Improving flood forecasting using conditional bias-aware assimilation of streamflow observations and dynamic assessment of flow-dependent information content

Abstract

We describe an adaptive extension of the conditional bias-penalized ensemble Kalman filter for conditional bias (CB)-aware data assimilation (DA) and comparatively evaluate with the ensemble Kalman filter (EnKF) for 6 headwater basins in Texas using the operational lumped hydrologic models from the National Weather Service. We then use CB-aware DA and the degrees of freedom for signal to assess the marginal information content of observations. We show that CB arises very frequently in varying magnitudes when assimilating streamflow observations during the catchment’s response to precipitation and subsequent drainage, and that, in general, larger discharges are associated with larger CB. CB-aware DA improves over EnKF by varying margins in times of significant flow, and the improvement is particularly large during sharp rises of the outlet hydrograph with large peak flows. For the 6 study basins, the average relative reduction in root mean square error of the ensemble mean streamflow analysis by CB-aware DA over EnKF is 31.5% for all ranges of observed flow and 32.1% for observed flow exceeding 200 cms. The flow-dependent marginal information content of the observations varies very significantly with the streamflow response of the catchment and the magnitude of CB, and tends to decrease and increase in the rising and falling phases of the hydrograph, respectively. The findings indicate that CB-aware DA with information content analysis offers an objective framework for dynamically balancing the predictive skill of hydrologic models, quality and frequency of hydrologic observation, and scheduling of DA cycles toward improving operational flood forecasting cost-effectively.