Transcript Document

The paleo-perspective: what can paleodata tell us about past extremes which
is useful for the future?
Dave Sauchyn
Prairie Adaptation Research Collaborative
University of Regina
National Workshop: Development of Scenarios of Climate
Variability and Extremes: Current Status and Next Steps
Victoria, BC, 16-17 October 2003
Acknowledgements
Collaborators: Dr. Elaine Barrow, Dr. Ge Yu
Graduate Students: Antoine Beriault, Jennifer Stroich
Funding:
Climate is Always Changing
From GSC Misc. Report 71 (2001)
Ice cores, tree rings,
lakes and oceans sediments:
windows on the past
(Leavitt, U of R)
Climatic Variability
A projected increase in climate variability,
including more frequent drought and major
hydroclimatic events, is the most challenging
climate change scenario. Social and biophysical
systems respond to short-term climate
variability and to extreme events long before
they respond to gradual changes in mean
conditions. More extreme climate anomalies
are likely to exceed natural and engineering
thresholds beyond which the impacts of
climate are more severe.
Ron Hopkinson, MSC
IPCC Workshop on Changes in Extreme Weather and
Climate Events: Workshop Report. 11 – 13 June,
2002, Beijing, China, 107 pp.
• assess whether recent changes in the intensity, frequency and
duration of extremes are unusual in the context of instrumental
and proxy records;
• more paleoclimatic records/analyses and proxy indicators of
pre-instrumental extremes,
• palaeo circulation records for the 'recent' 1000-2000 years
should be used to put recent trends and variations in
circulation-related extremes in the context of a longer history
of natural variations
• the available palaeo records and to assess the information they
contain on extremes
• whether indices calculated from model data have realistic
variability
• long coupled control simulations (1000 to as long as 10000
years in length) should be analysed for interannual, decadal
and centennial variations in simulated extremes
Mirror Lake, NWT
14
Mean July Temperature, Tungsten, NWT, 1751-2000
aR2: 39.5%
13
12
11
10
Reconstructed
Observed
11-yr Running Mean
9
1750
1800
1850
1900
1950
2000
Tree-Ring Chronologies
Near Outlook, Saskatchewan, May 2, 2002
Oldman River
Whaleback Ridge
White Spruce, Cypress Hills
Departures From Median Precipitation
June - July Precipitation
Medicine Hat, Alberta, 1754-2001
100
0
-100
-200
-300
1750
1800
1850
1900
1950
2000
600
August - July Precipitation,
Havre, Montana, 1727-2001
400
200
0
-200
1750
1800
1850
1900
1950
2000
Widespread dune activity induced by late 18th century dryness
Wolfe, et al. 2001
Fort Edmonton – HBC Archives
At Edmonton House, a large fire burned “all around us” on April 27th
(1796) and burned on both sides of the river. On May 7th, light canoes
arrived at from Buckingham House damaged from the shallow water.
Timber intended to be used at Edmonton House could not be sent to the
post “for want of water” in the North Saskatchewan River. On May 2nd,
William Tomison wrote to James Swain that furs could not be moved as,
“there being no water in the river.” (Johnson 1967: 33-39, 57)
In 1800 “Fine weather” continued into April at Edmonton House. On
April 18th, James Bird repeated his observation that the poor trade with
both the Slave and Southern Indians was the result of “the amazing
warmness of the winter” diminishing both the bison hunt and creating a
“want of beaver.” Bird reported “clear weather except for the smoke
which almost obscures the sun. The country all round is on fire.” On
June 15th, he noted that the “amazing shallowness of the water”
prevented the shipment of considerable goods from York Factory
(Johnson 1967: 240-248)
Monte-Carlo Probability Analysis
Reconstructed extremes can be compared in only
probability terms. Monte Carlo methods were used to
obtain a normally-distributed random sample of
10,000 errors for each year and thereby produce
10,000 error-added reconstructions (Touchan et al.,
1999; Meko et 2001). Applying sampling error
weighted probability considers the uncertainties in the
observed data
Using Monte Carlo random sampling to obtain erroradded reconstructions enables us to establish the
probability that reconstructed precipitation in any year
or group of years was lower than the record-low of
gauge precipitation.
Probability (x < Xc)
Drought criteria (Xc):
Lowest gauge precipitation
100%
80%CI:
75%
1812
1883 1902 1936
1961
16.4mm
m
50%
25%
0%
1700
1750
1800
1850
1900
1950
50% prob in the
dry-half of the
distribution is
equivalent to 100%
prob in two tails of
the probability
distribution.
2000
15-yr droughts
(running
average)
Probability (x < Xc)
Drought criteria (Xc):
Lowest gauge precipitation
100%
75%
90%CI: 1850, 1851
84.0mm
50%
25%
0%
1700
1750
1800
1850
1900
1950
2000
Markov Chain Monte Carlo Simulation (MCMC)
Markov Chain is a series where the realization of the next element in the
series, Y, is dependent only on the current state, X, and occurs with
probability, P(Y|X). i.e, a model of sequences of events where the
probability of an event occurring depends upon the fact that a preceding
event occurred (Papoulis, 1984).
Seven states were based on the probability distribution of the time
series in order to build Markov Chain.
1). Extreme dry: highest (<1th percentile)
2). Dry: < 10th percentile
3). Dry: < 20th percentile
4). Normal: between 20~80th percentiles
5). Wet: > 80th percentile
6). Wet: > 90th percentile
7). Extreme wet: highest(>99th percentile)
Sampling error- conditional
probability
Simple probability
b. Sample error-estimated 7-class
climate probability
a. 7-class climate probability
178
200
154.39
150
100
`
50
27
0
9
Frequency
Frequency
200
30
3
150
100
50
1
37.89
35.64
0.42 9.86
8.42 1.35
0
0
1
2
3
4
Class
5
6
7
1
2
3
4
Class
5
6
7
The sampling error weights can be considered as the influence on the class probability
through the entire time series Pr(A):
Pr(A) =  Pr(Bi) Pr(A|Bi)
Where  Pr(B) is a probability for event B, Pr(A|B) is a probability of event A under
condition of event B (Papoulis, 1984). Here we apply the sampling error weight as
Pr(Bi), Pr(A|Bi) is the probability of one classe throughout the series conditional on
the yearly sampling errors.
CGCM2 climate simulation for 1000-yr control run
versus proxy precipitation
The CGCM2 1000-year simulation with late 20th century atmospheric
concentration of greenhouse gases (Flato et al., 2000) downloaded from the
IPCC-DDC and the CCCMA webs (http://www.ipcc-ddc.cru.uea.ac.uk and
www.cccma.bc.ec.gc.ca/)
State
Sample/df
Frequency Dry (<1th %)
Dry (<10th %)
Dry (<20th %)
Normal
Wet (>80th %)
Wet (>99th %)
Wet (>90th %)
Mean
Median
Stdev
Skew
Kurt
Code
1
2
3
4
5
6
7
Observation
(sample=1000)
Simulation
(sample=20,000)
1000
0.011
0.089
0.110
0.591
0.100
0.089
0.010
3.987
4
1.054
0.005
0.866
20000
0.012
0.090
0.109
0.585
0.104
0.090
0.010
3.990
4
1.065
-0.021
0.825
T-test
for Mean
df1
df2
alpha
Lower limit
Upper limit
Statistics
Conclusion
F-test
for SD
20998
0.05
0.0627
1.9601
0.59636
H0: M1=M2
Accepted
999
19999
0.05
0.92595
1.07675
0.97981
H0: SD1=SD2
Accepted
Mean and standard deviation from 20,000 iteration MCMC simulation are equal
to mean and SD from TRI chronologies, gauge precipitation and GCM
precipitation simulation at 95% confidence level by T-test and F-test.
MCMC sim ulation for
Pjj (sam ple=20,000)
0.8
0.6
0.6
0.4
0.2
0.0
Probability
MCMC sim ulation for
TRI (sam ple=20,000)
0.8
Probability
Probability
Comparison of 7-state distribution in three climate series for the
Cypress Hills
0.4
0.2
0.0
1 2 3 4 5 6 7
State
MCMC simulation for
CGCM2 CTL-run Pjj
0.8
(sample=20,000)
0.6
0.4
0.2
0.0
1 2 3 4 5 6 7
State
1 2 3 4 5 6 7
State
Steady-state distribution probability:
C PH
TR I
G auge
CGCM 2
S tate 1
0.012336
0.016986
0.011011
S tate 2
0.089502
0.084648
0.089089
S tate 3
0.101555
0.102524
0.110110
S tate 4
0.598476
0.590239
0.590591
S tate 5
0.097218
0.102227
0.100100
S tate 6
0.088818
0.085537
0.089089
S tate 7
0.012095
0.017839
0.010010
The similarity among the probabilities suggests that the GCM modeling has simulated a
similar distribution to the real climate; probabilities of extreme dry (States 1) and
extreme wet (State 7) are slight smaller in GCM than the gauge precipitation or tree-ring
reconstructions, suggesting the GCM CTRL_run insufficiently simulates the two tails of
the distribution of events.
Spectral analysis
Cypress Hills
GCM
Spectrum
Log10 Spectrum
4.0
3.5
3.0
2.5
1
10
1000
600
400
200
0
0.0
0.1
Spectrum
3.5
3.0
0.2
0.3
0.4
0.5
0.3
0.4
0.5
0.4
0.5
Pjj
5000
T_max = 13.1 yr
4000
3000
2000
1000
0
2.5
1
10
100
0.0
1000
TRI
0.1
3.0
2.5
0.2
TRI
1000
Spectrum
3.5
Log10 Spectrum
T_max = 90.9 yr
T_2nd_max = 38.5 yr
T_3rd_max = 20.5 yr
Pjj
4.0
Log10 Spectrum
100
GCM
1200
1000
800
T_max =24.8 yr
T_2nd_max = 125.0 yr
T_3rd_max = 13.1yr
800
600
400
200
0
2.0
1
10
Period (yr)
100
1000
0.0
0.1
0.2
0.3
Frequency (1/yr)
For 20~25-yr and 10~13 yr timescales, the GCM model is consistent with TRI
and gauge records, but differs in longer timescale of hundred years.
GFDL-R15 climate simulation for the last
250-years
• MODEL GFLD (Geophysical Fluid Dynamics Laboratory, USA) is a
coupled ocean-atmosphere GCM, i.e. R15L9 (atmosphere) and 4degL12
(ocean).
• MODELING The emission scenario “Is92a-GS” was forced on the basis
of the IPCC-IS92a scenario, starting CO2 and aerosol forcing at 1766
levels, and running it through the present (historical equivalent CO2 +
aerosols from 1766 to 1990) and out to year 2065. A control integration
is also performed keeping concentrations of sulfate and carbon dioxide
fixed at 1765 levels (Haywood et al., 1997). The simulations investigate
changes in surface air temperature, hydrology and the thermohaline
circulation due to the radiative forcing of anthropogenic greenhouse
gases and sulfate aerosols in the GFDL coupled ocean-atmosphere
model.
Cypress Hill: 1754-2001AD
50.02~49.5N and -110~-110.72W
Precipitation (mm)
400
300
200
100
0
1750
1800
TRI
Gauge
1850
GFDL-Is92a-GS
1900
1950
2000
10 per. Mov. Avg. (GFDL-Is92a-GS)
• Simulations of the mean and maximum June-July precipitation are
much higher (ca. 70mm and 200mm) than both observed and proxy
climate due to model bias.
• The range of reconstructed precipitation is smaller (ca. 80mm) than the
observed value because regression methods causes the reconstruction
to be biased towards the calibration-period mean.
Timing/frequency of drought
Comparison of Drought Years (Criteria < 0.1 percentile)
400
1.4
1.2
1.0
200
0.8
100
0.6
0
0.4
-100
-200
1754
0.2
0.0
1779
1804
1829
1854
1879
GCM departure
1904
1929
1954
1979
TRI
Three drought years (1883, 1924 and 1980) in the modeling
are consistent with the tree-ring derived observations.
TRI
Precipitation (mm)
300
Comparison of 5-yr moving droughts
(Criteria < 0.1 percentile)
1.1
120
80
0.9
0.1 percentile
TRI
Precipitation (mm)
160
40
0
0.7
-40
0.1 percentile
GCM departure
1994
1974
1954
TRI
Comparison of 10-yr moving droughts
(Criteria < 0.1 percentile)
160
1.1
120
0.1 percentile
0.9
TRI
80
40
0
0.7
-40
0.1 percentile
GCM departure
TRI
1984
1974
1964
1954
1944
1934
1924
1914
1904
1894
1884
1874
1864
1854
1844
1834
1824
1814
1804
1794
1784
1774
0.5
1764
-80
1754
Precipitation (mm)
1934
1914
1894
1874
1854
1834
1814
1794
1774
0.5
1754
-80
(3) Spectral analysis
GCM
3.5
2.5
10
4000
100
1000
0.0
0.1
Pjj
4.5
3.5
0.2
0.3
0.4
0.5
0.3
0.4
0.5
0.4
0.5
Pjj
5000
Spectrum
Log10 Spectrum
T_max = 29.1yr
T_2nd_max = 49.7 yr
T_3rd_max = 14.9 yr
8000
0
1
T_max = 13.1 yr
4000
3000
2000
1000
0
2.5
1
10
100
0.0
1000
TRI
0.1
2.5
0.2
TRI
1000
Spectrum
3.5
Log10 Spectrum
GCM
12000
Spectrum
Log10 Spectrum
4.5
T_max =24.8 yr
T_2nd_max = 125.0 yr
T_3rd_max = 13.1yr
800
600
400
200
0
1.5
1
10
Period (yr)
100
1000
0.0
0.1
0.2
0.3
Frequency (1/yr)
Low-Pass Gaussian Filter (p>10 yr for 50% frequency response) show the
periods that for 13~15 yr timescales, the model is found to be consistent with
TRI and gauge records, but differs in timescale longer than 2 decades.
Comparison of GDFL-Is92a-GS historical climate
modeling with proxy climate observations
• A T-test suggests that modeled precipitation means are not equal to the TRI or
gauge data at a 0.95 confidence level. Correlation between the model and
observation is poor and F-test suggests that the modeled standard deviation is
not equal to those of the TRI or gauge data at a 0.95 CI. The model simulates
more variability than the observations.
• There is little agreement in the timing of annual precipitation in lowest 10th
percentile. There is more agreement when the drought criterion is relaxed to
the 20th percentile. These results suggest that the model poorly simulates the
most extreme events.
• There are obvious peaks of ca. 13 year period in spectral power of the TRI
and gauge series. The simulated series also has strong spectral power of this
period with 2 years. The strongest power at ca. 23 year period can be found
in both observation and modeling with 2 years. The power spectrum for the
simulated series diverges from the spectra for the TRI and gauge data at
timescales longer than the two decades.
Paleo Data (Products)
raw proxy data
filtered data (signal)
paleoclimatic and paleoenvironmental records
trends, variability,
frequencies, probabilities
temporal analogues
climate change
and impact scenarios