Hyperspectral imaging with machine learning for non-destructive prediction of quality attributes in Lanzhou frozen pears

Frozen pears undergo complex internal quality changes during frozen storage, while conventional evaluation methods are destructive and time-consuming, limiting rapid or continuous monitoring. This study employed hyperspectral reflectance imaging combined with spectral preprocessing, feature wavelength selection, and machine learning algorithms to model key quality indicators of frozen pears, including soluble solids content (SSC), titratable acidity (TA), and total sugar content (TSC). The results revealed distinct optimal modeling strategies for different quality attributes. Stable predictions of TSC were achieved using full-spectrum data or weak wavelength screening, whereas SSC and TA required targeted wavelength optimization to improve prediction robustness. The optimal models exhibited high coefficients of determination, low prediction errors, and acceptable RPD values across prediction sets. Further validation using the guidance method confirmed stable error distributions without evident overfitting. Overall, this study demonstrates that hyperspectral imaging can provide a reliable alternative for non-destructive quality assessment of Lanzhou frozen pears. • Hyperspectral imaging enables nondestructive evaluation of frozen pear quality. • Spectral preprocessing and wavelength selection influenced model performance. • PLSR models achieved acceptable prediction accuracy for SSC, TA, and total sugars. • Model uncertainty was quantified using cross-validation and bootstrap analysis.