Abstract:
Hyperspectral imaging (HSI) is a powerful tool for nondestructive quality assessment of fresh
produce, providing pixel-level spectral and spatial information. However, its performance
strongly depends on imaging configuration illumination geometry, background material, and
light polarization specifically in optically complex, highly scattering samples such as leafy
vegetables. Nevertheless, no clear criteria exist for tuning HSI setups according to sample type or
analytical goal. This study investigates six short-wave infrared (SWIR, 1000-1700 nm) HSI
configurations combining three polarization states (parallel, cross, and non-polarized), each
paired with either a black or white background, to examine their effects on reflectance spectral
quality and ascorbic-acid content interpretability in lettuce (Lactuca sativa L.) stored under 5°C,
20°C and 35°C for 6 days. Partial least-squares regression models were built from spectra
acquired under six configurations; informative wavelengths were selected with variable
importance-in-projection (VIP) scores > 1. Spectra recorded against a black background
consistently outperformed those from a white background. Across polarization states, predictive
power ranked cross-polarized > non-polarized > parallel-polarized, with the cross
polarized/black-background setup achieving the best performance (R² = 0.61, RMSE = 0.69 mg
100 g⁻¹). This was attributed to the suppression of background reflection interference and
surface glare, enhancing isolation of internal spectral information. Notably, the non
polarized/black-background configuration also yielded favorable performance, suggesting the
potential benefits of integrating both surface and subsurface spectral information. Conversely,
under white background conditions, re-entering reflected light and surface interference were
found to reduce model accuracy. These results demonstrate that optimizing HSI configurations in
accordance with the optical properties of the target sample improves the accuracy of non
destructive biochemical analysis. Configurations that minimize recursive reflection and surface
glare, thereby isolating internal spectral signals derived from internal components, were
validated as most effective. These findings make a significant contribution to the advancement of
sensing technologies in postharvest engineering from general HSI toward application-specific
biochemical assessment.
