Extended Abstract (28K)6A.4
Evaulation of rainfall forecasts from the operational GFDL hurricane model
Robert Tuleya, SAIC/NOAA/NWS/NCEP/EMC, Norfolk, VA; and M. DeMaria and R. J. Kuligowski
The devastating rains and resulting floods of landfalling tropical systems have been quite evident from the recent cases of Floyd (1999) and Allison (2001). Until recently, the primary guidance products were graphical images that did not get much circulation beyond the NHC. Evaluation of low resolution 1° gridded output for US landfalling cases of 1995-1999 indicated that the GFDL model exhibited some degree of skill in forecasting storm total precipitation. The GFDL model has been rerun to include higher resolution output for these 16 cases and these verifications have now been expanded to include nine additional cases spanning from 1995-2002. All cases were initialized at the same 12UTC times as the high density daily River Forecast Center data set. Results indicate a wide spectrum and variation of both qualitative and quantitative skill. Cases with strong synoptic forcing appear to be better forecast by the GFDL model including those of Fran(1996) and Floyd(1999).
One drawback to evaluation of rainfall skill for hurricanes at landfall is the problematic quantitative validation of results. Traditional methods may be ineffective and incomplete. Therefore a straight forward method primarily used so far has been the comparison of model data directly with gauge data at the observation location. Recently a baseline definition of skill has been developed, a rainfall Cliper which uses climatological rainfall rates moved along a forecasted track. Results indicate an overall correlation coefficient of ~0.5 between RFC gauges and GFDL model amounts for the storm total rainfall (primarily 72h) for these 25 cases. A version of rainfall Cliper run up the GFDL forecast tracks yields a correlation coefficient of ~0.33, thus indicating some degree of skill in predicting precipitation distribution. Both the GFDL and rainfall Cliper exhibited a rather large mean error of ~0.9in for the 32430 gauge observations. More traditional precipitation verification scores were also calculated including threat scores. These will also be presented along with some other objective techniques. Problems areas will be pointed out with prospects for improvements discussed.
Session 6A, tropical cyclone rainfall II
Tuesday, 4 May 2004, 10:15 AM-11:45 AM, Le Jardin Room
Previous paper Next paper