Integrating transfer learning and spectroscopy for enhanced pork spoilage assessment using correlation analysis

Area of Science:

• Food Science and Technology
• Analytical Chemistry
• Spectroscopy

Background:

• Accurate Total Viable Count (TVC) detection is crucial for ensuring food quality and safety.
• Traditional methods for assessing meat spoilage can be time-consuming and labor-intensive.
• Visible Near-Infrared (VNIR) spectroscopy offers a rapid, non-destructive analytical technique.

Purpose of the Study:

• To investigate the feasibility of using VNIR spectroscopy combined with transfer learning (TL) for monitoring pork chemical spoilage.
• To develop and optimize models for predicting TVC, total volatile basic nitrogen, pH, and color in pork.
• To enhance the accuracy of TVC prediction through advanced TL methods.

Main Methods:

• Utilized VNIR spectroscopy (400-1000 nm) to collect spectral data from pork samples.
• Developed initial predictive models using full spectral bands for TVC, total volatile basic nitrogen, pH, and color.
• Applied and compared various transfer learning (TL) techniques, including multiple correlation chain stacking-partial least squares, to optimize the TVC model.

Main Results:

• Base VNIR models showed good predictability for pork TVC, total volatile basic nitrogen, pH, and color (R_P: 0.821–0.916).
• The optimized TVC model using multiple correlation chain stacking-partial least squares achieved a high R_P of 0.947 and reduced RMSEP by 31.12%.
• The optimized TVC model achieved an RMSEP of 0.425 lg CFU/g and a relative percent deviation of 2.355%.

Conclusions:

• VNIR spectroscopy, when combined with transfer learning, is a feasible and effective method for monitoring the chemical spoilage of pork.
• The optimized TL approach significantly improved the accuracy of TVC prediction, offering a rapid alternative to traditional methods.
• This integrated approach holds promise for enhancing real-time food quality and safety assessment in the meat industry.