Relationships among ensiling, nutritional, fermentative, microbiological traits and contamination in corn silages addressed with partial least squares regression.

The objective of this work was to reduce the predictor dimensionality and to develop a model able to forecast contamination in corn silages. A survey on 33 dairy farms was performed, and samples from core, lateral, and apical parts of the feed-out face of corn silage bunkers were analyzed for chemical, biological (digestible and indigestible NDF), fermentative (pH, ammonia nitrogen, lactic acid, VFA, and ethanol), and microbiological (yeasts and molds) traits. Corn silage samples were analyzed for cell and spore counts by adoption of a molecular DNA-based method. A partial least squares (PLS) regression with a leave-one-out cross-validation method was used to reduce the dimensionality of the original predictors ( = 30) by projecting the independent variables into latent constructs. In a first step of the model development, the importance of independent variables in predicting contamination was assessed by plotting factor loadings of both dependent and independent variables on the first 2 components and by verifying for each predictor the variable influence on projection values adopting the Wold's criterion as well as the entity of standardized regression coefficients. Three ensiling characteristics (bunker type, presence of lateral wrap plastic film, and penetration resistance as a measurement of the ensiled mass density), a chemical trait (DM), 9 characterizations of the fermentative profile (pH, ammonia nitrogen, acetic acid, butyric acid, isobutyric acid, valeric acid, isovaleric acid, ethanol, and lactic acid), and 2 microbiological traits (yeasts and molds) were retained as important terms in the PLS model. Three reduced-variable PLS (rPLS) regressions-the first based on ensiling, chemical, fermentative, and microbiological retained important variables (rPLSecfm); the second based on chemical, fermentative, and microbiological retained important traits (rPLScfm); and the last based on only chemical and fermentative retained important variables (rPLScf)-were performed. The model that best fit the measurements was rPLSecfm. The rPLScfm and rPLScf models had similar regression performances but higher mean square errors of prediction than rPLSecfm. However, all tested models seemed adequate to rank corn silages for low, medium, and high risks of contamination. To avoid the visit on farm by trained people required to measure penetration resistance, the use of the rPLScf model is suggested as a useful tool to assess the risk of in corn silage.