A Comparison of Bayesian Methods for Uncertainty Analysis in Hydraulic and Hydrodynamic Modeling1. (15th June 2015)

• Record Type: Journal Article
• Authors: Camacho, René A.; Martin, James L.; McAnally, William; Díaz‐Ramirez, Jairo; Rodriguez, Hugo; Sucsy, Peter; Zhang, Song
• Is Part Of: Journal of the American Water Resources Association. Volume 51:Number 5(2015)
• Page Start: 1372
• Page End: 1393
• Publication Date: 2015-06-15
• Subjects: Water-supply -- Periodicals; Hydrology -- Periodicals; Water resources development -- Periodicals; Water resources development -- Environmental aspects -- Periodicals; 333.9100973
• DOI: 10.1111/1752-1688.12319
• Languages: English
• ISSNs: 1093-474X

Abstract

<abstract abstract-type="main" id="jawr12319-abs-0001"> <title>Abstract</title> <p>We evaluate and compare the performance of Bayesian Monte Carlo (BMC), Markov chain Monte Carlo (MCMC), and the Generalized Likelihood Uncertainty Estimation (GLUE) for uncertainty analysis in hydraulic and hydrodynamic modeling (HHM) studies. The methods are evaluated in a synthetic 1D wave routing exercise based on the diffusion wave model, and in a multidimensional hydrodynamic study based on the Environmental Fluid Dynamics Code to simulate estuarine circulation processes in Weeks Bay, Alabama. Results show that BMC and MCMC provide similar estimates of uncertainty. The posterior parameter densities computed by both methods are highly consistent, as well as the calibrated parameter estimates and uncertainty bounds. Although some studies suggest that MCMC is more efficient than BMC, our results did not show a clear difference between the performance of the two methods. This seems to be due to the low number of model parameters typically involved in HHM studies, and the use of the same likelihood function. In fact, for these studies, the implementation of BMC results simpler and provides similar results to MCMC. The results of GLUE are, on the other hand, less consistent to the results of BMC and MCMC in both applications. The posterior probability densities tend to be flat and similar to the uniform priors, which can result in calibrated parameter estimates centered in the parametric space.</p> </abstract>