METROFOOD-IT – Metrology for Food and Nutrition: development of chemometric models based on near-infrared spectroscopy and hyperspectral imaging for sourdough bread authentication

Sourdough bread commands a growing market premium, yet no harmonised European or international legislation defines the term or sets minimum process requirements for its use on labels. Products that reproduce the sensory character of sourdough without genuine lactic acid fermentation can therefore be marketed under the same denomination. Existing analytical methods able to verify the claim, from organic acid profiling to NMR metabolomics and quantitative PCR, are destructive, time-consuming and confined to specialised laboratories, and are unsuited to routine screening. This thesis develops and validates chemometric models for sourdough authentication on two near-infrared measurement modalities: single-point spectroscopy (NIRS) and hyperspectral imaging (HSI-NIR). Its contribution is primarily methodological, concerning the construction, interpretation and validation of multivariate models for spectroscopic data. Two studies share a common set of 108 bread slices spanning 18 formulations, so that the measurement modality is the only variable distinguishing them. On a benchtop FT-NIR instrument over 870–2500 nm, PLS-DA separated sourdough from non-sourdough breads with a sensitivity of 96.4%, a specificity of 100% and a non-error rate of 98.2%. On a line-scan camera over 950–1700 nm, each slice was segmented by K-means into five spectral populations and described by cluster-frequency descriptors, returning 88.1%, 62.5% and 75.3% respectively. Both platforms resolved the discriminant signal to the same chemistry: organic acids from lactic acid bacteria and the nitrogenous products of proteolysis. Reducing each slice to its mean spectrum abolished all discrimination, identifying intra-sample spatial heterogeneity, rather than average composition, as the carrier of the authentication signal. The work is presented as a laboratory proof of concept: the analytical procedure transfers, but the fitted models, calibrated on a single sample set without external validation, are not deployment-ready.