Incorporation of visible/near-infrared spectroscopy and machine learning models for indirect assessment of grape ripening indicators

Abstract

Abstract The assessment of grape ripeness is pivotal for optimizing harvest timing and ensuring high-quality fruit production. Traditional methods, relying on manual sampling and chemical analysis, are laborious and expensive. This study proposes an innovative approach combining Visible/Near-Infrared (VIS/NIR) spectroscopy with machine learning techniques—specifically, decision trees (DT) and gradient boosting regression (GBR)—to facilitate a rapid, non-destructive, and cost-effective prediction of key grape ripening indicators such as anthocyanin (An), total acidity (TA), total soluble solids (TSS), and the TSS/TA ratio. The performance of spectral reflectance indices (SRIs) in correlating with these ripening metrics across different grape ripening stages was examined. The study findings revealed notable variations in ripening indicators across stages, with the newly developed SRIs outperforming the existing indices. The application of dual and triple-band SRIs yielded strong correlations with An (R 2 = 0.75–0.88) and TSS (R 2 = 0.64–0.76), and moderate correlations with TA (R 2 = 0.63–0.70), but showed weaker associations with the TSS/TA ratio (R 2 = 0.15–0.52). Incorporating these SRIs into the DT and GBR models significantly enhanced the accuracy of ripening indicator predictions. The integration of multiple SRIs resulted in the most precise models. The DT model delivered outstanding performance for An (R 2 = 0.87, RMSE = 87.81) and TSS/TA (R 2 = 0.74, RMSE = 3.12). Meanwhile, the GBR model excelled in predicting TSS (R 2 = 0.82, RMSE = 0.92) and TA (R 2 = 0.70, RMSE = 0.05). Overall, the combination of VIS-NIR spectroscopy and machine learning offers a promising and efficient approach for assessing grape ripeness, providing a practical solution for the agricultural industry.