Overcoming Intrinsic Limitations of Fluorescence Spectroscopy Data in Food Quality and Safety with Machine Learning and Deep Learning

Ensuring food safety and quality remains a major challenge amid growing global food demands and complex supply chains. Fluorescence spectroscopy, owing to its speed and sensitivity, is a promising non-destructive method for food quality assessment, but its effectiveness is often hindered by the complexity and scarcity of spectral data. This study investigates the application of machine learning (ML) and deep learning (DL) techniques to overcome these limitations and presents novel methods for extracting chemical and physical insights, including improved feature selection tailored to continuous spectral ranges, to enhance the analysis of one-dimensional (1D) and two-dimensional (2D) fluorescence data. Two case studies are presented: (1) classification of multiple mycotoxin contamination in corn using 1D fluorescence spectra; and (2) prediction of extra virgin olive oil (EVOO) quality degradation using 2D excitation-emission matrices through DL models with domain adaptation and transfer learning. The results highlight the potential of ML/DL to improve spectral data interpretation, address limitations of small datasets, and uncover underlying physico-chemical insights, ultimately enabling faster, more cost-effective, and sustainable food quality monitoring.