Quantifying Extreme Precipitation Forecasting Skill in High-Resolution Models Using Spatial Patterns: A Case Study of the 2016 and 2018 Ellicott City Floods

Recent historic floods in Ellicott City, MD, on 30 July 2016 and 27 May 2018 provide stark examples of the types of floods that are expected to become more frequent due to urbanization and climate change. Given the profound impacts associated with flood disasters, it is crucial to evaluate the capability of state-of-the-art weather models in predicting these hydrometeorological events. This study utilizes an object-based approach to evaluate short range (<12 h) hourly forecast precipitation from the High-Resolution Rapid Refresh (HRRR) versus observations from the National Centers for Environmental Prediction (NCEP) Stage IV precipitation analysis. For both datasets, a binary precipitation field is delineated using thresholds that span trace to extreme precipitation rates. Next, spatial metrics of area, perimeter, solidity, elongation, and fragmentation, as well as centroid positions for the forecast and observed fields are calculated. A Mann–Whitney U-test reveals biases (using a confidence level of 90%) related to the spatial attributes and locations of model forecast precipitation. Results indicate that traditional pixel-based precipitation verification metrics are limited in their ability to quantify and characterize model skill. In contrast, an object-based methodology offers encouraging results in that the HRRR can skillfully predict the extreme precipitation rates that are anticipated with anthropogenic climate change. Yet, there is still room for improvement, since model forecasts of extreme convective rainfall tend to be slightly too numerous and fragmented compared with observations. Lastly, results are sensitive to the HRRR model’s representation of synoptic-scale and mesoscale processes. Therefore, detailed surface analyses and an “ingredients-based” approach should remain central to the process of forecasting excessive rainfall.