On the Prediction of Uncertainty in a Sediment Provenance Model

A Bayesian framework is created to tackle uncertainty in sediment provenance or fingerprint model. Two sources of uncertainty have been identified. One is from the physically based watershed sediment yield model runs due to spatial variation in the ground slopes. The other is from the use of long-range simulated episodic rainfall time series modeled using Markov Chains. The paper extends the Bayesian Markov Chain Monte Carlo (MCMC) algorithms of Fox and Papanicolaou (2008) for ensemble prediction of soil yield fraction or percentage from the floodplains adjacent to a stream, and the associated uncertainty. An erosion process parameter is the uncertain parameter of focus in this study because of its direct link with the physically based water erosion model. This link is identified in the Bayesian MCMC simulation runs. The study finds that less uncertainty is associated with sediment fraction yield estimation with increasing number of spatially distributed soil δ13C carbon isotope and Carbon/Nitrogen (C/N) atomic ratio tracer data, and low-range episodic rainfall time series when the Bayesian MCMC method is used. The drawbacks of the frequentist Monte Carlo Simulation method are discussed. The work compliments that of Fox and Papanicolaou (2008) via the introduction of prediction parameter uncertainty comparison based on the two aforementioned methods.