Application of an Ensemble Modeling Approach for Assimilating Observations to Improve Hydrologic and Streamflow Predictions

Most methodological studies supporting operational streamflow prediction have focused on the performance of new techniques in individual watersheds. Yet streamflow prediction is increasingly being performed in a large-domain context, where the applicability and efficacy of watershed-specific methods are less well understood. The move toward large-domain, higher resolution forecasting systems in a number of nations compels a new focus on applicability of ensemble methods in such regional or national systems, which may lack the agility and/or observational density of watershed-specific systems. In this presentation, we discuss practical challenges in this area and present findings from an effort to develop a real-time regional-scale system for medium-range and seasonal ensemble streamflow prediction. In particular, we use ensembles of hydrologic states and fluxes driven by meteorological forcing ensembles to assimilate watershed observations of streamflow and snow to improve forecast initialization. The data assimilation technique is the sequential importance resampling particle filter, and the forecast workflow also includes the downscaling of meteorological forecast ensembles, land surface modeling with a HUC12-based implementation of the SUMMA modeling framework, and streamflow forecast post-processing. This presentation describes the performance of ensemble flow forecasts produced using the NCAR System for Hydromet Analysis, Research and Prediction (SHARP) in comparison with operational forecasts from other sources. We also discuss key challenges inherent in the ensemble forecast approach, and more generally the potential of ensemble data assimilation for integrating Earth observations into model-based prediction systems.