The influence of synoptic airflow on UK daily precipitation extremes. Part I: Observed spatio-temporal relationships Douglas Maraun • Timothy J. Osborn • Henning W. Rust Received: 13 May 2009 / Accepted: 7 November 2009 / Published online: 1 December 2009 Ó Springer-Verlag 2009 Abstract We study the influence of synoptic scale atmospheric circulation on extreme daily precipitation across the United Kingdom, using observed time series from 689 rain gauges. To this end we employ a statistical model, that uses airflow strength, direction and vorticity as predictors for the generalised extreme value distribution of monthly precipitation maxima. The inferred relationships are connected with the dominant westerly flow, the orog- raphy, and the moisture supply from surrounding seas. We aggregated the results for individual rain gauges to regional scales to investigate the temporal variability of extreme precipitation. Airflow explains a significant fraction of the variability on subannual to decadal time scales. A large fraction of the especially heavy winter precipitation during the 1980s and 1990s in north Scotland can be attributed to a prevailing positive phase of the North Atlantic Oscillation. Our statistical model can be used for statistical downscal- ing and to validate regional climate model output. Keywords Extreme precipitation Synoptic airflow United Kingdom Climate variability Extreme value statistics Vector generalised model Covariates Statistical downscaling 1 Introduction Extreme precipitation is one of the major natural hazards in the United Kingdom. Hall et al. (2005) estimate the aver- age annual damage caused by flooding to be £1billion; and according to the Association of British Insurers, the 2007 summer floods alone cost £3billion (ABI 2007). Under global warming, the magnitude and pattern of extreme precipitation are expected to change (Trenberth et al. 2003). The increasing water holding capacity of the atmosphere will intensify precipitation events, and changes in the atmospheric circulation will redistribute moisture globally. Changes in the water cycle have already been observed. Trenberth et al. (2007) found an increase in annual-mean precipitation for the mid-latitudes, and a decrease for the subtropics; yet the intensity of precipita- tion has grown in most extra-tropical regions (Alexander et al. 2006). In agreement with global scale observations, Osborn et al. (2000) and Maraun et al. (2008) found positive trends in the contribution of heavy precipitation events to the total winter precipitation throughout the last century (to a lesser extend also for spring and autumn). Similarly, Fowler and Kilsby (2003) reported a positive 1961–2000 trend in extreme precipitation in North Scot- land. However, especially for precipitation on regional scales, it is difficult to distinguish these observed trends from natural variability and to attribute them to external forcing such as anthropogenic greenhouse gases (Hegerl et al. 2007). Electronic supplementary material The online version of this article (doi:10.1007/s00382-009-0710-9) contains supplementary material, which is available to authorized users. D. Maraun T. J. Osborn Climatic Research Unit, School of Environmental Sciences, University of East Anglia, Norwich NR4 7TJ, UK Present Address: D. Maraun (&) Department of Geography, University of Giessen, Giessen, Germany e-mail: [email protected]; [email protected]H. W. Rust Laboratoire des Sciences du Climat et de l’Environnement, 91191 Gif-sur-Yvette, France e-mail: [email protected]123 Clim Dyn (2011) 36:261–275 DOI 10.1007/s00382-009-0710-9
15
Embed
The influence of synoptic airflow on UK daily precipitation extremes: observed spatio-temporal relationships
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
The influence of synoptic airflow on UK daily precipitationextremes. Part I: Observed spatio-temporal relationships
Douglas Maraun • Timothy J. Osborn •
Henning W. Rust
Received: 13 May 2009 / Accepted: 7 November 2009 / Published online: 1 December 2009
� Springer-Verlag 2009
Abstract We study the influence of synoptic scale
atmospheric circulation on extreme daily precipitation
across the United Kingdom, using observed time series
from 689 rain gauges. To this end we employ a statistical
model, that uses airflow strength, direction and vorticity as
predictors for the generalised extreme value distribution of
monthly precipitation maxima. The inferred relationships
are connected with the dominant westerly flow, the orog-
raphy, and the moisture supply from surrounding seas. We
aggregated the results for individual rain gauges to regional
scales to investigate the temporal variability of extreme
precipitation. Airflow explains a significant fraction of the
variability on subannual to decadal time scales. A large
fraction of the especially heavy winter precipitation during
the 1980s and 1990s in north Scotland can be attributed to a
prevailing positive phase of the North Atlantic Oscillation.
Our statistical model can be used for statistical downscal-
ing and to validate regional climate model output.
statistics � Vector generalised model � Covariates �Statistical downscaling
1 Introduction
Extreme precipitation is one of the major natural hazards in
the United Kingdom. Hall et al. (2005) estimate the aver-
age annual damage caused by flooding to be £1billion; and
according to the Association of British Insurers, the 2007
summer floods alone cost £3billion (ABI 2007). Under
global warming, the magnitude and pattern of extreme
precipitation are expected to change (Trenberth et al.
2003). The increasing water holding capacity of the
atmosphere will intensify precipitation events, and changes
in the atmospheric circulation will redistribute moisture
globally. Changes in the water cycle have already been
observed. Trenberth et al. (2007) found an increase in
annual-mean precipitation for the mid-latitudes, and a
decrease for the subtropics; yet the intensity of precipita-
tion has grown in most extra-tropical regions (Alexander
et al. 2006). In agreement with global scale observations,
Osborn et al. (2000) and Maraun et al. (2008) found
positive trends in the contribution of heavy precipitation
events to the total winter precipitation throughout the last
century (to a lesser extend also for spring and autumn).
Similarly, Fowler and Kilsby (2003) reported a positive
1961–2000 trend in extreme precipitation in North Scot-
land. However, especially for precipitation on regional
scales, it is difficult to distinguish these observed trends
from natural variability and to attribute them to external
forcing such as anthropogenic greenhouse gases (Hegerl
et al. 2007).
Electronic supplementary material The online version of thisarticle (doi:10.1007/s00382-009-0710-9) contains supplementarymaterial, which is available to authorized users.
D. Maraun � T. J. Osborn
Climatic Research Unit, School of Environmental Sciences,
We model fl,s(si) and fl,v(vi) (and the corresponding func-
tions for r) as natural cubic splines (Press et al. 1992) with
two degrees of freedom.1 To ensure parsimony and circular
boundary conditions, the direction dependency is modelled
as a phase shifted sine fl,d(di) = al sin d ? bl cos d. The
annual cycle is represented by a phase shifted sine with a
period of one year, fl;tðdiÞ ¼ al sin 2pti365:25
þ bl cos 2pti365:25
:
The link function g-1 = exp(.) ensures positive values of rfor any values of the covariates. Apart from the link
function, the model for the annual cycle is the same as in
Maraun et al. (2009a). For each rain gauge, we assume a
constant shape parameter n, as motivated in Maraun et al.
(2009a) and Rust et al. (2009). This choice is common
practice in extreme value statistics, because the shape
parameter is difficult to estimate due to a very limited
number of observations in the tail. Estimating a variable
shape parameter usually leads to an unacceptably high
uncertainty (e.g. Coles 2001). The parameterisation of f(.)
as natural cubic splines with two degrees of freedom, and
the decomposition of the phase shifted sines in a linear
superposition of sine and cosine with zero phase, ensures a
model linear in the parameters. Choosing a VGLM allows
us to both model nonlinearities in the covariates, and to use
standard maximum likelihood estimation (MLE) to esti-
mate the parameters of this model. This approach provides
confidence intervals based on the curvature of the log-
likelihood at the maximum, and allows for a simple
1 Cubic splines are piecewise third order polynomial functions,
which are smoothly connected at a set of knots (i.e. the function and
its first and second derivatives are continuous). Natural splines require
vanishing curvature at the beginning and end of the data interval.
With only one knot (here in the centre of the data interval), a natural
cubic spline has two degrees of freedom.
264 D. Maraun et al.: Airflow and UK precipitation extremes
123
incorporation of covariates. It is also implemented in the
R-package VGAM (Yee and Stephenson 2007).
For non-independent iid processes, the confidence
intervals broaden accordingly. However, in the case at
hand, the auto-correlation of the residuals (precipitation
maxima minus mean of the predicted GEV) vanishes for
non-zero lags (not shown), such that the standard confi-
dence intervals can be used.
Maraun et al. (2009b) used the Akaike information
criterion (Akaike 1973) to select among several reasonable
model hypotheses, and evaluated the predictive power
using quantile verification scores (QVS) (Friederichs and
Hense 2007). For the validation, a particular model is fitted
to a training data set, and then covariates (si, di, vi) from an
independent validation data set are used to predict the a-
quantiles qa(si, di, vi) of the corresponding GEV. Given a
validation set of N precipitation observations, the quantile
verification score QVSa for this model is defined as
QVSa ¼XN
i¼1
qaðyi � qaðsi; di; viÞÞ; i ¼ 1. . .N; ð5Þ
with
qaðuÞ ¼au if u� 0
ða� 1Þu if u\0:
�ð6Þ
The QVSa is a proper score (Wilks 2006) in the sense that its
expected value is minimal if and only if the predicted
a-quantile equals the a-quantile of the distribution that gen-
erated the observations. The QVSa can be used to evaluate
and compare the ability of different model structures and
predictors to predict high quantiles. In contrast to, e.g. the
AIC, improvements relative to a reference model (e.g.
climatology) can be easily interpreted, and the predictive
power of a model fitted to the whole year can be assessed for
individual seasons. Maraun et al. (2009b) found that the
additive statistical model based on airflow indices with the
highest predictive power is the one described by Eq. (4),
which generally performs better than either the annual cycle
alone, or the airflow indices without an additional annual
cycle, or four independent VGLMs for each season.
In the block maxima approach, only the magnitude of
the block maximum is modelled by the GEV, but not its
occurrence time within the block. This does not pose a
challenge when fitting the statistical model to given
observations of the extremes and the covariates, nor to
investigating the behaviour during the historical record
when maxima occurrence times are known. Our GEV-
based statistical model is, however, limited in its utility for
making actual predictions: the times at which the covariate
ought to be evaluated are not known a priori. On the one
hand, evaluating the covariate on the day the maximum
occurred introduces information (the occurrence time)
which is not modelled and thus cannot be used for a pre-
diction in its classical meaning. On the other hand, using a
value representative of the entire block (e.g. the mean of
the covariate) that avoids introducing non-modelled infor-
mation weakens the relationship between the covariate and
the GEV parameters. This dilemma also occurs in the peak
over threshold approach, where exceedances over a high
threshold are modelled using the Generalised Pareto dis-
tribution (e.g., Coles 2001). In our case, the derived rela-
tionships hold—strictly speaking—only on those days
where precipitation maxima occur. However, the distribu-
tion of airflow on days of precipitation maxima is very
similar to the distribution of airflow on days where pre-
cipitation exceeds a high threshold, such that the results are
valid for all extreme precipitation situations (not shown).
The dilemma can only be resolved by modelling the full
extreme precipitation process as a point process, where
both threshold exceedance and occurrence time are mod-
elled (e.g. Coles 2001). This, however, is beyond the scope
of this manuscript: we do not intend to predict occurrence
and magnitude of daily extreme precipitation using daily
observed airflow indices. We rather aim to infer physical
relationships between airflow and the magnitude of pre-
cipitation extremes, as well as their spatio-temporal vari-
ability in general. Thus, when using the term ‘‘prediction’’
in the following context, we refer to the quantiles of the
fitted GEV distribution, given a specific value of the
covariates.
4 Results
For each of the 689 selected rain gauge time series of
monthly precipitation maxima, we estimated the parame-
ters of the VGLM (Eq. 4). As covariates, we used the
corresponding airflow time series and the annual cycle,
evaluated at the days of the precipitation maxima. Since at
different rain gauges, the maximum of a particular month
might occur on different days, the airflow time series
entering the fit will be different for each rain gauge. In the
following section, we present examples of the derived
relationships at individual stations, the spatial patterns of
these relationships, as well as the temporal and frequency
resolved variability of extreme daily precipitation on
regional scales.
4.1 Example stations
Figure 1 illustrates the structure of the VGLM for three
selected rain gauges: at Kinlochewe in North West Scot-
land (top row, 5�1802900W, 57�3604700N, Met Office Source
Identifier 66), Wick in North East Scotland (middle,
D. Maraun et al.: Airflow and UK precipitation extremes 265
123
3�0502000W, 58�2701400N, Src.Id. 32) and Sandringham
House in East Anglia (bottom, 0�3101600E, 52�4905200N,
Src.Id. 4698). The left column shows the relation between
airflow strength at the day of the precipitation maximum
and the precipitation maximum, the middle column shows
the relation between airflow direction and precipitation,
and the right column shows the corresponding relation
between airflow vorticity and precipitation. The annual
cycle has been analysed in detail in Maraun et al. (2009a)
and is not shown here; it has only been marginally affected
by including the airflow indices. The grey dots in Fig. 1
show the observed monthly maxima, plotted against the
covariate on the day of occurrence. The blue lines depict
the 50% quantile (median) of the predicted distribution,
given a specific value of the actual covariate plus the
average influence (in time) of the other covariates. The
green and red lines show the 90 and 99% quantiles,
respectively. The corresponding pointwise 95% confidence
intervals are indicated by colour shading. Showing the
quantiles instead of the covariates’ influence on individual
GEV parameters (i.e. the fl,.(.) and fr,.(.) in Eq. (4)) sum-
marises the information and concentrates on directly
observable and physically-relevant information.2
These example stations sustain clear relations during the
whole year. Kinlochewe is dominated by airflow strength:
the stronger the airflow, the higher the magnitude of
extreme precipitation. Almost all these extremes happen to
occur under westerly flow, reducing the power of airflow
direction to predict their magnitude.3 A typical weather
situation for extreme precipitation in Kinlochewe associ-
ated with strong westerly flow occurred on 6/7 Jan 2005,
where a low over the Greenland Sea and a high over south
western Europe drove strong and almost zonal flow across
the United Kingdom, causing heavy rain of 142.7 mm
within 24 h.
Interestingly, vorticity exhibits a negative influence on
extreme precipitation. Most heavy events occur for nega-
tive or zero vorticity, which is associated with anticyclones
south of the British Isles. Such a weather situation occurred
on 4 Feb 1999, when an anticyclone over the Bay of Biscay
together with a low over Iceland drove moderately strong
westerly flow across northern Scotland, causing heavy rain
of 34.6 mm within 24 h. For highly positive vorticity this
relationship might reverse. Given the width of the confi-
dence intervals, it is not obvious whether this relationship
is real or just due to sampling variability, yet similar
relationships occur for other rain gauges in the Scottish
Highlands, increasing the overall confidence.
Wick is dominated by airflow vorticity: the stronger the
vorticity, the stronger the expected precipitation. Given the
close distance to Kinlochewe (\160 km), the difference in
the airflow dependency between the two gauges illustrates
the importance of orography for local-scale precipitation.
Whereas Kinlochewe is close to the Scottish west coast and
almost directly exposed to the westerlies, Wick lies at the
rather flat Scottish east coast, in the rain shadow of the
Highlands. There are still heavy rains for westerly airflow,
but they are much weaker than in Kinlochewe, and the
strongest extremes are expected for northerly flow. The
influence of airflow strength is rather weak. Two examples
for weather situations causing extreme precipitation in
Wick occurred on 25 Oct 1998 and 2 Nov 2006. In the
former case, a strong cyclone passed Scotland from the
west, resulting in 26.2 mm of rain. The latter case is an
example for a direction-caused event in Wick: 31.2 mm of
rain fell, as an anticyclone over Ireland circled a warm
front from the north west over east Scotland.
Sandringham is dominated by airflow direction: highest
precipitation maxima are expected for easterly flow coming
from the North Sea. The vorticity relationship is rather
weak: the expected magnitudes increase with vorticity,
saturating for positive values, and with highest variability
around zero vorticity. Strength shows a slightly negative
influence on precipitation extremes.4
To check the validity of the statistical model, we cal-
culated diagnostic plots (e.g. Coles 2001); they are pre-
sented in Fig. 2, for Kinlochewe, Wick and Sandringham
House in the top, middle and bottom row respectively. The
left column depicts probability–probability (P–P) plots, the
right column quantile–quantile (Q–Q) plots. In both cases,
the model and the observations are rescaled to a standard
Gumbel distribution to account for the effects of the
covariates. The red lines show 95% confidence intervals
derived from 999 realisations of a Gumbel distribution.
P–P plots show the empirical frequency distribution against
the predicted probability distribution of the fitted VGLM
and focus on the centre of mass of the distribution. Q–Q
plots show the predicted quantiles against the empirical
quantiles and focus on the tail of the distribution. In all
three cases, the values lie well within the confidence bands,
confirming that the models are not considerably
misspecified.
2 Often, the estimators for l, r and n are correlated, and compensate
for each other. E.g. a low estimate for r could partly compensate for a
high estimate of n, leaving some quantiles basically unchanged. In
this context, only the quantiles have a unique and physically-relevant
meaning.3 This, however, does not imply that airflow in Kinlochewe almost
always flows from the west.
4 The density of data points for low airflow strength values is higher
than for high strength values. Consequently also the probability to
observe high precipitation values decreases with higher strength. This
creates a spurious impression of a negative relationship between
airflow strength and precipitation. For a general discussion refer to
Maraun et al. (2009b).
266 D. Maraun et al.: Airflow and UK precipitation extremes
123
4.2 Spatial variability
The examples presented in Sect. 4.1 indicate that the rela-
tionship between airflow and extreme precipitation varies
across the United Kingdom. To reveal and study these
spatial patterns, we summarised the information for all rain
gauges in geographical maps. Figure 3a–c depicts the time
average of the GEV parameters, i.e. the location parameter
Fig. 1 Estimated relation between airflow and quantiles of the
monthly maxima distribution. Top Kinlochewe (Src.Id. 66), middleWick (Src.Id. 32), bottom Sandringham House (Src.Id. 4698). Firstcolumn strength (natural spline with 2 d.o.f.), second column direction
(sine), third column vorticity (natural spline with 2 d.o.f.). Grey dots
observations, lines: (blue, bottom) 0.5, (green, middle) 0.9 and (red,
top) 0.99 quantiles of the statistical model. The shading depicts the
corresponding 95% confidence intervals. For units, see text and
Osborn et al. (1999)
D. Maraun et al.: Airflow and UK precipitation extremes 267
123
in panel (a), the scale parameter in panel (b) and the shape
parameter in panel (c). In both the location and the scale
parameters, a strong east-west gradient is visible, with
highest values in north-west Scotland and lowest values in
East Anglia. The shape parameter exhibits a different spa-
tial pattern: in the west, it is close to zero or even negative (a
finite tail), whereas it is positive in the east with highest
values along the coastal stretch north of the Scottish–Eng-
lish border. These patterns agree well with those derived
from the annual cycle model of Maraun et al. (2009a). In
addition, we found that the phase of the annual cycle agrees
between the VGLM and the annual cycle model. This
indicates that the phase is not determined by the airflow
dependency but rather by geography and orography.
A subtle difference exists, however, between the VGLM
and the annual cycle model. When comparing the spatial
patterns of the shape parameters of both models (see
Figs. 3(c), and Fig. 7 (g) in Maraun et al. 2009a), it turns
out that the shape parameter of the VGLM is considerably
lower than that of the annual cycle model (on average
approx. 0.05). This indicates that the high values of the
shape parameter in the annual cycle model can at least
partly be attributed to variability that can be explained by
the airflow indices.
The spatial patterns of the relationship between airflow
and extreme precipitation are presented in Fig. 3d–h. Panel
(d) summarises the spatial pattern of the influence of air-
flow strength on extreme precipitation: for each rain gauge,
we show the difference of predicted precipitation medians
for high and low strength values (90 and 10% quantile),
normalised with the predicted median for typical strength
values (50% quantile). A difference of zero indicates a
vanishing dependence on airflow strength.
A distinct pattern emerges, closely resembling the
orography of the UK. Along the west coast, a positive
relationship dominates with very strong relations in
northwest Scotland; further inland and towards the east
coast, this relationship reverses (compare also with the first
column in Fig. 1). The strongest positive relationship
occurs at mountain ranges along the west coast. Here,
landfalling fronts from the Atlantic are uplifted, leading to
a confined band of orographic rain (Roe 2005). Stronger
airflow then increases the moisture flow from the Atlantic
and intensifies precipitation. A secondary effect is the
friction of the rough surface, that slows the movement of
the passing front (Braun et al. 1999); this effect also
increases with stronger airflow. In the flatter regions east of
the main mountain ranges, the negative effect of airflow
strength on precipitation might be explained by a faster
movement of cloud systems for stronger airflow, and
therefore a distribution of precipitation over a larger area.
For vorticity, the relationship is not always monotonous
(see, e.g. Fig. 1). Therefore we depict the relationship for
anticyclonic (e) and cyclonic weather (f) separately. Panel
(e) depicts the vorticity dependence for vorticity values
below 20, i.e. for anticyclonic weather situations. Plotted is
the difference of predicted precipitation medians for a
vorticity of 20 and low vorticity values (10% quantile),
normalised with the predicted median for a vorticity of 20.
Apart from some outliers, zero and positive values domi-
nate, i.e. the magnitude of extreme rainfall increases with
increasing vorticity. This relationship is negative or weak
in the western Scottish Highlands and weak in East Anglia,
but strongly positive along the south coast, Wales, northern
England and southeast Scotland. This relationship is asso-
ciated with anticyclones reducing the likelihood of intense
rainfall. Panel (f) depicts the same relationship for vorticity
values above 20, i.e. for cyclonic weather situations.
Plotted is the difference of predicted precipitation medians
for high vorticity values (90% quantile) and a vorticity of
20, normalised with the predicted median for a vorticity of
20. A difference of zero indicates a vanishing dependence
on airflow vorticity for cyclonic weather. This pattern has a
distinctively lower range of values than the pattern for
anticyclonic weather types, indicating that anticyclones are
Fig. 2 Diagnostic plots for the statistical model. Top Kinlochewe
(Src.Id. 66), middle Wick (Src.Id. 32), bottom Sandringham House
(Src.Id. 4698). Left P–P plot, right Q–Q plot (re-normalised to time
independent Gumbel margins). Grey dots observations, red lines 95%
confidence intervals obtained from 999 realisations of this Gumbel
distribution
268 D. Maraun et al.: Airflow and UK precipitation extremes
123
much better predictors for low precipitation than cyclones
are for strong precipitation. However a pattern emerges,
with a saturation of the vorticity relationship in central and
eastern England, and sustained positive relationships for
the rest of the UK with highest values in north-east Scot-
land. This relationship is associated with depressions
moving over the United Kingdom.
Panel (g) depicts the relative amplitude of the airflow
direction dependency. It ranges from vanishing values in
northwest England, parts of Wales and Scotland, to positive
values in the rest of the country, with highest values in north
east England. Panel (h) depicts the direction for which the
highest precipitation is expected. From the Firth of Forth
along the English east coast to south west England, strongest
rainfall is expected for north easterlies and easterlies. In
central north Scotland around the Moray Firth, on the Ork-
neys and Shetlands, maximum rainfall is most probable
under northerly and northwesterly airflow conditions. Simi-
lar direction dependencies prevail in the Liverpool-Man-
chester area and parts of northwest Wales. On the Hebrides,
along the western Scottish coast, in the Lake District and
Snowdonia, precipitation magnitudes are highest for
westerlies. In southwest Scotland, south Wales and parts of
southwest England, strongest precipitation is expected for
southerlies. The direction patterns are linked to orography
and rain shadows as well as the relative locations of the
Fig. 3 a–c Time average of l, r and n. Plotted are not the actual
offsets l0 and r0, because the physical offset is split between the
model offset and an implicit offset in the splines varying from station
to station. Estimated relation between airflow and median of the
monthly maxima distribution. d strength of the relationship between
strength and precipitation (for details see text), strength of vorticity
relationship for e cyclonic and f anticyclonic weather (for details see
text), g relative amplitude of the direction dependence, h direction of
the expected maximum precipitation
D. Maraun et al.: Airflow and UK precipitation extremes 269
123
Atlantic Ocean, the Irish Sea and the North Sea, where air-
masses can absorb moisture. The eastern parts of the United
Kingdom lie within the rain shadow of the main mountain
ranges. Therefore, airflow direction plays an important role
in triggering heavy precipitation in these regions. North-
easterly moist airflow from the North Sea landfalls and
causes heavy precipitation. During winter, this will pre-
dominantly be associated with fronts, which might precipi-
tate already before landfall. However, low frontal clouds
might be uplifted in the higher elevated regions in north east