Uncertainty of simulated catchment runoff response in the presence of threshold processes:: Role of initial soil moisture and precipitation

This paper examines the effect of spatially variable initial soil moisture and spatially variable precipitation on predictive uncertainty of simulated catchment scale runoff response in the presence of threshold processes. The underlying philosophy is to use a physically based hydrological model named CATFLOW as a virtual landscape, assuming perfect knowledge of the processes. The model, which in particular conceptualizes preferential flow as threshold process, was developed based on intensive process and parameter studies and has already been successfully applied to simulate flow and transport at different scales and catchments. Study area is the intensively investigated Weiherbach catchment. Numerous replicas of spatially variable initial soil moisture or spatially variable precipitation with the same geostatistical properties are conditioned to observed soil moisture and precipitation data and serve as initial and boundary conditions for the model during repeated simulations. The effect of spatially soil moisture on modeling catchment runoff response was found to depend strongly on average saturation of the catchment. Different realizations of initial soil moisture yielded strongly different hydrographs for intermediate initial soil moisture as well as in dry catchment conditions; in other states the effect was found to be much lower. This is clearly because of the threshold nature of preferential flow as well as the threshold nature of Hortonian production of overland flow. It was shown furthermore that the spatial pattern of a key parameter (macroporosity) that determined threshold behavior is of vast importance for the model response. The estimation of these patterns, which is mostly done based on sparse observations and expert knowledge, is a major source for predictive model uncertainty. Finally, it was shown that the usage of biased, i.e. spatially homogenized precipitation, input during parameter estimation yields a biased model structure, which gives poor results when used with highly distributed input. If spatially highly resolved precipitation was used during model parameter estimation. the predictive uncertainty of the model was clearly reduced. (c) 2005 Elsevier B.V. All rights reserved.