Operational attribution |
Myles Allen | |
Department of Physics, University of Oxford | |
myles.allen@physics.ox.ac.uk | |
With thanks to: | |
Pardeep Pall & D‡ith’ Stone, University of Oxford | |
Peter Stott, The Met Office |
Why should we develop operational attribution? |
Policy-makers, lawyers, the public and the media: | ||
ÒWas this flood/drought/... due to climate change/ENSO/É?Ó | ||
No consensus on response: question is ambiguous. | ||
Essential element in costing climate change impacts. | ||
The insurance industry and its customers: | ||
Currently withdrawing from risks that may be influenced by climate change, even risks that may in fact be insurable. | ||
Understanding of the origins of current risks allows rational choices of what risks we accept in a changing climate. | ||
The scientific community: | ||
Because we want to know what is going on. |
How can we answer the question ÒIs climate change to blame for this flood?Ó |
Making sense of the question |
First formalize the question as one we can actually answer and convince the questioner that that is what they actually meant. | ||
Treat weather events as indivisible, like deaths. | ||
Use Òattributable riskÓ methods from epidemiology. | ||
P1: probability of event under current conditions. | ||
P0: probability of similar event with external driver removed. | ||
Fraction Attributable Risk FAR = 1 - P0/P1 | ||
Using coupled models to estimate attributable warming. | ||
Statistical methods to estimate FAR (Stott et al, 2004). | ||
GCM-based methods to estimate FAR (Pall et al, 2007). | ||
Extensions to other regions and variables (Hoerling, 2007). |
The European heat-wave of 2003 |
Modelling Southern European area-averaged June-July-August summer temperatures |
How we attribute different contributions to an observed change in expected temperatures |
Expected temperatures are not a random variable. | |
Treat this as a standard regression problem, with contributions from model-simulated patterns of response to different external drivers as the unknown parameters. | |
Conventional likelihoods for uncertainty estimates. | |
Complications arising from noise in model-simulated patterns: use a Total Least Squares or Errors in Variables approach. |
An example: attribution of hemispheric temperature changes with HadCM3 |
An example: attribution of hemispheric temperature changes with GFDL |
An example: attribution of hemispheric temperature changes with PCM |
External contributions to European summer temperatures, relative to pre-industrial |
Inferring external contributions to risk from external contributions to expected temperatures |
Stott et al (2004): assumed internal variability in large-area-averaged temperatures is unaffected by changing mean climate (no evidence to contrary). | |
Fit EV distribution to model-simulated variability, allowing for uncertainty in EV parameters. | |
Subtract possible offsets (change location parameter) due to anthropogenic influence to generate possible temperature distributions. | |
Only applies to variables like large-scale temperatures for which we can defensibly assume no change in variability. |
Changing risks of European
heatwaves (Stott et al, 2004) |
Anthropogenic contribution to the risk of the 2003 heat-wave |
Conclusions of Stott et al, 2004, study of the 2003 European heat-wave |
Human influence very likely increased the risk of such a heat-wave by at least a factor of two, with a best guess factor of 6-10. | |
But this study was based on a relatively coarse-resolution model (~250km grid over Europe), so conclusions only apply to large-scale, seasonally averaged temperatures. | |
Still work to be done for small-scale risks & for variables other than temperature. |
Autumn 2000 floods in the UK |
A template for operational
attribution (Pall et al, 2007) |
Aim: to quantify the role of increased greenhouse gases in precipitation responsible for 2000 floods. | |
Challenge: relatively unlikely event even given 2000 climate drivers and sea surface temperatures (SSTs). | |
Approach: large (multi-thousand-member) ensemble simulation of April 2000 – March 2001 using forecast-resolution global model (80km resolution near UK). | |
Identical Ònon-industrialÓ ensemble removing the influence of increased greenhouse gases, including attributable SST change. | |
Use several coupled models for SST pattern to allow for uncertainty. |
What, in climate research, is a relatively high-resolution model |
Patterns of SST signal removed: 1990s seasonal surface temperatures relative to control |
Large ensembles required, so we use distributed computing: http://attribution.cpdn.org |
Autumn 2000 as observed
(ERA-40 reanalysis)É Éand in one of the wetter members of our ensemble. |
Return-times versus total Sept-Nov precipitation in the ÒAutumn 2000Ó ensemble |
Return-times after removing estimated signal of greenhouse warming |
Return-times after removing estimated signal of greenhouse warming |
Contribution of greenhouse warming to risk of Autumn 2000 3-month precipitation anomalies |
Implications of Pall et al study |
Early studies with coarse-resolution driving models gave diametrically opposing conclusions re: anthropogenic trend in flood risk. | |
Large ensemble simulations with higher-resolution models give more modest, but still substantial, anthropogenic increase in flood risk. | |
We need large ensembles to capture events of interest, and we need relatively high resolution to simulate them realistically: ideal application for novel (ÒgridÓ) computing resources. |
Changing drought risk in Tripoli due to greenhouse gas increase 1900-2000 |
Link to work on attribution of trends in African rainfall (Hoerling et al, 2007) |
Impact of Indian Ocean warming on Southern African rainfall |