Benchmarking high-resolution hydrologic model performance of long-term retrospective streamflow simulations in the contiguous United States

Abstract. Because use of high-resolution hydrologic models is becoming more widespread and estimates are made over large domains, there is a pressing need for systematic evaluation of their performance. Most evaluation efforts to date have focused on smaller basins that have been relatively undisturbed by human activity, but there is also a need to benchmark model performance more comprehensively, including basins impacted by human activities. This study benchmarks the long-term performance of two process-oriented, high-resolution, continental-scale hydrologic models that have been developed to assess water availability and risks in the United States (US): the National Water Model v2.1 application of WRF-Hydro (NWMv2.1) and the National Hydrologic Model v1.0 application of the Precipitation–Runoff Modeling System (NHMv1.0). The evaluation is performed on 5390 streamflow gages from 1983 to 2016 (∼ 33 years) at a daily time step, including both natural and human-impacted catchments, representing one of the most comprehensive evaluations over the contiguous US. Using the Kling–Gupta efficiency as the main evaluation metric, the models are compared against a climatological benchmark that accounts for seasonality. Overall, the model applications show similar performance, with better performance in minimally disturbed basins than in those impacted by human activities. Relative regional differences are also similar: the best performance is found in the Northeast, followed by the Southeast, and generally worse performance is found in the Central and West areas. For both models, about 80 % of the sites exceed the seasonal climatological benchmark. Basins that do not exceed the climatological benchmark are further scrutinized to provide model diagnostics for each application. Using the underperforming subset, both models tend to overestimate streamflow volumes in the West, which could be attributed to not accounting for human activities, such as active management. Both models underestimate flow variability, especially the highest flows; this was more pronounced for NHMv1.0. Low flows tended to be overestimated by NWMv2.1, whereas there were both over and underestimations for NHMv1.0, but they were less severe. Although this study focused on model diagnostics for underperforming sites based on the seasonal climatological benchmark, metrics for all sites for both model applications are openly available online.