Improving Flood Forecasting Through Dynamic Assessment of Flow-Dependent Information Content and Conditional Bias-Informed Data Assimilation

Abstract:

To combine information in observation and model prediction optimally for real-time hydrologic forecasting, it is necessary to capture the flow-dependent information content therein. For example, a streamflow observation at the base of a rising limb has a much larger marginal value for flood forecasting than one at the lower recession limb. Because floods are associated with extreme states of the hydrologic system and their transitions, data assimilation (DA) for flood forecasting is generally subject to flow-dependent conditional bias (CB). Conventional DA methods, however, do not account for CB and hence are not likely to produce high-quality solutions in times of significant CB. In this work, we quantify the flow-dependent information content in observation and model prediction using CB-informed DA and the degrees of freedom for signal (DFS). The DFS serves as a skill score for information fusion and hence allows assessment of the marginal value of observation and its flow dependence given the hydrologic model used. We then comparatively evaluate CB-informed DA with the widely used ensemble Kalman filter. The effectiveness of DA depends strongly on the quality of the model and its parameters. If there are significant structural or parametric errors, one may expect strongly-constrained DA to inform their possible sources. In this presentation, we share the findings from twin experiments in which strongly- and weakly-constrained DA are carried out under identical conditions for a number of headwater basins in Texas, and the results from DA-informed adjustment of selected model parameters.