Discrimination of New Zealand wines by geographical origin and grape variety using 1H NMR metabolomic fingerprinting combined with chemometric and machine learning approaches

A total of 111 commercial New Zealand wines comprising Syrah ( n = 27), Pinot Noir ( n = 30), Sauvignon Blanc ( n = 28), and Chardonnay ( n = 26) produced in the North Island ( n = 41) or South Island ( n = 70) were analysed using untargeted proton nuclear magnetic resonance ( 1 H NMR) fingerprinting. Experimental results were modelled using Random Forest (RF), Support Vector Machines (SVM), Artificial Neural Networks (ANN), Automated Network Search (ANS), Data-Driven Soft Independent Modelling of Class Analogy (DD-SIMCA), Partial Least Squares – Discriminant Analysis (PLS-DA), and Soft Binary PLS-DA, aiming to authenticate wine geographical origin and varietal. Results show that varietal classification generally performed better than geographical-origin classification. SVM and RF achieved mean one-versus-rest accuracies of 97.7% and 96.0%, with mean class-specific Matthews correlation coefficients (MCCs) of 0.942 and 0.893, respectively. The best-performing ANS model correctly classified 110 of 111 wines, achieving 99.1% overall multiclass accuracy and a mean class-specific MCC of 0.988. Exploratory vintage classification achieved overall accuracies of 44.1% for SVM and 48.6% for RF; however, the uneven vintage distribution prevented these results from excluding potential vintage confounding. Hard-decision PLS-DA achieved a mean class efficiency (TEFF) of 99.4%, with one Pinot Noir sample misclassified as Syrah. For geographical-origin discrimination, SB-PLSDA achieved a TEFF of 98.2%, sensitivity of 97.2%, and specificity of 99.1% after excluding two outliers. Leave-One-Variable-Excluded-based selection improved DD-SIMCA specificity against the nearest alternative variety to 81.5–92.3%. These findings support the potential of 1 H NMR fingerprinting for wine discrimination and authentication.