Some abstracts do not have video files because ASAS was denied recording rights.
627
A Bayesian approach to unmix diet composition
Accurately measuring diet composition is key to delineating feed efficiency, but under grazing conditions doing so is challenging. Plant-wax markers can be used to estimate the composition of diet mixtures. However, traditional methodologies such as nonnegative least squares (NNLS) ignore variability in and relationships (covariances) between such markers. More recently, bayesian hierarchical models for linear unmixing (BHLU) have been successfully used for estimating mixing proportions in hyperspectral image analysis and could be applied to estimate diet composition. The objective of the present study was to determine the efficiency of BHLU under five scenarios i) BHLU with no covariance and uniform priors (BDU), ii) BHLU with no covariance and gaussian priors (BDG), iii) BHLU with covariances and uniform priors (BCU), iv) BHLU with covariances and gaussian priors (BCG), and for completeness v) NNLS. A simulation study was performed using n-alkane and long-chain alcohol (LCOH) concentrations measured on eight forage species in western rangelands: Bouteloua gracilis, Bromus tectorum, Amorpha canescens, Schizachyrium scoparium, Hesperostipa comata, Bouteloua curtipendula, Melilotus officinalis, and Pascopyrum smithii. Three mixtures were formed, i) the two plants M. officinalis and B. tectorum, ii) the three plants M. officinalis, B. tectorum, and B. gracilisiii) and all eight plants. For the two and three-plant mixtures, proportions were drawn from a Dirichlet distribution. The eight-plant mixture reflected the composition of a field in spring. Covariances between markers were specified based on observed correlations. In each case 900 samples were drawn. Efficiency was assessed through normalized mean squared error (NMSE) and coverage probability (CP). Accounting for covariances between markers increased efficiency of estimation, as shown by lower NMSE and increased CP (Table 1.). Gaussian priors increased errors and reduced coverage. Although not entirely comparable with Bayesian methods, NNLS performed reasonably well in terms of NMSE, but achieved lower CP. In the eight-plant scenario all methods performed poorly, suggesting delineating individual plants grazed in complex swards will be difficult.
Keywords: Diet composition, plant-wax markers, Bayesian inference