Spectroscopy-Based Machine Learning to Predict Stem-End Rot in Avocados with the Temporal Effects

Accurate prediction of postharvest fruit quality is crucial for diminishing food waste and enhancing supply chain efficiency. The current research explores the integration of spectral data and temporal dynamics to forecast stem-end rot in avocados using hyperspectral Vis-NIR information. The study compares three modeling strategies: a feedforward neural network relying solely on spectral data, an extended model that includes storage duration, and a physics-informed neural network (PINN) that incorporates a first-order kinetic constraint.

To assess the model's ability to extrapolate time, the models were trained on samples from the initial stages of storage and evaluated using data from day-12. The findings reveal that while spectral data demonstrate strong predictive power, integrating time as a basic input can result in overfitting. Conversely, the PINN enhances generalization by leveraging mechanistic knowledge to guide learning, ultimately yielding the most favorable performance on unseen temporal circumstances.

These insights underscore the significance of merging data-driven and physics-based methodologies, presenting a promising route towards more dependable and interpretable models for assessing time-dependent food quality.