Hyperspectral imaging and linear and nonlinear machine learning for tracing the geographical origin of pistachios

Authenticating the geographical origin of pistachios is crucial to protect premium varieties such as the PDO “Bronte Green Pistachio” from fraud. This study investigates the potential of visible/near-infrared (Vis/NIR) hyperspectral imaging (HSI) combined with chemometric and machine learning methods for discriminating pistachios from different geographical origins. Traceable batches of Pistacia vera L. from Iran, California (USA), Turkey, and Italy (Bronte and non-Bronte Sicily) were analysed in three sample forms: whole kernels, bulk samples, and ground powders using a 400–1000 nm HSI system. Spectral data were processed using Partial Least Squares–Discriminant Analysis (PLS-DA) and Multilayer Perceptron Artificial Neural Networks (MLP-ANN). PLS-DA achieved high discrimination among the four origins, with overall test set accuracies above 98% for bulk and powder samples, and slightly lower accuracies for individual kernels (86%). MLP-ANN models confirmed the high predictive potential, yielding comparable accuracies (>90%), particularly for ground samples (up to 100%). When focusing on the binary classification between Bronte and non-Bronte Sicilian pistachios, PLS-DA achieved satisfactory discrimination only for bulk/powder samples and not for single kernels, despite the high similarity between the two groups, achieving classification accuracies of 100% for Bronte and 97% for non-Bronte pistachios. Pixel-wise classification maps demonstrated the feasibility of spatially resolved origin prediction. The results indicate promising potential for laboratory-based quality control and traceability applications in the nut industry. • Vis/NIR hyperspectral imaging was applied to classify pistachios by geographical origin. • PLS-DA and ANN models were adopted as effective classification techniques. • Model accuracy ranged from 85% to 95% according to sample (bulk, ground, kernels). • Spatial analysis confirmed the consistent spatial distribution of classified pixels.