Impact of Dropwindsonde Observations on Hurricane Forecast Modeling
Results of Aberson and Franklin Analysis of Dropwindsonde Data on Hurricane Model Forecasts
Dropwindsondes are weather instruments that collect atmospheric data as they descend after being dropped from research aircraft. These dropwindsondes obtain vertical profiles of wind, temperature, and humidity from 400mb to the surface. This data is then ingested by numerical numerical models to come up with hurricane track and intensity forecasts. These observations provide grid points of observations over the tropical oceans which are generally devoid of weather observations. Aberson and Franklin (1999) state that "accurate modeling of tropical cyclone motion and intensity requires both realistic numerical models and accurate representation of meteorological fields through the depth of the troposphere on a variety of scales." While models have greatly improved over the past 20 years, significant forecast improvements are still possible by decreasing the analysis error. This has been the primary goal of the Hurricane Research Division (HRD) branch of NOAA. This is why NOAA has procured a new generation of dropwindsondes based on the Global Positioning System (GPS) as well a Gulfstream-IV jet aircraft (G-IV). A study of the impact of the new dropwindsondes' observations on hurricane forecast models was conducted by Aberson and Frankilin in 1997.
In the 1997 study about 30 dropwindsondes were used during each mission. These observations were then put into the Geophysical Fluid Dynamics Laboratory (GFDL) and VICBAR hurricane models and the Global Spectral Model (GSM) using the National Center for Environmental Prediction (NCEP) Global Data Assimilation (GDAS). GDAS uses a quality control algorithm, synthetic data and analysis procedures, and the Global Spectral Model in its data assimilation. Further information about the data assimilation can be found in Aberson and Franklin (1999).
For the study NCEP's GSM is run and is used as the boundary conditions for the HRD's barotropic model (VICBAR) and the GFDL model. The track forecasts of the GSM, VICBAR, and GFDL models were compared to the CLIPER model to calculate track errors and the SHIFOR model to calculate intensity errors. The CLIPER and SHIFOR models are statistical regression models which use only climatology and persistence in their forecasts.
Results of Aberson and Franklin Analysis of Dropwindsonde Data on Hurricane Model Forecasts
The results of the impact of dropwindsonde observations on hurricane forecast models are quite promising. All three models studied provided improved track forecasts. The track forecasts are considered to be improved if they have forecast errors smaller than the CLIPER model. The GFDL model showed the best forecast track error reductions. The GFDL model was improved by 32% within 48 hours of initialization. The VICBAR model had track error reductions of 10%. The GSM model performed the worst of the three models as it only had forecast improvements at 12 and 84 hours. However some attention must be paid to the fact that data in most of the G-IV missions was collected in an asymmetric matter. Better data collection is the probable reason that track forecasts were not more improved.
The GFDL model is the only one of the three models studied that provides intensity forecasts. The intensity forecasts of the GFDL model are considered to be improved if they have forecast errors smaller than those of the SHIFOR model. The GFDL model produced intensity errors greater than the SHIFOR model through 48 hours. However, from 48 hours through 72 hours the GFDL model's intensity forecasts were more than 20% better than the SHIFOR model.
Aberson and Franklin (1999) conclude that the assimilation of dropwindsonde observations improved the forecast models and that with better sampling of the atmosphere that even better track and intensity forecast improvements are possible. They state that "with the ability to sample the atmosphere in all quadrants of the tropical cyclone environment on a regular basis, and with further research into targeting and optimal sampling strategies, larger improvements in tropical cyclone forecasts may be achieved in the future."
The importance of dropwindsondes on tropical cyclone forecasting cannot be disputed. The figure below shows that the actual warning area for 1993's Hurricane Emily was much smaller than the area that would have been warned had dropwindsonde data not be available.
However, there are limitations to using dropwindsondes and aircraft reconnaissance. The below image shows the regions that a hurricane hunter aircraft would sample as a hurricane approaches. It also shows the main limitiation of aircraft reconnaissance. These aircraft have a limited range, and therefore could only sample a storm as it begins to near the US coast. No such limitation exists with remote sensing.
More info about the
Including GOES High-Density Multispectral Winds into GFDL Model
Soden et al. (2000) states that the "numerical prediction of hurricane forecasts require accurate representation of the current meteorological conditions." The accurate modeling of hurricanes is limited by the lack of atmospheric observations over the tropical oceans. Aircraft reconnaissance provides some information when tropical cyclones are development, however the observations recorded by the "Hurricane Hunters" are too few and far between. The availability of high-density satellite winds is a new and incredibly valuable tool that provides at least some data to be assimilated into numerical weather prediction models.
Soden et al. (2000) studied the impact of including GOES High-Density Winds into the GFDL hurricane forecast model. Satellite-derived GOES winds were included into experimental GFDL hurricane model forecast runs to study the impact of including the winds on increasing the accuracy of hurricane track predictions. These winds are found by analyzing cloud and water vapor imagery in GOES-8 multispectral radiance observations. A 3-dimensional optimum interpolation algorithm was then employed to relay the winds into the GFDL model. For the experiment, a series of forecasts with (WIND) and without (CTRL) the satellite-derived winds were produced.
In three hurricane seasons (1996-1998), 10 storms
were studied, for which over 100 forecasts were produced. The study
showed that the inclusion of the satellite winds in the GFDL model decreases
the track error in all forecast periods. In particular, 60% of the
predictions for the 24, 36, and 72 hours forecasts were improved.
This is attributed to the fact that the inclusion of GOES winds into the
model better represents the strength of vorticity gyes in the environmental
During the 1996 Atlantic hurricane season, Soden et al. looked at hurricanes Bertha, Edouard, Fran, and Hortense for forecast errors at 12, 24, 36, 48, and 72 hours. In all cases, the assimilation of GOES winds resulted in model improvements in track error. The wind forecasts also show a tighter recurvature for Bertha, Edouard, and Fran. As in the Aberson and Franklin dropwindsonde study, a track error less than that of the CLIPER model is an improvement.
The analysis for Bertha shows that a track error reduction at all times, with best model performances occurring at 12 and 72 hours. (Bertha Model Errors) The forecasts for Hurricane Edouard were also improved, especially from 12-36 hours, however the model performed worse than the CLIPER at 72 hours. (Edouard Model Errors) The model forecasts for Fran were degraded at 12-24 hours, but the 36 and 48 hour model runs had error reductions of 20%. (Fran Model Errors) The GOES winds model runs performed the best for Hurricane Hortense. The model errors reductions ranged from 6-32% for the entire forecast period. (Hortense Model Errors) A significant point to note is that the (CTRL) GFDL model runs without the GOES winds performed better than the CLIPER model at only one forecast time, while the GFDL model with GOES winds was better at all times. Soden et. al believes this to be a result of the improved initial conditions resulting from the GOES winds inclusion in the model.
For the 1996 season the largest track error improvements occurred at the 72 hour forecast period (17%). Also significant improvements were seen at the 12 and 36 hour forecasts with model error reductions of 15% and 11% respectively. "The assimilation of satellite winds led to a more accurate forecast in more than half of the cases, with the frequency of improved forecasts ranging from 13 of 25 cases (52%) at 36 hours to 13 of 20 cases (65%) at 72 hours.
The 1997 season had very few opportunities to do substantial studies due to the 1997 El Nino. Many of the storms were rather odd in nature and were bad candidates in that they usually dissipated before a full forecast period had passed. The only suitable storm of the 1997 season was Hurricane Erika. In 14 forecasts made for Erika, GOES winds improved 11 of them for a 79% improvement 72 hours.
During the 1998 season five storms were studied: Hurricanes Bonnie, Danielle, Georges, Ivan, and Jeanne. For these storms, the GOES winds improved track errors on average from 5% at 12 hours to 15% at 24 and 36 hours. Track error reductions of 20%-40% were seen in the 24-72 hour forecasts. (Bonnie Model Errors) (Georges Model Errors)
Figures showing model and observed tracks for:
Bertha and Edouard
Fran and Hortense
Bonnie and Georges
Overall Results of Soden et al. GFDL GOES Winds Study
The overall results of the Soden et al. GOES winds
experiments are promising towards decreasing model track reductions and
helping forecasters to predict with more accuracy a hurricane's movements
as it approaches landfall. The experiment showed that the inclusion
of GOES winds had the most significant improvements on the 72 hour forecasts
indicating that "the satellite winds are having the greatest impact on
the far-field environmental flow in the model rather than the near-storm
currents (Soden et al. 2000)." The lack of improvements in the near-storm
currents states Soden et al. is because the GFDL model eliminates much
of the circulation features near the hurricane's center and replaces it
with an artificial vortex. Another important point to mention is
that the (WIND) forecasts more accurately represented the storms' recurvature
and that successive model runs are less variable than the (CTRL) forecasts.
Discussion-Conclusions Back to Main Page