Intercomparison of lumped versus distributed hydrologic model ensemble simulations on operational forecast scales

Summary Distributed hydrologic models, with the capability to incorporate a variety of spatially-varying land characteristics and precipitation forcing data, are thought to have great potential for improving hydrologic forecasting. However, uncertainty in the high resolution estimates of precipitation and model parameters may diminish potential gains in prediction accuracy achieved by accounting for the inherent spatial variability. This paper develops a probabilistic methodology for comparing ensemble streamflow simulations from hydrologic models with high- and low-spatial resolution under uncertainty in both precipitation input and model parameters. The methodology produces ensemble streamflow simulations using well calibrated hydrologic models, and evaluates the distinctiveness of the ensembles from the high- and low-resolution models for the same simulation point. The study watersheds are of the scale for which operational streamflow forecasts are issued (order of a few 1000 km2), and the models employed are adaptations of operational models used by the US National Weather Service. A high-resolution (i.e., spatially distributed) model and a low-resolution (i.e., spatially lumped) model were used to simulate selected events for each of two study watersheds located in the southern Central Plains of the US using operational-quality data to drive the models. Ensemble streamflow simulations were generated within a Monte Carlo framework using models for uncertainty in radar-based precipitation estimates and in the hydrologic soil model parameters. The Kolmogorov–Smirnov test was then employed to assess whether the ensemble flow simulations at the time of observed peak flow from the high- and low-resolution models can be distinguished with high confidence. Further assessment evaluated the model performance in terms of reproducing the observed peak flow magnitude and timing. Most of the selected events showed the high- and low-resolution models produced statistically different flow ensembles for the peak flow. Furthermore, the high-resolution model ensemble simulations consistently had higher frequency of occurrence within specified bounds of the observed peak flow magnitude and timing.