Evaluation of GFDL and Simple Statistical Model Rainfall Forecasts for U.S. Landfalling Tropical Storms

To date, little objective verification has been performed for rainfall predictions from numerical forecasts of landfalling tropical cyclones. Until 2001, digital output from the operational version of the Geophysical Fluid Dynamics Laboratory (GFDL) hurricane forecast model was available only on a 1° grid. The GFDL model was rerun or reanalyzed for 25 U.S. landfalling tropical cyclones from 1995 to 2002 to obtain higher resolution (1/3°) output. Several measures of forecast quality were used to evaluate the predicted rainfall from these runs, using daily rain gauge data as ground truth. The overall quality was measured by the mean error and bias averaged over all the gauge sites. An estimate of the quality of the forecasted pattern was obtained through the correlation coefficient of the model versus gauge values. In addition, more traditional precipitation verification scores were calculated including equitable threat and bias scores. To evaluate the skill of the rainfall forecasts, a simple rainfall climatology and persistence (R-CLIPER) model was developed, where a climatological rainfall rate is accumulated along either the forecasted or observed storm track. Results show that the R-CLIPER and GFDL forecasts had comparable mean absolute errors of 0.9 in. (23 mm) for the 25 cases. The GFDL model exhibited a higher pattern correlation with observations than R-CLIPER, but still only explained 30% of the spatial variance. The GFDL model also had higher equitable threat scores than R-CLIPER, partially because of a low bias of R-CLIPER for rainfall amounts larger than 0.5 in. (13 mm). A large case-to-case variability was found that was dependent on both synoptic conditions and track error.