Constraining soil hydraulic parameter and output uncertainty of the distributed hydrological MIKE SHE model using the GLUE framework

Both calibration and uncertainty assessment are mandatory steps in today's modelling process. The former considers both the inputs (input variables and parameters) as well as model results. An exploratory investigation of the applicable parameter space results in a wide spectrum of values for a specific model output. By retaining only those model realizations that mimic reality in a sufficient way, inputs and associated response can be constrained, thereby quantifying the uncertainty involved. The generalized likelihood uncertainty estimation (GLUE) framework provides a structured methodology for this purpose. The study presented focuses on the applicability of the GLUE framework within the context of the distributed hydrological model MIKE SHE. Even though all significant processes involved are incorporated within the model, the problems of calibration and uncertainty assessment cannot be avoided. This has resulted in a quest for well-delineated effective parameters crucial for sound mechanistic model application. Being a complex model, the number of possible realizations using MIKE SHE is fairly small due to computing time, and an in-depth exploratory approach is impossible. On the other hand, several output variables are available aside from the hydrological river response, like water content in the soil profile or ground water level, which can be used for retaining realistic parameter sets. This study presents some preliminary results on the applicability of the GLUE-MIKE SHE framework and associated constrained preliminary parameter and output uncertainty values. Focus was given to the influence of soil hydraulic parameters on the hydrological behaviour of a small study basin. The soil hydraulic parameters were predicted using various pedo-transfer-function approaches and moisture retention measurements in the laboratory, and these distributions were then confronted with ranges of the constrained effective parameters. For the latter, the restrictions that were imposed based on behavioural acceptance were substantial. All the traditional methods show significant uncertainty, due to heterogeneity and model error, which cannot be disregarded. The first results suggest a reasonable match among effective parameters and laboratory measurements. The use of pedo-transfer-function distributions as effective parameters, however, may provide unacceptable results, even with liberal criteria. Copyright (C) 2002 John Wiley Sons, Ltd.