FEBRUARY 2004 TURNER ET AL.

231

Predictability of Precipitation from Continental Radar Images.

Part III: Operational Nowcasting Implementation (MAPLE)
B. J. TURNER AND I. ZAWADZKI
Department of Atmospheric and Oceanic Sciences, McGill University, Montreal, Quebec, Canada

U. GERMANN
MeteoSvizzera, Locarno-Monti, Switzerland

(Manuscript received 27 December 2002, in final form 8 August 2003)

ABSTRACT
Filtering of nonpredictable scales of precipitation can be used to improve forecast precision (rms). Previous
papers have studied the scale dependence of predictability of patterns of instantaneous rainfall rate and of
probabilistic forecasts. In this paper, motivated by the often localized, intermittent nature of rainfall, the wavelet
transform is used to develop measures of predictability at each scale. These measures are then used to design
optimal forecast filters. This method is applied to radar composites of rainfall reflectivity over much of the
continental United States and is developed to be appropriate for operational forecasts of rainfall rates and raining
areas. For the four precipitation events studied, the average correlation at 4-h lead time was increased from 0.50
for the original nowcasts to 0.62 with forecast filtering. This forecast filtering is incorporated into the McGill
Algorithm for Precipitation Nowcasting by Lagrangian Extrapolation (MAPLE), which now includes variational
echo tracking, a semi-Lagrangian advection scheme, scale-based filtering, and appropriate rescaling of the filtered
nowcast fields.

1. Introduction Extrapolation (MAPLE), we begin with a review of the

Weather radar is a crucial tool for the short-term (sev- development of nowcasting at McGill.
eral hours) quantitative prediction of precipitation. The The weather radar group at McGill has been active
automated production of such forecasts, called now- in the development of nowcasting algorithms over the
casting, is now widely implemented, and most radar past 30 years. Early study of the correlation structure
networks are equipped with specialized nowcasting soft- of patterns and echo motion by Zawadzki (1973) was
ware (Browning et al. 1982; Walton et al. 1985; Seed applied in the first nowcasting trials of Austin and Bel-
2003). In Part I of this series of papers (Germann and lon (1974). Useful short-term forecasts were produced
Zawadzki 2002, hereinafter Part I), the scale dependence by assuming Lagrangian persistence, such that precip-
of the predictability of precipitation patterns in conti- itation patterns would continue to move with a single,
nental-scale nowcasts was investigated. In Part II, pre- steady velocity without any growth or decay. The dis-
dictability was studied using probabilistic forecasts de- placement vector was determined by maximizing cross
veloped from nowcasts (Germann and Zawadzki 2004, correlation between current and previous observations.
hereinafter Part II). In this paper we take the additional The resulting implementation in the short-term precip-
step of applying measures of scale-dependent predict- itation forecasting procedure called SHARP for opera-
ability to the forecast method itself, designing forecast tional use by forecasters in Canada is described in Bel-
filters that reduce rms errors. With appropriate trans- lon and Austin (1978). Since then, the McGill group
formations of the filtered nowcasts, forecast skill ac- has worked on several enhancements to the basic meth-
cording to the critical success index can also be im- od. Bellon et al. (1980) applied nowcasting to geosta-
proved. Because these forecast filters and subsequent tionary satellite images of clouds in the visible and in-
transformations are incorporated into the McGill Al- frared in a system called RAINSAT. An important neg-
gorithm for Precipitation Nowcasting by Lagrangian ative result was the unsuccessful attempt of Tsonis and
Austin (1981) to improve forecasts by accounting for
rainfall tendencies. By 1982, the single displacement
Corresponding author address: Dr. Barry Turner, Department of
Atmospheric and Oceanic Sciences, McGill University, 805 Sher-
vector in the original SHARP was replaced by a nine-
brooke W, Montreal QC H3A 2K6, Canada. vector motion field. As a further development, a vari-
E-mail: [email protected] ational method was developed to derive the field of

q 2004 American Meteorological Society

Unauthenticated | Downloaded 04/18/21 01:39 PM UTC

232 JOURNAL OF APPLIED METEOROLOGY VOLUME 43

motion of radar patterns with any predetermined reso- TABLE 1. Summary of the separation of the mesoscale suggested
by Orlanski (1975).
lution (Laroche and Zawadzki 1995).
Predictability of Lagrangian persistence forecasts of Spatial scale Typical
precipitation was studied by Zawadzki et al. (1994). Name (km) Time scale weather feature
Predictability of each spatial scale was found to be time Meso-a 200–2000 Days Front
dependent and related to synoptic-scale parameters. Se- Meso-b 20–200 Hours Squall line
lective filtering of perishable scales improved forecast Meso-g 2–20 Minutes Localized thunderstorms
precision (Bellon and Zawadzki 1994). Seed (2003) in-
terpreted these results in the context of a cascading mul-
tifractal description of precipitation (Schertzer and predict? Lorenz (1969) made a theoretical study of the
Lovejoy 1987) and applied Fourier low-pass filters to predictability for a fluid, such as the atmosphere, that
improve forecast scores. Part I of the present series of possesses many scales of motion. He defined the range
papers also studied the scale dependence of predict- of predictability as the time until forecast errors exceed
ability for continental-scale Lagrangian nowcasts using some prechosen magnitude. He also noted that errors in
Fourier (discrete cosine transform) low-pass filtering, predicting a thunderstorm might become large within
but without invoking any assumptions of scaling. In Part 15 min, while errors in predicting larger-scale atmo-
II, scale dependence for probabilistic Lagrangian now- spheric features may not become large for days. This
casting using a high-resolution field of echo motion was is reflected in the time scales mentioned in Table 1.
examined. Now, in Part III, we take advantage of the Of course, the range of predictability (hereinafter re-
findings of Part I to improve operational nowcasting ferred to simply as the predictability) at each scale will
algorithms. By evaluating predictability at each spatial depend on the forecast method being used. Even a cli-
scale, nowcasts are filtered to reject unpredictable scales matological prediction has some limited skill, and a per-
and to optimize forecast precision. sistence forecast may be useful for short-range forecasts.
Here also, we take an approach that does not require With details on storm motion, a nowcast gives useful
any assumptions of scaling. Section 2 develops a frame- predictions for up to several hours. Numerical models
work for the multiscale analysis of predictability and containing details of atmospheric dynamics, combined
the design of forecast filters using the wavelet transform. with sufficient measurements as initial conditions, may
This local transform provides a more appropriate rep- have some skill up to several days. However, if the same
resentation of a precipitation field as a sum of localized forecast method is applied to various precipitation
rainfall pulses at each scale. In section 3 wavelet-based events, differences in forecast errors will reflect the pre-
forecast filters are applied to nowcasts based on radar dictability of each event. In some cases, frontal systems
composite observations of precipitation. Optimal and move for many hours over hundreds of kilometers with
more operationally useful near-optimal forecast filters little change. Nowcasting performs well in these situ-
provide significant gains in forecast precision as mea- ations. In other cases, a previously stable air mass reach-
sured by correlation or root-mean-square (rms) errors. ing a strong source of surface heating and moisture may
Appropriate transformation of filtered forecasts also im- result in rapid storm growth, which will not be predicted
proves skill scores for predictions of rain areas above by a Lagrangian extrapolation nowcast. With a partic-
a certain threshold. Section 4 discusses the potential for ular forecast method, some precipitation features are
these wavelet-based forecast filters, the effects of fil- predictable and others are not. Attempts to forecast un-
tering, and directions for future work. predictable features will only contaminate the skill of
the forecast; forecasts that focus on what is predictable
will have greater accuracy.
2. Method Selective filtering of smaller scales from nowcasts of
a. Motivation precipitation can reduce forecast errors. Although the
division between which scales are predictable will vary
Precipitation patterns include variability over a wide between cases and depend strongly on the forecast lead
range of scales. Various groups have sought classifi- time, the generally expected pattern is that smaller scales
cation schemes that would be convenient for defining will be less predictable. Bellon and Zawadzki (1994)
areas of study. One such classification for the mesoscale showed the reduced predictability of smaller scales in
(Orlanski 1975) is shown in Table 1. A localized, short- precipitation by studying spatially smoothed nowcasts.
lived thunderstorm at the meso-g -scale may be embed- More recently, in Part I and in work by Seed (2003),
ded in a more durable meso-b-scale squall line, which Fourier low-pass filtering has been applied for this pur-
may then be part of a meso-a-scale frontal system that pose.
continues for several days. Such classification schemes The Fourier transform can be used to represent an
are somewhat arbitrary and vary between individual au- image as a combination of waves (sines and cosines) of
thors and research groups. different wavelengths. Because sines and cosines are
We cannot predict every detail of the weather, es- extended in space, the Fourier transform is most useful
pecially precipitation. Which details can we hope to in representing wavelike properties. The forward Fou-

Unauthenticated | Downloaded 04/18/21 01:39 PM UTC

FEBRUARY 2004 TURNER ET AL. 233

rier transform (FT) produces a coefficient F(k) for each wavelet transform and forecast filters. When showing
wavenumber k: actual results, the length scale will be used because it
FT conveys physical meaning more directly. A more math-
f (x) → F(k), (1)
ematical introduction to the wavelet transform is found
21
while the inverse Fourier transform (FT ) reconstructs in Burrus et al. (1998) and other similar texts.
the original spatial field: A wavelet filter can be designed to selectively weaken
FT 21 or remove less predictable precipitation features from a
F(k) → f (x). (2)
forecast. Appropriate spatially localized filters are as-
High-wavenumber ks correspond to small scales, lower- sociated with particular wavelet functions. The simple
wavenumber ks correspond to larger scales, and k 5 0 block averaging of Bellon and Zawadzki (1994) is an
is associated with the average at the largest scale. Fou- example of a localized spatial smoothing function that
rier low-pass filtering eliminates contributions from is less susceptible to the Gibbs effect, as shown in ap-
wavenumbers larger than some cutoff k o . After per- pendix A. Block averaging is the two-dimensional
forming the forward Fourier transform, coefficients smoothing function associated with the Haar wavelet
F(k), k . k o are set to zero. The modified set of F(k) (described in appendix B), and acts as a wavelet low-
is then used in the inverse transform. In this way, only pass filter.
low-wavenumber (large scale) features pass through the Of course, the usefulness of forecast filtering depends
filter. on the selective weakening and removal of features only
However, rainfall fields usually do not have clear, at scales that have lost predictability. Using the wavelet
periodic extended patterns. Fields of rainfall rate (and transform for scale decomposition of forecast errors,
radar reflectivity) regularly show highly localized fea- optimal filters can be developed. This is described in
tures with sharp spatial gradients. Fourier low-pass fil- the next section, with more details concerning the wave-
tering is not appropriate in such circumstances, and re- let transform in appendix B.
sults in the Gibbs effect, which adds nonphysical ‘‘rip-
ple’’ features. The Gibbs effect (illustrated later in Fig.
A1) is described in greater detail in appendix A. b. Errors and filtering
The wavelet transform provides an alternative scale In this section we will design a forecast filter that is
decomposition in such circumstances. The localized optimized to reduce forecast errors. The loss of pre-
pulse functions used by the wavelet transform are more dictability for features of a given scale is not abrupt,
effective in representing localized, intermittent fields. but rather decreases smoothly in time. Thus, in opti-
Instead of the single Fourier transform coefficient F(k) mizing the scale filtering we take advantage of whatever
for each Fourier scale k, we obtain many wavelet co- predictability each scale has at a given forecast time. A
efficients W(m, x) at each wavelet scale m—one for each measure of forecast error is established, then expressed
position x. The forward wavelet transform (WT) pro- as a sum of contributions from different scales. After
duces the following coefficients W(m, x): defining a general forecast-filtering method using wave-
WT
f (x) → W(m, x). (3) let scale weights, the choice of scale weights is directed
by the need to reduce error contributions at all scales.
At a given position in space, x, small m will correspond Scale weights will be small for less predictable scales,
to smaller scales, and large m will correspond to larger reducing their influence.
scales. In practice, a long-term weighted average f (x) Forecast error is often measured as the mean squared
describes a (possibly nonzero) mean around each lo- difference between the forecast and later observed val-
cation. The inverse wavelet transform (WT 21 ) recon- ues. For a set of observations I o and a forecast I f pro-
structs the original spatial field, duced at an earlier time, this is
WT 21
W(m, x), f (x) → f (x). (4) (I f 2 I o ) 2 5 I f2 2 2I o I f 1 I o2 , (6)
A full set of coefficients W(m, x) for all scales and where the overbar indicates averaging over the forecast
positions is highly redundant. In this paper we use wave- area. Less predictable features of the precipitation will
let transform coefficients for a discrete set of wavelet contribute more to the forecast error. We expect smaller
scales m. In fact, we use m as a logarithmic scale index scales to be less predictable than larger scales. To verify
related to the wavelet length scale by this, we need to separate the contributions of different
LengthScale 5 2 m x o , (5) scales to the forecast error. Assuming the independence
of contributions to the error from different scales, we
where x o is the fundamental resolution or pixel length. should then construct a forecast filter that will minimize
The finest resolution x o corresponds to m 5 0, length error contributions from each scale.
scale 2x o corresponds to m 5 1, length scale 4x o cor- In one dimension, the contribution to I f2 from each
responds to m 5 2, and so on. The scale index m 5 0, scale interval m is given by the wavelet spectrum
1, 2, 3 . . . refers to the discrete set of wavelet scales
used, and is useful in describing and developing the S f (m) 5 22m [Wf (m, x)] 2 , (7)

Unauthenticated | Downloaded 04/18/21 01:39 PM UTC

234 JOURNAL OF APPLIED METEOROLOGY VOLUME 43

where the subscript f indicates the forecast, as opposed where the subscript ff denotes the filtered forecast. Be-
to observations that are denoted by a subscript o. The cause the filter is applied only to the forecast, the ob-
mean square I f2 is the sum of these contributions from servation spectrum S o (m) remains unchanged. By Eq.
each scale m and the square of the mean forecast value (9) the wavelet cospectrum of the same observations
I f2 5 I f 1
2
O S (m).
m
f (8)
and the filtered forecast Iff is
Coffo (m) 5 22m Wo (m, x)w(m)Wf (m, x)
A similar expression can be written for I o2 , the mean
square of observation values. In the same way, the con- 5 [w(m)]Co f o (m). (16)
tribution to I o I f from each scale interval m is given by
Following the same development as for Eq. (12), the con-
the wavelet cospectrum:
tribution of each scale m to the filtered forecast error is
Co f o (m) 5 22m Wo (m, x)Wf (m, x), (9)
Errorff (m) 5 [w(m)] 2 S f (m) 2 2[w(m)]Co f o (m)
and I o I f can be expressed as
Io If 5 Io If 1 O Co (m).
m
fo (10)
2 S o (m).
Equation (17) can be minimized by setting the first de-
(17)

Both the wavelet spectrum and cospectrum can be gen- rivative, with respect to w(m), to be zero,
eralized for our two-dimensional forecasts and obser-
vations. More details on the wavelet spectrum and co- ]Errorff (m)
5 2[w(m)]S f (m) 2 2Co f o (m) 5 0, (18)
spectrum are found in appendix B. ][w(m)]
Substituting Eqs. (8) and (10) into Eq. (6), then sep-
arating by scale m, we have and then solving for w(m). The optimal weights will
also depend on the lead time T between making a fore-
(I f 2 I o ) 2 5 (I f 2 I o ) 2 cast and the verification. For scales with S f (m) ± 0, the
1 O [S (m) 2 2Co (m) 1 S (m)].
m
f fo o (11)
optimal scale weights are given by
Co f o (m)
The first term on the right-hand side is the contribution w(m, T ) 5 . (19)
S f (m)
from the largest scale, while the second term is the error
contribution from each scale m: From here on, the scale weights, optimized to reduce
Errorf (m) 5 S f (m) 2 2Co f o (m) 1 S o (m). (12) filtered forecast errors at a particular forecast lead time
T, will be denoted by w(m, T ). These optimal scale
Imagine a perfect forecast, with I o 5 I f at all locations. weights may also vary from one case to another.
In this case the forecast spectrum S f (m) and the co- For a scale that is completely predictable, forecast
spectrum Cofo (m) would each be the same as the ob- patterns will exactly match the observations, so that
servation spectrum S o (m). As required, the error con-
tribution by scale as defined by Eq. (12) would be zero Co f o (m) 5 S f (m) ⇒ w(m, T ) 5 1. (20)
for all scales.
A filtered forecast Iff can be produced by multiplying For a scale that has become completely unpredictable,
the wavelet transform coefficients W(m, x) by a weight- Co f o (m) 5 0 ⇒ w(m, T ) 5 0. (21)
ing factor w(m) at each scale:
0 # w(m) # 1, (13) Scales with intermediate predictability will have optimal
weights w(m) between 0 and 1. For all m, as forecast
and then taking the inverse wavelet transform: lead time T → 0,
WT 21
w(m)Wf (m, x), f (x) → Iff . (14) w(m, T → 0) 5 1, (22)
If we set w(m) 5 1 at some scale, the wavelet coeffi-
cients for that scale are unchanged. If we set w(m) 5 assuming that the observations are compared with them-
0, the wavelet coefficients at that scale will be set to selves because no forecast is required for the time of
zero. What are the values of w(m) that will minimize observation. As T → `, the forecast loses skill and w(m,
the contribution to the forecast error from each scale? T ) decreases. A long forecast lead time renders the fore-
Assuming the independence of contributions to the error cast (on average) no better than climatology. For inter-
from different scales, this should minimize the total mediate times, w(m, T ) will usually increase with in-
forecast error. creasing m, decrease with increasing T, and take on
By Eq. (7), the wavelet spectrum Sff of the filtered values between 0 and 1.
forecast Iff is Appendix B provides more details on the implemen-
tation of these scale weights using a wavelet’s associated
Sff (m) 5 22m [w(m)Wf (m, x)] 2 5 [w(m)] 2 S f (m), (15) smoothing function.

Unauthenticated | Downloaded 04/18/21 01:39 PM UTC

FEBRUARY 2004 TURNER ET AL. 235

FIG. 1. The four precipitation events evaluated in this study. For each case, areas with radar reflectivity rates greater than 10 dBZ (about
0.1 mm h 21 ) are shown. Times are shown in UTC.

3. Observations and results of 2720 km 3 2720 km are used to illustrate the de-
velopment of optimal forecast filters using wavelet sta-
a. Composite radar data
tistics. Figure 1 shows one representative image of each
The U.S. national radar composites used in this anal- system, and Table 2 lists the duration, the spatial ex-
ysis are a Weather Services International (WSI) Cor- tension, and two summary statistics of the rainfall in-
poration NOWrad product that has benefited from three tensity distribution. The criteria for selecting the events
levels of quality control. The original resolution is 15 was a minimum extension of about 3 3 10 5 km 2 and
min in time, about 2 km in space (for east–west dis- sufficient data quality. Some further description of these
tances, there is some dependence on latitude), and 5- four cases may be found in Part I.
dBZ intervals of reflectivity. For the analysis presented
in this paper the spatial resolution has been reduced to
about 4 km, with latitude dependence of east–west dis- b. Motion estimation and nowcasts
tances ignored. Averaging was done in units of dBZ. Motion estimation by variational echo tracking (VET)
The spatial coverage extends from 208 to 528N latitude, was presented by Laroche and Zawadzki (1994, 1995).
and from 1308 to 608W longitude. Reflectivity corre- The technique was originally developed to retrieve the
sponds to the maximum value measured in the vertical three-dimensional wind field from single-Doppler clear-
over each pixel by any Weather Surveillance Radar- air echoes. It was adapted to continental-scale radar
1988 Doppler (WSR-88D) during the time period. composite images of precipitation with no Doppler in-
The accuracy of these reflectivity observations is not formation in Part I.
comparable to that of rainfall maps based on sophisti- The echo motion field in this study is retrieved in two
cated algorithms that include corrections for the vertical runs of the motion estimation algorithm. A first run
profile of reflectivity, visibility, attenuation, Z–R vari- produces an initial coarse guess on a 5 3 5 grid (of 520
ations, and others. The images certainly suffer from km 3 520 km squares). This is then used to initialize
brightband contamination and lack of visibility, which the subsequent minimization on a finer 25 3 25 grid
reduces the quality in terms of quantitative rainfall es- (of 104 km 3 104 km squares).
timates. However, the spatial coverage of the composites The extreme variability of reflectivity fields at the
is unique, making the product particularly suitable for high resolutions can make VET computations slow and
scale-dependent analyses of predictability up to near- unreliable. The VET algorithm may try to match un-
continental scales. related, intense short-lived thunderstorms if there is in-
As in Part I, four precipitation events in a subdomain sufficient spatial smoothing. If the time step between
images being compared is too short, the precipitation
TABLE 2. Statistics of the four precipitation events. The extension
may not have moved enough distance to be properly
is defined as the area with a precipitation rate larger than about 0.1 estimated. These problems are mitigated by appropriate
mm h21 (10 dBZ ). The last two columns indicate the fraction of that spatial smoothing and longer time steps between input
area with rates larger than about 1 mm h21 (25 dBZ ) and 10 mm h21 images. For motion estimation, the coarse 5 3 5 grid
(40 dBZ), respectively. The values are averages over the entire period. estimate was produced using three reflectivity fields 2
.1 mm .10 mm h apart (4, 2, and 0 h before the chosen time), smoothed
Start Duration Extension h21 h21 to 64 km 3 64 km. With the coarse-grid motion estimate
Date (UTC) (h) (10 5 km 2 ) (%) (%) as a reasonable initial guess to stabilize the VET al-
30 Jul 1998 0400 16 6.0 42 6.4 gorithm, the final computations on a 25 3 25 grid used
1 Jun 1999 1600 25 3.6 47 12 three reflectivity fields 30 min apart (60, 30, and 0 min
16 May 2000 0000 22 3.9 34 2.0 before the chosen time), smoothed to 16 km 3 16 km.
25 May 2001 0400 20 6.1 33 2.6
Experimentation showed that these parameters permit-

Unauthenticated | Downloaded 04/18/21 01:39 PM UTC

236 JOURNAL OF APPLIED METEOROLOGY VOLUME 43

ted the estimation of echo motion with sufficient spatial

detail, consistency between subsequent time periods,
and sufficient computational stability. The time periods
for the finer-grid calculations (including observations
for 1 h before chosen times) are listed in Table 2.
Nowcasts were produced with the semi-Lagrangian
backward scheme of Part I. As in that study, a forecast
lifetime can be defined using the correlation c(T ):
Io If
c(T ) 5 , (23)
ÏI o2 I f2
such that

Lifetime 5 E0
`

c(T ) dT. (24)

For an exponential correlation function, the lifetime is

the time for which the correlation drops to (1/e). Recall
that the mean value of our observations I o and the mean
value of the forecast I f are usually nonzero. For ex-
ample, the average rainfall rate over an area with any
rain will be nonzero and positive. In our definition of
the correlation and lifetime, we have not subtracted
mean values from the observations and forecast, because
zero rainfall is the only meaningful reference value.
When using radar reflectivities, a reference value for
zero rainfall must be chosen that is below the minimum
detectable rain threshold. In this paper, the value of 0
dBZ is assigned to nonraining areas.
This definition of lifetime was applied to forecast sets FIG. 2. Forecast lifetime as a function of low-pass cutoff scales
for (top) Fourier and (bottom) wavelet filters. Results are shown for
with up to 8 h lead time for the four precipitation cases. each of the four precipitation cases studied.
The correlation function was assumed to be approxi-
mately exponential. The increased predictability of larg-
er-scale features is apparent in Fig. 2, where lifetimes thresholding. With Fourier low-pass filters, such thresh-
increase dramatically with increasing cutoff scale for olding is necessary to eliminate widespread ripples due
either Fourier or wavelet low-pass filters. In each case, to the Gibbs effect, but this reduces forecast lifetimes
the same low-pass filter is applied to the forecasts and in a manner that depends on the threshold used.
the observations used for comparison. Removing small- As found in Part I, the cases of 16 May 2000 and 30
er, more perishable scales increases lifetime values. This July 1998 are most predictable. Forecast lifetimes are
differs in certain ways from a similar analysis in Part shortest for 1 June 1999, and the case of 25 May 2001
I. There, Fourier low-pass-filtered observations were shows an intermediate predictability. The curves in Fig.
used as input to the nowcasting method. Here, the fil- 2 are roughly linear, with similar slopes between the
tering is applied to the output of the nowcasting routine. four cases. Very few parameters may ultimately be re-
More important, Part I sought to avoid effects of the quired to describe the scale dependence of predictability.
ripples due to the Gibbs effect (as shown in Fig. A1b)
by applying a threshold of 10 dBZ to both the low-pass-
c. Spectra and cospectra
filtered forecasts and the observations used for com-
parison. With increasing smoothing, greater areas of re- The scale weights w(m, T ) for an optimal forecast
flectivity are reduced below this threshold, dramatically filter (OFF) will depend on the wavelet spectrum and
reducing correlation and lifetime values. cospectrum, as described in Eq. (19). Here we present
As explained in appendix A, Haar wavelet low-pass averaged Haar wavelet spectra and cospectra from the
filter results can be compared with Fourier low-pass four precipitation events.
filter results if a Ï2 correction factor is used to deter- Figure 3 shows the average Haar wavelet spectrum
mine an equivalent Fourier cutoff scale. This equivalent of logarithmic reflectivity (dBZ) for observations of the
cutoff scale is used in showing wavelet low-pass life- 25 May 2001 case. Spectral values are normalized by
times in Fig. 2. The results for Fourier and wavelet low- the mean square dBZ level, with larger scales to the left
pass filtering are very similar. However, with wavelet and smaller scales to the right. The average (solid line)
low-pass filtering there is no need to apply arbitrary curve in Fig. 3 was produced by averaging spectra from

Unauthenticated | Downloaded 04/18/21 01:39 PM UTC

FEBRUARY 2004 TURNER ET AL. 237

FIG. 3. Wavelet spectrum of observations of logarithmic reflectivity FIG. 4. Average wavelet spectrum of observations of logarithmic
(dBZ) for the case of 25 May 2001, for each hour from 0500 to 1600 reflectivity (dBZ ) for each of the four precipitation events of the
UTC. The solid curve indicates the average spectrum for the period. study. Spectral values are normalized by the mean square of loga-
Spectral values are normalized by the mean square of logarithmic rithmic reflectivity for each period.
reflectivity.

ever units are used. Here, logarithmic reflectivity units

all of the 1-h intervals during the period. The values (dBZ) are used.
for these individual spectra are also plotted to show the Returning to the specific case of 25 May 2001, Fig.
magnitude of variations with time. The spectral peak at 5 shows wavelet cospectra between predicted and ob-
scales of about 250 km indicates the approximate size served reflectivity measurements for different forecast
of the dominant precipitation patterns. This corresponds lead times. Cospectra in this figure are normalized by
to the width of the strong rainbands that are part of a the product of rms of the forecast and observations.
larger comma-shaped storm organization that contrib- Treating the observations as the ‘‘prediction’’ at zero
utes to spectral values at larger scales. Within individual lead time, we show the spectrum for the initial T 5 0
rainbands, a variety of more local storms contribute to min. Cospectral values at each scale decrease with in-
spectral values at smaller scales. creasing forecast lead time. The cospectral values de-
Figure 4 compares the average spectra for each of the crease most rapidly at smaller scales. For instance, less
four cases studied. In each case, the spectrum decreases than 1-h storms, with length scales less than 20 km,
with decreasing scale for scales less than about 150 km. have advanced through their life cycle. The Lagrangian
The slopes of the spectra are similar in this range, sug- nowcast of such storms for 1 h or more has negligible
gesting that some aspects of precipitation organization skill. Those features in the nowcast may contribute sig-
at these scales are independent of the larger-scale forc- nificantly to the spectrum, but will not contribute much
ing. The greatest differences between the four cases are to the cospectrum. As described in Table 1, larger storm
seen for scales greater than 250 km. As previously noted patterns evolve more slowly, and are predictable for
with the case of 25 May 2001, precipitation is com- longer times. Cospectral values for scales greater than
pressed in localized rainbands, giving a spectral peak
near 250 km. For the case of 30 July 1998, the spectral
peak is at scales closer to 500 km. In contrast, the spec-
trum for the 16 May 2000 case with more widespread
and uniform precipitation continues to increase with
scale length.
Of course, the shape of these spectra would be dif-
ferent if calculated using rainfall rates or linear reflec-
tivity instead of dBZ values. The logarithmic transfor-
mation from rainfall rate or linear reflectivity to loga-
rithmic reflectivity (dBZ) reduces the relative influence
of strong, localized peaks in reflectivity on the spectrum.
The strongest rainfall rates are often found in isolated,
small-scale convective cells. This will affect both the
spectra and cospectra, reducing the values at smaller
scales. This, in turn, will affect the optimal scale weights FIG. 5. Wavelet cospectra between predictions and observations of
w(m, T ) for our forecast filter. The optimal forecast filter logarithmic reflectivity (dBZ ) for the case of 25 May 2001, using
is designed to minimize the mean square error in what- predictions made each hour from 0500 to 1600 UTC.

Unauthenticated | Downloaded 04/18/21 01:39 PM UTC

238 JOURNAL OF APPLIED METEOROLOGY VOLUME 43

likely filters for a geographical region during different

seasons. Here we will derive a ‘‘near optimal’’ forecast
filter (NOFF) from scale weights for a single lead time,
using a forecast shortly before the time of interest.
We start by taking a 1-h forecast produced 1 h before
the chosen ‘‘current’’ time. From this forecast and cur-
rent observations, a set of recent scale weights
w(m, T ) can be determined for the OFF at 1-h lead
time. These weights, calculated for the previous hour,
are taken as near-optimal estimates for the current 1-h
forecast. As an example, suppose the current time is
0800 UTC. We first generate a 1-h lead time nowcast
using the observations and motion field that were avail-
able at 0700 UTC. Comparing the forecast produced
at 0700 for 0800 UTC with the current observations
(at 0800 UTC) allows us to compute a posteriori
smoothing scale weights for a 1-h lead time forecast.
We take these weights as still valid for the current
forecast (produced at 0800 UTC) for 0900 UTC.
Initial analysis of the decrease in scale weights w(m,
T ), such as those shown in Fig. 6, suggested a linear
transformation of weights between scale and time. This
was taken as a means of obtaining reasonable scale
weights for a NOFF based on available observations
and a recent forecast. For instance, the predictability of
features with a spatial scale of 100 km at 2 h may be
similar to the predictability of 50-km features at the 1-
h lead time. That is, we assume that
w(m, 2 h) 5 w(m 2 1, 1 h). (25)
FIG. 6. Scale weights w(m, T ) for the cases of (a) 30 Jul 1998 and
(b) 25 May 2001 for several forecast lead times (30 min–4 h). A More generally, we produce a full set of NOFF scale
posteriori, OFF scale weights are shown by solid lines. A priori, weights for different forecast lead times, as follows:
NOFF scale weights are shown as dashed lines, for comparison.
w(m, T ) 5 w(m 2 log 2T, 1 h), (26)
where the forecast lead time T is measured in hours.
200 km continue at over one-half of their T 5 0 min Because m is related to the length scale, as described
value for as long as 4 h, indicating that larger precip- in Eq. (5), exponentiating the argument (m 2 log 2T )
itation patterns were predicted with reasonable accuracy from the right-hand side of Eq. (26) gives
over such times. LengthScale
e m2log 2T 5 , (27)
T
d. Optimal and near-optimal forecast filters
with units of distance divided by time. This is the rate
Using spectra and cospectra, such as those shown in at which scale weights shift to larger length scales with
Figs. 4 and 5, OFF can be derived a posteriori using increasing forecast lead time. It is the parameter used
Eq. (19). The scale weights w(m, T ) for the 30 July in estimating a linear transformation of weights between
1998 and 25 May 2001 cases are shown as solid curves scale and time.
in Fig. 6. After verifying forecasts against later obser- As an initial test of Eq. (26), it was applied to the
vations, these would have been the optimal scale OFF scale weights (solid curves) for the 1-h forecast
weights for forecast filtering for different forecast lead lead time in Figs. 6a and 6b for the cases of 30 July
times. There are differences in optimal scale weights 1998 and 25 May 2001, respectively. The resulting
between cases. NOFF scale weights for forecast lead times of 30 min,
Some a priori approximation of optimal scale weights 2 h, and 4 h are shown as dashed lines. Equation (26)
w(m, T ) is required for an operational application to produced excellent estimates of OFF scale weights. The
forecasts. It is not enough to know the best forecast results for 25 May 2001 are more typical of the other
filter after the forecast has been verified! An appropriate cases, with overestimation of scale weights at longer
filter, as close to the best as possible, must be determined forecast lead times and larger scales. Clearly, Eq. (26)
and applied at the time that the forecast is made. With does not always provide accurate scale weights. In Fig.
sufficient analysis, it might be possible to determine 6b, OFF scale weights decrease more rapidly with fore-

Unauthenticated | Downloaded 04/18/21 01:39 PM UTC

FEBRUARY 2004 TURNER ET AL. 239

cast lead time than expected by the NOFF scale-weight reduced as the precipitation is smoothed over larger re-
estimates. The exact form of the scale transformation gions.
of weights varies from one case to another. However, As the lead time T → `, these forecast filters will
the NOFF approximation produces an a priori estimate average the predicted precipitation to a mean value for
of the best scale weights with a minimum of case-de- the entire area. This gives something like a climatolog-
pendent information. ical estimate, which is not really useful as a nowcast.
For forecast lead times that are not integral powers For each precipitation case there will be a forecast lead
of 1 h, scale weights must be interpolated between the time beyond which the filtered nowcast is of no practical
closest available lead times. A linear interpolation of value. This visible loss of forecast clarity can be initially
scale weights was used in this study. A logarithmic in- disturbing. How can we expect to improve forecast error
terpolation of scale weights from nearby lead times statistics by removing some of the detail? Figures 8a–
would be an alternative. 8d show the rms errors of original and (both OFF- and
We now have two smoothing methods designed to NOFF-) filtered forecasts as a function of forecast lead
reduce forecast errors. The a posteriori OFF gives the time. The rms values, in units of dBZ, were calculated
best possible reduction of errors. NOFF produces only using only locations with either a forecast or an ob-
a rough estimate of what will be the best scale weights. served intensity greater than 2 dBZ. The filtered fore-
However, because it is based on earlier forecast results casts have lower rms errors than the unsmoothed fore-
and requires a minimum of case-dependent information, casts, even for a 15-min lead time. Filtering has selec-
it can be applied to forecasts as they are produced. tively removed the generally inaccurate, small-scale de-
tails. The improvement becomes greater for longer lead
times. However, rms error increases with lead time for
e. Forecasts and error measures both the original and filtered forecasts. OFF usually has
lower rms errors than NOFF. The two only begin to
OFF was designed to minimize the contribution of differ by more than 1 dBZ for forecast lead times greater
each filtered forecast scale to the mean squared error. than a few hours. This is important because Fig. 6b
NOFF allows an estimate of the optimal scale weights showed some differences between average NOFF and
operationally as the forecasts are made. The result OFF scale weights for the 25 May 2001 case. Apparently
should be filtered forecasts that significantly improve the filtered forecast’s rms error is not very sensitive to
forecast accuracy, according to measures such as the these minor differences in scale weights. The near-op-
rms error and the correlation of the forecast with ob- timal forecast filtering, which can be applied operation-
servations. In this section we test how well these ob- ally at the time that forecasts are produced, is a suitable
jectives are met for the four precipitation events listed filtering method. At lead times greater than about 6 h,
in Table 2. the rms errors for OFF and NOFF may actually decrease,
The effect of filtering on the appearance of forecast but by this point the smoothing effects of forecast fil-
maps in dBZ units is shown in Fig. 7. Figure 7a shows tering have had dramatic effects on the forecast’s ap-
observations at 0500 UTC on 25 May 2001. NOFF pro- pearance. Forecast correlation with observations is also
duced from these observations are shown for the 30- strongly improved by filtering. The correlation results
min forecast (Fig. 7b), the 2-h forecast (Fig. 7c) and for the four precipitation events are shown in Figs. 9a–
the 8-h forecast (Fig. 7d). As previously described, de- 9d. Higher correlation of the original forecasts with ob-
tails for each scale are progressively removed as they servations indicates the relative predictability of the dif-
become less predictable. In Fig. 7b, the effects of fil- ferent cases. The event of 1 June 1999 was least pre-
tering are not noticeable. With a 30-min forecast, scales dictable, while the event of 16 May 2000 was most
greater than tens of kilometers are still largely predict- predictable. Forecast filtering by both OFF and NOFF
able. Average scale weights for the 25 May 2001 case improves correlation results in all four cases. Again, the
are shown in Fig. 6b. NOFF smoothing of scales of tens NOFF method provides almost as much improvement
of kilometers is not really visible on a map of this scale as OFF.
that covers a distance of thousands of kilometers. The These cases suggest that an operational application
effects of forecast filtering begin to be visible in the 2- of forecast filtering with NOFF will reliably improve
h NOFF results, shown in Fig. 7c. For this forecast lead correlation and reduce rms errors. Differences from OFF
time, even scales near 100 km suffer a significant loss results are not significant, and so it is reasonable to focus
of predictability and are affected by NOFF smoothing. instead on the effects of implementing NOFF.
These scales are large enough to be seen on this map.
For the 8-h forecast, shown in Fig. 7d, the dramatic
effects of filtering at scales up to hundreds of kilometers f. Smoothing and skill scores
are seen. There are two related effects. First, the forecast Because OFF and NOFF were designed to improve
area of (nonzero) precipitation is greater, because the forecast correlation and rms errors, their skill in pre-
exact future location of the rainfall is uncertain. Second, dicting areas of rainfall exceeding a threshold must be
the average forecast rainfall rate over raining areas is verified. In Fig. 7, we have already seen that NOFF

Unauthenticated | Downloaded 04/18/21 01:39 PM UTC

240 JOURNAL OF APPLIED METEOROLOGY VOLUME 43

FIG. 7. Effects of NOFF for the case of 25 May 2001, using predictions made at 0500 UTC. The figures
show (a) original observations at 0500 UTC from which the forecasts were produced, (b) the NOFF forecast
for 0530 UTC (30-min lead time), (c) the NOFF forecast for 0700 UTC (2-h lead time), and (d) the NOFF
forecast for 1300 UTC (8-h lead time).

forecasts contain less information on intensity distri- NOFF forecast values for better prediction of forecast
butions as the forecast lead time increases. The total areas, with some smoothing but not too much.
area for which precipitation is predicted increases with For filtered forecasts, there is an apparent contradic-
increasing forecast lead time. Predicted rainfall spreads tion between improvements in error measures and the
over larger areas, compensating for the increasing un- shift toward less intense, more widely spread areas of
certainty in the future position of precipitation patterns. forecast precipitation. In fact, the unsmoothed and fil-
In this respect, some smoothing is good. Coupled with tered forecasts are different types of estimates. The un-
this, the distribution of reflectivity or rainfall rates be- smoothed forecast represents one of many possible sets
comes more uniform, tending toward lower values. This of future observations. Forecasts filtered by OFF or
aspect of smoothing is not good, because it misrepre- NOFF give an average of the various most likely future
sents the type of rainfall that is expected. Understanding observations. Such a map of rainfall, averaging over
these effects of smoothing allows us to transform the possible outcomes, does not usually resemble any likely

Unauthenticated | Downloaded 04/18/21 01:39 PM UTC

FEBRUARY 2004 TURNER ET AL. 241

FIG. 9. Effects of OFF and NOFF forecast filters on c(T ) correlation

FIG. 8. Effects of OFF and NOFF forecast filters on rms errors for for the four precipitation events. For the four cases, the average
the four precipitation events. The rms errors (dBZ ) were calculated forecast correlation with 4-h lead time is 0.50 with no smoothing but
for locations where intensities greater than 2 dBZ were forecast and/ increases to 0.62 for the NOFF forecasts.
or observed.

fall rate. Results of the CSI will depend strongly on the

individual future rainfall field. Although OFF and NOFF threshold chosen.
were designed to reduce rms errors, this does not guar- The uncertainty in predicting areas with rainfall rate
antee that all other forecast error measures will be im- above a threshold increases with forecast lead time, but
proved. The spread of the precipitation area due to fil- the net effect of forecast filtering on this uncertainty
tering may be beneficial to the extent that it captures varies. For example, by spreading the forecast precip-
the uncertainty in the future position of precipitation. itation over a wider area, filtering may increase the num-
Alternatively, this spread could be optimized with re- ber of hits and decrease the number of misses for lower
spect to ‘‘hits’’ and ‘‘misses’’ of precipitation. thresholds, although it may also increase the number of
The critical success index (CSI) skill statistic is used false alarms. As well, the total number of locations fore-
here to study the ability of a filtered forecast to predict cast above a higher threshold decreases sharply with
areas with intensities above a threshold. Any image grid filtering as isolated strong features are smoothed to low-
point that is correctly forecast to be above the threshold er values over a wider area. In this situation, the number
counts as a hit. If a location is incorrectly forecast to of hits and false alarms decreases, and the number of
be below the threshold, this is a miss. A location that misses increases.
is incorrectly forecast to be above the threshold is a In Fig. 10, effects of OFF and NOFF on the CSI are
‘‘false alarm.’’ After determining the total number of shown for a 10-dBZ threshold. Naturally, the CSI de-
hits (H), misses (M), and false alarms (F) over all fore- creases with increasing lead time whether the original
cast grid points, the CSI is calculated as forecast (no smoothing), OFF, or NOFF are used. In
addition, for the first few hours of forecast lead time,
H either OFF or NOFF actually produce a small improve-
CSI(R t ) 5 , (28) ment in CSI for each of the four cases. However, for
H1M1F
each case there is a forecast lead time beyond which
where R t is the chosen threshold of reflectivity or rain- OFF and NOFF forecasts provide a poorer prediction

Unauthenticated | Downloaded 04/18/21 01:39 PM UTC

242 JOURNAL OF APPLIED METEOROLOGY VOLUME 43

FIG. 10. Effects of OFF and NOFF on the CSI at the 10-dBZ FIG. 11. Effects of OFF and NOFF on the CSI at 15-dBZ threshold
threshold for the four precipitation events. for the four precipitation events.

of the area exceeding 10 dBZ than the original forecast creases and the exact location of more intense storms
with no smoothing. For example, this crossover time is is more uncertain, increased smoothing avoids large er-
around 8 h for the case of 30 July 1998, while for the rors by predicting lower intensities over a wider area.
case of 25 May 2001 it is about 3 h. In Fig. 11, effects The same filtering that reduces rms errors also decreases
of OFF and NOFF on the CSI are shown for a 15-dBZ the CSI for higher reflectivities. As seen in Figs. 10 and
threshold. Again, both OFF and NOFF provide a slight 11, this eventually affects even the CSI at the 10- and
increase in CSI for the smaller lead times. For longer 15-dBZ thresholds.
lead times the CSI is worse than the original forecast Even if the future location of local, intense storms is
with no smoothing. The crossover time for each case uncertain, in most cases a forecaster will still need to
occurs earlier with the 15-dBZ threshold than with the know that such storms are likely to occur. These details
10-dBZ threshold. are gradually lost with increasing lead time in NOFF
How can OFF and NOFF dramatically reduce forecast forecasts, such as those shown in Fig. 7. Any modifi-
rms errors and increase correlation, as shown in Figs. cation of the filtered forecast is likely to increase the
8 and 9, respectively, while having so little effect on rms error, but this may be justifiable if the forecast pre-
the CSI at the 10- and 15-dBZ thresholds? This apparent cipitation areas and distributions of forecast values be-
contradiction forces us to reexamine the differences be- come more realistic. The NOFF forecast ‘‘spread’’ with
tween error measures and what is meant by forecast increasing smoothing can be studied by matching the
quality. The CSI is based on a small number of cate- area exceeding a certain reflectivity threshold. In Fig.
gories for a forecast at each location. All hits are con- 12, this type of matching is done for a set of forecasts
sidered equally important, and a near miss has the same from each of the four precipitation events. In Fig. 12a,
significance as any other miss. By contrast, a near miss the area exceeding 15 dBZ is first calculated for each
will contribute less to the rms error than an average observation image. For the corresponding NOFF
miss. The rms error measure, therefore, rewards the (smoothed) forecast, a second reflectivity threshold is
avoidance of large errors. NOFF and OFF are, therefore, found with the same total area exceeding that value. As
designed to avoid large errors, so that as lead time in- a result of filtering and spreading, the corresponding

Unauthenticated | Downloaded 04/18/21 01:39 PM UTC

FEBRUARY 2004 TURNER ET AL. 243

TABLE 3. Effects of several methods for rescaled NOFF forecasts

on rms error and the CSI at the 15-dBZ threshold.

Original S-PROG CSI-max Area

Forecast measure nowcast NOFF rescaled rescaled matching
2-h rms (dBZ ) 16.6 11.3 15.7 12.4 15.7
2-h CSI at 15 dBZ 0.325 0.320 0.371 0.356 0.372
4-h rms (dBZ ) 18.2 11.4 16.8 13.4 17.7
4-h CSI at 15 dBZ 0.229 0.174 0.258 0.232 0.257
8-h rms (dBZ ) 19.9 11.5 18.4 12.3 19.7
8-h CSI at 15 dBZ 0.109 0.040 0.125 0.060 0.123

that are suitable for operational implementation. Here

we suggest three possible forecast-rescaling methods,
and apply them on sample 2-, 4-, and 8-h lead time
nowcasts from 0500 UTC on 26 May 2001. The relative
effects of each method on rms error and the CSI for a
15-dBZ threshold are shown in Table 3. These statistics
are first given for the original nowcast and for the NOFF
version.
In the Spectral Prognosis (S-PROG) model of Seed
(2003), a rescaling method is applied to Fourier-filtered
nowcasts. The S-PROG rescaling is applied here at each
grid point of a NOFF forecast:
dBZ(S-PROG rescaled)

5 5 dBZ(NOFF) 1 (15 2 Z ) if dBZ(NOFF) . Z

0 otherwise,
(29)
FIG. 12. NOFF forecast thresholds with an area of greater reflec-
tivity equal to the corresponding area exceeding the 15-dBZ threshold where Z is the NOFF forecast threshold with the same
in (a) observations and (b) the original, unsmoothed nowcast inten- area exceeding it as the area exceeding 15 dBZ in the
sity. Results are shown for forecasts of 0500 UTC 30 Jul 1998, 1700
UTC 1 Jun 1999, 0100 UTC 16 May 2000, and 0500 UTC 25 May original, unfiltered nowcast. This rescaling counters
2001. some smoothing effects of filtering, removing widely
spread areas of low intensity, while uniformly increasing
higher intensities by the same amount to be 15 dBZ and
NOFF thresholds were less than 15 dBZ. In each case, greater. By reversing the spread of forecast areas of
as forecast lead time and NOFF smoothing increased, precipitation, the CSI for a threshold of 15 dBZ is higher
these matching NOFF thresholds decreased. The exact than for NOFF. For the forecasts used in Table 3, the
decrease with forecast lead time was different for each CSI values using S-PROG rescaling are actually greater
case. In Fig. 12b, the area exceeding 15 dBZ in each of than the values for the original nowcasts. This improve-
the original, unsmoothed nowcasts is first calculated. ment in CSI comes at a cost of increasing the rms errors,
The size of this area matched the area in the NOFF although not to the level of the original nowcasts. As
forecast exceeding some other, lower threshold. De- in section 3e, rms values (dBZ) are calculated using
creases in matching NOFF threshold values were again only locations with either a forecast or observed inten-
found with increasing forecast lead time. However, the sity greater than 2 dBZ.
decreases in matching thresholds in Fig. 12b were more Another rescaling method is based on first finding the
organized than those found for the same cases in Fig. single threshold in the original nowcasts and NOFF
12a. The matching thresholds for the four cases were forecasts that has the maximum CSI. For the example,
more similar, and showed roughly linear decreases with nowcasts used to produce Table 3, the CSI-maximizing
the logarithm of forecast lead time. (CSI max) thresholds were very low: 3.70 dBZ for 2-
Results such as those in Fig. 12a can only be deter- h lead time, 3.50 dBZ for 4 h, and 2.30 dBZ for 8 h.
mined after forecasts have been verified against obser- As for the S-PROG rescaling, forecast grid points with
vations. The results in Fig. 12b, however, can be de- NOFF values less than the threshold are set to zero.
termined when forecasts are produced. These area- Intensities above the CSI-max threshold are then in-
matching thresholds, reflecting the spatial characteris- creased by a constant value to preserve the average
tics of the forecast rainfall patterns, provide one basis intensity over the entire forecast area. Table 3 shows
for methods of ‘‘rescaling’’ NOFF forecast intensities the effects of CSI-max rescaling to be similar to S-

Unauthenticated | Downloaded 04/18/21 01:39 PM UTC

244 JOURNAL OF APPLIED METEOROLOGY VOLUME 43

PROG rescaling. As compared with NOFF, the CSI at for rainfall rates at each location described in Part II.
the 15-dBZ threshold is improved at the cost of greater Such products convey more information than filtered
rms errors. forecasts for longer lead times.
As a final, more extreme method, the example NOFF The NOFF method should now be applied to a long,
forecasts were rescaled to have the same area above continuous set of nowcasts and verified against obser-
each threshold as for the original nowcast. Even the vations. The same tests could also include further cal-
most isolated, intense, and unpredictable values of the ibration of NOFF rescaling factors for better rain-area
original nowcast appear after this area matching. This estimates. For hydrological applications, requiring pre-
contradicts the original design goal for forecast filter- cipitation estimates over larger areas, forecast filtering
ing—to improve forecasts by removing less predictable will be helpful. In other situations, it can be used along
features. As for S-PROG rescaling, the CSI at 15 dBZ with the original nowcast to illustrate forecast uncer-
is improved; but, with area matching, the rms errors are tainty.
almost the same as for the original nowcasts.
NOFF was designed to optimize forecast rms error Acknowledgments. This work was funded by a grant
statistics, which it does by unrealistically spreading the from the Canadian Foundation for Climate and Atmo-
forecast area of precipitation. Appropriately rescaled spheric Sciences.
NOFF forecasts can be used to improve the forecast of
rain area above a given threshold; however, rescaling APPENDIX A
sacrifices some or all of the reduction in rms errors
achieved by the NOFF method.
The Gibbs Effect

4. Discussion and conclusions Rainfall fields routinely exhibit highly localized fea-
tures and sharp spatial gradients. The Fourier represen-
In Parts I and II the scale dependence of precipitation tation, with its extended basis functions, has problems
nowcasts was investigated. In this paper, scale-depen- with such behavior. The Gibbs phenomenon occurs for
dent measures of predictability have been applied to the a Fourier series approximation of a piecewise smooth
forecast method itself to design forecast filters that re-
function with discontinuities. At each discontinuity, the
duce rms errors. Optimal forecast filters (OFF) and op-
Fourier approximation will either overshoot or under-
erationally useful near-optimal forecast filters (NOFF)
shoot the actual magnitude of the discontinuity by about
have been developed using measures of predictability
9% (Walker 1988, 60–63). This is accompanied by the
as a function of scale based on the wavelet transform.
Gibbs effect—a ‘‘ringing’’ or ‘‘ripple’’ extending away
These methods were applied to continental-scale La-
grangian nowcasts based on composite radar observa- from the discontinuity. The frequency of these ripples
tions of precipitation. The forecast filters, which adapt is equal to the first term omitted from the Fourier ap-
to changing conditions, provide significant improve- proximation (Weaver 1989, 125–127). Fourier low-pass
ments in forecast precision in terms of correlation and filters can be modified by various windowing functions
rms errors. For the four precipitation events studied, the to reduce (but not eliminate) the Gibbs effect (Ersoy
average correlation at the 4-h lead time was increased 1997, 281–287).
from 0.50 for the original nowcasts to 0.62 with forecast Although the Fourier and wavelet representations dif-
filtering. fer, one can compare Fourier and wavelet low-pass fil-
Filtered forecasts for longer lead times reflect the lack ters. The wavelet spectrum is equivalent to a smoothed
of predictability of small scales associated with higher Fourier spectrum (Hudgins et al. 1993), where the
intensities. Thus, they are understandably smoothed to smoothing depends on the wavelet used. Similarly, the
low values over large areas. The texture of precipitation Fourier low-pass cutoff frequency for best comparison
patterns is lost as small regions of higher rainfall rates with a wavelet low-pass filter depends on the wavelet
are averaged with adjacent areas. In this way, as forecast basis function used. Although not well localized in fre-
information about smaller storm structures becomes in- quency, the power spectrum of the fundamental Haar
accurate, NOFF removes these details in order to reduce wavelet (of length 1) has its main peak centered near a
rms errors. Where rms error is not the only relevant frequency of 0.74. This corresponds to a Fourier mode
forecast error measure, NOFF forecasts can be rescaled with a wavelength of about 1.35 times the length of the
to better predict rain areas above a threshold at the cost wavelet. For instance, a Haar wavelet of length 128 km
of some increase of rms errors. has a Fourier spectrum with strong components between
With increasing forecast lead time, forecasts become spatial scales of 128 and 256 km. Here, we will apply
less reliable in predicting rainfall rate at a point (or over (to the nearest pixel length) a Ï2 factor between Haar
smaller areas). Although the filtered forecasts presented wavelet low-pass and Fourier low-pass cutoff frequen-
here are based on this principle, each nowcast still dis- cies to render the results more comparable. A wavelet
plays a single value at each location. A more complete low-pass filter with its scale of 128 km will be compared
expression of this uncertainty is the probability forecast with a Fourier low-pass filter with a cutoff wavelength

Unauthenticated | Downloaded 04/18/21 01:39 PM UTC

FEBRUARY 2004 TURNER ET AL. 245

FIG. A1. Gibbs effect in Fourier low-pass filtering of radar observations of precipitation.

near 181 km (actually, 45 pixels of a 4-km length, or 3 180 km. The Gibbs effect appears as ripples that
180 km). extend well away from the discontinuities. Note that
Figure A1a shows a radar reflectivity field (in units white areas of these ripples take on values below the
of dBZ) for 0800 UTC on 25 May 2001. In Fig. A1b, original reflectivity threshold. If the same type of Fou-
a Fourier series approximation is shown, which omits rier low-pass filter is applied to the image, expressed as
Fourier components for spatial scales less than 180 km rainfall rates, nonphysical areas of ‘‘negative rainfall’’

Unauthenticated | Downloaded 04/18/21 01:39 PM UTC

246 JOURNAL OF APPLIED METEOROLOGY VOLUME 43

are produced. Although Fourier low-pass filtering of let coefficients is low. The associated inverse transform
rainfall rate R is, therefore, unacceptable, applying such (reconstruction) formula is given by

O E
a filter to log R will at least avoid negative-filtered val- ` 1`
ues for R. A truncated Haar wavelet approximation of f (x) 5 22m W(m, x9)c m (x 2 x9) dx9. (B3)
the same rainfall field, omitting wavelet components for m52` 2`
scales less than about 128 km 3 128 km, is shown in
Fig. A1c. This type of low-pass filter also smooths the The integral gives the contribution of each scale number
sharp spatial gradients seen in Fig. A1a but localizes m to the function. Each term in the summation acts as
the distortions closer to the discontinuities. The relative a wavelet bandpass filter of f (x). In practice the sum-
disadvantage of a Fourier low-pass filter relative to a mation is made for 0 , m # mmax , because we cannot
wavelet low-pass filter is determined by the magnitude use scales larger than the length of measurements. The
of sharp gradients and discontinuities in the field to be contribution of larger scales remains in a smoothed ver-
studied. sion of f (x), which is like a running mean.
The precipitation pattern in this case is not locally A wavelet power spectrum can be defined as
isotropic. In the northeast of the study area, it is S(m) 5 22m [W(m, x)] 2 , (B4)
stretched east–west. In the southeast, it is aligned from
southwest to northeast. In this case, any attempt to apply where the overbar denotes a spatial average and S(m)
appropriate anisotropic wavelets would need to adapt measures the importance of features with approximate
to the local precipitation pattern. length scales between 2 m21 and 2 m . Similarly for two
functions, say f (x) and g(x), one can define a wavelet
APPENDIX B cospectrum as
Co fg (m) 5 22m Wf (m, x)Wg (m, x), (B5)
Wavelets and Filtering where the subscripts f and g denote the the two func-
A family of wavelet functions (a basis) is produced tions f (x) and g(x), respectively. Hudgins et al. (1993)
from a fundamental wavelet c 0 (x) by dilations and showed that such a wavelet spectrum (or cospectrum)
translations: is equivalent to a Fourier spectrum (or cospectrum)
smoothed by the Fourier spectrum of the wavelet func-
tion at each scale. The wavelet spectrum (or cospectrum)
1 2
x 2 x9
c m (x 2 x9) 5 22m/ 2c 0 . (B1) gives the contribution to the variance (or covariance)
2m per logarithmic (roughly power of two) interval of scale.
Admissibility conditions for choosing a fundamental Each fundamental wavelet c 0 (x) has an associated
wavelet are discussed by Mallat (1989). In this paper orthogonal smoothing function f 0 (x). As for the wave-
discrete power-of-two dilations are used, with a scale lets, a set of smoothing functions are produced by di-
index m that increases with increasing scale. For in- lations and translations of f 0 (x). Greater mathematical
stance, a wavelet with scale index m 5 1 is 2 times as detail is given by Burrus et al. (1998) and other similar
wide as the fundamental wavelet, with a smaller am- texts. Convolution with the smoothing functions acts as
plitude. A wavelet with scale index m 5 2 is 4 times a low-pass filter:
as wide and one-half of the amplitude of the funda-
mental wavelet, and so on to the larger scales. By taking
such discrete scales, the efficient wavelet transform al-
f m (x) 5 E 1`

2`
f (x)f m (x 2 x9) dx9

gorithm of Mallat (1989) can be used.

Wavelet transform coefficients are given by 5 O `
22m9 E1`

W(m9, x9)c m9 (x 2 x9) dx9.

E
m95m11 2`
1`
(B6)
W(m, x) 5 f (x)c (x9 2 x) dx9.
m
(B2)
2` The smoothed function f m (x) contains only those fea-
tures represented by wavelet coefficients W(m9, . . .),
Each wavelet transform coefficient measures the simi- where m9 . m. The simplest admissible wavelet is the
larity between a particular wavelet and the function f (x) Haar function:
being studied. The wavelet is localized, and so the wave-
let coefficient will depend on what f (x) is like for that  1
local range of position. The wavelet is also of a certain
scale, so that the wavelet coefficient will describe the c 0(x) 5


11 2 ,x#0
2
1
0
21 0,x,1 (B7)
variability of f (x) at similar scales. Where there are 2
intense features of approximate length between 2 (m21)
 otherwise.
and 2 m near x, the absolute value of W(m, x) will be
high. Where there is little variability at a particular scale This can be used to represent a function as a sum of
and position, the absolute value of the associated wave- square wave pulses of different sizes at different posi-

Unauthenticated | Downloaded 04/18/21 01:39 PM UTC

FEBRUARY 2004 TURNER ET AL. 247

lets and one smoothing function are illustrated. Con-

volution of an image with the smoothing function of the
two-dimensional Haar basis gives uniformly weighted
block-average smoothing of an image for each scale.
Coefficients from the two-dimensional wavelet trans-
form are used to calculate the two-dimensional wavelet
spectrum and cospectrum:

O [W (m, x, y)] , and

3

S(m) 5 42m i 2
(B10)
i51

O W (m, x, y)W (m, x, y),

3

Co fg (m) 5 42m i
f
i
g (B11)
i51

where the overbar denotes a spatial (areal) average. The

normalization factor becomes 4 2m with the transition to
the two-dimensional transform.
Section 2b derives optimal weighting factors w(m, T )
for each scale m and a forecast lead time T. This filtering
can be implemented indirectly, avoiding explicit use of
the inverse wavelet transform. Recall that as a result of
the orthogonality of a wavelet and its smoothing function,
the convolution of the image with this smoothing function
acts as a low-pass filter:
I mf (x, y)

5 E 2`
1`

I f (x)f m (x 2 x9, y 2 y9) dx9

FIG. B1. Pattern of one- and two-dimensional Haar wavelets and

smoothing functions. The one-dimensional Haar wavelet is a single
5 O
m95m11
`

22m9 E1`

2`
W(m9, x9)c m9 (x 2 x9, y 2 y9) dx9,

square-wave pattern, and its smoothing function is a step function. (B12)

The two-dimensional Haar wavelets and smoothing function are the
possible products of the one-dimensional functions in the x and y where I (x, y) is a smoothed version of the original
m
f
directions. forecast I f (x, y) with details of scales m and smaller
removed. A more intuitive illustration of this, with sim-
pler notation, is shown in Fig. B2.
tions. The Haar wavelet and its associated smoothing The reconstruction formula can, therefore, be rewrit-
function, ten in terms of the uniformly weighted, block-averaged
versions of the forecast image:
 1 1
 11 2 ,x#1 I f s (x, y, T )
f (x) 5
0  2 2 (B8)

O [w(m 1 1, T ) 2 w(m, T )]I (x, y, T ).
`
0 otherwise,
5 m
f (B13)
also provide a computationally efficient transform. m50

Generalizing to two dimensions produces a set of Normally the scale weights at each scale will become
three wavelets and one smoothing function for each larger with increasing scale index m at any time T—
scale and position: small scales will be less predictable and decorrelate fast-
C1m (x, y) 5 c m (x)f m (y), er than larger-scale features. In practice, there will be
some maximum-scale index mmax beyond which wavelet
C2m (x, y) 5 f m (x)c m (y), coefficients are not calculated. For implementation with
0 , m # mmax , it is useful to set
C3m (x, y) 5 c m (x)c m (y), and
w(0, T ) 5 0,
F m (x, y) 5 f m (x)f m (y). (B9)
w(mmax , T ) 5 1, and
The one-dimensional Haar wavelet and smoothing func-
tions are shown in the upper part of the Fig. B1. Below I m50
f (x, y, T ) 5 I f . (B14)
this, the two-dimensional patterns of three Haar wave- The reconstruction formula is then restated as

Unauthenticated | Downloaded 04/18/21 01:39 PM UTC

248 JOURNAL OF APPLIED METEOROLOGY VOLUME 43

REFERENCES

Austin, G. L., and A. Bellon, 1974: The use of digital weather radar
records for short-term precipitation forecasting. Quart. J. Roy.
Meteor. Soc., 100, 658–664.
Bellon, A., and G. L. Austin, 1978: The evaluation of two years of
real-time operation of a short-term precipitation forecasting pro-
cedure (SHARP). J. Appl. Meteor., 17, 1778–1787.
——, and I. Zawadzki, 1994: Forecasting of hourly accumulations
of precipitation by optimal extrapolation of radar maps. J. Hy-
drol., 157, 211–233.
——, S. Lovejoy, and G. L. Austin, 1980: Combining satellite and
radar data for the short-range forecasting of precipitation. Mon.
Wea. Rev., 108, 1554–1566.
Browning, K. A., C. G. Collier, P. R. Larke, P. Menmuir, G. A. Monk,
and R. G. Owens, 1982: On the forecasting of frontal rain using
a weather radar network. Mon. Wea. Rev., 110, 534–552.
Burrus, C. S., R. A. Gopinath, and H. Guo, 1998: Introduction to
Wavelets and Wavelet Transforms: A Primer. Prentice Hall, 268
pp.
Ersoy, O. K., 1997: Fourier-Related Transforms, Fast Algorithms
and Applications. Prentice Hall, 522 pp.
Germann, U., and I. Zawadzki, 2002: Scale-dependence of the pre-
dictability of precipitation from continental radar images. Part
I: Description of the methodology. Mon. Wea. Rev., 130, 2859–
2873.
——, and ——, 2004: Scale dependence of the predictability of pre-
cipitation from continental radar images. Part II: Probability
forecasts. J. Appl. Meteor., 43, 74–89.
Hudgins, L. H., M. E. Mayer, and C. A. Friehe, 1993: Wavelet trans-
forms and atmospheric turbulence. Phys. Rev. Lett., 71, 3279–
3282.
Laroche, S., and I. Zawadzki, 1994: A variational analysis method
for retrieval of three-dimensional wind field from single–Doppler
radar data. J. Atmos. Sci., 51, 2664–2682.
——, and ——, 1995: Retrievals of horizontal winds from single-
Doppler clear-air data by methods of cross-correlation and var-
iational analysis. J. Atmos. Oceanic Technol., 12, 721–738.
Lorenz, E. N., 1969: The predictability of a flow which possesses
FIG. B2. Schematic illustration of low-pass and bandpass filtering many scales of motion. Tellus, 21, 289–307.
of the fast wavelet transform: S0 represents the image at finest res- Mallat, S. G., 1989: A theory for multiresolution decomposition.
olution, S1 after smoothing to scale index m 5 1, S2 after smoothing IEEE Trans. Pattern Anal. Mach. Intell., 11, 674–693.
to scale index m 5 2, and so on, and W1, W2, and W3 represent Orlanski, I., 1975: A rational subdivision of scales for atmospheric
image features described by wavelet coefficients at scales m 5 1, 2, processes. Bull. Amer. Meteor. Soc., 56, 527–530.
3, respectively. Schertzer, D., and S. Lovejoy, 1987: Physical modeling and analysis
of rain and clouds by anisotropic scaling multiplicative pro-
cesses. J. Geophys. Res., 92, 9693–9714.
Seed, A., 2003: A dynamic and spatial scaling approach to advection
forecasting. J. Appl. Meteor., 42, 381–388.
I f s (x, y, T ) Tsonis, A. A., and G. L. Austin, 1981: An evaluation of extrapolation

O
mmax21
techniques for the short-term prediction of rain amounts. Atmos.–
Ocean, 19, 54–65.
5 [w(m 1 1, T ) 2 w(m, T )]I mf (x, y, T ), (B15) Walker, J. S., 1988: Fourier Analysis. Oxford University Press, 440
m50
pp.
Walton, M. L., P. R. Ahnert, and M. D. Hudlow, 1985: Proposed on-
and the spatial smoothing function for the optimal filter site flash flood potential system for NEXRAD. Preprints, Sixth
is Conf. on Hydrometeorology, Indianapolis, IN, Amer. Meteor.

O
mmax21
Soc., 122–129.
Weaver, H. J., 1989: Theory of Discrete and Continuous Fourier
[w(m 1 1, T ) 2 w(m, T )]F m (x, y, T ). (B16) Analysis. John Wiley and Sons, 307 pp.
m50
Zawadzki, I. I., 1973: Statistical properties of precipitation patterns.
J. Appl. Meteor., 12, 459–472.
In this way, the optimally filtered forecast is constructed ——, J. Morneau, and R. Laprise, 1994: Predictability of precipitation
as a weighted sum of the Haar wavelet smoothed images patterns: An operational approach. J. Appl. Meteor., 33, 1562–
I mf at each scale m. 1571.

Unauthenticated | Downloaded 04/18/21 01:39 PM UTC