Understanding the re-infiltration process to simulating streamflow in North Central Texas using the WRF-hydro modeling system

WRF-Hydro (Weather Research and Forecasting model-Hydrological modeling system), as the core engine of the United States National Water Model (NWM), has now been used in many hydrometeorological applications throughout the world. One important feature that WRF-Hydro introduced is to allow infiltration excess (“ponded water”) for subsequent lateral re-distribution and soil re-infiltration, which is a major enhancement in terms of physical realism. Nonetheless, how well WRF-Hydro models re-infiltration is largely unknown, because this process is difficult to be directly measured. To gain an in-depth understanding of re-infiltration process under different hydrometeorological/geographical conditions with model parameter settings, we start conducting a series of idealized numerical experiments using 18 watersheds in North Central Texas as a testbed. Next, the model is automatically calibrated to best quantify re-infiltration amounts during two major storms (2010 Tropical Storm Hermine and 2015 May Event), which is accomplished by coupling the dynamically dimensioned search (DDS) algorithm with WRF-Hydro to achieve optimal calibration efficiency. The results show that re-infiltration has quite substantial impacts on streamflow simulation in WRF-Hydro, especially for areas with flat terrains and soils with high clay content. Among all examined factors, precipitation, saturated hydraulic conductivity (K_sat) and runoff partition parameter (REFKDT) are found to impose relatively higher impacts on both re-infiltration ratio and runoff coefficient. It is also found that the runoff coefficient and the re-infiltration ratio are positively correlated based on results from both hypothetical and real events, indicating re-infiltration effects can become more pronounced as flood potential increases. These findings collectively show the significance of representing the re-infiltration process in flood forecasting. Models that do not incorporate this process may be over-calibrated to compensate errors originated from the missing process.