VIS/NIR hyperspectral imaging with artificial neural networks to evaluate the content of thiobarbituric acid reactive substances in beef muscle

• Published in: Journal of food engineering Vol. 350; p. 111500
• Main Authors: Park, Seongmin, Yang, Myongkyoon, Yim, Dong Gyun, Jo, Cheorun, Kim, Ghiseok
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.08.2023
• Subjects: Beef; Hyperspectral imaging; Machine learning; Near infrared (NIR); Rancidity; Near infrared (NIR); Beef; Rancidity; Machine learning; Hyperspectral imaging
• ISSN: 0260-8774
• DOI: 10.1016/j.jfoodeng.2023.111500

Machine learning models were developed to predict the degree of rancidity of beef by a non-destructive method using a near infrared hyperspectral image acquisition system. The beef subject to the experiment was naturally oxidized during the 15-day cooling process. In a darkroom environment, hyperspectral data cubes were collected using a data acquisition device. Additionally, a technique was developed to selectively extract lean-meat spectra from hyperspectral data obtained from beef that was refrigerated for a variety of lengths of time. Thiobarbituric acid reactive substances (TBARS) experiment was performed in a traditional method to secure reference values for the rancidity level of the sample. Spectra were extracted through data selection and separated by training set and test set. PLSR, ANN, and 1D-CNN techniques were applied to model development. Variable Importance in Projection (VIP) score for the wavelength band was calculated, and the portion judged as valid was cut out to generate a reduced data set. Chemical maps were created for each developed model to visualize the performance of the model. As a result of the development, it was confirmed that the rancidity level of beef could be predicted through a model generated by hyperspectral data. •Proposal of non-destructive and real time method for predicting rancidity of beef using hyperspectral spectrum.•Suggestion of technical method for extracting only valid areas from hyperspectral data cubes.•ANN based predictive model showed the best results in both the full spectrum and selectively reduced spectrum.•Data reduction using significant wavelengths is expected to be efficient in application.