• Aberson, S. D., , S. J. Majumdar, , C. A. Reynolds, , and B. J. Etherton, 2011: An Observing System Experiment for tropical cyclone targeting techniques using the Global Forecast System. Mon. Wea. Rev., 139, 895907, doi:10.1175/2010MWR3397.1.

    • Search Google Scholar
    • Export Citation
  • Anderson, J. L., 2003: A local least squares framework for ensemble filtering. Mon. Wea. Rev.,131, 634–642, doi:10.1175/1520-0493(2003)131<0634:ALLSFF>2.0.CO;2.

  • Anderson, J. L., , T. Hoar, , K. Raeder, , H. Liu, , N. Collins, , R. Torn, , and A. Avellano, 2009: The Data Assimilation Research Testbed: A community facility. Bull. Amer. Meteor. Soc., 90, 12831296, doi:10.1175/2009BAMS2618.1.

    • Search Google Scholar
    • Export Citation
  • Bedka, K. M., , C. S. Velden, , R. A. Petersen, , W. F. Feltz, , and J. R. Mecikalski, 2009: Comparisons of satellite-derived atmospheric motion vectors, rawinsondes, and NOAA wind profiler observations. J. Appl. Meteor. Climatol., 48, 15421561, doi:10.1175/2009JAMC1867.1.

    • Search Google Scholar
    • Export Citation
  • Berger, H., , R. Langland, , C. S. Velden, , C. A. Reynolds, , and P. M. Pauley, 2011: Impact of enhanced satellite-derived atmospheric motion vector observations on numerical tropical cyclone track forecasts in the western North Pacific during TPARC/TCS-08. J. Appl. Meteor. Climatol., 50, 23092318, doi:10.1175/JAMC-D-11-019.1.

    • Search Google Scholar
    • Export Citation
  • Brennan, M. J., , and S. J. Majumdar, 2011: An examination of model track forecast errors for Hurricane Ike (2008) in the Gulf of Mexico. Wea. Forecasting, 26, 848867, doi:10.1175/WAF-D-10-05053.1.

    • Search Google Scholar
    • Export Citation
  • Brown, D. P., , J. L. Beven, , J. L. Franklin, , and E. S. Blake, 2010: Atlantic Hurricane season of 2008. Mon. Wea. Rev., 138, 19752001, doi:10.1175/2009MWR3174.1.

    • Search Google Scholar
    • Export Citation
  • Goerss, J. S., 2009: Impact of satellite observations on the tropical cyclone track forecasts of the Navy Operational Global Atmospheric Prediction System. Mon. Wea. Rev., 137, 4150, doi:10.1175/2008MWR2601.1.

    • Search Google Scholar
    • Export Citation
  • Harnisch, F., , and M. Weissmann, 2010: Sensitivity of typhoon forecasts to different subsets of targeted dropsonde observations. Mon. Wea. Rev., 138, 26642680, doi:10.1175/2010MWR3309.1.

    • Search Google Scholar
    • Export Citation
  • Kieu, C. Q., , N. M. Truong, , H. T. Mai, , and T. Ngo-Duc, 2012: Sensitivity of the track and intensity forecasts of Typhoon Megi (2010) to satellite-derived atmospheric motion vectors with the ensemble Kalman filter. J. Atmos. Oceanic Technol., 29, 17941810, doi:10.1175/JTECH-D-12-00020.1.

    • Search Google Scholar
    • Export Citation
  • Komaromi, W. A., , S. J. Majumdar, , and E. D. Rappin, 2011: Diagnosing initial condition sensitivity of Typhoon Sinlaku (2008) and Hurricane Ike (2008). Mon. Wea. Rev., 139, 32243242, doi:10.1175/MWR-D-10-05018.1.

    • Search Google Scholar
    • Export Citation
  • Kurihara, Y., , M. A. Bender, , and R. J. Ross, 1993: An initialization scheme of hurricane models by vortex specification. Mon. Wea. Rev.,121, 2030–2045, doi:10.1175/1520-0493(1993)121<2030:AISOHM>2.0.CO;2.

  • Landsea, C. W., , and J. L. Franklin, 2013: Atlantic Hurricane database uncertainty and presentation of a new database format. Mon. Wea. Rev., 141, 35763592, doi:10.1175/MWR-D-12-00254.1.

    • Search Google Scholar
    • Export Citation
  • Langland, R. H., , C. Velden, , P. M. Pauley, , and H. Berger, 2009: Impact of satellite-derived rapid-scan wind observations on numerical model forecasts of Hurricane Katrina. Mon. Wea. Rev., 137, 16151622, doi:10.1175/2008MWR2627.1.

    • Search Google Scholar
    • Export Citation
  • Majumdar, S. J., and et al. , 2011: Targeted observations for improving numerical weather prediction: An overview. Tech. Rep. WMO/WWRP/THORPEX 15, 37 pp. [Available online at http://www.wmo.int/pages/prog/arep/wwrp/new/documents/THORPEX_No_15.pdf.]

  • Pu, Z., , X. Li, , C. S. Velden, , S. D. Aberson, , and W. T. Liu, 2008: The impact of aircraft dropsonde and satellite wind data on numerical simulations of two landfalling tropical storms during the Tropical Cloud Systems and Processes Experiment. Wea. Forecasting, 23, 6279, doi:10.1175/2007WAF2007006.1.

    • Search Google Scholar
    • Export Citation
  • Sears, J., , and C. S. Velden, 2012: Validation of satellite-derived atmospheric motion vectors and analyses around tropical disturbances. J. Appl. Meteor. Climatol., 51, 18231834, doi:10.1175/JAMC-D-12-024.1.

    • Search Google Scholar
    • Export Citation
  • Skamarock, W. C., and et al. , 2008: A description of the Advanced Research WRF version 3. NCAR Tech. Note NCAR/TN-475+STR, 113 pp. [Available online at http://www.mmm.ucar.edu/wrf/users/docs/arw_v3_bw.pdf.]

  • Torn, R. D., , and C. Snyder, 2012: Uncertainty of tropical cyclone best-track information. Wea. Forecasting, 27, 715729, doi:10.1175/WAF-D-11-00085.1.

    • Search Google Scholar
    • Export Citation
  • Velden, C. S., , and L. M. Leslie, 1991: The basic relationship between tropical cyclone intensity and the depth of the environmental steering layer in the Australian region. Wea. Forecasting, 6, 244253, doi:10.1175/1520-0434(1991)006<0244:TBRBTC>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Velden, C. S., , C. M. Hayden, , S. J. Nieman, , W. P. Menzel, , S. Wanzong, , and J. S. Goerss, 1997: Upper-tropospheric winds derived from geostationary satellite water vapor observations. Bull. Amer. Meteor. Soc.,78, 173–195, doi:10.1175/1520-0477(1997)078<0173:UTWDFG>2.0.CO;2.

  • Velden, C. S., and et al. , 2005: Recent innovations in deriving tropospheric winds from meteorological satellites. Bull. Amer. Meteor. Soc., 86, 205223, doi:10.1175/BAMS-86-2-205.

    • Search Google Scholar
    • Export Citation
  • Wu, T.-C., , H. Liu, , S. J. Majumdar, , C. S. Velden, , and J. L. Anderson, 2014: Influence of assimilating satellite-derived atmospheric motion vector observations on numerical analyses and forecasts of tropical cyclone track and intensity. Mon. Wea. Rev., 142, 4971, doi:10.1175/MWR-D-13-00023.1.

    • Search Google Scholar
    • Export Citation
  • Yamaguchi, M., , and S. J. Majumdar, 2010: Using TIGGE data to diagnose initial perturbations and their growth for tropical cyclone ensemble forecasts. Mon. Wea. Rev., 138, 36343655, doi:10.1175/2010MWR3176.1.

    • Search Google Scholar
    • Export Citation
  • Yamaguchi, M., , T. Iriguchi, , T. Nakazawa, , and C.-C. Wu, 2009: An Observing System Experiment for Typhoon Conson (2004) using a singular vector method and DOTSTAR data. Mon. Wea. Rev., 137, 28012816, doi:10.1175/2009MWR2683.1.

    • Search Google Scholar
    • Export Citation
  • View in gallery

    The spatial distribution of superobbed AMVs assimilated in the HALL experiment (a) at 0000 UTC 11 Sep 2008 near the most intense stage of Sinlaku and (b) at 1200 UTC 4 Sep 2008 close to when Ike reached its strongest stage. The AMVs are grouped into three layers by their assigned heights and are shown by the corresponding colors with numbers indicating the amount of AMVs in each grouped layer. The black circle is centered at the JMA best-track location of Sinlaku (yellow star) at the time of the analysis and it marks the 1000-km radius (approximately 10°). (c),(d) As in (a),(b), but showing the spatial distribution of superobbed AMVs assimilated in the RSALL experiment. Note that the aspect ratios of the Sinlaku and Ike maps are not the same so that the 1000-km black circles in these two maps are in different sizes. (e) The spatial distribution of assimilated radiosondes and not assimilated but verified dropwindsondes (see Table 3) for the duration of Sinlaku. (f) As in (e), but for the duration of Ike.

  • View in gallery

    (a) Ensemble mean TC positions in analyses from the six parallel H experiments for the duration of Sinlaku. Ensemble mean (b) TC position error (km), (c) MSLP (hPa), and (d) 10-m maximum wind speed (m s−1) from the six parallel H experiments. (e)–(g) As in (b)–(d), but from the six parallel RS experiments. (h)–(m) As in (b)–(g), but showing ensemble spread instead of ensemble mean. The JMA best track is plotted in gray with a square marker and the JTWC advisory track is plotted in gray with an open-circle marker.

  • View in gallery

    (a)–(g) As in Figs. 2a–g, but showing analyses from the six parallel H and RS experiments for the duration of Ike. The NHC best track is plotted in gray with a square marker and the NHC advisory track is plotted in gray with an open-circle marker. (h)–(j) Ensemble spread of the six parallel RS experiments is shown.

  • View in gallery

    The averaged vertical profiles of (a) root-mean-square error, (b) bias, and (c) spread of estimated tangential winds (m s−1) relative to the TC center in the six H analyses verified with the 17 dropwindsondes during Sinlaku. (d)–(f) As in (a)–(c), but for radial winds (m s−1).

  • View in gallery

    As in Figs. 4a–f, but for estimated tangential winds and radial winds (m s−1) relative to the TC center in the six RS analyses verified with the 12 dropwindsondes during Ike.

  • View in gallery

    For the six H analyses at 1200 UTC 11 Sep during Sinlaku: pressure–radius profiles of ensemble mean (contours) and spread (color fill) of azimuthally averaged (a) tangential wind (m s−1), (b) radius of maximum wind (blue line as mean and horizontal bar as spread in km) and maximum tangential wind (red line as mean and horizontal bar as spread; for better visualization, wind speed is multiplied by 10), and (c) radial wind (m s−1).

  • View in gallery

    As in Figs. 6a–c, but for RS analyses at 0000 UTC 7 Sep during Ike.

  • View in gallery

    Averaged ensemble mean of (a) track error (km), (b) error of minimum MSLP (hPa), and (c) error of 34-kt wind radii of 72-h ensemble forecasts initialized with the HALL, HInterior, HExterior, HnoLL, HnoML, and HnoUL analyses at 0000 UTC 9–12 Sep and 1200 UTC 11 Sep for the Sinlaku case. (d)–(f) As in (a)–(c), but for ensemble spread. Colored stars in (a)–(c) refer to forecast hours at which the differences of errors between the HALL and the corresponding HInterior, HExterior, HnoLL, HnoML, and HnoUL ensemble members are statistically significant at the 95% level.

  • View in gallery

    Ensemble mean track of ensemble forecasts initialized with (a) the HALL, RSALL, RSInterior, and RSExterior analyses; (b) with the HALL, RSALL, RSnoLL, RSnoML, RSnoUL analyses at 0000 UTC 11 Sep for Sinlaku case. (c)–(f) As in (a),(b), but for ensemble forecasts initialized at (c),(d) 1200 UTC 11 Sep and at (e),(f) 0000 UTC 12 Sep.

  • View in gallery

    Ensemble mean 500-hPa geopotential height (m) of 24-h forecasts initialized with the RSALL (blue) and (a) RSInterior, (b) RSExterior, (c) RSnoLL, (d) RSnoML, and (e) RSnoUL (red) analyses at 1200 UTC 11 Sep for the Sinlaku case. The interval is 30 m. The color fill denotes the 500-hPa geopotential height difference between the blue and red contours (red minus blue). (f)–(j) As in (a)–(e), but showing 48-h forecasts.

  • View in gallery

    As in Fig. 8, but for averaged ensemble mean of 72-h ensemble forecasts initialized with (a)–(c) the HALL, HInterior, and HExteiror, HnoLL, HnoML analyses and (d)–(f) the RSALL, RSInterior, and RSExteiror, RSnoLL, RSnoML analyses at 0000 UTC 5–8 Sep for TC Ike. Colored stars in (a)–(f) refer to forecast hours at which the differences of errors between the HALL (RSALL) and the corresponding HInterior, HExterior, HnoLL, HnoML, and HnoUL (RSInterior, RSExterior, RSnoLL, RSnoML, and RSnoUL) ensemble members are statistically significant at the 95% level. (g)–(l) As in (a)–(f), but for ensemble spread.

  • View in gallery

    (a) Ensemble mean track forecasts initialized with the RS analyses at 0000 UTC 6 Sep for the Ike case. The orange bar in (a) denotes the 18-h track forecasts. Ensemble mean 500-hPa geopotential height (m) of 18-h forecasts initialized with the RSALL (blue) and (b) RSInterior, (c) RSExterior, (d) RSnoLL, (e) RSnoML, and (f) RSnoUL (red) analyses at 0000 UTC 6 Sep. The interval is 30 m. The color fill denotes the 500-hPa geopotential height difference between the blue and red contours (red minus blue).

  • View in gallery

    Ensemble mean track of forecasts initialized with the HALL (× symbol) and RSALL (open circle) analyses at 0000 UTC 9–10 Sep for the Ike case and corresponding 120-h GFS track forecast initialized at 0000 UTC 9 Sep (GFS09: blue square) and 0000 UTC 10 Sep (GFS10: blue triangle). The black dashed box highlights the track forecast differences between the GFS and the HALL and RSALL cases near Ike’s final landfall in southeastern Texas.

  • View in gallery

    (a) As in Fig. 13, but for the GFS analysis. The black dashed box in (a) outlines the forecast tracks of Sinlaku from the two different forecasts (red and green; initialized 1 day apart) during their overlapping period (10–14 Sep; total 72 h). (b) Time series of mean steering vectors of the GFS analyses and the WRF forecasts initialized with the HALL analyses at 0000 UTC 9 Sep (red) and at 0000 UTC 10 Sep (green) within the overlapped 72 h. (c) As in (b), but showing results from forecasts initialized with the RSALL analyses. The x-axis label in blue corresponds to mean steering vectors calculated from the GFS analysis, and the other two x-axis labels are colored for corresponding WRF forecasts.

  • View in gallery

    Ensemble mean 500-hPa geopotential (m) height of (a) 78-h forecasts initialized with the HALL (red) analysis at 0000 UTC 9 Sep and the GFS analysis (blue) valid at 0600 UTC 12 Sep. The interval is 30 m. (b) As in (a), but showing 54-h forecasts initialized with the HALL (red) analysis at 0000 UTC 10 Sep and the GFS analysis valid at the same time. The color fill denotes the 500-hPa geopotential height difference between the blue and red contours (red minus blue). The cyan boxes highlight the location of the northwestern edge of the ridge to the northeast of Ike.

All Time Past Year Past 30 Days
Abstract Views 0 0 0
Full Text Views 55 55 7
PDF Downloads 25 25 1

Understanding the Influence of Assimilating Subsets of Enhanced Atmospheric Motion Vectors on Numerical Analyses and Forecasts of Tropical Cyclone Track and Intensity with an Ensemble Kalman Filter

View More View Less
  • 1 Rosenstiel School of Marine and Atmospheric Science, University of Miami, Miami, Florida
  • | 2 Cooperative Institute for Meteorological Satellite Studies, University of Wisconsin–Madison, Madison, Wisconsin
  • | 3 Rosenstiel School of Marine and Atmospheric Science, University of Miami, Miami, Florida
  • | 4 National Center for Atmospheric Research, Boulder, Colorado
© Get Permissions
Full access

Abstract

Recent studies have shown that assimilating enhanced satellite-derived atmospheric motion vectors (AMVs) has improved mesoscale forecast of tropical cyclones (TC) track and intensity. The authors conduct data-denial experiments to understand where the TC analyses and forecasts benefit the most from the enhanced AMV information using an ensemble Kalman filter and the Weather Research and Forecasting Model. The Cooperative Institute for Meteorological Satellite Studies at the University of Wisconsin provides enhanced AMV datasets with higher density and temporal resolution using shorter-interval image triplets for the duration of Typhoon Sinlaku and Hurricane Ike (both 2008). These AMV datasets are then spatially and vertically subsetted to create six parallel cycled assimilation-forecast experiments for each TC: all AMVs; AMVs withheld between 100 and 350 hPa (upper layer), between 350 and 700 hPa (middle layer), and between 700 and 999 hPa (lower layer); and only AMVs within (interior) and outside (exterior) 1000-km radius of the TC center. All AMV subsets are found to be useful in some capacity. The interior and upper-layer AMVs are particularly crucial for improving initial TC position, intensity, and the three-dimensional wind structure along with their forecasts. Compared with denying interior or exterior AMVs, withholding AMVs in different tropospheric layers had less impact on TC intensity and size forecasts. The ensemble forecast is less certain (larger spread) in providing accurate TC track, intensity, and size when upper-layer AMVs or interior AMVs are withheld. This information could be useful to potential targeting scenarios, such as activating and focusing satellite rapid-scan operations, and decisions regarding observing system assessments and deployments.

Corresponding author address: Ting-Chi Wu, Cooperative Institute for Research in the Atmosphere, Colorado State University, 1375 Campus Delivery, Fort Collins, CO 80523-1375. E-mail: ting-chi.wu@colostate.edu

This article is included in the Sixth WMO Data Assimilation Symposium Special Collection.

Abstract

Recent studies have shown that assimilating enhanced satellite-derived atmospheric motion vectors (AMVs) has improved mesoscale forecast of tropical cyclones (TC) track and intensity. The authors conduct data-denial experiments to understand where the TC analyses and forecasts benefit the most from the enhanced AMV information using an ensemble Kalman filter and the Weather Research and Forecasting Model. The Cooperative Institute for Meteorological Satellite Studies at the University of Wisconsin provides enhanced AMV datasets with higher density and temporal resolution using shorter-interval image triplets for the duration of Typhoon Sinlaku and Hurricane Ike (both 2008). These AMV datasets are then spatially and vertically subsetted to create six parallel cycled assimilation-forecast experiments for each TC: all AMVs; AMVs withheld between 100 and 350 hPa (upper layer), between 350 and 700 hPa (middle layer), and between 700 and 999 hPa (lower layer); and only AMVs within (interior) and outside (exterior) 1000-km radius of the TC center. All AMV subsets are found to be useful in some capacity. The interior and upper-layer AMVs are particularly crucial for improving initial TC position, intensity, and the three-dimensional wind structure along with their forecasts. Compared with denying interior or exterior AMVs, withholding AMVs in different tropospheric layers had less impact on TC intensity and size forecasts. The ensemble forecast is less certain (larger spread) in providing accurate TC track, intensity, and size when upper-layer AMVs or interior AMVs are withheld. This information could be useful to potential targeting scenarios, such as activating and focusing satellite rapid-scan operations, and decisions regarding observing system assessments and deployments.

Corresponding author address: Ting-Chi Wu, Cooperative Institute for Research in the Atmosphere, Colorado State University, 1375 Campus Delivery, Fort Collins, CO 80523-1375. E-mail: ting-chi.wu@colostate.edu

This article is included in the Sixth WMO Data Assimilation Symposium Special Collection.

1. Introduction

Atmospheric motion vectors (AMVs) are proxies for the local horizontal wind, and are derived from sequential multispectral satellite images by tracking the motion of targets that include cirrus clouds, gradients in water vapor, and lower-tropospheric cumulus clouds (Velden et al. 1997). AMV data are assimilated routinely into operational global numerical weather prediction (NWP) systems, and have been found to improve forecasts of tropical cyclone (TC) tracks (e.g., Goerss 2009; Langland et al. 2009).

NWP systems have recently advanced dramatically with regards to model grid resolution and convective physics. As a result, much attention has turned to using state-of-the-art mesoscale models and data assimilation schemes to improve TC forecasting. Such improvements include more frequent updates with the use of higher density and temporal resolution observation datasets and the choice of thinning and superobbing,1 and error assignments of these observations that better suit the TC scales being analyzed (Bedka et al. 2009). Improvement of TC track and intensity forecasts has occurred in recent observing system experiments (OSEs) that assimilated high-density AMV data into a mesoscale model (e.g., Pu et al. 2008; Kieu et al. 2012; Wu et al. 2014). Nevertheless, these studies are case based and investigations on more TC cases are necessary to generalize their conclusions.

Advanced data processing methods provide enhanced coverage and improved quality over routinely available AMVs (Velden et al. 2005) and also more frequent AMV datasets if the rapid-scan imaging mode is activated. Using such data, Berger et al. (2011) demonstrated that global model forecasts of TC tracks can be improved through a more accurate representation of the environmental flow. The Cooperative Institute for Meteorological Satellite Studies (CIMSS) at the University of Wisconsin has prepared the enhanced AMV datasets during the life cycles of Typhoon Sinlaku (2008) and Hurricane Ike (2008). Assimilating these enhanced AMV data into a hurricane-scale model, Wu et al. (2014) found that initial analyses of TC vortex location, intensity, and structure are improved along with their forecasts.

Given the findings from the past decade of forecast impact experiments for selected sets of satellite and aircraft observations [summarized in a WMO report by Majumdar et al. (2011)], one may expect that the assimilation of AMVs in selected tropospheric layers or locations relative to the TC center would be particularly important for improving analyses and forecasts of TC track, structure or intensity. For example, Yamaguchi et al. (2009) and Aberson et al. (2011) showed that in selected case studies, the assimilation of dropwindsonde observations in targeted areas around the TC can result in improved track forecasts. Similarly, Harnisch and Weissmann (2010) found that dropwindsondes deployed in the vicinity of the TC are more useful for near-term track forecasting than those in the farther synoptic environment. Kieu et al. (2012) investigated the relative roles of AMVs located within the 300–100- versus 800–300-hPa layers, and found that despite the relatively small number of AMVs between 800 and 300 hPa, the track forecast is more sensitive to the AMVs in this layer. However, these studies put more emphasis on the TC track, and the impacts on TC structure and intensity is discussed only briefly. Given the greater availability of satellite observations, especially those observations with higher frequency and denser population (e.g., rapid-scan AMVs), a thorough study on the assimilation of high-resolution satellite observations is necessary. We suggest a better understanding of the impact of assimilating AMV dataset comes through identifying regions and locations of assimilated AMV data that improves forecast of TC track, intensity, and structure the most. Findings from this study can lead to informed decisions regarding observing system assessments and deployments. For example, this information could impact potential targeting scenarios, such as activating and focusing satellite rapid-scan operations, or directing aircraft assets.

In current operational NWP centers, AMV data from various geostationary and polar-orbiting platforms are routinely assimilated mostly via variational data assimilation systems in addition to conventional and additional satellite observations including radiance (for more information, please visit http://nwpsaf.eu/monitoring/amv/nwp.html). Because of relatively coarse grid spacing (analysis grid mostly >30 km), AMVs are thinned to ~(200 × 200) km2 × 100-hPa resolution before assimilation. With the focus on a high-density AMV dataset, this study follows the work of Wu et al. (2014) to investigate the relative role of spatially subsetted AMVs under a mesoscale framework with an ensemble data assimilation approach.

We quantify the impact of assimilating subsets of enhanced AMV data on mesoscale ensemble forecasts of TCs, via six parallel OSEs that assimilate the following: all AMVs; only AMVs near the TC (within 1000-km radius of the TC); only AMVs in the domain exterior to this region; and all AMVs but with each of three selected tropospheric layers removed. The two TC cases—Typhoon Sinlaku (2008) and Hurricane Ike (2008)—are described, followed by details of the assimilation experiments in section 2. The influence of assimilating the different subsets of AMV data on mesoscale ensemble analyses and forecasts is presented in sections 3 and 4, respectively, followed by a summary in section 5.

2. Tropical cyclone cases and methods

a. Typhoon Sinlaku (2008)

After its formation east of the Philippines on 0000 UTC 8 September 2008, the storm was upgraded to Tropical Storm Sinlaku by the Japan Meteorological Agency (JMA) at 1800 UTC on the same day. Sinlaku then underwent rapid intensification, reaching typhoon status at 1200 UTC 9 September, and intensifying further to 937 hPa by 1800 UTC 10 September with two concentric eyewalls. After this period, the inner eyewall dissipated, and the minimum sea level pressure did not fall back below 950 hPa over the subsequent two days. During the 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). In this study, we focus our attention on the period of 8–13 September 2008. Many publications have focused on Sinlaku, including a special collection titled “Targeted Observations, Data Assimilation, and Tropical Cyclone Predictability” in Mon. Wea. Rev.2

b. Hurricane Ike (2008)

Ike was one of the six consecutive cyclones that hit the United States in 2008 (Brown et al. 2010). Following its formation early on 1 September 2008, the tropical depression strengthened into Tropical Storm Ike later that day, west of the Cape Verde Islands. As it moved west-northwestward under the influence of a strong Atlantic subtropical ridge to its north, Ike intensified steadily over the next two days followed by rapid intensification to a category-4 hurricane by 0600 UTC 4 September. Over the following two days, Ike began to weaken and turned farther west. Around 0000 UTC 6 September, Ike then turned unexpectedly toward the west-southwest. Ike made two landfalls over Cuba, first on 8 September as a category-4 hurricane and then on 9 September as a category-1 hurricane before entering the Gulf of Mexico (Brown et al. 2010). Ike’s intensity continued to fluctuate as it entered the Gulf of Mexico on 10 September with an expanded wind field. Its track forecast was a particular challenge in the several days leading up to its eventual landfall in Galveston, Texas, on 13 September. Studies have been published about the sensitivity of Ike’s track forecast to the environmental flow (Brennan and Majumdar 2011; Komaromi et al. 2011). Our study considers the period during 5–10 September.

c. Modeling and data assimilation systems

The Advanced Research Weather Research and Forecasting Model (WRF-ARW) version 3.1.1 is employed in this paper (Skamarock et al. 2008). The type of ensemble Kalman filtering (EnKF) data assimilation scheme used here is the ensemble adjustment Kalman filter (EAKF; Anderson 2003), implemented 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 remote environment, the EAKF assimilation is performed on a large domain with 27-km grid spacing covering the western North Pacific (0°–50°N, 100°–150°E) for the Typhoon Sinlaku case, and the North Atlantic and continental United States (0°–45°N, 30°–110°W) for the Hurricane Ike case, with 36 vertical levels from the surface to 50 hPa. A nested moving domain with 9-km grid spacing is activated in the forecast cycling when the TC is present. However, due to computational limitations, no direct data assimilation is performed on the 9-km nested moving domain. Further details of the WRF-DART configuration and ensemble initial and boundary conditions are provided in Wu et al. (2014). Model error is neglected in this study.

d. Data: Enhanced and rapid-scan atmospheric motion vectors

High-resolution (space and time) AMV datasets were postprocessed and prepared at hourly intervals by CIMSS for each TC case using routinely available multispectral image triplets. For the Sinlaku case in 2008, MTSAT-1 scanned images every 30 min; for Ike in 2008, GOES-East scanned images every 15 min. These AMV datasets (derived from IR, VIS, and WV imagery) are denoted as H for hourly in this paper. [It should be noted that operational AMV datasets derived by national satellite data centers and made available over the GTS (not used or evaluated in this study) were only available at 6-h intervals in 2008.] The CIMSS automated AMV derivation algorithm is similar to that employed operationally at NOAA/NESDIS for GOES full-disk AMV dataset production. However the CIMSS processing method provides more enhanced coverage of AMVs in and around TCs when they are present (Sears and Velden 2012).

AMVs can be further enhanced in both quantity and quality if the time separation between images is reduced, which allows more coherent tracking of clouds (Velden et al. (2005). When a satellite rapid-scan mode is activated (denoted as RS in this paper), the more frequent sequential images yield a higher volume of AMVs than the hourly AMVs. The rapid-scan mode of the geostationary satellite MTSAT-2, which was operated experimentally by JMA in 2008, was activated shortly after 1200 UTC 10 September 2008. By this time, Sinlaku had already reached category-2 intensity, and was two days away from its landfall in northern Taiwan. During this period, RS AMV datasets (from IR and VIS only) were derived by CIMSS at hourly intervals using successive 15-min image triplets. In the North Atlantic region, RS from the GOES geostationary satellites is routinely used for operational tasking. In 2008, GOES-12 had an activated RS mode throughout the lifetime of Ike. The GOES RS allows 7.5-min image scanning, and these were employed by CIMSS to produce RS AMV datasets at hourly intervals during Ike.

All of the AMVs assimilated in this study are first passed through quality control steps, then superobbed by averaging them in 90 × 90 km2 × 25-hPa prisms as in Wu et al. (2014). An example of the horizontal and vertical distribution of superobbed AMVs in Sinlaku and its vicinity on 0000 UTC 11 September 2008 is illustrated in Figs. 1a,c. The activation of the RS mode yields many more low-level vectors, which is attributable to the enhanced tracking ability of low-level cumuliform type cloud tracers (especially from the visible channel) when higher-frequency imagery is available. An example of the horizontal and vertical distribution of superobbed AMVs for Ike on 1200 UTC 4 September 2008 is presented in Figs. 1b–d. As in the case of Sinlaku, more mid- to lower-layer AMVs are retrieved with the activation of RS mode.

Fig. 1.
Fig. 1.

The spatial distribution of superobbed AMVs assimilated in the HALL experiment (a) at 0000 UTC 11 Sep 2008 near the most intense stage of Sinlaku and (b) at 1200 UTC 4 Sep 2008 close to when Ike reached its strongest stage. The AMVs are grouped into three layers by their assigned heights and are shown by the corresponding colors with numbers indicating the amount of AMVs in each grouped layer. The black circle is centered at the JMA best-track location of Sinlaku (yellow star) at the time of the analysis and it marks the 1000-km radius (approximately 10°). (c),(d) As in (a),(b), but showing the spatial distribution of superobbed AMVs assimilated in the RSALL experiment. Note that the aspect ratios of the Sinlaku and Ike maps are not the same so that the 1000-km black circles in these two maps are in different sizes. (e) The spatial distribution of assimilated radiosondes and not assimilated but verified dropwindsondes (see Table 3) for the duration of Sinlaku. (f) As in (e), but for the duration of Ike.

Citation: Monthly Weather Review 143, 7; 10.1175/MWR-D-14-00220.1

e. Experimental design

For Sinlaku, six parallel WRF–EnKF experiments are initialized on 0000 UTC 1 September 2008, a week prior to its formation. The analyses with assimilated AMV subsets are produced using a 3-h cycle. The corresponding six parallel WRF–EnKF experiments for Hurricane Ike are initialized on 0000 UTC 25 August 2008, again a week prior to its formation. Conventional observational data assimilated include winds, temperature, and specific humidity from radiosondes situated at least 200 km away from the TC center, aircraft flight-level winds and temperature, surface altimeter data, and respective Joint Typhoon Warning Center (JTWC) or National Hurricane Center (NHC) advisory TC positions (latitude and longitude of TC at the analysis time). In addition, the uncertainty of the JTWC advisory TC position estimates are assumed to be dependent upon the TC maximum wind speed, and are assigned the following values: 90 km for maximum sustained wind <34 kt (1 kt = 0.5144 m s−1); 40 km for >85 kt; and 60 km for intermediate wind values (Torn and Snyder 2012). With more reconnaissance aircraft observations collected in the Atlantic basin, the uncertainty of the NHC advisory TC position estimates are assumed to be 25% lower than in the western Pacific basin (empirically determined). The available dropwindsonde observations in each case are reserved for verification and are therefore excluded from direct assimilation.

In addition to the conventional observations, the first experiment (denoted ALL) assimilates all of the available enhanced AMVs produced by CIMSS 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 central time of the window (no time adjustments). The analyses are used to initialize an ensemble of 3-h forecasts for the next analysis time. The second (third) experiment, denoted interior (exterior), is the same as ALL except that it assimilates the enhanced AMVs only within (outside) 1000-km radius from the TC center. The size of 1000 km is chosen as an approximation to the upper limit of the 1200-km filter radius that the GFDL vortex initialization method proposed by Kurihara and Bender (1993) uses to separate the TC flow from the environmental flow. The fourth to sixth experiments are the same as ALL except that they eliminate AMVs between 150–350, 350–700, and 700–999 hPa, respectively. These experiments are named noUL (no upper-layer AMVs), noML (no middle-layer AMVs), and noLL (no lower-layer AMVs), respectively. The six parallel experiments were conducted for hourly AMVs for both tropical cyclone cases. They were repeated for rapid-scan AMVs where available for Sinlaku (after its rapid intensification) and for the entire period of Ike. It is important to note that the H and RS AMV datasets are treated separately in this study, even though for much of the time in these two TC cases they were simultaneously available. This results in a total of 24 (6 denial experiments, 2 AMV datasets, and 2 TC cases) parallel experiments; a summary is given in Table 1.

Table 1.

WRF–EnKF experiments designed to understand the contribution of assimilating subsets of CIMSS hourly and rapid-scan AMVs during Typhoon Sinlaku (2008) and Hurricane Ike (2008). The experiments cover the lifetime of Sinlaku from 0000 UTC 8 Sep 2008 to 1200 UTC 13 Sep 2008 and the lifetime of Ike from 0000 UTC 1 Sep 2008 to 0000 UTC 10 Sep 2008. To differentiate hourly and rapid-scan experiments, H is added to the name of the experiments which assimilate CIMSS hourly AMVs, and RS is added to the rapid-scan experiments.

Table 1.

As shown in Fig. 1, AMVs are grouped into upper-, middle-, and lower-tropospheric layers and highlighted with colors corresponding to the experimental design. A black circle centered on the location of the TC (yellow star) indicates the separation between interior and exterior AMVs. From Table 2, it is first worth noting that the interior AMVs represent approximately 10% of the total number of AMVs in both TC cases. Second, as seen in Figs. 1a,b, most of the hourly AMVs are at the upper layers (73% upper, 11% middle, and 16% lower for Sinlaku and 56% upper, 10% middle, and 33% lower for Ike). The greater percentage with Sinlaku relative to Ike is due to the image sampling. The 15-min interval available from GOES for Ike is much better for tracking shorter-lived, low-level cumulus than the 30-min interval from MTSAT during Sinlaku, especially for the VIS AMVs. Third, the activation of rapid-scan mode increases the number of AMVs for all subsections as expected (e.g., Figs. 1c,d), but the percentages between the three layer subsets become more comparable, more so for Sinlaku than for Ike (31% upper, 40% middle, and 29% lower for Sinlaku and 48% upper, 24% middle, and 28% lower for Ike). The shorter image intervals allows for better tracking of low- and midlevel clouds, with the jump from 30- to 15-min scans having a bigger impact with Sinlaku than the 15–7.5-min images for the Ike rapid scan. Finally, the distribution of AMVs by layer in the interior section is comparable to the distribution of AMVs by layer over the whole domain.

Table 2.

Time-averaged percentages of different subsets of assimilated superobbed AMVs in the WRF–EnKF system for Typhoon Sinlaku (2008)/Hurricane Ike (2008). Rapid-scan AMVs are shown in boldface. Upper-layer, midlayer, lower-layer, and interior follow the definitions in Table 1. The first column is the average number of superobbed AMVs in ALL and interior experiments.

Table 2.

The distribution of radiosondes (assimilated) and dropwindsondes (not assimilated, and only within 100–1200 km from TC centers; see Table 3) is illustrated for Sinlaku and Ike in Figs. 1e and 1f, respectively.

Table 3.

Dropwindsondes selected for analysis verification (allow ±30-min difference from analysis time) during Typhoon Sinlaku (2008) and Hurricane Ike (2008). Dropwindsondes that are less than 100 km from each TC center are not included.

Table 3.

3. Analysis results from the AMV assimilation experiments

a. TC position, minimum sea level pressure, and 10-m maximum sustained wind speed

The ensemble mean analyses are first verified on the 27-km domain against the best-track position, minimum sea level pressure (MSLP), and 10-m maximum sustained wind speed (MSW) estimates from the JMA and NHC for Sinlaku and Ike, respectively. For Sinlaku, only the analyses from 0000 UTC 9 September 2008 onward are verified, given the ambiguity in locating the precise center prior to Sinlaku reaching tropical storm status. For Ike, only the analyses from 0000 UTC 5 September 2008 onward are verified because of the proximity of Ike to the eastern boundary of the domain at earlier times.3

1) Sinlaku

The ensemble mean TC positions, position errors, MSLP, and MSW of the ensemble analyses from the six H experiments during Sinlaku are presented in Figs. 2a–d. During the early stage of Sinlaku (0000 UTC 9 September), all the analyses are shifted westward of the JMA best-track positions (Fig. 2a) with an average error of around 140 km (Fig. 2b). It is noteworthy that the JTWC advisory TC position data were rejected in the first few assimilation cycles beginning from 0000 UTC 8 September. As mentioned before, the representation of Sinlaku in the 3-h WRF forecast is rather weak prior to 0000 UTC 9 September so that the prior observation operator was not able to locate the center. Without prior estimates of storm centers, no observation increment of storm center is available for the analysis update. Although AMVs were assimilated during the period when the advisory TC position data were rejected, improvement in reducing the TC position errors is limited because the majority of the near-storm AMVs are in upper layers associated with TC outflow. The position errors in five of the experiments then drop to 30 km after 0000 UTC 10 September and remain below 50 km at later times (the advisory TC position data are no longer rejected). The only exception is the HExterior experiment, where the position errors reduce much slower and stay above 50 km, until 0000 UTC 11 September. From 0000 UTC 11 September onward, the position errors from the six H analyses stay below 50 km and HnoLL has the smallest and very steady errors of 20 km. Since the H LL AMVs compose 22% of interior AMVs, which is a larger percentage than that of the interior midlevel AMVs (see Table 2), it would be intuitive to suspect that the position errors may increase more by withholding the H LL AMVs than the H ML AMVs. However, the opposite is true and a possible reason is that the interior lower-level AMVs are asymmetrically distributed4 and could act to pull the TC off center.

Fig. 2.
Fig. 2.

(a) Ensemble mean TC positions in analyses from the six parallel H experiments for the duration of Sinlaku. Ensemble mean (b) TC position error (km), (c) MSLP (hPa), and (d) 10-m maximum wind speed (m s−1) from the six parallel H experiments. (e)–(g) As in (b)–(d), but from the six parallel RS experiments. (h)–(m) As in (b)–(g), but showing ensemble spread instead of ensemble mean. The JMA best track is plotted in gray with a square marker and the JTWC advisory track is plotted in gray with an open-circle marker.

Citation: Monthly Weather Review 143, 7; 10.1175/MWR-D-14-00220.1

After the rapid-scan AMVs became available from 1800 UTC 10 September, the EnKF analyses of HALL were utilized to initiate the first cycles of the six parallel RS experiments at this time. The mean position errors, MSLP, and MSW from the six RS analyses at later times are also illustrated in Figs. 2e–g. Unlike in the H group, the differences between RSnoUL, RSnoML, and RSnoLL are very small, which is likely due to their more comparable AMV proportions and also because they all start from an analysis that has already been spun up with the hourly AMV data.

The ensemble mean MSLP and MSW are very similar among the six parallel experiments prior to 1200 UTC 10 September (Figs. 2c,d). However, although most of the experiments still show a steady intensification of Sinlaku over the following two days, the intensification is delayed about 1 day for HExterior and HnoUL analyses. Similarly, when HALL, HInterior, HnoML, and HnoLL undergo the weakening stage, there is also about 12-h delay in the case of HExterior and HnoUL analyses. These results indicate that the assimilation of AMVs within the general realm of the TC vortex can have an important influence on both the position and intensity analyses of Sinlaku, and that the upper-layer AMVs in particular are important to the intensity analyses, at least at earlier times of Sinlaku. From 0000 UTC 11 September, the mean MSLP and MSW of the six RS analyses are shown in Figs. 2f,g.

It is found that the ensemble spread (cf. ensemble mean error) is almost always underdispersive (see Figs. 2h–m), which is common for ensemble data assimilation. Despite the underdispersiveness, examining the spread of the analysis ensembles in addition to the ensemble mean error provides an appropriate representation of uncertainty. In Figs. 2h–j, not only do the HExterior and HnoUL ensemble mean possess less accurate analysis TC position and MSLP and MSW, but their ensemble spread is larger, particularly for HExterior. This result supports the findings above that without the vortex-domain AMVs, and to a lesser extent the upper-layer AMVs, the analyses are not only less accurate but also more uncertain in estimating the TC position and MSLP. The same situation is found in Figs. 2k–m where slightly larger ensemble spreads are also associated with RSExterior and RSnoUL.

2) Ike

Since the rapid-scan mode of GOES-East was activated during the full lifetime of Ike, the two groups (H and RS) of six parallel experiments are available for the entire duration of the assimilation experiments. In general, the analysis differences between the H and RS results are mostly not substantial for Ike, which is not that surprising given their similar AMV quantities (Table 2). Note that both HALL and RSALL are neither the best nor the worst matches to the best position in their groups. Unlike in Sinlaku, the exterior experiment at times yields the smallest ensemble mean position errors (see Figs. 3b,e). Curiously, interior analyses have larger position errors initially, and exterior errors grow after 0000 UTC 8 September, approximately when Ike made its first landfall on Cuba. This is evident in both H and RS groups. In both the H and RS experiments, the ensemble mean position errors of ALL, noLL, noML, and noUL analyses have similar evolutions, and it is not clear which analysis overall is most compromised when a particular layer of AMVs is removed.

Fig. 3.
Fig. 3.

(a)–(g) As in Figs. 2a–g, but showing analyses from the six parallel H and RS experiments for the duration of Ike. The NHC best track is plotted in gray with a square marker and the NHC advisory track is plotted in gray with an open-circle marker. (h)–(j) Ensemble spread of the six parallel RS experiments is shown.

Citation: Monthly Weather Review 143, 7; 10.1175/MWR-D-14-00220.1

In Figs. 3c,d and 3f,g, none of the H or RS ensemble mean MSLP and MSW estimates are close to the best-track MSLP, likely due to the limited model resolution. Despite this 20-hPa discrepancy, the ensemble mean MSLP and MSW estimates of RSInterior are closest to the best-track values prior to the landfall on Cuba on 8 September although the larger position errors are noted earlier (Figs. 3e–g). It is also noteworthy that the RSnoLL shows a ~1-day delay in the 6–7 September intensification while the other analyses show gradual intensification. This would seem to possibly be an artifact of this ensemble showing a 10-hPa weakening between 0000 and 1200 UTC 6 September, as many other ensembles similarly reflect a steady-state (or nearly so) intensity in the 1200 UTC 6 September–1200 UTC 7 September period. This delay of intensification is not seen in the corresponding HnoLL analyses, and could be due to the increased coverage of LL vectors produced by the rapid scans. After 8 September, as Ike makes landfall over Cuba, the six analyses in the H and RS groups have relatively small differences in the MSLP. Because of the similarity, only the ensemble spread of the RS group is shown (Figs. 3h–j).

Overall, for Ike, the effects of removing subsets of the AMV data from the assimilation experiments are subtler than for Sinlaku.

3) Discussion

In general, the AMVs provide an improvement to the analyses of Sinlaku, but the results are more mixed for Ike. Part of this may be due to the fact that Ike had very good storm information provided by the NHC, which was successfully used in the EnKF assimilation process. On the other hand, the analyses TC positions of Sinlaku are initially over 100 km off from the JMA best track for all the H AMV experiments, while the JTWC advisory positions are fairly close to the best track. Certainly the model resolution of 27 km accounts for some of the precision variability (especially noted in the MSLP and MSW analyses), but this also raises the question of near-core assimilation when you have 1) a lot of data that might not be horizontally or vertically symmetric in their distribution about the TC center (e.g., AMVs), and 2) TC advisory information of varying quality.

It is also important to mention that different storms raise different questions and react differently to data assimilation even if they are under the same experimental configuration. No direct comparison between the two TCs is intended be made. Sinlaku and Ike are not really comparable even if the similar points in their life cycle are considered, given the different AMV coverage, the different environmental influences, and the different sizes and structures of the TCs.

b. Analysis verification against dropwindsonde data

Vertical profiles of the ensemble mean storm-centered wind are verified against independent dropwindsonde observations within a 100–1200-km radius from the TC center. Prior to computing the mean, the wind profile in each ensemble member is first computed relative to the TC center in that member. Dropwindsondes were released in the region of Sinlaku during the Tropical Cyclone Structure (TCS-08)/THORPEX Pacific Asian Regional Field Campaign (T-PARC), and by NOAA in and near Ike. Only those dropwindsondes within 30 min of the analysis times at which the full ensemble fields were able to be stored (0000 and 1200 UTC) are considered, and are listed in Table 3.

For Sinlaku, the RMSE of tangential wind from 1000 to 200 hPa lies between 2 and 5 m s−1 for most of the experiments, except for HExterior and HnoUL where the RMSE lies between 3.5 and 8 m s−1 (Fig. 4a). There is a corresponding negative (weak) bias in the lower troposphere in all the data denial experiments, though this bias is nullified in HALL (Fig. 4b). The RMSE of radial wind between 1000 and 200 hPa is within the range of 3.5–8 m s−1 over all the six experiments. The larger RMSE is evident in HExterior (throughout the depth) and HnoUL (above 600 hPa) (Fig. 4d). A positive bias of radial wind up to 2 m s−1 is evident below 700 hPa, especially for HExterior, HnoLL, and HnoUL (Fig. 4e). In the upper troposphere, the negative bias in HnoUL contributes to large errors. Note that the dropwindsondes that are used to evaluate the ensemble analyses were deployed during Sinlaku’s peak intensity when HExterior and HnoUL analyses are overall much weaker than the other four analyses. Therefore, it makes sense that the ensemble mean biases of tangential and radial winds of HExterior and HnoUL are larger (weaker bias) for almost all levels.

Fig. 4.
Fig. 4.

The averaged vertical profiles of (a) root-mean-square error, (b) bias, and (c) spread of estimated tangential winds (m s−1) relative to the TC center in the six H analyses verified with the 17 dropwindsondes during Sinlaku. (d)–(f) As in (a)–(c), but for radial winds (m s−1).

Citation: Monthly Weather Review 143, 7; 10.1175/MWR-D-14-00220.1

The ensemble spread of tangential and radial wind between 1000 and 200 hPa is tightly clustered near 2 m s−1 with little vertical variation in four of the experiments (Figs. 4c,f). The two exceptions—HnoUL and HExterior—have larger spread, consistent with the vertical profiles of RMSE in Figs. 4a,d, and also the larger position, MSLP, and MSW spread in Figs. 2h–j.

For both the H and RS analyses of Ike, the vertical profiles of RMSE, bias, and spread of the tangential and radial winds are more tightly clustered in the six experiments than that of Sinlaku. While the profiles in H and RS are similar, the differences between the six experiments are more pronounced in the RS group, so only these are shown. The averaged profile of RMSE of tangential wind ranges approximately between 2 and 4 m s−1 (Fig. 5a), considerably smaller than of Sinlaku (Fig. 4a). The RMSE profiles of both tangential and radial winds in RSnoLL and RSExterior are found to be slightly larger than the other four analyses mostly in the lower troposphere below 700 hPa. In contrast to the Sinlaku case, the average tangential wind bias for Ike is small and positive in the lower troposphere, and consistently negative (too weak) and of magnitude 1.5 m s−1 above 700 hPa (Fig. 5b). The radial wind exhibits less than a 2 m s−1 negative bias below 500 hPa and a positive bias above 400 hPa (Fig. 5e), suggesting slightly stronger inflow in the lower layers and a much stronger outflow aloft in the ensemble analyses. One potential reason for the tighter clustering in Ike is these selected dropwindsondes are deployed on 7–8 September when the intensity estimations of Ike in the six analyses are much more similar than at earlier times.

Fig. 5.
Fig. 5.

As in Figs. 4a–f, but for estimated tangential winds and radial winds (m s−1) relative to the TC center in the six RS analyses verified with the 12 dropwindsondes during Ike.

Citation: Monthly Weather Review 143, 7; 10.1175/MWR-D-14-00220.1

The ensemble spread of both the tangential and radial wind profiles in the six parallel experiments is also smaller and more tightly clustered for Ike than for Sinlaku, not exceeding 2.5 m s−1 through the column. The spread is largest in RSExterior and RSnoLL, and is especially consistent with the largest RMSE in these two experiments below 800 hPa.

Overall, the error characteristics of the tangential and radial wind profiles show some differences between the Sinlaku and Ike cases, though with some common results.

c. Azimuthally averaged structure

A qualitative analysis of the ensemble mean and spread of the primary and secondary circulation of the TC via azimuthally averaged profiles is conducted for the six H and RS parallel analyses for Sinlaku and Ike, respectively, at selected times. In addition to the tangential, radial, and vertical winds, the radius of maximum wind (RMW) and maximum tangential wind are included. For Sinlaku, the six H analyses on 1200 UTC 11 September are presented in Fig. 6, at the time when Sinlaku had reached its maximum intensity and was in a steady state. The azimuthally averaged tangential wind speed in the ensemble mean of HExterior and HnoUL is weaker than that of HALL (Fig. 6a). This is also evident in Fig. 6b where the maximum tangential wind does not exceed 35 m s−1 in either HExterior or HnoUL, whereas the maximum value exceeds 45 m s−1 in the other four analyses. The ensemble spread in both HExterior and HnoUL is also relatively large. In most analyses, the ensemble mean RMW is a little less than 100 km through the column, and its slope is not steep (80–100 km from 1000 to 200 hPa) in most analyses. The spread of RMW is tight in many cases, except for HExterior and the upper levels of HnoUL. We, therefore, deduce that the assimilation of interior AMVs is responsible for reducing the uncertainty in estimating the RMW and also the maximum tangential wind. In other words, the assimilation of interior AMVs adjusts all ensemble members toward a particular portion of the spectrum of ensemble solution, thereby reducing ensemble spread.

Fig. 6.
Fig. 6.

For the six H analyses at 1200 UTC 11 Sep during Sinlaku: pressure–radius profiles of ensemble mean (contours) and spread (color fill) of azimuthally averaged (a) tangential wind (m s−1), (b) radius of maximum wind (blue line as mean and horizontal bar as spread in km) and maximum tangential wind (red line as mean and horizontal bar as spread; for better visualization, wind speed is multiplied by 10), and (c) radial wind (m s−1).

Citation: Monthly Weather Review 143, 7; 10.1175/MWR-D-14-00220.1

Comparing the vertical profiles of HALL with HnoLL, HnoML, and HnoUL, the assimilation of upper-layer AMVs has a significant influence on the profile of the maximum tangential wind. An examination of the radial wind structure reveals that HExterior and HnoUL are again showing much weaker lower-level inward flows and upper-level outward flows (Fig. 6c). Additionally, a region of low-level outflow centered on an 80-km radius and 700 hPa is evident in HALL, HInterior, HnoLL, and HnoML, with the strongest speed of 6 m s−1 in HALL and lower mean values in the other three analyses. This low-level outflow first appeared around 0000 UTC 10 September and intensified in the following days. It is less perceptible (at least in the azimuthally averaged perspective) in the ensemble mean of HExterior and HnoUL while a local patch of large ensemble spread near the same location is persistent in both cases throughout its presence. Given that HExterior and HnoUL exhibit a relatively weak lower-level inflow and upper-level outflow in the radial wind profile, it is not surprising that the vertical wind is also weak (figure not shown). Vertical profiles as in Fig. 6 were produced for the parallel six RS analyses at the same time, but are not shown here because of the similarity. The ensemble mean of tangential and radial wind profiles in the six RS analyses are much more comparable than the six H analyses, although RSExterior still exhibits slightly weaker profiles. The ensemble spread of tangential wind profiles in RSExterior and RSnoUL are the largest, but are slightly smaller than the corresponding ensemble spread in HExterior and HnoUL.

In Fig. 7, the ensemble mean and spread of tangential wind, RMW, maximum tangential wind, and radial wind are much more alike between the six RS analyses (figure for the six H analyses not shown because of the similarity) for Ike. In addition to RSExterior, RSnoLL also possess slightly weaker mean primary and secondary circulation (Figs. 7a,c). The averaged maximum tangential wind is ~35 m s−1 and takes place around the 50-km radius for the RSALL, RSInterior, RSnoML, and RSnoUL (Fig. 7b). It is noteworthy that the RSnoUL has the smallest RMW spread, but the spread of maximum tangential wind and radial wind are both slightly larger than the other five experiments.

Fig. 7.
Fig. 7.

As in Figs. 6a–c, but for RS analyses at 0000 UTC 7 Sep during Ike.

Citation: Monthly Weather Review 143, 7; 10.1175/MWR-D-14-00220.1

These results corroborate the previous verification against dropwindsonde data in concluding that the interior and upper-layer AMVs are especially important for modifying the primary and secondary circulation.

4. WRF ensemble forecasts

To assess model TC track and intensity forecast impact from the various AMV modifications to the analyses, 84-member ensemble forecasts are produced for each experiment by integrating the ensemble analyses with the same configuration as in the analysis experiment, together with a vortex-following inner grid of 9-km resolution.

For Sinlaku, two periods are focused upon (Table 4): 1) 0000 UTC 9–12 September and 1200 UTC 11 September, comparing the 3-day ensemble forecasts initialized with the six H analyses; 2) only the last 3 times in 1), 0000 UTC 11 September, 1200 UTC 11 September, and 0000 UTC 12 September, since RS AMVs are only available after 1800 UTC 12 September. This period specifically examines the recurvature in the track forecasts, which was briefly discussed in Wu et al. (2014) and further investigated here via comparing the 3-day ensemble forecasts initialized with the six RS denial experiments.

Table 4.

The 72-h ensemble forecasts initialized with WRF–EnKF analyses at different initial times during Typhoon Sinlaku (2008) and Hurricane Ike (2008). The ensemble forecasts initialized at 0000 UTC 9–10 Sep for Hurricane Ike (2008) are extended to 120 h to cover the landfall in southern Texas.

Table 4.

For Ike, six different initialization times are selected to focus on two periods (Table 4): 1) 0000 UTC 5–8 September, comparing the 3-day ensemble forecasts initialized with the six H and the six RS analyses respectively; 2) 0000 UTC 9–10 September, which specifically examines 5-day ensemble forecasts for the HALL and RSALL experiments during a period of operational model forecast uncertainty, and up to Ike’s landfall in Texas.

Ensemble forecasts of TC Sinlaku are verified against the modified version of JTWC best-track data5 because several datasets were not included in the original JTWC best track and are included retrospectively. Ensemble forecasts of TC Ike are verified against the NHC best-track data. The track, intensity, and 34-kt wind radii presented below are computed on the coarser 27-km nonmoving grid with the use of GFDL vortex tracker.6 The presented errors of wind radii are averaged over the four quadrants. A summary of the uncertainty of the NHC best-track data can be found in Landsea and Franklin (2013).

a. 3-day forecasts of Sinlaku: 9–12 September

The ensemble mean track error of HALL, averaged over the five forecasts initialized at 0000 UTC 9–12 September and 1200 UTC 11 September, is notably lower than that from the denial experiments (Fig. 8a). HInterior and HnoLL begin with a relatively small track error, but grow substantially after 18 h and reach more than 350 km in 3 days. Combining this with earlier findings on the impacts of the various subdatasets on the analyses of TC structure, this result suggests that the interior AMVs are especially important for initializing the TC structure and its initial motion.

Fig. 8.
Fig. 8.

Averaged ensemble mean of (a) track error (km), (b) error of minimum MSLP (hPa), and (c) error of 34-kt wind radii of 72-h ensemble forecasts initialized with the HALL, HInterior, HExterior, HnoLL, HnoML, and HnoUL analyses at 0000 UTC 9–12 Sep and 1200 UTC 11 Sep for the Sinlaku case. (d)–(f) As in (a)–(c), but for ensemble spread. Colored stars in (a)–(c) refer to forecast hours at which the differences of errors between the HALL and the corresponding HInterior, HExterior, HnoLL, HnoML, and HnoUL ensemble members are statistically significant at the 95% level.

Citation: Monthly Weather Review 143, 7; 10.1175/MWR-D-14-00220.1

HALL also has the smallest ensemble mean MSLP error in the first 18 h (Fig. 8b), but errors of HALL and corresponding denial experiments soon become similar as they reduce (forecast is weaker than observed) with time to 48 h. Thereafter errors increase (forecast is stronger than observed), except for HInterior. Interestingly, the HALL mean error of the 34-kt wind radii, averaged over all quadrants, is not the lowest to begin with. Many of the forecasts have relative small initial errors that increase rapidly with time (Fig. 8c). All forecasts have a larger storm than the observed one.

The statistical significance of these results is examined. A two-sample two-tailed t test at the 95% level is used to test the hypothesis that the forecast error differences between the HALL ensemble and the corresponding denial experiment ensembles are statistically significant, for all forecast lead times up to 72 h. The significant difference is marked by stars in Figs. 8a–c. Almost all five corresponding experiments have significantly different track errors from the HALL, except for the HNoLL and HNoML at the first 24 h. HInterior and HExterior have significantly larger track errors than that of the HALL from 36 h onward, while HExterior and HNoUL begin to show slightly larger errors than that of the HALL 12–18 h later. The large track errors of HnoLL is mainly from the forecasts initialized at 1200 UTC 11 September and 0000 UTC 12 September (not shown) in which HnoLL missed Sinlaku’s landfall in northern Taiwan while the others all captured the landfall.

The intensity error differences between the HALL and the corresponding denial experiments are mostly significant except for during the 30–36-h lead time when they become similar (Fig. 8b). Both the HInterior and HExterior have become very similar to the HALL during the 30–48-h lead time. From 48–54 h onward, they become significantly different from the HALL again, and HInterior has the smallest intensity errors than all others. The different errors of 34-kt wind radii are mostly significant between the HALL and the corresponding denial experiments, except for the HNoUL after the 42-h lead time and the HNoML for most of the lead times. The HNoLL has the smallest radii errors for the first 36 h, but in the last 36 h the errors go up to 165 km at 72 h while the majority of the five experiments stay around 130 km.

The corresponding ensemble error spread is illustrated in Figs. 8d–f. While HALL generally exhibits smaller spread, HExterior and HnoLL are often associated with larger ensemble spread, particularly HExterior. It is worth mentioning that HInterior shows more rapidly increasing ensemble spread in track errors and 34-kt wind radii errors than the others.

b. 3-day forecast of Sinlaku with a focus on RS dataset: 11–12 September

This set of forecasts focuses on Sinlaku’s track forecasts with respect to Taiwan landfall, and the subsequent recurvature with a focus on the impact of RS datasets (although HALL is included as a reference dataset). For the initial time of 0000 UTC 11 September (Figs. 9a,b), all forecasts are premature in turning Sinlaku toward Taiwan. The RSExterior provides the worst track forecast. The ensemble mean track forecasts in RSALL, RSnoLL, and RSnoUL then show signs of early recurvature back to the north beginning from ~42 h. In contrast, HALL, RSInterior, and RSExterior move Sinlaku more westward toward southeastern China after making landfall over northern Taiwan (Figs. 9a,b).

Fig. 9.
Fig. 9.

Ensemble mean track of ensemble forecasts initialized with (a) the HALL, RSALL, RSInterior, and RSExterior analyses; (b) with the HALL, RSALL, RSnoLL, RSnoML, RSnoUL analyses at 0000 UTC 11 Sep for Sinlaku case. (c)–(f) As in (a),(b), but for ensemble forecasts initialized at (c),(d) 1200 UTC 11 Sep and at (e),(f) 0000 UTC 12 Sep.

Citation: Monthly Weather Review 143, 7; 10.1175/MWR-D-14-00220.1

Considering the forecasts initialized on 1200 UTC 11 September, RSALL and RSExterior really pick up the impending recurvature to the northeast, albeit about 36 h too soon. HALL and RSInterior continue to move Sinlaku westward toward southeastern China (Fig. 9c). All of the other RS forecasts show recurvature to varying degrees, (Fig. 9d). For forecasts initialized 12 h later on 0000 UTC 12 September, HALL and RSInterior now show a late recurvature, while the other ensemble mean forecasts suggest recurvature (Figs. 9e,f), but far too early with minimal to no impact to Taiwan.

These results demonstrate the high sensitivity in the track forecasts during this period. To better understand the different forecast track behaviors, the steering flow for the forecast initialized at 1200 UTC 11 September are now visualized via difference fields of 500-hPa geopotential height (Fig. 10). First, both RSInterior and RSnoML weaken the trough to the north of Sinlaku compared with RSALL, leading to less of an northeastward steering flow (Figs. 10a,d). Also, the location of the 5880-m contour to the east of Sinlaku is found to be farther westward in RSInterior, RSExterior, RSnoLL, and RSnoML (more evident at 48-h forecast), compared with RSALL, which leads to a more westward track in the data-denial experiments. On the other hand, there is minimal difference in the vicinity of Sinlaku between the 500-hPa geopotential heights in RSnoUL and RSALL (Fig. 10e). At 48h (Figs. 10f–j), the difference field in 24 h is amplified and is consistent with the motion of Sinlaku.

Fig. 10.
Fig. 10.

Ensemble mean 500-hPa geopotential height (m) of 24-h forecasts initialized with the RSALL (blue) and (a) RSInterior, (b) RSExterior, (c) RSnoLL, (d) RSnoML, and (e) RSnoUL (red) analyses at 1200 UTC 11 Sep for the Sinlaku case. The interval is 30 m. The color fill denotes the 500-hPa geopotential height difference between the blue and red contours (red minus blue). (f)–(j) As in (a)–(e), but showing 48-h forecasts.

Citation: Monthly Weather Review 143, 7; 10.1175/MWR-D-14-00220.1

In this case of high track sensitivity to environmental steering, very different track forecasts result from the different AMV datasets denied in the assimilation experiments.

c. 3-day forecasts of Ike: 5–8 September

The average forecast errors for Ike during 5–8 September (four forecasts) exhibit different characteristics to those for Sinlaku. In contrast to Fig. 8, neither HALL nor RSALL consistently possess the lowest average errors in track or MSLP as compared with their data-denial counterparts (Fig. 11). A similar t test is also performed here for the forecast of Ike, except for a smaller sample size because there are only four forecast cases while there are five forecast cases for Sinlaku in Fig. 8. In general, the track error differences between HALL and the others are significant, especially from 30 h onward. The HnoML is very similar to the HALL in the first 24 h and the last 12 h. On the other hand, the differences between the RSALL and the corresponding denial experiments are more substantial and significant (Figs. 11a,d). In contrast to the ensemble forecasts for Sinlaku, the growth rate of track error in the interior tends to be close to that of ALL, and slightly larger (both H and RS). Curiously, noML performs the best for track forecasts, especially in the RS set for most of the times. While AMVs tend not to be overly abundant in this layer of the troposphere, this result suggests some sensitivity. In fact, it was found in Sears and Velden (2012) that AMV quality generally deteriorates in the middle levels due to uncertain vector height assignments. Since Ike was a well-developed vortex during this period, the steering flow was deep with a strong influence from the midlevels (Velden and Leslie 1991). Taking out the midlevel AMVs in this case actually improves the track forecast. It was also the case in Fig. 11d where RSnoML shows a more accurate track than RSALL, RSnoLL, and RSnoUL.

Fig. 11.
Fig. 11.

As in Fig. 8, but for averaged ensemble mean of 72-h ensemble forecasts initialized with (a)–(c) the HALL, HInterior, and HExteiror, HnoLL, HnoML analyses and (d)–(f) the RSALL, RSInterior, and RSExteiror, RSnoLL, RSnoML analyses at 0000 UTC 5–8 Sep for TC Ike. Colored stars in (a)–(f) refer to forecast hours at which the differences of errors between the HALL (RSALL) and the corresponding HInterior, HExterior, HnoLL, HnoML, and HnoUL (RSInterior, RSExterior, RSnoLL, RSnoML, and RSnoUL) ensemble members are statistically significant at the 95% level. (g)–(l) As in (a)–(f), but for ensemble spread.

Citation: Monthly Weather Review 143, 7; 10.1175/MWR-D-14-00220.1

Considering the MSLP mean forecast errors, HInterior and HnoLL outperforms HALL for the entire forecast period (Fig. 11b). The MSLP error difference between RSALL and the denial experiments is slightly smaller and less significant than of the differences between the HALL and the corresponding denial experiments, although RSInterior exhibits much smaller MSLP errors for the entire forecast period (Fig. 11e). Figures 11c,f both suggest that the interior generally has significantly larger 34-kt wind radii errors in both the H and RS experiments. Comparing to the interior and exterior AMV data-denial experiments (both H and RS in Figs. 11c,f), less changes to the storm size errors are found between forecasts initialized with different tropospheric layers removed, especially in the RS group. One exception is that the 34-kt wind radii errors in HnoLL are slightly larger, even though this HnoLL’s MSLP errors yielded the best results. In this case, taking out the hourly low-level AMVs created a stronger and bigger TC. It is worth mentioning here that the average NHC best-track uncertainty estimates for 34-kt wind radii is about 55 km, even with assistance from satellite and aircraft observations (Landsea and Franklin 2013).

Despite the underdispersiveness, the corresponding ensemble spread in Figs. 11g–l show that excluding the interior AMVs increases the ensemble spread (from both HALL and RSALL) in all three metrics—track, intensity, and size. The ensemble spread also increases when upper-layer AMVs is withheld, although it is a slightly less increase than excluding exterior AMVs.

To understand how the differences between the steering flows in the data-denial experiments change the forecast tracks, we further examine the ensemble forecasts initialized on 0000 UTC 6 September, which includes Ike’s unexpected turn to the southwest and first landfall in Cuba on 7 September. The H and RS track forecasts are relatively similar, so only the RS group is presented. In Fig. 12a, RSInterior (RSExterior) has the worst (best) track forecast with RSALL, RSnoLL, RSnoML, and RSnoUL in between. Figures 12b–f show the 18-h ensemble forecasts of 500-hPa geopotential height differences between RSALL and the other five experiments (around the time the track forecasts begin to diverge). RSInterior shows a substantially weaker westward extension of the ridge just north of Cuba (Fig. 12b), which would promote a more northerly track as seen in Fig. 12a. On the other hand, the small height differences just to the north of Ike in RSExterior and RSnoML (Figs. 12c,e) suggest a little stronger ridge and would promote the more southerly tracks.

Fig. 12.
Fig. 12.

(a) Ensemble mean track forecasts initialized with the RS analyses at 0000 UTC 6 Sep for the Ike case. The orange bar in (a) denotes the 18-h track forecasts. Ensemble mean 500-hPa geopotential height (m) of 18-h forecasts initialized with the RSALL (blue) and (b) RSInterior, (c) RSExterior, (d) RSnoLL, (e) RSnoML, and (f) RSnoUL (red) analyses at 0000 UTC 6 Sep. The interval is 30 m. The color fill denotes the 500-hPa geopotential height difference between the blue and red contours (red minus blue).

Citation: Monthly Weather Review 143, 7; 10.1175/MWR-D-14-00220.1

d. 5-day forecast of Ike: 9–10 September

The forecasts of Ike during this period presented a substantial challenge to NHC, given the run-to-run inconsistency in the operational models’ predictions of the landfall location in the United States (Brennan and Majumdar 2011). Here, we compare the 5-day HALL and RSALL track predictions over the same period.

Most of the operational numerical forecasts initialized at 0000 UTC 9 September took Ike south of the best track (GFS track forecasts are blue lines in Fig. 13), while the corresponding forecasts initialized at 0000 UTC 10 September brought Ike back toward its actual landfall in southeastern Texas (Brennan and Majumdar 2011). In contrast, we find that the ensemble means of our HALL and RSALL ensemble forecasts initialized on 0000 UTC 9 September were more accurate, while the corresponding forecasts initialized on 0000 UTC 10 September took Ike farther south toward the Texas–Mexico border (Fig. 13). To further understand this, a local environmental steering flow at pressure levels between 850 and 200 hPa is calculated for GFS analyses, HALL, and RSALL at each forecast time by averaging the ensemble mean wind vectors over a 1000 × 1000 km2 domain centered on the TC of each individual member.

Fig. 13.
Fig. 13.

Ensemble mean track of forecasts initialized with the HALL (× symbol) and RSALL (open circle) analyses at 0000 UTC 9–10 Sep for the Ike case and corresponding 120-h GFS track forecast initialized at 0000 UTC 9 Sep (GFS09: blue square) and 0000 UTC 10 Sep (GFS10: blue triangle). The black dashed box highlights the track forecast differences between the GFS and the HALL and RSALL cases near Ike’s final landfall in southeastern Texas.

Citation: Monthly Weather Review 143, 7; 10.1175/MWR-D-14-00220.1

For the 0000 UTC 9 September ensemble forecasts, the rapid-scan AMVs (RSALL) produce an excellent track forecast. HALL begins to depart from RSALL after 2 days, with stronger southwestward-pointed steering vectors versus GFS analyses above 500 hPa (red vectors inside the black box of Figs. 14b,c). The corresponding steering vectors a day later are now both indicating this phenomenon.

Fig. 14.
Fig. 14.

(a) As in Fig. 13, but for the GFS analysis. The black dashed box in (a) outlines the forecast tracks of Sinlaku from the two different forecasts (red and green; initialized 1 day apart) during their overlapping period (10–14 Sep; total 72 h). (b) Time series of mean steering vectors of the GFS analyses and the WRF forecasts initialized with the HALL analyses at 0000 UTC 9 Sep (red) and at 0000 UTC 10 Sep (green) within the overlapped 72 h. (c) As in (b), but showing results from forecasts initialized with the RSALL analyses. The x-axis label in blue corresponds to mean steering vectors calculated from the GFS analysis, and the other two x-axis labels are colored for corresponding WRF forecasts.

Citation: Monthly Weather Review 143, 7; 10.1175/MWR-D-14-00220.1

To further assess the behavior of these forecast tracks, the 500-hPa geopotential heights of the HALL and RSALL forecasts valid at 0600 UTC 12 September are compared against the corresponding NCEP GFS analysis. For the forecasts initialized on 0000 UTC 9 September, the 5940-m contours in HALL (red contours in Fig. 15a) shows the northwestern edge of the Atlantic subtropical ridge to the northeast of Ike is well aligned with its counterpart in the corresponding–validating NCEP GFS analysis (blue contour). On the other hand, in the forecasts initialized on 0000 UTC 10 September and valid at the same time (Fig. 15b), the northwestern edge of the ridge in HALL is shifted farther to the southwest, which acts to push Ike’s forecast track farther westward. Essentially, the same situation is found between RSALL forecasts initialized on 10 September and valid at the same time (not shown).

Fig. 15.
Fig. 15.

Ensemble mean 500-hPa geopotential (m) height of (a) 78-h forecasts initialized with the HALL (red) analysis at 0000 UTC 9 Sep and the GFS analysis (blue) valid at 0600 UTC 12 Sep. The interval is 30 m. (b) As in (a), but showing 54-h forecasts initialized with the HALL (red) analysis at 0000 UTC 10 Sep and the GFS analysis valid at the same time. The color fill denotes the 500-hPa geopotential height difference between the blue and red contours (red minus blue). The cyan boxes highlight the location of the northwestern edge of the ridge to the northeast of Ike.

Citation: Monthly Weather Review 143, 7; 10.1175/MWR-D-14-00220.1

5. Summary and discussion

The influence of assimilating specific subsets of enhanced AMV datasets on WRF/EnKF ensemble analyses and forecasts is investigated for Typhoon Sinlaku and Hurricane Ike (both 2008). Six different subsets are assimilated 3-hourly in parallel experiments (see Table 1) for hourly AMVs (normal geostationary images used, with datasets produced at hourly intervals) and repeated for rapid-scan AMVs (hourly datasets, but using special rapid-scan images to produce the wind vectors) for Sinlaku and Ike, resulting in a total of 24 parallel experiments.

About 10% of the total AMVs per dataset are located within the interior for both the hourly and rapid-scan datasets. For the hourly datasets, more than 50% of the AMVs are located in the upper layer and only about 10% are in the middle layer. The percentage of lower-layer AMVs is between 15% and 40%. For the rapid-scan datasets, a larger fraction of mid- to lower-layer AMVs are retrieved due to the greater availability of cloud tracers in the more frequent satellite images. Identical configurations of conventional observations are assimilated in each of the 24 experiments, without the use of vortex initialization or relocation schemes, but with the assimilation of TC advisory position data.

An examination of the analysis fields against best-track data suggests that mean errors and spread in the TC position, MSLP, and MSW are larger when interior AMVs and to a lesser extent when lower-layer AMVs are withheld. The verification of storm-centered vertical wind profiles against independent dropwindsonde data reveals some similarities and differences between the Sinlaku and Ike cases. In particular, the assimilation of AMVs in the interior and upper layers are particularly important in reducing analysis biases and ensemble spread. These main findings are corroborated by a qualitative investigation of the axisymmetric primary and secondary circulations, in which the upper-layer and interior AMVs are found to be especially important in appropriately modifying the TC circulations.

In summary, our findings show that the direct assimilation of AMVs provides an improvement to the analyses of Sinlaku, but the results are more mixed for Ike. This may be in part due to the fact that Ike was better covered by conventional and aircraft observations, and there was very good storm position information provided by the NHC, which was successfully used in the EnKF assimilation process.

In addition to the impact of the various AMV subsets on TC analyses, the performance of 3-day ensemble forecasts initialized with the modified analyses is also evaluated. For Sinlaku (9–12 September), the exterior AMVs are found to be important for accurately predicting the TC track via improved representation of the environmental flow and the interior AMVs influence the initial TC structure. A larger range of track errors in the ensemble is evident when interior AMVs are not assimilated. There is no clear signal in the forecasts of MSLP when the AMVs in different layers are removed. It is likely due to the coarse resolution of the 27-km grid that all forecasts have a larger storm than the observed one. For the subsequent stage of Sinlaku, between 11 and 12 September, the focus turned to track recurvature, which was more carefully investigated in Wu et al. (2014). Our study shows that the AMVs from rapid scans influence the steering flow correctly leading to forecast recurvature, albeit somewhat prematurely and, therefore, misses the landfall in northern Taiwan. This case presents a very sensitive situation, as interior AMVs, and AMVs at different layers produce substantially different forecast tracks.

The conclusions for the ensemble forecasts of Ike are not so clear, and they are not entirely consistent with those of Sinlaku. For Ike, more similarities are evident between the forecasts from which hourly versus rapid-scan winds are assimilated. This is perhaps expected, since unlike the Sinlaku case, the number of AMVs is not too different between hourly and rapid-scan datasets. Curiously, noML performs the best for track forecasts, especially in the RS set of runs. While AMVs tend not to be overly abundant in this layer of the troposphere, AMV quality generally deteriorates in the midlevels due to uncertain vector height assignments. Since Ike was a well-developed vortex during this period, the steering flow was deep with a strong influence from the midlevels. Taking out the “possibly suspect” midlevel AMVs in this case where the background steering flow is already good actually improves the subsequent track forecasts. Within five days of Ike’s landfall over northeastern Texas, the two WRF ensemble track forecasts examined with AMVs exhibit some variability, as did the operational numerical track forecasts issued at that time.

Commonalities in our findings between Sinlaku and Ike include the following: 1) interior AMVs are important for analyses and forecasts of MSLP; 2) excluding upper-layer AMVs generally results in larger track errors and ensemble spread; 3) compared with denying interior or exterior AMVs, withholding AMVs in different tropospheric layers has less impact on the forecasts of 34 kt wind radii; 4) without the assimilation of interior or upper-layer AMVs, the largest ensemble spreads are found in forecast track, MSLP, and 34-kt wind radii; and 5) withholding the midlayer AMVs can improve the track forecasts. For TCs at various stages, the AMV distributions can act in different ways to represent their contributions to the TC structure and near environments. These findings, therefore, provide suggestions for the data assimilation community not only on what specific attributes of AMVs may be more beneficial but also in the circumstances in which they were found to not be beneficial at improving the analyses of TCs and their near environments. These suggestions could be taken into account in the form of data selection and observation weighting.

With only two tropical cyclones considered in this study, it is not suitable to generalize the conclusions. However, some insights are found, that could influence future scenarios that involve the targeted acquisition and assimilation of high-density AMV observations in TC events. While beyond the scope of the present work, it might be insightful to consider data-denial experiments in which the lower-, mid-, and upper-troposphere subsets of the AMVs are applied to interior-only and exterior-only subsets.

AMVs can exhibit significant systematic errors and those are the main reason for spatial and temporal thinning–superobbing. While an increase in the resolution of the model and AMV datasets is essential to improve the representation of the tropical cyclone inner core, the operational use of AMVs would require a long-term evaluation. The appropriate treatment of correlated errors and biases in data assimilation, and significant improvements on AMV height assignment are of primary concern.

Acknowledgments

The authors gratefully acknowledge the support from National Oceanographic Partnership Program under ONR Marine Meteorology Program Award N00014-10-1-0123. The access to the NOAA T-Jet supercomputer was essential in enabling multiple data assimilation experiments and ensemble forecasts. The authors also thank the NCAR WRF-DART team for their help on various technical subjects, Dave Stettner of SSEC at University of Wisconsin–Madison for reprocessing hourly AMVs data for Hurricane Ike (2008), and Kathryn Sellwood of the Hurricane Research Division under the NOAA/Atlantic Oceanographic and Meteorological Laboratory for providing code to read NOAA dropwindsonde data. The authors wish to thank the three anonymous reviewers for their generous comments and careful reviews. Finally, great appreciation is expressed to the three anonymous reviewers for providing insightful comments in Wu et al. (2014), which later shaped the initial thoughts of this work.

REFERENCES

  • Aberson, S. D., , S. J. Majumdar, , C. A. Reynolds, , and B. J. Etherton, 2011: An Observing System Experiment for tropical cyclone targeting techniques using the Global Forecast System. Mon. Wea. Rev., 139, 895907, doi:10.1175/2010MWR3397.1.

    • Search Google Scholar
    • Export Citation
  • Anderson, J. L., 2003: A local least squares framework for ensemble filtering. Mon. Wea. Rev.,131, 634–642, doi:10.1175/1520-0493(2003)131<0634:ALLSFF>2.0.CO;2.

  • Anderson, J. L., , T. Hoar, , K. Raeder, , H. Liu, , N. Collins, , R. Torn, , and A. Avellano, 2009: The Data Assimilation Research Testbed: A community facility. Bull. Amer. Meteor. Soc., 90, 12831296, doi:10.1175/2009BAMS2618.1.

    • Search Google Scholar
    • Export Citation
  • Bedka, K. M., , C. S. Velden, , R. A. Petersen, , W. F. Feltz, , and J. R. Mecikalski, 2009: Comparisons of satellite-derived atmospheric motion vectors, rawinsondes, and NOAA wind profiler observations. J. Appl. Meteor. Climatol., 48, 15421561, doi:10.1175/2009JAMC1867.1.

    • Search Google Scholar
    • Export Citation
  • Berger, H., , R. Langland, , C. S. Velden, , C. A. Reynolds, , and P. M. Pauley, 2011: Impact of enhanced satellite-derived atmospheric motion vector observations on numerical tropical cyclone track forecasts in the western North Pacific during TPARC/TCS-08. J. Appl. Meteor. Climatol., 50, 23092318, doi:10.1175/JAMC-D-11-019.1.

    • Search Google Scholar
    • Export Citation
  • Brennan, M. J., , and S. J. Majumdar, 2011: An examination of model track forecast errors for Hurricane Ike (2008) in the Gulf of Mexico. Wea. Forecasting, 26, 848867, doi:10.1175/WAF-D-10-05053.1.

    • Search Google Scholar
    • Export Citation
  • Brown, D. P., , J. L. Beven, , J. L. Franklin, , and E. S. Blake, 2010: Atlantic Hurricane season of 2008. Mon. Wea. Rev., 138, 19752001, doi:10.1175/2009MWR3174.1.

    • Search Google Scholar
    • Export Citation
  • Goerss, J. S., 2009: Impact of satellite observations on the tropical cyclone track forecasts of the Navy Operational Global Atmospheric Prediction System. Mon. Wea. Rev., 137, 4150, doi:10.1175/2008MWR2601.1.

    • Search Google Scholar
    • Export Citation
  • Harnisch, F., , and M. Weissmann, 2010: Sensitivity of typhoon forecasts to different subsets of targeted dropsonde observations. Mon. Wea. Rev., 138, 26642680, doi:10.1175/2010MWR3309.1.

    • Search Google Scholar
    • Export Citation
  • Kieu, C. Q., , N. M. Truong, , H. T. Mai, , and T. Ngo-Duc, 2012: Sensitivity of the track and intensity forecasts of Typhoon Megi (2010) to satellite-derived atmospheric motion vectors with the ensemble Kalman filter. J. Atmos. Oceanic Technol., 29, 17941810, doi:10.1175/JTECH-D-12-00020.1.

    • Search Google Scholar
    • Export Citation
  • Komaromi, W. A., , S. J. Majumdar, , and E. D. Rappin, 2011: Diagnosing initial condition sensitivity of Typhoon Sinlaku (2008) and Hurricane Ike (2008). Mon. Wea. Rev., 139, 32243242, doi:10.1175/MWR-D-10-05018.1.

    • Search Google Scholar
    • Export Citation
  • Kurihara, Y., , M. A. Bender, , and R. J. Ross, 1993: An initialization scheme of hurricane models by vortex specification. Mon. Wea. Rev.,121, 2030–2045, doi:10.1175/1520-0493(1993)121<2030:AISOHM>2.0.CO;2.

  • Landsea, C. W., , and J. L. Franklin, 2013: Atlantic Hurricane database uncertainty and presentation of a new database format. Mon. Wea. Rev., 141, 35763592, doi:10.1175/MWR-D-12-00254.1.

    • Search Google Scholar
    • Export Citation
  • Langland, R. H., , C. Velden, , P. M. Pauley, , and H. Berger, 2009: Impact of satellite-derived rapid-scan wind observations on numerical model forecasts of Hurricane Katrina. Mon. Wea. Rev., 137, 16151622, doi:10.1175/2008MWR2627.1.

    • Search Google Scholar
    • Export Citation
  • Majumdar, S. J., and et al. , 2011: Targeted observations for improving numerical weather prediction: An overview. Tech. Rep. WMO/WWRP/THORPEX 15, 37 pp. [Available online at http://www.wmo.int/pages/prog/arep/wwrp/new/documents/THORPEX_No_15.pdf.]

  • Pu, Z., , X. Li, , C. S. Velden, , S. D. Aberson, , and W. T. Liu, 2008: The impact of aircraft dropsonde and satellite wind data on numerical simulations of two landfalling tropical storms during the Tropical Cloud Systems and Processes Experiment. Wea. Forecasting, 23, 6279, doi:10.1175/2007WAF2007006.1.

    • Search Google Scholar
    • Export Citation
  • Sears, J., , and C. S. Velden, 2012: Validation of satellite-derived atmospheric motion vectors and analyses around tropical disturbances. J. Appl. Meteor. Climatol., 51, 18231834, doi:10.1175/JAMC-D-12-024.1.

    • Search Google Scholar
    • Export Citation
  • Skamarock, W. C., and et al. , 2008: A description of the Advanced Research WRF version 3. NCAR Tech. Note NCAR/TN-475+STR, 113 pp. [Available online at http://www.mmm.ucar.edu/wrf/users/docs/arw_v3_bw.pdf.]

  • Torn, R. D., , and C. Snyder, 2012: Uncertainty of tropical cyclone best-track information. Wea. Forecasting, 27, 715729, doi:10.1175/WAF-D-11-00085.1.

    • Search Google Scholar
    • Export Citation
  • Velden, C. S., , and L. M. Leslie, 1991: The basic relationship between tropical cyclone intensity and the depth of the environmental steering layer in the Australian region. Wea. Forecasting, 6, 244253, doi:10.1175/1520-0434(1991)006<0244:TBRBTC>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Velden, C. S., , C. M. Hayden, , S. J. Nieman, , W. P. Menzel, , S. Wanzong, , and J. S. Goerss, 1997: Upper-tropospheric winds derived from geostationary satellite water vapor observations. Bull. Amer. Meteor. Soc.,78, 173–195, doi:10.1175/1520-0477(1997)078<0173:UTWDFG>2.0.CO;2.

  • Velden, C. S., and et al. , 2005: Recent innovations in deriving tropospheric winds from meteorological satellites. Bull. Amer. Meteor. Soc., 86, 205223, doi:10.1175/BAMS-86-2-205.

    • Search Google Scholar
    • Export Citation
  • Wu, T.-C., , H. Liu, , S. J. Majumdar, , C. S. Velden, , and J. L. Anderson, 2014: Influence of assimilating satellite-derived atmospheric motion vector observations on numerical analyses and forecasts of tropical cyclone track and intensity. Mon. Wea. Rev., 142, 4971, doi:10.1175/MWR-D-13-00023.1.

    • Search Google Scholar
    • Export Citation
  • Yamaguchi, M., , and S. J. Majumdar, 2010: Using TIGGE data to diagnose initial perturbations and their growth for tropical cyclone ensemble forecasts. Mon. Wea. Rev., 138, 36343655, doi:10.1175/2010MWR3176.1.

    • Search Google Scholar
    • Export Citation
  • Yamaguchi, M., , T. Iriguchi, , T. Nakazawa, , and C.-C. Wu, 2009: An Observing System Experiment for Typhoon Conson (2004) using a singular vector method and DOTSTAR data. Mon. Wea. Rev., 137, 28012816, doi:10.1175/2009MWR2683.1.

    • Search Google Scholar
    • Export Citation
1

Thinning and superobbing are both intended to lower the effect of correlated errors by reducing the data density and to have a spatial resolution the model can resolve. Observations are thinned to a predefined resolution by only selecting observation at a predefined spacing. Observations within a predefined prism are spatially averaged and weighted (in this study, the weight is uniform) to form a superob.

2

AMS Special Collections are available online at http://journals.ametsoc.org/page/collection.

3

Since Ike took a long and steady westward track through its life cycle, the analysis domain size is selected large enough to include all relevant atmospheric features throughout this period. Because of computational limitations, the domain’s eastern boundary is less than 1000 km from the center of Ike prior to 3 September 2008 and, therefore, limits the inclusion of subsets of enhanced AMV data in Ike’s earliest stages.

4

Observations with asymmetric distribution can cause problem in both objective analysis and data assimilation, especially satellite data assimilation, where observations are unevenly spaced and in regions of tight gradients (i.e., tropical cyclones).

5

Additional data include 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). Flight-level winds reduced to the surface together with Stepped Frequency Microwave Radiometer (SFMR) winds from reconnaissance aircraft were averaged.

Save