Shedding Light on the Ageing of Extra Virgin Olive Oil: Probing the Impact of Temperature with Fluorescence Spectroscopy and Machine Learning Techniques

Abstract: This work systematically investigates the oxidation of extra virgin olive oil (EVOO) under accelerated storage conditions with UV absorption and total fluorescence spectroscopy. With the large amount of data collected, it proposes a method to monitor the oil's quality based on machine learning applied to highly-aggregated data. EVOO is a high-quality vegetable oil that has earned worldwide reputation for its numerous health benefits and excellent taste. Despite its outstanding quality, EVOO degrades over time owing to oxidation, which can affect both its health qualities and flavour. Therefore, it is highly relevant to quantify the effects of oxidation on EVOO and develop methods to assess it that can be easily implemented under field conditions, rather than in specialized laboratories. The following study demonstrates that fluorescence spectroscopy has the capability to monitor the effect of oxidation and assess the quality of EVOO, even when the data are highly aggregated. It shows that complex laboratory equipment is not necessary to exploit fluorescence spectroscopy using the proposed method and that cost-effective solutions, which can be used in-field by non-scientists, could provide an easily-accessible assessment of the quality of EVOO.

• The paper demonstrates that fluorescence spectroscopy combined with ML can effectively monitor EVOO ageing by tracking key oxidation parameters.
• The study employs UV-Vis and excitation-emission measurements to quantify changes in pigments and tocopherols under thermal stress.
• The simplified sensor approach using dual excitation wavelengths shows over 90% accuracy in classifying EVOO quality based on oxidation metrics.

Optical and Machine Learning Approaches to Monitoring Extra Virgin Olive Oil Ageing

Introduction

The paper presents a systematic investigation into the thermal oxidation of extra virgin olive oil (EVOO), employing both spectroscopic and ML tools to quantify ageing-induced quality loss. The methodology combines UV absorption and detailed fluorescence spectroscopy, evaluated on 24 commercially sourced EVOOs across several European provenances subjected to accelerated storage at $60^{\circ}$C. The principal objective is to identify minimally complex, cost-efficient sensing strategies for in-field EVOO quality assessment, circumventing the laborious and resource-intensive protocols prescribed by regulatory frameworks.

Experimental Design and Data Acquisition

The study’s experimental structure encompasses a wide diversity of EVOO samples to capture matrix variability. Thermal stress over 53 days is imposed in minimal-oxygen hermetic vials, a critical step to approximate in-bottle storage conditions and mitigate artificial acceleration of oxidative changes. At 10 temporal intervals, each sample undergoes both UV-Vis spectroscopy for established European-regulation absorption parameters ($K_{232}$, $K_{268}$, $\Delta K$) and comprehensive excitation-emission fluorescence matrix (EEM) acquisition (300–650 nm excitation, 300–800 nm emission).

Spectroscopy and Regulatory Quality Metrics

The evolution of $K_{232}$ and $K_{268}$, quantifiers of primary and secondary oxidation products respectively, reveals strong inter-sample variability attributable to compositional heterogeneity. $K_{268}$ demonstrates higher sensitivity, with a 143% mean increment over 53 days, and only 4/24 samples remaining within the extra virgin legal threshold after ageing. $K_{232}$ is less discriminatory (26% mean increment; 7/24 below threshold). $\Delta K$ is deemed unreliable for ML due to signal-to-noise limitations, as it exhibits changes comparable to intrinsic measurement error.

Fluorescence Spectral Analysis

Fresh EVOOs display strong fluorescence bands near 650–750 nm at 350–400 nm and 600–650 nm excitation, coinciding with chlorophyll signatures. Variance is mostly observed at excitation 310–350 nm and emission 350–600 nm, highlighting tocopherols, flavins, and oxidation derivatives. Importantly, all measurements are performed on undiluted samples, emphasizing the goal of non-destructive, sample-preparation-free assessment suitable for on-site application.

During thermal ageing, characteristic spectral changes are observed:

• For 300 nm excitation, intensities at 350 nm (tocopherol-like) and 680 nm (chlorophyll) emissions decrease, supporting their oxidative degradation.
• At 400 nm excitation, 680 nm emission tends to increase, likely due to complex interplay between pigment photochemistry and product formation.
• Excitation at 480 nm consistently yields increased emission in the 500–625 nm and 680 nm regions, indicative of progressive product build-up and pigment transformations.

  Figure 2: Evolution of the fluorescence emission with ageing for two representative oils, showing excitation-dependent intensity changes and spectral redistribution.

Differential fluorescence maps (aged–fresh) at multiple time points highlight both the generalized and oil-specific nature of spectral evolution.

Figure 4: Difference in fluorescence intensity patterns at 9 and 53 days of ageing for two oils; red denotes increase, blue decrease, visualized on a log scale to emphasize subtle spectral shifts.

Feature Reduction and Sensor Practicality

A central hypothesis is that a simplified sensor measuring fluorescence at one or two excitation wavelengths suffices for binary EVOO versus non-EVOO classification. To operationalize this, an information theoretic criterion (relative error $RE$) is used to rank excitation wavelengths by their ageing-induced spectral divergence. Across samples, 480 nm and 300 nm emerge as those providing the maximal or next-maximal $RE$, supporting their discrimination power.

Correlation analysis shows $RE$ at selected excitation wavelengths and the canonical UV-absorption metrics ($K_{268}$, $K_{232}$) are monotonically related, validating the substitution of integrated fluorescence metrics for more complex spectral or chemical assessments.

Machine Learning Classification and Performance

AdaBoost (with decision stumps), Random Forests, logistic regression, and Naive Bayes are employed for the classification task—using only $RE$ features at 480 and 300 nm—against labels defined by the regulatory thresholds in $K_{232}$ and $K_{268}$. AdaBoost yields the strongest performance. A sequential validation regime corresponding to either (i) future time points (extrapolation) or (ii) immediately next temporal step (interpolation) is adopted.

The classifier achieves:

• Accuracy above 90% for $K_{268}$-based labels with three or more ageing points in training.
• Slightly reduced accuracy for $K_{232}$, reflecting its lower sensitivity to thermal oxidation in the experimental conditions.
• Sensitivity and specificity values in the range 0.5–0.95 (the scatter due to the inherent sparsity of "over-threshold" oils after ageing).

Notably, these results are robust despite the highly heterogeneous nature of the oils and minimal feature set—an explicit validation of the proposed “single or dual LED + photodiode” sensor architecture.

Theoretical and Practical Implications

The findings have both technological and methodological implications:

• Sensor Simplification: Demonstrated that fluorescence measurements at only two judiciously chosen excitation wavelengths can replace full spectrum acquisition, supporting the design of inexpensive, portable, non-destructive EVOO screening instruments.
• Aggregated Feature Sufficiency: While aggregation erases detailed compositional changes, it is adequate for the regulatory binary assignment—a relevant outcome for both producers and inspectors.
• Robustness to Product Variability: The approach generalizes across oils spanning a diverse compositional space, suggesting practical applicability beyond the specific experimental cohort.

Theoretical implications pertain to the sufficiency of low-dimensional fluorescence descriptors in capturing high-order chemical complexity relevant for binary quality thresholds. From an ML perspective, the paradigm exemplifies effective feature-space dimensionality reduction guided by unsupervised physical insights.

Prospective AI Directions

Further integration with advanced ML strategies, e.g., self-supervised pretraining on raw EEMs, domain adaptation to oils from different geographic or production contexts, and regression rather than classification targets (e.g., shelf-life prediction), could extend the method’s utility. Miniaturization and embedding of real-time inference pipelines into low-power embedded systems is a foreseeable development pathway.

Conclusion

The study rigorously demonstrates that oxidative ageing assessment for EVOO can be reliably conducted using highly-aggregated fluorescence data and simple ML models, bypassing the need for laborious chemical assays or full-spectrum analysis. This has direct significance for rapid, cost-effective EVOO quality control at both producer and distribution levels, and sets a precedent for the deployment of physically interpretable, ML-powered minimalistic sensors in food quality assessment.