Evaluation of NWS Multi-sensor Precipitation Estimates for the Southeast United States

Multi-sensor Precipitation Estimates (MPE) are operationally produced by the NWS River Forecast Centers. MPE are based on radar, satellite, and surface gage inputs of precipitation. MPE are used for flash flood forecasts and operational monitoring of precipitation patterns and intensity for locations where no surface gage is present. The available MPE in areas with no surface gages is also used for drought monitoring in this region. However, there is little published data evaluating the accuracy of MPE. An evaluation of MPE of the Southeastern US for the period 2002-2007 as compared to NWS Cooperative (COOP) Observing Network data is presented.

Raw MPE data were compared with data from the closest COOP station in the Southeast for 2002 through 2007 by season and year. While the accuracy of MPE generally improved with each year, and captured the overall precipitation intensity, it is highly variable by season. Summer MPE compared least favorably with COOP observations across the region with an average root mean square error of 0.39 inches, while the winter MPE compared most favorably with an average root mean square error of 0.27 inches.

A second evaluation was performed where the MPE data were interpolated to each COOP station using a bi-linear interpolation scheme. The interpolated MPE were also evaluated by season and year at each COOP station in the Southeast. The interpolated MPE showed the same patterns between season and year, but showed better accuracy as compared to the nearest neighbor MPE point.

A third evaluation of the nearest-neighbor and interpolated MPE data was also performed for different thresholds of observed precipitation. In cases where the observed precipitation was less then 0.1 inches, both MPE methods overestimated the precipitation. In cases with observed precipitation greater than 1 inch, MPE underestimated the precipitation at a given COOP site. For all thresholds, the interpolated MPE method generally performed better than nearest neighbor MPE method across the entire region.