• Aksoy, A., , D. C. Dowell, , and C. Snyder, 2009: A multicase comparative assessment of the ensemble Kalman filter for assimilation of radar observations. Part I: Storm-scale analyses. Mon. Wea. Rev., 137, 18051824.

    • Search Google Scholar
    • Export Citation
  • Aksoy, A., , D. C. Dowell, , and C. Snyder, 2010: A multicase comparative assessment of the ensemble Kalman filter for assimilation of radar observations. Part II: Short-range ensemble forecasts. Mon. Wea. Rev., 138, 12731292.

    • Search Google Scholar
    • Export Citation
  • Anderson, J. L., 2009: Spatially and temporally varying adaptive covariance inflation for ensemble filters. Tellus, 61A, 7283.

  • Barnes, S. L., 1964: A technique for maximizing details in numerical weather map analysis. J. Appl. Meteor., 3, 396409.

  • Bluestein, H. B., 2009: The formation and early evolution of the Greensburg, Kansas, tornadic supercell on 4 May 2007. Wea. Forecasting, 24, 899920.

    • Search Google Scholar
    • Export Citation
  • Caya, A., , J. Sun, , and C. Snyder, 2005: A comparison between the 4D-VAR and the ensemble Kalman filter techniques for radar data assimilation. Mon. Wea. Rev., 133, 30813094.

    • Search Google Scholar
    • Export Citation
  • Coniglio, M. C., , D. J. Stensrud, , and L. J. Wicker, 2006: Effects of upper-level shear on the structure and maintenance of strong quasi-linear mesoscale convective systems. J. Atmos. Sci., 63, 12311252.

    • Search Google Scholar
    • Export Citation
  • Dawson, D. T., , M. Xue, , J. A. Milbrandt, , and M. K. Yau, 2010: Comparison of evaporation and cold pool development between single-moment and multimoment bulk microphysics schemes in idealized simulations of tornadic thunderstorms. Mon. Wea. Rev., 138, 11521171.

    • Search Google Scholar
    • Export Citation
  • Dowell, D. C., , and L. J. Wicker, 2009: Additive noise for storm-scale ensemble data assimilation. J. Atmos. Oceanic Technol., 26, 911927.

    • Search Google Scholar
    • Export Citation
  • Dowell, D. C., , F. Zhang, , L. J. Wicker, , C. Snyder, , and N. A. Crook, 2004: Wind and temperature retrievals in the 17 May 1981 Arcadia, Oklahoma, supercell: Ensemble Kalman filter experiments. Mon. Wea. Rev., 132, 19822005.

    • Search Google Scholar
    • Export Citation
  • Dowell, D. C., , L. J. Wicker, , and C. Snyder, 2011: Ensemble Kalman filter assimilation of radar observations of the 8 May 2003 Oklahoma City supercell: Influences of reflectivity observations on storm-scale analyses. Mon. Wea. Rev., 139, 272294.

    • Search Google Scholar
    • Export Citation
  • Gaspari, G., , and S. E. Cohn, 1999: Construction of correlation functions in two and three dimensions. Quart. J. Roy. Meteor. Soc., 125, 723757.

    • Search Google Scholar
    • Export Citation
  • Gilmore, M. S., , J. M. Straka, , and E. N. Rasmussen, 2004: Precipitation uncertainty due to variations in precipitation particle parameters within a simple microphysics scheme. Mon. Wea. Rev., 132, 26102627.

    • Search Google Scholar
    • Export Citation
  • Jung, Y., , M. Xue, , G. Zhang, , and J. M. Straka, 2008: Assimilation of simulated polarimetric radar data for a convective storm using the ensemble Kalman filter. Part II: Impact of polarimetric data on storm analysis. Mon. Wea. Rev., 136, 22462260.

    • Search Google Scholar
    • Export Citation
  • Lemon, L. R., , and M. Umscheid, 2008: The Greensburg, Kansas tornadic storm: A storm of extremes. Preprints, 24th Conf. on Severe Local Storms, Savannah, GA, Amer. Meteor. Soc., 2.4. [Available online at http://ams.confex.com/ams/pdfpapers/141811.pdf.]

    • Search Google Scholar
    • Export Citation
  • Lin, Y.-L., , R. D. Farley, , and H. D. Orville, 1983: Bulk parameterization of the snow field in a cloud model. J. Climate Appl. Meteor., 22, 10651092.

    • Search Google Scholar
    • Export Citation
  • Mansell, E. R., , C. L. Ziegler, , and E. C. Bruning, 2010: Simulated electrification of a small thunderstorm with two-moment bulk microphysics. J. Atmos. Sci., 67, 171194.

    • Search Google Scholar
    • Export Citation
  • Milbrandt, J. A., , and M. K. Yau, 2006: A multimoment bulk microphysics parameterization. Part IV: Sensitivity experiments. J. Atmos. Sci., 63, 31373159.

    • Search Google Scholar
    • Export Citation
  • Morrison, H., , G. Thompson, , and V. Tatarskii, 2009: Impact of cloud microphysics on the development of trailing stratiform precipitation in a simulated squall line: Comparison of one- and two-moment schemes. Mon. Wea. Rev., 137, 9911007.

    • Search Google Scholar
    • Export Citation
  • Oye, R., , C. Mueller, , and S. Smith, 1995: Software for radar translation, visualization, editing, and interpolation. Preprints, 27th Conf. on Radar Meteorology, Vail, CO, Amer. Meteor. Soc., 359–361.

    • Search Google Scholar
    • Export Citation
  • Snook, N., , and M. Xue, 2008: Effects of microphysical drop size distribution on tornadogenesis in supercell thunderstorms. Geophys. Res. Lett., 35, L24803, doi:10.1029/2008GL035866.

    • Search Google Scholar
    • Export Citation
  • Stensrud, D. J., , and J. Gao, 2010: Importance of horizontally inhomogeneous environmental initial conditions to ensemble storm-scale radar data assimilation and very short-range forecasts. Mon. Wea. Rev., 138, 12501272.

    • Search Google Scholar
    • Export Citation
  • Stensrud, D. J., and Coauthors, 2009: Convective-scale warn-on-forecast system: A vision for 2020. Bull. Amer. Meteor. Soc., 90, 14871499.

    • Search Google Scholar
    • Export Citation
  • Sun, J., , and N. A. Crook, 2001: Real-time low-level wind and temperature analysis using single WSR-88D data. Wea. Forecasting, 16, 117132.

    • Search Google Scholar
    • Export Citation
  • van den Heever, S. C., , and W. R. Cotton, 2004: The impact of hail size on simulated supercell storms. J. Atmos. Sci., 61, 15961609.

  • Wicker, L. J., , and R. B. Wilhelmson, 1995: Simulation and analysis of tornado development and decay within a three-dimensional supercell thunderstorm. J. Atmos. Sci., 52, 26752703.

    • Search Google Scholar
    • Export Citation
  • Ziegler, C. L., 1985: Retrieval of thermal and microphysical variables in observed convective storms. Part I: Model development and preliminary testing. J. Atmos. Sci., 42, 14871509.

    • Search Google Scholar
    • Export Citation
  • View in gallery

    Initial sounding environment used in experiments after interpolation to the model vertical grid. The thermodynamic profile (black for temperature, blue for dewpoint temperature) is the same for all experiments and all members. The wind profile (shown here for the 0130 UTC KVNX VAD) is varied for different experiments (see Fig. 2). Each full (half) barb represents 10 (5) m s−1.

  • View in gallery

    Hodographs for the three sets of experiments, after interpolation to the model vertical grid. All hodographs have identical surface winds (taken from the KPTT observation at 0200 UTC) and winds above 3.5 km AGL. Between the surface and 3.5 km AGL, the KVNX VAD wind profile valid at 0130 (solid), 0200 (dashed), and 0230 UTC (dotted) are used in the 3 experiment sets, respectively.

  • View in gallery

    Schematic of sounding design. All heights are in m MSL. The construction of the wind profile is accomplished by interpolating the KPTT surface wind observation to the lowest KVNX VAD wind at 100 m AGL (750 m MSL, shifted upward from the original level of 470 m MSL). The KVNX VAD winds are then used up to 3500 m AGL, and the DDC raob winds above that level. The thermodynamic profile is taken from the original MSL sounding levels from the DDC raob, but in the lowest 900 m AGL the profile is replaced with a well-mixed layer consistent with the 0200 UTC KPTT surface thermodynamic conditions.

  • View in gallery

    Observation-space diagnostics for reflectivity for each of the assimilation and forecast experiments. The solid, dashed, and dotted lines denote experiments using the 0130, 0200, and 0230 UTC KVNX VAD wind profiles, respectively. Red, black, and blue lines indicate RMS innovation, mean innovation, and RMS standard deviation (ensemble spread), respectively. Thin lines represent the EnKF analysis cycle, in which the sawtooth pattern indicates that both prior and posterior statistics were included. The bold lines starting at ~0130, ~0145, and 0200 UTC represent the statistics computed for the free forecast ensembles, which were initialized at 0130 (no markers), 0145 (filled square markers) and 0200 UTC (filled circle markers), respectively. Statistics are calculated domain wide below 10 km AGL, and for all reflectivities greater than 15 dBZ. In addition, the free forecast ensemble statistics are those computed for the ensemble mean only. The vertical green bars indicate the initial times for the forecast ensemble experiments (0130, 0145, and 0200 UTC, respectively).

  • View in gallery

    As in Fig. 4, but for radial velocity (no thresholds applied).

  • View in gallery

    Individual 0–1-h vorticity swaths for the first 28 (of 30) ensemble members of experiment V0200I0200. For reference, overlaid in each are the tracks of the Greensburg tornado, as well as the two subsequent large tornadoes from the storm. The location of Greensburg, KS, is denoted by the yellow star. The scale is indicated in km in the lower left. Only a portion of the full model domain is shown.

  • View in gallery

    Probabilistic vorticity swaths for the ensemble forecasts starting at 0130, 0145, and 0200 UTC at 75 m AGL. Each plot is labeled with the corresponding experiment name as listed in Table 1 and in the text. Colors in the swath indicate the ensemble probability that a particular grid point exceeds the given vorticity threshold during the period of (top) 0130–0315 UTC, (middle) 0145–0315 UTC, and (bottom) 0200–0315 UTC. Also plotted in yellow are isochrones of the ensemble average time (in 15-min increments) that the vorticity threshold was exceeded within the swath time interval. For clarity, the isochrones are manually approximated versions of the actual isochrones. Each isochrone is labeled on the right with forecast time in minutes, and on the left with the valid UTC time. For reference, observed damage tracks from the Greensburg tornado and the two subsequent tornadoes are indicated, and the location of Greensburg is denoted by the yellow star. Finally, purple circles mark the approximate location of the low-level mesocyclone associated with the Greensburg tornado at 15-min intervals from 0200 to 0300 UTC.

  • View in gallery

    As in Fig. 7, but for 975 m AGL.

  • View in gallery

    The 45-min forecast simulated reflectivity (color fill) at 975 m AGL, valid at 0245 UTC 5 May 2007 for 24 of the 30 members of forecast experiment V0200I0200. (Member numbers are labeled.) Vorticity is contoured (thick black lines) in 0.005 s−1 intervals, starting at 0.01 s−1. Vertical velocity at ~2 km AGL is also contoured (10 m s−1; thick purple lines). (top left) Objectively analyzed [to a regular 1-km grid using the Barnes (1964) scheme] observed reflectivity from the 0.5° KDDC base scan valid at ~0246 UTC. In each, as in Figs. 68, tornado tracks and the location of Greensburg, KS (yellow star), are overlaid. (bottom left) The scale in km is shown.

  • View in gallery

    The 45-min forecast surface potential temperature perturbation (color fill) and reflectivity (black contours; 10-dBZ increment) for the first 12 members of (a) V0130I0200 and (b) V0200I0200.

  • View in gallery

    As in Fig. 5, but for V0200I0200 repeated at 2-km grid spacing (solid lines) and the corresponding experiment with reduced midlevel winds (dashed lines).

  • View in gallery

    As in Fig. 7, but for the (top) I0215 and (bottom) I0230 experiments, and swaths are integrated out to 0330 UTC.

  • View in gallery

    As in Fig. 12, but for 975 m AGL.

  • View in gallery

    As in Fig. 9, but for V0200I0215.

  • View in gallery

    As in Fig. 7h, but for the V0200I0200 experiment repeated with LFO microphysics.

  • View in gallery

    Surface (75 m AGL) perturbation potential temperature (color fill; 1-K increment), reflectivity (black contours, 10-dBZ increment), vorticity (0.005 s−1 increment, starting at 0.01 s−1) at the surface (red contours) and 975 m AGL (green contours), and horizontal wind vectors (every third grid point, scale at bottom left) for member 16 of the LFO version of V0200I0200 at (a) the initial time of 0200 UTC and (b) 0230 UTC (30-min forecast). (c),(d), As in (a),(b), but for member 22 of the original ZVD V0200I0200 experiment.

All Time Past Year Past 30 Days
Abstract Views 0 0 0
Full Text Views 144 144 17
PDF Downloads 79 79 13

Impact of the Environmental Low-Level Wind Profile on Ensemble Forecasts of the 4 May 2007 Greensburg, Kansas, Tornadic Storm and Associated Mesocyclones

View More View Less
  • 1 NOAA/National Severe Storms Laboratory, Norman, Oklahoma
  • | 2 National Weather Center, and Center for Analysis and Prediction of Storms, University of Oklahoma, Norman, Oklahoma
© Get Permissions
Full access

Abstract

The early tornadic phase of the Greensburg, Kansas, supercell on the evening of 4 May 2007 is simulated using a set of storm-scale (1-km horizontal grid spacing) 30-member ensemble Kalman filter (EnKF) data assimilation and forecast experiments. The Next Generation Weather Radar (NEXRAD) level-II radar data from the Dodge City, Kansas (KDDC), Weather Surveillance Radar-1988 Doppler (WSR-88D) are assimilated into the National Severe Storms Laboratory (NSSL) Collaborative Model for Multiscale Atmospheric Simulation (COMMAS). The initially horizontally homogeneous environments are initialized from one of three reconstructed soundings representative of the early tornadic phase of the storm, when a low-level jet (LLJ) was intensifying. To isolate the impact of the low-level wind profile, 0–3.5-km AGL wind profiles from Vance Air Force Base, Oklahoma (KVNX), WSR-88D velocity-azimuth display (VAD) analyses at 0130, 0200, and 0230 UTC are used. A sophisticated, double-moment bulk ice microphysics scheme is employed.

For each of the three soundings, ensemble forecast experiments are initiated from EnKF analyses at various times prior to and shortly after the genesis of the Greensburg tornado (0200 UTC). Probabilistic forecasts of the mesocyclone-scale circulation(s) are generated and compared to the observed Greensburg tornado track. Probabilistic measures of significant rotation and observation-space diagnostic statistics are also calculated. It is shown that, in general, the track of the Greensburg tornado is well predicted, and forecasts improve as forecast lead time decreases. Significant variability is also seen across the experiments using different VAD wind profiles. Implications of these results regarding the choice of initial mesoscale environment, as well as for the “Warn-on-Forecast” paradigm for probabilistic numerical prediction of severe thunderstorms and tornadoes, are discussed.

Corresponding author address: Daniel T. Dawson II, National Severe Storms Laboratory, 120 David L. Boren Blvd., Norman, OK 73072. E-mail: dan.dawson@noaa.gov

Abstract

The early tornadic phase of the Greensburg, Kansas, supercell on the evening of 4 May 2007 is simulated using a set of storm-scale (1-km horizontal grid spacing) 30-member ensemble Kalman filter (EnKF) data assimilation and forecast experiments. The Next Generation Weather Radar (NEXRAD) level-II radar data from the Dodge City, Kansas (KDDC), Weather Surveillance Radar-1988 Doppler (WSR-88D) are assimilated into the National Severe Storms Laboratory (NSSL) Collaborative Model for Multiscale Atmospheric Simulation (COMMAS). The initially horizontally homogeneous environments are initialized from one of three reconstructed soundings representative of the early tornadic phase of the storm, when a low-level jet (LLJ) was intensifying. To isolate the impact of the low-level wind profile, 0–3.5-km AGL wind profiles from Vance Air Force Base, Oklahoma (KVNX), WSR-88D velocity-azimuth display (VAD) analyses at 0130, 0200, and 0230 UTC are used. A sophisticated, double-moment bulk ice microphysics scheme is employed.

For each of the three soundings, ensemble forecast experiments are initiated from EnKF analyses at various times prior to and shortly after the genesis of the Greensburg tornado (0200 UTC). Probabilistic forecasts of the mesocyclone-scale circulation(s) are generated and compared to the observed Greensburg tornado track. Probabilistic measures of significant rotation and observation-space diagnostic statistics are also calculated. It is shown that, in general, the track of the Greensburg tornado is well predicted, and forecasts improve as forecast lead time decreases. Significant variability is also seen across the experiments using different VAD wind profiles. Implications of these results regarding the choice of initial mesoscale environment, as well as for the “Warn-on-Forecast” paradigm for probabilistic numerical prediction of severe thunderstorms and tornadoes, are discussed.

Corresponding author address: Daniel T. Dawson II, National Severe Storms Laboratory, 120 David L. Boren Blvd., Norman, OK 73072. E-mail: dan.dawson@noaa.gov

1. Introduction

Over the past decade, the ensemble Kalman filter (EnKF) method has been successfully applied in analyses of real-world severe convective storms, taking advantage of the rich source of data provided by Doppler radar (e.g., Dowell et al. 2004; Aksoy et al. 2009; Dowell et al. 2011). Much progress has been made in coupling the data assimilation system to cloud-resolving numerical models (e.g., Jung et al. 2008; Aksoy et al. 2009; Dowell and Wicker 2009), but many challenges remain. In general, the application of the EnKF technique to convective storm modeling and data assimilation has been focused on producing storm-scale analyses, rather than predictions. Comparatively little attention has been paid to ensemble forecasts of individual convective cells, let alone substorm-scale features such as mesocyclones and tornadoes, in no small part due to the substantial computational cost. Aksoy et al. (2010, hereafter ADS10) performed short-range (30 min) ensemble forecasts, at 2-km horizontal grid spacing, of three different convective events (a squall line, a multicell thunderstorm, and a supercell thunderstorm) that were initialized from a previous EnKF analysis cycle. They found, through examining observation-space diagnostic metrics based on reflectivity, that divergence of the forecasts among the various ensemble members occurred on relatively short time scales [O(10 min)] in all their experiments. In addition, for their supercell experiment, a tendency for rapid decay of the modeled supercell was seen across the entire ensemble, whereas the observed supercell was not observed to decay during this time period. ADS10 attributed this decay to errors in the analysis (especially an anomalous warm pool underneath the storm) and an initial environmental sounding that may not have been representative of the mature supercell’s surrounding environment. Stensrud and Gao (2010, hereafter SG10) discuss two storm-scale ensemble forecast experiments for the 4 May 2007 Greensburg, Kansas, tornadic supercell, based either on a series of horizontally homogeneous environments from point soundings derived from a mesoscale ensemble forecast [itself based on an intermittent three-dimensional variational data assimilation (3DVAR) analysis system], or on horizontally inhomogeneous environments with externally forced boundary conditions from the same mesoscale ensemble. They found that forecasted storm behavior across the ensemble was substantially improved when using the more realistic inhomogeneous mesoscale environment, rather than the traditional single-sounding environment used in most studies to date. They described resultant improvements in storm track and storm evolution [i.e., maintenance of supercell structures vs upscale growth into a mesoscale convective system (MCS)], including more realistic reflectivity structure and mesocyclonic circulations that were closer in strength to the observed (SG10).

Despite the advantages of the more sophisticated, horizontally inhomogeneous mesoscale environment approach, there are also some advantages to using a single-sounding, horizontally homogeneous environment. First, it is inherently simpler to implement than an inhomogeneous environment. Second, while an inhomogeneous environment may lead to better forecast accuracy (SG10), it is more difficult to assess impacts of the mesoscale environment. These impacts are more easily isolated in a single-sounding approach. Third, sensitivity studies are more easily conducted in such a framework. It is also important in the beginning states of a new research program to start with simplified experimental setups before ramping up the complexity to more realistic setups that would be suitable for an operational NWP system. A follow-up study is planned in which a full-physics inhomogeneous environment will be examined, but in the meantime we hold that such idealized experiments are still very useful in guiding the development of future experiments and real-time, possibly operational forecast frameworks for the prediction of high-impact events such as supercells and tornadoes. We therefore chose to implement the single-sounding approach in this study, with the goal of assessing the impact on and sensitivity of the forecasts to the low-level environmental wind profile.

Another important source of model error that affects both the free forecast and the filter performance during the assimilation period is the parameterization of cloud and precipitation microphysics. Several studies have shown significant sensitivity of storm behavior to the choice of fixed parameters in single-moment microphysics schemes (e.g., Gilmore et al. 2004; van den Heever and Cotton 2004; Snook and Xue 2008), as well as notable differences in behavior between single- and multimoment schemes (Milbrandt and Yau 2006; Morrison et al. 2009; Dawson et al. 2010). Milbrandt and Yau (2006) and Dawson et al. (2010) have shown an improvement in supercell storm structure and behavior when utilizing a double-moment (or higher) scheme over a single-moment scheme, partly owing to the greater flexibility in hydrometeor-size-distribution parameters and associated feedbacks to latent cooling within the storm downdrafts (and thus, cold pool intensity). In this study, we chose a double-moment microphysics parameterization (Ziegler 1985; Mansell et al. 2010) for the main experiments, which should help minimize the generation and accumulation of error from the microphysics.

A probabilistic forecast of possible storm tracks, evolution of storm mode, and (in the case of supercells) mesocyclone and tornado tracks, is one of the major goals of the Warn-on-Forecast (WoF) paradigm (Stensrud et al. 2009). Since the EnKF produces an ensemble of initial conditions, it naturally lends itself to this goal. Thus, the stated goal of WoF provides an overarching framework for this study, and this goal should be kept in mind when interpreting our conclusions. Within this larger context, the main goal of this study is to examine the numerical prediction of mesocyclone-scale circulation features (used as a proxy for tornado tracks) in the 4 May 2007 Greensburg tornadic supercell. We are particularly interested in the impact of the initial low-level wind profile and the forecast lead time on ensemble forecasts of the Greensburg storm. The forecast experiments are evaluated both subjectively from the standpoint of overlap with the observed Greensburg tornado track, and objectively from probabilities of significant rotation at a given location and time computed from the ensemble of mesocyclone tracks. Observation-space diagnostic statistics are examined for the forecast ensemble. The current study is similar in many ways to that of SG10, but the use of the EnKF technique for assimilating radar reflectivity and velocity data, rather than the 3DVAR technique used in that study, is a distinguishing feature. In addition, our study examines the sensitivity of the forecast mesocyclone tracks to the initial time, which was not investigated in SG10. We view this study, therefore, as being an important complement to SG10, and serves to provide a comparison between ensemble forecasts initialized with a 3DVAR-based analysis system to that of an EnKF-based system.

This paper is organized as follows. In section 2, we discuss the background of the Greensburg storm and the design of the ensemble analysis and forecast experiments. Section 3 covers the results of the experiments in light of the choices of environmental sounding and forecast lead time. We summarize our conclusions in section 4 and discuss future work.

2. Overview of the Greensburg storm and experiment design

a. Overview of the Greensburg storm and model initial environment

As discussed at length by Lemon and Umscheid (2008, hereafter LU08), the combination of strong vertical wind shear and thermodynamic instability characterizing the Greensburg storm’s environment was an extreme outlier on a classic CAPE versus 0–1-km storm-relative helicity (SRH) graph (see their Fig. 5). The Greensburg storm itself first developed after a long series of cell splits and mergers near the Oklahoma–Kansas border in the extreme eastern Oklahoma Panhandle between 0013 and 0038 UTC 5 May 2007, and first became tornadic around 0132 UTC (Bluestein 2009). The storm produced at least four small and relatively short-lived tornadoes (rated EF0–EF1; M. Umscheid 2009, personal communication) before producing its first significant long-track tornado (hereafter referred to simply as the “Greensburg tornado,” which was rated EF5). The Greensburg tornado began at approximately 0200 UTC, struck the town of Greensburg just after 0245 UTC, and finally dissipated at approximately 0300 UTC (LU08). The Greensburg tornado had a mean path width of approximately 2.0 km, and a maximum path width of 3.1 km, placing it among the largest tornadoes ever recorded (LU08). The prediction of the mesocyclone associated with this tornado is the main focus of our discussion, although the parent storm went on to produce several more long-track, large tornadoes. The remnant circulation of the Greensburg tornado became involved with the developing circulation of the next large tornado in the sequence, hereafter referred to as the “Trousdale tornado” because of its proximity to the town of Trousdale, Kansas. We chose to focus on the Greensburg tornado period, because of its closer proximity in time to the Dodge City, Kansas (DDC) sounding, proximity in space to the KDDC radar, and computational considerations. However, in section 3e, we also discuss results of “negative lead time” ensemble forecasts initialized after the genesis of the Greensburg tornado (at 0215 and 0230 UTC, respectively), which show evidence of predicting the cycling between the Greensburg and Trousdale tornadoes. For more details on the evolution of the mesoscale and storm-scale environment leading up to the development of the Greensburg storm see Bluestein (2009), and for further details on the Greensburg storm and tornado itself see LU08.

We used the National Severe Storms Laboratory (NSSL) Collaborative Model for Multiscale Atmospheric Simulation (COMMAS; Wicker and Wilhelmson 1995; Coniglio et al. 2006; Mansell et al. 2010) as the numerical simulation model for this set of analysis and forecast experiments. The horizontally homogeneous initial environment was based on 1 of 3 different soundings having identical thermodynamic profiles (Fig. 1), but different 0–3.5-km AGL wind profiles (Fig. 2). The thermodynamic profile was created using the 0000 UTC 5 May 2007 DDC rawinsonde above the boundary layer, and a well-mixed temperature and moisture profile in the boundary layer (approximately 900 m deep) based on Pratt, Kansas (KPTT), Automated Surface Observing System (ASOS) measurements taken at 0200 UTC 5 May 2007. This well-mixed layer was inserted because the 0000 UTC DDC rawinsonde was launched behind the dryline, while the Pratt surface observations were more representative of the low-level inflow environment of the Greensburg tornadic supercell.

Fig. 1.
Fig. 1.

Initial sounding environment used in experiments after interpolation to the model vertical grid. The thermodynamic profile (black for temperature, blue for dewpoint temperature) is the same for all experiments and all members. The wind profile (shown here for the 0130 UTC KVNX VAD) is varied for different experiments (see Fig. 2). Each full (half) barb represents 10 (5) m s−1.

Citation: Monthly Weather Review 140, 2; 10.1175/MWR-D-11-00008.1

Fig. 2.
Fig. 2.

Hodographs for the three sets of experiments, after interpolation to the model vertical grid. All hodographs have identical surface winds (taken from the KPTT observation at 0200 UTC) and winds above 3.5 km AGL. Between the surface and 3.5 km AGL, the KVNX VAD wind profile valid at 0130 (solid), 0200 (dashed), and 0230 UTC (dotted) are used in the 3 experiment sets, respectively.

Citation: Monthly Weather Review 140, 2; 10.1175/MWR-D-11-00008.1

The three vertical wind profiles were constructed using the 0200 UTC KPTT surface wind for all cases, the 0130, 0200, and 0230 UTC velocity-azimuth display (VAD) wind profiles up to 3.5 km AGL from the Vance Air Force Base, Oklahoma (KVNX), Weather Surveillance Radar-1988 Doppler (WSR-88D), respectively, and the wind profiles above 3.5 km AGL from the 0000 UTC DDC rawinsonde. The construction of the soundings is shown schematically in Fig. 3. To summarize, the three soundings used in the 3 sets of experiments were identical in all respects except for the surface-to-3.5-km AGL layer winds. The VAD wind profiles were chosen to be representative of the strengthening low-level shear environment during the development and evolution of the storm shortly before and while it was producing the Greensburg tornado.

Fig. 3.
Fig. 3.

Schematic of sounding design. All heights are in m MSL. The construction of the wind profile is accomplished by interpolating the KPTT surface wind observation to the lowest KVNX VAD wind at 100 m AGL (750 m MSL, shifted upward from the original level of 470 m MSL). The KVNX VAD winds are then used up to 3500 m AGL, and the DDC raob winds above that level. The thermodynamic profile is taken from the original MSL sounding levels from the DDC raob, but in the lowest 900 m AGL the profile is replaced with a well-mixed layer consistent with the 0200 UTC KPTT surface thermodynamic conditions.

Citation: Monthly Weather Review 140, 2; 10.1175/MWR-D-11-00008.1

b. Model configuration

The domain for all simulations was 140 km × 160 km × 20 km, with 50 vertical levels, a stretched vertical grid with spacing varying from 150 m near the surface to 600 m near the model top, and uniform 1-km horizontal grid spacing. The domain was positioned such that Greensburg (37.602°N, 99.292°W) was located 70 km east and 90 km north of the southwest corner of the grid. Owing to the relatively small domain, a flat-earth map projection was assumed. We included the Coriolis force, but operating only on the perturbation winds and using the f-plane approximation valid at the central latitude of the grid. The relatively dense vertical grid spacing near the surface was found to be important for resolving the low-level circulation and cold pool features of the simulated storm; tests with coarser resolution near the surface resulted in much weaker low-level circulations and a degraded track forecast (not shown). The two-moment 6-class bulk ice microphysics scheme described in Mansell et al. (2010) and Ziegler (1985) [i.e., the Ziegler variable density (ZVD)] was employed for most experiments. This scheme predicts mass mixing ratio and number concentration for two classes of liquid hydrometeors (cloud and rain) and four classes of ice (pristine ice, snow aggregates, graupel, and hail). In addition, the bulk densities of both graupel and hail are predicted. Further details of the scheme can be found in Mansell et al. (2010).

c. Ensemble analysis and forecast configuration

The COMMAS model has been recently coupled to an EnKF radar data assimilation system (Dowell and Wicker 2009; Dowell et al. 2011). We used this system to assimilate reflectivity and radial Doppler velocity data from the Dodge City, Kansas, WSR-88D radar (KDDC). First, the velocity data were manually dealiased using Solo II radar data-editing software (Oye et al. 1995). Both the reflectivity and velocity data were then objectively analyzed, using a single-pass Barnes (1964) objective analysis scheme, to a regularly spaced 2-km horizontal grid, but left at the elevations of the individual conical sweep surfaces (Sun and Crook 2001; Dowell et al. 2004; Dowell and Wicker 2009). No thresholding or special treatment of low-reflectivity observations was employed. The objectively analyzed data were then partitioned into 2-min bins and assimilated every 2 min from 0030 to 0300 UTC, covering the 90-min period prior to the genesis (0030–0200 UTC), and the 60-min life cycle (0200–0300 UTC) of the long-track Greensburg tornado. A compactly supported, fifth-order localization function (Gaspari and Cohn 1999) with a horizontal (vertical) radius of 6 (3) km was employed. The reflectivity and radial velocity observations were used to update all the hydrometeor variables within the filter, including the number concentrations, as well as all other state variables (except for the mixing coefficient and Exner function) through the covariance structures provided by the filter. The assumed observation errors for reflectivity and radial velocity were 5 dBZ and 2 m s−1, respectively. KDDC was operating in the volume coverage pattern (VCP) 12 mode, with approximately 4.1 min between volumes, so that the 2-min bins used in this study covered approximately half a volume scan. Thirty ensemble members were used for the EnKF cycles for all experiments. Random horizontal wind perturbations with maximum magnitudes of 2 m s−1 were applied to the initial wind profile of each ensemble member. This value of the maximum sounding wind perturbation is roughly of the same order as the difference in the magnitude of the low-level jet between the three VAD wind profiles used (cf. Fig. 2). Additive noise using the technique of Caya et al. (2005) was applied every 360 s in regions of observed reflectivity >30 dBZ at analysis times to enhance ensemble spread. This noise was applied between 250 and 2500 m AGL to the following fields: the horizontal wind components, the temperature, and the dewpoint (the latter two were then converted back to the model state variables of potential temperature and water vapor mixing ratio). These perturbations had standard deviations of 1.0 m s−1 for the horizontal wind components, and 0.5 K for the temperature and dewpoint fields. Finally, random thermal bubblelike perturbations were applied to regions (again every 360 s between 250 and 2500 m AGL) where observed composite reflectivity was 30 dBZ or greater than that in the model, after Dowell and Wicker (2009), in order to encourage updraft development in observed storm regions.

Domain-wide observation-space diagnostic statistics (including RMS innovation, mean innovation, and RMS standard deviation) for reflectivity were computed for regions with observed reflectivity >15 dBZ and height AGL <10 km (Fig. 4), and statistics for radial velocity were computed for all observed values and heights AGL <10 km (Fig. 5). The innovation d is defined as the difference between the observation and the model state mapped to that observation at a given point:
e1
where y0 is the observation, x is the model state vector (either prior or posterior), and H represents the forward operator that maps the state vector to the observation location and type. The mean innovation is given simply by , where the denotes an average over all the observations in a radar volume. The RMS innovation is given by
e2
and the RMS standard deviation (or spread) by
e3
where N is the number of ensemble members, the subscript n represents a given ensemble member, and the overbar represents the ensemble mean. These definitions are the same as used in Dowell et al. (2011). Briefly, for a given observed variable, the RMS innovation gives a measure of the overall fit of the observations to the forecast, the mean innovation yields information about possible sources of bias in the model and/or observations, and the RMS standard deviation gives information about the degree of spread of the ensemble.
Fig. 4.
Fig. 4.

Observation-space diagnostics for reflectivity for each of the assimilation and forecast experiments. The solid, dashed, and dotted lines denote experiments using the 0130, 0200, and 0230 UTC KVNX VAD wind profiles, respectively. Red, black, and blue lines indicate RMS innovation, mean innovation, and RMS standard deviation (ensemble spread), respectively. Thin lines represent the EnKF analysis cycle, in which the sawtooth pattern indicates that both prior and posterior statistics were included. The bold lines starting at ~0130, ~0145, and 0200 UTC represent the statistics computed for the free forecast ensembles, which were initialized at 0130 (no markers), 0145 (filled square markers) and 0200 UTC (filled circle markers), respectively. Statistics are calculated domain wide below 10 km AGL, and for all reflectivities greater than 15 dBZ. In addition, the free forecast ensemble statistics are those computed for the ensemble mean only. The vertical green bars indicate the initial times for the forecast ensemble experiments (0130, 0145, and 0200 UTC, respectively).

Citation: Monthly Weather Review 140, 2; 10.1175/MWR-D-11-00008.1

Fig. 5.
Fig. 5.

As in Fig. 4, but for radial velocity (no thresholds applied).

Citation: Monthly Weather Review 140, 2; 10.1175/MWR-D-11-00008.1

Overall, the statistics exhibit relatively small differences between the assimilation experiments with different VAD wind profiles. The experiment using the 0130 UTC VAD wind profile has the best observation-space statistics for both reflectivity and radial velocity, with the lowest overall RMS innovations for both (thin red lines in Figs. 4 and 5), the lowest mean innovation for reflectivity (thin black line in Fig. 4), and the least biased mean innovation for radial velocity (thin black line in Fig. 5). The experiment using the 0230 UTC VAD has the worst overall statistics (thin dotted lines in Figs. 4 and 5), although the differences are relatively small compared to the raw magnitudes of the various statistics. However, as will be discussed, this trend does not necessarily carry over to the free forecast ensemble statistics (bold lines in Figs. 4 and 5).

To test the sensitivity of forecasts to different lead times leading up to the genesis of the Greensburg tornado, sets of free forecasts were launched from each of the 30 ensemble analyses1 valid at 0130, 0145, and 0200 UTC (vertical green bars in Figs. 4 and 5), and run to 0330 UTC. In each experiment, the forecast period covers the entire lifetime of the Greensburg tornado (approximately 0200–0300 UTC), and continues into the early period of the Trousdale tornado. Additional forecast experiments at “negative” lead times (i.e., initialized after the genesis of the Greensburg tornado, at 0215 and 0230 UTC, respectively) were also performed for each of the three VAD profiles (for a total of 6 additional forecast experiments). These additional experiments will be discussed further in section 3e. The reflectivity innovation statistics in the analysis period bottom out between roughly 0130 and 0200 UTC (thin lines in Fig. 4), suggesting that the storm is relatively well “spun up” in the assimilation by this time frame. Visual inspection of both the mean and individual members (not shown) supports this conclusion.

To summarize, a total of three separate EnKF analysis experiments (where “experiment” here refers to the entire EnKF assimilation period, 0030–0300 UTC) were performed, one for each of the KVNX VAD wind profiles. For each of the three analysis experiments, 5 ensemble forecast experiments (where experiment here refers to a set of 30 ensemble forecasts) were performed, initialized from the ensemble member analysis states in 15-min intervals from 0130 to 0230 UTC, for a total of 15 ensemble “free” forecast experiments. These times were chosen to investigate the forecast sensitivity to lead time (30, 15, 0, −15, and −30 min, respectively, prior to genesis) of the Greensburg tornado. These forecast experiments, summarized in Table 1, will have the following nomenclature throughout the rest of the paper: VHHMMIHHMM, where HHMM is the 4-digit hour and minute, “V” stands for the VAD wind profile used, and “I” stands for the initial forecast time. When discussing groups of experiments by VAD profile or initial time, the nomenclature VHHMM and IHHMM will be used, respectively. As an example, the experiment using the 0200 UTC VAD profile, with ensemble forecasts initialized at 0145 UTC, is denoted by V0200I0145.

Table 1.

Experiment naming convention.

Table 1.

3. Results

a. General examination of probabilistic vortex swaths

To evaluate the free forecasts of the mesocyclonic circulations in the Greensburg storm, we examine time-integrated vertical vorticity “swaths” as a proxy for the track of the circulation center at different heights. The swaths are computed at a given height as the maximum vertical vorticity experienced by each grid point during the forecast period (i.e., 0130–0315 UTC for the I0130 forecasts, 0145–0315 UTC for the I0145 forecasts, and 0200–0315 UTC for the I0200 forecasts). The individual members vary significantly in the details of individual circulation tracks (shown for V0200I0200 in Fig. 6), with some members having coherent vortices whose tracks strongly resemble the observed Greensburg tornado track (overlaid in each panel in Fig. 6), others showing multiple individual swaths, and still others having relatively weak vorticity at the surface. We emphasize that since the model horizontal grid spacing is too coarse to resolve the actual tornadic circulations, these swaths should be considered only rough approximations or “proxies” to forecast tornado tracks. The magnitude of variability underscores the advantages of the ensemble approach for storm-scale forecasting versus the purely deterministic approach commonly used. Clearly, individual swaths vary substantially, but there is also a broad agreement on swath orientation and length. Caution must be exercised in the interpretation of these swaths, however, as it appears that the overall ensemble of initial conditions derived from the EnKF analyses exhibits too little spread (cf. Figs. 4 and 5). Future work may focus on optimizing the spread of the initial ensemble, perhaps by utilizing more advanced state-dependent covariance inflation techniques (Anderson 2009; Whitaker and Hamill 2011, manuscript submitted to Mon. Wea. Rev.). Nevertheless, by examining the ensembles of forecast vorticity in this manner, the variations across the individual members are “smoothed.” Then the effects from the large-scale environment, which is the focus of this paper, are made more apparent.

Fig. 6.
Fig. 6.

Individual 0–1-h vorticity swaths for the first 28 (of 30) ensemble members of experiment V0200I0200. For reference, overlaid in each are the tracks of the Greensburg tornado, as well as the two subsequent large tornadoes from the storm. The location of Greensburg, KS, is denoted by the yellow star. The scale is indicated in km in the lower left. Only a portion of the full model domain is shown.

Citation: Monthly Weather Review 140, 2; 10.1175/MWR-D-11-00008.1

Having an ensemble of vorticity swaths for each experiment, the probabilities of vorticity exceeding a given threshold can be readily computed from the ensemble. We chose to examine the model “surface” level (actually 75 m AGL—the first scalar level above ground), which represents the tornado track (Fig. 7), and the model level corresponding to 975 m AGL, which represents the low-level mesocyclone track (Fig. 8). Locations where vorticity is greater than a threshold value (0.01 s−1) are tabulated at each spatial grid point across the ensemble,2 and the ensemble probability of a point exceeding this threshold is then computed.

Fig. 7.
Fig. 7.

Probabilistic vorticity swaths for the ensemble forecasts starting at 0130, 0145, and 0200 UTC at 75 m AGL. Each plot is labeled with the corresponding experiment name as listed in Table 1 and in the text. Colors in the swath indicate the ensemble probability that a particular grid point exceeds the given vorticity threshold during the period of (top) 0130–0315 UTC, (middle) 0145–0315 UTC, and (bottom) 0200–0315 UTC. Also plotted in yellow are isochrones of the ensemble average time (in 15-min increments) that the vorticity threshold was exceeded within the swath time interval. For clarity, the isochrones are manually approximated versions of the actual isochrones. Each isochrone is labeled on the right with forecast time in minutes, and on the left with the valid UTC time. For reference, observed damage tracks from the Greensburg tornado and the two subsequent tornadoes are indicated, and the location of Greensburg is denoted by the yellow star. Finally, purple circles mark the approximate location of the low-level mesocyclone associated with the Greensburg tornado at 15-min intervals from 0200 to 0300 UTC.

Citation: Monthly Weather Review 140, 2; 10.1175/MWR-D-11-00008.1

Fig. 8.
Fig. 8.

As in Fig. 7, but for 975 m AGL.

Citation: Monthly Weather Review 140, 2; 10.1175/MWR-D-11-00008.1

Since the model forecast time corresponding to the maximum vorticity at each grid point for each member is known, we can also derive ensemble average isochrones of forecast time, allowing for evaluation of the temporal evolution of the vortex across the ensemble. These isochrones, in 15-min intervals, are also overlaid in each panel of Figs. 7 and 8. Approximate locations of the observed low-level mesocyclone (derived from the 0.5° objectively analyzed radial velocity data), are also plotted at 15-min intervals starting at 0200 UTC (purple circles in Figs. 7 and 8).

There are broad similarities across the ensemble forecast experiments. All experiments produced a coherent swath of vorticity at the surface (Fig. 7) that roughly parallels the observed Greensburg tornado track, but with an axis that tended to lie to the east of the observed track. Near 975 m AGL, however (Fig. 8), the probabilistic vorticity swaths show better overlap with the observed tornado track. This suggests that the vortices in the forecasts were, in general, tilted toward the west with height between the surface and 1-km layer, roughly perpendicular to the direction of motion, which was toward the north-northeast. Visual inspection of individual forecast members (not shown) indicate that this east–west tilting was most pronounced in the 0–500-m AGL layer, apparently caused by the relatively strong rear-flank outflow in most of the forecast members (cf. Fig. 16d, discussed further in sections 3e and 4), which could reveal a bias in the microphysics scheme. (As discussed later, future work will examine this issue in more detail.) In contrast the observed tornado was seen to tilt to the north-northeast with height, roughly parallel to the direction of motion (LU08).

Also of note is the fact that in many of the experiments, particularly in the I0200 experiments (bottom row of Figs. 7 and 8), the initial location of the vortex appears displaced to the southeast of the beginning of the Greensburg tornado track. Visual inspection of both the raw and objectively analyzed radial velocity data (not shown) suggests that this may be a result of the EnKF analyses leading up to the 0200 UTC start time emphasizing the remnant circulation of the small tornado immediately prior to the genesis of the Greensburg tornado (tornado number 3 in LU08). Between about 0156 and 0200 UTC, this remnant circulation, which was located southeast of the broader developing Greensburg circulation (at about the location where the probabilistic swaths begin as seen in the bottom rows of Figs. 7 and 8) rapidly moved north-northwestward and was absorbed into the contracting and intensifying Greensburg circulation. The ensemble forecasts display a similar evolution, although not as quick as in the observations, explaining the initial southeast displacement of the forecast swath that later comes into better overlap with the observed Greensburg track.

Another feature that is apparent in many of the experiments, especially at the 975-m AGL level, is a split in the probabilistic swath from 1 main axis to 2. (This can be most easily seen beginning between +30 and +45 min in the individual panels of Fig. 8.) Between about 0238 and 0306 UTC, the observed Greensburg storm went through a cycling phase (LU08), with the occlusion of the Greensburg tornado and associated mesocyclone completed by ~0306 UTC (represented by the cyclonic curl near the end of the tornado track in the various figures). The new eastern mesocyclone went on to produce the even larger Trousdale tornado (track beginning immediately to the northeast of the Greensburg track overlaid in the various figures). The split in the probabilistic swaths suggests a similar cycling process in the model, at least in a qualitative sense, albeit apparently beginning somewhat earlier (~10–15 min). The cycling behavior is seen by examining the low-level (975 m AGL) reflectivity and vorticity fields for individual forecast members (Fig. 9). Several of the members exhibit double-hook echoes and/or two distinct mesocyclones at this time (especially members 10, 13, 15, 17, 26, and 28), albeit in different locations and states of evolution. In general, as can also be seen by examining the +45-min isochrone in Fig. 8e (corresponding to V0200I0200), the forecast positions of the storm and associated mesocyclone are too far to the northeast at this time. This aspect of the timing of the mesocyclones will be discussed further in section 3c, and evidence for cycling will be further discussed in the context of the I0215 and I0230 experiments in section 3e.

Fig. 9.
Fig. 9.

The 45-min forecast simulated reflectivity (color fill) at 975 m AGL, valid at 0245 UTC 5 May 2007 for 24 of the 30 members of forecast experiment V0200I0200. (Member numbers are labeled.) Vorticity is contoured (thick black lines) in 0.005 s−1 intervals, starting at 0.01 s−1. Vertical velocity at ~2 km AGL is also contoured (10 m s−1; thick purple lines). (top left) Objectively analyzed [to a regular 1-km grid using the Barnes (1964) scheme] observed reflectivity from the 0.5° KDDC base scan valid at ~0246 UTC. In each, as in Figs. 68, tornado tracks and the location of Greensburg, KS (yellow star), are overlaid. (bottom left) The scale in km is shown.

Citation: Monthly Weather Review 140, 2; 10.1175/MWR-D-11-00008.1

b. Impact of forecast lead time

While qualitatively similar, there are still significant systematic differences between the forecast experiments discussed so far. First, we examine the impact associated with changing the forecast initialization time. A general trend of higher probabilities and better overlap with the observed Greensburg tornado track is seen when examining probabilistic vorticity swaths as the forecast lead time decreases from 30 min (I0130 forecasts; top row of Figs. 7 and 8) to 15 min (I0145 forecasts; middle row of Figs. 7 and 8) to 0 min (I0200 forecasts; bottom row of Figs. 7 and 8). The I0130 forecast surface circulations also tend to dissipate before reaching Greensburg, especially for V0130I0130 and V0230I0130 (Figs. 7a,c), while the duration is improved for the I0145 and I0200 experiments. It may also be pointed out that the I0145 (and to a lesser extent, the I0130 forecasts) appear to forecast well (15–30 min in advance) the timing of the initial development of significant near-surface vorticity across the ensemble, coinciding closely with the beginning of the observed Greensburg tornado track (top and middle rows of Fig. 7). Finally, the timing of the vortex is better predicted in the 0200 UTC forecasts, with a general tendency for a slower-moving vortex and better agreement with the timing of the severe tornado damage in Greensburg (0245–0255 UTC). To see this, note that the location of the isochrone in each of the panels of Figs. 7 and 8 that represents the average time of maximum vorticity for the forecast valid time of 0245 UTC (+75, +60, and +45 min for the I0130, I0145, and I0200 experiments, respectively), is closer in time and space to the observed end of the Greensburg tornado for the I0200 forecasts (bottom row of Figs. 7 and 8). In summary, an improvement in the forecast of even features such as mesocyclones is seen when the lead time of the forecast is decreased. While this result is far from surprising, as it is well known that this is the case for the numerical prediction of large-scale weather systems, it is not obvious a priori that this result holds for the storm scale. Therefore, the improvement seen across the board for all the experiments considered is encouraging.

c. Impact of low-level wind profile

The differences between results obtained across the different initial low-level wind profiles are also significant. For the purpose of this discussion, we will focus on the I0200 experiments (bottom row of Figs. 7 and 8). There is a tendency in all experiments for the ensemble probabilistic swath of vortex tracks to be too far to the east relative to the observed track, with the closest agreement and most overlap at the surface level in experiment V0200I0200 (Fig. 7h), while the vortex swaths are farther east for both the V0130I0200 and V0230I0200 experiments (Figs. 7g,i). For experiment V0130I0200 the surface vortex swath (Fig. 7g) shows a bend toward the east with time, while in contrast, the observed tornado track displayed a bend to the north with time. A closer examination of the individual forecast members for V0130I0200 indicates that the storm rear-flank outflow (associated with an intensifying and growing cold pool) intensified with time in most members (a subset is shown in Fig. 10a), causing an overall eastward drift to the storm and associated surface mesocyclone track. In contrast the outflow in V0200I0200 remains weaker (Fig. 10b), allowing for the more northward track. Thus, it appears that the low-level wind profile has a systematic impact on the cold pool intensity across the ensemble, but an investigation of the reason for these differences is beyond the scope of this study. The overall shape of the vortex swath is significantly better predicted for the V0200I0200 and V0230I0200 experiments (Figs. 7h,i), and the swath shapes are similar. However, the average timing of the vortex in V0200I0200 is significantly better than in V0230I0200, with the latter roughly 10 km farther north at the 0245 UTC valid time (cf. the +45-min isochrones in Figs. 7h,i).

Fig. 10.
Fig. 10.

The 45-min forecast surface potential temperature perturbation (color fill) and reflectivity (black contours; 10-dBZ increment) for the first 12 members of (a) V0130I0200 and (b) V0200I0200.

Citation: Monthly Weather Review 140, 2; 10.1175/MWR-D-11-00008.1

d. Observation-space diagnostics for the forecast ensembles

While the above examination of probabilistic vortex swaths is useful to evaluate the merits of the ensemble forecasts in a qualitative sense, it is desirable to examine more objective metrics. Therefore, we examine observation-space diagnostic statistics applied to the forecast ensemble mean for each experiment (Figs. 4 and 5). For reflectivity, we used the same threshold of 15 dBZ as was used for the analysis cycle (cf. thin lines in Fig. 4). In both these figures, the simple, unmarked (solid, dashed, or dotted) lines refer to the I0130 forecast experiments, the lines marked with filled squares refer to the I0145 experiments, and the lines marked with filled circles refer to I0200 experiments. In the following discussion we will focus on the period between 0200 and 0300 UTC (i.e., the duration of the Greensburg tornado). At 0200 UTC the RMS innovation for reflectivity for the I0130 experiments (bold red simple lines in Fig. 4) are roughly 25–26 dBZ (depending on the low-level wind profile used). The corresponding values for the I0145 experiments (bold red square-marked lines) are ~22–25 dBZ, and that for the I0200 experiments (bold red circle-marked lines) are ~18–21 dBZ. In all forecast experiments, the RMS innovation generally increases with time, but levels off (and even decreases slightly for V0130I0200 and V0200I0200) between 0215 and 0245 UTC. The magnitudes during this time are ~27–31 dBZ for the I0130 experiments, ~25–28 dBZ for the I0145 experiments, and ~22–27 dBZ for the I0200 experiments, with the overall lowest RMS innovations displayed by the V0200I0200 experiment (bold red, dashed, circle-marked line). V0200I0200 was also the experiment with the best overall probabilistic vortex swath forecast (cf. Figs. 7h and 8h). The worst overall statistics are seen for the V0230I0130 experiment (bold red dotted line), which also had one of the poorest vorticity swath forecasts, both in terms of orientation and speed (cf. Figs. 7c and 8c). The analysis cycle RMS innovations extended out to 0300 UTC for reflectivity for the three VAD experiments (thin red lines) are also shown for reference. In general, these had magnitudes between ~13–19 dBZ for this period, again depending on the VAD profile, with the lowest magnitudes overall being seen for the I0130 VAD profile.

For the mean reflectivity innovation, significant differences are seen between the I0130 and I0145 experiments on one hand, and the I0200 experiment on the other. The I0130 and I0145 experiments have magnitudes increasing steadily with time for all three low-level profiles (bold black simple and square-marked lines in Fig. 4). In contrast the magnitudes level off around 10–13 dBZ for most of the forecast period in the I0200 experiments (bold black circle-marked lines), until shortly before 0245 UTC, when they begin to rapidly increase to ~19–21 dBZ by 0300 UTC. This behavior in the domain-wide reflectivity innovation statistics is also consistent with the overall better vortex track forecast in I0200 experiments as opposed to I0130 and I0145 experiments. Again, the thin black lines show the corresponding extended analysis cycle mean innovations for the different VAD profiles and generally vary from ~5–11 dBZ during the period.

Finally, the spread (and its rate of growth) among the forecasts for reflectivity (bold blue lines in Fig. 4) are very similar across the experiments with different VAD profiles for a given initial forecast time. The V0200 experiments (bold dashed blue lines in Fig. 4) show the largest spread, while the V0230 experiments (bold dotted blue lines) show the smallest, though again the difference is small when compared to the magnitudes of the spreads. The spread in the extended analysis cycle (thin blue lines) are very similar across the VAD profiles and slowly decrease from about 2.5 to 2 dBZ during the 0200–0300 UTC period.

Statistics for radial velocity (bold lines in Fig. 5) show less difference overall (relative to those for reflectivity) between the various initial low-level wind profiles for a given forecast initial time. This may be partly because no threshold was applied to the radial velocity magnitudes for the purpose of calculating the statistics, while a 15-dBZ threshold was used for the reflectivity statistics, and some of the assimilated radial velocity observations come from the clear-air regions outside of the storm. Thus, the domain-wide statistics would be less sensitive to small and localized differences that may exist in and around circulation features in the storm. Nevertheless, there is still a trend toward lower RMS innovations in the I0200 forecasts as opposed to the I0145 and I0130 forecasts (circle marked vs square-marked and simple bold lines in Fig. 5). The best overall statistics (especially at later forecast times) were displayed by the V0200 experiments (bold red dashed lines), while the V0230 experiments (bold red dotted lines) showed the worst. Again, experiment V0200I0200 displays the overall smallest magnitude of negative mean innovation (bold dashed circle-marked black line in Fig. 5). These results are consistent with the results from the reflectivity statistics.

In addition the magnitude and trend of the mean innovation for radial velocity for all experiments (bold black lines in Fig. 5) decreases steadily from ~−2 to ~−6 m s−1 during the forecast periods in all 9 experiments. Since the storm was located southeast to east of the KDDC radar during most of this interval, most of the innovation from verification of radial velocity was applied to the zonal wind component. In time–height cross sections of the mean innovation of radial velocity (not shown), it can be seen that most of this innovation occurred in the midlevels of the troposphere, centered around 5–6 km AGL. Recall that for all experiments, the overall track of the mesocyclone was too far east, a result consistent with this negative mean radial velocity innovation (i.e., the simulated storm had too large an eastward component of motion). We speculate that this errant motion could reveal a bias in the base-state wind profile (i.e., too much westerly midlevel momentum), or a bias in the storm-induced midlevel winds, or both. As a preliminary investigation of this issue, two additional EnKF analysis and forecast experiments were performed at a reduced horizontal grid spacing of 2 km, based on the original V0200I0200 experiment. One was otherwise identical to V0200I0200, while the other was identical except that the zonal component of the winds was reduced between 4 and 8 km AGL in the initial sounding. That is, a maximum subtraction of westerly velocity of 5 m s−1 was applied at 6 km, linearly reduced both above and below to zero subtraction at 4 and 8 km. The radial velocity observation-space diagnostic statistics for these two experiments are shown in Fig. 11 (analogous to Fig. 5). The experiment in which the midlevel winds were reduced displays a small improvement (smaller magnitude) in the mean innovation statistics, as well as a slight reduction in the RMS innovation, suggesting that at least part of the negative bias is indeed due to a bias in the environmental midlevel winds. Further investigation of this issue is left to a future study.

Fig. 11.
Fig. 11.

As in Fig. 5, but for V0200I0200 repeated at 2-km grid spacing (solid lines) and the corresponding experiment with reduced midlevel winds (dashed lines).

Citation: Monthly Weather Review 140, 2; 10.1175/MWR-D-11-00008.1

e. Additional negative lead-time forecast experiments

In addition to the above forecast experiments, two additional sets of ensemble forecast experiments for each of the VAD profiles were performed with start times at 0215 and 0230 UTC, respectively. Analogous to Figs. 7 and 8, probabilistic vorticity swaths for these experiments at 75 and 975 m AGL are shown in Figs. 12 and 13, respectively. These swaths are plotted out to 0330 UTC to cover the first 30 min of the Trousdale tornado period. Compared to the previous experiments, the I0215 and I0230 forecast experiments have a tendency to produce a surface vorticity swath (Fig. 12) that is even farther east of the observed Greensburg and Trousdale tracks, especially for the V0130 and V0230 profiles (Figs. 12a,c), due again to the anomalous east–west tilt in the surface–500-m AGL layer, which is exacerbated at these later start times (cf. Fig. 16d). However, the forecast swaths at 975 m AGL show much better agreement with the observed tornado tracks, particularly for the I0215 experiments (top row of Fig. 13). Continuing the trend of the earlier experiments, experiments V0200I0215 and V0200I0230 have superior track forecasts overall as compared to the corresponding V0130 and V0230 experiments, and even show enhanced probabilities overlaying or in close proximity to the observed Trousdale tornado track (Figs. 12b,e and 13b,e). The split in the probabilistic swath noted in the V0200I0145 and V0200I0200 experiments at 975 m AGL (Figs. 8e and 8h, respectively) is even more prominent in V0200I0215 and V0200I0230 (Figs. 13b and 13e, respectively). The analysis states at 0215 and 0230 UTC for the V0200 experiments appear to be capturing circulations that may be associated with the parent circulations of satellite tornadoes of the main Greensburg tornado near and after 0210 UTC (LU08) transitioning into the circulation that would eventually become the Trousdale mesocyclone. The subsequent ensemble forecasts from these later start times thus appear to have some skill in predicting the cycling between the Greensburg and Trousdale tornadoes, which is corroborated by examining individual members of V0200I0215 at the 0245 UTC valid time (Fig. 14) and comparing with those of V0200I0200 (Fig. 9).

Fig. 12.
Fig. 12.

As in Fig. 7, but for the (top) I0215 and (bottom) I0230 experiments, and swaths are integrated out to 0330 UTC.

Citation: Monthly Weather Review 140, 2; 10.1175/MWR-D-11-00008.1

Fig. 13.
Fig. 13.

As in Fig. 12, but for 975 m AGL.

Citation: Monthly Weather Review 140, 2; 10.1175/MWR-D-11-00008.1

Fig. 14.
Fig. 14.

As in Fig. 9, but for V0200I0215.

Citation: Monthly Weather Review 140, 2; 10.1175/MWR-D-11-00008.1

4. Summary and discussion

We have described a series of ensemble storm-scale numerical forecast experiments for the 4 May 2007 Greensburg, Kansas, tornadic supercell case using an EnKF assimilation system to produce the initial analyses for the forecast members. The experiments were designed to examine the impact of both forecast lead time and the strength of the environmental low-level winds on ensemble probabilistic vorticity “swaths,” which served as proxies for the tornado track. These probabilistic swaths were compared directly with the observed tornado track to qualitatively evaluate the forecasts. Most forecast experiments produced realistic vortex swaths that were qualitatively similar to the observed Greensburg mesocyclone/tornado track, and the “negative lead time” experiments showed evidence of predicting the next major mesocyclone of the series, which were associated with the Trousdale tornado.

Observation-space diagnostic statistics were calculated for both the EnKF analysis cycle (0030 to 0200 UTC) and for the various forecast ensembles (0130, 0145, or 0200 UTC to 0300 UTC), and served as objective estimates of forecast accuracy. These statistics agreed well with the more qualitative vortex swath evaluations; larger overlap between the vorticity swaths and observed damage track generally corresponded to better statistics. As the forecast lead time (relative to tornadogenesis time) decreased, these experiments produced significantly better forecasts of the location and timing of the surface circulation, as well as higher ensemble probabilities.

Using our simple EnKF system, this work explored whether a horizontally homogeneous initial environment could be used to capture the short-term forecast evolution of the storm and to assess the sensitivity to at least one aspect of the analyses and forecasts to that environment. Since the steady increase of low-level shear, associated with the intensifying low-level jet, was one of the more pertinent characteristics of the Greensburg storm’s environment (LU08; Bluestein 2009), we chose to examine the sensitivity of the ensemble forecast to changes in the environmental low-level shear. Three different VAD wind profiles valid at 0130, 0200, and 0230 UTC respectively, were “transplanted” from the nearby KVNX radar, providing “snapshots” of the low-level shear during the beginning of the long-track tornado phase of the Greensburg storm. In addition to the increasing low-level shear, the boundary layer thermodynamic profile was slowly stabilizing with the loss of solar heating (though this was not investigated in the current study). Taken together, these considerations suggest that the use of a single initial environmental sounding is problematic. SG10, in their similar ensemble forecast study (though using a 3DVAR-based data assimilation system, in contrast to the EnKF system used herein) of the same case, reached the same conclusion. In particular, when they used a more realistic, inhomogeneous mesoscale environment derived from a mesoscale numerical analysis and forecast system as initial and external boundary forcing for their storm-scale forecasts, a marked improvement in forecast accuracy was seen over similar storm-scale forecasts using a homogeneous, one-time single-sounding environment as is typical of idealized cloud and storm modeling studies. In future work, we will examine the sensitivity of forecast results when a time-varying, inhomogeneous mesoscale environment is included.

Not surprisingly changing the environmental low-level wind profile had a significant impact on the forecast ensembles. The most obvious changes were in the overall storm and associated mesocyclone track and speed, with the best results found for the V0200 low-level wind profile (cf. dashed red line in Fig. 2). In both the V0130 and V0230 experiments, the mesocyclone track was significantly farther east than the observed track (as well as the track in the V0200 experiments). Furthermore, the vortex translational motion was significantly faster in the V0230 experiments, which may be explained by the overall stronger low-level southerly momentum in the environmental profile in this experiment (cf. dotted red line in Fig. 2). In all experiments, there was a tendency for storm motions to be too fast. LU08 discussed the possibility of migrating bird contamination of the VAD wind profile, possibly yielding stronger-than-actual wind speeds within the low-level jet, but no attempt was made to account for this possible source of error in constructing the soundings for this study. Altogether, the 0200 UTC KVNX VAD profile produced the best results in regard to the vortex swath, as well as in regard to the observation-space diagnostic statistics. These results suggest that this profile was a better match to the actual wind profile, thus leading to a better forecast of storm motion and behavior across the ensemble. However, systematic model errors, especially those due to microphysics, may bias the results such that a better forecast is obtained for the wrong reasons. Some hint of this possibility can be seen when comparing the vortex swaths at 75 (Fig. 7) and 975 m AGL (Fig. 8). As discussed previously the low-level vortex tended to tilt toward the west with height. This is in contrast to the observed Greensburg tornado vortex, which tilted primarily toward the north-northeast with height during most of its life (although the tilt of the model vortex from 1 km AGL and up more closely followed the observed tilt). In the model solutions this tilt can be traced to strong westerly rear-flank outflow that pushed the near-surface vortex to the east of the low-level mesocyclone. Similar behavior has been noted in past simulations of tornadic storms when the microphysics scheme yields relatively strong cold pools (Snook and Xue 2008).

As an initial test of the microphysics sensitivity, experiment V0200I0200 was repeated using a single-moment microphysics scheme (Lin et al. 1983, referred to as LFO; Gilmore et al. 2004). The rain intercept parameters varied between 4.0 × 105 m−4 and 2.0 × 106 m−4, the graupel/hail intercept parameters between 1.0 × 105 and 4.0 × 105 m−4, and the bulk graupel/hail density between 400 and 800 kg m−3 across the 30 ensemble members. The range of rain intercept parameters are consistent with relatively large drops, while that for graupel/hail is consistent with moderate-sized hail, and thus overall might be expected to yield relatively weak cold pools (see, e.g., Gilmore et al. 2004; van den Heever and Cotton 2004; Snook and Xue 2008; Dawson et al. 2010). The ensemble surface vortex swath resulting from this experiment is shown in Fig. 15. Clearly the change in microphysics scheme from a double-moment to a single-moment scheme (even with some ensemble diversity in the rain and graupel intercept and graupel density parameters) results in a very different overall probabilistic vortex swath. The timing of the vortex is similar to the original V0200I0200 experiment, but the orientation of the track (which shows curvature opposite to that of the observed track), as well as the overall probabilities of vorticity greater than 0.01 s−1 are inferior to that in V0200I0200. The latter appears to be associated with generally both weaker and smaller vortices in the LFO experiment (not shown). The progressively eastward curvature of the swath appears to be associated with an initially weaker cold pool (for a typical ensemble member) than in the corresponding V0200I0200 experiment discussed in section 3 (cf. Figs. 16a,c). This weaker cold pool initially produces a more northward track. During this early period, the vertical tilt of the vortex is more consistent with the observed tilt (LU08), in contrast to the original V0200I0200 experiment. With time, however, the cold pools in the LFO experiment developed to the point where they were comparable to their counterparts in the original ZVD experiment, yielding a typically more eastward track to the low-level circulation toward the end of the forecast period (Fig. 16b), approaching Greensburg from the southwest, rather than the south, as in the observations and most of the ZVD-based forecasts.

Fig. 15.
Fig. 15.

As in Fig. 7h, but for the V0200I0200 experiment repeated with LFO microphysics.

Citation: Monthly Weather Review 140, 2; 10.1175/MWR-D-11-00008.1

Fig. 16.
Fig. 16.

Surface (75 m AGL) perturbation potential temperature (color fill; 1-K increment), reflectivity (black contours, 10-dBZ increment), vorticity (0.005 s−1 increment, starting at 0.01 s−1) at the surface (red contours) and 975 m AGL (green contours), and horizontal wind vectors (every third grid point, scale at bottom left) for member 16 of the LFO version of V0200I0200 at (a) the initial time of 0200 UTC and (b) 0230 UTC (30-min forecast). (c),(d), As in (a),(b), but for member 22 of the original ZVD V0200I0200 experiment.

Citation: Monthly Weather Review 140, 2; 10.1175/MWR-D-11-00008.1

These results are broadly consistent with recent work that demonstrated an improvement in convective storm morphology and behavior when using a multimoment bulk microphysics scheme (Milbrandt and Yau 2006; Morrison et al. 2009; Dawson et al. 2010). In this experiment, however, the range of parameters in the LFO scheme was tuned for relatively weak cold pools in contrast to the studies of Milbrandt and Yau (2006) and Dawson et al. (2010). In any case, the microphysics scheme has clear and substantial impacts on the low-level mesocyclones in the simulations. In future work, we plan to examine microphysics sensitivity to ensemble storm-scale forecasts in more detail. We will use a triple-moment version of the ZVD scheme used herein, and perform more comparisons with single-moment schemes of various configurations. We will vary the shape parameter of the gamma distribution among the forecast members, in a manner similar to what was done for the intercept parameter and bulk densities for the LFO scheme in the current study.

In the larger context of storm-scale analysis and prediction, the results of this study provide reason for cautious optimism. On the one hand, the substantial variability across the ensemble members underscores the sensitivity of substorm-scale features to small differences in the initial conditions. Furthermore, the choice of model microphysics scheme (and tunable parameters therein) remains a substantial source of uncertainty and strongly impacts the forecasts. On the other hand, the ensemble-based metrics used in this study show coherent trends and tendencies for a given large-scale environment (such as the aforementioned improvement in the forecast as the lead time decreases), underscoring the advantages of the ensemble approach to overcome some of these difficulties. Therefore the main results of this study broadly support the envisioned “Warn-on-Forecast” paradigm for the U.S. National Weather Service of Stensrud et al. (2009). An important aspect of this study is the evaluation of the ensemble forecasts using a “dimension reducing” approach where temporal and spatial evolution of a storm feature, in this case the mesocyclone, is evaluated in an ensemble probabilistic sense across a given time frame. This approach allows for compact, “at-a-glance” qualitative forecast analysis (and subsequent verification) that we propose can be readily adopted in an operational forecast setting (see Fig. 6 in Stensrud et al. 2009).

Acknowledgments

Daniel Dawson was supported by the National Research Council (NRC) Postdoctoral Fellowship, awarded at the National Severe Storms Laboratory. Louis Wicker and Edward Mansell were partially supported by NSF Grant ATM-0802717. Robin Tanamachi was supported by NSF Grants ATM-0637148 and ATM-0934307, and NOAA Grant NA08OAR4320904. David Dowell provided many helpful comments and code for the preparation of the VAD wind profiles. Some of the code used to plot the observation-space diagnostics was adapted from that provided by Therese Thompson. Conversations with Tom Hamill and David Stensrud were also very helpful. Jeff Hutton of the Dodge City, Kansas, National Weather Service Forecast Office kindly provided the shape files of the tornado tracks.

REFERENCES

  • Aksoy, A., , D. C. Dowell, , and C. Snyder, 2009: A multicase comparative assessment of the ensemble Kalman filter for assimilation of radar observations. Part I: Storm-scale analyses. Mon. Wea. Rev., 137, 18051824.

    • Search Google Scholar
    • Export Citation
  • Aksoy, A., , D. C. Dowell, , and C. Snyder, 2010: A multicase comparative assessment of the ensemble Kalman filter for assimilation of radar observations. Part II: Short-range ensemble forecasts. Mon. Wea. Rev., 138, 12731292.

    • Search Google Scholar
    • Export Citation
  • Anderson, J. L., 2009: Spatially and temporally varying adaptive covariance inflation for ensemble filters. Tellus, 61A, 7283.

  • Barnes, S. L., 1964: A technique for maximizing details in numerical weather map analysis. J. Appl. Meteor., 3, 396409.

  • Bluestein, H. B., 2009: The formation and early evolution of the Greensburg, Kansas, tornadic supercell on 4 May 2007. Wea. Forecasting, 24, 899920.

    • Search Google Scholar
    • Export Citation
  • Caya, A., , J. Sun, , and C. Snyder, 2005: A comparison between the 4D-VAR and the ensemble Kalman filter techniques for radar data assimilation. Mon. Wea. Rev., 133, 30813094.

    • Search Google Scholar
    • Export Citation
  • Coniglio, M. C., , D. J. Stensrud, , and L. J. Wicker, 2006: Effects of upper-level shear on the structure and maintenance of strong quasi-linear mesoscale convective systems. J. Atmos. Sci., 63, 12311252.

    • Search Google Scholar
    • Export Citation
  • Dawson, D. T., , M. Xue, , J. A. Milbrandt, , and M. K. Yau, 2010: Comparison of evaporation and cold pool development between single-moment and multimoment bulk microphysics schemes in idealized simulations of tornadic thunderstorms. Mon. Wea. Rev., 138, 11521171.

    • Search Google Scholar
    • Export Citation
  • Dowell, D. C., , and L. J. Wicker, 2009: Additive noise for storm-scale ensemble data assimilation. J. Atmos. Oceanic Technol., 26, 911927.

    • Search Google Scholar
    • Export Citation
  • Dowell, D. C., , F. Zhang, , L. J. Wicker, , C. Snyder, , and N. A. Crook, 2004: Wind and temperature retrievals in the 17 May 1981 Arcadia, Oklahoma, supercell: Ensemble Kalman filter experiments. Mon. Wea. Rev., 132, 19822005.

    • Search Google Scholar
    • Export Citation
  • Dowell, D. C., , L. J. Wicker, , and C. Snyder, 2011: Ensemble Kalman filter assimilation of radar observations of the 8 May 2003 Oklahoma City supercell: Influences of reflectivity observations on storm-scale analyses. Mon. Wea. Rev., 139, 272294.

    • Search Google Scholar
    • Export Citation
  • Gaspari, G., , and S. E. Cohn, 1999: Construction of correlation functions in two and three dimensions. Quart. J. Roy. Meteor. Soc., 125, 723757.

    • Search Google Scholar
    • Export Citation
  • Gilmore, M. S., , J. M. Straka, , and E. N. Rasmussen, 2004: Precipitation uncertainty due to variations in precipitation particle parameters within a simple microphysics scheme. Mon. Wea. Rev., 132, 26102627.

    • Search Google Scholar
    • Export Citation
  • Jung, Y., , M. Xue, , G. Zhang, , and J. M. Straka, 2008: Assimilation of simulated polarimetric radar data for a convective storm using the ensemble Kalman filter. Part II: Impact of polarimetric data on storm analysis. Mon. Wea. Rev., 136, 22462260.

    • Search Google Scholar
    • Export Citation
  • Lemon, L. R., , and M. Umscheid, 2008: The Greensburg, Kansas tornadic storm: A storm of extremes. Preprints, 24th Conf. on Severe Local Storms, Savannah, GA, Amer. Meteor. Soc., 2.4. [Available online at http://ams.confex.com/ams/pdfpapers/141811.pdf.]

    • Search Google Scholar
    • Export Citation
  • Lin, Y.-L., , R. D. Farley, , and H. D. Orville, 1983: Bulk parameterization of the snow field in a cloud model. J. Climate Appl. Meteor., 22, 10651092.

    • Search Google Scholar
    • Export Citation
  • Mansell, E. R., , C. L. Ziegler, , and E. C. Bruning, 2010: Simulated electrification of a small thunderstorm with two-moment bulk microphysics. J. Atmos. Sci., 67, 171194.

    • Search Google Scholar
    • Export Citation
  • Milbrandt, J. A., , and M. K. Yau, 2006: A multimoment bulk microphysics parameterization. Part IV: Sensitivity experiments. J. Atmos. Sci., 63, 31373159.

    • Search Google Scholar
    • Export Citation
  • Morrison, H., , G. Thompson, , and V. Tatarskii, 2009: Impact of cloud microphysics on the development of trailing stratiform precipitation in a simulated squall line: Comparison of one- and two-moment schemes. Mon. Wea. Rev., 137, 9911007.

    • Search Google Scholar
    • Export Citation
  • Oye, R., , C. Mueller, , and S. Smith, 1995: Software for radar translation, visualization, editing, and interpolation. Preprints, 27th Conf. on Radar Meteorology, Vail, CO, Amer. Meteor. Soc., 359–361.

    • Search Google Scholar
    • Export Citation
  • Snook, N., , and M. Xue, 2008: Effects of microphysical drop size distribution on tornadogenesis in supercell thunderstorms. Geophys. Res. Lett., 35, L24803, doi:10.1029/2008GL035866.

    • Search Google Scholar
    • Export Citation
  • Stensrud, D. J., , and J. Gao, 2010: Importance of horizontally inhomogeneous environmental initial conditions to ensemble storm-scale radar data assimilation and very short-range forecasts. Mon. Wea. Rev., 138, 12501272.

    • Search Google Scholar
    • Export Citation
  • Stensrud, D. J., and Coauthors, 2009: Convective-scale warn-on-forecast system: A vision for 2020. Bull. Amer. Meteor. Soc., 90, 14871499.

    • Search Google Scholar
    • Export Citation
  • Sun, J., , and N. A. Crook, 2001: Real-time low-level wind and temperature analysis using single WSR-88D data. Wea. Forecasting, 16, 117132.

    • Search Google Scholar
    • Export Citation
  • van den Heever, S. C., , and W. R. Cotton, 2004: The impact of hail size on simulated supercell storms. J. Atmos. Sci., 61, 15961609.

  • Wicker, L. J., , and R. B. Wilhelmson, 1995: Simulation and analysis of tornado development and decay within a three-dimensional supercell thunderstorm. J. Atmos. Sci., 52, 26752703.

    • Search Google Scholar
    • Export Citation
  • Ziegler, C. L., 1985: Retrieval of thermal and microphysical variables in observed convective storms. Part I: Model development and preliminary testing. J. Atmos. Sci., 42, 14871509.

    • Search Google Scholar
    • Export Citation
1

Because of the manner in which the model state for each member was written to disk (at 1-min intervals), the forecasts were actually initialized from the prior states in the EnKF analysis cycle. For this reason, and because the 2-min data interval was used, the effective start times for the individual forecasts were 1–2 min prior to the times given here, depending on the time of the previous posterior analysis.

2

A value of 0.01 s−1 is roughly equivalent to a mesocyclone having a rotational velocity of ~25 m s−1 and a diameter of 5 km.

Save