Non-destructive firmness and near-infrared vibrational spectroscopy: a novel tool to predict consumer-derived sensory acceptability profiles of mango purees

Mango ( Mangifera indica L., cv. Cogshall) puree quality is strongly governed by the ripening stage of the raw fruit. This study assessed consumer acceptability of mango purees produced from fruits sorted by non-destructive firmness measurement, and evaluated the feasibility of sensory acceptability profiles from whole-fruit near-infrared spectroscopy (NIRS). Six puree batches representing contrasted ripening stages were produced and fully characterized. Physico-chemical, rheological and particle size properties were characterized instrumentally, and sensory evaluation was conducted with 100 consumers using Just-About-Right (JAR) scales and overall liking scores. Cumulative link mixed models (CLMM) were applied to identify the key sensory drivers of acceptability. Vis–NIR spectra acquired on 817 whole unpeeled fruits across three consecutive harvest seasons were used to train five classification models PLS-DA, OPLS-DA, support vector classification, Random Forest, and the transformer-based TabPFN to predict three sensory acceptability profiles (unbalanced, intermediate-balance, balanced). Purees from intermediate ripening stages were the most appreciated, driven by balanced sweetness and acidity and a smooth texture. Excessive acidity and over-sweetness were the main penalising attributes at the ripening extremes. TabPFN achieved the highest predictive performance on the Kennard–Stone prediction set (balanced accuracy: 0.88; overall accuracy: 0.93; macro F1-score: 0.90), reliably discriminating fruits yielding the unbalanced and the balanced profiles, while the intermediate-balance profile remained more challenging to predict. These results demonstrate the potential of NIRS combined with modern machine-learning approaches for the rapid, non-destructive prediction of consumer-derived sensory acceptability profiles in industrial mango puree production.