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

Abstract

Accurate prediction of postharvest fruit quality is essential for reducing food waste and improving supply chain efficiency. This study investigates the integration of spectral information and temporal dynamics for predicting stem-end rot in avocados using hyperspectral Vis-NIR data. Three modelling approaches were compared: a feedforward neural network using only spectral information, an extended model incorporating storage duration, and a physics-informed neural network (PINN) encoding a first-order kinetic constraint. To evaluate temporal extrapolation, models were trained on samples from early storage stages and tested on day-12 data. Results show that while spectral data provide strong predictive capability, incorporating time as a simple input can lead to overfitting. In contrast, the PINN improves generalization by guiding learning with mechanistic knowledge, achieving the best performance on unseen temporal conditions. These findings highlight the value of combining data-driven and physics-based approaches, offering a promising pathway toward more robust and interpretable models for time-dependent food quality assessment.