Uncertainty assessment of integrated distributed hydrological models using GLUE with Markov chain Monte Carlo sampling

• Published in: Journal of hydrology (Amsterdam) 2008-05, Vol.353 (1), p.18-32
• Main Authors: Blasone, Roberta-Serena, Madsen, Henrik, Rosbjerg, Dan
• Format: Article
• Language: English
• Subjects: Earth sciences; Earth, ocean, space; Exact sciences and technology; Freshwater; Generalized likelihood uncertainty estimation; hydrologic models; Hydrology; Hydrology. Hydrogeology; Integrated distributed hydrological model; Markov chain; Markov chain Monte Carlo; mathematical models; MIKE-SHE; model validation; Monte Carlo method; Multi-objective calibration; simulation models; Uncertainty assessment; watershed hydrology
• ISSN: 0022-1694; 1879-2707
• DOI: 10.1016/j.jhydrol.2007.12.026

In recent years, there has been an increase in the application of distributed, physically-based and integrated hydrological models. Many questions regarding how to properly calibrate and validate distributed models and assess the uncertainty of the estimated parameters and the spatially-distributed responses are, however, still quite unexplored. Especially for complex models, rigorous parameterization, reduction of the parameter space and use of efficient and effective algorithms are essential to facilitate the calibration process and make it more robust. Moreover, for these models multi-site validation must complement the usual time validation. In this study, we develop, through an application, a comprehensive framework for multi-criteria calibration and uncertainty assessment of distributed physically-based, integrated hydrological models. A revised version of the generalized likelihood uncertainty estimation (GLUE) procedure based on Markov chain Monte Carlo sampling is applied in order to improve the performance of the methodology in estimating parameters and posterior output distributions. The description of the spatial variations of the hydrological processes is accounted for by defining a measure of model performance that includes multiple criteria and spatially-distributed information. An initial sensitivity analysis is conducted on the model to avoid overparameterisation and to increase the robustness of the approach. It is demonstrated that the employed methodology increases the identifiability of the parameters and results in satisfactory multi-variable simulations and uncertainty estimates. However, the parameter uncertainty alone cannot explain the total uncertainty at all the sites, due to limitations in the distributed data included in the model calibration. The study also indicates that properly distributed information of discharge is particularly crucial in model calibration and validation.