Assessing the impact of climate change on extreme fire weather events over southeastern Australia

Abstract

Extreme fire weather events in southeastern Australia are frequently associated with strong cold fronts moving through the area. A recent study has shown that the 850 hPa temperature and the magnitude of its gradient over a small region of southeastern Australia provide a simple means of discriminating the most extreme cold frontal events during the last 40 yr from reanalysis data sets. Applying this technique to 10 general circulation models (GCMs) from the Coupled Model Intercomparison Project and calibrating the temperature gradient and temperature climatology of each model's simulation of the climate of the 20th century against the reanalysis climates allows estimates of likely changes in frequency of this type of extreme cold front in the middle and end of the 21st century. Applying this analysis to the output of 10 GCM simulations of the 21st century, using low and high greenhouse gas emissions scenarios, suggests that the frequency of such events will increase from around 1 event every 2 yr during the late 20th century to around 1 event per year in the middle of the 21st century and 1 to 2 events per year by the end of the 21st century; however, there is a great degree of variation between models. In addition to a greater overall increase under the high emissions scenario, the rate at which the increase occurs amplifies during the second half of the century, whereas under the low emissions scenario the number of extreme cases stabilizes, although still at a higher rate than that experienced in the late 20th century.

Full text

CLIMATE RESEARCH
Clim Res
Vol. 39: 159–172, 2009
doi: 10.3354/cr00817
Printed August 2009
Published online August 4, 2009
1. INTRODUCTION
Over recent years, and with increased emphasis since
the release of the 4th Assessment Report (AR4) of the
Intergovernmental Panel on Climate Change (IPCC) in
2007 (IPCC 2007), there has been intense interest in
likely trends in extreme weather events due to the an-
ticipated changes in climate. There has long been par-
ticular concern about the severity of future fire seasons
in southern Australia, where the expected trend in rain-
fall is negative and in temperature is positive (CSIRO &
Bureau of Meteorology 2007), and where recent sum-
mers have seen disastrous levels of wildfire activity.
Several studies have attempted to assess the likely
severity of fire seasons in Australia under future climate
scenarios, including those of Williams et al. (2001),
Hennessy et al. (2005), Pitman et al. (2007) and Lucas et
al. (2007). Because there is a limited range of variables
and frequency of output from the general circulation
models (GCMs) relied on for climate change simula-
tions, measures of daily fire danger, such as the forest
fire danger index (FFDI, Luke & McArthur 1978), can-
not be calculated as a daily time series. Accordingly,
the measures used in these studies generally relate the
current statistical distribution of fire weather indices
mapped onto broad measures of changed climates and
conclude that during the 21st century there will be a
clear increase in severity of fire seasons as measured by
duration and number of days for which very high and
extreme FFDI is expected.
A particular feature of the fire weather climate of
southern South Australia, Victoria and Tasmania are
the rare, extreme fires such as the Hobart fires in Feb-
ruary 1967 (Bond et al. 1967), Ash Wednesday 1983
© Inter-Research 2009 · www.int-res.com*Corresponding author: b.timbal@bom.gov.au
Assessing the impact of climate change on extreme
fire weather events over southeastern Australia
A. E. A. Hasson
1, 3
, G. A. Mills
1, 2
, B. Timbal
1,
*
, K. Walsh
3
1
Centre for Australian Weather and Climate Research, GPO 1289, Melbourne, Victoria 3001, Australia
2
Bushfire Cooperative Research Centre, Level 5, 340 Albert Street, East Melbourne, Victoria 3002, Australia
3
School of Earth Sciences, University of Melbourne, Swanston Street, Melbourne, Victoria 3010, Australia
ABSTRACT: Extreme fire weather events in southeastern Australia are frequently associated with
strong cold fronts moving through the area. A recent study has shown that the 850 hPa temperature
and the magnitude of its gradient over a small region of southeastern Australia provide a simple
means of discriminating the most extreme cold frontal events during the last 40 yr from reanalysis
data sets. Applying this technique to 10 general circulation models (GCMs) from the Coupled Model
Intercomparison Project and calibrating the temperature gradient and temperature climatology of
each model’s simulation of the climate of the 20th century against the reanalysis climates allows esti-
mates of likely changes in frequency of this type of extreme cold front in the middle and end of the
21st century. Applying this analysis to the output of 10 GCM simulations of the 21st century, using
low and high greenhouse gas emissions scenarios, suggests that the frequency of such events will
increase from around 1 event every 2 yr during the late 20th century to around 1 event per year in the
middle of the 21st century and 1 to 2 events per year by the end of the 21st century; however, there
is a great degree of variation between models. In addition to a greater overall increase under the high
emissions scenario, the rate at which the increase occurs amplifies during the second half of the cen-
tury, whereas under the low emissions scenario the number of extreme cases stabilizes, although still
at a higher rate than that experienced in the late 20th century.
KEY WORDS: Fire weather · Climate change · Strong cold front
Resale or republication not permitted without written consent of the publisher

Clim Res 39: 159–172, 2009
(Bureau of Meteorology 1984), the Canberra fire in
January 2003 (McLeod 2003) or the Lower Eyre Penin-
sula fires of 11 January 2005 (Bureau of Meteorology
2005). These have enormous social and economic con-
sequences, yet only occur on a few days each decade.
In a retrospective study of the meteorology of the Ash
Wednesday 1983 fires, Mills (2005a) showed that many
of the most extreme fire events in southeastern Australia
over the 40 yr to the end of the 2003 summer, and ~80%
of the bushfire-related deaths during that period, oc-
curred on days on which the magnitude of the east–west
850 hPa temperature gradient in a small rectangle over
southeastern Australia was in the highest 0.3% of its dis-
tribution. These calculations were based on the National
Centers for Environmental Prediction/National Center
for Atmospheric Research (NCEP/NCAR) reanalysis
(NNR) data set (Kalnay et al. 1996) using analyses at
00:00 and 12:00 UTC for the summer months. It was ar-
gued that while this measure of synoptic severity did not
identify a unique weather element sequence at any loca-
tion, it did indicate a strong front. The orientation of the
southeast Australian coastline, a hot, continental air
mass inland and the cooler maritime air mass over the
Southern Ocean leads to an intensification of cold fronts
as they approach and interact with the coastal tempera-
ture gradient of southeast Australia. As these fronts
intensify, the development of a northerly pre-frontal jet
advects hot, dry gusty winds southwards— all the in-
gredients for extreme fire weather (Luke & McArthur
1978). In addition, the wind-shift associated with these
fronts can cause the flank of a fire to become a much
longer head-fire, and this effect is exacerbated if there is
sufficient depth of cold air following these fronts to main-
tain strong, gusty, post-frontal winds that can greatly
enhance fire spread (Cheney et al. 2001, Mills 2005a).
As GCM output routinely includes the 850 hPa tem-
perature field, the fact that the 850 hPa temperature gra-
dient is such a simple calculation makes the potential
application to climate change models of Mills (2005a)
analysis worth investigating. The premise is that if there
is a change in the frequency of these strong temperature
gradient events under climate change simulations, then
this would translate to a change in the frequency of the
type of strong cold front that Mills (2005a) associated
with extreme fire weather events. To address this ques-
tion a number of steps need to be taken. First, Mills
(2005a) only used the NNR data set, so the same calcula-
tions are performed based on the European Centre for
Medium Range Weather Forecasts 40 yr reanalysis data
(ERA-40; Kallberg et al. 2005) to demonstrate that the
methodology is robust to choice of reanalysis data set.
Second, output fields from the GCM archived within
Phase 3 of the Coupled Model Intercomparison Project
(CMIP3) database used in the present study were only
available at 00:00 UTC, rather than the 00:00 and 12:00
UTC used by Mills (2005a); we show that applying the
technique to reanalysis data sets valid only at 00:00 UTC
still provides discrimination of the extreme fire events.
These aspects are addressed in the following section.
In the following section the application of the method
to GCMs, and in particular the manner in which the
temperature gradient calculations were made, is as-
sessed. The models are validated against the reanaly-
ses to evaluate their temperature gradient climatology
for the latter decades of the 20th century. Finally, the
resulting projected changes in the frequency of strong
cold fronts for the middle and the latter part of the 21st
century are described, and the results summarised.
2. DATA
2.1. Reanalysis data
Both the NNR and ERA-40 data sets are available on
the same 2.5 × 2.5° horizontal grid (although the model
used in the generation of the ERA-40 data had a native
grid of 1.8 × 1.8°), and at a number of pressure levels, in-
cluding 850 hPa. Mills (2005a) used NNR data from
1964–2003 inclusive, but in the present study the period
of analysis is restricted to 1 January 1964 to 28 February
2002, the last date available for the ERA-40 analyses.
2.2. Climate model data sets
Temperature fields from 10 coupled atmosphere–
ocean GCMs from the CMIP3 used in the recent IPCC
AR4 (IPCC 2007) were available for the present study.
These are listed in Table 1; more details of the configura-
tions of these models can be found in IPCC (2007). Data
were only available at 24 h intervals, and valid at 00:00
UTC. For each model the period 1960–1999 was used to
represent the climate of the 20th century (C20C) for com-
parison with the reanalysis climates, while the periods
2046–2065 and 2081–2100 were used to assess
changes that might be discernable by the middle and
late 21st century. As part of the IPCC AR4, a range of
emissions scenarios were used to force the GCMs and
provide estimates of a range of possible climate projec-
tions (IPCC 2000). For the present study, 2 of these pro-
jections were used, 1 each corresponding to one of the
higher (A2) and one of the lower (B1) emissions growth
scenarios. The A2 (B1) scenario leads to a global average
temperature rise of some 2.0 to 5.4°C (1.1 to 2.9°C) by the
end of the 21st century (IPCC 2007). For each scenario
the range of possible warming indicates the differing
sensitivities of the individual climate models to these ex-
ternal forcings. A measure of the sensitivity of the indi-
vidual GCMs used in the present study is indicated in
160

Hasson et al.: Fire weather events and climate change
the right-hand column of Table 1, where the global mean
temperature rise for each model forced by the A1B emis-
sion scenario (the closest to the mean of the scenarios
considered) and approximated by linear regression over
the 21st century (CSIRO & Bureau of Meteorology 2007,
their Table 4.1) is shown.
2.3. Extreme fire event data
The quality, temporal consistency and content of fire
event and/or activity databases in Australia is rather
varied, and this makes it difficult to quantitatively
relate fire activity or occurrence to purely meteorolog-
ical features. In addition, the probability of an ignition
occurring is unknown, and the effectiveness of mitiga-
tion efforts affects the subsequent fire behaviour and
consequences. Mills (2005a) used published reports as
an indicator of the socioeconomic impacts of a fire,
with the assumption that only ‘significant’ fires subse-
quently lead to documented post-event studies. That
study did not attempt to develop a causal relation-
ship between the temperature/temperature gradient at
850 hPa and fire activity, but did match published re-
ports and documented bushfire-related deaths to the
30 strongest temperature gradient days in the 40 yr
NNR data set analysed to demonstrate that these
days included a disproportionate number of major fire
events and fatalities. In the present study we use the
fire event data in a similar way — to qualitatively indi-
cate the robustness of the analysis technique to choice
of reanalysis data set, and to support the hypothesis
that changes in frequency of the strongest cold fronts is
likely to lead to an increase in damaging fire events.
The restriction of our comparison between the re-
analyses to the period of the ERA-40 data set precludes
using the fire events in the 2002–2003 summer (Mills
2005a, their Tables 1 & 2), which made some of the Mills
(2005a) event days unavailable for comparison with the
ERA-40 data. A few other event days have been added,
based on fire agency web-based data sets, and these are
indicated in Table 2. These data are only used in the
comparison of the NNR and ERA-40 data, as the GCMs
generate their own daily weather succession that cannot
be compared directly with the observed succession.
161
Table 1. List of acronyms used for IPCC AR4 coupled models,
the nation where they were developed and their resolution.
Temperature change indicates the change in temperature be-
tween 1980–1999 and 2090–2099 for each model under the
A1B emissions scenario
Acronym Nation Grid Temperature
spacing (°) change (°C)
CCM Canada 3.8 × 3.8 2.47
CNRM France 2.8 × 2.8 2.81
CSIRO Australia 1.9 × 1.9 2.11
GFDL1 USA 2.5 × 2.0 2.98
GFDL2 USA 2.5 × 2.0 2.53
GISSR USA 5.0 × 4.0 2.12
IPSL France 3.7 × 2.5 3.19
MIROC Japan 2.8 × 2.8 3.35
MPI Germany 1.9 × 1.9 3.69
MRI Japan 2.8 × 2.8 2.57
Table 2. List of fire events in southeast (SE) Australia over 1964–2000, adapted from Mills (2005a), featuring for each fire event its
location, date, values of the 00:00 UTC maximum thermal gradient (T
G
) and maximum temperature (T
max
) from NNR and ERA-40
data sets and from the ERA-40 modified data set as described in the text.
a
Events not listed in Mills (2005a),
b
event points in
Fig. 4b that have relatively high temperatures but T
G
values <2.5 K. NSW: New South Wales; SA: South Australia
Location Date NNR ERA-40 ERA-40 modified
(yr) (mo) (d) T
G
T
max
T
G
T
max
T
G
T
max
Longwood 1965 1 17 2.11 296.2 3.36 295.9 3.61 295.3
Brigalong
a,b
1965 2 22 3.27 292.2 1.56 298.3 2.70 292.1
Hobart
b
1967 2 7 3.26 296.0 1.77 295.4 2.58 294.5
SE Australia fires
b
1968 1 31 2.85 298.3 2.08 292.1 2.91 296.5
Yarra Junction
b
1972 12 2 3.86 293.1 1.38 294.9 2.74 287.9
Western District
a,b
1976 1 3 3.01 297.0 1.75 292.2 2.46 296.1
Streatham 1977 2 12 3.18 295.6 2.88 294.9 2.88 294.9
Paynesville 1978 1 15 3.79 298.7 3.01 292.2 3.01 292.2
Caroline Forest 1979 2 3 3.33 293.8 3.05 294.6 3.05 294.6
Mallee 1981 1 3 3.20 297.9 3.45 297.0 3.48 297.0
Central Victoria 1982 1 11 2.54 297.1 2.91 295.9 2.91 295.9
Yallourn 1982 1 24 3.21 302.1 3.29 300.8 3.29 300.8
Wombat State Forest
a
1983 1 9 3.27 291.1 3.47 291.3 3.47 291.3
SE Australia (Ash Wednesday) 1983 2 16 2.35 298.8 3.00 298.6 3.73 286.3
SE Australia, 33 fires
a,b
1984 2 26 2.88 292.5 2.03 289.0 4.00 292.3
Central Victoria 1985 1 14 2.92 299.7 3.27 297.8 3.27 297.8
SE Australia, 132 fires 1990 1 3 3.49 299.5 3.51 298.1 3.51 298.1
Tasmania 1996 12 25 2.98 289.5 3.37 288.0 3.37 288.0
NSW/Tasmania
a
1997 12 21 2.77 294.9 3.01 294.0 3.01 294.0
SA/Mt. Macedon 1998 2 26 3.31 299.2 3.72 299.0 3.72 299.0

Clim Res 39: 159–172, 2009
3. COMPARISON OF REANALYSES
Mills (2005a) characterised the synoptic character of
an extreme fire weather day over southeastern Aus-
tralia by the maximum value of the 850 hPa tempera-
ture gradient (T
G
) at the reanalysis gridpoints over a
small subsection (35–40° S, 135–150° E) of the analysis
domain; hereafter termed the ‘gradient box’. This is
shown in Fig. 1 overlaid on the NNR 850 hPa tempera-
ture field for the evening of the Ash Wednesday fires in
1983. The elongated gradient box was chosen so that if
an eastward-moving frontal zone was located some-
where in the gradient box, then a high value of the
temperature gradient would be diagnosed; it was
argued in Mills (2005a) that using reanalyses every
12 h was sufficient to capture most events. It was also
noted that many of the major fire events were associ-
ated with stronger east –west than north–south tem-
perature gradients: that is, more meridionally-oriented
isotherms. It was proposed that a parameter that iden-
tified these events was the highest temperature on the
east–west grid row through the middle of the gradient
box, termed T
max
. Mills (2005a) argued that the high
T
G
/high T
max
area of the entire T
G
/T
max
phase space
included the major events listed, and so discriminates
such events (Mills 2005a, their Fig. 16).
Fig. 2 shows the synoptic patterns associated with
the 8 strongest gradient ‘documented event’ days from
Mills (2005a, their Table 1), excluding the events after
the end of the ERA-40 data, from the ERA-40 analyses.
These show the mean sea level pressure and 850 hPa
isotherms, and in all cases there is a marked thermal
ridge and a marked baroclinic zone over southeast
Australia associated with the passage of a surface
trough. As noted in Mills (2005a), there is sufficient
case-to-case variation to mean that the particular sur-
face weather sequence at any location will differ from
case to case, but the strong 850 hPa thermal pattern is
common. Other more recent cases are documented in
Mills (2005b, 2008) and Hasson et al. (2008).
Since the publicly available NNR and ERA-40 data-
sets have the same spatial discretisation, identical gra-
dient boxes over southeast Australia can be used to
calculate T
G
and T
max
. Fig. 3 shows the T
G
versus T
max
scatter plots for the NNR and ERA-40 data sets with
both 12:00 and 00:00 UTC analyses included, and with
162
Fig. 1. Temperature field (K) at 850 hPa at 12:00 UTC 16 February 1983 from the NCEP/NCAR reanalysis data set. The rectangle
marks the gradient box used in the calculation described in the present study

Hasson et al.: Fire weather events and climate change
163
Fig. 2. ERA-40 analyses of 8 fire events in southeast Australia associated with strong 850 hPa temperature gradients. Black
contours: mean sea level pressure (MSLP), contour interval 4 hPa; gray contours: 850 hPa temperature, contour interval 4 K.
The descriptors are those used in Table 1 of Mills (2005a)

Clim Res 39: 159–172, 2009
the event dates shown in Table 2 highlighted. For
NNR, the events cluster strongly in the top right-hand
sector of the distribution, as seen in Mills (2005a).
While there is a greater degree of scatter in the events
in the ERA-40 data set, there is still a strong bias to
higher temperatures and an indication of enhanced
event frequency towards the stronger gradients.
The IPCC climate model datasets available for our
later analysis of the 21st century simulations were only
available at 24 h intervals, valid at 00:00 UTC (11:00 h
Eastern Australia Summer Time). Applying the analysis
to only 00:00 UTC reanalysis data on the day of the fire
leads to a considerably greater scatter of the event data
(Hasson et al. 2008, their Figs. 13–15). However, with
the normal diurnal variation of fire activity peaking in
the afternoon, this result is perhaps to be expected.
Sampling may also contribute to this result, as the east-
ward movement of a frontal system may well mean that
with 1 sample per day the major thermal gradient may
not lie within the gradient box at 00:00 UTC, particu-
larly as land–sea thermal gradients will also be
stronger later in the day. These issues are ameliorated if
the highlighted event points are matched to the higher
T
G
value either at 00:00 UTC on the calendar day of the
fire event, or on the subsequent 00:00 UTC analysis;
this analysis is shown in Fig. 4. There is a clear cluster-
ing of the event points towards the high T
G
/high T
max
zone of the phase diagrams.
An alternative way of illustrating this point is to
look more closely at the 6 event points in Fig. 4b that
have relatively high temperatures but T
G
values < 2.5 K
100 km
–1
. These events are marked ‘b’ in Table 2,
which lists, for each event, the 00:00 UTC values of T
G
and T
max
for NNR and ERA-40 on the calendar day of
the fire event, and also those values at 00:00 UTC on
the following day for ERA-40 (ERA-40 modified). Each
of these highlighted events shows a higher T
G
value at
00:00 UTC on the day following the date of the fire
event, although on only 2 of those 6 days was T
max
also
higher. This is consistent with the eastward movement
of a frontal system such that a baroclinic zone further
east in the gradient box would be more likely to have a
lower T
max
than one in the western part of the gradient
box.
164
y = 2.56x + 282.9
R
2
= 0.09
y = 2.45x + 282.1
R
2
= 0.08
270
275
280
285
290
295
300
305
0.5 1 1.5 2 2.5 3 3.5 4 4.5
0.5 1 1.5 2 2.5 3 3.5 4
Hobart
Hobart
Ash Wednesday
Ash Wednesday
T
max
(K)
270
275
280
285
290
295
300
305
T
G
(K 100 km
–1
)
T
max
(K)
Fig. 3. Scatter plot and its linear
regression of the 850 hPa maxi-
mum thermal gradient (T
G
) ver-
sus the maximum temperature
(T
max
) on 37.5°S over the Victo-
rian Box at 00:00 and 12:00
UTC during summer for the pe-
riod 1964–2002 for the NNR
(top) and ERA-40 data sets
(bottom). The squares are for
the events listed in Table 2

Hasson et al.: Fire weather events and climate change
It thus appears that, while less satisfactory than hav-
ing a 12 h interval between analyses, applying this
technique to a time sequence of analyses at 24 h inter-
vals does still show a clear association with extreme
fire events over southeastern Australia. This is an
important result in the context of the present study, as
it makes it possible to apply the technique to the
CMIP3 data set, for which only 00:00 UTC data are
readily available. The hypothesis can thus be devel-
oped that thresholds for T
G
and T
max
may be specified
such that if each is exceeded, then the part of the phase
space that exceeds both thresholds determines the
environments of this paradigm of extreme fire weather
environments in southeastern Australia. Thresholds for
the NNR data that that include all the fire events based
on Fig. 3a are 290 K for T
max
and 3.2 K 100 km
–1
for T
G
.
4. APPLICATION TO CLIMATE CHANGE MODELS
Any changes with time in the number of events
exceeding the joint thresholds may be interpreted as a
change in the frequency of extreme fire weather
events due to climate change. An inherent assumption
is that the threshold values of T
G
and T
max
have the
same physical significance under changed climate
regimes. There are, however, a number of aspects that
complicate the application of this technique to the cli-
mate models. Each model has a different resolution
(see Table 1), and so the calculations cannot be per-
formed on the same grid as the reanalyses, or on the
same grid for each model, and there is no a priori rea-
son to assume that the parameters simulated by the
GCMs will have the same statistical distribution as the
reanalyses. Accordingly, for each model the following
steps were taken in order to have a basis for assessing
the impact of climate change: (1) specification of gradi-
ent box bounds; (2) specification of latitude for T
max
calculations; (3) specification of T
G
and T
max
thresh-
olds; and (4) comparison of the T
G
/T
max
distribution for
C20C with the NNR and ERA-40 distributions. Only
general descriptions of these steps are given in the
following sections. Full details can be found in Hasson
et al. (2008).
165
270
275
280
285
290
295
300
305
270
275
280
285
290
295
300
305
y = 2.77x + 282.3
R
2
= 0.10
y = 2.29x + 282.1
R
2
= 0.07
0.5 1 1.5 2 2.5 3 3.5 4 4.5
0.5 1 1.5 2 2.5 3 3.5 4 4.5
T
max
(K)T
max
(K)
T
G
(K 100 km
–1
)
Hobart
Hobart
Ash Wednesday
Ash Wednesday
Fig. 4. Scatter plot and its linear re-
gression of the 850 hPa maximum
thermal gradient (T
G
) versus the
maximum temperature (T
max
) on
37.5°S over the Victorian Box at
00:00 UTC during summer for the
period 1964–2002 for the NNR (top)
and ERA-40 data sets (bottom). The
squares are for the events listed in
Table 2, with the higher of the 00:00
UTC T
G
values for the listed (Table 2)
and 24 h following dates highlighted

Clim Res 39: 159–172, 2009
4.1. Specification of bounds of the gradient box for
each model
Table 1 shows the range of grid spacings of the cli-
mate change models used in the present study; none
have the 2.5 × 2.5° spacing of the reanalyses. Accord-
ingly, careful decisions were needed when selecting
the appropriate gradient box and the latitude on which
T
max
was calculated for each model.
The size of the each individual gradient box was cho-
sen to be as close as possible to the bounds used for the
reanalyses (see Fig. 1). In most cases this resulted in
slightly larger geographic areas, but not necessarily a
larger number of gridpoints. While the east –west and
north–south dimension of the reanalysis gradient box
was 12.5 × 5°, these varied between the climate change
models from 11.1 (IPSL) to 15.0 (GFDL, GISSR) degrees
longitude and 7.4 (CCM) to 8.4 (MRI) degrees latitude
(Table 3).
4.2. Specification of T
max
latitude for each model
Selection of the latitude at which to calculate T
max
was often ambiguous due to the varying relationships
between the selected gradient box, the grid resolution
and the particular land –sea mask used in each model.
Accordingly, for each model, scatter plots similar to
Figs. 3 & 4 were constructed for T
max
calculated on the
latitude that was closest to the 37.5° used for the re-
analyses, and also using the next lower (equatorward)
latitude row. The next higher latitude was not tested as
it was generally located over the ocean south of the
modelled continent. The distribution of the points in
these diagrams varied, and for each model the latitude
for which the slope of the trend-line was closer to that
based on the NNR analyses was se-
lected, although it must be acknowl-
edged that the correlations, and thus
the statistical significance of these
slopes, are very low. This choice,
though, did select those latitudes
that showed a tendency for higher
temperatures at the higher range of
T
G
. The selected latitudes ranged
from 38.0 (GISSR) to 34.2°S (IPSL)
(Table 3). All scatter plots and a list-
ing of the 2 latitudes for each model
are found in Hasson et al. 2008.
4.3. Specification of thresholds
Inspection of the scatter plots in
Hasson et al. (2008) shows consider-
able variation in apparent distribution, together with
considerable range in absolute values, which may be
partly due to the normal latitudinal variation in tem-
perature and partly to the differing latitudes used in
the calculations. The approach to threshold specifica-
tion used with the reanalyses cannot be used, as the
climate models do not simulate actual events during
the C20C period, and so statistical methods were used.
A technique that determined approximately the same
number of events per year for each model for the C20C
was required, and after several approaches were
tested (details in Hasson et al. 2008) the thresholds
were specified based on the percentage of events that
exceeded the T
G
and T
max
thresholds in the NNR data
set. The data presented in Fig. 4 shows that 1.13% of
days exceeded the T
G
threshold of 3.2 K 100 km
–1
and
28.39% of days exceeded the T
max
threshold of 290 K.
Applying thresholds based on these percentages to
each of the climate models and to the ERA-40 reanaly-
sis in turn gave numbers of events in the C20C period
ranging between 12 (IPSL and GFDL1) and 32 (MPI).
These numbers can be compared with the reanalysis
numbers of 27 (NNR) and 28 (ERA-40). The individual
thresholds and numbers of events for each model are
listed in Table 3.
5. EVALUATION OF C20C SIMULATIONS
Before addressing future climates, this section
makes some comparisons between the C20C simula-
tions of the T
G
/T
max
phase space and the same distrib-
utions from each set of reanalyses, and a broad assess-
ment of the models is made on the assumption that
those that best simulate the reanalysis phase space for
their C20C simulations may be more reliable in their
166
Table 3. Maximum temperature gradient (T
G
) and maximum temperature (T
max
)
threshold values and number of cases where the thresholds are jointly exceeded for
each model, together with the selected latitude for the T
max
calculation and the
dimensions of the gradient box for each model. C20C: climate of the 20th century
Model Gradient box Selected T
max
threshold T
G
threshold No. events
(Long × Lat, °) latitude (°) (K 100 km
–1
) (K) (C20C)
NNR 12.5 × 5.0 –37.5 3.20 290.0 27
ERA-40 12.5 × 5.0 –37.5 3.53 289.6 28
MPI 13.3 × 7.6 –34.5 3.66 291.5 32
MIROC 14.0 × 8.4 –34.9 2.82 286.7 19
IPSL 11.1 × 7.5 –34.2 3.15 286.2 12
GFDL1 15.0 × 8.0 –35.0 4.63 292.1 12
CNRM 14.0 × 8.4 –37.7 2.29 283.1 15
GFDL2 15.0 × 8.0 –35.4 4.06 289.4 22
MRI 14.0 × 8.4 –37.7 2.96 285.6 23
CCM 11.4 × 7.4 –35.3 2.62 286.7 22
GISSR 15.0 × 8.0 –38.0 2.65 289.0 25
CSIRO 13.3 × 7.6 –34.5 3.20 289.5 27

Hasson et al.: Fire weather events and climate change
simulations of future climates. Although hard to prove,
it is an assumption often made when evaluating pro-
jections from climate models (CSIRO & Bureau of
Meteorology 2007, Whetton et al. 2007). The means
and variances of the T
G
and T
max
distributions are first
compared, and Fig. 5 shows the differences between
the means and variances of the T
G
distributions from
that of the NNR. The differences between the means
and variances of the 2 reanalyses provide some mea-
sure of the uncertainty of truth. We arbitrarily consid-
ered that a model’s simulation of the C20C
was close to that of the reanalyses if the
value of its mean T
G
was between those of
the 2 reanalyses (dashed lines in Fig. 5),
and a model was considered satisfactory if
its mean was within the range delimited by
the reanalyses’ means plus or minus their
difference (i.e. between –0.31 and +0.62 K
km
–1
; the dotted lines in Fig. 5). Models
with means outside these bounds were
considered distant. On this basis, MRI,
CSIRO and GISSR models were close, MPI,
MIROC, IPSL, CNRM and CCM models
were satisfactory and GFDL1 and GFDL2
models were distant.
It is difficult to apply the same methodol-
ogy to the difference between the variances
of the climate models and the variance dif-
ference between NNR and ERA-40, as this
latter difference (Δ = 0.012) is very small. It
was decided to consider a model as having
a distant T
G
distribution if the variance dif-
ference was > 50% of the NNR variance, i.e.
0.157. This test adds the CNRM model to
those considered distant, and the GFDL1
model is rated distant on this, as well as the
previous test.
Earlier in this paper the trend-line slope of
the scatter plots of T
G
versus T
max
was used to
select the latitude at which T
max
was calcu-
lated. There is still, however, considerable
variation in this slope from model to model
(Hasson et al. 2008, their Fig. 20 & Appen-
dix). In Fig. 6 the difference of the slopes of
the trend-lines in those scatter plots is com-
pared with the difference between the NNR
and ERA-40 trend-line slopes. Applying a
methodology analogous to that based on T
G
(Fig. 5), and using the same conventions,
classes the MIROC, CNRM, and GISSR mod-
els distant, while the GFDL1, GFDL2, CCM
and CSIRO models are classed close.
Simple statistical tests such as the com-
parison of mean and standard deviation do
not explore the entire data distribution, and
indeed the accurate reproduction of the mean and
variance of T
G
does not necessarily show that the
model probability density functions (PDFs) have reli-
able higher order moments. However, it is perhaps
these higher order moments that are more important
for the extreme events that are the focus of this inves-
tigation. Consequently, an attempt was made to use
higher-level statistical tests to strengthen our evalua-
tion of the models. Two tests, both based on the whole
data PDFs, were performed for our model evaluation
167
–0.6
–0.4
–0.2
0
0.2
0.4
0.6
0.8
1
ERA MPI MIROC IPSL CCM GISSR CSIROMRIGFDL2CNRMGFDL1
T
G
(K km
–1
)
Mean SD
Fig. 5. Difference between the mean and SD of NNR maximum thermal gra-
dient (T
G
) values and those of ERA-40 and the climate models. The boundary
between values of close and satisfactory models is shown by the dashed rect-
angle and between satisfactory and distant models by the dotted rectangle
–1.6
–1.4
–1.2
–1
–0.8
–0.6
–0.4
–0.2
0
0.2
0.4
0.6
0.8
1
Slope of regression (1/100 km)
ERA MPI MIROC IPSL CCM GISSR CSIROMRIGFDL2CNRMGFDL1
Fig. 6. Difference between the slopes of the linear regressions of the scatter
plots for the climate models and that of the NNR scatter plot. The dashed and
dotted rectangles have the same significance as in Fig. 5

Clim Res 39: 159–172, 2009
analysis: the Kolmogorov-Smirnov (KS) test (Wilks
1995) and a recent test proposed by Perkins et al.
(2007). These test the model PDFs against the PDFs of
the reanalyses, using the difference between the re-
analyses as an estimation of the uncertainty.
The KS test is used to assess by how much the mod-
els’ PDFs differ from those of the reanalyses. Unlike
most goodness-of-fit tests, this test makes no assump-
tions about the distribution of the data. The KS test
computes the D-statistic, which is the maximum verti-
cal deviation between the cumulative fraction plots of
2 sets of data, and assesses the differences in shape
and location of the cumulative distribution functions of
the 2 data sets. The best skill score for the KS test cor-
responds to the smallest value of its D-statistic. Due to
the dependence of the T
max
distribution on latitude (see
above), the KS test was only applied to the maximum
temperature gradient (T
G
) distributions over the
defined gradient boxes, using the R statistical package.
The D-statistic has been calculated with respect to
both reanalyses and therefore the test gives 2 values
per model (Fig. 7).
The KS test was also used to compare the 2 reanaly-
sis data sets, resulting in a D-statistic of 0.18; this value
is used for reference in the climate model comparison.
Much of the difference between the 2 reanalyses lies in
the different modes of the distributions, with the
modes of the NNR and ERA-40 distributions being 1.3
and 1.7 K 100 km
–1
, respectively. It is also worth noting
that the correlation between the 2 T
G
time series is
considerably stronger in the post-satellite period from
1979 (40% explained variance for 1964–1978, 62%
for 1979–2002). Averaging the 2 D-statistics for each
model (numbers and curve in Fig. 7) pro-
vides a basis on which each model can be
compared with the reanalysis D-value, D
era
.
In this assessment those models whose D-
value is lower than D
era
(MRI, CSIRO, CCM
and MIROC) are regarded as close, while
those whose D-value is >1.5 times D
era
(GFLD1, GFLD2 and MPI) are classed as
distant. Models whose D-values lie be-
tween these limits (CNRM and IPSL) are
considered satisfactory.
Stating that ‘it is not clear how to sum
across […] PDF-based statistics’, Perkins et
al. (2007, p. 4326) proposed an alternative
way to objectively assess the ability of
models to reproduce PDFs of selected
observed parameters by measuring the
common area under the PDFs of control
and test data sets. This test gives each
model a skill score (SS), with an SS of 1
indicating 2 identical frequency distribu-
tions. The SS is defined as:
where n is the number of bins used to calculate the
PDF for a given region, Z
m
is the frequency of values in
a given bin from the model and Z
o
is the frequency of
values in a given bin of observed data. Considering
the reanalyses to be 2 equivalently plausible versions
of the observed data, 2 SSs were computed for each
model (Fig. 8), where the models’ average SS have
been ranked in increasing order. There are no particu-
larly obvious break points in this series, but if the
reanalyses’ SS of 0.83 is used as a criterion for classing
a particular model as close, then MRI, CSIRO, CCM
and MIROC meet this criterion, while GFDL1, GFDL2
and CNRM models are distant.
Interestingly, apart from exchanging the positions of
the CNRM and MPI models, the ordering of the models
is the same using both the KS and the Perkins tests
(Figs. 7 & 8). The simple tests, based on mean and vari-
ance, and the higher order tests result in a very similar
ranking of the models, with CCM, CSIRO and MRI
models consistently being close, and the CNRM,
GFDL1 and GFDL2 models consistently distant. (It must
be acknowledged that the KS and Perkins tests are
applied to the whole data distribution, and so do not
test purely the upper tails of the distributions.) Al-
though this assessment is rather arbitrary, it does assist
the interpretation of the results from future climate
scenarios in the following section. In these future cli-
mate scenarios, all models are included but are differ-
entiated based on these assessments in order to better
describe the uncertainty attached to the future projec-
tions for these extreme weather events.
SS
mo
=
∑
min( , )ZZ
n
1
168
0.37
0.36
0.29
0.26
0.20
0.14
0.14
0.13
0.10
0.15
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
0.45
0.50
D-statistic
NNR ERA-40 Average
ERAMPI MIROCIPSL CCMGISSR CSIRO MRIGFDL2 CNRMGFDL1
Fig. 7. Kolmogorov-Smirnov (KS) test results (D-statistic) for each of the cli-
mate models using both NNR and ERA-40 as control sets. The average values
are linked by a dashed curve. Gray ERA bar: the KS test comparing the
ERA-40 with the NRR reanalysis

Hasson et al.: Fire weather events and climate change
6. PROJECTED CHANGES DURING THE
21ST CENTURY
In this section the changes in the number of strong cold
front days under the B1 and A2 SRES scenarios for two
20 yr periods centered on 2050 and 2090 are examined.
This analysis assumes that the thresholds defining an ex-
treme fire day will remain unchanged; therefore, any
changes in the number of days on which both T
G
and
T
max
thresholds are exceeded equate to changes in fre-
quency of these extreme fire weather frontal passages.
As an example, Fig. 9 shows the scatter plot of T
G
versus T
max
under the climate change scenarios from
the MRI model. Displayed are the results for the 20th
century (blue dots) and the A2 scenario (strongest
heating) for the 2081–2100 period (orange dots). The
trend lines for the C20C and the B1 and A2 scenarios
for the middle (2046–2065) and end (2081–2100) of the
21st century are overlayed. The slopes of all trend-
lines are virtually identical, but each line is displaced
to higher temperatures consistent with the simulated
average warming under the under the different emis-
sions scenarios and time extrapolations. It should be
noted that there is also a larger number of high T
G
val-
ues under the more extreme scenario (orange dots).
In Fig. 10 the percentage change in number of
extreme frontal systems for the assessment periods in
the middle and end of the 21st century are shown for
the 3 models (MRI, CCM and CSIRO) deemed in the
previous section to be consistently closest to the re-
analyses’ climate. These 3 models all show an increase
in the occurrence of the class of extreme fire weather
events as defined in the present study over the 21st
century, although there is considerable difference in
the amplitude of the signal between the 2 scenarios.
Under the B1 scenario, all 3 models show an increased
frequency of extreme events by the middle of the 21st
century. This frequency then declines slightly towards
the end of the 21st century, but still ranges between 13
and 99% greater than during the late 20th century.
Under the A2 scenario, there is a change of between
–25 and 102% by the middle of the 21st century, and an
increase of between 77 and 312% by the end of the 21st
century.
While the projections from these 3 models might be
hypothesized to be more reliable than the other models,
there is a need to assess the uncertainties attached to
these future projections, as their application to extreme
events, such as the intense frontal systems dealt with in
the present study, remains unproven. Given the spread
of results between models in Fig. 10, there is value in
presenting the projected changes for all the available
climate models in order to better interpret the uncer-
tainties already present between the closest models.
Such changes for all 10 models are shown in Fig. 11 for
the B1 and A2 scenarios in terms of number of events
per year. The 3 models whose climates are closest to the
reanalyses are in bold, and the 3 whose climates are
most distant are shown with dashed lines. For the B1
scenario, there is considerable qualitative consistency
in the projections, with all models showing an increase
in the number of events per year during the first half of
the 21st century, and all but one (IPSL) then either
showing a decrease, or a substantially lower rate of in-
crease, in numbers towards the end of the 21st century,
but with an increasing spread in the potential numbers.
Under the A2 scenario, all but 2 models show an in-
crease in the number of events during the first half of
the 21st century and an even greater increase
in the next 50 yr. Even if the IPSL model
(the model showing the most extreme rate
of increase) is excluded, the mean number of
events per year changes from 0.59 (C20C) to
0.93 (1.04) by the middle of the 21st century for
the B1 (A2) scenarios, and to 0.94 (1.54) by the
end of the 21st century.
The projected changes in strong frontal fre-
quency (Fig. 11) do not appear to be related to
the climate model sensitivity shown earlier
(Table 1). For example, the CSIRO model is
the least sensitive in terms of mean global
warming but shows the largest change in
number of extreme cold fronts under the B1
forcing scenario, while the second most sensi-
tive model (MIROC) shows the lowest A2 sig-
nal in terms of change in the number of strong
cold fronts. These associations suggest that
the changes in strong frontal frequency do not
relate to the model’s global sensitivity.
169
0.64
0.65
0.71
0.73
0.80
0.86
0.86
0.88
0.91
0.83
0.50
0.55
0.60
0.65
0.70
0.75
0.80
0.85
0.90
0.95
1.00
Perkins skill score
NNR
ERA-40
Average
ERAMPI MIROCIPSL CCMGISSR CSIRO MRIGFDL2 CNRMGFDL1
Fig. 8. Results of the skill score (SS) test (Perkins et al. 2007) for each cli-
mate model using both NNR and ERA-40 as control sets. Average values
are linked with a dashed curve. Gray ERA bar: the SS test comparing the
ERA-40 with the NRR reanalysis

Clim Res 39: 159–172, 2009
7. DISCUSSION
The 850 hPa maximum temperature/temperature
gradient phase space that Mills (2005a) suggested dis-
criminates the environments of the most extreme fire
weather cold fronts over southeastern Australia has
been applied to NNR and ERA-40 reanalyses, and to
the output of 10 GCMs’ simulation of both the climate
of the late 20th century and their simulations of cli-
mates in the middle and late 21st century. It was shown
that the 00:00 UTC data available from climate models
from the CMIP3 data set were sufficient to discriminate
these events.
However, the application to climate models required
careful selection of target areas and latitudes due to
their varying grid resolutions and careful evaluation of
their simulations of the 20th century climate. As any
changes in frequency in the extreme cold frontal syn-
optic paradigm under climate change scenarios are
dependent on the exceedences of temperature and
temperature gradient thresholds, the selection of these
threshold values is an important decision, and is
model-dependent since the models’ C20C climates are
different to those of the reanalyses. In the present
study, thresholds based on the percentage of excee-
dences in the C20C being the same as those percent-
ages in the reanalyses was used. While Hasson et al.
(2008) tested some alternative approaches, these were
found to be less satisfactory than the approach used,
but other techniques should be explored in the future.
When applied to climate simulations of time-slices in
the middle and the end of the 21st century, an increas-
ing frequency of strong cold frontal events under cli-
mate change conditions are indicated for the 2 forcing
scenarios investigated. Under both the low and high
emissions scenarios, the models used in the present
study show a general increase, but with very large
uncertainty, in the annual frequency of extreme frontal
fire weather events by the end of the 21st century.
Excluding the one model that showed a far stronger
trend than the others, the increases amount to a
change from around 1 event every 2 yr during the 20th
century to around 1 event per year in the middle of the
21st century, and 1 to 2 events per year by the end of
170
MRI summer TGRAD vs TMAX
275
280
285
290
295
300
305
0.5 1 1.5 2 2.5 3 3.5
C20C
A2_100
Linear (A2_50)
Linear (B1_50)
Linear (B1_100)
Linear (C20C)
Linear (A2_100)
T
max
(K)
T
G
(K 100 km
–1
)
Fig. 9. As Fig. 3 but for
the MRI model C20C
(blue dots). Values for
the period 2081–2100 are
displayed as orange dots.
Linear regressions for the
period 1964–2000 and for
the high and low emis-
sions scenarios over the
periods 2046–2065 (A2_50
and B1_50) and 2081–
2100 (A2_100 and B1_100)
are also shown
–50
0
50
100
150
200
250
300
350
MRI CCM CSIRO
% change
B1 (2046-2065)
B1 (2081-2100)
A2 (2046-2065)
A2 (2081-2100)
Fig. 10. Percentage change in the number of extreme cases
per decade from the 20th century to the periods 2046–2065
and 2081–2100 for the low and high emissions scenarios for
the 3 models whose 20th century climate is closest to that of
the reanalyses

Hasson et al.: Fire weather events and climate change
the 21st century; however, there is a great degree of
variation between models. Interestingly, there appears
to be little relation between the global mean tempera-
ture increases indicated by each model and the pro-
jected changes in the number of extreme cold frontal
events indicated by our analysis, suggesting that diag-
nosis of the type applied in the present study is com-
plementary to analyses of changes based on mean
global warming statistics. It is also interesting to note
that in addition to a greater overall increase under the
high emissions scenario, the rate at which the increase
occurs amplifies during the second half of the century,
whereas under the low emissions scenario the number
of extreme cases stabilizes. However, since no rela-
tionship is apparent between global temperature in-
crease (which is higher with the A2 than the B1 emis-
sions scenario) and extreme cold front occurrence, it is
a reasonable hypothesis that the difference between
the 2 scenarios is due to the different forcing and
warming rates. The difference in the overall increase
of the number of extreme danger episodes between
the 2 emissions scenarios, and between the different
models, should be taken in account when interpreting
the large relative increases that some models predict
by the end of the 21st century under the high emissions
scenario.
Previous studies (Hennessy et al. 2005, Lucas et
al. 2007) investigated the evolution of occurrence of
extreme fire danger days through FFDI statistics for
southeast Australia. Hennessy et al. (2005) found a
potential increase of very high and extreme FFDI days
by 15 to 70% by 2050, based on 2 CSIRO fine (60 km)
resolution climate models, scaled by IPCC (2001)
global warming ranges for the full range of emission
scenarios, including A1FI. Lucas et al. (2007) found a
potential increase of 5 to 100% by 2050, using the
same 2 CSIRO models scaled by the IPCC (2007)
global warming ranges for the full range of emissions
scenarios, including A1FI. In the present study, the
focus has been on projected changes in the number of
a particular type of synoptic weather pattern that has
been shown to be associated with extreme fire weather
events. Our conclusions that the GCMs indicate an
increase in the frequency of such events, combined
with the results of the studies based on FFDI, adds to
the consensus regarding future increased fire danger
in southeast Australia.
In the years since February 2002 there have been a
number of disastrous fires in southeastern Australia,
and Lucas et al. (2007) showed that these seasons show
an extraordinary increase in the number of observed
extreme fire danger days per season, as well as in-
creases in the median FFDI for those seasons. Repeat-
ing this exercise with updated reanalysis data to
include data to the end of the 2008–2009 summer, and
with higher-resolution climate model data, would be a
fruitful avenue for further research.
Acknowledgements. We acknowledge the Program for Cli-
mate Model Diagnosis and Intercomparison and the World
Climatic Research Program’s Working Group on Coupled
Modelling for their roles in making available the WCRP
CMIP3 multi-model data set. Support of this data set was pro-
vided by the Office of Science, US Department of Energy. A.
Moise and L. Hanson provided invaluable assistance reading
the data sets and their assistance is gratefully acknowledged.
The constructive reviews provided by K. Hennessy and R.
Colman improved the clarity of the paper.
LITERATURE CITED
Bond HG, McKinnon K, Noar PE (1967) Report on the me-
teorological aspects of the catastrophic bushfires in
south-eastern Tasmania on 7 February 1967. Bureau of
Meteorology, Melbourne
Bureau of Meteorology (1984) Report on the meteorological
171
B1 Scenario
0
0.5
1
1.5
2
C20C B1_2055 B1_2090
A2 Scenario
0
0.5
1
1.5
2
2.5
3
3.5
4
C20C A2_2055 A2_2090
Events per fire season
MPI
MIROC
IPSL
GFDL1
CNRM
GFDL2
MRI
CCM
GISSR
CSIRO
Fig. 11. Number of extreme fire weather frontal events per
season for the climate of the 20th century (C20C) and the
2046–2065 (B1_2055 and A2_2055) and the 2080–2100 peri-
ods (B1_2090 and A2_2090). Upper panel shows projections
for the B1 emissions scenario, and the lower panel for the A2
scenario. The heavy lines are for the 3 close models, the
dashed lines are for the 3 distant models and the thin solid
lines are for the 4 satisfactory models

Clim Res 39: 159–172, 2009
aspects of the Ash Wednesday fires – 16 February 1983.
Bureau of Meteorology, Canberra
Bureau of Meteorology (2005) Meteorological report on the
Wangary and Black Tuesday fires Lower Eyre Peninsula,
10–11 January 2005. Bureau of Meteorology, South Aus-
tralian Regional Office, Kent Town
Cheney P, Gould J, McCaw L (2001) The Dead Man Zone: a
neglected area of firefighter safety. Aust For 64:45–50
CSIRO (Commonwealth Scientific and Industrial Research
Organisation), Bureau of Meteorology (2007) Climate
change in Australia: technical report 2007. CSIRO Marine
and Atmospheric Research, Aspendale, Victoria. Avail-
able from the Australian Greenhouse Office, www.climat-
echangeinaustralia.gov.au
Hasson AEA, Mills GA, Timbal B, Walsh K (2008) Assessing
the impact of climate change on extreme fire weather in
southeast Australia. Centre for Australian Weather and
Climate Research, Melbourne. CAWCR Tech Rep 007, avail-
able at www.cawcr.gov.au/publications/technicalreports.
php
Hennessy K, Lucas C, Nicholls JB, Suppiah R, Ricketts J (2005)
Climate change impacts on fire-weather in south-east
Australia. CSIRO and Bureau of Meteorology, Canberra
IPCC (Intergovernmental Panel on Climate Change) (2000)
Emissions scenarios. Cambridge University Press, Cam-
bridge
IPCC (2001) Summary for policymakers. In: Houghton JT,
Ding Y, Griggs DJ, Noguer M,Van Der Linden PJ, Xioaosu
D (eds) Climate Change 2001: the scientific basis. Contri-
bution of Working Group I to the Third Assessment Report
of the Intergovernmental Panel on Climate Change. Cam-
bridge University Press, Cambridge
IPCC (2007) Climate Change 2007: the physical science basis.
Contribution of Working Group I to the Fourth Assess-
ment Report of the Intergovernmental Panel on Climate
Change, Cambridge University Press, Cambridge
Kallberg P, Berrisford P, Hoskins B, Simmons A, Uppala S,
Lamy-Thepaut S, Hine R (2005) ERA-40 Atlas, ERA-40
Project Report Series, No 19. European Centre for
Medium Range Weather Forecasts, Reading
Kalnay E, Kanamitsu M, Kistler R, Collins W and others (1996)
The NCEP/NCAR 40-year reanalysis project. Bull Am
Meteorol Soc 77:437–471
Lucas C, Hennessy K, Mills G, Bathols J (2007) Bushfire
weather in southeast Australia: recent trends and pro-
jected climate change impacts. Consultancy Report pre-
pared for the Climate Institute of Australia. Bushfire Coop-
erative Research Centre, East Melbourne
Luke RH, McArthur AG (1978) Bushfires in Australia. Aus-
tralian Government Publishing Service, Canberra
McLeod RN (2003) Inquiry into the operational response to
the January 2003 bushfires in the ACT. Department of
Urban Services, Canberra
Mills GA (2005a) A re-examination of the synoptic and
mesoscale meteorology of Ash Wednesday 1983. Aust
Meteorol Mag 54:35–55
Mills GA (2005b) On the sub-synoptic scale meteorology of
two extreme fire weather days during the Eastern Aus-
tralian fires of January 2003. Aust Meteorol Mag 54:
265–290
Mills GA (2008) Abrupt surface drying and fire weather. Part
1: overview and case study of the South Australian fires of
11 January 2005. Aust Meteorol Mag 57:299–309
Perkins SE, Pitman AJ, Holbrook NJ, McAneney J (2007)
Evaluation of the AR4 climate models’ simulated daily
maximum temperature, minimum temperature, and pre-
cipitation over Australia using probability density func-
tions. J Clim 20:4356–4376
Pitman AJ, Narisma GT, McAneney J (2007) The impact of
climate change on the risk of forest and grassland fires in
Australia. Clim Change 84:383–401
Whetton P, Macadam I, Bathols J, O’Grady J (2007) Assess-
ment of the use of current climate patterns to evaluate
regional enhanced greenhouse response patterns of cli-
mate models. Geophys Res Lett 34:L14701, doi: 10.1029/
2007GL030025
Wilks DS (1995) Statistical methods in the atmospheric sci-
ences. International Geophysics Series, Vol 59. Academic
Press, San Diego, CA
Williams AAJ, Karoly DJ, Tapper N (2001) The sensitivity of
Australian fire danger to climate change. Clim Change
49:171–191
172
Editorial responsibility: Balaji Rajagopalan,
Boulder, Colorado, USA
Submitted: November 22, 2008; Accepted: May 26, 2009
Proofs received from author(s): August 3, 2009
➤
➤
➤
➤
➤