Abstract

This paper considers the extent to which Bayesian networks may be utilised to determine plausible suspended sediment budgets under conditions of turbidity sensor failure. Directed acyclic graphs (DAGs) capable of representing the stochastic relationships between discharge and suspended sediment concentrations in both converging and diverging channel settings are presented. Variables within these DAGs are represented using discrete states. The computational procedure for marginalising variables within the DAG to create a Bayesian network is described. The technique allows a probabilistic assessment of the state of any variable conditioned on the array of dependent variables within in the network. Hence, it offers a technique by which values of failed sensors may be inferred from the values of functioning sensors. An example application of the technique is made using data from a proglacial setting in the Ecrins National Park, France and limitations due to the discrete nature of the network are considered.

Keywords: Bayesian network, sediment budget, turbidity, stochastic, random variables