Flash Flood Forecasting for the Semi-Arid Southwestern United States

Abstract

Flash flooding in the semi-arid United States poses a significant danger to life and property. One effective way to mitigate flood risk is by implementing a rainfall-runoff model in a real-time forecast and warning system. This study investigated the feasibility of using the mechanistic, distributed semi-arid rainfall-runoff model KINEROS2 driven by high resolution radar rainfall input estimates obtained from the NEXRAD WSR-88D DHR reflectivity measurements in such a system. The original procedural paradigm-based KINEROS2 Fortran 77 code with space-time looping was recoded into an object-oriented Fortran 90 code with time-space looping for this purpose. The recoded form is now applicable to large basins, is easily future-extensible, and individual modules can be incorporated into other models.Sources of operational uncertainty in the above system were investigated for their influence over several events within a sub-basin of the USDA-ARS Walnut Gulch Experimental Watershed. Uncertainties considered were in the rainfall estimates, the model parameters, and the initial conditions. The variance-based Sobol' method of global sensitivity analysis conditioned on the observed streamflow showed that the uncertainty in the modeled response was heavily dominated by the operational variability of biases in the radar rainfall depth estimates. Sensitivities to KINEROS2 parameters indicates the need for improved representation of semi-arid hillslope hydrology in small basins, while pointing to specific influential, but poorly identified model parameters towards which field investigations should be directed. The significant influence of initial hillslope soil moisture showed the requirement of a sophisticated inter-storm model component for a continuous forecasting model.A synthetic study data was used to further explore the phenomena seen in the above real data study, of behavioral modifier set inconsistency across all events and of irreducibility in the spatial modifier ranges. The former was found to be attributable to wide uncertainty ranges in the sources of uncertainty, and the latter to the high distributed model non-linearity with associated interactions. These contribute towards a high predictive uncertainty in operational forecasting.Overall, the GLUE-based predictive uncertainty method with behavioral classification and accommodation of wide operational source uncertainty ranges is recommended as a simple and effective setup for operational flash flood forecasting.