a. GFDL model description

Kurihara et al. (1995) described the GFDL model, which is one of the operational hurricane forecast models at NCEP. It is a primitive equation model formulated in latitude, longitude, and sigma coordinates with 18 vertical levels (Table 1 of Kurihara et al. 1990). The grid system consists of two movable inner meshes of 1/6° and 1/3° resolution within a fixed 75° × 75° outer domain of 1° resolution (Bender et al. 1993). The GFDL forecast system consists of three steps: 1) an interpolation of the NCEP global analysis onto the MMM regional domain, 2) a vortex specification to produce a realistic storm-scale initial condition near the hurricane, and 3) a model forecast from the above-obtained initial conditions. The GFDL model typically produces a 72-h forecast from the global NCEP analysis with the vortex specification derived from a storm message supplied by the National Hurricane Center. Based on semioperational and operational forecasts for the 1993–95 Atlantic and eastern Pacific seasons, the GFDL track forecasts appear superior to other NOAA operational products especially beyond one day (e.g., Aberson and DeMaria 1994).

b. NCEP data assimilation for the dropwindsonde cases

The NCEP Global Data Assimilation System consists of a global model, an analysis procedure, a quality control algorithm, and a synthetic data (wind bogusing) procedure. More details of the NCEP model can be found in Kanamitsu (1989) and Kanamitsu et al. (1991). The ∼1° global analysis utilizes spectral statistical interpolation (SSI, Parrish and Derber 1992) as configured for operational use in June 1991. Information on the synthetic data procedure and initial analyses used in the present paper can be found in Lord (1991, 1993). In the synthetic data procedure, wind pseudo-observations based on operationally estimated intensity and storm position are ingested at mandatory pressure levels from 1000 to 300 mb over an area 300 km on a side near the storm center. The synthetic data scheme was designed to give a more realistic initial state of storm structure and steering current to the global NCEP analysis and, therefore, to improve the subsequent NCEP forecast. The ODW data (wind, temperature, and moisture) were assimilated in a similar fashion to the current operational method (Burpee et al. 1996). The cases and analyses used in this study are from Lord (1993) and correspond to his CNTL (control, no ODW or synthetic data), ODW (ODWs only), SYN (synthetic data, no ODWs), and ODW/SYN (synthetic data ODWs) for each of the 14 cases listed in Table 1. These cases are shown in Fig. 2 and include storms in the western North Atlantic or Gulf of Mexico with four making landfall in the continental United States.

c. GFDL model results without dropwindsondes

To examine the impact of ODWs using the GFDL model, one may use the synthetic or control runs of the global analyses. In addition the GFDL modeling system can be used with and without its specified vortex procedure. It is interesting to examine first the sensitivity of the GFDL model to these variations in initial conditions. The GFDL track errors are shown in Fig. 3 for the four possible cases: 1) control case, experiment GCNTL (no ODW or synthetic vortex); 2) NCEP synthetic vortex in the global analysis, experiment GSYN; 3) control case with the GFDL-specified vortex, experiment GSPE; and 4) NCEP synthetic vortex and GFDL-specified vortex, experiment GFDL. The results are shown relative to CLIPER, the standard climatology and persistence model. The liability of running the GFDL MMM with an ill-defined vortex in the global analysis is clearly shown in the control experiment suite with no bogus vortex specification. Some cases exhibit skill [e.g., Hurricane Gloria (1985), Kurihara et al. (1990)] when the global analysis is able to resolve a well-defined tropical system in the appropriate location without adding any idealized storm data. Overall for the total of these 14 control cases, however, the control experiments GCNTL show no relative skill until after 2 days. On the other hand, the implementation of either idealized vortex (NCEP synthetic, GFDL specified, or both) demonstrated considerable skill in track forecasting. This can be seen in Fig. 3 in which skill relative to CLIPER is shown from 12 h onward. If the synthetic NCEP vortex is used together with the GFDL-specified vortex (i.e., the case for the present operational GFDL system in which the scheme first filters out the global NCEP synthetic system and replaces it with the GFDL specified vortex), a high degree of skill is achieved (generally 40% better than CLIPER after 1 day), similar in performance to when either the NCEP synthetic or the GFDL-specified vortex is used alone. One major reason to install the NCEP synthetic system operationally was to have a reasonable, trackable system from the start. It appears that the GFDL model is capable of credible track forecasts using the NCEP synthetic vortex without utilizing its own vortex specification method. Notice that in these cases, forecasts with the NCEP synthetic vortex alone yield slightly better results than when the GFDL-specified vortex was used, but there is no statistical difference in track error among GSYN, GSPE, and GFDL. However, in order to track weak systems in operational mode, the GFDL-specified vortex was necessary in the day-to-day implementation of the GFDL system (Bender et al. 1993).

The GFDL MMM is the first operational dynamical model to forecast intensity. Its feasibility was demonstrated in Bender et al. (1993), but consistent predictive skill relative to SHIFOR for a large case size has yet to be demonstrated. Despite the lack of skill, Fig. 4 shows the advantage in the GFDL-specified vortex system in reducing the intensity error in the first 2 days of the forecast. The GFDL-specified system forecasts were about the same for the case with or without the NCEP synthetic vortex (i.e., GFDL and GSPE). On the other hand the suite of experiments using the NCEP synthetic vortex alone (GSYN) shows improvement over the control (GCNTL), which lacks skill for the entire 72-h forecast period. Both GCNTL and GSYN are limited by the coarse resolution of the global assimilation system; GCNTL, without the synthetic vortex, also suffers from data sparsity near the storm center. Skill relative to SHIFOR, the standard climatology and persistence intensity technique, is not reached until 36 h in GSYN but is reached at ∼30 h in GFDL. In general, significant skill relative to SHIFOR is achieved in GSYN and GFDL at and after these times for this limited case study. In the first day or so, the level of forecast intensity skill is not large in either GFDL system as one sees evidence of an initial adjustment bias in Fig. 4. Values of 400% worse and 200% worse than SHIFOR are shown at 12 h, decreasing to 200% and 50% worse at 24 h for the NCEP synthetic and the GFDL-specified series, respectively. The GFDL system overintensifies weak systems and underpredicts intense systems regardless of the forecast hour (Fig. 5). Intensity errors range from as high as 25 m s⁻¹ overprediction to as low as 23 m s⁻¹ underprediction. Note that there are only four cases of overprediction and four cases of underprediction for observed wind speeds greater than and less than 25 m s⁻¹, respectively. This bias remains if one examines the intensity errors relative to the initial intensity as well. Research and model development is continuing on removing or alleviating this initial bias, which is also evident from the operational GFDL forecast system (Kurihara et al. 1995).

a. Impact of dropwindsondes on the GFDL forecast tracks

In this section, the impact of ODWs will be shown using the GFDL model for four suites of initial conditions, which are divided into two pairs. The first pair uses the NCEP model initial conditions with synthetic data (without and with ODWs), and the second pair is the same as the first except that the GFDL specified vortex is used as well (Table 2). The two suites without ODWs correspond to the GSYN and GFDL experiments discussed earlier. The mean track errors for each 12-h forecast period are smaller with ODWs than without ODWs (Fig. 6). The mean improvement ranges from 12 km at 12 h, ∼50 km for 24–60 h, to 127 km for the nine cases at 72 h. Although the relative errors of GSYN and GFDL change with forecast period, both suites of experiments indicate similar improvement with ODWs despite the different method of vortex specification. Since the track forecasts are better, environmental wind and basic steering current improvements apparently are retained with both bogusing techniques.

One can also examine the case-to-case variability in addition to the mean results. This is important for testing the robustness of the ODW impact on track forecast errors whose case-by-case variation is known to be large. A majority of cases indicate improvement using ODWs for all forecast periods. Improvements ranged from 9 of 14 cases at 12 h to all cases at 72 h, with improvement overall at ∼77% of the verifying times. The mean track error for all forecast periods showed a considerable difference at the 95% significance level except at 12 and 48 h, which indicated improvement at the 90% significance level. Figure 7 shows a scatter diagram of forecast errors for experimental suites GFDL and GSYN for the 14 cases at all forecast periods. Most of the cases lie above the diagonal line, indicating improvement using ODWs.

While the mean errors show an overall improvement with ODWs, there is considerable case-to-case variation (Fig. 7). A comparison of improvement using GFDL and GSYN shows positive correlation with ∼2/3 of the verifying times of the 14-case ensemble indicating the same sign of improvement. For both vortex initialization methods the most improvement was shown for the case of the 72-h forecast hour of Hurricane Floyd on 11 October with improvements of 504 and 457 km for GSYN-WD and GFDL-WD, respectively. On the other hand the worst responses were for GSYN-WD for Hurricane Florence at 48 h (158-km degradation) and GFDL-WD at 48 h for Hurricane Floyd on 12 October (261-km degradation). Note that all forecasts with ODWs that were worse had errors of less than 300 km for GSYN and 750 km for GFDL. Overall only 15 and 13 of the 63 total verifying times were degraded for GSYN-WD and GFDL-WD systems, respectively.

b. Comparison of improvement with previous studies

It is useful to compare previous results of ODW impact with the current study. Table 3 lists the forecast track error for a homogeneous sample of the 14 synoptic cases without ODWs for the NCEP model with the synthetic vortex, VICBAR, and the GFDL model experiments, GFDL and GSYN, with the NCEP synthetic vortex implemented (with and without the GFDL-specified vortex). Consistent with error statistics from larger sample size, the GFDL model suites display substantially superior forecasts beyond 12 h. It is therefore more instructive to examine the relative improvement that ODWs have on each model by comparing the fractional improvement in the mean forecast error (Fig. 8). The mean improvement using the two GFDL model suites is >10% at 12 h and 15%–30% thereafter. The NCEP model improves about as much as the GSYN model suite for the first 36 h, increasing from 10% to >20% beyond 24 h. The model suite GFDL improves somewhat less initially with values approaching 20%. Apparently the GFDL vortex specification system has some influence on the impact of the ODW data. This technique may dilute the ODW data signal near the storm center. It is left for further study to evaluate the exact impact of different filtering and bogusing strategies on limiting the usefulness of additional high-density data such as ODWs. Although both indicating improvement beyond 36 h, the NCEP and GFDL model suites behave quite differently. One main difference between the NCEP and both GFDL experiments is that at the 72-h forecast period where the GFDL forecasts continue to improve at the same or even greater degree while the NCEP global model improvement drops off below 10%. This is due primarily to the difficulty the T126 global model has in sustaining a viable vortex past 60 h.

One can also examine the individual case improvement between the NCEP and the two GFDL suites. The GSYN and the NCEP global runs are highly correlated with ∼44 of 60 verifying times of the 14-case ensemble indicating the same sign of improvement. This is even more evident for the 12- and 24-h periods when 23 of 28 verifying times show trends of the same sign. As expected, the correlation between the GFDL and the NCEP global series is not as strong due to the GFDL-specified vortex changing the initial conditions. During later periods the GSYN and GFDL series of experiments show a more similar response to ODWs than that of the NCEP response. This suggests that the results become more model sensitive and less dependent on details of the initial conditions for later times in the forecast period.

In addition, a positive relationship was found between the magnitude of the 850–200-mb vertical wind shear and the magnitude of the ODW track improvements in the NCEP, GSYN, and GFDL runs with corresponding correlation coefficients of 0.44, 0.40, and 0.10, respectively. The correlations were calculated at 12, 24, 36, 48, and 72 h using the shear calculated for a 10° × 10° area following the storm in the GSYN forecasts. Large ODW improvements were found in highly sheared, rapidly moving storms such as Hurricane Emily for the NCEP, GSYN, and GFDL runs. In contrast, the VICBAR series shows a relatively small fractional improvement of ∼10% to 36 h, decreasing after that, with little correlation (correlation coefficient of 0.04) found between ODW improvement and vertical shear. This indicates that both the barotropic and baroclinic structures are improved by ODWs and that this is subsequently manifested in more improved track forecasts in the baroclinic models. The VICBAR model can improve only through barotropic changes induced by ODWs and its forecast skill is limited by two-dimensional assumptions beyond ∼36 h. This can be further seen in individual case comparisons (J. L. Franklin 1995, personal communication), which indicate 98% of the ODW improvements are confined to within 100 km in VICBAR, whereas in the NCEP and GFDL models greater than 20% of occurrences had improvement greater than 100 km. Additionally in a manner similar to the GFDL suite, the VICBAR system may not fully utilize the ODW data near the storm center. Nevertheless ∼45 of 60 occurrences indicate coincident improvement of the VICBAR and the GSYN experiments. Overall it appears that the NCEP, VICBAR, and GFDL suites were able to benefit from the additional ODW data and that for most cases the relative improvement is at least qualitatively similar. At the later forecast periods, however, the GFDL model shows absolute skill levels above those of either VICBAR and the NCEP model experiments, and thus should be more able to retain the initial information from the ODWs. This supposition is confirmed by the fact that the GFDL model suites show ODWs having a significantly large positive influence into day three.

c. Comparison of individual cases with and without dropwindsondes

Section 3a indicated that tracks of the 14 cases were improved for ∼70% of the forecast verification times when using ODWs and the GFDL model. The initial analyses with and without ODWs and their corresponding differences are now examined for particular cases. For this purpose the initial analyses and forecast tracks will be analyzed using the GSYN series. In this study the ODW impact on storm track was traced directly to the initial analysis, although clearly as section 3b indicates, model forecast skill has a strong bearing on the evolution of this impact.

The case for Hurricane Floyd (0000 UTC 11 October 1987) had large forecast errors without ODWs but showed considerable improvement due to the well-sampled ODW distribution (Fig 1). Figure 9 shows the observed track of Hurricane Floyd and the forecast with and without ODWs. The ODW case better forecasts the initial north-northwest movement and subsequent recurvature northeastward. Figure 10 (top) displays both the initial sea level pressure and wind analysis near 500 mb. At this level differences of ∼5 m s⁻¹ were typical between wind analyses with and without ODWs for this and all other cases. The movement of Hurricane Floyd was affected by a southward extension of the midlatitude upper-level trough above 500 mb and related weakness in the subtropical ridge. In order to analyze this case further, the analyses with and without ODWs were then spatially filtered to remove the disturbance field (see Kurihara et al. 1993). The vertical mean difference between these two filtered analyses (Fig. 10, bottom) indicates near-storm differences of ∼0.4 m s⁻¹ with larger differences to the north in the direction of storm movement. The analysis with ODWs shows an enhanced weakness in the subtropical ridge with the increased westerlies being consistent with the more initial eastward and subsequently more rapid northeastward movement of the storm.

The case for Hurricane Hugo (0000 UTC 20 September 1989) was analyzed with and without ODWs in a similar manner. In contrast to the upper-level trough interaction in Hurricane Floyd, the track of Hurricane Hugo was dictated by a deep northwestward flow on the southwest side of both surface and midlevel highs. Figure 11 shows the observed and two forecast tracks with and without ODWs, and Fig. 12 (top) displays the synoptic situation. Notice the more accurate initial northwestward movement and subsequent more rapid movement of Hurricane Hugo toward the coast in the experiment with ODWs. Although ∼6 h too slow in making landfall, the ODW experiment predicted landfall ∼4 h earlier than the no-ODW experiment. The vertical mean difference field (Fig. 12, bottom) clearly showed an enhanced flow around the Atlantic high with the ODW data steering the storm more rapidly northward toward the coast. Near-storm differences of ∼0.5 m s⁻¹ are observed with larger values in the path of the storm again being consistent with the differences in forecast storm track between ODW and no-ODW experiments.

Similar results were also noted for other cases including Hurricane Debby (0000 UTC 16 September 1982), which took a jog in response to an upper-level trough. This small-scale movement (Fig. 4 of Burpee et al. 1996) was forecast particularly well by the GFDL model using the analysis with ODWs. On the other hand, the worse case of GSYN ODW degradation was examined: that of Hurricane Florence, which moved northward from the southern gulf making landfall in the Louisiana delta region. As in the case of Hurricane Floyd, Hurricane Florence was influenced by an upper-level trough in the westerlies. With ODWs, this trough was not as sharp leading to a more relatively westerly mean flow across the path of the storm. This led to an erroneous eastward movement with ODWs. A key unanswered question is whether such degradations can be traced to errors in the data, biases in either analysis scheme or model, or are due to the random fluctuations of predictability.

Steering currents may be defined in a gross sense (e.g., Chan and Gray 1982; Dong and Neuman 1986; Carr and Elsberry 1990; Franklin et al. 1996) as a vertically averaged flow along the perimeter of the storm. This was done for an azimuthal ring, vertically integrated throughout the atmosphere, 600 km from the center of each of the 14 storms at the initial time for the analyses with and without ODWs. In the case of Hurricane Floyd on 11 October the ring analysis had a steering direction 70° north of west at 2.3 m s⁻¹ for the analysis without ODWs. The ODW analysis had a steering flow more toward the north, 83° north of west at 2.3 m s⁻¹. This again was consistent with the model track evolution in which the storm in experiment GSYN-WD moved more eastward than that in experiment GSYN-ND. In the case of Hurricane Hugo on 20 September the ring analysis had a steering magnitude of 3.4 m s⁻¹ at ∼73° north of west for the analysis without ODWs compared to 4.5 m s⁻¹ in the analysis with ODWs. This again was consistent with the model track evolution in which the storm in experiment GSYN-WD moved northwestward more rapidly than that in experiment GSYN-ND. A similar relationship existed for most of the 14 cases with the steering current differences reflecting the ultimate difference in the model storm tracks. The mean magnitude of the difference in the steering current between the two analyses was ∼0.6 m s⁻¹ for those cases ranging from a difference of 0.1 m s⁻¹ in Hurricane Jerry to 1.4 m s⁻¹ for Hurricane Emily (24 September 1987). In summary, forecast track differences between analyses with and without ODWs can be explained qualitatively either by differences in steering currents defined in a traditional fashion (ring average) or by differences in the filtered fields of the vertical mean current over the storm. Vertically averaged wind changes due to ODWs are very subtle, typically ∼0.5 m s⁻¹, but a sustained difference of this magnitude can amount to more than 40 km in forecast track over 1 day. Given that errors in forecast wind fields grow rapidly with time (Lorenz 1965), mean improvements considerably larger than 40 km day⁻¹ can be expected for forecasts of several days or more.

d. Impact of dropwindsondes on intensity

Although the physical factors controlling intensity are not well understood, and achievable skill relative to SHIFOR is difficult, it is nevertheless informative to investigate whether there is any sign of increased skill when ODWs are used. Note that for these 14 cases skill relative to SHIFOR was achieved in forecasting without ODWs after ∼30 h using the GFDL model (section 2, Fig. 4). ODWs produce improvement in intensity forecasts, especially for the GSYN experiments (Fig. 13). The mean error reductions of 2, 3.5, 1.5, 2, 1, and 1.5 m s⁻¹ at 12, 24, 36, 48, 60, and 72 h, respectively, using ODWs indicate 95% statistical significance for all forecast periods except at 48 h. The GSYN experiments show positive impact of ODWs for the first 2 days when skill is low relative to SHIFOR (Fig. 4). The addition of ODW data in the hurricane environment combined with the NCEP synthetic vortex improves intensity forecasts when no mesoscale vortex specification method such as Kurihara et al. (1993) is used. The GFDL vortex specification system, however, initializes and forecasts a more realistic storm intensity even without ODWs in the early forecast periods. When ODWs are utilized in the GFDL vortex specification system, the mean error shows little significant increase in skill except at 60 and 72 h. For this period 75% of verifying times in the 14-case ensemble show improvement over an already skillful forecast, with a mean error reduction of 20%. Changes due to ODWs may be partially obscured by the known spinup bias at early periods in the GFDL forecast and possibly filtered out by the GFDL vortex specification technique.

Individual forecasts (Fig. 14) show improvement in intensity with ODWs on a case-by-case basis. This is indicated by the majority of solid dots (52 of 75) above the diagonal line for the GSYN forecasts. However as mentioned above, many of the improved cases occur when the intensity forecast is worse than SHIFOR (see Figs. 4 and 13). For the GFDL forecasts, the impacts of ODWs are not as apparent, although the majority (43 of 75 verifying times) indicate improvement. One can look at individual intensity forecasts and their relationship to forecasted large-scale fields. One factor known to be related to storm intensity is vertical shear of the environmental wind (DeMaria and Kaplan 1994). The shear was calculated from 200 to 850 mb for a 10° × 10° area surrounding the forecast storm position for the GSYN forecasts with and without ODWs. This quantity was then compared to both model forecast intensity and observed storm intensity. The results are shown in Table 4 for the 60-h forecast time for the 11 cases when positions were available. Notice that 7 of 11 cases showed improvement in intensity forecasts when ODWs were used and that in 5 of these 7 cases the vertical shear apparently was partially accountable for the improved forecast. Increased (reduced) shear contributed to reduced (increased) intensity including those cases with the three largest improvements in GSYN-WD: error reductions of 60%, 80%, and 15% for the cases of Hurricanes Floyd (11 October), Florence (9 September), and Hugo (21 September), respectively. On the other hand, note the degradation of forecast intensity for Hurricane Emily (24 September) in which GSYN-WD prematurely decayed the storm. One might expect that intensity prediction and track prediction may be correlated. No general relationship was found, however, for these experiments.