A comprehensive sensitivity and uncertainty analysis for discharge and nitrate-nitrogen loads involving multiple discrete model inputs under future changing conditions

Abstract

Abstract. Environmental modeling studies aim to infer the impacts on environmental variables that are caused by natural and human-induced changes in environmental systems. Changes in environmental systems are typically implemented as discrete scenarios in environmental models to simulate environmental variables under changing conditions. The scenario development of a model input usually involves several data sources and perhaps other models, which are potential sources of uncertainty. The setup and the parametrization of the implemented environmental model are additional sources of uncertainty for the simulation of environmental variables. Yet to draw well-informed conclusions from the model simulations it is essential to identify the dominant sources of uncertainty. In impact studies in two Austrian catchments the eco-hydrological model Soil and Water Assessment Tool (SWAT) was applied to simulate discharge and nitrate-nitrogen (NO3--N) loads under future changing conditions. For both catchments the SWAT model was set up with different spatial aggregations. Non-unique model parameter sets were identified that adequately reproduced observations of discharge and NO3--N loads. We developed scenarios of future changes for land use, point source emissions, and climate and implemented the scenario realizations in the different SWAT model setups with different model parametrizations, which resulted in 7000 combinations of scenarios and model setups for both catchments. With all model combinations we simulated daily discharge and NO3--N loads at the catchment outlets. The analysis of the 7000 generated model combinations of both case studies had two main goals: (i) to identify the dominant controls on the simulation of discharge and NO3--N loads in the two case studies and (ii) to assess how the considered inputs control the simulation of discharge and NO3--N loads. To assess the impact of the input scenarios, the model setup, and the parametrization on the simulation of discharge and NO3--N loads, we employed methods of global sensitivity analysis (GSA). The uncertainties in the simulation of discharge and NO3--N loads that resulted from the 7000 SWAT model combinations were evaluated visually. We present approaches for the visualization of the simulation uncertainties that support the diagnosis of how the analyzed inputs affected the simulation of discharge and NO3--N loads. Based on the GSA we identified climate change and the model parametrization as being the most influential model inputs for the simulation of discharge and NO3--N loads in both case studies. In contrast, the impact of the model setup on the simulation of discharge and NO3--N loads was low, and the changes in land use and point source emissions were found to have the lowest impact on the simulated discharge and NO3--N loads. The visual analysis of the uncertainty bands illustrated that the deviations in precipitation of the different climate scenarios to historic records dominated the changes in simulation outputs, while the differences in air temperature showed no considerable impact.

Simultaneously Determining Global Sensitivities of Model Parameters and Model Structure

“…It is notable that several publications that are considering structural sensitivities so far are limited by the disjoint definition of parameter groups (Baroni and Tarantola, 2014;Schürz et al, 2019;Francke et al, 2018;Günther et al, 2019). They would hence show a similar behavior as presented here for the example Baroni method results.…”

Simultaneously determining global sensitivities of model parameters and model structure

“…In this section we briefly describe the Sobol' method (Sobol', 1993) that is traditionally used to derive model parameter sensitivities (Sect. 2.2.1).…”

“…E(f |x ∼i ) is the expected model output when all model parameter except x i are fixed. Sobol' (1993) proposed an elegant and efficient method to approximate the variances V i , V ∼i , and V (f ). We have used the implementations proposed by Cuntz et al (2015) (Appendix D therein).…”

“…A key purpose of model sensitivity analysis is to inform model calibration or model uncertainty analysis so as to focus either of these analyses on only the model inputs/model structural choices the model outputs are most sensitive to. While the literature on sensitivity analysis for model parameters is rich (Morris, 1991;Sobol', 1993;Demaria et al, 2007;Foglia et al, 2009;Campolongo et al, 2011;Rakovec et al, 2014;Pianosi and Wagener, 2015;Cuntz et al, 2015Cuntz et al, , 2016Razavi and Gupta, 2016a, b;Borgonovo et al, 2017;Haghnegahdar et al, 2017), and there has likewise been a good deal of research into the influence of model input uncertainty (Baroni and Tarantola, 2014;Abily et al, 2016;Schürz et al, 2019), sensitivity to model structural choice has received far less attention (McMillan et al, 2010;Clark et al, 2011). With model structure we refer to various process conceptualizations within a model rather than, for example, model discretization.…”

“…Günther et al (2019) applied the discrete value method to determine the sensitivity of a multi-physics snowpack model regarding model parameters, forcing data, and 32 distinct model structures, but individual model parameter sensitivities were not determined. Schürz et al (2019) proposed a comprehensive sensitivity analysis regarding alternative model inputs, climate scenarios, and model setups, where the model setups varied in the number of sub-basins and hydrologic response units (but not process representation). The analysis, based on pre-sampled behavioral parameter sets and 7000 model combinations, could assess the relative impact of the different sources of uncertainty but could not be used to examine the linkages between different types of uncertainty.…”