n. w. watkins with d. a. stainforth , s. c. chapman · 3 papers •chapman, stainforth, watkins,...
TRANSCRIPT
Quantifying the time dynamics of the full local
distribution of daily (temperature &) precipitation
and its uncertainties
N. W. Watkins1,3,5 with
D. A. Stainforth3,4,1, S. C. Chapman1,2
1CFSA, Physics, Univ. of Warwick; 2Mathematics and Statistics, UIT; 3CATS, LSE, 4Grantham Institute LSE;
5Open University, UK
We acknowledge the E-OBS dataset from the EU-FP6 project ENSEMBLES (http://ensembles-eu.metoffice.com) and the data providers in the ECA&D project (http://eca.knmi.nl)
3 Papers• Chapman, Stainforth, Watkins, Phil. Trans. R.
Soc. A. 2013
• Stainforth, Chapman, Watkins, Environ. Res. Lett., 2013
• Chapman, Stainforth, Watkins, Environ. Res. Lett., 2015
• All at Sandra’s webpage, my Researchgate etc
Temperature
Pap
Temperature
P Precip
Temperature
TEMPERATURE: Is it just hot today, or is it climate change?- why this particular question?
Perception of past climate change as experienced locally
A local phenomenon- varies with geographical location, quantile/variable (less cold days or hotter warm days?)
Even locally, trend in the extremes may not follow trend in the mean (average days are no warmer, but the hottest days are much hotter?)
Is local trend very different to global mean temperature trend?
[Later on, more heavy tailed climate variables (precipitation)]
Planning thresholds- crops, ecosystems, building overheating…
Fundamental statistical constraints- to what extent can we answer this question using only the data?
Is it just hot today, or is it climate change?
Quantify likelihood (of daily temperature) and how it is changingcumulative density function (cdf) C(T)
Measure a daily temperature T
Fraction of days observed to be colder than T
1
0T
Compare two different time periods t1 and t2
C(T,t1) C(T,t2)
Fundamental constraints on the number of observations
Quantify likelihood (of daily temperature) and how it is changing
Fraction of days observed to be colder than T
Only 92 days per summer season
multistation data only from 1950
C(T,t1), C(T,t2) have uncertainties
1
0
T
C(T,t1)C(T,t2) Can we measure how
fast C is changing?
-uncertainties depend on unknown functional form of C!
warming
Constraint on number of observations → different samples of C will fluctuate in the absence of time dependence‘random’ error in C
The daily NAO index is constructed by projecting the daily
(00Z) 500mb height anomalies over the Northern
Hemisphere onto the loading pattern of the NAO (shown)
defined as the first leading mode of Rotated Empirical
Orthogonal Function (REOF) analysis of monthly mean500mb height during 1950-2000 period.
NAO-Hurrell et al Science 1995, plots courtesy
UCAR, NOAA
NAO
AO
AMO
Natural oscillations (which may also be changing with the climate): North Atlantic Oscillation (NAO), Arctic Oscillation (AO), Atlantic MultidecadalOscillation (AMO), El Nino… index is deviation from mean pattern
Plotted ~1860-2009
precip
temp
Probability of daily temperature, precip. and how it’s changing-depends on where you live
Distribution of stations in Eobsgridded set
Data 1950-now
Coverage not uniform. Not always the same stations. Data not continuous.
Stations are moved, or city ‘heat islands’ move near them.
Not always the same instrumentation.
Observation constraints→ ‘random’ and systematic errors
pdf P(T,t):Likelihood of T observed in the range T-T+dT
cdf C(T,t):Fraction of days observed to be colder than T
4 locations in E-OBSis C changing with time?at what temp T is it changing?
T
T
0
1
0
<1
<T>
0.5
0.75
0.25
each cdf, pdf is made using 9 successive summers9x92 days in each 1950- now, so 63 years…
Remember this plot you will see it again!
MethodParameterize how different quantiles are changing with time (forcing)
Observe for example daily temperature Tfrom a given location over τ seasons (years) centred on an epoch t (year).
Form cdf C for different epochs giving C(t,T) for each t
Now, cdf C can change with t on timescales that are much longer than τ*separation of timescales*This climate state has some unknown forcing parameter g(t). Then:
C(t,T)≡C(g(t),T)
C(g,T)
T
C(g(t1))C(g(t2))
warming
At t2 we observe a temperature T*> Tq
Does this observation arise from forcing, i.e. the change in C with g(t)?
q
Tq
( , )
set 0
q
q
q
q xg
q C x g
C Cdq dx dg
x g
dq
since the pdf is the differential of the cdf w.r.t.
x this is:
1 (
( ))
)
(q
q
q x
Cdx dg
x
x
PP
C
xg
Provided C can be treated as a smooth, differentiable
function we can write, with
dxq=x*-xq
The change dxq=x*-xq due solely to
forcing is obtained by setting dq=0.
Estimating the change at a quantile (temps)
Now to estimate ∆x and its uncertainty…
N yearly seasons dataper cdf
( )T TQ Q P
Estimate ∆Q=Q(t2)-Q(t1)
Each pair of cdfsgives one estimate
We can repeat this procedure for different t1,t2 to get several estimates of ∆(Q=any quantity that depends on the distribution).
Estimate changes and uncertainties solely from dataConsider some Q(x) that captures a property of the distribution C(x)…
pdf<threshold
1
0 x
C(x,t1)C(x,t2)
time
Use all 10 estimates to quantify uncertainty
TQ
x
Show you “light tailed” example-temperaturedaily mean, max, min
-show you how uncertainties can be due to restrictions on the observations-look at some data(temperatures: E-OBS gridded dataset since 1950, Europe)
STEP
1Aggregate
observations over several consecutive
years. STEP
2
Construct the empirical CDF and PDF for each of the time slices. ST
EP 3
Estimate the change either in a specific threshold or
specific quantile.
STEP
4
The significance of
the trend is estimated
based on the statistical
uncertainty in the input data.
2
0 0
For each 'year' , that is, each 100 values
specify shift, scale parameters ( ), ( ).
Mean , varian
no time
ce
v
,skewness 2 / ,excess Kurtosis 6 / .
(i) : ariation
mean changes:
3; 5
(ii) as (
t x
a t b t
ab ab b b
a a b b
0
0
i) with shift 5 /100
(iii) ( ) [1 1 / 2( /100)]; consmean, variance change:
mean, variance, skew, kurtosis change
tant
(iv) ( ) [1 1 / 2( /100)]; constant:
x x t
a t a t b
b t b t a
Test how well this works:Illustrativemodel dataSame size sample as actual data‘fair dice’
cdf
time
( ) 1 ( )
( )
1( , )
( ) ( ( ))
x
a t b t
a tP x t x e
b t a t
Gamma distributed iid
Technical details:
Change dTq=T*-Tq
Shown:∆C=C(t=80)-C(t=20)
P all ‘years’, estimate:
cdf
( )
(
0
)
)
0
(P T
T
C
C
T
constant varies , varies , , , variesS k Model data3 ‘years’ data(3 consecutive ‘seasons’)per pdf, cdf
( )
( )q
C
P TT
T
cdf
( )
(
0
)
)
0
(P T
T
C
C
T
Change dTq=T*-Tq
( )
( )q
C
P TT
T
constant varies , varies , , , variesS k
So maybe ok to use 9 ‘years’ data (9 consecutive ‘seasons’) per cdf…
Model data9 ‘years’ data(9 consecutive ‘seasons’)per pdf, cdf
constant varies , varies , , , variesS k
P>0.005
( )
( )q
C
P TT
T
3 years data per cdf
9 years..
Model data10 estimates of ∆TIllustrates likely ‘spread’ in results just due to sample size and rate of convergence
∆T
∆T
Data: E-OBS max daily temperatures at 4 locations
look at four locations at longitude and latitude(i)[4.75 52.25] Leiden Holland (ii) [-4.75 51.75] west Wales (iii) [-4.75 42.75] Leon, north Spain(iv) [11.25 43.75] Florence, Italy.summer aggregate daily temperatures over 3 month intervals within each year, June, July and August, 92 samples per year.
cdf
E-OBS dataset (http://eca.knmi.nl)
cdf
( )
(
0
)
)
0
(P T
T
C
C
T
Change dTq=T*-Tq
Shown:∆C=C(t=2005)-C(t=1955)
P all ‘years’, estimate:
( )
( )q
C
P TT
T
E-OBS data3 ‘years’ data(3 consecutive ‘seasons’)per pdf, cdf
Leiden, NL W-Wales N-Spain Florence, IT
Leiden, NL W-Wales N-Spain Florence, IT
cdf
( )
(
0
)
)
0
(P T
T
C
C
T
( )
( )q
C
P TT
T
Change dTQ=T*-TQ
Shown:∆C=C(t=2005)-C(t=1955)
P all ‘years’, estimate:
E-OBS data9 ‘years’ data(9 consecutive ‘seasons’)per pdf, cdf
Combine the results and uncertainties on one map…2 kinds of uncertainty (i) ‘random’ fluctuations (ii) systematic trends, cycles
big trend, small spread
small trend, small spread
big trend, big spreadsmall trend, big spread
plot the minimum ∆T from the 10 estimates for each T and location
∆C 1955-1995, 1956-1996 …
Leiden, NL W-Wales N-Spain Florence, IT
P>0.005
( )
( )q
C
P TT
T
9 years..
3 years data per cdf
E-OBS data10 estimates of ∆TIllustrates likely ‘spread’ in results due to all data issues
Indication of where results are robust and where not!
∆C 1955-1995, 1956-1996 …
Leiden, NL W-Wales N-Spain Florence, IT
P>0.005
( )
( )q
C
P TT
T
9 years..
3 years data per cdf
E-OBS data10 estimates of ∆TIllustrates likely ‘spread’ in results due to all data issues
Indication of where results are robust and where not!
Use minimum of all 10 samples
Summer daily max temp minimum change at quantile
Summer daily maximum temperature signed smallest abs(∆T) of 10 estimates
• Tendency for upper quantilesto show largest ∆T
• Highest ∆T across a band from northern France to Denmark at the highest quantiles.
• At mid-to-high quantiles, largest ∆T in a band slightly further south.
• Scandinavia- little robust signature.
Winter daily min temp minimum change at quantile
Winter daily minimum temperaturesigned smallest abs(∆T) in 10 estimates
Winter night time temperatures- smaller changes in ∆T EXCEPT:• At the very lowest
quantiles in central western Europe, and
• At low to mid-quantiles in Scandinavia
NOW PRECIP: Estimate change in the extremes-‘heavy tailed’ pdfs, daily precipitation
-look at uncertainties just due to restrictions on the observations-look at some data(temperatures-Eobs gridded dataset since 1950, Europe)
II: Capturing the behaviour of the extremes (precip)mean total season precip above a threshold (on all days wetter than) x=PT-how wet are the downpour days
Change in mean total season precip on all days wetter than PT-how much more rain now falls on downpour days
% change in total … on all days wetter than PT-is that a big change?
Change in fraction of total season precip that occurs on days wetter than PT-all days are wetter? or same total season precip, now in downpours?
PT
( )T TP x x P
Generalizes indices at fixed threshold eg:Alexander, L. V. et al. Global observed changes in daily climate extremes of temperature and precipitation. JGR-Atmos. 111, doi:D05109 10.1029/2005jd006290 (2006).Climdex. Datasets for Indices of Climate Extremes, http://www.climdex.org/indices.html (2015).
TP
% %TT
T
PP
P
( 0)
TT
T T
PF
P P
TP
x
Shifting mean
Shifting mean and changing
shapeChanging
shape
Explore these scenarios:
Can these changes be distinguished in the data?(effect of fundamental constraints/uncertainties)
(a) Change in precip at a quantile
(b) Change in mean total season precip on all days wetter than PT
(c) % change in total season precipon all days wetter than PT
(d) Change in fraction of total season precip that occurs on days wetter than PT
Modelling uncertainties: Gamma distributed iid RV
Use the min of 10 samples..Throw away if not all of one sign(zero spanning values)
(a) Change in precip at a quantile
(b) Change in mean total season precip on all days wetter than PT
(c) % change in total season precipon all days wetter than PT
(d) Change in fraction of total season precip that occurs on days wetter than PT
(a) Change in precip at a quantile
(b) Change in mean total season precip on all days wetter than PT
(c) % change in total season precipon all days wetter than PT
(d) Change in fraction of total season precip that occurs on days wetter than PT
E-OBS daily precipitation4 locations
(i) SW Scotland(ii) Central Scotland(iii) Dordogne(iv) SW Wales
Precipitation: consider wet days only x>1mm
(a) Change in precip at a quantile
(b) Change in mean total season precip on all days wetter than PT
(c) % change in total season precipon all days wetter than PT
(d) Change in fraction of total season precip that occurs on days wetter than PT
E-OBS daily winter precipitation4 locations
(i) SW Scotland(ii) Central Scotland(iii) Dordogne(iv) SW Wales
more total precip
and shift to more intense days
and shift to more intense days
(a) Change in precip at a quantile
(b) Change in mean total season precip on all days wetter than PT
(c) % change in total season precipon all days wetter than PT
(d) Change in fraction of total season precip that occurs on days wetter than PT
E-OBS daily winter precipitation4 locations
(i) SW Scotland(ii) Central Scotland(iii) Dordogne(iv) SW Wales
Not much more total precip
but shift to more intense days
(a) Change in precip at a quantile
(b) Change in mean total season precip on all days wetter than PT
(c) % change in total season precipon all days wetter than PT
(d) Change in fraction of total season precip that occurs on days wetter than PT
E-OBS daily winter precipitation4 locations
(i) SW Scotland(ii) Central Scotland(iii) Dordogne(iv) SW Wales
Clearly identifiable drying
(a) Change in precip at a quantile
(b) Change in mean total season precip on all days wetter than PT
(c) % change in total season precipon all days wetter than PT
(d) Change in fraction of total season precip that occurs on days wetter than PT
E-OBS daily winter precipitation4 locations
(i) SW Scotland(ii) Central Scotland(iii) Dordogne(iv) SW Wales
No clear single trend
(a) Change in precip at a quantile
(b) Change in mean total season precip on all days wetter than PT
E-OBS daily winter precipitationMaps of min of all 10 estimates, zero spanning values not plotted
(a) Change in precip at a quantile
(b) Change in mean total season precip on all days wetter than PT
E-OBS daily winter precipitationMaps of min of all 10 estimates, zero spanning values not plotted
+2mm on a q=0.95 day
-2mm on a q=0.95 day
+40% on all days >15mm
(a) % change in total season precip on all days wetter than PT
(b) Change in fraction of total season precip that occurs on days wetter than PT
E-OBS daily winter precipitationMaps of min of all 10 estimates, zero spanning values not plotted
(a) % change in total season precip on all days wetter than PT
(b) Change in fraction of total season precip that occurs on days wetter than PT
E-OBS daily winter precipitationMaps of min of all 10 estimates, zero spanning values not plotted
1
In time observe events which in ascending size are
.. ..
in there are 1 events of si
,
'Return Time' for event i
ze
( )( 1) ( )
Now ( ) /
so (1 )
or in term
s
k N
k
k
k
t N
x x x
t N k x x
x x
t tR x
N k N k
C x x k N
RR C
C
s of the change at quantile
( 1)( 0)
C C Rdx dq
P P R
Finally, this can all be recast in terms of conventional ‘Return Times’…
t
Summary and Next Steps
Our method is entirely data-based: not about prediction or attribution, but rather characterising recent rate of change at particular quantiles.
Focus at particular quantiles in particular locations.
With temperature looked a d(temperature) at given temp. Can generalise idea as we did in precip to derived quantities which are application specific e.g. all days which exceed a threshold for flood management.
Talking to end users will enable us to find these quantities.
Summary and Next Steps
Decisions are usually threshold-specific. Planning will be sensitive to how likely such a threshold crossing is in future e.g. potential collaboration with Zaid Chalabi of Public Health England.
Uncertainties are subtle and depend on where you are geographically and in the distribution.
Need for modelling that goes beyond iid into dependence (and even long range dependence), and into more realistic, data-determined distributions-not just Gaussian and gamma.