Separation and prioritization of uncertainty sources in a raster based flood inundation model using hierarchical Bayesian model averaging

Abstract

Uncertainty in a hydrodynamic model originates from input data, model structure and parameters. In order to provide the robust model predictions, the Bayesian model averaging (BMA) approach could be used as a multi-model combining method to account for the compound effects from various uncertainty sources. However, BMA cannot provide a clear picture of the impact from individual uncertainty sources. Hierarchical Bayesian model averaging (HBMA) is a recently developed approach to study the relative impact of different uncertainty sources, which explicitly considers various sources of uncertainty in the hierarchical structure (BMA tree) for analysing uncertainty propagation. In this study, HBMA is tested over the Black River watershed in Missouri and Arkansas based on water stage predictions from 243 LISFLOOD-FP model configurations that integrate five sources of uncertainty including channel shape, channel width, channel roughness, topography and flow forcing. To compare, the model perturbation approach is also applied in this study to investigate the influence of individual uncertainty sources on model prediction. Overall, the results indicate that HBMA provides an alternative way for flood modellers to deal with modelling uncertainty in data sparse region when multiple choices of uncertainty sources are considered. Findings from Black River watershed point out that without considering the model weight (model perturbation approach), channel width and topographical data resolution have the largest impact on the hydrodynamic model predictions followed by flow forcing, which has a relatively greater influence than channel cross-sectional shape and model parameter. However, when model weights are taken into account (HBMA), model input (topography and flow forcing) and model parameter have a larger impact on prediction variance than model structure (channel width and cross-sectional shape). Moreover, as results move up the hierarchy along the BMA tree, the accuracy of deterministic mean prediction also increases in general whereas the 95% confidence interval associated with the deterministic mean prediction might become larger.