Adv. Geosci.
, 16, 3–9, 2008
www.adv-geosci.net/16/3/2008/
© Author(s) 2008. This work is distributed under
the Creative Commons Attribution 3.0 License.
A multiscale approach for precipitation verification applied to the
FORALPS case studies
A. Lanciani1,2 , S. Mariani2,3 , M. Casaioli2 , C. Accadia4 , and N. Tartaglione5
1 Inter-UniversitiesNational Consortium for Physics of Atmospheres and Hydrospheres (CINFAI), Camerino, Italy
2 Agency for Environmental Protection and Technical Services (APAT), Rome, Italy
3 Mathematics Department, University of Ferrara, Ferrara, Italy
4 EUMETSAT, Darmstadt, Germany
5 Physics Department, University of Camerino, Camerino, Italy
Received: 25 August 2007 – Revised: 14 December 2007 – Accepted: 23 January 2008 – Published: 9 April 2008
Abstract. Multiscale methods, such as the power spectrum,
are suitable diagnostic tools for studying the second order
statistics of a gridded field. For instance, in the case of Nu-
merical Weather Prediction models, a drop in the power spec-
trum for a given scale indicates the inability of the model to
reproduce the variance of the phenomenon below the corre-
spondent spatial scale. Hence, these statistics provide an in-
sight into the real resolution of a gridded field and must be ac-
curately known for interpolation and downscaling purposes.
In this work, belonging to the EU INTERREG IIIB Alpine
Space FORALPS project, the power spectra of the precipita-
tion fields for two intense rain events, which occurred over
the north-eastern alpine region, have been studied in detail.
A drop in the power spectrum at the shortest scales (about
30 km) has been found, as well as a strong matching between
the precipitation spectrum and the spectrum of the orography.
Furthermore, it has also been shown how the spectra help un-
derstand the behavior of the skill scores traditionally used in
Quantitative Precipitation Forecast verification, as these are
sensitive to the amount of small scale detail present in the
fields.
1 Introduction
Mesoscale meteorology, dealing with phenomena ranging
from rain bands to thunderstorms, is deeply concerned with
small scale detail. Nowadays Numerical Weather Predic-
tion (NWP) models are capable of producing such detail.
However, their verification is nontrivial. For instance, tra-
ditional skill scores (which measure point-to-point match-
ing) are sensitive to small displacement errors, resulting in
a double penalty effect, so that increasing forecast detail
Correspondence to: A. Lanciani
(alexandre.lanciani@apat.it)
Published by Copernicus Publications on behalf of the European Geosciences Union.
Advances in
Geosciences
may decrease the score. This kind of first order verification
can be deceptive. Higher order moments must be studied to
make sure that the fields being compared are defined on grids
with the same real resolution, and even if they are, that they
have the same amount of small scale detail (see, e.g., Beck
and Ahrens, 2004; Chèruy et al., 2004; Grasso, 2000; Harris
et al., 2001; Tartaglione et al., 2002).
In this work, performed within the EU INTERREG IIIB
Alpine Space FORALPS project (http://www.foralps.net),
we have studied in detail the power spectra of the precipi-
tation fields for two intense rain events which occurred over
the north-eastern alpine arc. The first event took place on
16–20 November 2001, whilst the second one occurred on
8–10 September 2005. Namely, we have studied the spec-
tra of three operational Limited Area Models’ (LAMs) fore-
casts, and attempted to compare these results with those from
a Quantitative Precipitation Forecast (QPF) intercomparison
study by means of traditional skill scores (Mariani and Ca-
saioli, 2008). We have also investigated how these spectra are
affected by different interpolation techniques that are usually
employed in QPF verification.
The rest of the paper is arranged as follows: in the next
section we will define the methodology, then we will describe
the observations and models data set. Afterward we will dis-
cuss some results we have obtained, and finally we will draw
some conclusions.
2 Methodology
2.1 Power Spectrum
The power spectrum (Wilks, 1995) can be an effective diag-
nostic tool to study the higher order moments of a gridded
field and its scale dependency (Goody et al., 1998). The 2-D
spectrum E(kx , ky ) of a real field φ(x, y), where kx and ky
 configuration at the Regional Meteorological Observatory of
average Friuli-Venezia Giulia (OSMER). Moreover, the 0.5◦ T511
ectrum L60 ECMWF global model (http://www.ecmwf.int/) is also
4used in the comparison. A. Lanciani et al.: A multiscale approach for precipitation verification
be writ-
shows 2.2 QPF Verification
The ab-
dicator In this study, for QPF verification, we have used the equi-
r, more table threat score (ETS; Schaefer, 1990). This score is tallied
up on 2×2 contingency tables (Wilks, 1995), which summa-
rize in a categorical way possible combinations of forecast
and observed events above or below a given precipitation
threshold. For each selected threshold, four categories are
uitable then defined: hits; false alarms; misses and correct non-rain
d up on
forecasts (a, b, c and d respectively). To reduce the sensitiv-
arize in
ity of skill scores to small changes into the population of the
and ob-
contingency table elements, they have been calculated on a
eshold.
sum of daily contingency tables depending on the time period
efined: Fig. 1. LAMs’ domains: ALADIN (red), QBOLAM (green), and considered. Thus, for the November 2002 event scores have
asts (a, Fig. 1.(blue).
WRF LAMs domains:
The ALADIN
gray shaded area is(red), QBOLAMarea.
the verification (green), and been calculated on five contingency tables, whereas for the
of skill WRF (blue). The gray shaded area is the verification area.
September 2005 event they have been calculated over three
contin-
areThe
the models
wavenumber tables.
sum of differ components, is formally
in parameterization and defined as the
discretization
Fourier The ETS is an accuracy measure for events, that is, it mea-
idered. schemestransform
(ALADINofand
the ECMWF
autocorrelation function:
are spectral models, while
Z sures how well the forecast “yes” events correspond to the
1
E(kx , ky ) = √ dlx dly e−i(kx lx +ky ly ) K(lx , ly ), (1) observed “yes” events. ETS allows also for the number of
2π hits that would be obtained purely by chance (random fore-
where cast). It is defined by:
Z
a − ar
K(lx , ly ) = dxdyφ(x + lx , y + ly )φ(x, y), (2) ETS = , (4)
a + b + c − ar
is the autocorrelation function and lx and ly are respectively where ar =(a+b)(a+c)/(a+b+c+d) is the number of
the lag in the x direction and in the y direction. model hits expected from a random forecast.
However, it can also be computed according to the Wiener- A score equal to one represents a perfect score; whilst a
Khinchin theorem by multiplying the 2-D Fourier transform value close to zero or negative means that the model has a
by its complex conjugate: questionable forecasting ability.
Z 2
1  −i(xkx +yky )

E(kx , ky ) = dxdye φ(x, y) . (3)
2π   3
We have chosen the latter method, as it suppresses some
computational noise. Moreover, a Hanning window was pre-
viously used to filter the data and to reduce aliasing (Press
et al., 1992).
The relationship between the model domain grid size 1
and the wavenumber grid size 1̃ is given by 1̃ = (N 1)−1 ,
where N is the number of the model grid points. The largest
wavenumber from which we can extract meaningful infor-
mation is given by the Nyquist frequency (21)−1 . Since
the models are defined on different grids with different grid
steps, the wavenumber ranges will be different, too. The
2-D spectrum can be presented as an isotropic power spec-
trumq E(k), if it is averaged angularly and k is defined as
k= kx2 + ky2 . The width of the bands where the average is
made is chosen in order to smooth the isotropic spectrum
without losing any significant information.
Scaling of the power spectrum occurs when it can be writ-
ten as E(k)∼k −β . In other words, the spectrum shows scale
invariance if it is linear in k on a log-log plot. The absolute
value of the spectral slope, that is, β, is an indicator of the
fields smoothness. The higher β, the smoother, more orga-
nized, is the structure.
Adv. Geosci., 16, 3–9, 2008
Models data set and observations
3.1 Models
Forecasts were modelled by three LAMs (whose domains
are shown in Fig. 1) the 11-km Aire Limitée Adaptation
dynamique Développement InterNational (ALADIN; http:
//www.cnrm.meteo.fr/aladin/), operational at the Environ-
mental Agency of the Republic of Slovenia (EARS); the
0.1◦ QUADRICS BOlogna LAM (QBOLAM) operational
at the Agency for Environmental Protection and Technical
Services (APAT; Speranza et al., 2007); the hydrostatic ver-
sion of the 10-km Weather Research and Forecasting (WRF;
http://wrf-model.org/) model running in a research configu-
ration at the Regional Meteorological Observatory of Friuli-
Venezia Giulia (OSMER). Moreover, the 0.5◦ T511 L60
ECMWF global model (http://www.ecmwf.int/) is also used
in the comparison.
The models differ in parameterization and discretization
schemes (ALADIN and ECMWF are spectral models, while
QBOLAM and WRF are finite difference models) and in both
initial and boundary conditions. Global analyses and fore-
casts from ECMWF are employed as initial and boundary
conditions, respectively, by QBOLAM and WRF, whereas
www.adv-geosci.net/16/3/2008/
 4 A. Lanciani et al.: A multiscale approach for precipitation verification

4 A. Lanciani et al.: A multiscale approach for precipitation verification

A. Lanciani et al.: A multiscale approach for precipitation verification 5

(a) ALADIN (b) QBOLAM (c) WRF

(a) ALADIN (b) QBOLAM (c) WRF

Fig. 2. Power spectrum of the 16 November 2002 24h accumulated precipitation forecast: original grid (solid line), bilinear interpolation
(dotted
Fig. line) and
2. Power remapping
spectrum (dashed
of the line). A linear
16 November fit on
2002 24 the first part of
h accumulated the originalforecast:
precipitation spectrum original
is also shown.
grid (solid line), bilinear interpolation
(dotted line) and remapping (dashed line). A linear fit on the first part of the original spectrum
Fig. 2. Power spectrum of the 16 November 2002 24h accumulated precipitation forecast: original is alsogrid
shown.
(solid line), bilinear interpolation
(dotted line) and remapping (dashed line). A linear fit on the first part of the original spectrum is also shown.

(a) ALADIN (b) QBOLAM (c) WRF

(a)2,
Fig. 3. Same as Fig. ALADIN
but for the 9 September 2005 forecast. (b) QBOLAM (c) WRF
Fig. 3. Same as Fig. 2, but for the 9 September 2005 forecast.

ALADIN
Fig. 3. Same employs
as Fig. 2,ARPEGE
but for the (Action
9 September de 2005
Recherche
forecast.Petite 3.2 Observations
interesting
Echelle to note
Grande that thehttp://www-pcmdi.llnl.gov/)
Echelle; spectrum of ALADIN’s orography anal-
drops
yses andsharply
forecasts. for Moreover,
scales smaller 24 hthanALADIN3 gridruns, stepsstarting
(Fig. 4). at
Orography
interesting
00:00 UTC, to is note
are spectrally
used, therepresented
thatwhile spectrum
the other ofin ALADIN,
ALADIN’s
two LAMs are which
orography is
initial- For the November 2002 event, precipitation data have been
whyatits
drops
ized power
sharply
12:00 UTC spectrum
for of the should
scales smallerbethan
previous equal
day 3 to
for a zero
grid h below
36steps run,(Fig.about
and 4).
the obtained from 766 working rain gauges belonging to the
30 12h
first km. of
Orography The isresidual
each spectrally
run are power arises as
represented
discarded from back
ainspin-up.
ALADIN, transformation
which is networks of APAT [former Italian National Hydrographic
and
whyTheinterpolation
itsmodels
power spectrum to the physical
considered shoulddifferbegrid.
equal
also to in zero below about
horizontal grid and Marigraphic Service (SIMN) network], Agenzia Re-
30Itkm.
size. is
Since Thethe residual
interesting power
at this
intercomparisonpoint arises from back
to consider
results alsobe
may transformation
thesensitive
ETS score to gionale per la Protezione dell’Ambiente (ARPA) of Emilia-
and
such interpolation
(Fig.difference,
5). Here we to
willtheonly
precipitationphysical
discuss grid.
forecaststhe 2002
have case.
been The alsogreat-
post- Romagna, ARPA of Liguria, OSMER, EARS and Zen-
estItdifference
processed is interesting
on two is on at18
this
common November
point 2002 (see
to consider
verification also
also
grids the
(withFig.
ETS6 for
grid an
score
size tralAnstalt für Meteorologie und Geodynamik (ZAMG).
0.1◦5).
eyeball
of(Fig. and 0.5◦we
comparison
Here willbetween themeans
only discuss
respectively) by forecast
the 2002 fields
case.and
of bilinear The the ob-
great-
interpola-
servation
est
tion difference
and field).
also As 18
is on
remapping canNovember
1be seensee
(e.g., from
2002Fig.
(see
Accadia 5(a),
also
et ALADIN
al.,Fig. ap-
6 forand
2003), an For the September 2005 event, precipitations have been
pears to comparison
eyeball
accumulated haveata24 better ETS than
between
h, starting the
fromtheforecast
other two
00:00 models,
fields
UTC. at least
and the ob- collected from 781 working rain gauges of APAT, ARPA of
upAlltothese
servation a 30field).
mm As −1
24h
differences canthreshold.
be seen from
notwithstanding, However,
Fig. all its
5(a), performance
theALADIN
models (in- ap- Emilia-Romagna, ARPA of Liguria, ARPA of Lombardia,
◦
seems
pears tothe
cluding to worsen
have if weETS
a better
hydrostatic look
WRF) at the
than the0.1
considerother grid
two(Fig.
models,
precipitation 5(b)). at least
a diag- OSMER, EARS, and ZAMG.
−1
upAs
nostic tovariable.
awe30anticipated
mmAs 24h such, inthreshold.
itthe However,
introduction,
is calculated its
this
after allcanperformance
the be partly
prognos- Fig. 4. Spectra of the models’ orography: ALADIN (dotted line),
seems to worsen if we look spectral
at the This ◦
0.1behavior In order to produce an adequate 24-h observed rainfall
ticexplained
variablesby havethe been
different
advected. isgrid (Fig.
of the
particularly 5(b)).
models.
impor- QBOLAM (dashed line) and WRF (dot-dashed line).
gridded analysis over the two common verification grids, a
Indeed,
tant As
forwe theanticipated
ALADIN, of βinfor
valuesbecause theasboth events can
introduction,
a spectral be obtained
this
model can from
be partly
its dynamics Fig. 4. Spectra of the models’ orography: ALADIN (dotted line),
a linear
explained fit on the first
byspectral (scaling)
the different part of the spectra. At least for two-pass Barnes objective analysis scheme has been used
take place in space,spectral
whereasbehavior of the models.
the microphysics take QBOLAM (dashed line) and WRF (dot-dashed line).
(Barnes, 1964,with
1973).
the
Indeed,
place 2002 event,
the values
in physical of β for both events can be obtained from
ALADIN’s
space. forecast on the original domain contradiction our This is adiscussion
earlier Gaussian weighted-averaging
on the orography
aalways
linear fithasonmore structure
the first (scaling) thanpartWRF’s, which in
of the spectra. Atturn
leasthas
for technique that assigns a weight
spectrum. The contradiction could to be
each rain gauge
explained observa-
by several
more
the structure
2002 event, than QBOLAM’s.
ALADIN’s forecast This
on would
the seem to
original be in
domain tion as
reasons, a function of
both physical
contradiction the distance
with our(absence between
of prognostic
earlier discussion the gauge and the
on microphysical
the orography
1 Remapping is performed by subdividing each verification grid grid box center.
always has more structure than WRF’s, which in turn has spectrum. The contradiction could be explained by several
box
more intostructure
n×n sub-boxes (in this case,This
than QBOLAM’s. n=13).would Then seem the to
value of
be in reasons,
the nearest native grid point is assigned to each sub-grid point. The More both physical
precisely, call(absence of prognostic
x the position of themicrophysical
analysis point
average of these sub-grid point values produces the remapped value and xk , k=1, . . . , K, the positions of the gauges within its
of the verification grid point. region of influence. A first pass is performed to produce a

www.adv-geosci.net/16/3/2008/ Adv. Geosci., 16, 3–9, 2008

 6 A. Lanciani et al.: A multiscale approach for precipitation verification

aphy The spectrum of the fields on their native grids displays

g. 4). scale invariance down to about 30 km, after which there is
ch is a fall off (we show two examples in Figs. 2 and 3). Hence,
about 30 km can be taken as the minimal resolution of the grids
ation insofar as the precipitation is concerned. Actually, taking
into account numerical implementation issues, such as su-
score perdiffusion operators that ensure computational stability, it
great- is probably even lower (for a more throughout discussion on
or an effective model resolution, see Frehlich and Sharman, 2007,
e ob- and the references therein).
N ap- Both interpolation methods result in smoother fields, al-
least though bilinear interpolation slightly more so. This is more
mance striking in the case of QBOLAM (see Figs. 2b and 3b).
The correlation between the precipitation spectra on the
Fig.4.4.Spectra
Spectraofofthe
themodels’
models’orography:
orography: ALADIN
ALADIN (dotted
(dotted line), models’ native grid and the spectra of the models’ orography
partly Fig. line),
QBOLAM(dashed
(dashedline)
line)and
andWRF
WRF(dot-dashed
(dot-dashed line).
line). is very high, always higher than 0.98. But it is perhaps more
odels. QBOLAM
interesting to note that the spectrum of ALADIN’s orography
from
drops sharply for scales smaller than 3 grid steps (Fig. 4).
st for
first-guess precipitation analysis (Daley, 1991): Orography is spectrally represented in ALADIN, which is
main contradiction with our earlier discussion on the orography why its power spectrum should be equal to zero below about
n has spectrum. K contradiction could be explained by several
The
(0) 30 km. The residual power arises from back transformation
X
be in f
reasons,
A (x) = wk fO (x(absence
both physical k ), (5)
of prognostic microphysical
k=1 and interpolation to the physical grid.
It is interesting at this point to consider also the ETS score
where fO (xk ) is the precipitation measured at the k-th gauge (Fig. 5). Here we will only discuss the 2002 case. The great-
and the weights are defined as est difference is on 18 November 2002 (see also Fig. 6 for an
eyeball comparison between the forecast fields and the obser-
!
|xk − x|2
wk = exp − . (6) vation field). As can be seen from Fig. 5a, ALADIN appears
4aR 2 to have a better ETS than the other two models, at least up to
In the previous equation |xk −x| is the distance between a 30 mm 24 h−1 threshold. However, its performance seems
the analysis point and the position of the gauge, R is the av- to worsen if we look at the 0.1◦ grid (Fig. 5b).
erage data spacing and a is a proportionality coefficient that As we anticipated in the introduction, this can be partly
results in the optimal response function. explained by the different spectral behavior of the models.
This is followed by a second pass that is needed to increase Indeed, the values of β for both events can be obtained from
the amount of detail in the first guess: a linear fit on the first (scaling) part of the spectra. At least for
the 2002 event, ALADIN’s forecast on the original domain
K h i always has more structure than WRF’s, which in turn has
(1) (0) (0)
X
fA (x) = fA (x) + wk fO (xk ) − fA (xk ) . (7) more structure than QBOLAM’s. This would seem to be in
k=1
contradiction with our earlier discussion on the orography
The algorithm’s convergence is very fast: Koch et al. spectrum. The contradiction could be explained by several
(1983) have shown that only these two passes through the reasons, both physical (absence of prognostic microphysical
data are needed to achieve the convergence of the analysis to fields implies that none of the orographic rain lands in the lee
the observations, provided that a numerical convergence pa- of the mountains, amplifying small scale features; Dr. Mark
rameter γ ∈ (0, 1) is used to redefine R in Eq. (6) as R 0 =γ R Žagar, 2007, personal comunication) and numerical (spectral
in the second pass. We have used γ =0.3 and R=0.15◦ . truncation entails a Gibbs effect, see Lindberg and Broccoli,
1996). However, this aspect needs further study to be fully
understood.
4 Results Moreover, if we look at the spectra of the fields remapped
on the same 0.1◦ and 0.5◦ common grids, where the skills
The spectra were calculated, using the IDL2 software and its
scores were calculated (see Figs. 7 and 8), we find that the
Fast Fourier Transform (FFT) and Hanning algorithms, over
differences among the models’ spectra are greater on the for-
the intersection of all the models’ domains, i.e.: between
mer grid than on the latter.
8.7◦ W and 18.4◦ W in longitude and 42.9◦ N and 48.9◦ N
Accordingly, ALADIN’s ETS is more penalized on the
in latitude (see the shaded area in Fig. 1).
0.1◦ grid and the intercomparison should not be performed
2 Interactive Data Language, version 6.4, © ITT Visual Informa- there but rather on the 0.5◦ grid. On the other hand, from
tion Solutions, Boulder, CO, http://www.ittvis.com/. Fig. 5a we see that the models’ scores are much more

Adv. Geosci., 16, 3–9, 2008 www.adv-geosci.net/16/3/2008/

A.A.Lanciani
Lancianietetal.:
al.:AAmultiscale
multiscaleapproach
approachfor
forprecipitation
precipitationverification
verification 55
A. Lanciani et al.: A multiscale approach for precipitation verification 7

◦ ◦ verification grid ◦
(a)(a)0.5
0.5 verification grid (b) 0.1 verification
verification grid
grid

Fig.
Fig.
Fig. 5.
5. 5. The
The
The ETSscore
ETS
ETS score for
scorefor the
forthe 2002
the2002 event
2002event reported
eventreported as
asaaafunction
reportedas function of
functionof the
ofthe selected
selectedthresholds.
theselected thresholds.
thresholds.

Fig. 6. Contours (in mm 24 h−1 ) of precipitation observed (a) and forecast by ALADIN (b), QBOLAM (c) and WRF (d) on 18 Novem-
Fig. 6. Contours (in mm 24h−1 ) of precipitation observed (a) and forecast by ALADIN (b), QBOLAM (c) and WRF ◦(d) on 18 November
ber6.2002. Forecasts are 24h
remapped
−1
of on the 0.1◦ common verification grid. For the observations, the contours of the 0.1 ◦(d)
Barnes
on 18analysis is
Fig.
2002. Contours
Forecasts(inaremm
remapped ) on the 0.1◦ common
precipitation observed (a) and forecast
verification grid. Forbythe
ALADIN (b), QBOLAM
observations, (c) of
the contours andthe
WRF
0.1 Barnes November
analysis is
masked over the sea. ◦ ◦
2002.
maskedForecasts
over theare
sea.remapped on the 0.1 common verification grid. For the observations, the contours of the 0.1 Barnes analysis is
masked over the sea.

www.adv-geosci.net/16/3/2008/ Adv. Geosci., 16, 3–9, 2008

66 A.A.Lanciani
Lancianietetal.:
al.:AAmultiscale
multiscaleapproach forfor
approach precipitation verification
precipitation verification
8 A. Lanciani et al.: A multiscale approach for precipitation verification

Fig. 7. Spectra
Fig. of the 18 November2002
2002 24 accumulated
h accumulated precip- Fig.8.8.Spectra
Spectraofofthe
the1818November
November2002
200224h24 haccumulated
accumulatedprecipi-
precip-
Fig.7.7.Spectra
Spectraofofthe
the1818November
November 200224h
24h accumulatedprecipi-
precipi-
◦ ◦ common grid: ALADIN (dotted line),
Fig.
Fig. 8.interpolated
itation Spectra of on
thea18 November 2002 24h accumulated
◦ ◦ common grid: ALADIN (dotted line),
0.5 precipi-
itation
tation interpolated
interpolated on on
a a 0.1
0.1 ◦common grid: ALADIN (dotted line), tation interpolated on a 0.5 common
◦ grid: ALADIN (dotted line),
tation interpolated on a 0.1 common grid: ALADIN (dotted line), tation interpolated
QBOLAM(dashed on
(dashedline), a 0.5
line),WRF common
WRF(dot-dash grid:
(dot-dash line) ALADIN
line) and
and ECMWF (dotted
ECMWF (solid line),
(solid
QBOLAM (dashedline)
QBOLAM line) andWRF
WRF (dot-dashedline).
line).
QBOLAM(dashed (dashed line)andand WRF(dot-dashed
(dot-dashed line). QBOLAM
QBOLAM
line). (dashed line), WRF (dot-dash line) and ECMWF (solid
line).
line).
homogeneous, although ALADIN seems to behave slightly
fields implies
better. In a that
similarnone way,of the orographic rain is lands in the lee 5allyConclusions
coarser than the grid mesh-size. Moreover, interpolation
fields implies that none of the
the coarser
orographic field rain more
landspenalized
in the lee 5 Conclusions
ofasthethemountains, amplifying small scale features; Dr. Mark techniques further worsen the resolution by oversmoothing
of the mountains, amplifying small scale features; Dr. inMark
threshold increases, and this is again reflected the
Žagar,
ETS 2007,
behavior personal
– and, comunication)
one more time, andthenumerical
difference (spectral
is more Inthethis
precipitation
work, we have fields.studied
Between thethe two techniques
power spectra of the we con-
pre-
Žagar, 2007, personal comunication) and numerical (spectral In this work, we have studied the power spectra of the pre-
truncation
macroscopic entails on athe Gibbs0.1◦ effect,
grid seeon
than Lindberg
the 0.5◦and Broccoli,
grid, which is sidered, that
cipitation fieldsis, forecast
remapping by and threebilinear
NWP LAMs, interpolation, namely: theAL-lat-
truncation entails a Gibbs effect, see Lindberg and Broccoli, cipitation fields forecast by three NWP LAMs, namely: AL-
1996).
again However,
coherent this our aspect needs further study to be fully ter fares
ADIN the worst.atThis
operational EARS, first QBOLAM
result is consistent operational with at previous
APAT
1996). However,with this aspect results on the
needs spectra.
further study to be fully ADIN on operational at EARS, QBOLAM operational at APAT
understood.
That increasing the threshold entails a reduction in the studies
and the subject.
WRF operational at OSMER. For comparison and com-
understood. and WRF operational at OSMER. For comparison and com-
ETS of the more detailed field
Moreover, if we look at the spectra of the fields remapped can be understood in the Second, we we
prehensiveness showed have that also there consideredis a strongthe ECMWF matchingfore- be-
Moreover, if ◦weIf look at◦ theisspectra of thetwo fields remapped prehensiveness we have also considered the ECMWF fore-
following way: the field
on the same 0.1 ◦ and 0.5 ◦ common grids, where the skillsnot smooth, neighboring tween
casts andthetheprecipitation
observationsspectra available andfor thethe spectra of the 16-20
two events: orog-
on the same 0.1 very anddifferent
0.5 common grids, where the over- skills casts and
raphy the
used2001,in observations
theand models. available
Althoughfor thistheshould
two events:
not come 16-20
points
scores can have
were calculated (see Figs. values.
7 andIn8), particular,
we find the that the November 8-10 September 2005. We focused on
scores were calculated (see Figs.peaks,
7 andone 8), belonging
we find that the November
as regions
a surprise, 2001,
as the and 8-10
two Spaceevents asSeptember
wetheystudied 2005. have We focused
a clearoforo- on
lap between two slightly shifted
differences among the models’ spectra are greater on the for- to the the of the Alpine are the subject the
differences among the models’ spectra are greater on the for- the regions
graphic component,
FORALPS
of the
project within
Alpine Space
it is important as
which thistostudy
they are
recognize the subject
this aspect
was performed.
of the
forecast field and
mer grid than on the latter. the other to the observed field, will be lesser
mer FORALPS project inwithin whichprecipitation
this study was is a performed.
thangridfor than on the latter.
a smoother forecast field. Since whatever is below explicitly.
Our results were Indeed, threefold. this study diagnos-
Accordingly,
the threshold ALADIN’s
gets cut, for aETS is morehigh
sufficiently penalized
threshold on whatthe Our
ticFirst,results
variable,
we found werethat
deriving threefold.
from
therecomplexis a dropprocessesof the power in NWP mod-
spectrum
Accordingly,
◦
0.1remains
grid and the ALADIN’s
intercomparison ETS should
is morenot penalized
be performedon the First, we found that there isbea drop of temperature.
the power spectrum
0.1◦ but inand
gridrather thetheformer case◦ are two isolated
intercomparison should peaks,
not be entailing
performed a atelsthewhich
shortestinvolvescales specific
that humidity
can and
theoretically resolved The
by
there
double penalty on
error theto0.5 the grid. On the other
◦ETS. On the contrary, if the fore-
hand, from at the
two
these shortest
latter
models,variables scales
confirming that thatcan
are prognostic be
their theoretically
variables
real and as such
resolution resolvedtheirby
is actu-
there5(a)butwe rather on the 0.5 grid. On the are other hand, from these models,
Fig.
cast5(a)
Fig. is smooth
we
see there
see
that
that
the
will
the
models’
still be scores
models’ anscores
overlap are
much
resulting
much
more in ho-
more both
ho-
equations
ally coarser than confirming
include, thefor gridreasons that
mesh-size. their
owning real
to
Moreover, resolution
numerical interpolationis actu-
stability,
mogeneous,
a false alarm although
(the forecastALADIN peak) seems
but also to behave
a hit (whereslightlythebet- two ally coarser
diffusion
techniques than the
operators
further thatgrid
worsen cause mesh-size.
the spectralMoreover,
aresolution drop. interpolation
Instead,
by oversmoothing a clear
mogeneous,
ter. In a overlap).
similar although
way, theALADIN coarser field seems to behave
is more penalized slightlyas thebet- techniques further worsen the resolution by oversmoothing
fields and direct action of diffusion operators
the precipitation fields. Between the two techniques we con- on precipitation has
ter. In a similar way, the coarser field
threshold increases, and this is again reflected in the ETS be- is more penalized as the the
still precipitation
to be fully fields.
understood. Between the two techniques we con-
threshold increases, and this is again reflected in the ETS be- sidered, that is, remapping and bilinear interpolation, the lat-
havior — and, one more time, the difference is more macro- ter fares the worst. This first result is consistent with previouslat-
sidered,
Finally, that
we is,
have remapping
related and
the bilinear
spectra of interpolation,
the fields with the
the
havioron
scopic —the and, 0.1one
◦ more time, the difference
grid than on the 0.5◦ grid, which is more is macro-
again ter fares
5 Conclusions ◦ ◦ ETS,
studies on the
one of the
the worst. This first
traditional
subject. QPF result is consistent
verification skillwith
scores previous
that
scopic on the 0.1 grid
coherent with our results on the spectra. than on the 0.5 grid, which is again studies
are often onusedthe subject.
in model intercomparison. As we expected,
coherent with our results on the spectra. Second, we showed that there is a strong matching be-
In thisincreasing
That work, we have studied the
the threshold powera spectra
entails reduction of thein thepre- the Second,
tween lesser we showed
smoothness
the precipitation of that
ALADIN’s
spectra there
and the isfield
aspectra
strong ofmatching
was penalized the orog- bybe-
That
cipitation increasing
fields the
forecast threshold
by three
ETS of the more detailed field can be understood in the entails
NWP a
LAMs, reduction
namely: in AL-the tween
the ETS the
on precipitation
the higher spectra
resolution
raphy used in the models. Although this should not come and
(in the
fact, spectra
unrealistically of the orog-
high
ADIN
ETS of operational
the more at EARS,
detailed QBOLAM
field
following way: If the field is not smooth, two neighboring can be operational
understood at APAT
in the asresolution
raphy used
a surprise, – seeas above)
in thethe two grid.
models. events This wesuggests
Althoughstudied this that
have for
should QPF
a clear notveri-
come
oro-
and WRF
following
points can have operational
way: If the
very at
field
different OSMER. not For
isvalues. Incomparison
smooth, andover-
two neighboring
particular, the com- fication
as a purposes
surprise, as the
the comparison
two events
graphic component, it is important to recognize this aspect should
we be
studied performed
have a on
clear theoro-
prehensiveness
points
lap can have
between two verywe
slightly have
differentalso considered
shiftedvalues.peaks, In one the ECMWF
particular,
belongingthe tofore-
over-
the coarser
graphic grid only.
component, Otherwise
it is particular
important
explicitly. Indeed, in this study precipitation is a diagnos- attention
to recognize must be
this paid
aspect
casts
lap between
forecast and the
field and observations
twothe slightly
other shifted available for
peaks, one
to the observed the two
belonging
field, events:
will be lesser to16–the inexplicitly.
tic the discussion
variable, Indeed,
deriving of the in results,
from this study
complex namely with regard
precipitation
processes is to
in NWP which
a diagnos-
mod-
20 November
forecast
than for afield and
smoother 2001, and 8–10
theforecast
other theSeptember
tofield. observed
Since whatever 2005.will
field, Weisbefocused
lesser
below model
elsticwhich performs
variable,involve better.
deriving specific Even
fromhumidity so, a
complexand comparison
processes at
temperature. a finer
in NWPThe res-
mod-
onthreshold
than
the the
for regions getsofcut,
a smoother the Alpine
forecast
for SpaceSince
field.
a sufficiently ashigh
they are theissubject
whatever
threshold below
what olution
two elslatter
whichcan provide
involveare
variables physical
specific
prognostic insight
humidity into
variables the
and and model
temperature.behavior,
as such theirThe
of the
the threshold
remains FORALPS project
gets cut,case
in the former forare within which
a sufficiently
two isolatedhigh this
peaks,study
thresholdwas what
entailing per-a such
two as
equations ALADIN’s
latter variables
include, fortendency
are prognostic
reasons toowning
produce to small
variablesnumerical scaleasstability,
and features.
such their
formed.
remainspenalty
double Our results
in the error
former were
to case
the ETS.threefold.
are two Onisolated
the contrary, peaks,if entailing
the fore- a ALADIN’s
equations
diffusion ETS
include,that
operators was further
forcause
reasons penalized
owningdrop.
a spectral when the
to numerical thresholds
Instead, astability,
clear
double
cast First, we found
penalty
is smooth errorthat
there to there
will the
stillETS. is aanOn
be drop of contrary,
the
overlap theresulting
powerifspectrum
theboth
in fore- indiffusion
and the contingency
direct operators
action oftablethatcalculation
diffusion causeoperators became
a spectral ondrop. veryInstead,
high. ahas
precipitation clear
at
acast the
falseisalarmshortest
smooth (the scales
there
forecast that
will peak) can
still bebutbe theoretically
analso
overlap
a hit resulting
(where resolved
the intwo by
both still Ittoshould
and direct
be fully beunderstood.
action noted that, as inoperators
of diffusion any other onwork restrictedhas
precipitation
these models,
a falseoverlap).
fields confirming that their
alarm (the forecast peak) but also a hit (where the two real resolution is actu- tostill
caseto be
Finally, studies,
wefullyhave what
understood.
related we the have analyzed
spectra of thehere fieldsarewithmerelythe
fields overlap). Finally, we have related the spectra of the fields with the
Adv. Geosci., 16, 3–9, 2008 www.adv-geosci.net/16/3/2008/
A. Lanciani et al.: A multiscale approach for precipitation verification
snapshots. The subject of precipitation second order statistics
deserves further study. For example, other multiscale meth-
ods (e.g., wavelet decomposition) could be used. Perhaps
more importantly one could try to understand the physical
and mathematical reasons behind our results. Further studies
along this direction, that include the use of the observations’
spectrum, are on the way.
Acknowledgements. This work was developed and funded under
the framework of the EU INTERREG IIIB Alpine Space FORALPS
project (contract n. I/III/3.1/21). We thank the FORALPS partners
ARPA of Lombardia, EARS, OSMER and ZAMG for providing us
the observed and forecast data used in this study. Our thanks also
go to the ARPA of Emilia-Romagna and of Liguria for precipitation
data over their territory. Finally, we wish to thank B. Lastoria
(APAT) for the help on data processing, the two anonymous
reviewers for their suggestions and, last but not least, A. Speranza
(University of Camerino) and G. Monacelli (APAT) for the support.
Edited by: S. C. Michaelides
Reviewed by: two anonymous referees
References
Accadia, C., Mariani, S., Casaioli, M., Lavagnini, A., and Speranza,
A.: Sensitivity of Precipitation Forecast Skill Scores to Bilinear
Interpolation and a Simple Nearest-Neighbor Average Method
on High-Resolution Verification Grids, Weather Forecast., 18,
918–932, 2003.
Barnes, S. L.: A technique for maximizing details in numerical
weather map analysis, J. Appl. Meteorol., 3, 396–409, 1964.
Barnes, S. L.: Mesoscale objective analysis using weighted time-
series observations, p. 60, tech. memo. ERL NSSL-62, 1973.
Beck, A. and Ahrens, B.: Multiresolution evaluation of pre-
cipitation forecasts over the European Alps, Meteorologische
Zeitschrift, 13, 55–62, 2004.
Chèruy, F., Speranza, A., Sutera, A., and Tartaglione, N.: Surface
winds in the Euro-Mediterranean area: the real resolution of nu-
merical grids, Ann. Geophys., 22, 4043–4048, 2004,
http://www.ann-geophys.net/22/4043/2004/.
www.adv-geosci.net/16/3/2008/
9
Daley, R.: Atmospheric data analysis, Cambridge University Press,
1991.
Frehlich, R. and Sharman, R.: The use of structure functions
and spectra from numerical model output to determine effective
model resolution, Mon. Wea. Rev., in press, 2007.
Goody, R., Anderson, J., and North, G.: Testing Climate Models:
An Approach, Bull. Am. Meteor. Soc., 9, 2541–2549, 1998.
Grasso, L. D.: The differentiation between grid spacing and res-
olution and their application to numerical modeling, Bull. Am.
Meteorol. Soc., 81, 579–580, 2000.
Harris, D., Foufoula-Georgiou, E., Droegemeier, K. K., and Levit,
J. J.: Multiscale statistical properties of a high-resolution precip-
itation forecast, J. Hydrometeor., 2, 406–418, 2001.
Koch, S. E., desJardins, M., and Kocin, P. J.: An interactive Barnes
objective map analysis scheme for use with satellite and conven-
tional data, J. Climate Appl. Meteor., 22, 1487–1503, 1983.
Lindberg, C. and Broccoli, A. J.: Representation of topography in
spectral climate models and its effect on simulated precipitation,
J. Clim., 9, 2641–2659, 1996.
Mariani, S. and Casaioli, M.: Forecast verifivication: A summary
of common approaches and examples of application, FORALPS
Technical Report 5, Univerità degli Studi di Trento, Dipartimento
di Ingegneria Civile e Ambientale, Trento, Italy, 60 pp., 2008.
Press, W. H., Flannery, B. P., Teukolsky, S. A., and Vetterling, W. T.:
Numerical Recipes in FORTRAN: The Art of Scientific Comput-
ing, Cambridge University Press, 1992.
Schaefer, J. T.: The critical success index as an indicator of warning
skill, Weather Forecast., 5, 570–575, 1990.
Speranza, A., Accadia, C., Mariani, S., Casaioli, M., Tartaglione,
N., Monacelli, G., Ruti, P. M., and Lavagnini, A.: SIMM: An
integrated forecasting system for the Mediterranean Area, Mete-
orol. Appl., 14, 337–350, 2007.
Tartaglione, N., Lanciani, A., and Speranza, A.: Analysis of the ef-
fects of numerical diffusion on some meteorological fields fore-
casted by NWP models in the Mediterranean area, in: Proceed-
ings of the 4th European Geophysical Soc. Plinius Conference
on “Mediterranean Storm”, Mallorca, Spain, 3–5 October 2002,
4 pp., 2002.
Wilks, D. S.: Statistical methods in the atmospheric sciences: an
introduction, Academic Press, San Diego (USA), 1995.
Adv. Geosci., 16, 3–9, 2008