IMPACT ON HURRICANE TRACK AND INTENSITY FORECASTS OF GPS DROPWINDSONDE OBSERVATIONS FROM THE SYNOPTIC SURVEILLANCE MISSION INTO HURRICANE DANIELLE ON 28 AUGUST, 1998.

Sim D. Aberson

Hurricane Research Division
Atlantic Oceanographic and Meteorological Laboratories
National Oceanic and Atmospheric Administration

4301 Rickenbacker Causeway
Miami, Florida


Click here for catalog of sonde drops.
Click here for zipped post-processed sonde data.


1. Synoptic situation

Hurricane Danielle developed from an easterly wave midway between the Cape Verde Islands and the Leward Islands on 24 August, 1998. It moved westward, strengthening rapidly into a hurricane the next day. Due to the potential threat to the Carolina coastline, a synoptic surveillance mission was tasked for nominal time 29 August 1998 0000 UTC. At that time, Danielle was moving slowly west-northwestward about 600 km north of western Puerto Rico, to the south of the strong subtropical ridge (Fig. 1). A weak cold low was located over Jamaica, and another about 1000 km east of the Lesser Antilles. The circulation south of Long Island was weakening Tropical Storm Bonnie. A strong tropical wave was located over the northwestern Caribbean Sea, and that wave eventually developed into Hurricane Earl in the Gulf of Mexico.

2. General Assessment of dropwindsonde impact

A. GFDL model

Figure 2 shows the GFDL forecast tracks for Hurricane Danielle, and Table 1 shows the errors and impact of the synoptic surveillance mission. The results are positive, with large improvements at all times through 120 h when the center could not be followed in some of the model runs. The upper-tropospheric data improved the forecasts after 36 h, and seemed to provide almost all of the large improvements late in the forecast.

B. VICBAR

Figure 3 shows the VICBAR forecast tracks for Hurricane Danielle, and Table 2 shows the errors and impact of the synoptic surveillance mission. The dropwindsonde data improved the forecasts at all times except 72 h, even though the forecasts without the dropwindsondes were exceptional through 48 h. The upper-tropospheric data degraded the forecast at all times except 48 and 72 h, likely because Danielle was a weakening hurricane with mainly shallow convection throughout much of the forecast period.

C. GSM

Figure 4 shows the GSM forecast tracks for Hurricane Danielle, and Table 3 shows the errors and impact of the synoptic surveillance mission. The dropwindsonde data improved the GSM forecast track at all times, even when the forecast without the dropwindsondes was very good. The upper-tropospheric data improved the forecast at all forecast times after 36 h.

D. Intensity

Table 4 shows the GFDL intensity forecast errors and impact of the synoptic surveillance mission. The dropwindsonde data had a mixed impact, improving the forecasts only at 72 and 84 h. Tuleya and Lord (1997) also showed modest improvements to GFDL model intensity forecasts in the HRD synoptic flow cases.

3. Targeting

The atmosphere has long been recognized as a chaotic system (Lorenz 1963), e.g. very small perturbations to initial conditions result in increasingly large differences in the evolution of the atmosphere with time. Since the exact state of the atmosphere can never be measured, all analyses contain errors whose magnitudes can only be estimated. An indeterminate number of initial conditions consistent with the observational data can therefore be used in numerical weather prediction, and single model runs at any synoptic time only give one possible solution to the evolution of the atmosphere. Many operational forecast centers around the world, therefore, now employ ensemble forecasting as a means of quantifying the uncertainty in the evolution of the atmospheric system. Small perturbations from a best "control" state are calculated and added to and subtracted from this control to allow for different integrations starting from theoretically equally likely initial states. These perturbations are designed so as to mimic the fastest growing modes in the model and to create the largest envelope of possibilities in the forecast. Therefore, they generally correspond to locations where large analysis errors will most impact the forecast. Those features corresponding to perturbations which most impact the tracks of tropical cyclones must be found, properly sampled, and thoroughly tested, to prove the efficacy of targeting techniques.

Figure 5 shows the variance of the size of the perturbations in the National Centers for Environmental Prediction (NCEP) global model ensemble forecasting system (Toth and Kalnay 1993). The two largest perturbations are associated with the Tropical Storm Bonnie south of New England and Hurricane Howard in the North Eastern Pacific Ocean. Near Danielle, large perturbations exist in the region between Danielle and the subtropical ridge to the north and in the trailing anticyclone behind Danielle. Another area of large perturbation is in the subtropical ridge axis to the northwest of Danielle. Two other smaller areas correspond to the two cyclonic circulations in the western Caribbean Sea and the slight ridging just off the eastern coast of Nicaragua. Of these, the only two areas of perturbation which were well-sampled were the ones to the northwest of the center of Danielle corresponding to the axis of the subtropical ridge, and the area between the subtropical ridge and the circulation of Danielle.

Two sets of model runs have been performed. The first, the TG run, includes the dropwindsonde taken in and around well-sampled areas of large ensemble perturbation. Because the southeastern portion of the perturbation corresponding to Danielle was not sampled, only those observations within and around the two regions northwest of Danielle were included (all the dropwindsondes represented by closed circles in Fig. 5). The other, the NT run, includes the complement of the TG run, with dropwindsondes represented by open circles in Fig. 5. Results are shown in Table 1, Table 2, Table 3, and Table 4 and Fig. 2, Fig. 3, and Fig. 4. The TG run provided better forecasts than the run including all dropwindsonde data through 36 h in the GFDL, at 24 and 36 h and from 72 h onward in VBAR, and at 24 and 36 h in the GSM. Since the two southeasternmost dropwindsondes were not included in the TG run, another set of runs included these two dropwindsondes to test whether the lack of data completely surrounding the circulation of Danielle led to the degradation of the forecast. These forecasts (not shown) were only slightly different from those provided with the TG dropwindsondes.

Figure 6 shows the difference in the 850 - 200 hPa averaged winds between the runs in which all the dropwindsonde data and none of the dropwindsonde data are included. The largest differences form an almost complete ring 500 km in radius around Danielle, roughly corresponding to the locations of the soundings. However, this impact is just outside the dropwindsonde locations to the south of Danielle, and rotates inside the sounding locations to the northwest, and much of the maximum is either inside or outside the dropwindsonde locations, and extends more than 500 km away from these locations in some places. A second large difference is located off the Carolina coastline. The spread of the impact away from the dropwindsonde locations suggests that the data assimilation has allowed the dropwindsonde data from surrounding areas to influence the initial conditions there. Since the difference maximum to the south of Danielle is in an area of large ensemble perturbation, this aliasing may amplify in time and negatively impacted the forecasts. Figure 7 shows that, by 24 h into the forecast, the differences between the forecasts with and without the dropwindsonde data have rotated around Danielle. After amplifying for the first 12 h (not shown), they have started to decay slightly. The large difference initially off the Carolina coastline in an area of relatively low ensemble perturbation has moved northeastward and rapidly decayed. A new difference maximum located over Central Florida, originated in the relative difference maximum about 300 km east of the northern Bahamas. These results generally confirm that the largest (smallest) ensemble perturbations correspond to amplifying (decaying) modes in the model.

4. Conclusion

The dropwindsonde data obtained during the synoptic surveillance mission for Hurricane Danielle at nominal time 29 August 1998 0000 UTC has provided excellent results throughout the five day forecast of Hurricane Danielle in all three models. The MRF ensemble forecasting system suggested that data surrounding Danielle and in the subtropical ridge axis to the north of Danielle would have the greatest impact on the Danielle forecast. Despite sampling in all quadrants of Danielle, the southeasternmost of these target locations was not completely sampled, and some aliasing in this area may have prevented even larger improvements, as suggested by the results of model runs without dropwindsonde data in this area.


Table 1
Track forecast errors for the no dropwindsonde GFDL control (GFNO), the all dropwindsonde run (GFAL), the lower-level dropwindsonde run (GFP3), and the run with only targeted observations (GFTG), and the percent improvement of the latter three over the control.
Forecast
time (h)
GFNO
Error (km)
GFAL Error (km)
(% Improvement)
GFP3 Error (km)
(% Improvement)
GFTG Error (km)
(% Improvement)
1252.39.( 25%)30.( 42%)35.( 33%)
24104.77.( 26%)75.( 28%)63.( 39%)
36186.119.( 36%)116.( 38%)112.( 40%)
48280.169.( 40%)194.( 31%)177.( 37%)
72481.275.( 43%)382.( 79%)338.( 30%)
84749.428.( 57%)617.( 18%)560.( 25%)
961136.617.( 46%)906.( 20%)824.( 27%)
1201111.1481.
Table 2
Track forecast errors for the no dropwindsonde VBAR control (VBNO), the all dropwindsonde run (VBAL), the lower-level dropwindsonde run (VBP3), and the run with only targeted observations (VBTG), and the percent improvement of the latter three over the control.
Forecast
time (h)
VBNO
Error (km)
VBAL Error (km)
(% Improvement)
VBP3 Error (km)
(% Improvement)
VBTG Error (km)
(% Improvement)
1242.40.( 5%)40.( 5%)42.( 0%)
2461.41.( 33%)41.( 33%)40.( 34%)
3660.32.( 47%)23.( 62%)20.( 67%)
4841.10.( 76%)20.( 51%)24.( 41%)
72266.264.( 1%)273.( -3%)235.( 11%)
84328.302.( 8%)288.( 12%)271.( 10%)
96524.473.( 10%)452.( 14%)453.( 14%)
108781.676.( 13%)638.( 18%)654.( 16%)
1201154.996.( 14%)926.( 20%)984.( 15%)
Table 3
Track forecast errors for the no dropwindsonde GSM control (GSNO), the all dropwindsonde run (GSAL), the lower-level dropwindsonde run (GSP3), and the run with only targeted observations (GSTG), and the percent improvement of the latter three over the control.
Forecast
time (h)
GSNO
Error (km)
GSAL Error (km)
(% Improvement)
GSP3 Error (km)
(% Improvement)
GSTG Error (km)
(% Improvement)
1256.15.( 73%)15.( 73%)39.( 30%)
24112.102.( 9%)93.( 17%)83.( 26%)
36123.111.( 10%)101.( 18%)95.( 23%)
48192.163.( 15%)171.( 11%)163.( 15%)
7294.49.( 48%)88.( 6%)52.( 45%)
84560.450.( 20%)551.( 2%)470.( 16%)
96787.587.( 25%)780.( 1%)623.( 21%)
1081056.787.( 25%)1109.( -5%)837.( 21%)
1201213.961.( 21%)1371.(-13%)1003.( 17%)
Table 4
Intensity forecast errors for the no dropwindsonde GFDL control (GFNO), the all dropwindsonde run (GFAL), the lower-level dropwindsonde run (GFP3), and the run with only targeted observations (GFTG), and the percent improvement of the latter three over the control.
Forecast
time (h)
GFNO
Error (kn)
GFAL Error (kn)
(% Improvement)
GFP3 Error (kn)
(% Improvement)
GFTG Error (kn)
(% Improvement)
1205( und%)3( und%)2( und%)
2424(-100%)0( 100%)1( 50%)
36411(-175%)9(-125%)11(-175%)
4869( -50%)6( 0%)4( 33%)
7298( 11%)7( 22%)7( 22%)
8456( -20%)8( -60%)3( 40%)
9622( 0%)0( 100%)4(-100%)
120511


Return to 1998 assessment page
Return to HRD home page