Prediction of remaining shelf life using hyperspectral imaging

Summary

This project aimed to develop, implement, and test a robust algorithm to quantify the remaining shelf life
of cod using hyperspectral imaging. The project addressed the need for rapid, objective, and non
destructive quality-assessment methods in the seafood industry, where shelf-life evaluation currently
relies on sensory assessment and time-consuming laboratory analyses.
Initially, a sensory reference method was established to determine whether cod fillets are saleable,
marginal, or expired. This method was then applied in storage trials that covered relevant variations in
raw material history, capture method, handling conditions, and storage practices. Hyperspectral images
acquired during storage were used to develop regression models to estimate remaining shelf life, with
the sensory evaluation as the reference standard.
The results showed that VNIR (visible Near infrared) hyperspectral images contain relevant and robust
information for estimating the remaining shelf life of cod fillets. A neural network-based model achieved
promising predictive performance while maintaining generalizability across storage experiments not
included in the training set. Model interpretation using partial dependence analysis showed that the
model relied on stable spectral indicators, supporting the conclusion that predictions were based on
quality-related information rather than experiment-specific artefacts. Model stress testing under non
ideal storage conditions demonstrated that the model captured the expected reduction in remaining
shelf life for fillets from fish exposed to elevated pre-filleting temperatures. In an independent validation
experiment, the model identified clear differences between raw material groups using the images
acquired immediately after filleting, including variation associated with vessel, capture method, and fish
condition. Pixel-level prediction maps further revealed spatial differences in predicted remaining shelf
life within individual fillets, even though the model was trained using only fillet-level sensory reference
values. These results show that variability associated with raw material characteristics, handling
methods, and storage history can be detected using hyperspectral imaging.
In addition to the main project goals, the project investigated temporal changes in spectral signals during
storage and explored benchtop NMR (nuclear magnetic resonance) spectroscopy for measuring TMA
(trimethylamine) and TMAO (trimethylamine oxide), linked to cod spoilage. These analyses showed
systematic spectral and chemical changes during storage, including fillet-specific degradation patterns
and spatial differences between anatomical regions. NMR measurements supported the hyperspectral
observations by indicating faster deterioration in the belly and tail regions than in the loin. These results
provide new and useful information for future work aimed at improving model interpretation, reducing
reference uncertainty, and further strengthening remaining-shelf-life prediction models.
Overall, the project demonstrates the potential of hyperspectral imaging as a useful tool for non
destructive shelf-life prediction, product sorting, and improved quality management in the industrial
production of cod fillets.

Read publication here:

NVA : hdl.handle.net/11250/5595982