Revised manuscript submitted to
Corresponding Author:Harold E. Brooks, NOAA/NSSL, Norman, OK, 73069, USA.
Reports
of tornadoes, broken down by damage, from seven countries have been examined.In
particular, the long-term relatively high-quality dataset from the United
States is used to develop distributions which indicate that the number
of tornadoes decreases log-linearly with increasing F-scale.Two
distinct distributions, one apparently associated with supercell tornadogenesis
processes and the other with non-supercell processes, are found in both
the United States data and in other countries.The
similarity of the distribution in the United States prior to the 1950s,
when an official, organized collection effort began, and the French record,
suggests that only 15% of French tornadoes are being reported currently.In
addition, we can use the simple statistical distributions to estimate the
return period of violent tornadoes in France (approximately one every 5-10
years) and the United Kingdom (approximately one every 250-300 years).
1.Introduction
Tornadoes
have been observed on all continents except Antarctica.Prior
to the 20th century, most of the reports were anecdotal in nature
and systematic evaluation of data was extremely rare.In
the 1960s, studies of the damage associated with individual tornadoes in
the United States led to the development of the Fujita damage scale (Fujita
1971), a method to classify tornadoes based on the maximum level of damage.Although
Fujita estimated windspeeds associated with the different levels of damage[1],
in practice, the scale only provides information on damage.Some
of the difficulties this leads to have been discussed by Doswell and Burgess
(1988) and Grazulis (1993).
All
tornadoes in the United Sates have been assigned F-scale values since 1973.It
is possible to estimate F-scale values for tornadoes prior to that date
if sufficient documentation (e.g., newspaper accounts, photographs) is
available.This has been done by
Tecson et al. (1979) for all United States tornadoes back to 1916 and Grazulis
(1993) for United States strong and violent tornadoes as far back as 1640,
and by other researchers for a number of other countries (e.g., Paul 1999
for France).Information from many
of those countries can be found in this volume (e.g., Reynolds 2000 for
the United Kingdom).
The
data from the United Kingdom were originally given according to the TORRO
scale V = 2.365 (T+4)1.5 (Meaden 1976), where V is the velocity
in m s-1 and T is the T-scale number.From
the two velocity-based definitions, an F value can be found from a T value
by F = .52T + .08.For simplicity,
we have approximated this by F = .5T and truncated values to the nearest
integer, so that T0 and T1 correspond to F0, T2 and T3 to F1, etc., in
the manner of Elsom and Meaden (1982).The
two scales differ in their assignment of velocity values by 6%.Given
the inherent difficulties in assigning F and T values, this is a valid
approximation.In operational practice,
the two scales differ only in the details of the defining equation.We
have chosen to use the scale that is in more widespread use.In
addition, the problems associated with damage surveys and uncertainties
associated with estimating windspeed from observed damage make highly precise
assignments dubious.
Several
fundamental questions can be addressed by looking at the results from different
countries.Chief among them is:What
similarities and differences can be found in the distribution of tornadoes
by damage around the world?If tornadoes
have similar characteristics in different parts of the world, then it may
be possible to use data from areas of relatively high frequency and quality
of reports to make estimates of threats in other parts of the world.In
addition, it may be possible to develop estimates of the degree of underreporting
of tornadoes in different countries.
2.The
United States record
The
longest tornado record collected by an official national agency at the
time of the events occurs is that from the United States, beginning in
1953 with the creation of the National Severe Storms Project.Through
efforts at researching previous events, a relatively high quality record
was extended back to 1916 by Tecson et al. (1977).The
number of low F-scale tornadoes has increased dramatically since the 1920s
(Fig. 1), with approximately 700 F0 tornadoes reported
annually in the 1990s compared to only 7 in the 1920s.
Although
the mean annual number of F3 tornadoes has ranged from about 20 to 60 depending
on the decade, it is important to note that there is no long-term secular
increase across the length of the record.Indeed,
the maximum values (about 50-60 F3 tornadoes per year) are found in 1950-1979,
with the remainder of the record averaging between 25-35 F3 tornadoes per
year.Changes in the total number
of tornado reports in the United States are almost entirely the result
of changes in the low-end reports.It
is also of importance to note that the largest changes between any two
decades occurred between the 1940s and 1950s, at a time when it became
the responsibility of a governmental weather forecasting and research agency
to collect the data when the events occurred.The
mean annual number of F0-F2 tornadoes reported in the United States increased
from about 150 to 450 from the 1940s to the 1950s.
3.Distributions
by F-scale
A
feature of interest in the United States record is that the distribution
of tornadoes by F-scale has been approaching log-linear (see Fig.
1).This distribution is consistent
with standard statistical distributions of rare events, such as the Gumble
distribution, that show a nearly log-linear decline as the intensity of
the event increases and the frequency at which it is observed decreases.This
log-linear behavior has been seen in other weather records, such as extreme
hourly precipitation amounts (Brooks and Stensrud 2000).
It
is important to consider sources of error in the distributions.In
general, there are at least four sources of error in the collection of
data and classification of tornadoes by damage scale.First
of all, there are times when no or very few reports at all are collected.There
is evidence of this in periods such as the 1940s in France and Germany,
as well as the mid-nineteenth century in Germany (Fig. 2).Similarly,
from 1905-1995, there were five tornadoes reported in the Eastern Cape
province of South Africa.After two
well-publicized tornadoes elsewhere in South Africa in the early summer
of 1998-1999, there were ten tornadoes reported in the province the rest
of that season (E. de Coning, personal communication.)It
seems likely that the total of five reports in the earlier 90 years is
an underestimate of the true number of events.In
those cases, tornadoes at all F-scales are missed.Second,
low F-scale tornadoes are likely to be missed in the reporting because
they typically have short lifetimes and path lengths.Third,
given that the assignment of an F-scale rating depends upon adequate structures
being present to be damaged, it is likely that the number of tornadoes
at the highest F-scales is underestimated.For
example, if there are no structures present in the path of a tornado, it
is impossible, in practice, to rate it as a violent tornado.In
general, this kind of problem moves tornadoes from higher F-scale values
towards lower F-scale values.Finally,
there may be random errors in the assignment of F-scale.Interpretation
of the exact cause and extent of damage is an extremely difficult task
and uncertainties in knowledge of the construction of a building or the
debris that struck a building lead to questions about the assignment that
are often hard to answer.It is
the experience of the authors that damage surveyors may often disagree
over the value to assign to event by one F-scale.
The
first type of error does not affect the probability distribution function
of tornado damage classification, but obviously it will affect the total
number of tornadoes.The second
and third types of errors �move� tornadoes out of the ends of the distribution
towards the middle classes.If there
truly are more F0 tornadoes than higher classes, random errors actually
move tornadoes preferentially out of F0 toward higher F-values.A
simple hypothetical example illustrates this.Consider
the case where there are 1000 �true� F0 tornadoes, 400 F1 tornadoes, and
160 F2 tornadoes.If 10% of tornadoes
are misassigned one class too high and 10% are misassigned one class too
low, the fact that no tornadoes come into the F0 class from �below� means
that the distribution of assigned F-scales will be 940 F0, 436 F1, 168
F2, and 16 F3 tornadoes.
Since
the 1950s, the slope for United States tornadoes has been relatively constant
for F2-F4 tornadoes.In the limiting
case that the �true� distribution is characterized by a log-linear distribution,
it can be shown that, for large numbers of reports, the slope of the line
on a log-scale will not be affected by random classification errors except
at the ends of the F-scale.Since
the other three kinds of errors do not affect the probability distribution
function, the slope of the distribution between F2 and F4 is a basic parameter
of the distributions seen in Fig. 1.Between
27 and 35 F3 tornadoes have been reported annually in the United States,
on average, for every 100 F2 tornadoes, depending on decade, and between
5 and 8 F4 tornadoes have been reported for each 100 F2 tornadoes.
Additional
insight into the nature of the distribution can be gained by looking at
the number of tornadoes in different parts of the United States for the
period 1950-1995.In the Central
Plains (the states of Oklahoma, Kansas, and Nebraska), a region roughly
corresponding to an area sometimes called �Tornado Alley�, the number of
F3 tornadoes per 100 F2 tornadoes is slightly more than 38, with 13 F4
tornadoes per 100 F2s (Fig. 3).For
the remainder of the United States east of that region, except for Florida,
the corresponding numbers are 34 and 13.There
are almost 7000 tornadoes in the Central Plains region and over 17000 in
the Eastern United States region over the time period.The
Eastern United States region is almost ten times as large as the Central
Plains, so that per unit area, there are about four times as many tornadoes
in the Central Plains.Nevertheless,
the probability of a violent tornado, given that a tornado is reported,
is approximately the same in each region.The
unconditional probability of a violent tornado is much lower in
the Eastern United States because the overall probability of a tornado
is much lower.
Other
regions of the United States show a very different distribution.Tornadoes
in Florida, the Front Range region of Colorado just east of the Rocky Mountains,
and the West Coast states have a much steeper slope as F-scale increases
(8-11 F3s per 100 F2s and 1 F4 per 100 F2s in Florida, with no F4 tornadoes
in the 344 F2 or greater tornadoes in the Front Range and West Coast) (Fig.
3).Florida, particularly in
its southern part, has a tornado record dominated by waterspouts coming
on shore, and the Front Range of Colorado has many so-called �landspouts�
(e.g., Brady and Szoke 1989), which are generally considered to be weaker
in intensity than supercell-produced tornadoes.
We
hypothesize that the difference in the two regimes in the United States
is a result of the physical processes leading to tornadogenesis in those
regimes.The Central Plains and Eastern
United States regions appear to be dominated by processes associated with
supercells, while the other regions are dominated by non-supercell processes.To
the extent that the slopes represent the �true� distributions, it appears
that there may be two limiting slopes-with the number of tornadoes at F(n+1)
being about 36% of the number at F(n) for supercell processes and about
10% for non-supercells.This distribution
would imply that about 1 out of every 70 supercell tornadoes is violent
(F4 or F5), and about 1 out of every 7000 reported non-supercell tornadoes
in the United States is violent.
With
the background of the record from the United States, we want to look at
the distribution of tornadoes in other countries.We
have used data from eight countries (Argentina, Australia, Canada, Germany,
France, Italy, South Africa, and the United Kingdom) with more than 100
tornadoes in each reported by F-scale.The
results have been scaled to 100 F2 tornadoes as before and plotted in comparison
to the United States� data from the 1990s (Fig. 4).
The Argentine, Australian, Canadian, German, and South African records are all similar to the United States records for F3 and higher intensity events.The similarity between the Argentine and United States records continues below F2, with even the F0 report frequency being reasonably similar.The Argentine and Canadian F0 records lie between the United States in the 1980s and 1990s (see Fig. 1).Given the historical underreporting in the United States, the apparent completeness of the their records is remarkable.Although it is likely that tornadoes are missed in the reporting database there, it appears that, if so, they are missed across all ranges of intensity, with only a slight preference for the lowest end of the F-scale.Thus, the records seem to reflect an unbiased sample of the true distribution.
The South African sample shows more apparent underreporting of F0 tornadoes than the Argentine, Canadian, or modern American records.Similarly, Italian tornadoes also have a relative lack of weak tornadoes.In addition, the Italian record for F3 tornadoes lies between the two limiting distributions in the United States record.Although the sample size is relatively small (approximately 18 tornadoes per year), it is not unreasonable to believe that this reflects an important aspect of the Italian climatology.With the long coastline extending into the warm Mediterranean Sea, it seems likely that a large number of Italian tornadoes may be waterspouts that have moved on shore.The geographic distribution of Italian tornadoes supports this notion, particularly in the southern part of Italy and along the Gulf of Genoa (Giovannani 1999).Thus, the Italian record may represent both extreme limiting processes-the supercell process seen in the central and eastern United States and the non-supercell process that appears to dominate the record in Florida and the Front Range of Colorado.
The United Kingdom record, on the other hand, does not resemble any of the records from the United States.F3 reports are less than 3% of the number of F2 reports, in comparison with the 8-11% values seen from Florida, the Colorado Front Range, and the West Coast.Although this implies that the United Kingdom record may be dominated by non-supercell processes, the extremely low value is curious.Whether it implies that the apparent limit in the data from the United States is too high, or that fundamental differences exist in the basic nature of the datasets, is not clear.Given that procedures for collecting information on tornadic damage are very different in the two countries, it is not obvious that we will ever be able to resolve the reasons for the differences in the records.
The
French record is also of particular interest, especially when it is considered
using the United States record as a background.The
raw data show a less steep slope from F2 to F4 than the central and eastern
United States record, with the peak number of reports in the F2 range.A
different impression is gathered, however, by comparing the French record
to the pre-1950s United States record (Fig. 5).The
distributions are remarkably similar.Comparing
the pre-1950s United States record to the 1990s United States record shows
an increase in the total number of tornado reports by a factor of seven.Thus,
it seems reasonable to expect that if an official effort began to collect
reports in France, seven times as many reports would be collected as are
currently found.Since Paul (1999)
has reported three tornadoes per year in modern France, this leads to an
estimate that approximately 20 tornadoes per year actually occur in France.
4.Estimating
violent tornado occurrence
The
existence of what appear to be regular distributions of tornadoes with
increasing damage allows us to make estimates of the return periods of
extremely rare events such as violent tornadoes.This
is important for assessment of the threat of rare, potentially devastating
events.Assuming that the �true�
distribution of French tornadoes by F-scale would look like the 1990s in
the United States, we can make an estimate of the return period of violent
tornadoes in France.Given that,
for the entire United States in the 1990s, between 0.5% and 1% of all tornadoes
have been violent, a total of 20 tornadoes per year would lead to one violent
tornado per every 5-10 years in France.Four
were reported in the 33 years from 1967-1999, consistent with that estimate.
We can make a similar estimate for the United Kingdom, which appears to be dominated by non-supercell processes.While we do not have the evidence to make a quantitative estimate of the underreporting in the United Kingdom, we can provide some bounds on it.Reynolds (2000) reports 33 tornadoes per year in the United Kingdom.If we assume that 65-80% of actual tornadoes are reported, then using the non-supercell F(n+1)/F(n) ratio of 0.1 from the United States leads to an estimate of one violent tornado every 250-300 years in the United Kingdom.TORRO (1997) report two violent tornadoes in the United Kingdom historical record, one in 1091 and the other in 1810, both producing F4 damage.Given the great antiquity of the reports, caution must be used in their interpretation, but the modern record contains no F4 tornadoes and only 1 or 2 F3 tornadoes in the United Kingdom from 1950-1997.The rarity of even F3 tornadoes makes it seem likely that the return period of F4 tornadoes in the United Kingdom must be very long.
5.Discussion
The
long and extensive record of tornadoes by damage classification from the
United States provides a background for reports of tornadoes from other
parts of the world.In particular,
we have identified two limiting kinds of distributions when the reports
are plotted versus F-scale.One,
exemplified by the central and eastern United States, Argentina, and Canada,
is characterized by a ratio of about 36% between reports at a particular
F-scale value and the next smaller F-scale value.We
hypothesize that this distribution is dominated by processes associated
with supercell thunderstorms producing tornadoes.The
second distribution, exemplified by Florida, the Front Range of Colorado,
and the United Kingdom, is characterized by a F-scale ratio of about 10%
or less.We hypothesize that this
distribution is dominated by processes associated with non-supercell tornadogenesis.
The
existence of these two distributions provides a powerful check on the reasonableness
of tornado datasets, given that they are of sufficient size.If
the observed slope is much steeper or shallower than the �limiting� cases,
that may be indicative of certain problems with the dataset.For
instance, the relatively low number of F0 and F1 tornadoes in France are
likely the result of underreporting.Comparison
with the American record prior to the development of an official, organized
collection effort indicates that it is probable that only about 15% of
tornadoes are being reported in France.It
is not unreasonable to assume that similar (or worse) underreporting problems
occur in other countries where we don�t have large enough datasets to make
the kinds of comparisons that we have made here.
The
existence of an official, organized severe weather report collection effort
is critical for many reasons.First,
it helps identify the true nature of severe weather.The
presence of what appear to be consistent distributions in a variety of
locations makes it possible even to estimate the likelihood of extremely
rare, devastating severe weather events based upon the observed frequency
of less rare, but less devastating severe weather events.Thus,
data collection is the first step in identifying hazards for groups such
as public planning and insurance interests.
On
a longer time scale, questions of possible changes in severe weather frequency
and intensity as a result of global climate change cannot be addressed
without reasonable estimates of the �true� baseline climatology.A
reasonably long, stable record of reports is an important aspect of this
effort.It may be possible to use
climatologies of environmental observations from upper-air soundings (e.g.,
Rasmussen and Blanchard 1998), coupled with high-quality reports, to develop
covariates relating the well-observed environmental variables and the poorly-observed
severe weather occurrences (Brown and Murphy 1996).If
that can be done, then changes in the environmental variables can be tested.The
disadvantage to this is the need to develop a strong relationship between
the environment and severe weather, which may be problematic.
Severe
thunderstorms are, by their very nature, rare at any particular location.As
a result, awareness by weather forecasters, emergency managers, and the
general public may not be very good.In
the absence of good estimates of the climatology, based fundamentally on
a high-quality database of reports, if is unlikely that awareness can be
developed in any of those groups.Without
that awareness and the concomitant preparedness, the likelihood of major
disasters occurring somewhere is high.The
historical record in Europe (e.g., Boscovich 1749, Wegener 1917) demonstrates
that significant tornadoes have occurred in the past in the vicinity of
metropolitan areas.With population
increases and growth of urban areas since many of those events, the possibility
of a large-fatality tornado cannot be ignored.
Boscovich,
R. G., 1749:Sopra il turbine
che la notte tra gli XI, e XII Giugno del MDCCXLIX danneggiò una
gran parte di Roma.Niccolò,
e Marco Pagliarini, Rome, 230 pp. (In Italian)
Brady,
R. H. and E. J. Szoke, 1989:A case
study of nonmesocyclone tornado development in northeast Colorado:Similarities
to waterspout formation.Mon.
Wea. Rev., 117, 843-856.
Brooks,
H. E., and D. J. Stensrud, 2000:Climatology
of heavy rain events in the United States from hourly precipitation observations.Mon.
Wea. Rev., 128, 1194-1201.
Brown,
B. G., and A. H. Murphy, 1996:Verification
of aircraft icing forecasts:The
use of standard measures and meteorological covariates.Preprints,
13th Conf. Probability and Statistics in the Atmospheric Sciences
(San Francisco, California, USA) Amer. Meteor. Soc., 251-252.
Doswell,
C. A. III, and D. W. Burgess, 1988:On
some issues of the United States tornado climatology.Mon.
Wea. Rev., 116, 495-501.
Elsom,
D. M., and G. T. Meaden, 1982:Tornadoes
in the United Kingdom. Preprints, 12th Conf. Severe Local
Storms (San Antonio, Texas, USA), Amer. Meteor. Soc., 55-58.
Fujita,
T. T., 1971:Proposed characterization
of tornadoes and hurricanes by area and intensity.SMRP
Res. Paper 97, Univ. of Chicago, 42 pp.
Giovannoni, M., 1999:I Tornado in Italia.<URL: http://members.xoom.it/_XOOM/tornadoit/> (In Italian).
Goliger, A. M., R. V. Milford, B. F. Adam,
and M. Edwards, 1997: Inkanyamba: Tornadoes in South Africa.CSIR
Building Technology and South African Weather Bureau. 77 pp.
Grazulis,
T. P., 1993:Significant Tornadoes,
1680-1991.Environmental Films,
St. Johnsbury, Vermont, USA, 1326 pp.
Meaden,
G. T., 1976:Tornadoes in Britain:Their
intensities and distribution in space and time.J.
Meteor., U.K., 1, 242-251.
Paul,
F., 1999:An inventory of tornadoes
in France. Weather, 54, 217-219.
Rasmussen,
E. N., and D. O. Blanchard, 1998:A
baseline climatology of sounding-derived supercell and tornado forecast
parameters.Wea. Forecasting,
13, 1148�1164.
Reynolds,
D.J., 2000: A revised UK tornado climatology, 1960-1990.Atmos.
Research, this volume.
Tecson,
T. T. Fujita, and R. Abbey, 1977:Statistics
of U.S. tornadoes based on the DAPPLE (Damage Area Per Path Length) tornado
tape.Preprints, 11th
AMS Conference on Severe Local Storms, Amer. Meteorol. Soc., (Kansas City,
Missouri, USA), 227-234.
TORRO,
1997:British and European tornado
and hailstorm extremes.<URL:http://www.torro.org.uk/records.htm>.
Wegener,
A., 1917:Wind- und Wasserhosen
in Europa.Druck und Verlag von
Friedr. Vieweg and Sohn, Braunschweig, 301 pp.(In
German)
Australia
(variable), courtesy Phil Alford, Australian Bureau of Meteorology (N=239)
Canada
(1950-98), courtesy David Etkin, Environment Canada (N=625)
France
(1680-1998), courtesy Francois Paul (1999) (N=294)
Germany
(1594-1999), courtesy Nikolai Dotzek (2000) (N=136)
Italy
(1991-9), courtesy Mauro Giovannoni (N=158)
South
Africa (1905-96, 1998-9), Goliger et al. (1997) and Estelle de Coning,
South Africa Weather Bureau (N=195)
United
Kingdom (1950-97), courtesy David Reynolds, Tornado and Storm Research
Organization (N=942)
United
States (1920-98), courtesy United States National Weather Service (N=44417)
Figure
Captions
Fig.
1:Annual average of tornado
reports by decade in United States by F-scale.
Fig.
2:Annual average of tornado
reports by decade since 1800 in France (light bars) and Germany (dark bars).
Fig.
3:Tornado reports by F-scale
for different regions of United States (period of record 1950-1995).Reports
have been normalized to 100 F2 tornadoes.
Fig.
4:Same as Fig. 3 except for
different countries.
Fig.
5:Annual average of tornado
reports by F-scale for United States during 1920s (light solid line), 1930s
(light short-dashed line), and 1940s (light long-dashed line) and reports
from France for 1680-1998 divided by two (heavy solid line).