Multivariate calibration and regionalization of a conceptual hydrological model using streamflow and groundwater level

Hydrological models are an important tool in understanding the complex interactions of a catchment's water balance and for supporting water resource management. The widespread practice of calibrating these models is based on streamflow, which causes problems like inaccurate representation of other important fluxes and equifinality, where different parameter sets yield similar modelling results. These problems reduce the model interpretability, robustness, and propagate uncertainty in processes like regionalization.

Using 935 German catchments from the CAMELS-DE dataset, supported by groundwater level records and hydrogeological descriptors, we compared univariate (streamflow-only) and multivariate (streamflow + groundwater) calibration strategies. Several groundwater representation approaches and objective functions were tested. Correlation-based evaluation of groundwater storage outperformed bias-insensitive KGE, yielding higher median streamflow KGE values during calibration (0.75 vs. 0.71) and validation (0.64 vs. 0.60), confirming groundwater levels as reliable indicators of groundwater storage. Hydrogeological characteristics also showed a strong influence on model performance.

When the resulting parameter sets were used in the PASS regionalization framework, both models were found to reduce the equifinality. The multivariate-based regionalization performed significantly better, even with parameters that showed greater variability during the calibration phase. We also found that low-land catchments showed lower model efficiency during local calibration phase. During regionalization, we find the slow-draining porous catchments to show greater variability for the nonlinear parameter of groundwater β_GW for the multivariate model in comparison to the univariate model. Overall, the approach underscores the importance of having additional constrained to improve the physical interpretability of the model and reduce the uncertainty and equifinality of produced parameter sets.