On the difficulty to optimally implement the Ensemble Kalman filter: An experiment based on many hydrological models and catchments

- Intensive testing of Ensemble Kalman filter tuning is carried out on 20 models.
- Conclusions drawn from single event/catchment/model are likely to be misleading.
- Hydrological models do not equally benefit from EnKF updating.
- The specification of hyper-parameter is unintuitive if the structural error is not explicitly taken into account.
- The identification of the state variables that should be updated is not straightforward.

SummaryForecast reliability and accuracy is a prerequisite for successful hydrological applications. This aim may be attained by using data assimilation techniques such as the popular Ensemble Kalman filter (EnKF). Despite its recognized capacity to enhance forecasting by creating a new set of initial conditions, implementation tests have been mostly carried out with a single model and few catchments leading to case specific conclusions. This paper performs an extensive testing to assess ensemble bias and reliability on 20 conceptual lumped models and 38 catchments in the Province of Québec with perfect meteorological forecast forcing. The study confirms that EnKF is a powerful tool for short range forecasting but also that it requires a more subtle setting than it is frequently recommended. The success of the updating procedure depends to a great extent on the specification of the hyper-parameters. In the implementation of the EnKF, the identification of the hyper-parameters is very unintuitive if the model error is not explicitly accounted for and best estimates of forcing and observation error lead to overconfident forecasts. It is shown that performance are also related to the choice of updated state variables and that all states variables should not systematically be updated. Additionally, the improvement over the open loop scheme depends on the watershed and hydrological model structure, as some models exhibit a poor compatibility with EnKF updating. Thus, it is not possible to conclude in detail on a single ideal manner to identify an optimal implementation; conclusions drawn from a unique event, catchment, or model are likely to be misleading since transferring hyper-parameters from a case to another may be hazardous. Finally, achieving reliability and bias jointly is a daunting challenge as the optimization of one score is done at the cost of the other.