Prediction of milk mid-infrared spectrum using mixed test-day models

Mid-infrared (MIR) analysis of milk currently allows the measurement of many variables of interest for the dairy sector related to milk nutritional quality, milk technological properties, cow’s status or environmental fingerprint. The aim of this study was to explore the ability of a test-day model to predict milk MIR spectra, and therefore all the resultant variables, for a future test day of a known cow or for a new cow based on easily known characteristics of cows. This is useful for instance for herd management (e.g. detecting problems, predicting potential of heifers) or to predict future environmental impacts of a dairy herd. A total of 467,496 milk MIR spectra from 53,781 Holstein cows in first lactation were used for the calibration data set. First, 323 wavelengths out of the 1,060 wavelengths of the milk spectra were conserved. This spectral information was reduced by using principal component analysis (PCA). A total of 8 principal components (PC) were kept, representing 99% of the spectral information. Then 8 univariate test-day models including the day in milk, herd×year and herd×month as fixed effects and herd×test date, permanent environment and genetics as random effects were applied for each PC. From the solutions of the models and by using a back reversing operation using eigenvectors of the PCA, the predicted 323 wavelengths of the spectra were re-obtained. The calibration correlations between observed and predicted spectral data ranged from 0.76 to 0.93. Correlations between observed and predicted milk fat and protein contents obtained from the modelled spectra were 0.83 and 0.89, respectively. These findings demonstrate the moderate ability of a test-day model to predict milk MIR spectra.