Appropriate use of chemometrics for feasibility study for developing low-cost filter-based multi-parameter detection spectroscopic device for meat proximate analysis

Developing filter-based spectroscopic devices is crucial to reduce the instrument's cost for a dedicated application. Selecting informative spectral features (spectral wavelengths or bands) is particularly important to develop a low-cost filter-based spectroscopic device. In this study, only a set of informative bands were selected from NIR spectroscopy to determine proximate compositions (moisture, fat, and protein) in meat. Chemometrics models were developed using partial least squares regression with full spectral range, selected individual informative spectral bands for each attribute, and common informative spectral bands for all attributes. Calibration models developed using common informative spectral bands were superior (moisture: R²_p = 0.97, RMSEP = 1.67%, RPD = 6.07; fat: R²_p = 0.98, RMSEP = 1.99%, RPD = 6.59; protein: R²_p = 0.96, RMSEP = 0.61%, RPD = 4.97) to the individual informative spectral bands (moisture: R²_p = 0.97, RMSEP = 1.80%, RPD = 5.63; fat: R²_p = 0.97, RMSEP = 2.11%, RPD = 6.22; protein: R²_p = 0.95, RMSEP = 0.66%, RPD = 4.59) and full spectral range (moisture: R²_p = 0.96, RMSEP = 1.94%, RPD = 5.22; fat: R²_p = 0.97, RMSEP = 2.09%, RPD = 6.28; protein: R²_p = 0.96, RMSEP = 0.67%, RPD = 4.52). The prediction accuracy using common informative spectral bands for all attributes significantly improved when non-linear calibration was used. The results indicated the efficacy of common informative spectral bands for predicting multiple quality attributes in meat using NIR spectroscopy. These informative bands can be used as a basis for designing and developing a dedicated low-cost and custom-designed device for proximate composition in meat.