Nondestructive Testing Of Apple Quality Based On Visible/near Infrared Spectroscopy

As the income level and health awareness of Chinese residents continue to rise,people are paying more attention to the nutritional content of foods such as vegetables and fruits.High-quality apples have shown stronger competitiveness in the market.At present,the demand for apples is also increasing,but traditional manual sorting methods are laborintensive,time-consuming,technically demanding,and destructive.In recent years,visible/near-infrared spectroscopy analysis technology has been widely used in nondestructive testing of agricultural products due to its low cost,fast speed,non-destructive,accurate,and sustainable characteristics.However,problems such as single detection index,low accuracy,and poor system robustness still exist.In order to solve these problems,this paper takes Yantai Red Fuji apples as experimental samples,uses sugar content,acidity,and hardness as inspection indices,and conducts research on visible/near-infrared spectroscopy analysis technology from three aspects: the construction of spectral acquisition platform,spectral data processing,and apple quality prediction model optimization.The main research work is as follows:1)Construction of spectral acquisition platform and data processing: This paper independently constructs an apple spectral acquisition platform,studies the spectral characteristic bands that can reflect the sugar content,acidity,and hardness indices of apples,and studies the internal principles of apple spectral detection.Using the visible/near-infrared diffuse reflection mode,1563 sampling points from the spectral range of 400.42-1100.11 nm were retained.Ten spectral preprocessing algorithms were compared in the preprocessing stage,and the results were visualized.Then,the KS sample selection method was adopted,and the principles of the Cars and Uve feature wavelength selection algorithms were compared and analyzed.Finally,the optimal spectral processing combination method was selected based on the above different preprocessing stages to lay the foundation for subsequent modeling.2)Construction of apple sugar content,acidity,and hardness prediction models: This paper establishes quantitative analysis models for apple sugar content,acidity,and hardness using three quantitative analysis methods(PLS,SVR,RF)respectively.The experimental results show that the D1+KS+Cars+PLS model has good fitting effects on the three quality characteristics of apples,which indicates that the D1+Cars method is more suitable for PLS modeling,showing the universality of this method and better extracting the feature information of samples for apple quality prediction.The determination coefficient of sugar content reaches 0.93,the root mean square error of prediction is 0.29,and the mean absolute error is 0.23;the determination coefficient,root mean square error of prediction,and mean absolute error MAE of acidity and hardness models are 0.96,0.04,0.04,and 0.97,0.37,0.28,respectively.The D1+KS+Cars+PLS model of acidity and hardness can better extract the feature information of samples,facilitating the prediction of apple acidity and hardness with strong accuracy and robustness.3)Optimization model of apple sugar content prediction based on one-dimensional convolutional neural network: The performance of the apple sugar content prediction model based on traditional machine learning still needs to be improved.Therefore,this chapter studies the basic principles of convolutional neural networks and constructs an apple sugar content prediction model based on a one-dimensional convolutional neural network according to the collected apple spectral data features.The proposed method avoids the wavelength selection process in traditional methods,reducing the loss of information to a certain extent.The experimental part verifies the performance of this sugar content model and compares the performance of the sugar content prediction model proposed under different preprocessing methods.Finally,D1 is selected as the optimal preprocessing method.Compared with traditional machine learning regression models,the proposed sugar content prediction model has been significantly improved in all three evaluation indicators,and the determination coefficient,root mean square error of prediction,and mean absolute error have been increased from 0.93,0.29,0.23 to 0.97,0.23,0.19,respectively,with greatly improved accuracy and robustness.