Developing Transferable Fourier Transform Mid-Infrared Spectroscopy Predictive Models for Buffalo Milk: A Spatio-Temporal Application Strategy Analysis Across Dairy Farms.

A robust model of buffalo milk based on Fourier Transform Mid-Infrared Spectroscopy (FT-MIRS) is lacking and is difficult to complete quickly. Therefore, this study used 614 milk samples from two buffalo farms from south and central China for FT-MIRS to explore the potential of predicting buffalo milk fat, milk protein, and total solids (TS), providing a rapid detection technology for the determination of buffalo milk composition content. It also explored the rapid transformation and application of the model in spatio-temporal dimensions, providing reference strategies for the rapid application of new models and for the establishment of robust models. Thus, a large number of phenotype data can be provided for buffalo production management and genetic breeding. In this study, models were established by using 12 pre-processing methods, artificial feature selection methods, and partial least squares regression. Among them, a fat model with PLSR + SG (w = 15, = 4) + 302 wave points, a protein model with PLSR + SG (w = 7, = 4) + 333 wave points, and a TS model with PLSR + None + 522 wave points had the optimal prediction performance. Then, the TS model was used to explore the application strategies. In temporal dimensions, the TS model effectively predicted the samples collected in a contemporaneous period (RPD (Relative Analytical Error of Validation Set) = 3.45). In the spatial dimension, at first, the modeling was conducted using the samples from one farm, and afterward, 30-70% of a sample from another farm was added to the debugging model. Then, we found that the predictive ability of the samples from the other farm gradually increased. Therefore, it is possible to predict the composition of buffalo milk based on FT-MIRS. Moreover, when using the two application strategies that predicted contemporaneous samples as the model, and adding 30-70% of the samples from the predicted farm, the model application effect can be improved before the robust model has been fully developed.