10.4122/1.1000000370
Blasone, Roberta-Serena
Roberta-Serena
Blasone
rsb@er.dtu.dk
Vrugt, Jasper A.
Jasper A.
Vrugt
vrugt@lanl.gov
Madsen, Henrik
Henrik
Madsen
hem@dhi.dk
Rosbjerg, Dan
Dan
Rosbjerg
dr@er.dtu.dk
Blasone, Roberta-Serena
Roberta-Serena
Blasone
rsb@er.dtu.dk
Uncertainty Assessment in Hydrologic Modeling: Comparison of GLUE and Markov Chain Monte Carlo methods
XVI International Conference on Computational Methods in Water Resources
2006
2006
In recent years, predictive uncertainty analysis in hydrologic modeling has become
an active area of research. Many different methods have been proposed to obtain
meaningful confidence intervals on the model predictions. These methods rely on
different assumptions on how input, output, parameter and model structural error
are made explicit, and therefore generate different output prediction uncertainty
ranges. One of the most commonly used methods to quantify output uncertainty is the
Generalized Likelihood Uncertainty Estimation technique, GLUE advocated by Beven
and coworkers (1992). Despite its features of generality and ease of
implementation, the main drawbacks of GLUE are the high level of subjectivity in
determining the threshold cutoff value and likelihood measure, and the high
computational cost of its implementation. Recently, Vrugt et al. (2005) proposed
the Shuffled Complex Evolution Metropolis (SCEM-UA) algorithm for computationally
efficient estimation of the high probability density region (HPDF) of the parameter
space. Contrary to GLUE, the SCEM-UA method is a Markov Chain Monte Carlo sampler
that periodically updates the proposal distribution of the parameters to converge
to the HDPF in the parameter space. In this study we compare the efficiency and
effectiveness of GLUE and SCEM-UA for parameter uncertainty assessment in
hydrologic modeling. Both methods are applied to a set of increasingly complex
conceptual watershed models. The results demonstrate that the adaptive nature of
the SCEM-UA method significantly reduces the computational burden to obtain a
behavioral sample set of points from the HPDF of the parameter space.