Unveiling varietal specificity in non-destructive grape quality monitoring: Explainable AI and feature selection for sugar and organic acid prediction using NIR spectroscopy.

Sugar and organic acid content are crucial factors determining grape quality. Non-destructive testing of these components aids in accurately determining optimal harvest timing and wine-making potential. However, few studies have addressed how varietal differences impact the universality of predictive models. This study combines near-infrared spectroscopy with machine learning, specifically partial least squares regression (PLSR) and convolutional neural networks (CNN). It compares single-variety and mixed-variety modeling strategies for grape sugars (glucose, fructose) and organic acids (malic acid, tartaric acid, shikimic acid). The results indicate that PLSR models constructed based on single varieties demonstrate superior performance in predicting malic acid, glucose, and fructose, with model R2P ranging from 0.835 to 0.923, notably outperforming PLSR and CNN models based on mixed varieties. The competitive adaptive reweighted sampling (CARS) and successive projections algorithm (SPA) algorithms successfully compressed the full-spectrum variables to 6-29 key wavelengths. The simplified models maintained high accuracy (R2P = 0.777-0.927) while substantially improving model efficiency. Mechanistically, SHapley Additive exPlanations (SHAP) analysis revealed the significance of key variables. The critical variables for glucose and fructose models occur around 1150 nm and 1450 nm, respectively. In contrast, the feature variables for the malic acid model exhibit broader distribution, spanning multiple bands including 1150 nm, 1200 nm, 1600 nm, and 1650 nm. This study provides a solid foundation and mechanistic explanation for non-destructive grape quality assessment, while also offering theoretical support for developing specialized spectral sensors.