12.4.3 Optimal Fingerprint Methods
The use of “optimal” techniques can increase the detectability of
forced climate changes. These techniques increase the signal-to-noise ratio
by looking at the component of the response away from the direction of highest
internal variability (see, e.g., Hasselmann, 1979, 1997, 1993; North et al.,
1995; see also Box 12.1 on optimal detection and
Appendix 12.1). Several new approaches to the optimal
detection of anthropogenic climate change have been undertaken since the SAR.
We focus on optimal detection studies that use a single pattern of climate change
in the following section. Attribution (see Section 12.1.1),
which requires us to consider several signals simultaneously, will be considered
in Sections 12.4.3.2 and 12.4.3.3.
12.4.3.1 Single pattern studies
Since the SAR, optimal detection studies of surface temperature have been extended
(Hegerl et al., 1997, 2000; Barnett et al., 1999) and new studies of data other
than surface air temperature have been conducted (Allen and Tett, 1999; Paeth
and Hense, 2001; Tett et al., 2000).
Surface temperature patterns
The Hegerl et al. (1996) optimal detection study was extended to include more
recent estimates of internal variability and simulations with a representation
of sulphate aerosols (Hegerl et al., 1997). As in the previous study, different
control simulations were used to determine the optimal fingerprint and the significance
level of recent temperature change. The authors find significant evidence for
a “greenhouse gas plus sulphate aerosol” (GS) fingerprint in the most
recent observed 30-year temperature trends regardless of whether internal variability
is estimated from models or observations. The 30-year trend ending in the 1940s
was found to be significantly larger than expected from internal variability,
but less so than the more recent trends. This work has been extended to include
other models (Figure 12.10a; see also Barnett et al.,
1999: Hegerl et al., 2000), examining whether the amplitude of the 50-year summer
surface temperature trends in the GS simulations is consistent with that estimated
in the observations. In eleven out of fourteen cases (seven models each evaluated
using the fingerprints from the two original models), the model trends are consistent
with observations. The greenhouse gas only simulations are generally not consistent
with observations, as their warming trends are too large. Berliner et al. (2000)
detect a combined greenhouse gas and sulphate signal in a fixed pattern detection
study of temperature changes using Bayesian techniques.
|
Figure 12.10: Comparison between the amplitude of anthropogenic
signals from observed and modelled JJA trend patterns using fingerprints
from two different climate models (ECHAM3/LSG and HadCM2) and data from
five climate models. (a) Comparison of the amplitude of a single greenhouse
gas + sulphate aerosol (GS) signal (expressed as change in global mean
temperature [°C] over 50 years). Results show that a significant GS
signal can be detected in observed trend patterns 1949 to 1998 at a 5%
significance level (one-sided test), independent of which pair of fingerprints
was used. The observed signal amplitude is consistent with contemporaneous
GS amplitudes for most models’ GS simulations. 90% confidence intervals
are shown by solid lines for estimates using ECHAM3/LSG fingerprints and
by dashed lines for estimates based on HadCM2 fingerprints. Cases where
a model’s and the observed amplitude disagree are marked by a cross
on the axis. (b) and (c) show an estimate of the observed amplitude of
a greenhouse gas signal (horizontal axis) and a sulphate aerosol signal
(vertical axis) estimated simultaneously. Both signal amplitudes can be
estimated as positive from observations based on ECHAM3/LSG fingerprints
shown in (b) while only the greenhouse gas signal is detected based on
HadCM2 fingerprints shown in panel (c). The amplitudes of both signals
from the observations are compared with those from model simulations forced
with various forcing histories and using different climate models (1:
HadCM2; 2: ECHAM3/LSG; 3: GFDL; 4: ECHAM4/OPYC; 5: CCCma1; 6: CCCma2).
Simulations with symbols shown in black are consistent with observations
relative to the uncertainty in observations (grey ellipse) and that of
the model simulations (not shown). Simulations which are inconsistent
are shown in grey. Model simulations where only a single ensemble member
is available are illustrated by thin symbols, those based on ensembles
of simulations by fat symbols.
Results from consistency tests indicate that most greenhouse gas only
simulations (G, shown by "x") are inconsistent with observations.
Ten of the GS simulations in both panels are in agreement with observed
trend patterns, discrepancies arise mostly from the magnitude of a sulphate
signal (vertical axis). The failure to detect a sulphate signal as well
as a greenhouse gas signal in panel (c) is due to the two signals being
very highly correlated if only spatial patterns are used- this makes separation
of the signals difficult. These results show that estimates of a sulphate
aerosol signal from observations are model dependent and quite uncertain,
while a single anthropogenic signal can be estimated with more confidence.
All units are in °C/50 year, values in the upper right quadrant refer
to a physically meaningful greenhouse warming and sulphate aerosol cooling
signal. The consistency test establishes whether the difference between
a model’s and the observed amplitude estimate is significantly larger
than the combined uncertainty in the observations (internal variability
+ observational uncertainty) and the model simulation (internal variability).
The figure is derived by updating the data used by Barnett et al. (1999)
(for details of the analysis see Hegerl et al., 2000) and then applying
a simple linear transformation of the multi-regression results (Hegerl
and Allen, 2000).
Results for 1946 to 1995 period used by Barnett et al. (1999) are similar,
except fewer of the models in b and c agree with observations and the
case of both signals being zero in c is not rejected. Simulations of natural
forcing only ending before 1998 are also rejected in that case.
|
Vertical patterns of temperature
Allen and Tett (1999) use optimal detection methods to study the change in the
vertical profile of zonal mean temperature between 1961 to 1980 and 1986 to
1995. Estimated signals from ensemble AOGCM simulations with greenhouse gas
alone (G), greenhouse gas plus direct sulphate (GS), and also including stratospheric
ozone forcing (GSO; Tett et al., 1996) are considered. The G and GSO signals
are detected separately. The amplitude of the GSO fingerprint estimated from
observations is found to be consistent with that simulated by the model, while
the model-simulated response to greenhouse gases alone was found to be unrealistically
strong. The variance of the residuals that remain after the estimated signal
is removed from the observations is consistent with internal variability estimated
from a control run.
Other climatic variables
Schnur (2001) applied the optimal detection technique to trends in a variety
of climate diagnostics. Changes in the annual mean surface temperature were
found to be highly significant (in agreement with previous results from Hegerl
et al., 1996, 1997). The predicted change in the annual cycle of temperature
as well as winter means of diurnal temperature range can also be detected in
most recent observations. The changes are most consistent with those expected
from increasing greenhouse gases and aerosols. However, changes in the annual
mean and annual cycle of precipitation were small and not significant.
Paeth and Hense (2001) applied a correlation method related to the optimal
fingerprint method to 20-year trends of lower tropospheric mean temperature
(between 500 and 1,000 hPa) in the summer half of the year in the Northern Hemisphere
north of 55°N. Greenhouse gas fingerprints from two models were detected.
The combined greenhouse gas plus (direct) sulphate (GS) fingerprints from the
two models were not detected.
Summary
All new single-pattern studies published since the SAR detect anthropogenic
fingerprints in the global temperature observations, both at the surface and
aloft. The signal amplitudes estimated from observations and modelled amplitudes
are consistent at the surface if greenhouse gas and sulphate aerosol forcing
are taken into account, and in the free atmosphere if ozone forcing is also
included. Fingerprints based on smaller areas or on other variables yield more
ambiguous results at present.
