Photo: F. Zwiers Assessing Human Influence on Changes in Extremes Francis Zwiers, Climate Research Division, Environment Canada Acknowledgements – Slava Kharin, Seung-Ki Min , Xiaolan Wang, Xuebin Zhang, Bill Hogg Photo: F. Zwiers
Photo: F. Zwiers
Assessing Human Influence on
Changes in Extremes
Francis Zwiers, Climate Research Division, Environment Canada
Acknowledgements – Slava Kharin, Seung-Ki Min , Xiaolan Wang, Xuebin Zhang, Bill Hogg
Photo: F. Zwiers
• Introduction• Some approaches• Can climate models simulate
extremes?• What changes are projected?• Have humans influence on extremes?• Conclusions
Outline
Photo: F. Zwiers
• Language used in climate science is not very precise– High impact (but not really extreme)– Exceedence over a relatively low threshold
• e.g., 90th percentile of daily precipitation amounts– Rare events (long return period)– Unprecedented events (in the available record)
• Space and time scales vary widely– Violent, small scale, short duration events (tornadoes)– Persistent, large scale, long duration events (drought)
What is an extreme?
Photo: F. Zwiers
Simple Indices
Simple indices• Examples include
– Day-count indices• eg, number of days each year above 90th percentile
– Magnitude of things like warmest night of the year• Easily calculated, comparable between locations if the underlying
data are well QC’d and homogenized– ETCCDI and APN have put a lot of effort into this
• Peterson and Manton, BAMS, 2008• http://cccma.seos.uvic.ca/ETCCDI/
• Can be analysed with simple trend analysis techniques and standard detection and attribution methods
• Have been used to– Assess change in observed and simulated climates– Understand causes of observed changes using formal detection
and attribution methods
Indices of temperature “extremes”
Alexander, Zhang, et al 2006
DJF Cold nights Trend in frequency Tmin below 10th percentile
JJA Warm days Trend in frequency Tmax above 90th percentile
Photo: F. Zwiers
Extreme value theory
Photo: F. Zwiers
Extreme value theory• Statistical modelling of behaviour of either
– Block maxima (eg, the annual extreme), or– Peaks over threshold (POT, exceedances above a high
threshold)
• Relies on limit theorems that predict behaviour when blocks become large or threshold becomes very high– A familiar limit theorem is the Central Limit Theorem
• Predicts that sample average Gaussian distribution– Similar limit theorems for extremes
• Block maxima Generalized Extreme Value distribution• Peaks above a high threshold Generalized Pareto
Distribution
• Used to estimate things like long-period return values– Eg, the magnitude of the 100-year event
• Can be used to – Learn about climate model performance– Identify trends in rare events (e.g., 10- or 20-yr event)– Account for the effects of “covariates”
• New research is venturing into detection and attribution – Fully generalized approach is not yet available
Extreme value theory …
Photo: F. ZwiersPhoto: F. Zwiers
Can climate models simulate extremes?
Zonally averaged 20-yr 24-hr precipitation extremesRecent climate - 1981-2000
Reanalyses (black, grey)CMIP3 Models (colours)
Kharin et al, 2007
Zonally averaged 20-yr 24-hr temperature extremesRecent climate - 1981-2000
Reanalyses (black, grey)CMIP3 Models (colours)
Kharin et al, 2007
Photo: F. Zwiers
What changes are projected?
Expected waiting time for 1990 event, 2081-2100
Increase in frequency (for N. America) B1: ~66% (33% - 166%) A1B: ~120% (66% - 233%) A2: ~150% (80% - 300%)
20-years
10-years
5-years
Projected waiting time for late 20th century 20-yr 24-hr precipitation extremes circa 2090
Kh
ari
n e
t a
l, 2
00
7
Projected change in 20-yr temperature extremes
°C
10
8
6
4
2
1
Kharin et al, 2007
A1B~2090 vs ~1990
20-yr extreme annual maximum
temperature
20-yr extreme annual minimum
temperature
Have humans influenced extremes?
Photo: F. Zwiers
Changes in background state related to extremes
• Regional mean surface temperature – Global, continents, many
regions– Area affected by European 2003
heatwave (Stott et al, 2004)– Tropical cyclogensis regions
(Santer et al, 2006; Gillett et al, 2008)
• Global and regional precipitation distribution (Zhang et al, 2007; Min et al 2008)
• Atmospheric water vapour content (Santer et al, 2007)
• Surface pressure distribution (Gillett et al, 2003, 2005; Wang et al, 2009)
RO
BE
RT
SU
LLIV
AN
/AF
P/G
etty
Im
ages
scrapetv.com
Detection of human influence on extremes• Temperature
– Potential detectability (Hegerl et al, 2004)
– In observed surface temperature indices (Christidis et al, 2005; Brown et al, pers. comm., others)
• Precipitation– Potential detectability (Hegerl,
et al, 2004; Min et al, 2009)• Drought
– In area affected based on a global PDSI dataset (Burke et al, 2006)
• Extreme wave height– In trends of 20-yr events
estimate used a downscaling approach (Wang et al, 2008)
HadSLP2 hindcast
2
0
-2
Simulated (9 models)
0.8
0
-0.8
Trend in 20-yr extreme SWH(1955-2004)
cm/yr
cm/yr
Wang et al, 2009
AR4 basis for assessment
Current status
Formal study
Formal study
Expert judgement ??
Expert judgementGlobal precip and
water vapour results
Formal study
Expert judgement Supporting SST detection results
Expert judgement Formal study on waves
• New idea introduced during the IPCC AR4 process• Can’t attribute specific events…• ..... but might be able to attribute changes in the risk of extreme events• Approach to date has been
– Detect and attribute observed change in mean state – Use a climate model to estimate change in risk of an extreme event
• Stott et al (2004) estimated that human influence had more than doubled the risk of an event like the European 2003 heat wave
• Would like to constrain this estimate observationally …
Attributing changes in the risk of extremes …
Schar et al, 2004
Photo: F. ZwiersPhoto: F. Zwiers
Conclusions
Photo: F. Zwiers
Conclusions/Discussion• The evidence on human influence on extremes is beginning to
emerge, albeit it slowly• Pushing into the tails reveals weaknesses in observations, models
and analysis techniques• We have done / are doing the easy stuff on extremes
– Indices (3D space-time optimal detection)– Trends in return values (2D optimal detection)– Bayesian decision analysis approaches
• Concept of attributable risk is extremely useful– Available estimates of attributable risk are currently very limited,
and not observationally constrained• Data will continue to be a limitation• Scaling issues will continue to be a concern
The End
Photo: F. Zwiers