Sensitivity and uncertainty analysis of a surface runoff model using ensemble of artificial rainfall experiments

Abstract Surface runoff models are essential for designing water and soil protection measures. However, they often exhibit uncertainty in both parameterization and results. Typically, uncertainty is evaluated by comparing model realizations with measured data. However, this approach is constrained by limited data availability, preventing comprehensive uncertainty assessment. To overcome this limitation, we employed the generalized likelihood uncertainty estimation (GLUE) methodology to conduct sensitivity and uncertainty analyses on a series of surface runoff models. These models were based on an ensemble of artificial rainfall experiments comprising 77 scenarios with similar settings. We utilized the rainfall-runoff-erosion model SMODERP2D to simulate the experiments and employed Differential Evolution, a heuristic optimization method, to generate sets of behavioural models for each experiment. Additionally, we evaluated the sensitivity and uncertainty with respect to two variables; water level and surface runoff. Our results indicate similar sensitivity of water level and surface runoff to most parameters, with a generally high equifinality. The ensemble of models revealed high uncertainty in bare soil models, especially under dry initial soil water conditions where the lag time for runoff onset was the largest (e.g. runoff coefficient ranged between 0–0.8). Conversely, models with wet initial soil water conditions exhibited lower uncertainty compared to those with dry initial soil water content (e.g. runoff coefficient ranged between 0.6 – 1). Models with crop cover showed a multimodal distribution in water flow and volume, possibly due to variations in crop type and growth stages. Therefore, distinguishing these crop properties could reduce uncertainty. Utilizing an ensemble of models for sensitivity and uncertainty analysis demonstrated its potential in identifying sources of uncertainty, thereby enhancing the robustness and generalizability of such analyses.