Deep learning domain adaptation to understand physico-chemical processes from fluorescence spectroscopy small datasets and application to the oxidation of olive oil

Abstract

Fluorescence spectroscopy is a fundamental tool in life sciences and chemistry, with applications in environmental monitoring, food quality control, and biomedical diagnostics. However, analysis of spectroscopic data with deep learning, in particular of fluorescence excitation-emission matrices (EEMs), presents significant challenges due to the typically small and sparse datasets available. Furthermore, the analysis of EEMs is difficult due to their high dimensionality and overlapping spectral features. This study proposes a new approach that exploits domain adaptation with pretrained vision models, along with a novel interpretability algorithm to address these challenges. Thanks to specialised feature engineering of the neural networks described in this work, we are now able to provide deeper insights into the physico-chemical processes underlying the data. The proposed approach is demonstrated through the analysis of the oxidation process in extra virgin olive oil (EVOO), showing its effectiveness in predicting quality indicators and identifying the spectral bands and thus the molecules involved in the process. This work describes a significantly innovative approach to deep learning for spectroscopy, transforming it from a black box into a tool for understanding complex biological and chemical processes.

Keywords: Deep learning; Domain adaptation; Excitation emission matrices; Fine tuning; Fluorescence spectroscopy; Food quality; Olive oil; Transfer learning.

Fig. 1

Overview of the phases of the machine learning approach. (a) The data preprocessing phase consists of splitting the dataset for the LOO approach, normalisation of pixel values, and preparation for the MobileNetv2 network input layer by reshaping and creating the three necessary layers. (b) The domain adaptation phase consists of transfer-learning and fine-tuning of the network using the training dataset. The trained network is then evaluated on the test dataset and its performance assessed through the Mean Absolute Error (MAE). (c) Information Elimination Approach (IEA) process diagram. doi references indicate the papers that describe some of the used components.

Fig. 2

Comparison of the true (blue) and predicted (red) values of the quality indicator $K_{232}$ for all the oils at all oxidation stages (vertical scale on the left axis). The corresponding absolute error (AE) is shown as an area in yellow (vertical scale on the right axis). The Mean Absolute Error (MAE) obtained as average over all the oxidation stages is displayed in each panel for each oil.

Fig. 3

Comparison of the true (blue) and predicted (red) values of the quality indicator $K_{268}$ for all the oils at all oxidation stage (vertical scale on the left axis). The corresponding absolute error (AE) is shown as an area in yellow (vertical scale on the right axis). The Mean Absolute Error (MAE) obtained as average over all the oxidation stages is displayed in each panel for each oil.

Fig. 4

(a) Comparison of predicted and measured (actual) values of the quality indicators $K_{232}$ and $K_{268}$ for all oils at all oxidation stages. The gray area in each plot marks the limit set by the Food and Agriculture Organisation of the United Nations and by the European Union. Oil C is marked in blue as the $K_{232}$ value was was already above this limit at the beginning of the study and, therefore, is not well predicted by the model. (b) Violin plots of the AE for each oil for $K_{232}$ (above) and $K_{268}$ (below). The dashed lines indicate the 3$\sigma$ statistically estimated experimental error.

Fig. 5

Average of the heatmaps obtained for all oils in the last oxidation stage showing the spectral band of relevance for the prediction of the $K_{232}$ and $K_{268}$. R1 marks the absorption and emission bands of chlorophylls, R2 those of oxidation products.

Fig. 6

IEA approach showing the spectral bands of the fluorescence spectrum which most significantly contribute to the prediction for two selected oils in the last oxidation stage. (a) EEM of Oil J and region importance heatmap overimposed for quality indicator $K_{232}$. The fluorescence spectrum at ${\lambda }_{\text{ex}}=416$ nm is shown in the top panel. (b) EEM of oil V and region importance heatmap overimposed for quality indicator $K_{268}$. The fluorescence spectra at ${\lambda }_{\text{ex}}=344$ nm and ${\lambda }_{\text{ex}}=392$ nm are shown in the top panel.

Fig. 7

IEA approach showing the evolution during the oxidation process on two selected oils: “Fresh” marks the oil just after opening the bottle, “Oxidized” the oil at the latest oxidation stage. (a) Oil P EEM and region importance heatmap overimposed for the prediction of the quality indicator $K_{232}$; (b) Oil D and region importance heatmap overimposed for the prediction of the quality indicator $K_{232}$.