Infrared spectroscopy for farming system classification and beef quality assessment: a multi device chemometric approach

One of the aims of the EU INTAQT project was to develop innovative and cost‑efficient analytical tools for the rapid prediction of intrinsic quality traits and the authentication of husbandry systems across Europe. We evaluated and compared the performance of two portable spectroscopic devices, the IndiGo Vis (visible region, 400-700 nm) and the FlameNIR (NIR region, 935-1665 nm), with that of a benchtop reference instrument (NIRS™ DS2500, combined Vis/NIR region) for authentication and quality assessment in the beef production chain. The first objective of this study was to classify beef samples according to the intensification level of the farming system, previously defined based on farming input data. For this purpose, we used a linear discriminant analysis with previous variable selection algorithm based on covariance among wavelengths. The IndiGo Vis and FlameNIR devices showed classification abilities comparable to the NIRS™ DS2500, achieving testing accuracies of 0.72-0.78 when considering the farm as the experimental unit and 0.82-0.90 when considering the animal as the experimental unit. The second objective was to predict a wide range of beef biochemical quality attributes, including proximate composition, fatty acids profile, oxidation products, B-group vitamins (and others), using partial least squares regression. In this context, the NIRS™ DS2500 achieved the highest overall performance (R² = 0.65-0.93, depending on the trait), whereas the portable devices generally produced lower prediction accuracy. Among the latter, only the IndiGo Vis yielded reliable prediction models for total and heme iron (R² = 0.69 and 0.77, respectively), likely because these traits are linked to fresh meat colour and therefore detectable within the visible spectral region. As part of this objective, we also compared spectral data collected on pooled samples (Pooled) with those obtained on individual samples and subsequently averaged (Ind_Aver). Prediction accuracy differed significantly between the two strategies (p<0.001), with consistently higher R² values for Ind_Aver (0.65) compared to Pooled (0.22) across all traits. The third objective was to predict beef carcass characteristics measured at the slaughterhouse, such as pH and marbling score, plus meat tenderness (shear force) and sensory attributes. Although the NIRS™ DS2500 again outperformed the portable devices, the resulting prediction equations were unsatisfactory (R² < 0.30), regardless of the instrument used. However, some classification approaches might provide more robust results than exact value prediction (analysis ongoing). The studies showed that the tested portable devices could support specific applications for authentication and for prediction of some biochemical traits, offering promising perspectives for practical field use. Nevertheless, developing robust calibrations for a broader set of quality traits remains challenging and further methodological refinement will be necessary to improve prediction accuracy and expand the applicability of portable spectroscopic tools.