Accounting for model structure, parameter and input forcing uncertainty in flood inundation modeling using Bayesian model averaging

Reliability of flood stage and inundation extent predictions are affected by the performance of a hydraulic model. However, uncertainties at all times exist in the model setup process. Therefore, prediction from a single hydraulic model implementation may be subject to huge uncertainty. Bayesian model averaging (BMA) is applied in this study to combine ensemble predictions from different hydraulic model implementations and to develop a robust deterministic water stage prediction as well as the prediction distribution. The BMA approach is tested over the Black River watershed in Missouri and Arkansas based on water stage predictions from 81 LISFLOOD-FP model configurations that integrate four sources of uncertainty including channel shape, channel width, channel roughness and flow input. Model ensemble simulation outputs are trained with observed water stage data during one flood event to obtain the weight and variance for each model member, and BMA prediction ability is then validated for another flood event. The results indicate that the BMA approach is able to provide consistently good and reliable deterministic flood stage prediction across the basin, though it does not always outperform the best model in the ensemble. The BMA water stage prediction has better performance than the ensemble mean prediction. Additionally, high-chance flood inundation extent derived from a BMA probabilistic flood map is more accurate than the probabilistic flood inundation extent based on the equal model weights in the Black River watershed.