Detection and attribution of climate change Global temperature

1
Detection and attributionof climate change
Global temperature

• Nathan Gillett,
• n.gillett_at_uea.ac.uk

2
Introduction

• The last report of Working Group 1 of the IPCC
  concluded that
• There is new and stronger evidence that most of
  the warming observed over the last 50 years is
  attributable to human activities.
• On what evidence is this conclusion based?

3
Observed temperature change
Updated Jones and Mobert (2003) global mean
surface temperature, defined using station
temperatures and area weighting.
4

• How can we determine whether the observed warming
  is due to human influence or natural climate
  variability?

5
Paleo temperature record
Reconstructions of temperature from the past 1000
years based on paleo-data. The 2s uncertainty
associated with the Mann et al. (1999)
reconstruction is also shown. Taken from IPCC,
2001.
6

• How else could we determine whether the trend
  could have occurred due to internal climate
  variability?
• Definitions
• Internal climate variability Variability due to
  the internal dynamics of the climate system.
• Natural climate variability Variability due to
  internal dynamics and natural external influences
  (stratospheric volcanic aerosol and changes in
  solar irradiance).

7
Simulated internal variability
Global mean surface air temperature anomalies
from 1,000-year control (constant forcing)
simulations with three different climate models,
HadCM2, GFDL R15 and ECHAM3/LSG (labelled HAM3L),
compared to the recent instrumental record
(Stouffer et al., 2000). No model control
simulation shows a trend in surface air
temperature as large as the observed trend. Taken
from IPCC (2001).
8
Is the simulated variability realistic?
Spectra of the variability of global mean
temperature in a range of climate models and
observations. While some (starred) models
underestimate the observed variability, most
provide realistic simulations of internal
variability. The spectrum of observed global
temperature is shown with a linear trend removed
(solid black line) and with an independent
estimate of the response to changes in natural
and anthropogenic forcing removed (dotted line).
Taken from IPCC (2001).
9
Updated spectra
Spectra of observed global mean temperature
compared to simulated spectra of global mean
temperature in model runs with anthropogenic and
natural forcings.
Source Stone et al. (2005)
10

• How do we know whether the observed trends are
  due to natural variability?
• If we assume that volcanic and solar influences
  during the past century have been comparable with
  those in the the past 1000 years, we can use the
  paleo-record.
• Alternatively we can use climate model
  simulations

11
Simulated response to natural forcing
Simulated temperatures from HadCM3 forced with
changes in volcanic aerosol and solar irradiance,
compared with Jones et al. (2001) observations.
Although changes in natural forcing help to
explain early 20th century warming, they cannot
explain the recent warming (Stott et al., 2000).
Taken from IPCC (2001).
12

• What else do we need to test if we are to
  attribute observed changes to anthropogenic
  forcing?

13
Simulated temperatures from HadCM3 forced with
natural (volcanoes and solar) and anthropogenic
forcing (greenhouse gases, sulphate aerosol,
ozone depletion), compared with observations
(Stott et al., 2001). Taken from IPCC (2001).
14
Other models

• Global mean temperature from individual
  simulations of 12 coupled climate models with
  anthropogenic and natural forcing.

Source Stott et al. (2005) and Stone et al.
(2005).
15
Other models

• Top panel shows global mean temperature in 12
  models with all forcings.
• Lower panel shows global mean in models with
  natural forcings only.

Source Stott et al. (2005) and Stone et al.
(2005).
16
Observed trends and trends simulated in response
to anthropogenic and natural influences (Stott et
al., 2001).
17
Testing the significance of the observed changes

• Simulations of the response to anthropogenic
  forcing look consistent with observed changes,
  but how can we test this formally?
• Correlation coefficients Calculate correlation
  between observed temperatures, and the simulated
  response to anthropogenic forcing, then repeat
  with sections of control integration to determine
  significance. Used in early detection and
  attribution studies, but dont give us any
  information about whether the amplitudes of the
  simulated and observed responses are consistent.
• Regression Gives us information about the
  relative amplitudes of simulated and observed
  changes. This is the technique commonly used for
  detection and attribution.

18
Regresssion

• Suppose we represent a vector of observed
  temperatures with y (e.g. 10 5-yr means of
  temperature).
• We use x to represent the simulated response to
  anthropogenic forcing (also 10 5-yr means of
  temperature).
• We may relate the two with the regression
  equation
• y ßx u
• where u represents internal climate variability,
  and ß is regression coefficient which we want to
  estimate.
• We can estimate ß with a least squares fit. If x
  and y are expressed as anomalies

19
Example Tropical Atlantic September SST

• SST observations are compared with HadCM3 all
  forcings simulations.

ß 0.703
20

• Replace observations with a segment of control
  simulation and repeat many times to derive a
  595 uncertainty range on ß.

21
Example Tropical Atlantic September SST

• This gives ß 0.703 0.467
• Since ß is significantly positive, this indicates
  a detectable influence of external forcing on
  tropical Atlantic SST in September.
• This is a simple example of a detection and
  attribution or fingerprinting calculation.

22
Detection and attribution

• If ß is found to be significantly greater than
  zero, we say that climate change is detected.
  This means that the observed change is
  inconsistent with internal variability.
• If after considering all plausible contributions
  to the observed change we find that ß is
  inconsistent with zero and consistent with one
  then we say that the change may be attributed to
  the forcing in question. This means that the
  observed change is also consistent in amplitude
  with the simulated change.

23
Refinements Time-space detection and attribution

• In example plot of observed temperature anomalies
  against simulated temperature anomalies, each
  data point represented a five-year mean
  temperature over the tropical Atlantic.
• But if we have a latitude-longitude grid of
  observed temperature trends and a similar grid of
  simulated temperature trends, we could make a
  similar scatter plot, where each data point
  represented a grid cell.
• We could carry out a regression to estimate ß as
  before.

24
Refinements Time-space detection and attribution

• Or if we had, say, gridded decadal means, we
  could plot one data point for every grid point in
  every decade and carry out the regression.
• This would be known as a time-space detection
  calculation.

25
Refinements Time-space detection and attribution
Source Weaver and Zwiers (2000).
26
Refinements Multiple regression

• We may want to distinguish the contributions of
  more than one forcing to observed change (for
  example natural and anthropogenic influence).
• We can regress the observations on more than one
  pattern at once in a multiple regression.
• If we represent each response pattern as a column
  of matrix X, then

27
Refinements Multiple regression

• ß is now a vector of scaling factors, which we
  can estimate by

28
Refinements Signal-to-noise optimisation

• We want to maximise our chances of distinguishing
  the response to climate forcing from internal
  variability.
• Therefore we can give less weight to patterns
  with large internal variability and more weight
  to patterns with small internal variability.
• This is called signal-to-noise optimisation, and
  the resulting regression is called an optimal
  regression.
• Analogy If we wanted to understand a voice on a
  bad radio link, we might amplify those
  frequencies we know have little noise, and damp
  those frequencies with large noise.

29
Refinements Signal-to-noise optimisation

• Using signal-to-noise optimisation the best
  estimator of ß is
• where C is a covariance matrix, representing
  the internal variability in the pattern
  considered.

30
Detection in global tempeature
Estimates of the regression coefficients ß for a
detection analysis based on global decadal mean
temperature anomalies 1946-1996 and a range of
models. Large scale (gt5000 km) spatial
information was also retained in the analysis.
Although results vary depending on the model
used, a greenhouse gas influence is detected in
every case. Taken from IPCC (2001).
31
Uncertainties

• Observational uncertainty Although there have
  been many studies of this separately, it has not
  been included in detection and attribution
  studies to date.
• Sampling uncertainty in model response Internal
  climate variability contributes to the
  uncertainty in the simulated response pattern to
  a forcing. This may be reduced by averaging over
  an ensemble of simulations, each with slightly
  perturbed initial conditions. This source of
  uncertainty has also been explicitly included in
  detection and attribution studies.
• Model uncertainty Climate models have
  uncertainties associated with model structure and
  uncertainties in model parameters. The latter
  source of uncertainty is being explored by the
  climateprediction.net experiment.

32
Summary

• The attribution of observed global temperature
  change to human influence requires
• Global temperature observations.
• A good estimate of internal climate variability.
• Simulations of the response to anthropogenic and
  natural forcings.

33
Summary

• Optimal detection and attribution uses these
  three elements
• To test whether observed climate change is
  consistent with internal variability, and if not
• To attribute observed change to external climate
  influences.