Use of a portable NIR instrument as a rapid tool for cattle feeding control.

Avoid feed sorting and providing a homogeneous ratio is a crucial aspect of cattle feeding. We evaluated Near Infra-Red (NIR) as a tool for assessing TMR’s homogeneity (Hi), animals’ sorting (Si) and their relationship with mixer wagon (MW) and ration features. A dataset of 311 TMR samples was split into 55% testing and 45% validation sets. Samples were analyzed for chemical composition, particles size (PS) distribution through a 6-strata sieve and geometric mean length (GML). A portable NIR instrument was calibrated in a range of 902–1660 nm on the testing set by a partial least square (PLS) algorithm and validated on the training set (v). An observational open cohort study of 19 dairy farms was conducted over two years. MW characteristics, TMR formulation, and milk yield were collected through a survey and TMR analysis was performed by NIR. The Hi (range 0–100%, 100 = perfect homogeneity) was calculated as the weighted sum of the standard deviation to mean ratios of TMR composition recorded along the feed alley at 16 points. In these 16 points, the Si (range 0–1, 0 = no selection) was calculated by comparing values of fresh distributed TMR with those collected after 2 h from feed distribution, by a t-test and a weighted coefficient correction. Parameters used for Hi and Si calculation were PS =3.8 and 1.8 mm, pan, GML, crude protein (CP), aNDF, and starch. The median was used as a threshold for Hi and Si binary classification as inhomogeneous (Ibhi, Hi ≤79%) or homogeneous (Hbhi), and as negligible selection (NSbsi ≤0.30) or evident selection (ESbsi). Logistic and linear models assessed the outcomes (Hi–Si) for predictors. ROC curves were drawn. In NIRs calibration, the coefficient of determination of validation (Rv2) was 0.95, 0.74, 0.93, and 0.82 for DM, CP, aNDF, and starch, respectively. The PS traits showed all Rv2 between 0.68 and 0.82. The logistic regression for binary Hi showed odds =1.72 (p = .03) and 0.39 (p = .06), Akaike’s information criterion (AIC) = 16.6 and 20.8, Area under the curve (AUC) = 0.89 and 0.79 for DM and aNDF, respectively. The Hi regressed for MX load fullness with intercept =96.0, coefficient = −0.19, adjusted R2 = 0.374 (p = .02). The logistic regression for binary Si showed odds =2.57 (p = .06), AIC =19.5, AUC =0.82 for aNDF. These findings confirm NIR as a reliable instrument for at farm TMR analysis and suggest a relation between aNDF and both Hi and Si and between DM and MX load with Hi.