Nondestructive detection of major components in fresh pork using data fusion of visible-near infrared and short-wave infrared spectra

To achieve rapid non-destructive detection of the main components of pork, a partial least squares regression (PLSR) model was developed to predict the moisture, fat, and protein content in the longissimus dorsi muscle (PLM) using visible near-infrared spectroscopy (VNIR) and short-wave infrared spectroscopy (SWIR) combined with a data fusion strategy. The results showed that the low-level data fusion strategy was most effective for predicting moisture content (R P = 0.971, RMSEP = 0.192), whereas the high-level data fusion strategy demonstrated superior performance in predicting fat content (R P = 0.961, RMSEP = 0.770) and protein content (R P = 0.929, RMSEP = 0.281). Specifically, compared with the best results obtained from single spectral source models, the RPD values for moisture, fat, and protein content models established by data fusion improved by 1.206, 0.333 and 0.300, respectively. In conclusion, combining VNIR and SWIR spectral data with data fusion strategies effectively addressed the limited detection range of a single spectrum and offered a promising alternative to traditional analytical methods for efficiently quantifying the chemical composition in pork. Additionally, it also provided the core algorithm for developing handheld meat analyzers and supported the industrialization of rapid field-testing equipment. • Combining VNIR and SWIR data improves prediction of the main components of PLM. • Three fusion strategies (LLDF/MLDF/HLDF) compared for pork component prediction. • LLDF is best for moisture, HLDF excels for fat/protein prediction in pork analysis. • The potential of data fusion to improve PLSR prediction was demonstrated.