How well can hydrological models simulate drought-to-flood transitions?

The impacts of floods can be enhanced when they occur shortly after droughts. Although these types of events have been widely studied separately, it is yet unclear to what extent hydrological models can capture these drought-to-flood transitions and what are the most important modeling decisions to achieve accurate simulations. To address this research gap, we calibrated four conceptual bucket-style hydrological models (GR4J, GR5J, GR6J, and TUW) for 63 catchments in Chile and Switzerland. We assessed the relative importance of different methodological choices, including model structure and calibration metric (based on the Kling-Gupta efficiency - KGE), on the model's capability to capture streamflow transitions. Further, we explored the link between the detection of transitions and the representation of different processes (e.g., snow, soil moisture, and evaporation) during these events. Our results show that i) a satisfactory KGE does not guarantee a good performance in terms of detecting streamflow extremes, ii) the choice of model structure and catchment characteristics play a relatively more important role in the model’s capability to capture transitions (compared to calibration metrics), and iii) the detection of streamflow extremes and transitions primarily depends on streamflow timing rather than other hydrological signatures or variables (e.g., evapotranspiration, snow water equivalent, etc.). We conclude that the model’s capability to simulate transitions depends on how well the streamflow response to high snowmelt or precipitation rates is represented. We showed that a model that adequately simulates individual drought and flood events does not necessarily capture observed transitions. Based on our results, we hypothesize that parsimonious models such as GR4J seem to be more suitable for simulating drought-to-flood transitions. Finally, our work highlights the importance of assessing the model’s ability to detect and simulate streamflow extreme transitions, and not purely relying on the overall model performance retrieved from the calibration or verification period.