An integrated machine learning framework for accurate, real-time and non-destructive vitamin C quantification from Vis/NIR spectral data

This study aimed to develop and evaluate a hybrid framework integrating feature selection and regression modeling for the nondestructive prediction of vitamin C, soluble solids content, and firmness in dried kiwifruit. Within this framework, the spectra of 250 dried kiwifruit samples were acquired using visible–near-infrared spectroscopy over the wavelength range of 470–980 nm. Four feature selection methods, including the ant colony optimization algorithm, competitive adaptive reweighted sampling, genetic algorithm, and minimum redundancy–maximum relevance, were employed to identify and compare the most informative wavelengths. Thirty wavelengths selected by each method were subsequently used as input variables for four regression models, namely partial least squares regression, random forest, support vector regression, and extreme gradient boosting. Model performance was evaluated during the validation stage using the coefficient of determination (R 2 ), root mean square error, and residual predictive deviation. The results showed that, for the prediction of vitamin C and soluble solids content, the combination of SVR with ACO achieved the highest validation performance, with RPD values of 5.52 and 5.17, respectively. In contrast, for firmness prediction, the combination of PLSR and CARS exhibited more stable predictive performance, yielding a validation R 2 of 0.91, RMSE of 0.43, and RPD of 3.50. Comparison of the feature selection methods also enabled an assessment of the degree of overlap among the selected wavelengths and their potential roles in predicting different quality attributes of dried kiwifruit. These findings demonstrate that the effectiveness of feature selection depends on both the target quality attribute and the type of regression model, and that integrating appropriate feature selection methods with suitable regression algorithms can improve the nondestructive prediction of quality attributes in dried kiwifruit. Future studies should employ larger and independent datasets and evaluate the proposed framework under real-world and online conditions to further assess its generalizability and practical applicability