Cross-cultivar prediction of apple SSC: Mitigating biological variability through optical properties and calibration transfer methods.

Non-destructive detection of apple soluble solids content (SSC) is challenged by biological variability. In this study, the visible/near-infrared (Vis/NIR) spectroscopy of the intact fruit and the optical properties of skin and flesh tissue were compared for SSC prediction across four cultivars ('Aksu', 'Fuji', 'Smith' and 'Ruixue'). Moreover, the performance of model updating and calibration transfer methods, including slope and bias correction (SBC) and parameter-free calibration enhancement (PFCE), was evaluated in cross-cultivar prediction. Results indicated that the models based on flesh-skin-μa outperformed the Vis/NIR models in both single-cultivar modeling and cross-cultivar prediction. Model updating with 25-30 newly added samples improved the cross-cultivar prediction performance, with root mean square error of prediction (RMSEP) reduction to 23.0-50.7%. Among the calibration transfer methods, multitask-PFCE demonstrated the optimal performance, outperforming model updating with determination coefficient of prediction (Rp2) of 0.800-0.900 and RMSEP of 0.300-0.450°Brix. This study provides insights and methodological guidance for addressing biological variability in non-destructive fruit quality detection.