In this study, we perform a data assimilation (DA) experiment on a very large number (>700) of small‐ and medium‐scale (150–10,000 km²) European catchments to assess the impact of satellite soil moisture (SM) DA on streamflow simulations for different climatic and hydrologic conditions. In the experiment, Climate Change Initiative SM active, passive and combined products are assimilated over a time period 2003–2016 via an Ensemble Kalman Filter (EnKF). The results show that, on average, the assimilation of the three products provides relatively small improvements as compared to analogous open loop (OL) results (i.e., with an increase on median Kling‐Gupta Efficiency equal to 0.0048, 0.0033, and 0.0022 [−] for the active, the passive, and the combined products, respectively). OL performance itself is found to be a significant driver of the assimilation results: greater improvements are observed in catchments with poor OL streamflow predictions and inaccurate precipitation estimates. The remotely sensed product accuracy also emerges as relevant for assimilation efficiency, while factors affecting SM retrievals such as vegetation density, topographic complexity and basin area are found to have only a limited impact on the spatial pattern of performance. Small and detrimental effects of SM assimilation are observed over fully humid catchments and at high latitudes where pre‐storm soil moisture has reduced control on runoff generation as well as in basins where the hydrological model contains structural limitations.

Although not matching the formal definition of the predictive probability distribution, meteorological and hydrological ensembles have been frequently interpreted and directly used to assess flood‐forecasting predictive uncertainty. With the objective of correctly assessing the predictive probability of floods, this paper introduces ways of taking into account the measures of uncertainty provided in the form of ensemble forecasts by modifying a number of well‐established uncertainty postprocessors, such as Bayesian Model Averaging and Model Conditional Processor. The uncertainty postprocessors were developed on the assumption that the future unknown quantity (predictand) is uncertain while model forecasts (predictors) are given, which imply that they are perfectly known. With this in mind, we propose to relax this assumption by considering ensemble predictions, in analogy to measurement errors, as expressions of errors in model predictions to be integrated in the postprocessors coefficients estimation process. The analyses of the methodologies proposed in this work are conducted on a real case study based on meteorological ensemble predictions for the Po River at Pontelagoscuro in Italy. After showing how improper can be the direct use of ensemble predictions to describe the predictive probability distribution, results from the modified postprocessors are compared and discussed.