The influence of assimilating enhanced atmospheric motion vectors (AMVs) on mesoscale analyses and forecasts of tropical cyclones (TC) is investigated. AMVs from the geostationary Multifunctional Transport Satellite (MTSAT) are processed by the Cooperative Institute for Meteorological Satellite Studies (CIMSS, University of Wisconsin–Madison) for the duration of Typhoon Sinlaku (2008), which included a rapid intensification phase and a slow, meandering track. The ensemble Kalman filter and the Weather Research and Forecasting Model are utilized within the Data Assimilation Research Testbed. In addition to conventional observations, three different groups of AMVs are assimilated in parallel experiments: CTL, the same dataset assimilated in the NCEP operational analysis; CIMSS(h), hourly datasets processed by CIMSS; and CIMSS(h+RS), the dataset including AMVs from the rapid-scan mode. With an order of magnitude more AMV data assimilated, the CIMSS(h) analyses exhibit a superior track, intensity, and structure to CTL analyses. The corresponding 3-day ensemble forecasts initialized with CIMSS(h) yield smaller track and intensity errors than those initialized with CTL. During the period when rapid-scan AMVs are available, the CIMSS(h+RS) analyses offer additional modifications to the TC and its environment. In contrast to many members in the ensemble forecasts initialized from the CTL and CIMSS(h) analyses that predict an erroneous landfall in China, the CIMSS(h+RS) members capture recurvature, albeit prematurely. The results demonstrate the promise of assimilating enhanced AMV data into regional TC models. Further studies to identify optimal strategies for assimilating integrated full-resolution multivariate data from satellites are under way.
Improving forecasts of tropical cyclone (TC) track and intensity remains a challenge in numerical weather prediction (NWP). This is partially due to the difficulty in prescribing accurate initial conditions of the TC structure and its surrounding environment. Since a TC spends most of its lifetime over the ocean, initial conditions are crucially dependent on the accurate assimilation of data from geostationary and polar-orbiting satellites. Given the increasing quantity, variety, and resolution of satellite data available, it is necessary to seek optimal methods to exploit the use of these multiple and integrated satellite datasets.
One type of satellite data that is expected to improve the representation of TC structure and its environmental flow is atmospheric motion vectors (AMVs). AMVs are derived from sequential satellite images by tracking targets including cirrus cloud edges, gradients in water vapor, and small cumulus clouds (Velden et al. 1997). They provide unique detailed coverage of mid- and upper-tropospheric wind data over the ocean. AMVs are assimilated routinely in all operational global numerical weather prediction systems, and they have been shown to produce significant positive impacts on the accuracy of global model initial conditions (e.g., Le Marshall et al. 2008a) and forecasts of tropical cyclone track (e.g., Goerss 2009). Methods to process and improve the accuracy of AMVs are evolving (Velden et al. 2005), and the Cooperative Institute for Meteorological Satellite Studies (CIMSS) at the University of Wisconsin–Madison provides specially processed hourly AMV data, together with more frequent AMV data if the “rapid scan” mode is activated on the satellite. Using these AMV data, it has been demonstrated that track forecasts in the Navy Operational Global Atmospheric Prediction System (NOGAPS) can be improved through a more accurate representation of the environmental flow (Langland et al. 2009; Berger et al. 2011). It was also evident in Soden et al. (2001), where the assimilation of AMVs in the Geophysical Fluid Dynamics Laboratory (GFDL) hurricane model led to improved TC track forecasts due to a more accurate representation of the steering flow. Le Marshall et al. (2008b) have also documented the impacts of AMVs in the Australian regional model, especially the Multifunctional Transport Satellite-1R (MTSAT-1R) hourly AMVs. To operationally use rapid-scan AMVs in their NWP system in 2013, the Japan Meteorological Agency (JMA) found that the assimilation of MTSAT Rapid-Scan AMVs in their operational mesoscale and four-dimensional variational data assimilation (4D-Var) system provided slight improvements to forecasts of typhoon intensity but not the track, for three typhoon cases in 2010 (Yamashita 2012). However, the qualitative and quantitative influence of assimilating hourly and rapid-scan AMV data on regional model predictions of TC structure and track still need to be investigated thoroughly.
The field of data assimilation specific to TCs has evolved rapidly in recent years. In particular, the ensemble Kalman filter (EnKF) has become widely used in the research community, due in part to the practicality of its application and the physical realism of the flow-dependent error covariance structure. The assimilation of a variety of data including airborne Doppler radar, surface best track data, global positioning system (GPS) radio occultation refractivity, and targeted dropwindsonde observations has been explored in several studies. These studies have demonstrated that the EnKF is effective in providing a realistic initial representation of the TC structure and environment, and therefore an improved forecast (e.g., Chen and Snyder 2007; Torn and Hakim 2009; Wu et al. 2010; Weissmann et al. 2011; Zhang et al. 2011; Aksoy et al. 2013; Jung et al. 2012; Liu et al. 2012; Weng and Zhang 2012). To offer more robust statistics, Torn (2010) employed a mesoscale EnKF in the Weather Research and Forecasting Model (WRF) framework for a sample of 10 Atlantic basin TCs and concluded that the EnKF-initialized WRF forecasts possessed lower errors in track and TC wind radii than corresponding WRF forecasts initialized from the National Centers for Environmental Prediction (NCEP) Global Forecast System (GFS) and GFDL analyses. The recent introduction of the EnKF into the GFS framework yielded forecasts of TC track that were competitive with those from the European Centre for Medium-Range Weather Forecasts (ECMWF), and superior to those provided by several other modeling systems including the version of the NCEP GFS that was operational at the time (Hamill et al. 2011).
Given the promise of AMV data over the ocean and the EnKF data assimilation scheme, it is expected that the assimilation of high-resolution AMVs will improve regional model analyses and forecasts of TC structure and track. In this paper, an in-depth investigation is performed to demonstrate this concept for Typhoon Sinlaku (2008), a widely studied tropical cyclone whose track and structural evolution were difficult to predict.
The paper is organized as follows. An overview of Typhoon Sinlaku, the AMV data and its quality control, and the design of the data assimilation experiments are presented in section 2. Verifications and diagnostics of the influence of assimilating the AMV data on EnKF analyses are presented in section 3. The corresponding influence on 72-h ensemble forecasts is illustrated in section 4, followed by a summary and discussion in section 5.
2. Typhoon Sinlaku, AMV data, and assimilation framework
a. Typhoon Sinlaku (2008)
To examine questions on TC predictability and data assimilation, the Office of Naval Research (ONR) Tropical Cyclone Structure 08 (TCS-08)/The Observing System Research and Predictability Experiment(THORPEX) Pacific Asian Regional Campaign (T-PARC) field experiment was conducted in the western North Pacific basin in 2008. The most intensively observed TC during the field campaign was Typhoon Sinlaku, which underwent two phases of intensification during its lifetime. Many papers have been published on Sinlaku, including the evolution of its structure (e.g., Wu et al. 2012; Huang et al. 2012), sensitivity of the track forecast to the environmental flow (e.g., Komaromi et al. 2011; Majumdar et al. 2011; Yamaguchi and Majumdar 2010) and the impact of assimilating data from dropwindsondes and in situ observations (e.g., Harnisch and Weissmann 2010; Kim et al. 2010; Kunii et al. 2012). Among the limited studies completed on the influence of assimilating satellite data are Liu and Li (2010) who assimilated water vapor and temperature soundings from the Atmospheric Infrared Sounder (AIRS) in the WRF/EnKF system, and Berger et al. (2011) who assimilated hourly and rapid-scan AMVs processed by CIMSS in the Navy’s global 4D-Var scheme.
As reported in these and many other studies, the evolution of Sinlaku was complicated. After its formation east of the Philippines on 0000 UTC 8 September 2008, Sinlaku was upgraded to a tropical storm by the JMA at 1800 UTC on the same day. Sinlaku then underwent a period of rapid intensification, reaching typhoon status on 1200 UTC 9 September, and intensifying further to 937 hPa by 1800 UTC 10 September with two concentric eyewalls (Figs. 1a–c). After this period, the inner eyewall dissipated, and the minimum sea level pressure of Sinlaku did not fall back below 950 hPa over the subsequent 48 h. During the entire period between its formation and landfall in Taiwan on 13 September, the motion of Sinlaku was slow and meandering, with high uncertainty in its track forecasts (Yamaguchi and Majumdar 2010). After landfall and subsequent recurvature on 15 September, Sinlaku decayed to tropical storm status. As it traveled northeastward toward Japan, Sinlaku then reintensified into a typhoon on 18 September. Through the rest of this paper, attention is focused on improving the analysis and forecast of Sinlaku’s track, structure, and intensity during the period of rapid intensification, prior to its landfall in Taiwan.
b. Atmospheric motion vectors
During our study period in 2008, AMVs derived from the geostationary satellite MTSAT-1R were produced operationally by JMA every 6 h over East Asia and the western North Pacific Ocean. These datasets were accessible to NCEP for real-time assimilation into the GFS model. However, only a fraction of the available JMA AMVs that pass through the NCEP thinning and quality control processes are accepted and included in the operational assimilation. These AMV datasets constitute what is assimilated in our study control runs, since our assimilation experiments employ the operational NCEP GFS as background fields. Additionally, during the TCS-08/T-PARC field campaign in 2008, high-resolution (space and time) AMV datasets were prepared hourly by CIMSS using routinely available 30-min image triplets (Berger et al. 2011). The CIMSS AMV automated derivation algorithm is similar to what is employed operationally at National Oceanic and Atmospheric Administration/National Environmental Satellite, Data, and Information Service (NOAA/NESDIS) for the Geostationary Operational Environmental Satellite (GOES) AMV dataset production. Although operational AMV processing centers and CIMSS generally derive a similar volume of AMVs from common geostationary satellite imagery [infrared (IR), visible (VIS), and water vapor channel (WV)] over the same region, the CIMSS processing method provides more detailed coverage of AMVs over tropical cyclones when they are present (Velden et al. 1998; Sears and Velden 2012). We found this to be the case with Sinlaku (not shown). Finally, the rapid-scan (RS) imaging mode on the geostationary satellite MTSAT-2 was specially activated for several days by JMA shortly after 1200 UTC 10 September 2008, at which time TC Sinlaku was a category-2 typhoon two days prior to its landfall in northern Taiwan. During this period, rapid-scan AMVs (from IR and VIS) were derived by CIMSS using successive 15-min image triplets. The shorter image intervals allow for higher-quality AMVs (Velden et al. 2005).
All of the CIMSS AMVs assimilated in this study are first passed through quality control steps (next section), then “superobbed” (see section 2c). An example of the horizontal and vertical distribution of superobbed AMVs in TC Sinlaku and its vicinity on 0000 UTC 11 September 2008 is illustrated in Figs. 1d–f. Figure 1d shows the spatial distribution of the AMVs from the NCEP dataset. Figure 1e shows the CIMSS hourly dataset. As previously indicated, the volume of AMVs in Fig. 1d is greatly reduced as a result of the strict quality control employed by the NCEP system. In contrast, the AMV coverage from the CIMSS hourly dataset is much broader and the number of AMV superobs is almost 10 times that present in the NCEP dataset. Although the majority of the CIMSS hourly AMVs are located above 400 hPa, about 20% are located below 700 hPa. Figure 1f shows the spatial distribution of AMV superobs from the combined CIMSS hourly and rapid-scan datasets. The activation of the rapid-scan mode yields many more vectors below 400 hPa. This dataset feature is attributable to the enhanced tracking ability of low-level cumuliform type cloud tracers, mainly from the visible channel, when higher-frequency imagery is available.
c. AMV quality parameters and observation error
The CIMSS AMV data records come with two appended parameters that estimate each vector’s quality: a quality indicator (QI) and an expected error (EE). The QI consists of vector coherency and consistency checks (Holmlund 1998), with normalized values ranging from 0 to 1. The EE is a modified QI that converts to the more physically intuitive units of meters per second, and also includes consideration of the vector speed, the environmental wind shear and temperature, and the assigned vector pressure level (Le Marshall et al. 2004). In this study, the hourly CIMSS AMVs are assimilated only if the QI is equal to or larger than an empirically determined 0.5, and for the RS AMVs of slightly higher quality the QI threshold is ≥0.6. In addition, AMVs meeting these QI thresholds but with EE values ≥4.5 m s−1 are filtered out unless the AMV is >25 m s−1 and with an attending QI ≥0.7. This latter constraint is invoked since the EE is sensitive to higher wind speeds. This is most evident in Fig. 2, which shows the distribution of the two CIMSS AMV datasets as a function of QI and EE. The RS VIS AMVs are more prevalent near the storm center where higher wind speeds exist, leading to higher EE values. No AMVs are derived from the WV in RS mode since WV AMVs do not benefit from the more frequent imaging (Velden et al. 2005).
At this point, the superobbing is performed on those AMVs that pass the above QC tests. The horizontal dimension of the prism within which the AMVs are averaged is chosen to be 90 km × 90 km, reflecting 2–3 times the model grid size that is empirically used (R. Torn 2012, personal communication). The vertical dimension of the prism is 25 hPa. We did different tests on the sensitivity of the horizontal dimension of the superob prism and found that the size of 90 km provided the most accurate analyses of initial track and intensity of Sinlaku. AMVs within this 90 km × 90 km × 25 hPa prism are averaged with uniform weight (i.e., arithmetic mean). The final preprocessing step is the assignment of observation errors for the superobbed AMVs. The NCEP and CIMSS hourly AMV datasets follow the operational NCEP observation error statistics for AMV as shown in Table 1. The assigned errors are a function of AMV height. Because of the greater ambiguity in the broad height assignment of WV AMVs, the observation errors are relatively higher than those AMVs that are derived from infrared or visible channels that mainly reflect more discrete cloud tops. For the rapid-scan AMVs, there is a greater chance of high EE values (Fig. 2), and therefore the observation errors in Table 1 are multiplied by 1.5 when the respective EE exceeds 3.0 (the error statistics of AMVs at NCEP can also be found online at http://research.metoffice.gov.uk/research/interproj/nwpsaf/satwind_report/amvusage/ncepmodel.html). With these varying characteristics of AMVs, a more appropriate strategy might be a flow-dependent error assignment scheme (a subject for future investigation).
d. Modeling and data assimilation framework
The Advanced Research WRF (ARW-WRF) version 3.1.1 is employed in this paper (Skamarock et al. 2008). The specification of the ARW-WRF includes the following: the WRF single-moment 6-class microphysics scheme (WSM6), Rapid Radiative Transfer Model (RRTM) longwave radiation, the Dudhia shortwave scheme, the Kain–Fritsch cumulus scheme, the Yonsei University (YSU) boundary layer scheme, and the Noah land surface model. The data assimilation cycles use the ensemble adjustment Kalman filter (EAKF; Anderson 2001), hosted by the Data Assimilation Research Testbed (DART; Anderson et al. 2009). An 84-member ensemble is used in the cycling. To account for interactions between the TC and the far and near environments, EAKF assimilation is performed on a large domain with 27-km resolution grid covering the western North Pacific (0°–50°N, 100°–150°E) with 36 vertical levels from the surface to 50 hPa. A nested moving domain with 9-km resolution grid is activated in the forecast cycling when the tropical cyclone is present. This field is then fed back to the larger domain. However, because of the limited computer resources, no assimilation is performed on the 9-km nested moving domain. Ensemble mean initial and lateral boundary conditions are obtained from the 1° × 1° global analysis produced by NCEP. Random draws from a distribution of the forecast error covariance generated from the WRF three-dimensional data assimilation system (Barker et al. 2004) are added to ensemble mean fields to generate 84 initial and lateral boundary ensemble perturbations (Torn et al. 2006).
A fixed horizontal and vertical localization (Gaspari and Cohn 1999) is applied to all of the analysis increments from observations to reduce the impact of spurious covariance estimates. The half-width of the fixed horizontal localization is 650 km and the fixed vertical localization is 8 km. This means that at any model grid point, an observation will have no impact on the model state variables if it is located 1300 km beyond and 16 km above that grid point. The temporally and spatially varying adaptive inflation algorithm of Anderson (2009) is used to inflate the ensemble background error covariance. This algorithm uses the difference between the background ensemble mean estimate of an observation and the observed value along with the background and observational error variances to estimate the necessary inflation. The localized ensemble covariances that are used to distribute the impact of an observation to surrounding state variables are also used to distribute the updated inflation from an observation to surrounding state variables. Given that the background error covariance should be particularly large but is often underestimated in a tropical cyclone (because of the deficiency in ensemble size, nonlinearity, and model error), enhanced inflation is necessary to retain a necessarily high value of analysis error covariance after assimilation. Furthermore, if the observational density is large, enhanced inflation is also necessary to avoid an inappropriately small ensemble variance (Anderson 2009). As the tropical cyclone moves away, both the background error variance and observation density are expected to be smaller, and the need to inflate the background error covariance is less necessary. In other words, if the inflation is kept constant, a tropical cyclone can leave a trail of enhanced inflation that can persist for long periods in its wake. To avoid overinflation after a tropical cyclone passes, the inflation factor is damped by 10% (a value that was chosen by empirical testing) after each assimilation time. This procedure has little impact in locations that are routinely well observed since the observations consistently reinforce appropriate inflation values.
Three parallel WRF/EnKF experiments are initialized on 0000 UTC 1 September 2008, a week prior to the formation of Sinlaku. Conventional observational data assimilated include radiosonde winds, temperature, and specific humidity that are at least 200 km away from the TC center; aircraft winds and temperature; and surface pressure data and the Joint Typhoon Warning Center (JTWC) advisory TC position (latitude and longitude of minimum sea level pressure). In addition, the uncertainty of the JTWC advisory TC position estimates are assumed to be dependent upon wind speed, and are assigned the following values: 90 km for the maximum sustained wind <34 kt (1 kt = 0.5144 m s−1), 40 km for the wind >85 kt, and 60 km for intermediate wind (Torn and Snyder 2012). In addition to the conventional observations, the first experiment, denoted “CTL” also assimilates AMVs from the NCEP operational analysis dataset in a 3-h-wide window centered on 0000 UTC, 0300 UTC, etc. All observations in each 3-h window are assimilated as if they were taken at the center time of the window. The analyses are used to initialize an ensemble of 3-h forecasts for the next analysis time. The second experiment, denoted “CIMSS(h)” is the same as CTL except that it assimilates the hourly AMVs prepared by CIMSS. A third experiment, denoted “CIMSS(h+RS),” is the same as CIMSS(h) until the first time that rapid-scan data are assimilated on 1800 UTC 10 September 2008, and thereafter all the hourly and quality-controlled rapid-scan AMVs are assimilated in a 3-h wide window centered on 0000 UTC, 0300 UTC, etc. A summary of these three experiments is given in Table 2.
For the CIMSS(h+RS) experiment, rapid-scan AMV datasets are available each hour from 1700 UTC 10 September to 0400 UTC 13 September, with a short gap between 0900 and 1500 UTC 12 September. The number of superobbed AMVs assimilated in the CIMSS(h+RS) experiment as a function of analysis time is shown in Fig. 3. A sharp increase in the number of observations assimilated after 1800 UTC 10 September 2008 marks the beginning of the inclusion of rapid-scan AMVs. Also, a diurnal cycle can be observed from Fig. 3 during the inclusion of the rapid-scan AMVs, since the major contribution of extra AMVs is from the visible channel (see Figs. 2d–f). The majority of the enhanced rapid-scan AMVs are below 400 hPa, while the majority of CIMSS hourly AMVs are above 400 hPa.
3. Influence on WRF/EnKF analyses
a. Track and intensity analyses
The ensemble mean analyses are first verified on the 27-km domain against the position and intensity estimates from the JMA best track data (available online at http://www.wmo.int/pages/prog/www/tcp/documents/JMAoperationalTCanalysis.pdf) in which the Dvorak analysis method is utilized on cloud images from the MTSAT satellite. Prior to Sinlaku reaching tropical storm intensity, the ensemble members possessed very broad minima of mean sea level pressure (MSLP), and there was also ambiguity in distinguishing between the low-level circulation maxima associated with Sinlaku and those of neighboring atmospheric features. We therefore consider the analyses from 0000 UTC 9 September 2008 onward. To track Sinlaku in each of the ensemble members, the GFDL vortex tracker (available online at http://www.dtcenter.org/HurrWRF/users/downloads/index.tracker.php) is used throughout this paper. Between 0000 UTC 9 September and 1200 UTC 10 September, considerable differences exist between the CTL analyses and the CIMSS(h) analyses (Figs. 4a–c). The ensemble mean center of Sinlaku in the CTL is too far west of the best track, resulting in track errors above 100 km (Fig. 4a, blue line). In contrast, the ensemble mean center in CIMSS(h) is closer to the best track (although still to the west) with track errors consistently reduced by up to 100 km (Fig. 4a, green line). This example also highlights the difficulty in providing accurate analyses of the track when a tropical cyclone is in its early stages. JMAs operational TC analysis uniquely uses an early-stage Dvorak analysis (EDA) for TCs in generation stage, and the conventional Dvorak analysis for developing or mature TCs. Therefore, it is also worth noting that the error of the JMA best track is likely to be larger in weaker storms than in stronger storms. During the intensification period, the track errors are reduced in both analyses, although the errors in the CTL ensemble members remain larger than those of CIMSS(h). In its mature stage, Sinlaku moved slowly toward northeastern Taiwan from 0000 UTC 11 September to 1200 UTC 13 September with a well-defined center. The mean tracks in the three analyses CTL (blue), CIMSS(h) (green), and CIMSS(h+RS) (red) are very similar at this stage, with errors less than 50 km (Fig. 4a).
Prior to the intensification period, the majority of ensemble members in both the CTL and CIMSS(h) analyses exhibit very similar values of minimum MSLP (not shown). During the intensification between 0000 UTC 9 September and 0000 UTC 10 September, the ensemble mean of the CIMSS(h) MSLP analyses is about 5 hPa too high and that of maximum surface wind speed analyses is about 5–10 m s−1 too weak, but it importantly has a rate of deepening similar and closer to the JMA best track (Figs. 4b,c). In contrast, the mean analysis of MSLP and maximum surface wind speed in the CTL experiment is far weaker than that in the JMA best track and does not adequately capture the strength and timing of the deepening. After 0000 UTC 10 September, as Sinlaku enters its mature stage, the mean of the analysis values of MSLP in both the CTL and CIMSS(h) experiments does not approach the best track value of 937 hPa. One possible reason for this inconsistency is that the grid size of 27 km is not fine enough to adequately resolve the inner-core processes associated with the continued deepening. Between 0000 UTC 11 September and 1200 UTC 13 September, Sinlaku continues through its mature stage and then weakens. This is also the period during which the CIMSS(h+RS) experiment is computed. During this period, there are a few times when the ensemble mean MSLP and maximum surface wind speed in CTL analysis is closest to the JMA best track. However, the ensemble mean tracks of CIMSS(h) and CIMSS(h+RS) analyses are more accurate than that of CTL analysis. It is also worth noting that without the rapid-scan AMVs assimilated, the mean of the CIMSS(h) analysis values of MSLP and maximum surface wind speed are much weaker than that of CIMSS(h+RS) during this period. Overall, the representation of Sinlaku in the CTL experiment appears less accurate than in both CIMSS experiments for most of the analysis times.
The spread of the analysis ensembles reveals the estimation of the analysis uncertainty. Prior to Sinlaku entering its relatively steady state (1200 UTC 10 September), the CTL analyses have a broad track spread over 122°–127°E (Fig. 4d), with a general left bias relative to the best track. As expected, the spread decreases substantially as Sinlaku develops a more well-defined center and enters its mature stage. In contrast to the CTL analyses, the CIMSS(h) analyses of the track are more tightly clustered close to the best track, although still with a left-of-track bias (Fig. 4e). Figures 4b,c reveal that the CIMSS(h) analyses have a smaller spread of MSLP and maximum wind speed than CTL throughout the entire experiment. During the 36 h of intensification, the standard deviation steadily increases from 15 to almost 25 hPa for the CTL analyses. However, the corresponding standard deviation only stays at 8 hPa for the CIMSS(h) analyses. This indicates that the larger uncertainty associated with Sinlaku’s genesis was greatly reduced by the assimilation of CIMSS hourly AMVs. From 0000 UTC 11 September, it is also found that the track spread of CIMSS(h+RS) is fairly close to CIMSS(h), and the MSLP spread of CIMSS(h+RS) gradually increase to 15 hPa at 1200 UTC 13 September (Figs. 4a,d,e).
b. Analysis diagnostics
1) Against independent observations
The realism of the WRF/EnKF analyses for the three experiments is investigated via a comparison against independent observations that are not assimilated. First, radii of 34-kt winds in the WRF/EnKF analyses are compared against a modified version of the JTWC best track dataset,1 especially for the early stage of Sinlaku. Figure 5a shows the time series of radii of 34-kt winds in three WRF/EnKF analyses and radii of 34-kt winds from the modified JTWC best track data in the southeastern quadrant. Performance on the other three quadrants is similar. Prior to 1200 UTC 10 September, only a few CTL ensemble members possessed winds exceeding 34 kt, and 34-kt winds were absent in the ensemble mean field.
Comparisons were also performed between the spatial structures of the surface wind fields in the WRF/EnKF analyses and observations from the Quick Scatterometer (QuikSCAT), which was operational in 2008. The 2.5-km resolution wind fields produced by Brigham Young University (BYU; Halterman and Long 2006) in a 20° × 20° domain covering the TC are used here. A few passes of QuikSCAT over Sinlaku are close to the analysis time (allowing ±3-h difference from analysis time). An example of QuikSCAT sea surface wind image on 0944 UTC 9 September 2008 is shown in Fig. 5b. To eliminate the impacts of position spread on the spatial structure of wind fields, the 10-m horizontal wind from CTL and CIMSS(h) ensemble analyses at 1200 UTC 9 September 2008 (Figs. 5c,d) are relocated in a storm-centered perspective. The 27-km analysis is not expected to exactly reproduce the features identified in the 2.5-km QuikSCAT image. However, it is evident that the magnitude of the surface wind of the storm are better described by CIMSS(h) analysis. Three further QuikSCAT passes with adequate coverage of Sinlaku at later times (2141 UTC 11 September, 1005 UTC 12 September, and 2115 UTC 12 September) are compared with the corresponding WRF/EnKF analyses (not shown). However, given that the three WRF/EnKF analyses produced similar surface wind fields at these later times when Sinlaku was a mature typhoon, no significant differences between the respective 10-m wind structures were diagnosed.
Vertical profiles of tangential and radial wind in the WRF/EnKF analyses are also compared against targeted dropwindsonde data from Dropsonde Observations for Typhoon Surveillance near the Taiwan Region (DOTSTAR). Only those soundings that are within 30 min of the analysis times at which the full ensemble fields were able to be stored (0000 and 1200 UTC) are considered. Four dropwindsondes deployed fairly close to Sinlaku (about 200 km) on 0006 UTC 10 September, 0015 UTC 10 September, 2339 UTC 9 September, and 2355 UTC 9 September 2008 are selected for comparison with the CTL and CIMSS(h) analyses at 0000 UTC 10 September 2008. The vertical profile of tangential and radial wind relative to the TC center from the first of these dropwindsondes is illustrated in Fig. 6a alongside those produced by the corresponding WRF/EnKF analyses for CTL (blue) and CIMSS(h) (green). A qualitative inspection reveals that the dropwindsonde profile mostly lies in the middle of the ensembles for both the CTL and CIMSS(h) analyses, although the tangential wind in the dropwindsonde profile does not decay as sharply with height as in the analysis fields. The low-level radial inflow is generally stronger in the CIMSS(h) analyses than in the dropwindsonde data. On the other hand, the low-level radial inflow in the CTL analyses is generally weaker. The spread of the CTL analyses is higher throughout the column. The averaged root-mean-square (RMS) error of the CIMSS(h) tangential wind at this time, verified against the four aforementioned dropsondes, is reduced by over 55% compared with CTL, due to the strengthening of the tangential wind near the core with the assimilation of the extra AMVs (first column of Table 3). The RMS error of the radial wind is reduced by about 5% due to the extra AMVs, and a 17% improvement of temperature through the column is also realized. Near 0000 UTC 11 September 2008, another three dropwindsondes about 400–600 km from Sinlaku are selected for verification of the EnKF analyses, with an example shown in Fig. 6b. A qualitative comparison (see Table 3) indicates that the vertical profile of tangential wind best resembles that of CIMSS(h+RS), compared with CIMSS(h) and CTL. The RMS error of CIMSS(h) and CIMSS(h+RS) estimated tangential wind (radial wind) (temperature) are improved by 28% (34%) (25%) and 36% (38%) (39%), respectively, compared with the CTL estimations (Table 3). Given that there is already a large number of AMVs in the vicinity of Sinlaku in the CIMSS(h) dataset, the room for improvement due to the addition of rapid-scan winds is smaller. Accordingly the reduction of errors of tangential wind, computed by comparing the RMS errors in CIMSS(h+RS) to CIMSS(h), is 12%, and the corresponding error reduction in the radial wind is 7%. The addition of rapid-scan winds contributes to a 18% improvement in the temperature through the vertical over CIMSS(h).
2) Wavenumber-0 structure
Given that the 27-km horizontal grid size is not suitable for quantitatively understanding the inner-core structures of the TC, the following analysis diagnostics are instead mostly focused on a qualitative analysis of the broad primary and secondary circulation of the TC, via azimuthally averaged profiles. Figure 7a illustrates the ensemble mean azimuthally averaged tangential wind (black contour), radial wind (shading), and vertical wind (green contour) plus the radius of maximum wind (RMW, gray line) of the CTL and CIMSS(h) analyses at 1200 UTC 10 September. This time is close to the most intense stage of Sinlaku. It had been shown in Fig. 4c that the CIMSS(h) analysis is able to capture the rapid intensification in the early stage of Sinlaku from a tropical storm to a category-2 typhoon, while the ensemble mean minimum MSLP in the CTL analysis only deepened by about 10 hPa over the same period. The azimuthally averaged profile (Fig. 7a) suggests that the lower MSLP in CIMSS(h) is associated with a more intense primary and secondary circulation. A stronger tangential wind with a maximum value of 36 m s−1 situated about 70 km from the center is exhibited in the mean of the CIMSS(h) analyses. This sharp wind maximum is located near 850 hPa, whereas in CTL the broad weak wind is spread almost homogeneously from the top of the boundary layer to about 300 hPa. Also, the low-level radial inflow wind is about 2 times stronger than CTL ensemble mean analysis at the same time. The azimuthally averaged vertical wind is generally strongest at the uppermost levels, and its mean position is generally aligned with or lies within the average RMW. This analysis time is later chosen as one of the initial times for the 72-h ensemble forecasts, to be presented in section 4.
A day later at 1200 UTC 11 September, the mean analysis of CIMSS(h) still exhibits a stronger tangential wind (maximum value 4 m s−1 larger than the maximum value in CTL), a stronger and more radially tilted structure of vertical wind, and a tighter RMW than the mean analysis of CTL (Fig. 7b). Additionally, a stronger low-level outflow at 700 hPa is evident in CIMSS(h). However, it is at this time that the outflow near 100–200 hPa begins to weaken in CIMSS(h). At the same time, with the introduction of the rapid-scan AMVs over the previous 18 h, the mean analysis of CIMSS(h+RS) keeps the general structure as CIMSS(h) does, but with a more pronounced outflow and vertical wind aloft.
Another day later, at 1200 UTC 12 September, the ensemble mean vertical structure in the CTL analyses has now intensified. However, Sinlaku had been slowly weakening over the previous day (Figs. 4b,c), and the CIMSS(h) and CIMSS(h+RS) analyses are accordingly not intensifying the vertical structure (not shown).
3) Analysis increments
To further understand the role of the AMVs as Sinlaku developed from a tropical storm into a category-2 typhoon during its rapid-intensification period, the structure and magnitude of analysis increments for horizontal wind and temperature fields are examined in a vertical plane that shows the cross section of the TC.
Even though only a few AMVs are available in CTL (Figs. 8a,c), a large number are present in the vicinity of Sinlaku in CIMSS(h) especially in the upper levels (Figs. 8b,d). At 0000 UTC 9 September, when Sinlaku was a tropical storm, an anticyclonic increment is present to the west of the center and the associated cold increment is evident in CTL (Figs. 8a,c). On the other hand, a cyclonic increment throughout the depth of the storm is evident in CIMSS(h), suggesting that the assimilation of the AMVs at that time serves to strengthen the storm in the analysis (Fig. 8b). It is worth mentioning that other observation types may also contribute to the increments. For example, the assimilation of the TC positions may contribute to the analyses increments in the lower troposphere. However, by comparing the analysis increments between CTL and CIMSS(h) analyses, the AMVs coverage and quality are the only differences that significantly contribute to the analysis increments.
In addition to the dynamic fields being modified by the assimilation when many AMVs are present, the corresponding thermodynamic fields are also modified. This is evident in the CIMSS(h) analysis at the same time, in which the assimilation of mostly upper-level AMVs produces a moderate but deep warming increment through the troposphere (Fig. 8d), consistent with the deep cyclonic increment in Fig. 8b.
Similar structures in the analysis increment of meridional wind and temperature fields are evident through 0000 UTC 10 September in CIMSS(h), when analysis increments in CTL first starts to show sign of strengthening. In summary, together with sufficient upper-level AMV coverage in the vicinity of the tropical cyclone, the covariance structure in the EnKF enables the production of a physically consistent modification to the thermodynamic field through the depth of the troposphere.
4) Environmental wind fields
In addition to the modification to the storm structure, the effect of assimilating the AMVs on the local environmental flow is now investigated for two representative times, at the 200-, 500-, and 850-hPa levels. Because of the three-dimensional nature of (i) the distribution of AMV observations derived from moving clouds and water vapor gradients, (ii) the background error covariance in the EnKF, and (iii) the propagation of the impact of observations over time, we expect the AMV observations to influence the environmental analysis fields at all levels.
At 1200 UTC 9 September, when a large group of ensemble members of CTL analyses was found to move the TC erroneously far westward, the most significant difference between the CTL and CIMSS(h) AMV coverage in the near environment of Sinlaku is in the upper-level anticyclonic outflow (not shown, but similar to Figs. 1d,e). In Sinlaku itself, the ensemble mean analysis of CTL exhibits relative vorticity of up to 1.5 × 10−4 s−1 at 850 hPa, and closed isoheight in the vicinity up to 500 hPa. However, no clear center can be identified at 200 hPa in either the geopotential height or relative vorticity fields (Figs. 9a–c). On the other hand, the ensemble mean analysis of CIMSS(h) at this time exhibits relative vorticity of up to 2.5 × 10−4 s−1 at 850 hPa (Fig. 9f). A more distinct center associated with stronger vorticity is also visible at both 500 and 200 hPa (Figs. 9d,e). At 200 hPa, a larger area of negative vorticity (−1.5 × 10−4 s−1) associated with the outflow suggests that the CIMSS(h) analysis possesses a stronger vortex throughout the tropospheric levels (Fig. 9d). Differences between the CTL and CIMSS(h) analyses are also evident in the far environment of Sinlaku. They include the short-lived upper-level trough 15° to the north of Sinlaku (visible in both 200 and 500 hPa) and the low-level trough over the Southeast Asian peninsula. However, the differences are relatively small and their influences on the movement of Sinlaku are very subtle.
Given that the most significant wind difference exists in the vicinity of Sinlaku instead of the far surrounding environment, a local environmental “steering flow” is calculated by averaging the wind vectors over a 1000 × 1000 km2 domain centered on Sinlaku from 850 to 200 hPa for each analysis time (Fig. 10a). By 0000 UTC 10 September, the mid- to lower-tropospheric steering vectors in the CTL analyses exhibit northward to northwestward flows, whereas the steering vectors in the CIMSS(h) analyses are northeastward. This is consistent with the differences in Sinlaku’s motion between the CTL and CIMSS(h) ensemble members in Figs. 4d,e. The steering flow weakens by 1200 UTC 10 September in both runs as Sinlaku slows down.
At 1200 UTC 11 September, in addition to the upper-level outflow, the rapid-scan AMVs provide enhanced coverage at the mid- and lower levels. Additional AMVs are derived in the broad tropical area of clouds and easterly winds far to the southeast of Sinlaku, and further AMVs exist between 400 and 700 hPa on the southern side of the midlatitude jet between Korea and Japan, and in the Malaysian and Indonesian regions (not shown, but similar to Fig. 1f). At this time, Sinlaku is already a mature TC, and the analysis fields between the CIMSS(h) and CIMSS(h+RS) experiments show similar steering flows throughout the troposphere (Fig. 10a). We therefore only show the CIMSS(h) and CIMSS(h+RS) analyses at one level: 500 hPa (Figs. 10b,c). The shape and strength of the subtropical ridge to the immediate east of Sinlaku is slightly different between CIMSS(h) and CIMSS(h+RS); however, this difference apparently does not appear to contribute substantially to the steering flows and motion of Sinlaku at this time. In general, from 1200 UTC 11 September onward, the steering vectors are similar between CIMSS(h) and CIMSS(h+RS) analyses throughout the tropospheric layers, with the CIMSS(h+RS) vectors slightly farther eastward than the CIMSS(h) vectors above 500 hPa.
4. Influence on WRF ensemble forecasts
Given that the WRF/EnKF analyses provide improved representations of track and structure of Typhoon Sinlaku with the assimilation of AMVs, the forecasts initialized from these analyses are expected to produce superior results. Seven different initialization times are chosen as in Table 4 to study the influence of assimilating the CIMSS hourly and rapid-scan winds on 3-day forecasts. Group one comprises the first four initial times: 0000 UTC 9 September, 1200 UTC 9 September, 0000 UTC 10 September, and 1200 UTC 10 September (hereafter referred to as FT0900, FT0912, FT1000, and FT1012, respectively). This group is used to compare CTL forecasts versus CIMSS(h) forecasts. Group two comprises the last three initial times: 0000 UTC 11 September, 1200 UTC 11 September, and 0000 UTC 12 September (FT1100, FT1112, and FT1200, respectively). This group is used to compare forecasts from the CTL, CIMSS(h), and CIMSS(h+RS) experiments. For each of these seven initial times, ensemble forecasts are produced for each experiment by integrating the entire 84 ensemble analyses. Consistent with the configuration of the analysis experiment, a fixed outer grid of 27-km resolution and a moving inner grid of 9-km resolution is used for the ensemble forecasts.
For group one, the averaged ensemble mean track error of forecasts initialized with CIMSS(h) stays below 100 km in the first 42 h and increases up to about 200 km in 72 h (Fig. 11a). The averaged ensemble mean track error of forecasts initialized with CTL begins at about 140 km, and increases to approximately 400 km at 72 h. As an example, the initially large errors and incorrect westward movement in several of the CTL ensemble members for the FT1012 case leads to many of the CTL ensemble members making landfall in central and southern Taiwan, south of the actual track (Fig. 11c). On the other hand, the CIMSS(h) ensemble forecast members are generally distributed farther to the north, with the best track lying near the mean track. As for the MSLP forecasts, it is first worth noting that the initial MSLP in all CTL and CIMSS(h) is far from the true value, due in part to the coarse grid resolution of the analysis (27 km). As a result, over all four initial times in this group, the average MSLP errors of the ensemble members are generally decreasing with forecast time for both the CTL and CIMSS(h) experiments (Fig. 11b). The errors are clearly smaller for the CIMSS(h) ensemble forecast than for CTL, being approximately 5 hPa more accurate for 72-h forecasts. Additionally, the average 54–72-h forecast errors for CIMSS(h) are very low, less than 5 hPa. An example of one of the ensemble forecasts of MSLP is shown in Fig. 11d. In the CTL ensemble forecast, many members deepen and reach their peak intensity around 42–48 h after the initial time, before weakening. In the CIMSS(h) ensemble forecast, the ensemble members deepen less rapidly, due in part to their superior initial structure, and reach a near steady state around 36–42 h after the initial time. The MSLP in the CIMSS(h) ensemble members remains close to the best track MSLP, and the spread of values is considerably smaller than that in the CTL ensemble forecast.
The forecast track error differences between forecasts initialized with CTL and CIMSS(h) are a combination of initial position errors in earlier lead times and changes of the environmental steering flows at later times. At 0 h, the mean initial position differences between CTL and CIMSS(h) analyses are approximately 70 km (Fig. 4a,d,e) where the mean initial position of the CIMSS(h) analyses is shifted more eastward and closer to the best track (Figs. 11c). The changes in the environmental steering flows can be explained in part by examining the ensemble mean 500-hPa geopotential height associated with the FT1012 track forecast (Fig. 11c). Between 0 and 24 h, the CTL and CIMSS(h) forecasts of 500-hPa geopotential height associated with the environment of Sinlaku are similar (Figs. 12a,b, intermediate times not shown for brevity). By 24 h, the ensemble mean forecast of Sinlaku is a little to the northeast in CIMSS(h) compared with CTL. Concurrently, the subtropical ridge to the immediate east of Sinlaku (5880 m isoheight) exhibits a less westward intrusion, which provides less westward steering for Sinlaku (Fig. 12b) than is the case with CTL. From 48 to 72 h, this westward-intrusive subtropical ridge in CTL continues to provide a stronger westward steering flow to Sinlaku in many ensemble members, leading to those members incurring larger track errors by making landfall in central to southern Taiwan (Figs. 12c,d). To confirm this, we also compute local environmental steering vectors by averaging the wind vectors over a 1000 × 1000 km2 domain centered on the TC of each individual member from 850 to 200 hPa for each 0–72-h forecast initialized with the different analyses (similar to Fig. 10a, but for forecast times instead of analysis times). As is evident in Fig. 10a for the analysis times, the forecast steering vectors also exhibit a stronger westward component in CTL compared with CIMSS(h), for forecasts of 24 h and beyond (not shown here due to a close similarity to Fig. 10a).
For group two, the averaged ensemble mean track error of CIMSS(h) is again lower than that of CTL with an average improvement of about 50–125 km for 30–72-h forecasts (Fig. 13a). However, in the last 24 h, the averaged track error of CIMSS(h+RS) unexpectedly grows from 200 to 500 km while CTL has a maximum track error of 375 km and CIMSS(h) has a maximum track error of 225 km at 72 h. These results are illustrated with an example of ensemble track forecasts from FT1112 (Fig. 13b). While Sinlaku actually recurved toward the northeast after leaving northern Taiwan on 14 September, many of the ensemble members in CTL make landfall in central-to-north Taiwan and then erroneously make another landfall in southeastern China without recurvature. On the other hand, a few of the northern members in CIMSS(h) demonstrate their potential to recurve after 48 h without making landfall in either Taiwan or southeastern China beforehand.
Furthermore, all of the members in CIMSS(h+RS) initially travel close to the best track and then recurve toward the northeast without making landfall in Taiwan, followed by a subsequent rapid track toward the northeast, which amplifies the forecast errors at 72 h. The capture of the recurvature is an important issue in the track forecast of Sinlaku, especially when Sinlaku left Taiwan and was on its way toward Japan. Unfortunately, the recurvature in the CIMSS(h+RS) members was premature.
For the same case (FT1112), the differences between the environmental 500-hPa geopotential height fields in the CIMSS(h) and CIMSS(h+RS) analyses are very subtle from 1200 UTC 11 September onward [as mentioned in section 3b(4)]. The same is true for their respective forecasts up to 18 h, and the interpretation is unclear. Instead, the time–height diagram of local environmental steering vectors is presented in Fig. 14. The difference between the averaged local environmental steering vectors in the CIMSS(h) and CIMSSS(h+RS) ensemble forecasts becomes clear in the mid- and upper troposphere from 24 h onward. These vectors in the CIMSS(h) forecasts are northwestward to northward, whereas the steering vectors in the CIMSS(h) forecast are northeastward. This is consistent with the premature recurvature found in ensemble forecast tracks of CIMSS(h+RS). In other words, by 24 h, the influence of the rapid-scan AMV data on the midlatitude flow has begun to influence the track forecasts. At later times, an examination of the 500-hPa geopotential height fields shows that the track of Sinlaku begins to be influenced by a deeper midlatitude trough to its north in the CIMSS(h+RS) forecasts (not shown).
5. Summary and discussion
The influence of assimilating geostationary satellite-derived atmospheric motion vectors (AMVs) on WRF/EnKF ensemble analyses and forecasts of Typhoon Sinlaku (2008) was investigated. Three different groups of AMV data were assimilated in parallel experiments: CTL, the same dataset that is assimilated in the NCEP operational analysis; CIMSS(h), hourly datasets processed by CIMSS; and CIMSS(h+RS), the corresponding rapid-scan datasets. All of the AMVs assimilated in this study were superobbed over 90 km × 90 km × 25 hPa prisms. The CIMSS(h) dataset comprised an order of magnitude more AMVs than the CTL dataset, and when rapid-scan mode was activated the CIMSS(h+RS) dataset offered approximately 5 times more AMVs than CIMSS(h) below 400 hPa. Identical configurations of conventional observations (excluding radiances) were also assimilated in each of the three experiments, and no artificial vortex initialization or relocation schemes were used.
Compared against the CTL analyses, the CIMSS(h) analyses were mostly superior in depicting the initial storm positions, and the timing of the early intensification of Sinlaku. The more accurate initial intensification rate corresponded directly with a stronger three-dimensional primary and secondary circulation, and a tighter vortex with a sharper wind maximum during the period prior to Sinlaku reaching peak intensity (1200 UTC 10 September). Comparisons with a modified best track, independent QuikSCAT ocean surface winds and vertical wind profiles from dropwindsondes also suggested that the CIMSS(h) analyses were more consistent with observations than the CTL analyses during the same period. Furthermore, the larger quantity of AMV data in CIMSS(h) allowed the EnKF to produce more dynamically and thermodynamically consistent analysis increments through the depth of Sinlaku in its developing stage. Finally, during the same period, the influence of assimilating the hourly AMV data on the analyses was most pronounced in the TC and its near environment, and less so in the far environment. On the negative side, the peak intensity was not reached in any of the analysis ensemble members in either experiment, likely due to the 27-km resolution on the assimilation grid. The CIMSS(h) analysis ensemble also exhibited an unappealingly low spread in the intensity, which was clearly insufficient when compared with the error. After Sinlaku had reached peak intensity, the vertical profiles of tangential wind, radial wind, and temperature in the CIMSS(h) analyses remained more consistent with available dropwindsonde profiles than CTL, but other improvements were less clear.
Given that the MTSAT rapid-scan mode was activated after TC Sinlaku had already reached its peak intensity and was in a relatively steady state, the impact of adding and assimilating the rapid-scan AMV data on the TC structure analyses was not expected to be as substantial as compared with CIMSS(h). However, some minor improvements are noted. First, the CIMSS(h+RS) analyses produced average minimum MSLP and maximum surface wind values closer to the best track. Additionally, the vertical profiles of tangential wind, radial wind, and temperature were 12%, 7%, and 18%, respectively, superior to CIMSS(h) when evaluated against profiles from the three dropwindsondes. Finally, the assimilation of the rapid-scan AMV data were found to make some subtle changes to the TC structure in the analyses, and minor changes to the local and remote environmental fields.
The 0–72-h WRF ensemble forecasts initialized with CIMSS(h) analyses outperformed the forecasts initialized with CTL data in both track and intensity. The averaged ensemble mean forecast track error of CIMSS(h) initialized forecasts were found to be 100–150 km smaller than CTL initialized forecasts. During the 36-h period in which TC Sinlaku was intensifying, the MSLP errors for 30–72-h forecasts initialized during this period were on average reduced by at least 50%. During the 24-h period after TC Sinlaku had reached peak intensity and the rapid-scan AMV data were available, the CTL and many of the CIMSS(h) ensemble members tended to predict an erroneous landfall in mainland China. In contrast, the CIMSS(h+RS) ensemble members predicted recurvature away from mainland China, although unfortunately prematurely. The influence of assimilating the rapid-scan AMV data on the track forecasts was very subtle, with initially small differences in the CIMSS(h+RS) versus CIMSS(h) analyses leading to large differences in the forecasts 2–3 days later.
While some results from this case study are promising, a larger sample of TC cases will be necessary in order to evaluate the consistency of the forecast impacts provided by the enhanced AMV data. Furthermore, understanding where the TC analyses are benefiting most from the enhanced AMV information can lead to potential targeting scenarios, such as activating and directing rapid-scan operations. In this regard, studies are currently under way to investigate the optimal tropospheric layers and TC-centric environments in which AMV data are most beneficial to the high-resolution analyses and subsequent TC forecasts. Additionally, the optimal integration and assimilation of the AMVs must be considered with other full-resolution satellite data including temperature and moisture radiances/soundings, total precipitable water, and ocean surface winds.
In addition to better data handling, advanced model numerics and physics, and higher resolution in the assimilation domain should improve the representation of the tropical cyclone inner core. Improvements to increase the dispersiveness of the ensemble are essential, in order to provide more accurate error covariance information for data assimilation (and reduce filter divergence and reliance on inflation factors), and to yield a larger spread in the track and structure of the TC in the ensemble forecast members. Innovations in data assimilation to handle non-Gaussianity in tropical cyclone data assimilation are also expected to improve ensemble prediction. Such improvements in mesoscale modeling and data assimilation are expected to further exploit the utility of high-resolution satellite data including AMVs.
This work was supported through the National Oceanographic Partnership Program, representing a partnership between the Office of Naval Research (ONR) and the NOAA/Hurricane Forecast Improvement Program (HFIP), under ONR Marine Meteorology Program Award Number N00014-10-1-0123. The access to the NOAA T-Jet supercomputer provided by NOAA/HFIP was essential in enabling multiple data assimilation cycles and ensemble forecasts. The authors thank Tim Hoar, Glen Romine, Nancy Collins, and Chris Snyder of NCAR for providing great help with learning WRF-DART. Derrick Herndon of CIMSS provided a modified best track dataset for the study, and David Raymond and Carlos Lopez Carrillo provided wind fields from the airborne Electra Doppler Radar (ELDORA) for comparison studies. The authors also thank William Komaromi, Daniel Stern, and David Nolan of RSMAS/University of Miami for providing diagnostic code, and Ryan Torn of SUNY Albany for valuable discussions and encouragement.
In the modified version of the best track, several datasets not used in the original JTWC best track are included retrospectively. First, flight-level winds reduced to the surface together with Stepped Frequency Microwave Radiometer (SFMR) winds from reconnaissance aircraft were averaged. Additionally, QuikSCAT and Advanced Scatterometer (ASCAT) imagery and Cooperative Institute for Research in the Atmosphere (CIRA) Advanced Microwave Sounding Unit (AMSU) objective estimates were used to recalculate the wind radii (D. Herndon, CIMSS, 2010, personal communication).