Mahlstein et al. (2011). Early onset of significant local warming in low latitude countries

1. Introduction

Detecting and attributing a climate change signal using
the optimal fingerprinting method which includes signal to
noise considerations globally and regionally (Barnett and
Schlesinger

, Hegerl et al

, Santer et al

has been done for a number of variables and on different
regional scales (Stott and Tett

1998

Such studies have

shown that attribution of surface warming is now possible
not only on global but also on continental scales and few
studies exist focusing on smaller scales (Stott et al

, Stott

, Gillett et al

, Zhang et al

), and improved

approaches have allowed attribution not only of warming but

3

Address for correspondence:

Universit¨atstrasse 16, 8092 Z¨urich,

Switzerland.

also of some other climate-related variables and impacts such
as regional rainfall patterns and wildfires (Zhang et al

Santer et al

, Gillett et al

The anthropogenic

trends are superimposed on natural fluctuations, and the ratio
of the two depends on spatial and temporal scales.

For

example, while the largest anthropogenic warming is expected
at high latitudes, the largest variability also occurs there. Here
we use climate model output to approach the question of
the emergence of significant anthropogenic climate changes
from a local and scenario independent point of view in a
manner that complements global attribution studies: taking
account of spatial and temporal variability, we evaluate at what
global temperature increase a particular place experiences a
significantly different temperature regime compared to early
20th century conditions. Large-scale detection and attribution

1748-9326/11/034009+06

$33.00

Printed in the UK

1

Environ. Res. Lett. 6 (2011) 034009

I Mahlstein et al

studies rely on regions with a high signal to noise ratio in
order to detect an anthropogenic warming signal as early as
possible (Hansen et al

1998

, Baettig et al

, Giorgi and

Bi

2009

, Williams et al

, Barnett and Schlesinger

1987

,

Karoly and Wu

2005

, Madden and Ramanathan

1980

, Wigley

and Jones

1981

For the first time we consider both the

‘noise’ of local variability and the ‘signal’ of global warming
to specifically identify the locations and seasons where new
temperature regimes emerge. We show that tropical regions
display the earliest significant warming based on current
models, broadly consistent with available data.

Such an

analysis is important not only for understanding how humans
experience climate change in different locations, but also
because a range of climate impacts (e.g. on ecosystems) can
generally be expected to increase as climate changes exceed
past variability.

Ecoregions in the tropics are likely to be

particularly vulnerable to anthropogenic climate change, as
a smaller increase in temperature will lead to larger future
extremes in tropical ecoregions based on an A1B scenario
as shown by Beaumont et al (

2011

) and may lead to food

insecurity (Battisti and Naylor

2009

2. Results and discussion

Twenty-three atmosphere ocean general circulation models
(AOGCM) are used in the analysis; only one run from
each model is used, to avoid heavier weights for models
that ran ensembles. As will be shown, a key focus is the
low latitudes in the summer season.

A subset of models

was also chosen according to the quality of their ability
to simulate tropical variability (see supplementary material
available at

iop.org/ERL/6/034009/mmedia

). However,

the results are not changed significantly by the selection
of models (see supplementary material (available at

stacks.

) for a comparison of results

from the two model sets) and therefore all results shown are
based on all models available. The models are compared to
gridded global observations for the period from 1900–99 from
the NASA analysis (GISTEMP, see

http://data.giss.nasa.gov/

gistemp/references.html

and section

stacks.iop.org/ERL/6/034009/

). Note that the results

of GISTEMP are consistent with other observational datasets
(see supplementary available at

mmedia

The noise of interannual variability and the signal of

anthropogenic warming are calculated separately for each grid
cell (all models are regridded to a common grid of T42).
The spatial scale of a grid cell is small enough to detect
local changes, because (unlike rainfall) regional temperature
variations extend over areas bigger than a single grid cell
(Mahlstein and Knutti

2010

, Jones et al

1997

). One measure

of the noise in summer surface temperature (TAS) is the
interannual summer variability, shown in the top panel of
figure

(see section

). Note that the left column of figure

shows the model results whereas the right column represents
the observational data (which is gridded at a resolution of
2

.

8

◦

×

2

.

8

◦

).

The summer (defined as the three warmest

consecutive months during the time period 1900–50) signal of
anthropogenic warming is indicated by the difference between

1990–99 and 1900–99, as shown in the middle panel of
figure

. Models indicate that the warming signal is more

pronounced over land than ocean, and is stronger in the
Northern high latitudes, the US and the subtropics.

But

the noise of interannual variability of surface temperature
of the models is also highest in the higher latitude land
areas and lowest in the tropics (panel (a)).

Figure

(c)

indicates that many middle and high latitude places showing
a strong absolute summer warming signal (including the high
latitudes and much of the US) are characterized by a low
warming to variability ratio, due to large local variability.
During the winter season, variability is higher at middle
and high latitudes, leading to lower warming to variability
signals in nearly all regions despite higher warming; thus we
focus here on summer.

The low latitudes, with very low

interannual variability, contain large areas where people live
that show high warming to variability ratios (particularly in
many coastal and island regions). A region of high warming
to variability is also obtained in the data and models in
some uninhabited or sparsely inhabited parts of the Southern
Ocean.

For annual average conditions, similar results to

those of summer are obtained in the tropical regions, but the
high latitudes show even lower warming to variability ratios
(see supplementary material available at

stacks.iop.org/ERL/6/

034009/mmedia

). The observed summer variability is broadly

similar to that obtained in the models in many regions, with
notable exceptions such as the behaviour obtained in Australia
and southern Africa. Note that the observed local variability
is smaller than that generally found in the models, as noted in
previous studies (Gleckler et al

2008

). Generally the patterns

in the observations are less smooth than those obtained in the
models, but this is not surprising since the observations provide
only one realization, whereas the model results are based on
more than twenty realizations. The warming to variability ratio
of the models compares well with the observations over most
regions with higher warming to variability ratios (greater than
1), although differences are seen over parts of the Atlantic,
northern North America, and near the Antarctic Peninsula.

The results are robust across all models considered, the

only exception being the northwest coast of South America
(probably due to differences in how models represent the
topography of the Andes). Note that the qualitative results are
not sensitive to the time periods used in the 20th century. For
example, the patterns do not change if data are restricted to
the period after 1950 (see supplementary material available at

The signal to noise ratio is an important parameter for

climate impacts. For example, changes in climate in a location
with large variability may have less severe impacts on those
ecosystems and species that are adapted to large changes on
short timescales (Williams et al

). In contrast, species

with limited ability to adapt to environmental changes such as
some tropical plants, insects, and reptiles (Deutsch et al

2008

,

Sinervo

2010

) may be quite vulnerable to climate change.

Changes in the mean temperature are relevant to ecosystems as
many species show climate change impacts due to the warming
trends of the past (Parmesan and Yohe

2003

The increase in global temperature required for the signal

of change to emerge from the noise can be estimated by

2

Environ. Res. Lett. 6 (2011) 034009

I Mahlstein et al

Figure 1. (a)–(c) are based on model results and show (a) interannual variability (noise) of summer average surface temperature (TAS) in

◦

C,

(b) signal of summer TAS trend (

◦

C), calculated as the difference between 1990–99 and 1900–99, (c) warming to variability ratio of the

summer trend to the summer variability. (d)–(f) are the same as (a)–(c) but for the observations (GISTEMP). Note that there are small regions
outside the colour scale on (e) and (f).

comparing the TAS distributions of different time periods.
Two samples of summer or warm season temperatures of two
different time periods are significantly different when they do
not originate from the same distribution at the 95% confidence
level (see section

and supplementary material (available

at

) for three different

statistical tests used, and the robustness of results among
them). Figure

depicts the mean global temperature increases

for which 80% of the models show a significant summer
warming over different countries (see section

). The pattern of

the results is similar to the interannual warming to variability
ratio shown in figure

(c).

During summer, a number of

countries in the low latitudes (e.g. Indonesia, parts of the
Middle East and Central America, and large parts of Africa)
should already be experiencing significant changes based upon
this analysis. These results are robust to how the interannual
variability is characterized and to the rate of warming in
models (see supplementary material available at

stacks.iop.

org/ERL/6/034009/mmedia

). Figure

also demonstrates that

for many locations worldwide, less than 1

◦

C global annual

mean warming compared to 1900–29 is sufficient for local
summer warming to exceed natural variability. A constant

composition or ‘commitment’ scenario (IPCC

) projects

a global average warming of about 1

.

2

◦

C, but studies of

economically plausible scenarios predict a warming of near
2

◦

C or more in the long-term (Rogelj et al

2010

, Meinshausen

et al

2009

), which would cause significant summer changes

worldwide. Thus, these results demonstrate that the globe is
committed to significant local changes in summer TAS in all
countries, with the earliest emergence occurring in tropical
regions.

The question of whether a significant local warming

has already been measured is complicated by the fact that
many local high quality observational records extend only to
about 1950. Data are also sparse in many tropical countries.
Therefore, when working with limited observations the local
detectability may only be achieved with a larger global
warming as part of the signal is already included in the time
series which is used for testing. If records are even shorter, this
would further delay the point when changes can be identified
as significant.

However, the pattern of the results is not

sensitive to changes in the baseline time period or the transient
climate response of the scenario, nor is the pattern significantly
changed when annual means are used instead of summer means

3

Environ. Res. Lett. 6 (2011) 034009

I Mahlstein et al

Figure 2. The map shows the global temperature increase (

◦

C) needed for a single location to undergo a statistically significant change in

average summer seasonal surface temperature (TAS), aggregated on a country level. The black line near the colour bar denotes the committed
global average warming if all atmospheric constituents were fixed at year 2000 levels. The small panels show the interannual summer-season
variability during the base period (1900–29) (

±

2 standard deviations shaded in grey) and the multi-model mean summer surface temperature

(red line) of one arbitrarily chosen grid cell within the specific country. The shading in red indicates the 5% and 95% quantiles across all
model realizations.

(see supplementary material available at

stacks.iop.org/ERL/

6/034009/mmedia

). One recent observational study (Collins

2011

) suggests that significant changes are indeed emerging in

parts of tropical Africa, especially in summer.

Figure

shows the warming to variability of the

summer warming on an aggregated country level versus the
emission per capita for the countries shown in figure

http://www.iea.org/publications/

.

The striking result is that the two happen to be anti-
correlated.

Most of the countries showing the most

significant changes at a small global warming have low
CO

2

emissions per capita (

free new Desc.asp?PUBS ID=2143

In contrast, some of

the largest emitters in the past and present are among the
least affected countries thus far as measured by warming to
variability.

In short, those countries affected most by the

warming are not the ones that are the most responsible for it
(Schneider

2003

). The fact that locally significant warming

emerges first in countries with low emissions has no underlying
economic or societal cause. To estimate the risks associated
with climate changes, the local impacts, vulnerability and
adaptive capacity need to be considered. It is frequently stated

that the low latitude countries are the most vulnerable due to
limited financial resources and hence low adaptive capacity, but
here we have documented a key aspect of tropical vulnerability
that is physically based (rather than linked to the economical
issues in these countries): the size of the warming signal
relative to interannual noise.

3. Methods

3.1. Climate model data

The climate model data are available from the World
Climate Research Programme (WCRP) Coupled Model
Intercomparison Project Phase 3 (CMIP3) (Meehl et al

All data are regridded to a T42 grid, and for each model only
the first run available in the CMIP3 archive is used for the
analysis.

3.2. Observations

The gridded temperature data from the NASA analysis
(GISTEMP, see Hansen et al

2006

and

http://data.giss.nasa.

4

Environ. Res. Lett. 6 (2011) 034009

I Mahlstein et al

Figure 3. Warming to variability ratio (local summer temperature
signal relative to interannual summer temperature variability) for
each country versus CO

2

emissions per capita in year 2009.

gov/gistemp

) are shown in figure

. The data were gridded

to T42 resolution using the provided Fortran program. Since
GISTEMP only provides the anomalies, the summer season
at each grid point was determined using the NCEP/NCAR
reanalysis surface temperatures (Kalnay et al

1996

) during

1950–60.

3.3. Interannual summer variability

The interannual summer variability is estimated by first
linearly detrending the model and data values for the time
period 1900–99.

For each year the average summer TAS

is then calculated and the standard deviation of the time
series is obtained. The interannual variability is calculated
individually for each model before averaging across models.
Other approaches to quantifying variability do not change
the basic conclusions (see supplementary material, especially
figure S5, available at

3.4. Significance tests

Whether temperatures of two time periods are significantly
different is tested in several different ways. The primary test
on summer TAS is applied using 30-year moving windows
at 10-year steps, starting with 1900–29 as a baseline and
ending in 2070–99 for the SRES A1B model runs.

This

procedure is applied to each model and grid cell. The results
are not significantly different for different lengths of the
moving window. The local warming is considered statistically
significant (i.e. detected) when a Kolmogorov–Smirnov test
rejects with 95% significance that the samples of the two 30-
year windows are drawn from the same distribution. The last
year of the moving window is taken as the year of emergence
in one model. Changes are considered significant in the year
when the signal is detected in 80% of the models.

This

procedure is done for each grid point. The year is then used to
estimate the corresponding global temperature change based
on the A1B simulation in each model. Figure

averages the

global temperature threshold across countries for comparison
with figure

. Alternative statistical tests considered include

the Student t-test and a simple test which detects a significant
change when the difference of the mean temperature of the
two time windows exceeds twice the interannual variability.
The results are not materially affected by the test applied (see
supplementary material available online at

stacks.iop.org/ERL/

6/034009/mmedia