The dilemma of objective function selection for sensitivity and uncertainty analyses of semi-distributed hydrologic models across spatial and temporal scales

• Temporal scale outweighs spatial scale in determining parameter importance. • Objective function (OF) selection may only favor certain processes. • Multi-OFs improve parameter identifiability and reduce uncertainty more than other OFs. • Stricter OF thresholds improve sharpness but reduce reliability of model predictions. • Important parameters should be constrained from coarser to finer temporal scales. Semi-distributed hydrologic models have been extensively utilized for simulating watershed-scale hydrologic processes, given their fast execution time and flexibility in implementing management practices. However, they have many parameters, several of which cannot be accurately measured, leading to equifinality. Sensitivity analysis (SA) is often used to identify the most important parameters and reduce the dimensionality of calibration. Here, we applied a global SA method based on the Variogram Analysis of Response Surfaces (VARS) to investigate parameter importance and the influence of parameter identifiability on streamflow prediction uncertainty. In our experiments, we considered the impacts of space and time scales of interest, the choice of objective functions (single- versus multi-objective), and the direct significance of parameters in driving models, independent of observed data on model outputs (regardless of objective functions). We set up the Soil and Water Assessment Tool (SWAT) for three catchments (< 200 km 2 ) with different physiographic characteristics in the United States to simulate daily streamflow at three spatial discretization scales. SA was performed at daily, monthly, and yearly aggregation scales. Our findings, across all experiments, indicate that the runoff curve number is the most important parameter in SWAT. Results also reveal that parameter importance is largely dependent on the temporal aggregation scale of outputs but almost independent of spatial discretization. Further, we show that the choice of objective function and performance threshold significantly influences the number and configuration of important parameters, highlighting the benefit of incorporating multiple objectives in SA, and their trade-off between improving sharpness and reliability.