In 1997, the National Oceanic and Atmospheric Administration’s National Hurricane Center and the Hurricane Research Division began operational synoptic surveillance missions with the Gulfstream IV-SP jet aircraft to improve the numerical guidance for hurricanes that threaten the continental United States, Puerto Rico, the U.S. Virgin Islands, and Hawaii. The dropwindsonde observations from these missions were processed and formatted aboard the aircraft and sent to the National Centers for Environmental Prediction and the Global Telecommunications System to be ingested into the Global Forecasting System, which serves as initial and boundary conditions for regional numerical models that also forecast tropical cyclone track and intensity. As a result of limited aircraft resources, optimal observing strategies for these missions are investigated. An Observing System Experiment in which different configurations of the dropwindsonde data based on three targeting techniques (ensemble variance, ensemble transform Kalman filter, and total energy singular vectors) are assimilated into the model system was conducted. All three techniques show some promise in obtaining maximal forecast improvements while limiting flight time and expendables. The data taken within and around the regions specified by the total energy singular vectors provide the largest forecast improvements, though the sample size is too small to make any operational recommendations. Case studies show that the impact of dropwindsonde data obtained either outside of fully sampled, or within nonfully sampled target regions is generally, though not always, small; this suggests that the techniques are able to discern in which regions extra observations will impact the particular forecast.
From 1982 to 1996, the National Oceanic and Atmospheric Administration (NOAA) P-3 aircraft released dropwindsondes in data-sparse tropical cyclone (TC) environments to improve numerical track forecasts. These experiments yielded significant improvements in the primary numerical guidance for National Hurricane Center (NHC) track forecasts (Burpee et al. 1996) and led to the procurement of a Gulfstream-IV-SP jet aircraft (G-IV) in 1996 for use in operational daily or, since 2002, twice-daily “synoptic surveillance” missions. The G-IV is deployed during the critical watch and warning period 24–60 h prior to anticipated landfall in the continental United States, Puerto Rico, the U.S. Virgin Islands, or Hawaii. Between 25 and 30 dropwindsondes are deployed during each mission, sometimes supplemented with additional dropwindsondes from missions conducted with the U.S. Air Force C-130 or NOAA P-3 aircraft, thus providing broader environmental sampling than the G-IV accomplishes alone. Similar missions have been regularly conducted in the western Pacific since 2003 as part of Dropwindsonde Observations for Typhoon Surveillance near the Taiwan Region (DOTSTAR), with additional participation of other international agencies during 2008 as part of The Observing System Research and Predictability Experiment (THORPEX) Pacific Area Regional Campaign (T-PARC; Wu et al. 2005, 2007; Weissmann et al. 2011; Harnisch and Weissmann 2010).
During the first 10 yr of surveillance (1997–2006), 175 missions were conducted (Aberson 2010), leading to a 10%–15% improvement in 12–60-h track forecasts from the NOAA Global Forecast System (GFS). Since a complete mission costs about $40,000, far less than the estimated $1 million average needed to evacuate just one mile of coastline for landfall (Aberson et al. 2006), such missions, regularly conducted, are very cost effective.
Questions remain as to whether the dropwindsondes are deployed optimally and as to whether larger error reductions are possible. If the TC environment is sampled symmetrically in all directions from the storm center, for example, the various sensitive areas in which additional observations would most improve the forecast may not be sampled. One strategy to find these sensitive “targets” was based on the hypothesis that regularly spaced observations collected in areas of high deep-layer-mean (DLM; 850–200 hPa) wind variance from the National Centers for Environmental Prediction (NCEP) Global Ensemble Forecasting System valid at the observing time would reduce growing analysis errors in the steering flow, thereby reducing track forecasts errors (Aberson 2003). The advantage of targeting observations based on this ensemble variance (EV) versus uniform sampling has been demonstrated in simple models (Lorenz and Emanuel 1998; Morss et al. 2001). Aberson (2003) showed that the assimilation of the data subset from fully sampled1 targets produced a statistically significant GFS track forecast error reduction of up to 25%, thereby yielding larger improvements than were possible by assimilating all available surveillance data. The deficiencies of symmetric sampling were attributed to suboptimal data assimilation schemes and their impact near targets that were bisected or otherwise not fully sampled.
Advanced techniques, such as the ensemble transform Kalman filter (ETKF; Bishop et al. 2001) and singular vectors (SVs; Palmer et al. 1998), have been evaluated in idealized and operational models. These techniques have now been extended for use in tropical cyclone prediction. Their respective guidance has been compared in the Atlantic and western Pacific basins (Majumdar et al. 2006; Reynolds et al. 2007; Wu et al. 2009). Unlike the EV, they focus on specific forecasts and seek optimal sampling locations in which the assimilation of new data would modify that particular forecast. Yamaguchi et al. (2009) showed, in one case with one targeting technique, that the assimilation of a small subset of dropwindsonde data in a well-chosen target area may provide significant track forecast improvements.
The applicability of these three techniques for improving TC forecasts is presented in the framework of an Observing System Experiment. Either all the dropwindsonde data or only those data within the fully sampled targets are included in the model runs described below. Given that dropwindsonde data obtained in tropical cyclone environments lead to an average 10%–15% reduction2 in track forecast error during the first 60 h (Aberson 2010), the resulting track forecasts from these model runs are compared to those obtained from model runs in which all the dropwindsonde data have been assimilated. The focus of this study is to examine whether, for each targeting technique, using the subsample of dropwindsonde data obtained only in fully sampled targets provides the same average results as using all the dropwindsonde data (suggesting that the particular technique is able to accurately define sensitive regions). The model and the three targeting techniques are briefly described in the next section. General results from the three targeting techniques and individual case studies are provided in section 3, followed by a discussion and the conclusions.
2. Overview and procedures
The G-IV released 25–30 dropwindsondes at 150–200-km intervals during each mission. The dropwindsondes sample the atmosphere below the aircraft flight level (near 150 hPa). In those cases in which additional aircraft supplemented the G-IV, 20–25 dropwindsondes were released at the same horizontal resolution from around 400 (P-3) or 300 hPa (C-130). The G-IV did not penetrate the TC core during surveillance missions; when other aircraft participated in the missions, at least one usually gathered data near the center. Hurricane Research Division (HRD) meteorologists aboard the NOAA aircraft validated the wind and thermodynamic data and generated standard (TEMPDROP) messages for transmission to NCEP and the Global Telecommunications System in preparation for assimilation into numerical models until 2006. This task on the G-IV has been carried out by NOAA/Aircraft Operations Center meteorologists in most cases since then.
The version of the GFS operational at the time of each mission was used to assess the impact of the dropwindsonde data. Several upgrades to the quality control algorithm, vortex initialization, data assimilation, and global spectral model were made during the 3 yr covered in this study. The quality control algorithm involves optimal interpolation and hierarchical decision making to evaluate observations before they are input to the analysis (Woollen 1991). Vortex relocation (Liu et al. 2000) places TCs in the first guess (background) field to their operationally analyzed positions (as in Kurihara et al. 1995) ensuring that they are initialized in the proper locations. The data assimilation scheme was the spectral statistical interpolation (Parrish and Derber 1992): the background field (the previous 6-h forecast) is combined with observations with a three-dimensional variational multivariate formalism. The global spectral model horizontal resolution was spectral triangular 254 (T254), and the vertical coordinate extended from the surface to about 2.7 hPa with 64 (L64) unequally spaced sigma levels on a Lorenz grid (Caplan et al. 1997; Surgi et al. 1998) in 2004; the resolution was increased to T382L64 in 2005. Minor changes to model physics were made during the 3 yr encompassed in the study.
c. Choice of data to assimilate
Multiple GFS runs were made for each mission. The first run has all dropwindsonde observations assimilated (hereafter operational). To demonstrate the effectiveness of each of the three targeting techniques (described below) to define sensitive regions, additional GFS runs have been performed when less than two-thirds of the dropwindsonde observations fully sample a particular target, as in Aberson (2003). All dropwindsonde data outside target regions (for any technique), or within targets that are not fully sampled, are removed from the data assimilation system. The numerical values provided by each technique are ignored, as in Aberson (2003) since any maximum is expected to be an area of relative sensitivity. The choice of which dropwindsondes to include in each test run is therefore subjective, but consistent from case to case and between techniques. The key is to sample targets so as to limit potentially spurious increment spread outside the observed region into locations where they are likely to grow. In this way, the growth of errors introduced into the model by the suboptimal data assimilation scheme is minimized. This agrees with Bergot et al. (1999), who state that with then-current data assimilation techniques, the entire target must be sampled, not only the extremum. All other observations from the NCEP “final” archive were ingested into the assimilation system for both sets of runs.
d. Targeting techniques
1) Ensemble variance
The EV from the NCEP Global Ensemble Forecast System (GEFS) provides a general approach to TC targeting (Aberson 2003). Random perturbations are evolved and rescaled using a low-resolution (T126) version of the GFS (Toth and Kalnay 1993; Lorenz and Emanuel 1998). Since a forecast is the first guess for a subsequent GFS cycle, locations in which perturbations are large are those in which initial condition errors have recently grown. The TC tracks depend on the environmental DLM flow, so the DLM wind variance at the targeting (observing) time is used. Aberson (2003) showed that GFS DLM increments grow where the EV is large, and decay where it is small, and that the growth is independent of the increment size.
Numerous upgrades to the GEFS have been implemented since the Aberson (2003) study. Most involve perturbation rescaling, masking, and sizing. On 16 August 2005, the breeding cycle was reduced from 24 to 6 h, and a TC vortex relocation technique was instituted; the latter change increased the GEFS effectiveness by requiring that the TC existed near the diagnosed location. On 30 May 2006, the ensemble size was increased from 10 to 14 every 6 h, and the ensemble transform technique replaced breeding (Wei et al. 2008). However, the essence of the original breeding technique remains within the GEFS.
The ETKF uses ensemble-based data assimilation theory (Bishop et al. 2001) to predict the reduction in forecast error variance within a verification area for feasible deployments of targeted observations. To mimic operations, the NCEP GEFS output available when flight track planning must be completed that initialized 48 h prior to the observing time ta is used. The ETKF computation is a two-stage process. First, the analysis error covariance matrix at ta pertaining to the routine observational network of rawinsonde and satellite-based temperature observations 𝗣r(ta) is found by solving the Kalman filter error statistics equations. Next, the analysis error covariance matrix for the observational network augmented by the qth hypothetical targeted observation 𝗣q(ta) is computed. The associated “signal covariance” matrix, or the reduction in forecast error covariance valid at the verification time (here, tυ = ta + 48 h), is deduced. The trace of this matrix localized within a verification region, defined as a 500-km radius circle centered at the Official NHC forecast location at tυ, is referred to as the “signal variance.” The ETKF guidance represents this as a function of the central location of adjacent targeted observations. The area with the highest signal variance within the verification region (the target) is deemed optimal for sampling. Since TC tracks depend on the environmental DLM flow, that flow is used as the metric in this study. Further details are provided in Majumdar et al. (2006).
3) Navy Operational Global Atmospheric Prediction System SVs
Localized SVs are a type of targeted analysis error covariance (AEC) optimals that have been used to identify observation target areas for both midlatitude storms and TCs (Palmer et al. 1998; Buizza and Montani 1999; Majumdar et al. 2006; Reynolds et al. 2007). The optimals of the tangent forward propagator 𝗟(ta; tυ) and its adjoint 𝗟T sample the directions of maximum growth during the optimization time interval between ta and tυ, evolving into the leading eigenvectors of the forecast error covariance matrix 𝗣(tυ). For an analysis error covariance metric = 〈v(ta); [𝗣a(ta)]−1v(ta)〉 and verification-time metric = 〈v(tυ); [𝗣υ(tυ)]−1v(tυ)〉, where 〈·;·〉 indicates an inner product, the analysis-time optimals vi(ta), where i indicates an index number, are computed via the eigenvalue problem:
where 𝗚 localizes perturbations within the verification region (Buizza 1994). The square roots of the eigenvalues σi are the singular values, and the eigenvectors vi(ta) are the right SVs, of 𝗚𝗟 with respect to the metrics (Noble and Daniel 1977).
Targeted AEC optimals in which 𝗣a and 𝗣υ are identical, diagonal, and elements equal to the fixed total energy weights, such that (𝗣a)−1 = (𝗣υ)−1 = 𝗘 and ‖v‖𝗘2 = 〈v; 𝗘v〉, are known as total energy SVs (TESVs). The sensitivity pattern s is a composite of the vertically integrated total energy of the leading SV weighted by the singular values:
where ej(x, t) is the vertically integrated total energy of the jth SV at location x, and the first three SVs are used. The TESVs are calculated using the Navy Operational Global Atmospheric Prediction System (NOGAPS; Hogan and Rosmond 1991; Peng et al. 2004) tangent and adjoint models. The SVs are calculated at a reduced T79L30 resolution, although the linearization is based on the trajectory from the full physics, high-resolution (T239L30) operational NOGAPS forecast. The models include surface drag and horizontal and vertical diffusion, but not moist processes (Rosmond 1997). The optimization time and lead time are both 48 h. A local projection operator is used to define the verification region, here a large fixed area covering the Gulf of Mexico eastward (i.e., the box in Fig. 4c). Examination of the final-time SVs confirms that the targets are relevant to the forecast of interest.
Although the SVs used here are from the NOGAPS forecasting system, the similarity noted in previous studies of SVs calculated using different forecasting systems supports their relevance for NCEP forecasts. For example, Majumdar et al. (2006) find that for 78 guidance products for 2-day forecasts for the 2004 Atlantic hurricane season, the European Centre for Medium-Range Weather Forecasts (ECMWF) and NOGAPS SV guidance identifies similar synoptic-scale target regions in 90% of the cases. Wu et al. (2009) likewise find strong similarities between SVs calculated using the ECMWF, NOGAPS, and the Japan Meteorological Agency (JMA) systems for typhoons in the western Pacific for the 2006 season (e.g., the products on the large domain were similar to each other in about 80% of the 84 cases considered). Although a comparison with NCEP SVs is not possible, the strong similarity between the ECMWF, NOGAPS, and JMA SV guidance suggests that the sensitive regions identified by the NOGAPS system will be relevant to other large-scale (i.e., global) forecast systems.
Table 1 shows homogeneous comparisons of GFS track forecast errors for runs in which all dropwindsonde observations (operational), or subsets of observations that fully sample each of the three sets of targets, are assimilated. The two sets of forecast errors are statistically significantly different at the 85% level at 12, 24, 48, 60, 84, and 108 h in the EV cases, though only at the last two times does the subset improve the forecasts. Minor improvements to the forecasts during the critical watch and warning period (24–60 h before landfall when preparations must be completed) are produced when all data are assimilated.3 The two sets of forecast errors are statistically significantly different at the 85% level at 12, 24, 84, and 108 h in the ETKF sample; the first two forecast times are degradations, and the last two are improvements, so these results are slightly better than those for the EV cases. Fewer cases are available for the SV targets than for the others because the relatively large SV target size prohibits full sampling of the entire target regions with the usual one-aircraft missions. The two sets of forecast errors are statistically significantly different at the 85% level from 24 to 72 h, and the differences are improvements at all these forecast times. As a result of the small sample size, any conclusive statement that SVs are a superior targeting technique is premature. The important result is that, for the ETKF and SV techniques, the data outside the target regions have no, or very little, positive impact on the ultimate track forecasts; for the EV, the data outside the target regions has some positive impact early in the forecast. A discussion of two examples in which all three targeting techniques are available, and of another illuminating example, follows.
a. Hurricane Emily: 0000 UTC 18 July 2005
Figure 1 shows the operational GFS forecast track, as well as those from runs in which data subsets, depending upon the targeting techniques, were assimilated. The removal of the nontargeted data had a small positive impact during the first 2 days of the forecast, though by the approximate landfall time, the forecast spread from the various model runs increased to about 250 km. This suggests that all three targeting techniques are effective in discerning the regions in which additional data will have a large impact on the subsequent forecast.
Emily was located in the northwestern Caribbean Sea moving westward under the influence of a strong subtropical high centered over the southeastern United States at the targeting time (Fig. 2). Some of the guidance suggested a northward turn before a possible Texas landfall since the ridge extended only to the western edge of the Gulf of Mexico. NHC tasked a two-aircraft mission, with the G-IV sampling the region around Emily and the southeastern Gulf of Mexico, and a C-130 sampling the remainder of the Gulf of Mexico (Fig. 3). The two-plane mission allowed for adequate sampling of the target regions from all three techniques (Fig. 4).
Runs using only those data within fully sampled targets provided better forecasts than the operational run. The assimilation of data outside target regions that were not fully sampled caused a southward forecast bias (Fig. 1). The EV target was nearly symmetric around Emily (Fig. 4a), so that most of the data in the Gulf of Mexico, except those closest to the Yucatan peninsula, those from two dropwindsondes released near 25°N, and from three in the Caribbean Sea, were removed from the EV assimilation. The ETKF and SV targets were both located on the eastern side of Emily in the region of anticyclonic curvature of the DLM flow (Figs. 4b,c, and 2). Data from three dropwindsondes along the northern and western coasts of the Yucatan peninsula and from the easternmost two dropwindsondes were removed from the SV assimilation, in addition to the Gulf of Mexico dropwindsonde data removed in the EV run, resulting in the northernmost forecast. The ETKF Gulf of Mexico data were the same as for the EV run, but the data from the two easternmost and one southernmost dropwindsondes were removed. The EV and ETKF forecast tracks were closest to each other because their targets were the most similar to each other, though they slowly diverged through the forecast. The removal of the dropwindsonde data in the Gulf of Mexico appears to cause subtle differences in the initial analyses leading to the northward motion in the targeted runs. The EV dropwindsonde analysis had a region of high relative vertical vorticity in the upper troposphere over the north-central Gulf of Mexico (Fig. 5a); the operational analysis with all the dropwindsonde data had a weaker feature than the EV analysis (Fig. 5b). The circulation moved slowly northwestward in the targeted runs (EV being representative; Fig. 5c), and its resultant inverted trough was located over Louisiana by 36 h into the forecast; the circulation moved slowly northward and weakened in the operational run (Fig. 5d). The inverted trough in the targeted runs allowed for the enhanced northward component of motion resulting in improved forecasts versus the operational run.
The nontarget dropwindsonde data in the Gulf of Mexico were obtained from the C-130 flying near 350 hPa, below the normal G-IV flight altitude. The relatively strong upper-level feature that was stronger in the operational run than in the targeted runs resulted from the data assimilation system spreading the good dropwindsonde information from below 350 hPa upward into a region with no data. The vertical spread of information into this data-sparse region is similar to the horizontal spread that can lead to forecast failures (Aberson 2002, 2003, 2008). Though sampling the entire troposphere in this region may not have led to a forecast track improvement, this case illustrates that the data assimilation can spread information vertically into data-sparse regions. Obtaining soundings through the depth of the troposphere for improving hurricane track forecasting may be important since the DLM flow usually steers tropical cyclones. This case confirms that the removal of data outside fully sampled target regions can lead to the larger forecast track improvements than the inclusion of all the data (Aberson 2003).
b. Hurricane Ophelia: 0000 UTC 11 September 2005
Figure 6 shows the model forecast tracks for Hurricane Ophelia initialized at 0000 UTC 11 September 2005. Ophelia was meandering slowly off the southeastern United States coast between two anticyclones: one over the central Mississippi River Valley and the other in the central Atlantic Ocean (Fig. 7). Other features complicating the forecast were a trough extending southwestward from Ophelia across central Florida, and three shortwave troughs moving through the longwave trough off the East Coast (one moving southward over New England, one over New Brunswick, and the other at the base of the trough south of Newfoundland). Two saddle points were between Ophelia and nearby troughs: one northeast and one southwest of the storm. Finally, a wind speed maximum and associated relative vorticity maximum are seen rotating around the western edge of the subtropical ridge southeast of Ophelia.
A one-plane mission was tasked for the G-IV because of the proximity and potential threat to the coastline. The targeted data in all three cases caused the model to accelerate Ophelia northward more than in the operational run, degrading the forecasts. The main difference is that the targeted data projected a landfall in New England, whereas the run with all the dropwindsonde data did not forecast landfall there.
The EV found three targets near Ophelia (Fig. 8a): 1) Ophelia itself, 2) the saddle point between Ophelia and the midlatitude trough to the northeast, and 3) the wind speed maximum along the western edge of the mid-Atlantic subtropical high; the third feature was not fully sampled. The ETKF identified three targets (Fig. 8b), the first and third of the EV targets and the base of the trough northeast of Ophelia; the last of these was not sampled as it was outside aircraft range. Three SV target regions were seen (Fig. 8c): 1) the eastern half of Ophelia, 2) the trough extending southwestward from Ophelia to central Florida, and 3) the midlatitude shortwave trough over New England; the first two (with concurrent rawinsonde releases in Florida and the Bahamas) were fully sampled, but the lack of data off the New England coast prevented full sampling of the third feature. In the discussion of this case, the impacts in the SV run are representative of the others, so only the SV impact results are shown.
A large amount of the dropwindsonde data within target regions were removed in the SV and EV runs because the respective targets did not meet the criterion of being fully sampled. Though the data outside the target regions had only a small impact on the track forecasts in the Emily case, the data removed in the current case were within partially sampled target regions, and their removal led to large short-range track forecast improvements. The cause of these improvements was that the runs in which the targeted subsets of data were assimilated initialized the midlatitude vorticity maximum/confluence region northeast of Ophelia (identified as an ETKF and EV target) and the southerly flow between Ophelia and the subtropical ridge to its east (identified as an ETKF and EV target) to be stronger than in the operational run initial condition. The vorticity maximum/confluence northeast of Ophelia (the cyclonic difference streamlines in Fig. 9a) was advected westward, and the enhanced southerly flow to the east of Ophelia moved northward. The two features combined and acted to advect Ophelia northward more rapidly in the targeted runs than in the operational one during the early part of the forecast. These increments are expected to grow because each of these regions was identified as targets; the actual growth might be erroneous since the targets were not fully sampled during the mission. This case shows that the targeting techniques can be effective in discerning which regions will have error growth that affect the TC track forecast; it also shows the importance of fully sampling the target regions, because removing data that only partially sampled targets improved the track forecasts.
c. Hurricane Ivan: 0000 UTC 11 September 2004
Relatively small track forecast differences between the targeted and symmetrically sampled data are shown in the above two cases. The Hurricane Ivan case initialized at 0000 UTC 11 September 2004 shows that large differences can occur. Figure 10 shows the ETKF and operational forecast tracks for this case. Neither SV nor EV cases were run because few or no data outside their targets were gathered. Ivan was moving toward the west-northwest just south of Jamaica. A ridge extended east-northeastward from central Florida, and a large mid- and upper-level cyclonic circulation was centered about 2500 km northeast of Ivan (Fig. 11). The operational forecast track called for Ivan to recurve and make landfall near Tampa Bay.
A one-plane mission was tasked for the G-IV because of the potential threat to Florida and the Gulf Coast. The ETKF target was centered on Hurricane Ivan (Fig. 12) with weak secondary maxima located outside the sampled region south of Guatemala and on the western side of the cyclonic circulation northeast of Ivan. The data obtained between 67° and 70°W were removed from the ETKF run; the impact of these data was spread northeastward by the data assimilation into the cyclonic circulation northeast of Ivan, where rapid error growth is expected by the ETKF. This led to a westward extension of the low, resulting in a northward component of motion to Ivan (Fig. 13). This case again shows the effectiveness of the ETKF in finding where errors affecting the particular forecast are expected to be large and grow and suggests that data should be obtained mainly close to targets since the then-current data assimilation techniques may erroneously spread the data impact into other regions degrading the forecast (Aberson 2003).
Using oversampled datasets obtained during synoptic surveillance during 2004–2006, three targeting techniques to find regions in which extra dropwindsonde observations should improve TC track forecasts are investigated with a series of Observing System Experiments. Since all the dropwindsonde data obtained in these cases led to substantial average forecast track improvements, in the current study, either all the dropwindsonde data or only those data within the fully sampled targets are included. This study was performed to examine whether using the particular subsample of data obtained only in fully sampled target regions provides the same average results as using all the dropwindsonde data for each of the three targeting techniques. An affirmative result would suggest that the particular technique tested is able to accurately define sensitive regions. This information could be used to optimize the sampling strategy for tropical cyclone targeting. Because all three techniques show that the removal of at least 33% of the dropwindsonde data, those data outside fully sampled target regions, rarely leads to forecast degradations and may lead to improvements, they all show some promise in obtaining maximal forecast improvements while limiting flight time and dropwindsonde expendables. The SVs provided the best forecasts, though the sample size is too small to make any operational recommendations.4
Individual cases are examined to deduce the impacts of the removal of at least one-third of the data, those residing outside fully sampled target regions, from the model data assimilation cycle. The increments within (without) target regions amplify (decay) in all cases, suggesting that each of the techniques studied has promise in identifying regions in which the biggest forecast impact will be obtained. Negative impacts of data removal are generally due to imperfect data assimilation systems (Aberson 2008), and they can be mitigated with improved sampling strategies. Spread of the data vertically and horizontally into relatively data-sparse regions, in some cases into secondary target regions, caused the data outside target regions to degrade the forecasts. The results continue to suggest that data should be obtained mainly within and immediately around target regions so as to adequately sample them and to limit spread of data, both horizontally and vertically, into regions with relatively sparse data, and which the targeting techniques suggest will have amplifying errors.
These results can be extended to other, similar observing systems, such as rawinsondes and satellite wind and radiance data. During threatening situations, NHC requests that regular observing sites in the southeastern United States release off-time (0600 and 1800 UTC) rawinsondes. Since rawinsonde data are obtained through the troposphere in regularly spaced grids like dropwindsonde data, the techniques reported here may be used to decide whether to request off-time releases, to maximize the data impact if such a request is made, and to include additional regions outside the southeastern United States.
The first author thanks NCEP/EMC for help in running the models and in providing the computer resources to make this study possible. Bob Kohler and Bill Barry provided computer support at HRD. Altug Aksoy, Sundararaman Gopalakrishnan, Mike Jankulak, Tomislava Vukicevic, and an anonymous reviewer improved previous versions of this manuscript. The authors thank the NOAA/Aircraft Operations Center (AOC) flight crews, AOC G-IV project manager Jack Parrish, and HRD personnel who participated in the flights, in addition to Air Force C-130 crews that also provided surveillance data over the years. CAR gratefully acknowledges the support from the sponsor, ONR PE-0601153N, and computer resources provided by the Department of Defense High-performance Computing program. The Joint Hurricane Testbed funded SJM and BJE for some of this work.
Corresponding author address: Sim D. Aberson, NOAA/AOML/Hurricane Research Division, 4301 Rickenbacker Cswy., Miami, FL 33149. Email: firstname.lastname@example.org
This article included in the Targeted Observations, Data Assimilation, and Tropical Cyclone Predictability special collection.
Each of the targeting subsamples shows similar improvements, though they are slightly larger (18%–29%) during the first 24 h of the forecasts. This could be because many of these cases include sampling by multiple aircraft; Aberson (2002) showed that the number of aircraft sampling the tropical cyclone environment was related to the amount of forecast improvement in each case.
Contrary to these results, Aberson (2003) reported that the assimilation of those dropwindsonde data within and around target regions as defined by the EV provided statistically better forecasts than those using all the data. Since 1999, the sampling strategy employed during missions has largely followed recommendations reported in Aberson (2003). The cases reported here are generally those in which additional aircraft increased the data coverage, improving forecasts (Aberson 2002), or in which targets could not be fully sampled with one aircraft, leading to degradations (Aberson 2003). These results should therefore not be construed as meaning that the EV technique is no longer valid, nor that it is less capable than the other techniques to define sensitive regions.