Multivariate calibration can increase simulated discharge uncertainty

Hydrological models are typically calibrated against discharge data. However, the resulting parameterization does not necessarily lead to a realistic representation of other simulated variables, such as actual evapotranspiration, soil water storage, or total water storage. Since a variety of hydrological variables are now freely available globally, multivariate model calibration has become a popular method to overcome the aforementioned limitation of a discharge-based calibration. Given the improved process representation after multivariate calibration, it seems reasonable to expect that such a calibration also leads to reduced hydrograph uncertainty, associated with more constrained flux maps (i.e., combinations of streamflow generation mechanisms). However, this expectation assumes that an intersection exists within the parameter space between the separate behavioural clouds of the two or more variables considered in multivariate calibration. Here, we tested this assumption in twelve Australian catchments located in five different climate zones. We calibrated the SIMHYD model using a Monte Carlo-based approach in which an initially large sample of parameter sets was constrained using discharge only, actual evapotranspiration only, and a combination of both variables (combined into a single objective function). As could be expected, considering both variables in model calibration resulted in the best overall model performance in all catchments. However, adding actual evapotranspiration to a discharge-based calibration caused hydrograph uncertainty to increase for 11 of the 12 study sites, whereby increases tended to be larger for low flows than high flows. Similarly, flux map areas increased on average by 27% as a result of less constrained streamflow generation mechanisms under multivariate calibration relative to univariate calibration. Analysis of behavioural parameter sets suggests that these symptoms could be caused by non-overlapping behavioural parameter distributions among the different variables. By separately considering both locally observed and remote sensing-based evapotranspiration in the analysis, we could demonstrate that the source of the information did not affect our findings. This has implications both for model parameterization and model selection, emphasising that the value of non-discharge data for improving process representation through calibration is contingent on the suitability of the model structure.