Property-dependent predictive performance of linear and nonlinear models in NIR-based monitoring of apple maturity

Abstract This study aimed to evaluate the capability of near-infrared (NIR) spectroscopy combined with multivariate regression algorithms to predict physicochemical properties and characterize ripening-stage progression of Fuji apple. Spectral preprocessing included SNV and smoothing, and three regression algorithms (PLSR, PCR, and SVM) were applied. Results demonstrated that predictive performance strongly depended on the property type. For physical properties, particularly firmness and L*, SVM achieved the highest prediction accuracy ( R 2 = 0.890 and 0.891, RMSE = 3.951 and 3.449, RPD = 3.00 and 2.92, respectively), indicating excellent model robustness and industrial applicability. In contrast, a* and b* showed poor predictability ( R 2 < 0.40, RPD ≈ 1.0–1.2). For chemical properties, starch-related properties showed the best performance (Starch%: R 2 = 0.854, RMSE = 8.518, RPD = 2.33 using SVM; TA: R 2 = 0.858, RMSE = 0.087, RPD = 2.66 using PLS), whereas TSS and pH showed consistently weak predictive performance across all algorithms ( R 2 < 0.32, RPD < 1.3). Overall, NIR spectroscopy combined with appropriate regression modeling proved highly effective for predicting structural and starch-related maturity indicators, whereas prediction of TSS and pH requires further methodological refinement.