The GLUE Methodology for Model Calibration with Uncertainty

Abstract

The Origins of GLUEThe Generalised Likelihood Uncertainty Estimation (GLUE) methodology was first introduced by Beven and Binley (1992) as a way of trying to assess the uncertainty in model predictions when it is difficult to formulate an appropriate statistical error model. It was effectively an extension of the Generalised Sensitivity Analysis (GSA) approach of Hornberger and Spear (1981) which uses many model runs with Monte Carlo realisations of sets of parameters and criteria to differentiate between models that are considered acceptable (behavioural models) and those that are not (non-behavioural models). Where the posterior parameter distributions of the behavioural and non-behavioural sets are strongly differentiated, it suggests that the results are sensitive to a particular parameter. Where they are not, it suggests insensitivity. The results, of course, depend on choices of the prior ranges of the parameters and the criteria used to separate the two sets of models.If a model is considered behavioural then we expect that the predictions of that model will have some utility in practice. But the multiple behavioural models, even after conditioning on some calibration data in this way, might well give widely different predictions. Beven (1993Beven ( , 2006aBeven ( , 2009 later called this the equifinality problem, chosen to indicate that there may be many ways of reproducing the calibration data acceptably 87 Applied Uncertainty Analysis for Flood Risk Management Downloaded from www.worldscientific.com by KAINAN UNIVERSITY on 02/07/15. For personal use only.

Neglecting Model Parametric Uncertainty Can Drastically Underestimate Flood Risks

“…Hydrologic models are commonly used to understand hydrological processes, predict the response of hydrological systems to changing stresses, and provide boundary conditions to estimate flood hazards and risks (Bates et al., 2021; Judi et al., 2018; Koren et al., 2004; Rajib et al., 2020). However, hydrologic projections are subject to uncertainties such as from model structures, parameters and forcings (Beven, 2014; Fisher & Koven, 2020; Gupta et al., 2012; Hu et al., 2019; Kavetski et al., 2006; Mendoza et al., 2015). Parametric uncertainty can arise, for example, from the epistemic uncertainties about model parameters (Vrugt et al., 2003), the associated prior distributions (Tang et al., 2016), spatial‐resolutions and objective functions (Melsen et al., 2019), and different choices of calibration approaches (Kavetski et al., 2018).…”

Neglecting uncertainties surrounding model parameters can drastically underestimate flood risks

“…Hydrologic models are commonly used to understand hydrological processes, predict the response of hydrological systems to changing stresses, and provide boundary conditions to estimate flood hazards and risks (Bates et al, 2021;Brunner et al, 2020;Judi et al, 2018;Koren et al, 2004;Rajib et al, 2020;Thorstensen et al, 2016). However, hydrologic projections are subject to uncertainties such as from model structures, parameters and forcings (Gupta et al, 2012;Kavetski et al, 2006;Beven, 2014;Fisher & Koven, 2020;Hu et al, 2019;Mendoza et al, 2015).…”