Testing the Adequacy of 7533 KGE Calibrated Conceptual Model Structures

Discussions calling for more rigorous evaluation practices for hydrologic models have recently increased. In addition to the widely used aggregated objective functions, hydrologic signatures are becoming common evaluation metrics for testing the adequacy of hydrologic models for specific application purposes.

This work calibrates 7533 conceptual model structures using KGE as an objective function. These structures are evaluated based on their accuracy (KGE performance) and their adequacy. We defined adequacy as showing less than a +/- 50% percentage bias on inter- and intra-annual flow representation as well as on ten selected signatures. These signatures represent five aspects of the hydrological regime (magnitude, frequency, duration, rate of change, and timing). The large number of model structures, calibrated to the streamflow of 12 hydro-climatically differing MOPEX catchments, allows general insight into how well common conceptual model structures can represent observed hydrological behavior evaluated by signatures.

Results show that a large number of model structures perform accurately (high KGE performance) but almost none of these may be considered adequate (poor signature performance). In nine catchments not a single model can be considered adequate. In the remaining three catchments, only between 1 (0.1%) and 49 (0.7%) of all tested model structures are adequate according to all testing requirements. While inter-annual mean flow representation is typically represented well, the number of models able to represent intra-annual mean flow and/or individual signatures rapidly decreases.

This study presents overwhelming evidence that traditional single-objective function-based calibration is unlikely to return model structures that adequately represent complete hydrologic regimes. We therefore recommend that any model intercomparison or evaluation study needs to be constrained with additional data and/or evaluated by more meaningful metrics than traditional objective functions alone.