Number 34, June 2002

Response to April 2002 WeatherZine Editorial

"Tropical Cyclone Landfall Forecasting: Making Research Relevant"

Powell and Aberson recently considered National Hurricane Center (NHC) forecasts in which the forecast track of the center of a tropical storm or hurricane intersected the coastline of the United States during the period 1976-2000. Their study, and their recent guest editorial on this topic (see Tropical Cyclone Landfall Forecasting: Making Research Relevant, April 2002 WeatherZine), may lead some readers to conclude that the accuracy of NHC forecasts close to the United States has not improved over time. In fact such forecasts have improved.

It had been previously shown that official NHC track forecasts over the period 1970-1998 improved at an annual average rate of 1.0%, 1.7%, and 1.9% for the 24-h, 48-h, and 72-h forecast periods, respectively, for the Atlantic basin as a whole. Powell and Aberson noted that "although these trends [were] promising, neither forecast landfall position nor time error trends [had] been quantified." Investigating these specific parameters (by interpolation from NHC forecasts) at time periods roughly 12, 24, 36, 48, and 60 h prior to landfall, they showed that none of the landfall location error trends, and only the 24 h landfall timing error trend, showed a statistically significant improvement. In the AMS Bulletin article describing their research, they attributed the overall lack of improvement in landfall forecasts to a "conservative least-regret" forecast philosophy for storms threatening to make landfall, or to deficiencies in numerical models or the observing network in the Caribbean and Central America.

One might conclude from their discussion that this apparent lack of improvement in landfall forecasts would also characterize the trend of actual official NHC forecasts for storms near or threatening the coastline, since their suggested explanations should apply equally well to the 12, 24, 36 h, etc., official forecast positions from which the coastline intersections were inferred. Two questions that naturally arise are: (1) what are the long-term trends of NHC forecast errors for storms near the coastline, and (2) are these forecast trends detectably different from basin-wide trends?

Powell and Aberson restricted their analysis to forecast tracks making landfall or passing within 75 km of the coastline. However, this strategy omits cases of keen interest for which a storm is forecast to remain offshore but come close enough to the coast to require the issuance of a tropical cyclone watch or a warning, such as 2001's Hurricane Michelle. Interest on the part of emergency managers and the general public was extremely high for this event, and evacuations were ordered, even though no (United States) landfall was ever forecast.

Our interest, then, is in forecasts issued for storms threatening land, whether or not a landfall is specifically forecast. A simple way to identify such threats is to consider those forecasts issued when watches or warnings (either hurricane or tropical storm) were in effect. Three statistical regressions were performed for the period 1970-2001, one for those forecasts issued when U.S. mainland watches or warnings were in effect, one for those forecasts issued when watches and warnings were not in effect, and one for the entire sample of forecasts. The trend lines at each forecast period indicate that NHC forecasts for storms that threaten land have been improving, with annual average percentage improvements of 0.8%, 1.7%, and 1.9% at 24, 48, and 72 h, respectively. The 24 and 48 h trend lines were significant at or above the 95% level, while the 72 h trend line exceeded the 90% significance level. The improvement trends at 48 and 72 h were about 75%-80% as large as those for the non-threats, while the 24 h trend was about half as large.

Statistical tests confirmed that forecasts for non-land-threatening storms improved more rapidly than those for storms threatening land, at least at 24 and 48 h. Examination of the trend lines showed that forecast accuracy is currently comparable for the threat and non-threat samples, whereas early in the period, non-land-threatening forecasts lagged in accuracy. In our view, changes in observing systems over time, in particular the increasing use of satellite observations in numerical models over data-sparse oceanic regions, could account for the trend more plausibly than a "conservative" forecast philosophy on the part of the NHC.

In terms of forecaster philosophy, data availability, or model performance, there is not likely to be much difference between the sample of forecasts that contain a landfall and those issued when watches and warnings are in effect. Given that NHC forecasts near the coast are improving, what explains Powell and Aberson's assessment that there has been a lack of improvement in landfall forecasts? One possibility is the very small sample that occurs when only the landfall point is considered. For the roughly 24 h forecast period, they verify 129 landfalls during the interval 1976-2000, or an average of just 5 verifications per year. At 48 h, there are fewer than 3 forecasts per year being verified. These numbers represent only about 3% of the total number of forecasts issued, and form an extremely small data set from which to extract long-term trends.

Finally, we disagree with Powell and Aberson's assertion that landfall is the most important part of the forecast, and are troubled that emergency managers are "really paying attention to the landfall forecast." If this is true, then we are failing in our efforts to focus attention away from the precise forecast track of the center. Coastal residents under a hurricane warning are risking their property and lives if they fail to respond adequately because they see an official forecast indicating landfall in some community other than their own. While the timing and location of landfall of the center are important, particularly to the news media, it is not clear that these "forecast" parameters are of any more practical importance than, say, the 24-36 h forecast positions, which help determine the coastal warning zones.

While the most severe hazards generally occur fairly close to the center, dangerous conditions associated with tropical cyclones cover a large area and may last for a day or more. This distribution of hazards is difficult to assess from the official forecast track alone. The mean landfall position error 24 h in advance is about 120 km. This current level of uncertainty in tropical cyclone track forecasts, and the distribution of hazards, dictate that actions to protect life and property should be more closely tied to the threats defined by the tropical storm or hurricane watches and warnings.

James L. Franklin
Hurricane Specialist,
National Hurricane Center
NOAA/NWS/Tropical Prediction Center
James.Franklin@noaa.gov