You are on page 1of 15

DECEMBER 2002 GERMANN AND ZAWADZKI 2859

Scale-Dependence of the Predictability of Precipitation from Continental Radar Images.


Part I: Description of the Methodology
URS GERMANN AND ISZTAR ZAWADZKI
Department of Atmospheric and Oceanic Sciences, McGill University, Montreal, and
J. S. Marshall Weather Radar Observatory, Sainte Anne de Bellevue, Quebec, Canada

(Manuscript received 28 November 2001, in final form 14 March 2002)

ABSTRACT
The lifetime of precipitation patterns in Eulerian and Lagrangian space derived from continental-scale radar
images is used as a measure of predictability. A three-step procedure is proposed. First, the motion field of
precipitation is determined by variational radar echo tracking. Second, radar reflectivity is advected by means
of a modified semi-Lagrangian advection scheme assuming stationary motion. Third, the Eulerian and Lagrangian
persistence forecasts are compared to observations to calculate the lifetime and other measures of predictability.
The procedure is repeated with images that have been decomposed according to scales to describe the scale-
dependence of predictability.
The analysis has a threefold application: (i) determine the scale-dependence of predictability, (ii) set a standard
against which the skill for quantitative precipitation forecasting by numerical modeling can be evaluated, and
(iii) extend nowcasting by optimal extrapolation of radar precipitation patterns. The methodology can be applied
to other field variables such as brightness temperatures of weather satellites imagery.

1. Introduction: Predictability of precipitation


The predictability of precipitation, in the sense of a
EE V
Ĉ(t 0 1 t , x)C(t 0 1 t , x) dx
c(t ) 5

[EE EE
nonlinear response and sensitivity to small perturbations

]
0.5
in initial and boundary conditions in fluid dynamics, is Ĉ(t 0 1 t , x) dx
2
C(t 0 1 t , x) dx
2
a complex problem. It is a fundamental and intrinsic V V
property of nonlinear systems, but in practice it will
depend as well on the particular forecasting model. (2)
Thus, as pointed out by Lorenz (1973), predictability is
and integrating over t
a concept better taken in a relative sense—relative to a
method of forecasting.
One of the simplest forecasting methods is using cli-
matological values such as the arithmetic mean, the me-
L5 E0
`

c(t ) dt (3)

dian, or the most likely value. For the arithmetic mean,


the predictability is given by the variance, and is in- we get the lifetime (or decorrelation time) L. This def-
dependent of forecast time. inition of correlation functions without subtraction of
A second approach is keeping the most recent ob- the mean [Eq. (2)] has been introduced by Zawadzki
servation frozen (Eulerian persistence) (1973). If c(t) follows an exponential law, L is equal
to the time constant in the exponent, and corresponds
Ĉ(t 0 1 t, x) 5 C(t 0 , x), (1) to the time when the correlation falls below 1/e 5 0.37.
The decorrelation time as defined in Eq. (3) is a practical
where C is the observed precipitation field, t 0 is the measure of the predictability of precipitation patterns.
start time of the forecast, t is the lead time, and Ĉ(t 0 Equation (1) can be expressed in differential form
1 t, x) is the forecasted rate at time t 0 1 t and position using the two-dimensional conservation equation of C,
x. By correlating the forecast with observations in the written as follows
domain V,
dC ]C ]uC ]y C
5 1 1 (4)
Corresponding author address: Dr. Urs Germann, Atmospheric and dt ]t ]x ]y
Oceanic Sciences, McGill University, 805 Sherbrooke W., Montreal,
QC, H3A 2K6, Canada. Neglecting the compressibility term C(]u/]x 1 ]y /]y)
E-mail: urs@zephyr.meteo.mcgill.ca we get

q 2002 American Meteorological Society

Unauthenticated | Downloaded 04/18/21 01:37 PM UTC


2860 MONTHLY WEATHER REVIEW VOLUME 130

dC ]C ]C ]C lead time advection forecasts are useful (Pierce et al.


5 1u 1y . (5)
dt ]t ]x ]y 2000).
So far, precipitation nowcasting by radar was com-
In the case of Eulerian persistence, the local rate of monly based on data of one single radar. This puts an
change ]C/]t is set to zero. upper limit to the scales of observations, and, conse-
A third approach is setting the source-sink term to quently, also to the timescale over which the forecast
zero is useful. The lead times of useful forecasts reported in
]C ]C ]C the literature are usually less than 2 h often less than 1
05 1u 1y (6) h (Mecklenburg et al. 2000). The recent development
]t ]x ]y
of radar networks covering entire continents opens a
and advecting C following the motion of precipitation new opportunity. Simple visual observations of radar
parcels (Lagrangian persistence) composite patterns suggest that a certain degree of per-
sistence is present on timescales of many hours, and in
Ĉ(t 0 1 t, x) 5 C(t 0 , x 2 a), (7) some cases even days. Carbone et al. (2002) observed
where a is the displacement vector. Analogously to the coherent rainfall events, of order 1000 km in zonal span
previous case, we can calculate the decorrelation time and 1-day in duration, occurring nearly one per day.
and thus get another measure of predictability, namely Numerical models are not going to replace precipitation
the predictability obtained by Lagrangian persistence. nowcasting by advection of radar rainfall fields unless
There are several ways of defining the Lagrangian per- there is a breakthrough in appropriately measuring and
sistence (section 4). assimilating into the model the current state of the at-
As a refinement of Lagrangian persistence, we first mosphere.
introduce a source-sink term SC 5 dC/dt, which allows The objective of this paper is to explore the use of
for growth and dissipation of precipitation, and second continental-scale radar composite images to determine
replace C on the right-hand side of Eq. (7) by C̃ the predictability of precipitation. As a measure of pre-
ˆ 0 1 t , x) 5 C(t
˜ 0 , x 2 a) 1 SC (t 0 , x 2 a) dictability we take Eulerian and Lagrangian decorrela-
C(t (8)
tion times (Zawadzki et al. 1994). In its most simplified
Here, C̃ is obtained from C by any type of transfor- version, the proposed procedure has three steps: vari-
mation, such as spectral filtering or averaging in time ational echo tracking, advection of radar reflectivity, and
and/or space. Equation (8) is the general equation of the correlation of the forecast with the observation. This
Lagrangian forecast procedures used in this paper. paper gives a detailed description of the methodology,
A practical aspect of Lagrangian persistence is in and discusses first results.
nowcasting. Nowcasting precipitation has historically Spectral and spatial filters are needed to decompose
been based on the advection of ‘‘frozen’’ radar reflec- precipitation patterns according to scales. Decomposi-
tivity fields. Operational applications began in 1976 tion according to scales is done for two reasons: First,
when the McGill Weather Radar Observatory started Bellon and Zawadzki (1994) have shown that applying
sending 1–6-h rainfall forecasts to the local weather a spatial filter to the forecast image improves the root-
office (Bellon and Austin 1978). Attempts at including mean-square error; increasing smoothing windows for
growth and dissipation failed to improve the skill of the increasing lead times prevents structures from being re-
forecast. A recent refinement has been presented in Bel- produced beyond their lifetime. Second, by calculating
lon and Zawadzki (1994). They show that eliminating the correlation between forecast and observation as a
the small-scale perishable information by means of spa-
function of scale, we can determine the scale-depen-
tial averaging leads to a consistent reduction of the root-
dence of predictability.
mean-square error of the forecast. A power law relates
An example of Eulerian and Lagrangian persistence
the length scale of the spatial filter to the forecast time.
A similar technique proposed by Seed and Keenan is given in Fig. 1. It shows the correlation between the
(2001) uses a Fourier-notch-filter for scale decompo- true observed precipitation field and Eulerian and La-
sition, and advects each scale with its own decorrelation grangian forecasts for lead times up to 8 h. Spatial res-
time. An attempt to relate the Lagrangian persistence of olution of the echo field is 4 km, as further discussed
rainfall patterns to larger-scale meteorological param- in section 2b. A description of the weather situation and
eters such as convective available potential energy the motion field is given later. For the selected forecast
(CAPE) or storm-relative helicity was done in Zawadzki run, the decorrelation time of Lagrangian persistence is
et al. (1994). A limiting factor in the analysis was the approximately 1.8 times that of Eulerian persistence.
maximum range properly sampled by a single radar. In Looking at large-scale features with wavelengths of 64
the United Kingdom, a high correlation was found be- km and more, and thus ignoring all the scales smaller
tween Lagrangian persistence and storm-relative helic- than mesobeta, further extends the range of predict-
ity. Storm-relative helicity introduced by Lilly (1986) ability by about 1 h.
provides, therefore, a guide to the forecaster as to what Finally the analysis will have a threefold application:

Unauthenticated | Downloaded 04/18/21 01:37 PM UTC


DECEMBER 2002 GERMANN AND ZAWADZKI 2861

FIG. 2. Setup of forecast mode. In this paper the assimilation win-


dow and the forecast period are fixed to 1 and 8 h, respectively, and
time step is 15 min.

two stages are automated and the third is manual in-


tervention. As a result, most of the residual clutter and
signals from anomalous propagation are removed. The
resolution of the composite images is 15 min in time,
FIG. 1. Correlation as defined in Eq. (2) between forecast and 2 km in space, and 5 dBZ in reflectivity. The spatial
observation for Eulerian persistence (dotted), Lagrangian persistence coverage extends from 208 to 538N and from 1308 to
(solid), and for Lagrangian persistence of low-pass filtered images 608W. Reflectivity corresponds to the vertical maximum
with cutoff wavelength of 64 km (dashed). Data is from one 8-h
forecast, starting at 1100 UTC 30 Jul 1998. The considered domain value measured by any Weather Surveillance Radar-
has a side length of 2720 km. The precipitation area is about 6 3 1988 Doppler (WSR-88D) radar in the given space–time
10 5 km 2 . The triangle marks the lifetime L of the solid curve cal- frame of one pixel. The accuracy in terms of rainfall
culated with Eq. (3). rates is not comparable to that of rainfall maps based
on sophisticated algorithms that include corrections for
the vertical profile of reflectivity, visibility, attenuation,
Z–R variations and others. The images certainly suffer
1) Determine the predictability of precipitation as a from brightband contamination, different calibration,
function of scales. and lack of visibility, which reduces the quality in terms
2) Set a standard against which the skill for quantitative of quantitative rainfall estimates. But the spatial cov-
precipitation forecasting (QPF) by numerical mod- erage of the composites is unique, and makes the prod-
eling can be evaluated. uct suitable to analyze the scale-dependence of pre-
3) Serve as a basis for extended nowcasting of precip- dictability up to the continental scale, as outlined earlier.
itation by optimal extrapolation of past observations. For the analysis presented in this paper, the spatial
resolution has been reduced to 4 km. A lower threshold
of 10 dBZ (0.1 mm h 21 ) is used. Measurements below
2. Setup of the experiment this value are considered as ‘‘no rain.’’
a. Forecast mode
Strictly speaking, Lagrangian space is defined by the c. Four events
trajectories of the precipitation parcels, and Lagrangian
decorrelation only depends on the rate of growth and To illustrate and test the individual steps of the pre-
dissipation (source-sink term) of the moving parcel. Our sented methodology, four precipitation events in a sub-
case is a bit different, because we are looking at pre- domain of 2720 km 3 2720 km are selected. Figure 3
dictability relative to a forecast method. The exact tra- depicts one representative image of each system, and
jectory of a precipitation parcel is unknown at the mo- Table 1 lists the duration, the spatial extent, and two
ment the forecast is issued, but is estimated from past summary statistics of the intensity distribution.
observations in the assimilation window. With this set- The only criteria for selecting the events are a min-
up, illustrated in Fig. 2 (and hereafter referred to as imum extent of about 3 3 10 5 km 2 and sufficient data
forecast mode), precipitation decorrelates more rapidly quality. The minimum extent criterion makes sure that
than it does in the strict Lagrangian space following the the correlation function c(t), the lifetime L, and other
true trajectories. The difference in decorrelation times scores are based on a sufficiently large sample, and thus
is due to the error of the estimated trajectories (e.g., have meteorological meaning. Echo tracking and ad-
nonstationarity of the echo motion field). vection of rainfall patterns can be done with much small-
er systems, but the evaluation of the forecasts may give
noisy results because a sample size is too small. To
b. Data obtain a meaningful estimate of the Lagrangian persis-
The U.S. national radar composites used in this anal- tence of small-scale systems, we would have to pool
ysis are a WSI Corporation NOWradTM product, which several events of this type before calculating lifetimes.
benefits from three stages of quality control. The first Occasionally the composites contain strong residual

Unauthenticated | Downloaded 04/18/21 01:37 PM UTC


2862 MONTHLY WEATHER REVIEW VOLUME 130

FIG. 3. The four precipitation events, used in this study. Indicated is the area with precipitation rates
higher than 0.1 mm h 21 (10 dBZ ). Time is in UTC.

clutter, or large data voids because of missing radars. the shape of the system changes rapidly. At a later point
Such situations are easily detected by visual inspection in this paper we will see that the corresponding Eulerian
of animated images. and Lagrangian decorrelation times are relatively small.
The last column of Table 1 is a measure of the im- In the 1998 case, the rates .10 mm h 21 make up 6.4%
portance of convection. In the 1999 case, 12% of the of the precipitation area. This is still a considerable
raining area has rainfall rates higher than 10 mm h 21 . fraction, but throughout the entire period of 16 h, the
Precipitation is predominantly formed in convective system is more coherent, and the cells are embedded in
cells, the temporal evolution is fast (dC/dt high), and a larger area of weak to moderate rates. Not surprisingly,
we will find longer persistence for this event than for
TABLE 1. Statistics of the four precipitation events used in this
the 1999 event. The echo motion field of both the 2001
study. The extent is defined as the area with a precipitation rate larger and the 1999 case exhibits rotation, while the motion
than 0.1 mm h21 (10 dBZ). The last two columns indicate the fraction fields of the other two cases are almost free from ro-
of that area with rates larger than 1 mm h21 (25 dBZ) and 10 mm tation.
h21 (40 dBZ), respectively. The values are averages over the entire
period. The fractional coverage of rainfall rates greater than 10 mm
h21 over the 16-h duration of the 1998 storm was, for instance, 6.4%.
d. Summary of the procedure
Start Duration Extent .1 mm .10 mm
Date (UTC) (h) (10 5 km 2 ) h21 (%) h21 (%)
30 Jul 1998 0400 16 6.0 42 6.4 1) Determine the motion field by variational echo track-
16 May 2000 0000 22 3.9 34 2.0 ing in the assimilation window (section 3).
25 May 2001 0400 20 6.1 33 2.6 2) Advect the radar reflectivity using a modified semi-
1 Jun 1999 1600 25 3.6 47 12
Lagrangian backward scheme (section 4).

Unauthenticated | Downloaded 04/18/21 01:37 PM UTC


DECEMBER 2002 GERMANN AND ZAWADZKI 2863

3) Compare the forecast and observation and calculate Experiments have shown that the vector field can be
the correlation and skill scores (section 6a). accurately retrieved if a ‘‘close’’ first guess of the con-
4) Repeat steps 1 to 3 with spectral-filtered images to trol variables is available. To obtain a first guess, Lar-
determine the scale-dependence of predictability oche and Zawadzki (1994) developed a scaling-guess
(section 6b). procedure, in which the field is iteratively retrieved with
increasing grid resolution. This strongly reduces the risk
that the minimization converges towards secondary
3. Variational echo tracking minima. Figure 7 illustrates the scaling guess of the
The echo motion field is estimated by variational echo retrieval shown in Figs. 4 and 5. The left-most plot
tracking (VET), first presented by Laroche and Zawa- shows a single motion vector for the whole domain.
dzki (1994, 1995). The technique was originally de- This vector can be obtained either by means of the con-
veloped to retrieve the three-dimensional wind field jugate-gradient algorithm as is done for the other scales,
from single-Doppler clear-air echoes. We adapted it to or by a simple search over a set of predefined vectors.
continental-scale radar composite images of precipita- In an operational context the scaling-guess procedure
tion fields with no Doppler information. The spatial var- can be replaced by using as a first guess the latest motion
iation of the error structure of the radar data is taken field, or the 700-hPa winds from numerical models. This
into account by means of a weighting vector. Regions will reduce the processing time of the constrained non-
with frequent clutter or poor visibility—for instance, in linear minimization by at least one order of magnitude.
the mountains—are given small weights. In a first step the weighting vector b (x) is set to 1
A cost function with two constraints is minimized for the whole domain. Data quality is thus assumed to
be constant, and all pixels are given the same weight.
JVET (u) 5 JC 1 J 2 (9)
Data of bad quality, such as partially blocked rays or
The first constraint, JC , is the sum of squares of residuals isolated cluttered pixels, cause little harm in the retrieval
of the conservation equation [Eq. (6)], and is calculated of the motion field, as long as their spatial extent is
with an upstream semi-Lagrangian scheme small. First, the grid scale of the motion field is between

JC 5 EE V
b(x)[C(t 0 , x)
one and two orders of magnitude larger than the pixel
resolution of the radar image (here, 104 km compared
to 4 km). Second, the smoothing constraint does not
permit strong discontinuities between close vectors,
2 C(t 0 2 Dt, x 2 uDt)] 2 dx dy, (10)
which further reduces the influence of small regions of
where b (x) is the weighting vector representing the data bad data.
quality. The second term of Eq. (9), J 2 , is a smoothness- A next step is to determine b (x) by examining the
penalty function (Wahba and Wendelberger 1980) spatial distribution of data quality of the U.S. radar com-

EE 1
posites. This is particularly important if the domain of
2 1 2 1 2 1 2
]2u ]2u ]2u ] 2y
2 2 2 2

J2 5 g 1 12 1 the analysis is extended to regions with lower data qual-


V
]x 2 ]y 2 ]x]y ]x 2 ity and more data voids, such as over the Rocky Moun-
tains.

1 2 1 2
] 2y ] 2y
2 2
The scale of the features to be tracked is controlled
1 12 dx dy, (11) by the scaling-guess procedure and the smoothing con-
]y 2 ]x]y
straint. Here, the scale of the motion field is chosen
where g is a constant weight, and u and y are the x- small enough to allow for large-scale rotation, defor-
and y-component of u, respectively. All the control var- mation, and differential motion between different sys-
iables (u and y of the domain) are retrieved simulta- tems. But it is large enough to avoid tracking the fast,
neously in one global minimization. Several techniques perishable convective scales. A smaller scale may be
can be found in the literature to minimize large-scale suitable for nowcasts of a few 10s of minutes.
nonlinear cost functions (Navon and Leger 1987). In Once the motion field is determined, it is taken un-
VET, we use the conjugate-gradient algorithm to deter- changed during the entire forecast period (Fig. 2). That
mine the search direction and the step length. is, rainfall patterns are advected assuming stationary
Figure 4 shows an example of the echo motion field motion. In a forecast mode the motion field is updated
obtained by VET using three radar images of a 1-h every hour.
assimilation window. The lower two plots are combined
and enlarged in Fig. 5. An example of a second case is
given in Fig. 6. 4. Advection of radar reflectivity
The retrieved vector field is most reliable in the re-
gions of precipitation echoes, whereas far away from We tested several advection schemes and finally end-
any precipitation area the vectors must be interpreted ed up with a modified version of the semi-Lagrangian
with care. They are either extrapolated values, or result scheme presented by Sawyer (1963) and further devel-
from tracking a few isolated cluttered pixels. oped by Robert (1981). Figure 8 illustrates the displace-

Unauthenticated | Downloaded 04/18/21 01:37 PM UTC


2864 MONTHLY WEATHER REVIEW VOLUME 130

FIG. 4. Three radar composite images and corresponding echo motion field obtained by variational echo
tracking (VET). Time step is 30 min, the last image is for 0800 UTC 25 May 2001. Indicated is the area
with precipitation rates larger than 0.1 mm h 21 (10 dBZ ). For better legibility only one gray shade is shown,
whereas echo tracking is done with the full range of dBZ values. The vectors are most reliable in regions
with precipitation echoes (black arrows), whereas far away from any precipitation they must be interpreted
with care (light arrows). For enlargement see Fig. 5.

ment vectors of four schemes: two backward schemes Strictly speaking, only forward schemes are mass
and two forward schemes. conservative, which means that the mean of the ad-
vected field is conserved. In practice, the divergence of
the echo motion field is weak, and also backward
a. Backward versus forward schemes are, to a first approximation, mass conserva-
In a forward scheme we start at grid point P and tive.
advect the parcel downstream up to point Q. Thus, ‘‘for- The problem of forward schemes is choosing the op-
ward’’ means forward in time and downstream in space. timum radius of influence when redistributing (spread-
Whereas in a backward scheme, we move upstream and ing) the advected value to the neighbouring grid points.
determine the origin O of a parcel that would end up The radius must be large enough to smooth ripples
at grid point P. Normally, neither O nor Q exactly co- caused by divergence, but small enough to limit the loss
incide with a grid point. Therefore, in a forward scheme of power at small scales.
the advected value is redistributed to the neighbouring
grid points, while in a backward scheme interpolation b. Constant vector versus semi-Lagrangian
is required to determine the value at point O. Both re-
distribution and interpolation result in a loss of power The constant-vector scheme is straightforward: the
at small scales, as is evidenced in the power spectra displacement vector a of Eq. (8) is simply
discussed in section 4c. a 5 t u(t 0 , x P ), (12)

Unauthenticated | Downloaded 04/18/21 01:37 PM UTC


DECEMBER 2002 GERMANN AND ZAWADZKI 2865

c. Advection schemes and numerical diffusion


In all of the schemes discussed previously precipi-
tation must be either interpolated from or redistributed
to neighboring grid points. For forward schemes we use
Gaussian spreading with a radius of influence that is
proportional to the length of the displacement vector.
Whereas in the backward schemes a four-point bilinear,
or a 16-point cubic, interpolation is applied.
Both spreading and interpolation result in the deg-
radation or loss of small-scale features, known as im-
plicit numerical diffusion or false dissipation. Simula-
tions of mass conservation, numerical diffusion, dis-
persion, and monotonicity of commonly used advection
schemes can be found in Ostiguy and Laprise (1990).
Numerical diffusion changes the power spectrum of the
precipitation field (Fig. 10) in an uncontrollable way,
and thus interferes with one of the objectives of the
project, which is to determine the scale-dependence of
predictability. For this reason, we have to keep numer-
ical diffusion as small as possible. One way to com-
FIG. 5. Enlargement of Fig. 4: Radar composite image and cor- pletely avoid this problem is to use one single translation
responding echo motion field of 0800 UTC 25 May 2001. Levels of vector for the entire domain. Certainly, this way we
gray shading correspond to reflectivity between 10–25 dBZ, 25–40
dBZ, and larger than 40 dBZ, respectively. To improve legibility the
preserve all the small scales of the precipitation field
radar pattern has been low-pass filtered with a cutoff wavelength of itself, and the forecast will have identically the same
32 km. power spectrum as the field at time t 0 . But by ignoring
the small scales of the motion field, it again prohibits
analysis of the predictability of small-scale precipitation
patterns. That is, the small scales artificially decorrelate
where u(t 0 , x P ) is the echo motion at the grid point P. faster than they would in a real Lagrangian space, be-
The constant-vector approach uses, for each grid point, cause they are advected to the wrong place. The solution
one constant translation vector and does not allow for to this dilemma is a modified version of the semi-La-
rotation. To overcome this drawback we can use a semi- grangian backward scheme. In the conventional semi-
Lagrangian scheme: The advection is divided up into N Lagrangian scheme interpolation is done for each time
steps of length Dt with NDt 5 t, and for each time step step Dt. The result is a serious loss of small-scale power,
a is iteratively determined as follows: which increases with the number of time steps. The
lowest curve of Fig. 10 shows the power spectrum of

1 2
a a 6-h forecast obtained by means of the conventional
a 5 Dt u t 0 , x 2 (13) semi-Lagrangian scheme with a time step Dt of 15 min
2
and bilinear interpolation. The third-lowest spectrum is
starting with a 5 0. The final displacement vector is from the same scheme but with cubic interpolation. In
the vectorial sum of the N vectors of the individual time our context, we can always reduce the number of in-
steps. Thus, in the semi-Lagrangian scheme we deter- terpolations to one, independently of the number of time
mine the trajectory of a parcel by following the stream- steps, by following the streamlines back to the origin
lines either upstream or downstream assuming station- of the parcel by means of Eq. (13), and then interpolate
arity, that is u(t, x) 5 u(t 0 , x). For the echo motion the precipitation rate once. Since both the constant-vec-
fields obtained by VET (section 3), Eq. (13) converges tor backward bilinear and the modified semi-Lagrangian
rapidly and the iteration can be aborted after 2 or 3 backward bilinear scheme use one interpolation, they
loops. Figure 8 shows the vectors for semi-Lagrangian have almost identical power spectra. Numerical diffu-
forward (slf ), semi-Lagrangian backward (slb), con- sion can be further reduced by replacing the bilinear
stant-vector forward (cvf ), and constant-vector back- with a cubic interpolation.
ward (cvb). If rotation is not negligible, as on synoptic We decided to use the modified semi-Lagrangian
scales, the semi-Lagrangian approach is obviously to be backward scheme (semi-Lagrangian IO, ‘‘IO’’ stands for
preferred. Figure 9 shows a forecast obtained with a interpolate once) because first, it allows for rotation,
semi-Lagrangian scheme for a rotating system. It qual- second, it is almost conservative in mass, and third, it
itatively shows that the scheme is able to conserve the limits the loss of power at small scales that results from
shape of the precipitation system even in the presence interpolation. An example of a forecast based on the
of strong rotation. semi-Lagrangian IO scheme is given in Fig. 9.

Unauthenticated | Downloaded 04/18/21 01:37 PM UTC


2866 MONTHLY WEATHER REVIEW VOLUME 130

FIG. 6. Same as Fig. 4 except for 1100 UTC 30 July 1998.

d. Source-sink term al. 1998; Pierce et al. 2000). The role of daytime forcing
becomes evident when looking at the frequency of thun-
A source-sink term is incorporated in the forecasting
derstorms versus daytime. Based on U.S. national radar
procedure to allow for growth and dissipation of pre-
composites, Carbone et al. (2002) found diurnal cycles
cipitation. One way to estimate this term is determining
over and near the eastern and western cordillera, and
in the assimilation window the residuals of the conser-
semidiurnal cycles between the cordillera.
vation equation by means of an upstream semi-Lagrang-
ian scheme. It thus corresponds to the residuals that are
minimized with the first constraint JC of the cost func- 5. Use of logarithmic reflectivity as the prognostic
tion of VET (section 3). This improves the forecast if variable
growth and dissipation, following a precipitation parcel
(dC/dt), exhibit a certain degree of persistence, as can So far we referred to C as the precipitation field with-
be observed on synoptic scales. out having specified exactly which variable we are look-
In particular for convection at the storm-scale, how- ing at. There was no need to do so up to this point,
ever, the persistence of the total derivative is low, and because the proposed methodology can be applied to
other factors such as daytime and synoptic forcing dom- assess the predictability of virtually any field variable,
inate over advection. Tsonis and Austin (1981) have provided that advection is a reasonable forecasting
shown that extrapolating changes in intensity and size method for that variable. Examples of such variables
yield little if any improvement at the storm scale. Thus, are the rainfall rate, precipitation liquid water, cloud
recent techniques developed for nowcasting the initia- liquid water, water vapor, or ‘‘dry’’ quantities such as
tion and dissipation of intense convection, combine ad- aerosol concentrations, or, if radiative forcing is sub-
vection of radar echoes with external information ac- tracted, also air temperature.
counting for daytime and synoptic forcing (Wilson et From the instrumental point of view we have to

Unauthenticated | Downloaded 04/18/21 01:37 PM UTC


DECEMBER 2002 GERMANN AND ZAWADZKI 2867

FIG. 7. Scaling-guess procedure: The echo motion field is retrieved in three runs with increasing resolution.
The retrieval starts with (left) a uniform field, which is used as a first guess to retrieve (middle) the field
on a 5 3 5 grid, which in turn is the first guess of (right) the final minimization with a 25 3 25 grid.

choose a variable that can be measured over a large or when determining the correlation between the fore-
domain at a sufficiently high temporal and spatial res- cast and the observation).
olution, like radar reflectivity or brightness temperature For several reasons we decided to use logarithmic
derived from satellite data. radar reflectivity (dBZ) throughout the analysis: First,
Weather radars are a unique source of observations the strength of the continental radar composites is its
to investigate the predictability of precipitation. Yet, we ability to give an overall view of the precipitation pat-
have to decide whether to use logarithmic reflectivity tern up to the synoptic scale in terms of multiplicative
(dBZ), or any derived quantity like linear precipitation rainfall rates (using classes such as 1–2 mm h 21 , 2–4
rates (mm h 21 ), or precipitation liquid water (kg m 23 ). mm h 21 , etc.), rather than to provide precise measure-
In some steps of the proposed chain of techniques, it ments of the rainfall amounts (section 2b). In the linear
does not matter what variable we are using (for instance domain too much attention is paid to the upper end of
when calculating skill scores like the false-alarm rate), intensities (strong rainfall rates), which is affected by
whereas in other steps it does (e.g., when calculating brightband contamination and remaining ground clutter.
JC in VET, when interpolating in the advection scheme, In the logarithmic domain, on the other hand, equal
weights are given to the intervals 0.1–1 mm h 21 , 1–10
mm h 21 , and 10–100 mm h 21 . Consequently, logarith-
mic reflectivity better matches with the information con-
tent and quality of the continental radar composites.
Second, relative errors are often of more importance
than absolute errors. An error, for instance, of 5 mm
h 21 between a predicted value of 6 mm h 21 and an
observed rate of 1 mm h 21 is large. And yet the same
absolute error becomes almost negligible in a cell of
heavy precipitation with rates above 50 mm h 21 . One
way to look at relative errors is to work in the loga-
rithmic domain.
Third, reflectivity is the quantity directly measured
by the radar.

6. Results
a. Predictability of precipitation
Figure 1 depicts the correlation, as defined in Eq. (2),
between forecast and observation for one single 9-h
period (1-h assimilation window plus 8-h forecast). In
order to obtain a better estimate of Lagrangian persis-
tence we average the correlation of several 9-h periods
FIG. 8. Displacement vectors of four advection schemes: semi- by extending the domain V of Eq. (2) and integrating
Lagrangian forward (slf ), semi-Lagrangian backward (slb), constant-
vector forward (cvf ), and constant-vector backward (cvb). P is the over # # #V . . . dx dt 0 . Figure 11 shows the Lagrangian
grid point of a parcel for which one wants to find out where it comes and Eulerian persistence of four events, each lasting
from (backward), or where it goes (forward). between 16 and 25 h. In the 25-h period, for instance,

Unauthenticated | Downloaded 04/18/21 01:37 PM UTC


2868 MONTHLY WEATHER REVIEW VOLUME 130

FIG. 9. Observation and forecast obtained by semi-Lagrangian advection for lead times of 1, 2, and 3 h. Depicted are the regions with
reflectivity larger than 10 dBZ (0.1 mm h 21 ). The forecast starts at 0800 UTC 25 May 2001 (Fig. 5). For better legibility only one gray
shade is shown, whereas advection of patterns and evaluation of the forecast is done with the full range of values.

we can fit 17 9-h periods, when issuing one forecast per b)(a 1 c)/(a 1 b 1 c 1 d). We can thus define the
hour. To a first approximation, Eulerian persistence de- skill scores as follows:
correlates twice as fast as Lagrangian persistence. In POD 5 a/(a 1 b) (14)
other words, by advecting echoes following a stationary
motion field the forecast time for a given level of con- FAR 5 c/(a 1 c) (15)
fidence can be doubled. For all four events the solid a
lines are almost straight, and can therefore be approx- CSI 5 (16)
imated by an exponential law. The triangles in Fig. 11 a1b1c
indicate the lifetime, defined as the integral of the cor- a2w
relation function over t [Eq. (3)], for Lagrangian per- ETS 5 . (17)
a1b1c2w
sistence. The exact values are, from left to right, 7.7,
8.7, 5.2, and 4.4 h, respectively. In some cases, we We prefer the ETS to the more common CSI for the
observed lifetimes far beyond 8 h, as illustrated in Fig. following reasons: the ETS is 0 for constant forecasts
12. The high variability of persistence agrees with what of either category (either a 1 c or b 1 d is 0) as well
we expect from visual observation of large-scale radar as for random forecasts (a/b 5 c/d), while a perfect
composites. Small isolated thunderstorms typically have forecast gets a value of 1. Figure 13 depicts the POD,
short lifetimes of a few 10s of minutes, and exhibit low FAR, CSI, and ETS for a threshold of 0.1 mm h 21 for
persistence, both Eulerian and Lagrangian. Whereas the same datasets as used in Fig. 11. At the point where
large coherent systems, which on an overall average the POD and FAR reach 0.5 the ETS is approximately
usually have smaller rainfall rates, are much more per- 0.3.
sistent in time, and predictable, by simple advection, up Neither the correlation depicted in Fig. 11 nor the
to several hours ahead. skill scores of Fig. 13 provide a direct measure of how
The correlation as defined in Eq. (2) is one way to accurate the forecast is. We thus calculate another pa-
assess the predictability. Another way is to look at skill rameter, the conditional mean absolute error (CMAE)
scores, such as the probability of detection (POD), the defined as follows:
false-alarm rate (FAR), the critical success index (CSI),
and the equitable threat score (ETS). These scores de-
scribe the skill of predicting the occurrence of precip-
CMAE 5
1
a
O |C(t
a
ˆ 1 t , x) 2 C(t 1 t , x)|
0 0 (18)

itation above a given threshold rate. The forecast image A difference in terms of logarithmic reflectivity can be
is of binary type: for each pixel we say event or no- attributed to a factor in terms of linear precipitation
event and define terms as follows: a 5 hits, b 5 misses, rates. Therefore, CMAE in dBZ corresponds to a relative
c 5 false alarms, d 5 correct negatives, and w 5 (a 1 error in mm h 21 . Assuming an exponent of 1.5 in the

Unauthenticated | Downloaded 04/18/21 01:37 PM UTC


DECEMBER 2002 GERMANN AND ZAWADZKI 2869

FIG. 12. Eulerian persistence (dotted) and Lagrangian persistence


(solid) for a subdomain of the 16 May 2000 data. The area of the
precipitation system is about 3 3 10 5 km 2 . Lifetime L is 15.4 h.
FIG. 10. Numerical diffusion of six advection schemes for a 6-h
forecast. OBS is the image at time t 0 . Here, IO stands for interpolate
once, as opposed to the conventional semi-lagrangian scheme where
interpolation is done several times. what extent can the predictability be extended by ig-
noring the small scales and focusing on features at larger
scales? To know the scale-dependence of predictability
Z–R relation (Smith and Joss 1997) a CMAE of 9 dBZ of precipitation is crucial for several applications, for
corresponds to a factor of 4 for precipitation rates. The instance, (i) when designing automated nowcasting and
CMAE combined with the POD provide crucial infor- warning systems, (ii) to compare the skill of two fore-
mation to the user of the forecast. For the 16 May 2000 casts with different resolutions, and (iii) in advection
case and a lead time of 3 h, Figs. 13 and 14 indicate a nowcasting procedures to prevent features from being
POD of 60% and a CMAE of 7.6 dBZ. Thus, the oc- reproduced beyond their lifetime (Bellon and Zawadzki
currence of precipitation has been correctly predicted 1994).
for 60% of the precipitation area, and for this area the A simple method of scale decomposition is to convert
average error is a factor of 3.2 in terms of rain rates. the observation and forecast into binary maps (non-
event–event) using a threshold in intensity (Austin and
Houze 1972), and compare the binary maps by means
b. Scale-dependence of predictability
of skill scores. The reason why a threshold in intensity
The results presented in the previous section are all can be used to separate different scales lies in the in-
based on pixel-by-pixel comparisons at a resolution of tensity–size relationship. Figure 15 shows the ETS of
4 km. Features at scales of a few kilometers typically two events and a threshold of 0.1 mm h 21 , 1 mm h 21 ,
have short lifetimes (a few 10s of minutes, Orlanski and 10 mm h 21 . During the 1998 event there is a large
1975) and significantly contribute to the decorrelation gap between the lifetimes of 1 mm h 21 and 10 mm h 21
of Lagrangian persistence. A legitimate question is, To features. For the 10 mm h 21 threshold the ETS reaches

FIG. 11. Correlation between forecast and observation for Eulerian persistence (dotted) and Lagrangian persistence (solid) of four events.
The triangles mark the lifetimes L of the solid curves estimated by means of Eq. (3).

Unauthenticated | Downloaded 04/18/21 01:37 PM UTC


2870 MONTHLY WEATHER REVIEW VOLUME 130

FIG. 13. Skill scores for a threshold of 0.1 mm h 21 .

0.3 already at a lead time of 0.5 h, while the same level 7. Discussion and outlook
of confidence is reached at about 3 h for the 1 mm h 21
scale, and at 4 h for the 0.1 mm h 21 scale, which contains A methodology is presented to determine the pre-
the entire precipitation system. dictability of precipitation in terms of the Eulerian and
To obtain a more complete picture of the predict- Lagrangian persistence of patterns from the storm scale
up to synoptic scales.
ability as a function of the threshold rate, we determine
The motion field of radar echoes is retrieved from a
for nine thresholds the lead time at which the ETS falls
time series of images by minimizing a cost function
below 0.3. Figure 16 shows the corresponding curves
with two constraints: the residuals of the reflectivity
for the four events. On the basis of this type of diagram, conservation equation and a smoothing penalty func-
the skill for quantitative precipitation forecasting by nu- tion. Because of growth and dissipation of precipitation,
merical modeling can be evaluated against Lagrangian the conservation equation of reflectivity is not perfectly
persistence. fulfilled. Therefore, it is used here as a weak constraint
A more rigorous way to determine the scale-depen- that allows for model errors. The degree of smoothing
dence of predictability is to decompose the radar images and the resolution of the motion field can be easily
according to scales and to repeat the procedure discussed controlled by parameters. The error structure of the radar
in the previous sections. The images are decomposed data is considered by utilizing a weighting vector. Re-
by means of a low-pass filter in the spectral domain gions where radar measurements have low quality are
using discrete-cosine-transforms (DCT; Ahmed et al. given small weights. The variational approach as op-
1974; Gonzalez and Woods 1992). Then, decomposed posed to simple tracking by correlation has the advan-
images are advected and compared to decomposed ob- tage of incorporating several constraints, each weighted
servations. The scale-dependence of the lifetimes of the according to its importance. External information can
four events is shown in Fig. 17. be easily incorporated, either by adding a new con-
A next step (not taken here) is to replace the DCT straint, or by using it as a first guess from which the
by wavelet transforms and spatial smoothing, possibly minimization starts.
anisotropically, and to repeat the same analysis. Radar precipitation patterns are then advected by

FIG. 14. Conditional mean of absolute error (CMAE) between forecast and observation [Eq. (18)].

Unauthenticated | Downloaded 04/18/21 01:37 PM UTC


DECEMBER 2002 GERMANN AND ZAWADZKI 2871

FIG. 15. Equitable threat score (ETS) for 30 July 1998 data (16 h,
thick lines) and for 16 May 2000 data (22 h, thin lines).

means of a modified semi-Lagrangian advection scheme


assuming stationary motion. The combination of the
variational echo tracking with semi-Lagrangian advec-
tion allows for rotation, which is mandatory when work- FIG. 16. Lead time at which ETS drops below 0.3 as a function of
ing on synoptic scales. threshold rate.
When using single-radar images, the lead time of use-
ful forecasts obtained by extrapolation of radar echoes
is strongly limited by the maximum range of the radar. suffer to some extent from brightband contamination,
A net gain in terms of forecast quality is observed when remaining ground echoes, and sampling error due to the
going to continental composites, which entirely reveal 15-min time step. Particular difficulties occur when
mesoscale precipitation systems such as frontal systems compositing data of several instruments, here more than
or lines of organized convection moving over the con-
tinent. Using an ETS of 0.3 as a limit of predictability
we get lead times for the four analysed systems of 4,
6, 2.5, and 2.5 h, respectively (Fig. 13).
The lifetime of precipitation patterns in Eulerian and
Lagrangian space are compared. According to our pre-
liminary results, Eulerian and Lagrangian persistence
differ by roughly a factor of 2 (Fig. 11). The time of
predictability for a given level of confidence can there-
fore be doubled by simple Lagrangian advection. In
other words, advection explains a significant part of the
variation of the precipitation rate at a given location.
The range of predictability increases with increasing
scale (Figs. 15, 16, and 17). Using a cutoff wavelength
of 64 km, and thus ignoring all the variation at the
thunderstorm scale, increases predictable Lagrangian
lifetimes over that of unfiltered images by 57, 53, 47,
and 39 min for the 1998, 2000, 2001, and 1999 events,
respectively. The slope of the curves in Fig. 17 reflect
to what extent the lifetime increases as a function of
scale. If the predictability is predominantly limited by
the smallest scale present in an image, then filtering
efficiently increases the lifetime. This is the case for the
1998 event, where most of the small-scale features are
embedded in larger rain areas. Note that, based on the FIG. 17. Lifetime as a function of scale. Scales are decomposed
by means of a low-pass filter and discrete-cosine-transforms. The
analyses presented here, nothing can be said about the images at scale 128 km, for instance, contain all variations at wave-
lifetime of scales with wavelengths smaller than 10 km. lengths of 128 km and larger. Power at wavelengths smaller than 128
The U.S. national radar composites used in this study km has been filtered out.

Unauthenticated | Downloaded 04/18/21 01:37 PM UTC


2872 MONTHLY WEATHER REVIEW VOLUME 130

100 radars. Missing radars and different calibration, for Acknowledgments. The authors express their gratitude
instance, cause data gaps and discontinuities in the com- to the Global Hydrology and Climate Center GHRC for
posited image. All these sources of measurement error providing access to the WSI radar composites, and to
contribute to the nugget variance, and result in under- Dr. R. Carbone and his group at NCAR for putting an
estimating the Eulerian and Lagrangian lifetimes. The excellent case selection kit on the web. A special thank-
nugget variance is defined as the discontinuity at zero you goes to Dr. Stéphane Laroche and Alain Caya for
lag of variograms or autocorrelation functions (Math- the fruitful discussions on variational echo tracking.
eron 1963; Germann and Joss 2001). It can be deter-
mined by extrapolating the correlation function c(t) to
REFERENCES
zero lag. If we then shift the correlation function by the
amount of the nugget variance, we get estimates of per-
Ahmed, N., T. Natarajan, and K. R. Rao, 1974: Discrete cosine trans-
sistence that are not affected by measurement errors of form. IEEE Trans. Comput., C23, 90–93.
the radar composites. Only preliminary tests have been Austin, P. M., and R. A. Houze Jr., 1972: Analysis of the structure
performed to do this correction. It has not been applied of precipitation patterns in New England. J. Appl. Meteor., 11,
to the material presented in this paper. In spite of the 926–935.
Bellon, A., and G. L. Austin, 1978: The evaluation of two years of
somewhat lower accuracy inherent to large-scale com- a real-time operation of a short-term precipitation forecasting
posites there is considerable benefit from using conti- procedure (SHARP). J. Appl. Meteor., 17, 1778–1787.
nental radar composites both to investigate the predict- ——, and I. Zawadzki, 1994: Forecasting of hourly accumulations
ability and to improve nowcasting. of precipitation by optimal extrapolation of radar maps. J. Hy-
drol., 157, 211–233.
The results of the few experiments that we performed Carbone, R. E., J. D. Tuttle, D. A. Ahijevych, and S. B. Trier, 2002:
with the source-sink term are mixed: In some cases Inferences of predictability associated with warm season pre-
decorrelation is slower and the forecast skill increases cipitation episodes. J. Atmos. Sci., 59, 2033–2056.
when introducing a source-sink term that has been es- Germann, U., and J. Joss, 2001: Variograms of radar reflectivity to
timated in the assimilation window, while in others no describe the spatial continuity of alpine precipitation. J. Appl.
Meteor., 40, 1042–1059.
significant improvement is observed. For the results pre- Gonzalez, R. C., and R. E. Woods, 1992: Digital Image Processing.
sented in section 6, the source-sink term has been set Addison-Wesley, 716 pp.
to zero. The predictability of growth and dissipation has Laroche, S., and I. Zawadzki, 1994: A variational analysis method
yet to be analyzed before its incorporation in the ex- for retrieval of three-dimensional wind field from single-Doppler
radar data. J. Atmos. Sci., 51, 2664–2682.
trapolation procedure leads to consistent improvement. ——, and ——, 1995: Retrievals of horizontal winds from single-
The analysis will finally be performed with a large Doppler clear-air data by methods of cross correlation and var-
set of data. The lifetime will be determined separately iational analysis. J. Atmos. Oceanic Technol., 12, 721–738.
for different systems, and compared to parameters of Lilly, D. K., 1986: The structure, energetics and propagation of ro-
tating convective storms. Part II: Helicity and storm stabilization.
synoptic forcing (Zawadzki et al. 1994). The results will J. Atmos. Sci., 43, 126–140.
set a standard against which the skill for quantitative Lorenz, E. N., 1973: On the existence of extended range predict-
precipitation forecasting by numerical modeling can be ability. J. Appl. Meteor., 12, 543–546.
evaluated. Lagrangian persistence fails at the point Matheron, G., 1963: Principles of geostatistics. Econ. Geol., 58,
where the source-sink term or nonstationarities in the 1246–1266.
Mecklenburg, S., J. Joss, and W. Schmid, 2000: Improving the now-
motion field become more important than advection. casting of precipitation in an Alpine region with an enhanced
Hence, to beat radar echo extrapolation the model has radar echo tracking algorithm. J. Hydrol., 239, 46–68.
to correctly predict the source-sink term and nonsta- Navon, I. M., and D. M. Leger, 1987: Conjugate-gradient method for
tionarities in the motion field, a rather difficult task. large-scale minimization in meteorology. Mon. Wea. Rev., 115,
1479–1502.
So far, we were talking about the predictability of Orlanski, I., 1975: A rational subdivision of scales for atmospheric
precipitation intensity. There may be other quantities of processes. Bull. Amer. Meteor. Soc., 56, 529–530.
interest. In the same way we can study the predictability Ostiguy, L., and J. P. R. Laprise, 1990: On the positivity of mass in
of the temporal evolution of precipitation (persistence commonly used numerical transport schemes. Atmos.–Ocean, 28,
147–161.
of growth and dissipation), the predictability of the pow- Pierce, C. E., P. J. Hardaker, C. G. Collier, and C. M. Haggett, 2000:
er spectrum, of the intensity probability distribution, or GANDOLF: A system for generating automated nowcasts of
of the spatial variability as described by variograms convective precipitation. Meteor. Appl., 7, 341–360.
(Germann and Joss 2001). Predicting these statistics, in Robert, A., 1981: A stable numerical integration scheme for the prim-
addition to the precipitation intensity, gives a more com- itive meteorological equations. Atmos.–Ocean, 19, 35–46.
Sawyer, J. S., 1963: A semi-Lagrangian method of solving the vor-
plete forecast of the precipitation pattern that has to be ticity advection equation. Tellus, 15, 336–342.
expected at a certain lead time. It crucially depends on Seed, A. W., and T. Keenan, 2001: A dynamic and spatial scaling ap-
the application of the quantities which are most impor- proach to advection forecasting. Preprints, 30th Conf. on Radar
tant. Meteorology, Munich, Germany, Amer. Meteor. Soc., 492–494.
Smith, P. L., and J. Joss, 1997: Use of a fixed exponent in ‘‘adjustable’’
Another interesting experiment will be to repeat the Z-R relationships. Preprints, 28th Conf. on Radar Meteorology,
study with satellite images in order to assess the pre- Austin, TX, Amer. Meteor. Soc., 254–255.
dictability of clouds. Tsonis, A. A., and G. L. Austin, 1981: An evaluation of extrapolation

Unauthenticated | Downloaded 04/18/21 01:37 PM UTC


DECEMBER 2002 GERMANN AND ZAWADZKI 2873

techniques for the short-term prediction of rain amounts. Atmos.– 1998: Nowcasting thunderstorms: A status report. Bull. Amer.
Ocean, 19, 54–65. Meteor. Soc., 79, 2079–2099.
Wahba, G., and J. Wendelberger, 1980: Some new mathematical meth- Zawadzki, I., 1973: Statistical properties of precipitation patterns. J.
ods for variational objective analysis using splines and cross Appl. Meteor., 12, 459–472.
validation. Mon. Wea. Rev., 108, 1122–1143. ——, J. Morneau, and R. Laprise, 1994: Predictability of precipitation
Wilson, J. W., N. A. Crook, C. K. Mueller, J. Sun, and M. Dixon, patterns: An operational approach. J. Appl. Meteor., 33, 1562–1571.

Unauthenticated | Downloaded 04/18/21 01:37 PM UTC

You might also like