algorithm were used respectively to predict the content of chlorophyll b from chlorophyll b.The characteristic sensitive bands were screened from the 176-dimensional spectral data obtained by the experiment.Finally,6 CARS-SVM and 5 SPA-SVM prediction models were established for chlorophyll b content by using the support vector machine model(SVM).The coefficient and root mean square error are obtained:The SG1-CARS-SVM prediction model with the highest accuracy was selected from the six CARS-SVM models for chlorophyll b,the test set determination coefficient R~2=0.864,the root mean square error RMSECV=0.565,and The SNV-SPA-SVM prediction model with the highest accuracy was selected from the 5 SPA-SVM models,the coefficient of determination of the test set was R~2=0.917,and the root mean square error RMSECV=0.478.Finally,by comparing the test results of the SG1-CARS-SVM model and the SNV-SPA-SVM model on the prediction set of chlorophyll b content,it is shown that the fitting effect of the measured value and the predicted value of the prediction set of the SNV-SPA-SVM model is better than that of the SG1-CARS-SVM model.SVM,and finally the SNV-SPA-SVM model was used to predict the chlorophyll b content of cured tobacco leaves.(3)The rapid prediction of nicotine content in cured tobacco leaves is realized.First,five preprocessing methods are used to preprocess the spectral data of cured tobacco leaves,and then a competitive adaptive algorithm is used to detect feature sensitivity from the176-dimensional spectral data of nicotine content.Finally,six CARS-SVM prediction models were established for the nicotine content of cured tobacco leaves by using the support vector machine model(SVM).Since the accuracy of the model did not meet expectations,the MOA algorithm was used to optimize the parameters c and g of the SVM model,and a MOA-SVM model with better coefficient of determination and higher prediction accuracy was established.Comparing the model output results with the detection index determination coefficient and root mean square error that we can get.1)The parameter optimization of the MOA algorithm significantly improves the prediction accuracy of the model.The determination coefficient R~2 of the six MOA-SVM prediction models has reached more than 0.9,which has a good ability to predict the nicotine content of cured tobacco leaves.2)Finally,after comparing and analyzing the six prediction models of nicotine content in cured tobacco leaves,the SG1-MOA-SVM model with the highest prediction accuracy was selected,and its test set determination coefficient R~2=0.958 and root mean square error RMSECV=0.141.The data of the prediction set was substituted into the SG1-MOA-SVM model for verification,and it was concluded that the observed and predicted values of the validation set had a good fitting effect.Finally,the SG1-MOA-SVM model was used to predict the nicotine content of cured tobacco leaves.