A Global Calibration Model For Predicting Fatty Acids In Red Meat Based On Visible Near Infrared Hyperspectral Techniques

This study aims to establish a global prediction model for the fatty acid content of red meat based on hyperspectral imaging combined with deep learning methods.In the experiment,we used instruments and equipment to conduct statistical analysis of the fatty acid content of different types of red meat and collected and extracted spectral information.Four methods,namely,support vector regression(SVR),random forest(RF),convolutional neural network(CNN),and fully convolutional network(FCN),were used for modeling in this study.Data augmentation techniques were utilized to enhance the original data,and a fatty acid content model for red meat based on data augmentation was established.We explored the quality analysis of enhanced data,mixed data augmentation,and data extension experiments and analyzed the results.This model provides strong support for non-destructive detection of various indicators of red meat and prediction of meat quality,and has certain reference value for future related research.The specific research content and results are presented belowThis study aims to establish a global prediction model for red meat fatty acid composition based on the combination of hyperspectral imaging and deep learning methods.In the experiment,we used experimental instruments and equipment to perform statistical analysis on the fatty acid content of different types of red meat,and collected and extracted spectral information.Four methods,namely Support Vector Regression(SVR),Random Forest(RF),Convolutional Neural Network(CNN),and Fully Convolutional Network(FCN),were employed in this study for modeling.By utilizing data augmentation techniques,the original data was enhanced to establish a red meat fatty acid content model based on data augmentation.We explored various aspects including enhanced data quality analysis,mixed data augmentation,and data extension experiments,and conducted result analysis accordingly.This study provides strong support for non-destructive detection of various indicators of red meat and prediction of meat quality,and it also has significant reference value for future related research.The following are the specific research contents and results:(1)Statistical analysis of red meat fatty acidsThis chapter analyzes the fatty acid content of lamb,beef,and pork,and finds that lamb and beef are predominantly composed of monounsaturated fatty acids(MUFA),while pork is predominantly composed of saturated fatty acids(SFA).Although the minor fatty acid contents differ among the three types of meat,each has its specific role.Lamb has higher levels of linoleic acid and arachidonic acid,while beef has higher levels of oleic acid,and pork has higher levels of cis-8,cis-11,and α-linolenic acid.Choosing the appropriate type of meat depends on individual health conditions and dietary needs.(2)Research on prediction models for major fatty acid content in red meatThis study explores and establishes prediction models for the fatty acid content in red meat,including spectral curve analysis,prediction models based on raw spectra,preprocessing modeling analysis,and modeling analysis based on feature wavelengths.In the spectral curve analysis,the spectral curves of beef and lamb are similar,while pork shows significant differences.The models based on raw spectra can predict SFA and MUFA,but face challenges in predicting polyunsaturated fatty acids(PUFA).CNN and FCN are suitable for PUFA prediction,while RF and SVR are suitable for SFA and MUFA.In the preprocessing models,Baseline and SG smoothing perform well,while OSC performs poorly.In the modeling analysis based on feature wavelengths,IRIV performs optimally in the CNN model for predicting MUFA and PUFA(with R2 values of 0.6603 and 0.6375,respectively)while SFA performs best in the CNN model with raw data,achieving R2 values of 0.7498 and 0.6875 for the training and testing sets,respectively.The optimal feature wavelengths differ for different fatty acids,with raw data generally performing well.(3)Research on red meat fatty acid content models based on data augmentationThis study aims to explore the impact of data augmentation on the models predicting red meat fatty acid content and analyze its effects on data quality and modeling results.We conducted a quality analysis of augmented data,which included spectral curve analysis,data distribution analysis,and mixed spectral curves.Subsequently,we trained and tested red meat fatty acid prediction models using four different modeling methods(SVR,RF,CNN,and FCN),comparing the results obtained with augmented data and non-augmented data.The results indicate that augmented data improves the performance of deep learning models,but it may also lead to issues such as overfitting and excessive learning.Furthermore,we analyzed the prediction results for different types of fatty acids to assess the role of augmented data in various predictions.Overall,augmented data significantly enhances the performance of most models,particularly deep learning models.These findings provide robust support for future related research and offer valuable insights.