Quantifying Uncertainty in Distributed Flash Flood Forecasting for a Semiarid Region

Abstract:

Reliability of semiarid flood forecasting is affected by several factors, including rainfall forcing, the system input-state-output behavior, initial soil moisture conditions and model parameters and structure. This study employed Bayesian frameworks to enable the explicit description and assessment of parameter and predictive uncertainty for convective rainfall-runoff modeling of a semiarid watershed system in Iran. We examined the performance and uncertainty analysis of a mixed conceptual and physical based rainfall-runoff model (AFFDEF) linked with three Markov chain Monte Carlo (MCMC) samplers: the DiffeRential Evolution Adaptive Metropolis (DREAM), the Shuffled Complex Evolution Metropolis (SCEM-UA), and DREAM- ZS, to forecast four potential semiarid convective events with varying rainfall duration (<24 hrs) and amount (>20 mm). Calibration results demonstrated that model predictive uncertainty was heavily dominated by error and bias in the soil water storage capacity which reflect inadequate representation of the upper soil zone processes by hydrological model. Furthermore, parameters associated with infiltration and interception capacity along with contributing area threshold for digital river network were identified the key model parameters and more influential on the modeled flood hydrograph. In addition, parameter inference in the DREAM model showed a consistent behavior with the priori assumption by closely matching the inferred error distribution to the empirical distribution of the model residual, indicating that model parameters are well identified. DREAM result further revealed that the uncertainty associated with rainfall of lower magnitudes was higher than rainfall of higher magnitudes. Uncertainty quantification of semiarid convective events provided significant insights into the mathematical relationship and characteristics of short-term forecast error and may be applicable to other semiarid watershed systems with the similar rainfall-runoff processes.