Tornadoes contain some of the strongest winds on earth, causing death and damage when impacting man-made and natural objects, such as buildings and trees1. Quantifying tornado windspeeds near the surface is critical to characterizing tornado hazards. In situ observations are exceedingly rare, and wind speeds are nearly always only inferred approximately, and likely with bias, from damage surveys2,3,4. Occasionally, proximate mobile radars map tornado windspeeds, usually >100 m above the ground (AGL), well above typical building heights4. The representativeness of these measurements to windspeeds closer to the surface is unknown4,5,6, severely limiting their interpretation and our understanding of near-surface tornado kinematics.

Here we show for the first time that the strongest supercell tornado windspeeds usually occur very near the ground and that radar observations at >100 m AGL usually are substantial underestimates of tornado intensity. Using rare, very low-level mobile radar observations of 73 different supercell tornadoes observed in the U.S. Great Plains we find that most such tornadoes exhibit a decrease in windspeed with increased altitude, with the strongest winds usually at the lowest observed levels. The median maximum tornado windspeed is 31% stronger at 15 m above the ground compared to radar measurements from higher altitudes. This is in contrast to the vertical distribution of the horizontal windspeeds in non-tornadic wind storms, in which windspeeds increase with height. Our results allow a more accurate quantification of tornado winds near the surface. We anticipate that these results will be used to update tornado hazard models6 and improve community risk assessments1. Furthermore, knowledge of tornado windspeed profiles, and, in particular constraining the usual height of the maximum windspeed to <15 m AGL, provide a basis to refine numerical and theoretical wind models with observational constraints, increasing the understanding of near-surface tornado vortex kinematics and applicability of model results7,8,9,10,11.



Damage-based inferences of near-surface tornado peak windspeeds using the Fujita2 and Enhanced Fujita Scales3 exhibit significant errors, biases4 and uncertainties, and do not reveal the vertical wind speed profile of tornadoes since nearly all tornado-impacted structures are very near the surface. Proximate radars12,13,14,15 have measured tornado wind speeds up to ~140 m s−1, but these measurements usually are from >50 m above ground level (AGL), well above building heights. Wind speeds in nearly all atmospheric phenomena are stronger aloft compared to near the ground. Measurements from aloft can be reduced to infer windspeeds at heights relevant to human and structural impacts, e.g., 10 m AGL, using well established methods6. However, a tornado’s winds might exhibit different behavior, the degree to which may depend on its structure and the turbulent characteristics of the boundary layer8,11, so it has been unclear how to relate radar measurements of tornado winds at >50 m to near-surface values.

Mobile radars deploying proximately to tornadoes have been able to resolve tornado structures16,17,18,19,20,21,22,23,24 and, in rare instances, wind observations at multiple levels within 4–100 m above radar level (ARL) (Tornadoes observed by DOWs nearly always occur in very flat terrain, so ARL and AGL are nearly identical.) have been obtained25,26,27. In rare, individual-case analyses of unknown generality, proximate radar data at multiple levels have revealed that wind speeds were strongest very near the ground27,28, suggesting that even radar measurements from only several tens of meters ARL may be underestimates of true intensity.

Characterizing how these winds vary with height is necessary for understanding how winds observed aloft map to near-surface, surface-, and structure-level winds. The Doppler on Wheels (DOWs)12,15 mobile radars have collected data in over 250 tornadoes spanning over three decades, and a broad range of intensities, vortex structures, and sizes23, providing an expansive dataset through which near-surface tornado wind characteristics can be assessed (Fig. 1). In 73 of these tornadoes, DOW data were obtained at multiple levels, at least one of which was below 100 m ARL. These were used to construct vertical profiles of quasi-horizontal wind speed, allowing for the quantification of the general variation of tornado wind speeds as a function of height, and exploration of how these profiles vary from tornado to tornado, and with tornado intensity.

Fig. 1: DOWs observe tornado wind speeds.
figure 1

A DOW observing a tornado. B Map of locations of DOW-observed tornadoes 1995-2006. C Example of DOW Vd fields observed at three levels in a tornado. Blue/red = in/outbound. Note that the peak wind speed observed at each level is not necessarily immediately above the peak observed in the next lower level.

Doppler on wheels tornado windspeed profiles

The maximum wind speed at each observation height was extracted from the DOW data following standardized procedures4 (see methods section). Only time periods which had nearly contemporaneous measurements from at least two heights, at least one of which was <100 m ARL, were used for this analysis. At the time of peak tornado intensity, DOW-measured Doppler wind velocity (Vd) maps (Radars measure nearly instantaneous “Doppler Velocities” (Vd)29. Radar-measured wind speeds in this analysis represent a very short period, <<1 s, spatially averaged, wind measurements, not average wind speeds, 1–3 s, or other averaging period wind gusts as measured by stationary anemometers. Critically, damage from tornadoes is believed to be caused by short period wind gusts2,3,6.) were used to determine tornado center locations, velocity difference across the tornado (DV), propagation velocity (Vp), calculate maximum ground-relative wind speeds in each radar slice through the tornado (Vg), and the maximum Vg observed below 100 m AGL (Vgmax). In order to compare tornadoes with different intensities, Vg were normalized by Vgmax to obtain |Vg|. Vertical profiles of |Vg| and d|Vg|/dz, where z is height AGL/ARL, were extracted for 73 tornadoes.

Analysis of all |Vg| profiles suggest that there is considerable variation in wind speed from the lowest observed levels to about 100 m ARL (Fig. 2A), and many of the profiles exhibit values near 1.0 near 100 m ARL. However, the height of the lowest observation level in each profile is different, ranging from 6 m to 98 m ARL, potentially skewing this result. Nearly all profiles that contained observations both at low levels (<15 m ARL) and at high levels (between 80-100 m ARL) exhibited the strongest winds near the lowest level, as indicated by the color coding of the profiles. Thus, the profiles containing data only near 100 m ARL likely did not sample the stronger wind speeds below, resulting in spuriously high normalized values aloft.

Fig. 2: Normalized windspeed |Vg| profiles observed by DOWs.
figure 2

Vg normalized by Vgmax observed <100 m AGL. Dashed lines: median profile from the 2012 Russell, KS tornado27. A All profiles containing observations <100 m ARL. Line colors are height of Vgmax in each profile. B Only profiles containing observations <15 m ARL. Line colors are DV observed at lowest level in each profile.

In order to mitigate the effects of this sampling bias, data were filtered to include only those profiles that began at low levels, <15 m ARL, resulting in profiles from nine tornadoes (Fig. 2B). There was considerable case-by-case variation in the profiles. Most of the profiles exhibit the strongest |Vg| at the lowest levels. However, there are two exceptions, each associated with very weak tornadoes, with DV < 45 m s−1, typical of vortices described as “marginal tornadoes”15. These exhibit |Vg| which increase with height in the lowest observed levels in the manner of non-tornadic windstorms. It is not clear if strengthening windspeeds with height is a property of marginal tornadoes since the 3rd profile associated with a marginal tornado does exhibit decreasing |Vg| with height. Profiles from some slightly more intense tornadoes, with 50 < DV < 60 m s−1, exhibit |Vg| that weakens with increasing height.

Analysis of dVg/dz allows examination of the vertical dependence of wind speeds and derivation of generalized vertical Vg profiles. Due to possible contamination of the lowest elevation radar scans by scattering from foliage, structures, and the ground29,30, which can spuriously reduce Vd, d|Vg|/dz was calculated using the 2nd and 3rd lowest height observations in each profile (Fig. 3). This revealed d|Vg|/dz <0 from 15–90 m ARL. Median d|Vg|/dz was approximately −0.005 (variance = 0.00007) from 17–35 m ARL, less negative (−0.002 to −0.0014, variance = 0.00002) from 40–96 m ARL, and near zero (variance = 0.00001) above 96 m ARL. Integrating these median profile \(\int d\left|{V}_{g}\right|/{dz\; dz}\) resulted in generalized DOW-observed tornado windspeed profiles (Fig. 4). These reveal that median |Vg| decreases rapidly from 15 to 40 m AGL, then more slowly from 40 to 100 m AGL, and then very slowly from 100 to 140 m AGL.

Fig. 3: Median and all samples of slope of DOW-observed tornado windspeeds d|Vg|/dz|.
figure 3

Normalizations based on Vg observed at lowest level in each profile. Icon colors represent tornado vortex intensity, DV, at lowest level in each profile. Hollow triangles represent values with dz >50 m. Hollow squares represent values with nearly identical radar beam elevations which are excluded from medians. Black circles are seven-sample medians of d|Vg|/dz| plotted at the mean z of the seven samples. Gray boxes outline d|Vg|/dz| and z spread of all values in each median.

Fig. 4: DOW-observed tornado windspeed profile.
figure 4

Median |Vg| profile based on DOW observations, and example of profile calculated using measured Vg(140 m AGL) = 40 m s−1 implying Vg(15 m AGL) = 52 m s−1.


Theoretical, numerical, and laboratory modeling studies suggest that the strongest tangential winds occur near the top of the tornado boundary layer7,9. Since boundary layer depth may vary as a function of surface roughness and/or tornado morphology, the height at which the maximum tangential winds occur is not expected to be the same for every tornado. Numerical modeling10 revealed the height of the maximum average tangential wind speed to range from ~30–70 m AGL, depending on model configuration, with average speed values 15% lower at 10 m AGL. Model analysis examined 10 m AGL wind gusts but only calculated the average windspeeds at 30–70 m, and the vertical dependency of gusts is not discussed. Critically, since radars measure wind gusts, our current radar-based windspeed profile analysis directly addresses measured wind gusts as a function of height.

Because  tornado winds are transient at any given point, damage resulting from tornadoes, and tornado intensity ratings, are associated with the effects of short period peak wind gusts on natural and man-made structures2,3,6. Our results show that Vg at 15 m AGL average 31% stronger than Vg observed by radars at 100–140 m AGL, but with considerable case-by-case variation. The amount of underestimation depends on the minimum observation height. The profile in Fig. 4 can be used to calculate impactful near-surface (15 m AGL) tornado wind gusts from radar observations obtained aloft. For example, if radar indicated Vg(140 m AGL) = 40 m s−1, low-level tornado Vg can be calculated as Vg(15 m ARL) = Vg(140 m AGL)/(1–0.235) = 52 m s−1.

The wind speed profile reported in Fig. 4 is consistent with, but slightly steeper than, those previously reported27,28. Critically, the individual profiles, representing snapshots of Vg, reveal considerable variation among tornadoes and it has been demonstrated27,28 that there was considerable temporal variability of profiles during individual tornadoes. Therefore, using the current generalized results to extrapolate individual radar observations to infer the near-surface intensity of individual tornadoes should be conducted with caution. DOW observations at different levels are not truly simultaneous, and peak Vg in each radar scan are not necessarily immediately above peak Vg measured in immediately lower scans. Therefore, none of the profiles illustrated in Fig. 2 represent instantaneously-measured truly vertical profiles of Vg in precisely vertical columns.

These results have substantial implications for tornado hazard analysis and for validation and refinement of the wind-damage relationships that inform intensity determinations. However, we do not suggest that average structure-impacting tornado wind speeds are necessarily 31% (1/(1–0.235) = 1.31) stronger than measured by radars at 100–140 m ARL or those analyzed statistically4. Many, if not most, structures impacted by tornadoes suffer damage caused by winds at heights <15 m AGL. There are not yet sufficient tornado wind speed observations from <15 m ARL to extend the profile shown in Fig. 4 to the standard meteorological observation height of 10 m AGL. Residential structures common in tornado-prone areas typically extend from the ground to only about 3 m (single-story manufactured homes), 5 m (single-story house) or 8 m (two-story houses), with first stories centered at about 2 m AGL. Critically, the strength of winds in the 0–15 m AGL layer in the presence of man-made structures and trees are nearly impossible to sample by radar due to blockage from those same objects. Therefore, very near ground tornado Vg in built environments remain rarely sampled. Case study results suggest that maximum intensity winds may sometimes occur as low as 5 m AGL, but the radar data used in those studies were obtained over open land, not near or between houses and trees27. Observations very near the ground by anemometers at 3 m AGL27,31 and at 1 m AGL32 suggest that peak wind gusts (Vg) in the 1–3 m AGL layer are less than or comparable to those measured by radars above the anemometers, but the impact of the built environment on these observations in unknown.


Navigation of DOW data

DOW radars have collected data in tornadoes since 1995. Vd data from several individual case studies, typically of intense or otherwise unusual tornadoes, have been navigated, processed, and analyzed in detail prior to this study14,16,18,21,23,27,28,31,32,33,34,35, and a statistical analysis of ~100 tornadoes observed from 1995-2006 has been conducted4. In addition, dual-Doppler36 and Ground-Based Velocity Track Display (GBVTD)35,37 analyses have been conducted for some tornadoes using navigated, processed and analyzed data. For this study, analyzed data from these previous studies, as well as from all additional tornadoes observed by DOWs during 2007–2009, are included.

Navigation of DOW data was achieved through a combination of Global Positioning System (GPS) locations and mapping of ground clutter targets. In most cases, data are from stationary leveled deployments when antenna pitch and roll were less than 0.2° as measured by precision bubble levels inside and near the antenna pedestals. During 1995–1999, and for a few cases during 2001-2005, data collected while DOWs were mobile were included. Vd were filtered using signal quality measures, such as signal-to-noise-ratio and returned power, and dealiased, as needed; data from non-meteorological targets were removed. No additional filtering was applied and Vd was analyzed in the native radar grid. When radar beam elevations are less than the radar beam width (B), a Gaussian-weighted correction raises the effective elevation angle such that they asymptote at 0.3° to account for terrain, tree, and structural blockage of the bottom portion of beams, with the effective elevation angle of a beam nominally pointing at 0.0° elevation being adjusted to 0.34°.

In this analysis, radar data are characterized by height ARL. When the ground height under a tornado is different than that under the DOW, AGL and ARL values will be different. This is rarely significant in the relatively flat terrain of the Plains, and at the close ranges over which most DOW tornado data are collected. This is especially true for the data included in this analysis, in which it was required that at least one observation be below 100 m ARL. Also, given the typically short ranges, non-linear and other anomalous propagation of radar beams is neglected, except for the adjustment for blockage of the lower portion of ground-skimming beams discussed above.

The GURU software suite, developed by the Center for Severe Weather Research, was used to analyze several thousand cross-sections through candidate tornado vortices, most of which had been processed4, with tornadoes observed during 2007–2009 added. For each cross-section containing data in a candidate tornado vortex, the authors enter a first-guess center location and outline a ~2 km region enclosing the vortex. These are then used by GURU to calculate refined center locations, tornado diameter (Xd), and Vp. Automatically-determined center positions and Xd are reviewed by an expert team and refined as necessary. Sometimes this was required for complex vortices, containing sub-vortices or multiple wind speed maxima. In some cases, the determination of whether vortices were independent tornadoes or sub-vortices within multiple vortex mesocyclones16,18,23,34 was subjective. Centers, Vd fields, and Vp were used to automatically identify cross-sections containing Vgmax, which were then used in this study.

Tornadoes spawned from quasi-linear convective systems and waterspouts were excluded from this study. Vortices were characterized as tornadoes when the maximum velocity difference across them (DV) was > = 40 m s−1 within a < 2 km region4,23, were associated with a mesocyclone and/or hook echo of a supercell thunderstorm and were not close to and/or associated with other tornadoes, thus excluding vortices along proximate gust fronts (“gustnadoes”) and nearby anti-cyclonic vortices23. Vortices not observed below 100 m ARL were also rejected.

Calculation of Tornado metrics

The time of maximum intensity (defined as peak Vgmax) and various metrics of the wind field, including Vp, Vgmax, were calculated for each tornado. Some DOW-observed tornadoes may not have been observed at the time of maximum actual intensity. In some cases, Vp could not be determined from DOW measurements. This could occur due to irregular cross-section scheduling or the existence of only a single cross-section through the tornado. When needed, Vp was calculated from mesocyclone motion as measured by operational NWS WSR-88D radars. Vgmax at each observed level in a tornado was not required to be immediately above the next lower level Vgmax.

Adjustment for observation resolution and unobserved components of wind velocity

Following standardized procedures4,6 our analysis adjusts Vd for the effects of coarse spatial sampling and inferred unobserved Vgmax due to Vp. Vd are corrected for aspect ratio sampling errors4,38 by multiplying by 1/(1–0.48(B/Xd)), where B is beamwidth, capped at 1.086, equivalent to what would be applied for B/Xd > 6, to arrive at a temporary product, adjusted velocity (Vda). Vd is a measure of only the towards/away component of the full wind velocity vector. For vortices propagating at any appreciable angle to the radar beams, the unobserved component of the wind velocity vector, Vp sin(θ) where θ is the angle between Vp and the beam pointing angle, is significant. This is added to Vda to calculate Vgmax.

Eligible vertical profiles

Only data from tornadoes for which there were at least one DOW observation below 100 m ARL were included in this analysis. In some instances, observations were obtained at nearly identical levels, so that differences in height were much less than a radar beam width. Data in Fig. 4, for example, are flagged where the difference in radar beam elevation is « B, specifically 0.2°. Profiles were calculated for sweeps collected during a radar volume (usually a sequence of radar slices at increasing beam elevation) containing the time of peak Vgmax. However, in a few cases, there was no eligible profile in this volume (e.g., the volume was degenerate, containing only one slice). In these cases, the profile was extracted using the previous or following eligible volume, if the time of the low-level slice was within 60 s of the slice containing Vgmax.

Calculation of median statistics and profiles

Median d|Vg|/dz statistics were calculated after eliminating values calculated from radar scans with differences in elevation of less than 0.2° or difference in height of 5 m or less, since these resulted in unreliable values of vertical derivatives since differences in Vd are likely dominated by temporal evolution. Medians were calculated from successively increasing mean height groups of seven d|Vg|/dz values and plotted at the mean height of the values contributing to the medians. Profiles were calculated by piece-wise integrating the median d|Vg|/dz profiles from 17-139 m AGL.