Improving NIRS predictions of ingredient composition in compound feedingstuffs using Bayesian non-parametric calibrations

The selection of an appropriate algorithm for modelling the relationship between spectral and reference data is one of the factors that influence the accuracy of predictions using near infrared (NIR) spectroscopy. For the prediction of ingredient composition in compound feeds, several modelling approaches, both linear and non-linear, have shown the possibilities of NIR technology although with some limitations. The present work evaluates the use of Bayesian non-parametric calibrations for the prediction of 26 ingredients, with different nature, distribution and behaviour. One advantage of this approach is that it avoids the bias towards the mean of the reference data that is not appropriate for the non-Gaussian data distributions of the ingredients of compound feeds. Another is the provision of confidence intervals that quantify the uncertainty in individual predictions. These confidence intervals were used in the tuning process, to help avoid over fitting. The results obtained with the Bayes method for most of the ingredients were similar to those obtained by local regression, both methods improving on the linear approach by around 50%. However, unlike the local regression, the Bayes approach predicted all samples and moreover provided prediction intervals that appear to be very realistic for this dataset.

► To improve prediction ability of ingredient composition in compound feeds in a Bayesian framework.
► A critical challenge is the improvement of the Bayesian approach used for calculating a confidence interval to represent the uncertainty in the predictions.
► Similar results were obtained in comparison with other non linear methods as LOCAL algorithm.
► Bayes' method provides prediction intervals that appear to be very realistic.