Research On The Updating Method Of Agricultural Product Quality Detection Model

Non-destructive detection of agricultural product quality is of great significance for ensuring the quality and safety of agricultural products.In recent years,Optical detection methods including near-infrared spectroscopy techniques,hyperspectral image techniques,etc.have been widely used in non-destructive detection of agricultural product quality.The optical detection method for agricultural product quality is a soft sensing(or pattern recognition)problem,and its essence is to establish a mathematical model of spectral information and quality parameters of agricultural products.However,the agricultural products are easily affected by the producing area and the year,and the extracted spectral characteristic parameters also change.When the origin or year of test samples changes,the predictive accuracy of the original model will be reduced.It is very expensive to build a completely new model;how to update the existing model makes that the updated model is robust and does not change with the changes of origin and year,which is of great research value for improving the accuracy of non-destructive detection and reducing the actual cost.The purpose of this paper was to solve the problem of reducing the accuracy of existing models in detecting cross-year agricultural products by combining the methods of spectral analysis and model updating;for classification and regression models,different model updating strategies were proposed to improve the generalization ability of the model.The research contents of the paper were as follow:1.A method for updating the maize seed classification model based on similarity measurement was proposed.Maize seed classification belongs to pattern recognition,and the target value is a discrete variable.The paper drew on the idea of semi-supervised learning,the existing models were used to predict the unlabeled samples in other years and obtain the predictive labels.The reliability of the predictive labels was determined by calculating the similarity between the unlabeled samples and the training set samples.The few samples that the predictive labels of them were more reliable were selected and added to the training set in order to update the model each iteration.In the paper,a total of 3600 maize seeds harvest in 3 years(4 varieties)were used to verify the method of updating the model.The experimental results show that for the three different test sets,the average classification accuracy of the updated model were 8.0%,27.6%,and 7.2% higher than that of the non-updated model,respectively.It indicated that combining semi-supervised learning with similarity measures provided a possible technical approach to improve the robustness of classification models.2.A method for updating the maize seed classification model based on pre-labeling was proposed.The sample selection strategy of the method for updating the maize seed classification model based on similarity measurement mainly relied on the similarity between unlabeled set and training set samples,resulting in that the selected samples cannot adequately represent the test set,and the updated model had a limited performance improvement.In the paper,the samples with the same prediction labels by the initial model of or the updated model in the update process were considered as one class.After getting the class center,the distances between each sample in the class and the center of the class were calculated.The samples with small distances not only had reliable prediction labels and were more representative.The model was gradually updated by selecting few samples with smaller distances in the iterative way.The experimental results showed that for the three different test sets,the average classification accuracy of the updated model were 8.0%,27.6%,and 7.2% higher than that of the non-updated model,respectively.And the results were better than that of the method for updating the maize seed classification model based on similarity measurement,it indicated that the method for updating the maize seed classification model based on pre-labeling was more effective.3.A method for updating the regression model of apple SSC based on distance measurement and semi supervised learning was proposed.The prediction of apple soluble solid content was a soft sensing problem,and the target value was a continuous variable.The distances between samples cannot be accurately measured only by apple spectral information.Therefore,a distance measurement method based on spectral information and sugar value was built in the paper.Based on the distance measurement method couple with differences between different models,the few samples were selected to update the initial models in the iterative way and the test set was predicted by the updated model.The paper uses 2373 ‘Gold Delicious'(GD)apples harvest in 2 years to verify the method of updating the model.The experimental results showed that the relation coefficient between the predictive value and the actual value of the SSC of the test set based on the updated least squares support vector machine regression model(LSSVR)reaches 0.853,which was 27.5% higher than that of the non-updated model.