Fusion vs. Isolation: Evaluating the Performance of Multi-Sensor Integration for Meat Spoilage Prediction.

High-throughput and portable sensor technologies are increasingly used in food production/distribution tasks as rapid and non-invasive platforms offering real-time or near real-time monitoring of quality and safety. These are often coupled with analytical techniques, including machine learning, for the estimation of sample quality and safety through monitoring of key physical attributes. However, the developed predictive models often show varying degrees of accuracy, depending on food type, storage conditions, sensor platform, and sample sizes. This work explores various fusion approaches for potential predictive enhancement, through the summation of information gathered from different observational spaces: infrared spectroscopy is supplemented with multispectral imaging for the prediction of chicken and beef spoilage through the estimation of bacterial counts in differing environmental conditions. For most circumstances, at least one of the fusion methodologies outperformed single-sensor models in prediction accuracy. Improvement in aerobic, vacuum, and mixed aerobic/vacuum chicken spoilage scenarios was observed, with performance enhanced by up to 15%. The improved cross-batch performance of these models proves an enhanced model robustness using the presented multi-sensor fusion approach. The batch-based results were corroborated with a repeated nested cross-validation approach, to give an out-of-sample generalised view of model performance across the whole dataset. Overall, this work suggests potential avenues for performance improvements in real-world, minimally invasive food monitoring scenarios.