Optical spectra, and particularly fluorescence spectra, contain a large quantity of information about the substances and their interaction with the environment. It is of great interest, therefore, to try to extract as much of this information as possible, as optical measurements can be easy, non-invasive, and can happen in-situ making the data collection a very appealing method of gathering knowledge. Artificial neural networks are known for their feature extraction capabilities and are therefore well suited for this challenge. In this work, inspired by convolutional neural network (CNN) architectures in 2D and their success with images, a novel approach using one-dimensional convolutional neural networks (1D-CNN) is used to extract information on the measured spectra by using explainability techniques. The 1D-CNN architecture has as input the entire fluorescence spectrum and takes advantage in its design of prior knowledge about the instrumentation and sample characteristics as, for example, spectrometer resolution or the expected number of relevant features in the spectrum. Even if network performance is good, it remains an open question if the features used for the predictions make sense from a physical and chemical point of view and if they match what is known from existing studies. This work studies the output of the convolutional layers, known as feature maps, to understand which features the network has effectively used for the predictions, and thus which part of the measured spectra contains the relevant information about the phenomena at the basis of what has to be predicted. The proposed approach is demonstrated by applying it to the determination of the UV absorbance at 232 nm, K232, from fluorescence spectra using a dataset of 18 Spanish olive oils, which were chemically analysed from certified laboratories. The 1D-CNN successfully predicts the parameter K232 and enables, by studying feature maps, the clear identification of the relevant spectral features. The main contributions of this work are two. Firstly, it describes how designing the neural network architecture with prior knowledge (spectrometer resolution, etc.) will help the network in learning features that have a clear connection to the chemical composition of the substances, and thus are clearly explainable. Secondly, it shows how, in the case of olive oil, the identified features match perfectly the relevant features known from existing previous studies, thus confirming that the network is learning from the underlying chemical process.
