Reliability of Semiarid Flash Flood Modeling Using Bayesian Framework

Abstract

A case study examining Bayesian techniques for assessing parameter and predictive uncertainty of semiarid flash flood events is<br><br>presented here. The focus is on testing a fully distributed rainfall-runoff model (i.e., AFFDEF) linked with Markov chain Monte Carlo<br><br>(MCMC) samplers to simulate four semiarid flash flood events with varying rainfall durations (&lt;24 h) and amounts (&gt;20 mm). MCMC<br><br>samplers showed consistent behaviors with the a priori assumption and successfully improved performances on complex and multivariate<br><br>search problems of semiarid flood simulation over the Abol-Abbas watershed, Iran. Analysis suggests that parameters associated with infiltration and interception capacity along with the contributing area threshold for the digital river network were the key model parameters<br><br>and were more influential on the shape and volume of the flood hydrograph. Model predictive uncertainty was heavily dominated by error<br><br>and bias in the soil water storage capacity, which reflects inadequate representation of the upper soil zone processes in the AFFDEF distributed model. Overall, the modeling results revealed that a fat-tailed Gaussian distribution using the standard least-squares (SLS) error<br><br>assumption yielded improved estimates of parameter and predictive uncertainty for the semiarid flood events. This case study emphasizes<br><br>the importance of proper statistical representation of the residual error distribution as a basis to improve parameter and predictive<br><br>uncertainty. DOI: 10.1061/(ASCE)HE.1943-5584.0001482.