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

A physics-informed neural network using hyperspectral Vis-NIR data predicts stem-end rot in avocados better on unseen temporal conditions than spectral-only and time-input models, especially for day-12 testing.

• Why it matters: Postharvest quality forecasting is needed to cut food waste, but models often fail when storage time shifts beyond the training window.
• What they did: The study compared three models—spectral feedforward neural network, an extended network adding storage duration, and a PINN with a first-order kinetic constraint—training on early storage stages and testing on day-12.
• The result: Adding time as an input can overfit, whereas the PINN’s mechanistic constraint improves generalization for temporal extrapolation, enabling more robust and interpretable time-dependent assessment.

AI summary of the abstract · may contain errors · not medical advice · read the paper.