Image-based modeling of three-way data for butter authentication

In recent years, the issue of food authentication has received increasing attention due to the growing consumer demand for high quality and regionally distinctive food products [1]. As a result, food safety and authenticity detection play a crucial role in safeguarding consumers’ interest, especially in cases involving high value-added ingredients [2]. Among these, butter, is practically susceptible to intentional adulteration with low-quality additives for economic gain.<br/><br/>Currently, some of the most effective analytical techniques for detecting food adulteration rely on three-way methods, such as fluorescence spectroscopy and GC-MS. Classical approaches for analysing such multiway data include well-established methods like PARAFAC and Tucker decomposition. However, these techniques are not always suitable for building robust discrimination or authentication models. <br/><br/>This study explores an alternative strategy: treating three-way datasets as images, so two modes (e.g. excitation and emission wavelength) serve as spatial dimensions and the spectral intensities represent pixels values. This transformation allows the application of powerful image analysis models, particularly Convolutional Neural Networks (CNN) – that are widely used for classification and anomaly detection. These models can be pre-trained on large image datasets and subsequently fine-tuned for specific cases, thereby reducing the need for extensive domain-specific training data.<br/><br/>To validate this approach, a butter authentication case study was conducted. Samples<br/>from six different commercial brands of butter were collected from various retail sources. A subset of these samples was adulterated with 5% and 10% w/w of coconut oil, resulting in a total of 91 samples. Excitation emission matrices (EEMs) were acquired for each sample using two acquisition setups. As a preliminary step, a CNN was trained to perform a discriminant analysis for detecting the presence or absence of adulterants. Subsequently, since the focus is on identifying a single target class (pure butter), a one-class modelling approach was adopted [3]. Specifically, a Variational Autoencoder (VAE), a class of deep learning models designed for image<br/>representation - was combined with the SIMCA algorithm to construct a one-class classifier [4]. All experimental results and methodological specifics will be disclosed during the presentation.<br/><br/>References<br/>[1] Rodionova, O. Y., Oliveri, P., Malegori, C., et al. Trends Food Sci. Tech., 2024 (147), 104429<br/>[2] Wang, Y., Gu, H. -W., Yin, X. -L., et al. Trends Food Sci. Tech., 2024 (146), 104396<br/>[3] Oliveri, P., Anal. Chim. Acta, 2017 (982), 9-19<br/>[4] Petersen, A., Kucheryavskiy, S., Chemometr. Intell. Lab., 2025 (256), 105276<br/>