Beyond the browning index: RGB radiomics enables non-destructive prediction of fresh-cut "Red Fuji" apple browning and enzyme activities.

Fresh-cut apples undergo rapid enzymatic browning, which compromises visual quality and causes substantial postharvest losses. Most RGB vision approaches rely on handcrafted color/texture descriptors that are sensitive to acquisition conditions and provide limited linkage to the underlying enzyme dynamics. This study proposes a radiomics-enhanced RGB imaging framework to non-destructively quantify browning progression and predict key browning-related enzymatic activities in fresh-cut "Red Fuji" apples. Apple slices were imaged from 0 to 24 h after cutting using a custom high-resolution RGB setup under controlled illumination. Browning regions were automatically cropped from the original photo, and high-throughput radiomic features were extracted from regions of interest with PyRadiomics, including first-order statistics, 2D shape descriptors, and multi-scale texture features from original, wavelet-transformed, and 2D local binary pattern (LBP2D) images. A variance filtering plus Elastic Net selection pipeline produced a compact set of 58 informative and weakly correlated features dominated by wavelet textures. Five regressors (KNN, SVM, EN, DT, and MLP) and four ensemble strategies (Bagging, Boosting, Voting, and Stacking) were evaluated using repeated cross-validation. The best single models achieved R2 of 0.909 (Time), 0.966 (BI), 0.950 (PPO), and 0.918 (POD), while ensemble learning consistently delivered robust performance with R2 generally >0.90 (e.g., voting: 0.913 for Time, 0.972 for BI, 0.924 for POD). These results suggest that radiomic features extracted from low-cost RGB images may provide a useful approach for estimating browning and selected enzyme activities in fresh-cut 'Red Fuji' apples under the conditions investigated.