Sara Wilhelmina Erasmus

Consumer demands and EU policies on animal welfare have resulted in the extensification of animal husbandry systems reflected in increased space, different breeds, feed adaptations and enrichment options. Apart from impacting animal welfare, such changes may impact the quality of the meat. However, since extensive farming comprises a range of husbandry systems and factors, it is difficult to draw generalized conclusions on husbandry-meat quality relations. The complex nature asks for a data-driven approach. Therefore, the aim of this study was to use machine learning models to establish relations between pork quality characteristics and husbandry extensification factors grouped into: space, genetics, feed, and enrichment. The models were trained to distinguish the different levels of the four extensification factors using the physicochemical properties of the meat and the country of origin. The highest performance of the model was achieved by the model for the genetics factor, followed by the diet factor, the space allowance and the quality of space. The model performance used as a proxy for the strength of the relation revealed the most pronounced differences between indigenous and conventional breeds. Further analysis of the genetics factor model using model-agnostic explainable artificial intelligence uncovered that its predictions were mostly based on the country of origin, moisture, L*, and EZ-DripLoss. These results provide insights into likely links between the physicochemical properties of pork and the extensification factors for which the causal relations can be tested in future research.

This study introduces an Artificial Intelligence (AI)-based proof-of-concept approach to tackle food fraud by using convolutional neural networks (CNNs) and citizen science-generated imagery to predict the country of origin of Cavendish banana cultivars (Musa spp.). A total of 6000 images were collected from iNaturalist, and a CNN classifier was trained to distinguish bananas sourced from six countries. Transfer learning was leveraged, and among nine pre-trained models tested, MobileNetV1 demonstrated the best trade-off between performance and computational efficiency. Following model fine-tuning, data augmentation was implemented to mitigate class imbalance and ensure a dense feature space. The model achieved an average accuracy of 0.86 with Monte Carlo Cross Validation and 0.77 with a 5-Fold Cross Validation. The final selected model attained a validation accuracy of 0.79. Accordingly, this study should be viewed as a foundational proof-of-concept demonstrating the potential of AI for origin detection at the cultivation stage. While the current evaluation framework reflects an early-stage experimental setting, the findings introduce a promising new dimension for proactive food fraud detection. Moving forward, this pipeline provides a foundation that can be expanded and independently validated.