The strongest winds in tornadoes are very near the ground

Tornadoes produce extreme winds that cause significant damage and loss of life, yet near-surface wind speeds are poorly quantified because direct observations close to the ground are rare. Mobile radars typically sample winds well above building height (>100 m AGL), raising questions about how representative those measurements are of the winds that impact structures and people. This study addresses the key research question: how do tornado wind speeds vary with height very near the ground, and do radar measurements at >100 m AGL underestimate true near-surface intensity? Using rare, very low-level mobile radar data from 73 supercell tornadoes in the U.S. Great Plains, the authors show that the strongest tornado winds generally occur very close to the ground, with median maximum wind speeds 31% stronger at 15 m AGL than at higher sampled altitudes. This behavior contrasts with non-tornadic windstorms, where wind speed typically increases with height, and has important ramifications for hazard characterization and model development.

Damage surveys using the Fujita and Enhanced Fujita (EF) scales can be biased and uncertain and do not constrain vertical wind profiles because structures sampled are near the surface. Proximate mobile radars have recorded tornado winds up to about 140 m s−1 but typically at heights >50–100 m AGL, above building tops. In most atmospheric phenomena, wind speed increases with height, allowing extrapolation downward, but tornado boundary layer behavior may differ and depend on vortex structure and turbulence. Prior case studies with multi-level mobile radar data have occasionally shown strongest winds near the ground, suggesting that even tens-of-meters AGL radar measurements may underestimate intensity. The Doppler on Wheels (DOW) program has amassed observations for >250 tornadoes across three decades, providing a unique dataset to examine near-surface tornado wind characteristics.

Dataset and selection: DOW mobile radars collected Doppler velocity (Va) data in >250 supercell tornadoes; 73 tornadoes with multi-level observations including at least one level <100 m above radar level (ARL; approximately AGL in flat terrain) were used to build vertical profiles of quasi-horizontal wind speed. Waterspouts and quasi-linear convective system tornadoes were excluded. Vortices were classified as tornadoes if the velocity difference across the vortex (ΔV) was ≥40 m s−1 within <2 km, associated with a supercell mesocyclone and/or hook echo, and not part of nearby non-tornadic vortices. Data navigation and quality control: Radar data were navigated using GPS positions and mapped ground clutter. Most data were from stationary, leveled deployments (antenna pitch/roll <0.2°). Va was filtered using signal quality metrics, dealiased, and analyzed on the native radar grid. For near-ground beams, an effective elevation adjustment (minimum 0.34° for nominal 0.0°) accounted for blockage of the lower beam by terrain/structures. ARL heights were used as the vertical coordinate; ARL≈AGL given flat Plains terrain and close ranges. Analysis steps: At times near peak tornado intensity, Va fields were used to determine tornado center, DV (velocity difference across the tornado), propagation velocity Vp, and to compute ground-relative maximum wind speed V in each radar slice. The maximum V observed below 100 m AGL was denoted Vgmax. To compare tornadoes of different intensities, V at each height was normalized by Vgmax to form V profiles. Only profiles with nearly contemporaneous measurements from at least two heights (one ≤100 m) were used. To mitigate contamination in the lowest scan by ground/foliage backscatter that may spuriously reduce Va, vertical gradients d|V|/dz were computed using the 2nd and 3rd lowest height observations in each profile. Profiles were also filtered to exclude nearly identical elevation angles (difference <0.2°) or small height differences (≤5 m) when computing derivatives, to reduce temporal evolution effects. Resolution and unobserved component corrections: Va was adjusted for aspect ratio sampling errors via multiplication by 1/(1−0.48(B/Xa)), capped at 1.086, where B is beamwidth and Xa is tornado diameter, yielding adjusted velocity Vda. Because Va measures only the along-beam component, the unobserved component due to tornado propagation was added: Vgmax = Vda + Vp sin(θ), where θ is the angle between Vp and beam pointing. Profile construction and statistics: Vertical profiles were extracted for each tornado during the radar volume containing peak Vgmax (or within ±60 s if needed). A subset analysis included only profiles beginning <15 m ARL (nine tornadoes) to reduce sampling bias from higher-only lowest levels. Median d|V|/dz statistics were generated from groups of seven samples binned by mean height and plotted at mean z, then piecewise integrated from 17–139 m AGL to derive generalized median windspeed profiles. Supplementary details: The GURU software suite was used to semi-automatically determine centers, Xa, and Vmax, with expert review for complex multi-vortex cases.

• Across 73 supercell tornadoes with multi-level DOW observations, most vertical profiles showed decreasing wind speed with height from the lowest observed levels up to about 100 m AGL, indicating strongest winds very near the ground. - Sampling bias assessment showed that profiles containing both very low levels (<15 m ARL) and higher levels (80–100 m) almost always had maximum winds at the lowest levels; profiles lacking very low-level data tended to overstate normalized speeds aloft. - In the subset of nine profiles starting <15 m ARL, most showed peak |V| at the lowest level; two exceptions (marginal tornadoes with DV <45 m s−1) exhibited increasing |V| with height at the lowest levels, akin to non-tornadic windstorms. - The generalized vertical gradient of windspeed, d|V|/dz, was predominantly negative from 15–90 m ARL: approximately −0.005 (variance 0.00007) from 17–35 m, −0.002 to −0.0014 (variance 0.00002) from 40–96 m, and near zero (variance 0.00001) above 96 m ARL. - Integrating median d|V|/dz yielded a generalized DOW-observed windspeed profile: median |V| decreases rapidly between 15–40 m AGL, more slowly between 40–100 m, and very slowly between 100–140 m. - Quantitatively, the median maximum tornado windspeed is 31% stronger at 15 m AGL compared to measurements at higher altitudes (e.g., a measured |V| of 40 m s−1 at 140 m AGL implies ~52 m s−1 at 15 m AGL). - Overall, radar observations at 100 m AGL typically substantially underestimate true near-surface tornado intensity.

The findings provide multi-event observational evidence that, contrary to the typical atmospheric boundary layer where wind speeds increase with height, tornado wind speeds are generally strongest very near the ground. This supports and refines theoretical and numerical expectations that peak tangential winds occur near the top of the tornado boundary layer, whose depth varies with surface roughness and vortex morphology. Large-eddy simulations have placed the height of maximum average tangential winds around 30–70 m AGL; the DOW-based median profiles indicate strong near-ground maxima and a rapid decrease from ~15 to ~40 m AGL, constraining model representations of near-surface vortex structure. From a hazard perspective, because many structures experience winds below 15 m AGL, reliance on radar-derived winds >100 m AGL likely underestimates damaging near-surface winds. The results can improve tornado hazard models, community risk assessments, and inform building design standards. However, substantial case-to-case and temporal variability exists, and vertical profiles assembled from sequential scans are not instantaneous vertical columns. Therefore, extrapolating individual radar observations to infer near-surface winds for a given tornado should be done with caution.

Using rare, very low-level mobile radar observations from 73 supercell tornadoes, the study demonstrates that the strongest tornado winds typically occur very near the ground, and that radar measurements at >100 m AGL generally underestimate near-surface intensity by about 31% at 15 m AGL on median. The derived generalized vertical profiles and gradients offer observational constraints for refining theoretical and numerical models and for updating tornado hazard and risk assessment methodologies. Future work should expand observations below 15 m AGL, especially in built environments where radar sampling is limited, obtain more truly simultaneous multi-level measurements, and integrate these constraints into high-resolution simulations and damage-wind relationships to improve intensity estimation and building codes.

• Vertical profiles are constructed from sequential radar scans; peaks at different levels are not necessarily sampled simultaneously or aligned vertically, so profiles are not true instantaneous vertical columns. - There is considerable tornado-to-tornado and temporal variability; generalized profiles may not represent individual events accurately, cautioning against simple extrapolation for a single case. - Very near-ground observations (<15 m AGL) remain sparse due to radar beam blockage by terrain, trees, and structures; the 0–15 m layer, most relevant to many structures, is poorly sampled. - Lowest elevation scans may be contaminated by non-meteorological scattering, potentially reducing Va; measures were taken to mitigate this in gradient calculations, but uncertainty remains. - Sample selection excludes non-supercell tornadoes (e.g., QLCS, waterspouts) and vortices with ΔV <40 m s−1, limiting generalizability to supercell tornadoes meeting inclusion criteria. - AGL versus ARL differences are assumed negligible in flat terrain, but small mismatches can occur. - Some cases lacked directly measured propagation velocities (Vp) and required estimates from WSR-88D data, introducing additional uncertainty.