Abstract: The contents of protein and fat in liquid milk were determined by near infrared (NIR) spectroscopy combined with chemometric methods. Four training set and test set samples partitioning methods: random sampling (RS), kennard-stone (KS), Duplex, and sample set partitioning based on joint x-y distance (SPXY) were used and compared. Haaland’s criterion was employed for outlier detection. Eleven spectral pretreatments were used to eliminate the slope-background and noise. Under optimized conditions, the RMSECV, RMSEP, Rc2, and Rp2 of the established model for fat content were 2.434, 2.099, 0.958 and 0.964, respectively. These four parameters of the protein content prediction model were 2.270, 2.564, 0.942, and 0.940, respectively. The results show that this method can provide a useful strategy for quality of liquid milk owing to its accuracy and better generalization.

Key words: milk, near infrared (NIR), set partitioning based on joint x-y distance (SPXY) algorithm, Haaland’s criterion