Discrimination of plant-based milk alternatives: a pilot study comparing mid and near-infrared spectroscopy

Plant-based milk alternatives (PBMAs) are expanding rapidly, motivating the development of fast, minimal-prep analytical tools for their analysis. We collected and analyzed (n = 154) commercial PBMAs (almond, oat, rice, soy, coconut) from Austrian and Croatian supermarkets using attenuated total reflection mid‑infrared (ATR‑MIR) and diffuse reflection near‑infrared (DR‑NIR) spectroscopy to compare classification performance across spectral regions. Spectra were acquired without notable sample preparation and processed using second-derivative transformation. Exploratory principal component analysis (PCA) and partial least squares discriminant analysis (PLS‑DA) were applied in one‑class (PLS1‑DA) and multi‑class (PLS2‑DA) modes. Comparison of using either NIR or MIR was done on the Austrian samples only, as only legacy MIR data was available for the Croatian samples. A significant difference (p < 0.05) in the obtained predictions was found between the usage of MIR and NIR for one-class PLS1-DA identification of oat and almond models as well as for the discrimination using the PLS2-DA multi-class approach. Cross‑validated multi‑class macro‑F1-score reached 0.90 (MIR) versus 0.81 (NIR), a F1-score of 0.98 for one-class almond classification using MIR data and 0.90 for soy one-class classification using the NIR data was obtained. The MIR models trained on Austrian data showed excellent generalization on Croatian samples, for almond, oat, rice, and soy, while coconut showed poor generalization. Overall, our results suggest that ATR‑MIR offers slightly better discrimination to DR‑NIR for PBMA classification, though NIR remains attractive for straightforward portable screening. In the future, expanding sample diversity and exploring advanced machine learning modelling could further improve performance, particularly for coconut PBMAs.