Which model do you trust? And, Arctic sea ice reaches a record minimum

Hurricane Dean, now a major Category 3 hurricane with 125 mph winds, continues to intensify and grow larger in size. Dean pounded Martinique and St. Lucia this morning, and claimed its first victim when a 62-year old man died on St. Lucia while trying to save his cow from raging flood waters.

Dean's eye is now visible on long range radar out of Puerto Rico. Buoy 42059 is in Dean's path, and should be interesting to watch.

We're fairly confident of the 1-2 day forecast, which has Dean headed west to west-northwest over the Central Caribbean, very close to Jamaica and the Cayman Islands, then into the Western Caribbean. After that, things become murkier. The latest 12Z runs of the NOGAPS, UKMET, GFS, and HWRF computer models all show Dean hitting the Yucatan Peninsula, and continuing on into the Gulf of Mexico towards a second landfall near or south of the Texas border. The HWRF run is slower, and does not take Dean to the coast at the end of its forecast period. The big outlier is the GFDL model, which now takes Dean northwest into central Louisiana. Which model is correct? The problem is that each model has a different solution for the behavior of an upper-level low pressure system expected to be over the Gulf of Mexico early next week. Which model should we trust?

In 2006, the official NHC forecast performed better than any of the individual computer forecast models. However, several "consensus" forecasts made using an average of the "big four" computer models (GFDL, GFS, UKMET, and NOGAPS), slightly outperformed the official forecast at some time periods (Figure 1). The Florida State Super-Ensemble (FSSE), for example, combines the "big four" models on the basis of past performance in an attempt to correct for biases in those models. (The FSSE is owned by a private company, which makes it available to NHC but not the general public). The Florida State Super Ensemble slightly out-performed the official NHC forecast at most forecast times.

The "big four" models are plotted on wunderground.com's computer model page for Dean, (along with the inferior BAMM model, which is plotted since it is always available quickly, and has done well at longer range forecasts in the past). We do not get tracking points for the ECMWF or HWRF models at this point, so you'll have to go the raw plots to see those forecasts. Note that three of the "big four" models performed well in 2006, with the GFDL and GFS performing the best. The UKMET had a very poor showing in the Atlantic in 2006. However, the UKMET was the best-performing model in the Eastern Pacific in 2006, along with the GFDL and BAMM models.

The European Center's model (ECMWF) outperformed the "big four" consensus models for 72, 96, and 120 hours forecasts in the Atlantic. However, the ECMWF model was generally not available in time to be used by forecasters. Efforts are being made to make the ECMWF available in a more timely fashion for the 2007 season, which would be a big help. We also have the new HWRF (Hurricane Weather Research Forecast) model this year. In tests done on a number of hurricanes for past years, the HWRF performed about as well as the GFDL (Figure 2).



Figure 1. Track forecast skill in 2006 of the official forecast and the various models, compared to a "zero skill" forecast using NHC's CLIPER5 model. The CLIPER model (short for CLImatology and PERsistence) is a model that makes a forecast based on historical paths hurricane have taken, along with the fact that hurricanes tend to keep moving in the direction they are going (i.e., their current motion persists). Note that many models had a negative skill for their 120 hour (5 day) forecast. The official NHC forecast had about 10% skill at 5 days. Image credit: NHC.

Figure 2. Track errors for 48-hour forecasts from the 2006 version of the GFDL model (black) and the new HWRF model (red). The HWRF model performed better on some hurricane than the GFDL, and worse on others. Overall, the two models had about the same performance on the cases tested. Image credit: Naomi Surgi, NOAA Environmental Modeling Center.

In conclusion, the official NHC forecast outperforms all the individual models, particularly at long ranges. Looking at the individual model plots can be helpful to determine the uncertainty in the forecast, but it's tough to beat the NHC. In the case of Dean, where one model is an outlier from the rest, it is usually better to believe the consensus of the other models.

If you want to look at plots of the individual models, I've written a description of the various models and where to find these plots on our tropical weather page.

Arctic sea ice shrinks to record low
The National Snow and Ice Data Center announced today that Arctic sea ice has just surpassed the previous single-day (absolute minimum) record for the lowest extent ever measured by satellite. Satellite measurements began in 1979. Sea ice extent has fallen below the 2005 record low absolute minimum and is still melting. Sea ice extent is currently tracking at 5.26 million square kilometers (2.02 million square miles), just below the 2005 record absolute minimum of 5.32 million square kilometers (2.05 million square miles). This new record was set a full five weeks before the usual late September minima in ice extent, so truly unprecedented melting is occurring in the Arctic. The most recent images from the North Pole webcam show plenty of melt water and rainy conditions near the Pole.


Figure 2. Current extent of the polar sea ice, compared to the normal for this time in August (pink line). Image credit: National Snow and Ice Data Center.

I'll have an update Saturday morning.
Jeff Masters