Validation of Gridded Precipitation Datasets for Flood Typing in Select Conterminous US Basins

Gridded precipitation datasets are required for flood-typing (i.e., classifying floods by causal mechanisms) historical annual peak streamflow events in basins across the conterminous United States. Selected gridded precipitation data sets were validated during 1981–2013 through comparisons with NOAA Global Historical Climatology Network daily (GHCNd) gauge data. The ability of each gridded data set to capture the spatiotemporal characteristics of daily precipitation, including multiday extremes over six selected basins, was assessed using the Kling–Gupta efficiency metric and its component statistics. Overall, the Parameter-elevation Regression on Independent Slopes Model and Livneh-unsplit were found to best match the spatiotemporal variability of the GHCNd precipitation data, including extremes. The Analysis of Record for Calibration was found to be the third best-performing data set in most basins, except for those in the western United States. The performance of the evaluated reanalysis data sets is poor compared to that of gauge-based data sets, possibly because they are unable to skillfully capture precipitation amounts at the correct location and time. Gauge- and radar-based data sets were found to have relatively small biases (within +/−10% on an annual basis), whereas reanalysis data sets were found to have larger positive apparent biases, especially in winter and spring in most basins. It is possible that the apparent overestimation of winter and spring precipitation in reanalysis data sets reflects snow undercatch at gauges, especially in the central United States. An overall decline in performance for correlation and/or variability was also observed for the summer season compared to other seasons in the reanalysis data sets. Various precipitation data sets might need to be combined for flood typing during different periods from the late 19th century to present. Data sets from different sources have different biases and errors and might have to be homogenized using downscaling and bias-adjustment methods. Alternatively, precipitation thresholds used in some flood-typing schemes might have to be adjusted as a function of time.