Image-based modelling and augmentation of three-way data: butter authentication example

This paper demonstrates that Convolutional Neural Network (CNN) models, originally developed for image classification, can be applied to three-way data originating from second-order analytical techniques represented in form of digital images. Excitation-emission matrices (EEMs) of Danish butter samples, both pure and adulterated with Palmin – bleached and purified coconut oil – at two concentration levels (5% and 10%), were used as a benchmark dataset. A recently proposed VAESIMCA method, combining Variational Autoencoders with the Data-Driven SIMCA (DD-SIMCA) approach for one-class classification, was employed and compared against conventional DD-SIMCA as a baseline. Two strategies for data augmentation based on Procrustes cross-validation (PCV) were proposed: full unfolding and multiblock unfolding. The results show that multiblock unfolding approach preserves the original data structure without introducing artifacts and is most suitable for image-based models, while full unfolding strategy benefits matrix-based methods. Both approaches achieved satisfactory classification performance, with VAESIMCA reaching efficiency of up to 100%. Furthermore, reconstruction error maps from the VAESIMCA models provided interpretable visualization of spectral regions responsible for the adulteration, highlighting the potential of Deep Learning approaches for chemometric interpretation of three-way data.