Non-destructive discrimination of broiler chicken (Gallus gallus domesticus) feeding regimes using VIS–NIR hyperspectral imaging and machine learning

Food fraud and the lack of reliable dietary traceability systems in meat products represent a growing challenge for food safety, consumer trust, and the competitiveness of the poultry sector, particularly in contexts where verification of animal feeding relies on documentary records that may be prone to error or manipulation. In this framework, this study aimed to evaluate the capability of VIS–NIR hyperspectral imaging combined with machine learning to discriminate feeding regimes in broiler chickens (Gallus gallus domesticus), specifically assessing the influence of anatomical region on classification performance. A controlled experiment was designed using 60 broilers distributed into three contrasting feeding groups, and hyperspectral images were acquired from four anatomical regions of the carcass (breast, tail/uropygial region, thigh, and drumstick/leg). Spectra were preprocessed using Savitzky–Golay filtering and SNV normalization, and several supervised classification models, including linear and non-linear algorithms, were trained under strict individual-wise validation schemes. The results showed that the breast provided the highest discriminative capability, reaching an external accuracy close to 0.98 with an optimized Ridge model, whereas the tail/uropygial region, thigh, and leg exhibited considerably lower performance, reflecting greater structural variability and weaker diet-related spectral signatures. Furthermore, band reduction techniques successfully compressed the input space from 300 to 120 spectral bands without compromising predictive capability. Overall, these findings highlight the strong potential of VIS–NIR hyperspectral imaging as a non-destructive tool to support dietary traceability in chicken meat.