Classification of grape ripeness in the vineyard using hyperspectral imaging: comparison of cultivar-specific and cross-cultivar models

Abstract

Determining the optimal harvest date is critical because it directly affects wine quality and profitability. Conventional destructive sampling limits temporal frequency and spatial representativeness. This study compared cultivar-specific and cross-cultivar machine-learning models for classifying ripeness using hyperspectral images acquired under commercial vineyard conditions. Images covering 400–1700 nm were collected from a mobile platform between veraison and harvest in Tempranillo Tinto and Tempranillo Blanco vineyards in La Rioja. Total soluble solids, pH and total acidity from 200 berries per vine were integrated by principal component analysis to generate a global ripeness index and classify vines into early ripening, advanced ripening and harvest maturity. Preprocessed spectral data were used to train Random Forest and Support Vector Machine models. Cultivar-specific models achieved accuracies above 95 % for early ripening and harvest maturity, while the intermediate stage reached 70–80 %, reflecting its progressive nature. The cross-cultivar model performed consistently, with only a slight reduction in accuracy, indicating both conserved spectral patterns and cultivar-specific variability. Hyperspectral imaging is therefore suitable for operational ripeness classification, while transferable models offer a promising basis for scalable decision-support systems in precision viticulture.

References

1. Tejada-González, A. Classification of Grape Ripeness in the Vineyard Using Hyperspectral Imaging: Comparison of Cultivar-Specific and Cross-Cultivar Models. Recording of the presentation delivered at Enoforum Spain 2026, Zaragoza, 20–21 May 2026. Available soon on Infowine Premium: https://vinidea.it/es/brand/infowine/