Chemical analysis of olive oils from fluorescence spectra thanks to one-dimensional convolutional neural networks

The chemical analysis of food is essential to monitor and guarantee its quality. The determination of the chemical parameters, like the concentration of particular substances, is performed by specialized laboratories and is a time-consuming and costly process. Therefore, alternative methods with easier handling are of great interest. Among these fluorescence spectroscopy offers great opportunities. Fluorescence spectra are one-dimensional arrays of values already successfully employed together with artificial neural networks for classification problems in chemistry, physics, and other fields. However, the extraction of specific quantities from the spectra poses a much harder challenge. This work analyzes and compares the ability of feed-forward neural networks (FFNN) and one-dimensional convolutional neural networks (1D-CNN) to extract relevant features from fluorescence spectra of olive oils. The results indicate that 1D-CNN, contrary to FFNN, successfully predicts the chemical parameters with high accuracy. The great advantages of the proposed method are: 1) the possibility of using optical methods instead of time-consuming chemical ones, like chromatography, 2) the lack of any special sample handling, like dilution and 3) the lack of any pre-processing of the data. The problem of small datasets, which may arise for novel techniques like the proposed one, is also addressed statistically by using the leave-one-out resampling technique.