Near infrared spectroscopy for biomonitoring: cow milk composition measurement in a spectral region from 1,100 to 2,400 nanometers.

Abstract

The potential of near infrared spectroscopy (NIRS; 1,100 to 2,400 nm) to measure fat, total protein, and lactose content of nonhomogenized milk during milking and the influence of individual characteristics of each cow's milk on the accuracy of determination were studied. Milk fractions were taken during milking, twice per month, for 6 mo. Samples were taken every 2nd and 4th wk at the morning and the evening milkings. Teatcups were removed at each 3 L of milk yield as determined with a fractional sampling milk meter. A total of 260 milk samples were collected and analyzed with an NIRSystem 6500 spectrophotometer with 1-mm sample thickness. Partial least squares (PLS) regression was used to develop calibration models for the examined milk components. The comparison with the reference method was based on standard error of cross validation (SECV). The obtained SECV varied from .107 to .138% for fat content, from .092 to .125% for total protein, and from .066 to .096% for lactose content, and the accuracy of the reference method (AOAC, 1990, method No 972.16) was .05% for all measured milk components. The obtained models had lower SECV when an individual cow's spectral data were used for calibration. The reduction of SECV for each cow's individual calibration, when compared with SECV for the set of all samples, differed with the different constituents. For fat content determination, the reduction reached 22.46%, for protein 26.40%, and for lactose 31.25%. This phenomena was investigated and explained by principle component analysis (PCA) and by comparing loading of PLS factors that account for the most spectral variations for each cow and the measured milk components, respectively. The results of this study indicated that NIRS (1,100 to 2,500 nm, 1-mm sample thickness) was satisfactory for nonhomogenized milk compositional analysis of milk fractions taken in the process of milking.