MARCH 2020 CHUNG AND YAO 1099

Improving Radar Echo Lagrangian Extrapolation Nowcasting by Blending Numerical

Model Wind Information: Statistical Performance of 16 Typhoon Cases

KAO-SHEN CHUNG AND I-AN YAO

Department of Atmospheric Sciences, National Central University, Taoyuan City, Taiwan

(Manuscript received 13 June 2019, in final form 4 October 2019)

ABSTRACT

Severe weather nowcasting is a crucial mission of atmospheric science for the betterment of society to save
life, limb, and property. In this study, composite radar data from the Central Weather Bureau of 16 typhoons
are collected to examine the statistical performance of the McGill Algorithm for Precipitation nowcasting
using Lagrangian Extrapolation (MAPLE) over Taiwan, an extrapolation algorithm that predicts future
precipitation based on current radar echoes. In addition, instead of mixing the precipitation between radar
extrapolation and numerical model forecast as in previous studies, a blending system is formed by synthe-
sizing the wind information from model forecast with the echo extrapolation motion field via a variational
algorithm to improve the nowcasting system. The statistical results of the radar echo extrapolation for 16
typhoon cases show that while the quantitative precipitation nowcasting skill can persist for up to 2 h, sig-
nificant distortion for the rotational system is found after 2 h. On the other hand, the blending system helps to
capture and maintain the rotation of typhoon rainband structures. The blending system extends the now-
casting skill by 1 h to a total of 3 h. Furthermore, the blending scheme performs especially well after the
typhoon makes landfall in Taiwan. For disaster prevention and mitigation, this blending nowcasting technique
may provide effective weather information immediately.

1. Introduction amount of time. However, model forecasts face spinup

or spindown issues, whereby a certain amount of time
Heavy rain from typhoons is one of the most severe
(usually 1–3 h) is required after warm-start initializa-
weather systems in Taiwan. Heavy rains and strong wind
tion to reach a stable model state (Shrestha et al. 2013;
accompany typhoons, and these phenomena may last for
Chung et al. 2013; Jacques et al. 2017). On the other
several hours before and after landing in Taiwan, lead-
hand, by assimilating radar observations in different
ing to disasters such as flooding and landslides. To
data assimilation systems, as previous studies have
mitigate the disastrous impacts of heavy rain from
shown, the improvement of QPF sometimes could last
typhoons, quantitative precipitation forecasts (QPF)
up to 6 h (Kain et al. 2010; Sun et al. 2014), and some-
in the time scale of the very short term, known as
times the impact may remain confined to around 1–2 h
‘‘nowcasting’’ (less than 6 h), play an important role in
(Aksoy et al. 2010; Chang et al. 2016; Chang et al.
disaster prevention in Taiwan.
2014; Chung et al. 2009). Therefore, to forecast pre-
In recent years, numerical weather models have
cipitation of severe weather in the very short-term
shown significant improvements in forecast skill due
effectively, radar echo extrapolation remains a pow-
to the development and refinement of physical mod-
erful and highly relevant method.
eling and computational techniques. These models
Because weather radar is capable of providing high
provide increasingly fine spatial and temporal reso-
temporal and spatial resolution over a certain area, it re-
lution forecast fields for users in an ever-decreasing
mains the most useful instrument to survey severe weather
over land and near coasts where such radar systems are
Denotes content that is immediately available upon publica- available. Radar extrapolation is a well-established tech-
tion as open access. nique to perform quantitative precipitation nowcasting
(QPN). For instance, Thunderstorm Identification,
Corresponding author: Dr. Kao-Shen Chung, [email protected]. Tracking, Analysis and Nowcasting (TITAN) tracked
edu.tw and extrapolated storms linearly by identifying the

DOI: 10.1175/MWR-D-19-0193.1
Ó 2020 American Meteorological Society. For information regarding reuse of this content and general copyright information, consult the AMS Copyright
Policy (www.ametsoc.org/PUBSReuseLicenses).
Brought to you by SEJONG UNIVERSITY LIBRARY | Unauthenticated | Downloaded 04/16/21 05:58 PM UTC
1100 MONTHLY WEATHER REVIEW VOLUME 148

centroids of the weather system (Dixon and Wiener replaces the motion field estimated by echo movement.
1993). The Short-range Warnings of Intense Rainstorms In a summer case study, their results showed that a
in Localized System (SWIRLS) developed by the Hong composite approach can perform effective nowcasts for
Kong Observatory uses the technique named Tracking up to a 3-h lead time. Sokol et al. (2017) included the
Radar Echoes by Correlations (TREC) (Rinehart and information of tracking errors in the motion field.
Garvey 1978; Tuttle and Foote 1990). TREC calculates When focusing on strong convection in warm seasons,
the correlation coefficients between consecutive radar their results demonstrated that the forecast skill is
echo images and proceeds to compute an extrapola- useful up to 60 min. Ryu et al. (2019) introduced an
tion forecast. When the information of SWIRLS is advection-diffusion model to modify motion vectors,
used to modify the hydrometeor variables of the nu- and concluded that when motion vectors are time de-
merical model, it can extend the forecast ability (Li pendent in the nowcast lead time, better nowcast per-
and Lai 2004). Instead of finding the maximum cor- formance can be achieved.
relation to get the motion vector, such as TREC, the Recently, probabilistic nowcasting systems have also
McGill Algorithm for Precipitation nowcasting using been used to improve the capability of nowcasting.
Lagrangian Extrapolation (MAPLE; Germann and Based on the stochastic perturbation of the Lagrangian
Zawadzki 2002, 2004) uses the variational method to extrapolation, Atencia and Zawadzki (2014) generated
minimize a cost function to define the motion field that an ensemble to alleviate the errors of growth and decay.
then advects the radar echo images for nowcasting. Their results have shown that this approach can better
The major drawback of the extrapolation-based reproduce the spatial structure of the precipitation sys-
nowcasting based on extrapolation is that it is difficult tem and extend the nowcast lead time beyond 2–3 h
to capture the growth and decay of the weather system statistically. Sokol et al. (2017) used random perturba-
and the uncertainty of the displacement. To overcome tions generated by the historic radar data from the warm
this issue, several studies have applied blending tech- season (May to September) of years 2009–12 to conduct
niques to improve nowcasting systems. For instance, an ensemble nowcast. Their results showed that the
the Short-Term Ensemble Prediction System (STEPS; forecast skill in summertime (July 2012) can last ap-
Seed 2003), developed by the Met Office and Bureau of proximately 30–40 min.
Meteorology in Australia, utilized the Fourier filter to The purpose of this study is to examine and improve the
decompose precipitation information (radar echoes) performance of MAPLE for nowcasting heavy precipi-
into different spectral scales, and then combined the tation in the Taiwan area. Taiwan, a subtropical island,
extrapolation nowcast and downscaled NWP forecast regularly experiences heavy precipitation exceeding
to produce a probabilistic precipitation forecast. Based 15 mm h21 and 50 mm day21 (Chen et al. 2007). Since the
on this algorithm, the forecasting skill can persist for up average mountain height in Taiwan is around 2 km with
to 6 h in Australia (Bowler et al. 2006). The Rainstorm peaks up to 4 km, it is a natural environment for evalu-
Analysis and Prediction Integrated Data-processing ating MAPLE’s performance over complex terrain. In
System (Li et al. 2005) merges the precipitation infor- previous studies, MAPLE has shown its capability for
mation from the Nonhydrostatic Model (NHM) with nowcasting up to 2–6 h depending on the regions being
SWIRLS echo extrapolation. With a rainstorm case implemented (Germann and Zawadzki 2002, 2004;
study, the results show an improvement of 3-h now- Turner et al. 2004; Bellon et al. 2010; Lee et al. 2010,
casts. Similar to SWIRL, the Adjustment of Rain from Mandapaka et al. 2012). MAPLE was first applied and
Models with Radar (ARMOR; Lee et al. 2009) com- configured over Taiwan in 2018 (Pan et al. 2018). In their
bines the Weather Research and Forecasting (WRF) study, two typhoon cases and one frontal system case
Model with MAPLE. By estimating the phase error were selected to examine the performance of MAPLE.
of precipitation between WRF and observations, the The results showed that MAPLE outperformed persis-
blending system modifies the rainfall rate of the model. tence and it produced reasonable skill for nowcasting up
It is able to improve the distribution of the precipita- to 2 h. In this study, 16 typhoon cases are selected to ex-
tion and reduce the false alarm area from NWP, but amine the performance of MAPLE statistically over
echo extrapolation still outperforms the nowcast in the Taiwan’s complex terrain. In addition, to alleviate the
first few hours (DuFran et al. 2009). Instead of blending distortion of the rotational weather system during ex-
information between radar observations and NWP trapolation, wind information from a numerical model is
models, several studies focus on mixing model forecast included to improve the nowcasting system. Each ty-
winds with the nowcast system. By applying a fuzzy phoon is divided to three stages (before landing, after
algorithm, Liang et al. (2010) proposed a composite landing, and leaving Taiwan) to investigate the results of
approach by which model-predicted wind gradually nowcasting in each stage.

Brought to you by SEJONG UNIVERSITY LIBRARY | Unauthenticated | Downloaded 04/16/21 05:58 PM UTC
MARCH 2020 CHUNG AND YAO 1101

The organization of this study is as follows: the source

of data and the selected cases are described in section 2;
the methodology of MAPLE, the blending algorithm,
and verification scores are illustrated in section 3; the
results of echo extrapolation nowcasting with and without
blending as well as a statistical analysis of the perfor-
mance of nowcast skill are illustrated and discussed in
section 4; section 5 presents the conclusion.

2. Data and cases overview

a. Radar network
Through a multiagency effort, Taiwan developed a QPE
system called Quantitative Precipitation Estimation and
Segregation Using Multiple Sensors (QPESUMS; Gourley
et al. 2002; Zhang et al. 2009). QPESUMS integrates radar
observations over Taiwan. The composite radar obser-
vations are from the Taiwanese Central Weather Bureau
radars RCWF, RCCG, RCKT, and RCHL as well as
the Taiwan Air Force radars RCCK and RCMK. RCCG,
RCKT, and RCHL are S-band Doppler weather radars,
and RCWF is an S-band dual-polarization radar. RCCK
and RCMK are C-band dual-polarization radars. The
positions of the six radars and the coverage of QPESUMS
are shown in Fig. 1a.
The composite reflectivity of the integrated radar
network is selected to evaluate the performance of
MAPLE. The horizontal grid resolution of the composite
reflectivity is 0.01258 and the time interval is 10 min.
Because of Taiwan’s complex terrain, quality control for
the radar data is conducted to reduce the contamination
from ground and sea clutter (Chang et al. 2009).
b. Rain gauge data
There are over 400 stations with rain gauges (Fig. 1b)
over Taiwan Island, and the time resolution is once per
hour. In this study, the nowcast performance is examined
by comparing to surface observations. As the radars used
by QPESUMS are not all dual-polarized radars, the Z–R
relationship is applied to convert reflectivity (Z) to the
rainfall rate (R). In this study, since Taiwan is subtropical,
commonly used Z–R relationships such as the Marshall–
Palmer formula of Z 5 200R1.65 or Z 5 300R1.4 derived in
the midlatitudes may not be appropriate. Therefore, the
following Z–R relationship developed using 7 years of
precipitation data in Taiwan (Chen et al. 2017) is applied:

Z 5 223R1:51 . (1) FIG. 1. The positions of (a) six radars with radar coverage and
(b) the surface observation station in Taiwan.
c. ECMWF ERA-Interim
The European Centre for Medium-Range Weather
Forecasts (ECMWF) interim reanalysis (ERA-Interim)

Brought to you by SEJONG UNIVERSITY LIBRARY | Unauthenticated | Downloaded 04/16/21 05:58 PM UTC
1102 MONTHLY WEATHER REVIEW VOLUME 148

TABLE 1. List of 16 real cases, including the arrival and departure time according to the Central Weather Bureau (CWB). The number
with parentheses indicates the total times of the nowcast.

Case Period of the nowcast T1 T2

2008_Fung-Wong 0600 UTC 27 Jul–0300 UTC 29 Jul (46) 2300 UTC 27 Aug 2008 0700 UTC 28 Jul 2008
2009_Morakot 1300 UTC 6 Aug–1800 9 Aug (78) 1600 UTC 7 Aug 2009 0600 UTC 8 Aug 2009
2010_Namtheun 1500 30 Aug–0900 31 Aug (19) — —
2010_Fanapi 0600 UTC 18 Sep–0700 UTC 20 Sep (50) 0100 UTC 19 Sep 2010 1000 UTC 19 Sep 2010
2013_Soulik 0700 UTC 12 Jul–1500 UTC 13 Jul (33) 1900 UTC 12 Jul 2013 0000 UTC 13 Jul 2013
2014_Matmo 0400 UTC 22 Jul–1500 UTC 23 Jul (36) 1600 UTC 22 Jul 2014 2000 UTC 22 Jul 2014
2014_Fung-Wong 0800 UTC 20 Sep–2100 UTC 21 Sep (38) — —
2015_Soudelor 1300 UTC 7 Aug–0100 UTC 9 Aug (37) 2100 UTC 7 Aug 2015 0300 UTC 8 Aug 2015
2015_Dujuan 1600 UTC 27 Sep–2300 UTC 28 Sep (32) 1000 UTC 28 Sep 2015 1700 UTC 28 Sep 2015
2016_Nepartak 0200 UTC 7 Jul–0600 UTC 9 Jul (53) 2200 UTC 7 Jul 2016 0700 UTC 8 Jul 2016
2016_Meranti 0700 UTC 13 Sep–0300 UTC 15 Sep (46) — —
2016_Megi 1200 UTC 26 Sep–0900 UTC 28 Sep (46) 0600 UTC 27 Sep 2016 1300 UTC 27 Sep 2016
2017_Nesat 2000 UTC 28 Jul–0600 UTC 30 Jul (35) 1100 UTC 29 Jul 2017 1500 UTC 29 Jul 2017
2017_Haitang 1600 UTC 29 Jul–0400 UTC 31 Jul (37) 0100 UTC 30 Jul 2017 1700 UTC 30 Jul 2017
2018_Maria 0800 UTC 10 Jul–0400 UTC 11 Jul (21) — —
2018_Lowpressure 1800 UTC 22 Aug–1300 UTC 24 Aug (44) — —

global reanalysis is used to examine the relationship 3. Methodology

between real wind and the motion vector through vari-
a. Description of MAPLE
ational Echo tracking. The ERA-Interim is produced
by a four-dimensional variational analysis (4D-Var) The nowcasting system, MAPLE, is developed by the
data assimilation system with a 6-h analysis window J. S. Marshall Radar Observatory of McGill University.
and a 0.1258 resolution. In this study, only the wind fields By using the variational algorithm, the forecasting pro-
on the vertical levels of 400, 500, 600, 700, and 850 hPa cedure of MAPLE uses Variational Echo Tracking
are used. (VET) to calculate the motion vector according to the
previous observations and uses semi-Lagrangian ad-
d. Cases description
vection to proceed with the extrapolation. These two
A total of 16 events from 2008 to 2018, including 15 components are introduced in the following section.
typhoons and 1 low pressure system, are selected for this
study (Table 1). The MAPLE system is launched hourly 1) VARIATIONAL ECHO TRACKING TECHNIQUE
to produce nowcasts up to 4-h lead time, amounting to a The original VET technique, developed by Laroche
total of 650 times of nowcast. The best track of all ty- and Zawadzki (1995), used Doppler data in a small re-
phoon cases is shown in Fig. 2. To better evaluate the gion to generate a nowcast. The maximum reflectivity of
performance of the typhoon nowcasts, various tracks that the composite radar network is used to calculate the VET
landed over all regions of Taiwan are included in these motion field. The cost function J of the VET technique can
16 typhoon cases. There are 11 typhoons that landed in be expressed in terms of these two components as follows:
Taiwan. Four of these landed in northern Taiwan, namely
Marokot (2009), Soulik (2013), Dujuan (2015) and Nesat JVET (V) 5 JC 1 JV , (2)
(2017); four typhoons landed in the middle of Taiwan,
namely Typhoons Fung-Wong (2008), Matmo (2014), where V presents the two-dimensional motion vector
Soudelor (2015) and Megi (2016); and three landed in that is evaluated by iteratively minimizing the cost func-
south Taiwan: Typhoons Fanapi (2010), Nepartak (2016) tion. The first term in Eq. (2) is the conservation of re-
and Haitang (2017). The other five typhoons that did not flectivity constraint JC. It represents the sum of squares of
make landfall are also examined. They are the following: the echo residuals in the domain. The equation of JC is
Namtheun (2010) and Maria (2018), which passed the
ðð
north of Taiwan; Fung-Wong (2014), which passed by the
JC 5 b(x)[C(t0 , x) 2 VC(t0 2 Dt, x 2 uDt)]2 dx dy,
east coast of Taiwan; and Meranti (2016) and the low
pressure system (2018), which passed the Bashi Channel (3)
southwest of Taiwan. These typhoons or tropical de-
pression (TD) systems brought heavy rain in a short pe- where b(x) denotes the weight of the reflectivity con-
riod of time over Taiwan. straint that is related to data quality. To comply with the

Brought to you by SEJONG UNIVERSITY LIBRARY | Unauthenticated | Downloaded 04/16/21 05:58 PM UTC
MARCH 2020 CHUNG AND YAO 1103

bilinear interpolation. Therefore, the actual forecasts can

be generated using the semi-Lagrangian scheme as pro-
posed by Germann and Zawadzki (2002). The equation is

t 5 NDt , (5)
 a
a 5 Dtu t0 , x 2 , (6)
2

where t is the entire time of the forecast, and a is

the displacement vector. According to Eqs. (5) and (6),
the entire time is divided into N steps of length Dt. The
advantage of this method is that it allows for different
motion velocities during the forecasting process and
gives the ability to simulate the rotation of the system.
The choice of advection scheme is ‘‘backward in time
and upstream in space,’’ which can also be interpreted to
mean that the information at the current time is decided
by the upstream information, as Germann and Zawadzki
FIG. 2. The best track of 15 typhoon cases. (2002) demonstrated. However, as Bellon et al. (2010)
pointed out, as the center grid point (i, j) is in a divergent
conservation of the reflectivity means the echo does not region, the neighboring grid points may have the same
decay or grow during the forecast, and therefore the in- source as (i0, j0). Therefore, the grid point (i0, j0) may
tegrated domain must be smaller than the actual com- cause stretching or an increase in the area of the forecast
posite map in order to ensure the offset displacement precipitation. Conversely, if the grid point (i, j) is in a
C(t0 2 Dt, x 2 uDt) remains within in the domain. The convergent region, the result may be compressing and
second term of Eq. (2) is a penalty function that smooths decreasing in the area of the forecast precipitation. Due
the motion field of the reflectivity by the second deriva- to these two effects, the results may generate distortions
tive in space. The function of JV can be expressed as according to the magnitude of divergence/convergence of
the VET motion field, which will be shown in section 4.
ðð " 2 2  2 2  2 2  2 2
›u ›u ›u ›y b. Model setup
JV 5 g 1 1 2 1
›x2 ›y2 ›x›y ›x2
The WRF (Powers et al. 2017; Skamarock et al. 2008)
 2  2 2 # model version 3.7.1 is used in this study. The WRF domain
›2 y ›y of each case is displayed in Fig. 3. Three nested domains
1 12 dx dy, (4)
›y2 ›x›y with 52 vertical layers and 27, 9, and 3 km horizontal grid
spacings are used with 181 3 181, 301 3 301, 451 3 451
where g is the weight of the smoothness constraint. To grid points, respectively. The physical parameterizations
minimize the cost function efficiently, VET uses the used are the Rapid Radiative Transfer Model (RRTM)
conjugate-gradient algorithm described by Navon and longwave scheme (Mlawer et al. 1997), the Dudhia short-
Legler (1987) to locate the minimum. To avoid the wave scheme (Dudhia 1989), the Yonsei University (YSU)
probability of converging toward a secondary minimum, planetary boundary scheme (Hong et al. 2006), the Grell–
the scaling guess developed by Laroche and Zawadzki Freitas cumulus scheme (Grell and Dévényi 2002), and the
(1994) is used to determine the best motion field. Goddard Cumulus Ensemble (GCE) for the microphysics
There are several user-selectable parameters that can scheme (Tao et al. 2003). Note that the cumulus scheme is
affect the performance of MAPLE. The meaning of the replaced by the microphysical scheme in domain 3.
parameters and the major settings used in this study are In this study, the initial conditions are taken from
shown in Table 2. ERA-Interim (0.758 3 0.758), and the simulation length
for each case is 30 h. The first 6 h are not used to avoid
2) SEMI-LAGRANGIAN ADVECTION
the spinup issue.
After the retrieval of the VET technique is applied,
c. Combination of VET with other sources
the VET motion field is obtained in the (m 3 m) sub-
domain, where (m 3 m) 5 (48 3 48) or (144 3 144) in Since VET is estimated by the advection of radar
this study, and is then interpolated to every grid point by echoes, the motion field outside of the echo area may not

Brought to you by SEJONG UNIVERSITY LIBRARY | Unauthenticated | Downloaded 04/16/21 05:58 PM UTC
1104 MONTHLY WEATHER REVIEW VOLUME 148

TABLE 2. The settings of MAPLE.

VET parameter Brief description Pan et al. (2018) In this study

No. of maps No. of maps used by the VET algorithm 3
Time difference between each map The time interval of the consecutive 30
radar map
Amount of smoothing It is used to remove random noise 333
Reflectivity threshold The threshold of tracking 5
Number of scaling guesses It is used for the nonecho area to retrieve 4 5
the motion field
Vector density The resolution of the motion field 48 3 48 144 3 144
Relative weightings for b and g The user-selectable weightings in 0.5 and 1000
Eqs. (3) and (4)

be reliable. Once the distortion of the nowcast becomes å(F 2 F)(O 2 O)

SCC 5 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi , (8)
significant, the performance will degrade. To improve
å(F 2 F) (O 2 O)
2 2
the performance of the VET motion field, we attempt to
include the steering flow information using a blending
technique. The formula of the blending system can be where F and O are the accumulated rainfall of the predic-
expressed as tion and the observation, and F and O denote an area av-
erage over a two-dimensional plane of the field. The RMSE
ðð
evaluates the rainfall difference between the forecast and the
Jcom 5 [w1 (u 2 uVET )2 1 w1 (y 2 y VET )2
observation quantitatively. The equation of the RMSE is
1 w2 (u 2 uref )2 1 w2 (y 2 y ref )2 ] dx dy, vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
(7) u
uN
u å (F 2 O)2
where ‘‘ref’’ means the information from a source other t i51
RMSE 5 , (9)
than the VET, such as an analysis or numerical model, N
and w1 and w2 are the weighting coefficients for the
VET and ‘‘ref’’, respectively. In this study, in order to where F and O are the same definition as in Eq. (8), and N
find the best set of weightings, a sensitivity test is stands for the total number of grid points used for the calculation.
conducted in section 4. 2) CATEGORICAL VERIFICATION
d. Verification scores The categorical verification scores are formulated as
followed:
To examine the accuracy of the quantitative pre-
cipitation nowcasting (QPN) by MAPLE, the 0–4 h
nowcasts of reflectivity are verified every hour. In this
study, we evaluate the performance by continuous
verification scores, namely the spatial correlation
coefficient (SCC) and the root-mean-square error
(RMSE), as well as categorical scores, namely the
probability of detection (POD), the false-alarm rate
(FAR), the bias, the critical success index (CSI), and
the equitable threat score (ETS). When comparing
the results with rain gauges, only the grid points at the
positions of the rain gauge stations are verified with
the ETS score. In addition, to alleviate the issue of
double-penalty when examining the QPN at high
resolution, the neighborhood method is also used to
assess the performance of MAPLE.

1) CONTINUOUS VERIFICATION
The SCC can evaluate the extent of the similarity
between the echo distribution of the forecast and the
observation. The SCC is defined as FIG. 3. The experimental domain setting of WRF.

Brought to you by SEJONG UNIVERSITY LIBRARY | Unauthenticated | Downloaded 04/16/21 05:58 PM UTC
MARCH 2020 CHUNG AND YAO 1105
a exceeding the given threshold in the various search
POD 5 , (10)
a1b distances. The FSS is defined as
c
FAR 5 , (11) 1
a1c åå[O(L) (i, j) 2 F(L) (i, j)]2
N i j
a FSS(L) 5 1 2 , (16)
CSI 5 , (12) 1 1
a1b1c å å
N i j
O(L) (i, j)2 1 ååF(L) (i, j)2
N i j
a1c
Bias 5 , (13)
a1b where N is the total number of grid points over the
verified domain, and the double sum indicates the total
where a, b, c, and d represent the hit, miss, false alarm, number of the grid points in the verified domain.
and correct negative, respectively, defined by the O(L)(i, j) and F(L)(i, j) are the observation and the
occurrence when the events are predicted and veri- forecast fractions at the grid point (i, j), where the
fied by the observation, the occurrence when the fractions are the ratio of the number of grid points
events are forecast but not verified by the observa- exceeding the given threshold within the radius of
tion, the occurrence when events are not predicted search L. Thus, we can understand the spatial distri-
but verified by the observations, and the occurrence bution of the model skill by the FSS under various radii
when the events are not predicted and verified by the and thresholds. The value of the FSS ranges from 0 to 1,
observation. where FSS equal to 1 represents a perfect forecast.
The scores of these parameters are between 0 and 1 Roberts and Lean (2008) provided the reference value
except BIAS. A POD or CSI equal to 1 represents a of the FSS, which is named FSSref in this study. It de-
perfect forecast, and if POD or CSI equals 0 the model fined as FSSref 5 0.5 1 fo/2, where fo is the fraction of
does not have the ability to forecast. CSI represents the the number of grid points exceeding the threshold in
probability of the successful forecast but does not ac- the verified domain. Thus, when the FSS less than
count for correct negatives. FAR is the number of false FSSref, the model does not have the skill to provide a
hits per the total number of forecast points above a useful forecast. To examine the sensitivity of the search
threshold, and a FAR of 1 means a perfect forecast. The radius (L) to the FSS, L 5 1, 3, 5, 7, 10, 17, and 21 grid
bias is the ratio of the precipitation forecast to its ob- points, according to the distances of 1, 4, 7, 10, 14, 23,
servation. If the bias is greater than 1, it means the and 29 km, respectively, are used and the threshold of
forecast overestimated the precipitation events and the reflectivity is 15, 35, and 45 dBZ.
vice versa.
The ETS is similar to the CSI, but also considers the
number of hits by chance. This index is defined as 4. Results and discussion
In this section, we first examined the correlation be-
a2R
ETS 5 , (14) tween the motion field of the VET and the reanalysis
a1b1c2R wind field of ECMWF in section 4a. We then dem-
(a 1 b) 3 (a 1 c) onstrate with one typhoon case that blending the VET
R5 , (15) wind with reanalysis wind has the potential to miti-
(a 1 b 1 c 1 d)
gate the distortion problem in MAPLE nowcasts. In
where R means a random forecast. The ETS can eval- section 4b, we further show the feasibility of blend-
uate the capability of the model forecast beyond a ing the VET wind with wind derived from a WRF
random guess. If a is equal to R, the ETS shows a forecast. In addition, a series of sensitivity tests are
forecast skill score of zero; whereas ETS equal to 1 done to obtain the optimal settings of the blending
implies a perfect forecast. When the ETS is negative, it system. The results of nowcasts for Typhoon Fanapi
means the performance of the forecast is worse than (2010) and all cases are shown in sections 4c and 4d,
the reference measure. In this study, following Kato respectively.
et al. (2017), the predictability limit for ETS is defined
a. Comparing VET with the reanalysis wind field
as the forecast time when the ETS drops to the value of
of ECMWF
0.3 or lower.
Chan and Gray (1982) showed that the relationship
3) NEIGHBORHOOD METHOD
between the steering flow wind field and the track of the
The fractions skill score (FSS) is used to verify the typhoon is, to a first order, quite instructive. When the
adjacent grid points between the ratio of the echo skill of nowcasting by echo extrapolation could last for a

Brought to you by SEJONG UNIVERSITY LIBRARY | Unauthenticated | Downloaded 04/16/21 05:58 PM UTC
1106 MONTHLY WEATHER REVIEW VOLUME 148

couple of hours, the question naturally arises whether

Fung-Wong Morakot Namtheun Fanapi Soulik Matmo Fung-Wong Soudelor Dujuan Nepartak Meranti Megi Nesat Haitang Maria Low pressure
there is any correlation between the motion field of VET

(2018)
0.72
0.81
0.82
0.90
0.86
0.91
0.85
0.90
0.84
0.85
and steering flow. We begin by examining the wind field

TABLE 3. The correlation of u and y component with the reanalysis wind (at 400, 500, 600, 700, and 850 hPa) of ECMWF and the VET field for 16 typhoon cases.
from ECMWF reanalysis in comparison to the VET
wind. The latitude and longitude of the domain are from

(2016) (2016) (2017) (2017) (2018)

0.68
0.72
0.69
0.71
0.69
0.71
0.68
0.71
0.66
0.70
208N, 1188E to 278N, 123.58E. Table 3 shows the corre-
lation of the u and y components for 16 typhoon cases.
Results show that for the selected pressure levels, mod-

0.60
0.65
0.61
0.71
0.66
0.70
0.70
0.67
0.69
0.58
erate to high correlation coefficients are found between
the VET wind and the reanalysis wind. It can be inferred
that since the typhoon is an equivalent barotropic system,

0.78
0.89
0.80
0.88
0.82
0.88
0.84
0.86
0.85
0.83
there exist fairly good correlations as long as the rotation
circulation continues.

0.71
0.78
0.73
0.77
0.74
0.76
0.73
0.77
0.67
0.73
Figures 4a–f compare the radar echoes between ob-
servation (Figs. 4a,d) and nowcast at different lead times
for Typhoon Fanapi. It is found that the radar extrapo-

0.58
0.86
0.64
0.86
0.67
0.87
0.65
0.84
0.61
0.80
lation of MAPLE begins to exhibit shape distortions
from the 2nd hour nowcast (Fig. 4b), and the distortion
becomes significant at the 4 h lead time (Fig. 4e). This is

(2016)
0.66
0.78
0.77
0.83
0.77
0.84
0.77
0.83
0.76
0.78
because the motion field from VET can represent the
circulation of typhoon well in the echo region, but it
cannot capture the rotation structure outside of the

(2015)
0.57
0.68
0.52
0.66
0.52
0.65
0.52
0.65
0.48
0.64
rainband area (Fig. 4g versus Fig. 4h). Therefore, the
advection by VET cannot accurately maintain the ro-
tation structure and causes shape distortion. However,
(2015)
0.77
0.65
0.76
0.66
0.76
0.64
0.77
0.63
0.74
0.63
as the wind field of ECMWF is based on a full dy-
namical model along with a reanalysis of all available
observations, the wind field of ECMWF can accurately
(2014)

reflect this phenomenon (Fig. 4g). Moreover, since there

0.62
0.68
0.60
0.68
0.66
0.71
0.68
0.72
0.72
0.63
are high correlations between the VET wind and the
analysis wind, we investigate if combining the reanalysis
wind field with the VET could improve the nowcasting
(2010) (2013) (2014)
0.83
0.79
0.82
0.80
0.84
0.81
0.85
0.80
0.83
0.75
system. This may be useful, for example, if a forecast
system is reliable after a given amount of spinup or
0.65
0.79
0.67
0.81
0.67
0.82
0.63
0.82
0.60
0.79
spindown time in order to merge in the nowcast.
Furthermore, during spinup, the wind field of a forecast
model may be reliable even if the precipitation is not.
0.68
0.68
0.69
0.74
0.68
0.75
0.66
0.74
0.65
0.72

Therefore, this type of investigation allows the ‘‘best of

both worlds’’ by combining the strengths of MAPLE
and a full numerical weather prediction model. By in-
(2010)
0.58
0.47
0.81
0.84
0.88
0.80
0.86
0.74
0.85
0.71

corporating the reanalysis wind field of ECMWF

(Fig. 4g), with equal weighting coefficients for VET
and reanalysis, the blended motion field indeed has less
(2009)
0.73
0.76
0.73
0.75
0.72
0.73
0.70
0.72
0.69
0.68

distortion and more accurately reproduces the circu-

lation in the nonecho region (Figs. 4c,f,i). When further
examining the 4-h accumulated rainfall on the surface
in Fig. 5, results show that, compared to the observa-
(2008)
0.90
0.89
0.90
0.90
0.91
0.91
0.91
0.92
0.90
0.90

tions in Fig. 5a, nowcasts from the blending motion

field (Fig. 5c) can capture the heavy rainfall pattern
better than the nowcast from echo extrapolation alone
400 hPa u

500 hPa u

600 hPa u

700 hPa u

850 hPa u
y

(Fig. 5b). This indicates that the modification of the

motion field outside of the rainband region can po-
tentially improve the capability of QPN.

Brought to you by SEJONG UNIVERSITY LIBRARY | Unauthenticated | Downloaded 04/16/21 05:58 PM UTC
MARCH 2020 CHUNG AND YAO 1107

FIG. 4. The image of radar echo: (a),(d) observation at 0800 and 1000 UTC 19 Sep 2010. (b),(e) MAPLE forecast after 2 and 4 h from
0600 UTC. (c),(f) MAPLE and EC forecast after 2 and 4 h from 0600 UTC. Motion fields at 0600 UTC: (g) the VET, (h) the ECMWF
reanalysis wind field at 400 hPa, (i) the blending VET obtained from (g) and (h).

The results of the blending system demonstrate the track prediction by numerical models has recently shown
improvement of the nowcast. However, of course, the significant improvement (Halperin et al. 2016; Wu et al.
utility of blending the VET wind with the reanalysis wind 2007). Therefore, we investigate blending the VET with
is limited for nowcasting due to the lack of reanalysis data the wind from a forecast WRF model in this study. The
at operational times. On the other hand, tropical cyclone following section will evaluate if combining VET with the

Brought to you by SEJONG UNIVERSITY LIBRARY | Unauthenticated | Downloaded 04/16/21 05:58 PM UTC
1108 MONTHLY WEATHER REVIEW VOLUME 148

FIG. 5. The rain map of (a) the observation, (b) MAPLE, and (c) MAPLE_EC after 4 h accumulated rainfall.

wind field of WRF output has the same beneficial effect high resolution VET (144 3 144), low resolution blending
as combining VET with the wind field of ECMWF. between VET and WRF (48d02), and high-resolution
blending (144d03). The weightings of VET and WRF
b. Sensitivity test of the combination of VET and
outputs are equal and set to 0.5. Results show that: 1) the
WRF wind fields
SCC score is almost overlapping in different resolutions
To obtain higher temporal resolution of the steering of VET. As Lee et al. 2010 and Pan et al. 2018 point out,
flow around the typhoon system, a forecast wind field improvement from increasing the resolution of VET is
is used. The variational method in Eq. (7) is applied insignificant; and 2) the blending motion field has a cer-
to combine the VET with the wind field of WRF. tain improvement after a 2-h nowcast lead time. The
According to the studies by Bellon et al. (2010) and Pan reason is as shown in Fig. 4i that the blending motion field
et al. (2018), there is little improvement to be found in can capture the rotation of the typhoon structure better
modifying the resolution of VET. However, different than the motion field from echo tracking alone, especially
resolutions of the VET and numerical model outputs in the nonecho area. Therefore, the SCC reveals some
are investigated here to find the optimal blending ra- improvement after approximately a 2-h nowcast lead
tios. For the resolution of VET, vector densities of time. Overall, the combination of both high resolution of
144 3 144 (approximately 7-km resolution) and 48 3 48 VET and WRF output (144d03) demonstrates the best
(approximately 20-km resolution) are selected. For the performance for nowcasting.
WRF model, resolutions of 9- and 3-km are chosen. In Second, different vertical levels (850, 700, 600, 500,
addition, to determine the weighting of w1 and w2 in and 400 hPa) of WRF horizontal wind are examined
Eq. (7), and to decide which vertical level is used for (Fig. 6b), and the weightings of VET (echo tracking
blending, a series of sensitivity tests are conducted. In only) and WRF outputs are still set to have equal weight.
addition, four typhoon cases are selected to evaluate In general, there is no significant difference between
the tests, and they are as follows: Fung-Wong (2008), extracting wind from different levels.
Soulik (2013), Nepartak (2016), and Meranti (2016). Third, echo tracking only (w1 5 1) and WRF output
The four typhoon cases passed the Taiwan area via only (w2 5 1) are tested. In addition, since the result
Bashi Channel, the south of Taiwan, the middle of in Fig. 4h indicates that echo and nonecho regions
Taiwan and the north of Taiwan, respectively. The may have different accuracy and performance, the
averaged SCC score of 4 typhoon cases is used to verify weighting is further differentiated in these two regions
the results. (i.e., the echo area and nonecho area). Therefore, the
Different resolutions of VET and WRF outputs are echo area and the nonecho area are separated by the
first examined. Figure 6a shows the tests of different threshold of 15 dBZ. Table 4 displays the experimental
resolutions of VET and blending with WRF outputs. design of different weightings in Eq. (7). Results of Fig. 6c
The experiments include: low resolution VET (48 3 48), show that in the first 1–2 h, the purely MAPLE weighting

Brought to you by SEJONG UNIVERSITY LIBRARY | Unauthenticated | Downloaded 04/16/21 05:58 PM UTC
MARCH 2020 CHUNG AND YAO 1109

of Exp1 has the best SCC score compared to Exp2 and

Exp3. When differentiating the weightings between
the echo and nonecho area, Exp4 illustrates a better
performance than Exp5 of the blending system after
the 2-h nowcast lead time. From the results above, it
is confirmed that the motion field estimated by echo
tracking is superior in the echo regions, but it is
beneficial to combine the model forecasting wind
field to modify the motion field in the nonecho
regions.
According to a series of sensitivity tests illustrated in
this section, the optimal setting of the blending model is
established, which is called MAPLE_WRF. Before ex-
amining the statistical result of 16 typhoon cases, one
typhoon case is first examined to demonstrate the im-
provement under the optimal settings.
c. An example of the typhoon cases
Before examining the results statistically, Typhoon
Fanapi is selected to demonstrate the performance of
MAPLE and MAPLE_WRF. Fanapi had an ex-
tremely heavy rainfall event occur at 0600 UTC
19 September 2010 in southwest Taiwan. Strong
reflectivity was observed in southern Taiwan, and the
rainband was almost stationary (Fig. 7). The maxi-
mum value of the 4-h accumulated precipitation was
over 100 mm h21 (Fig. 7a), which caused heavy flooding
and numerous downed trees in Tainan, Kaohsiung, and
Pingtung in southern Taiwan.
From the results of the nowcasts in Figs. 7b and 7c,
both MAPLE and MAPLE_WRF are able to re-
produce the strong reflectivity in southern Taiwan.
However, as mentioned in section 3a(2), severe shape
distortion of reflectivity occurred on the MAPLE sys-
tem after 2 h of nowcast. When blending WRF hori-
zontal wind with the nowcasting system, the shape
distortion is alleviated and the rotation structure of the
typhoon is well maintained (Fig. 7c). This result shows
that the motion field of MAPLE_WRF can improve the
circulation of the typhoon. Furthermore, comparing
total accumulated rainfall estimated by radar echo
with surface observations in Fig. 7a, the nowcasting
results of MAPLE and MAPLE_WRF have a similar
precipitation pattern as the observations in south
Taiwan, but the total rainfall from both results are
underestimated in the nowcast lead time in general.
However, MAPLE_WRF (Fig. 7c) represents better
accumulated rainfall pattern after a 2-h nowcast lead
time compared to MAPLE (Fig. 7b). Overall, blend-
FIG. 6. The average SCC score of 4 typhoon cases on (a) the tests ing WRF horizontal winds produces similar effects
of different resolutions. 48 3 48 and 144 3 144 mean the resolution
of the VET. 144d03 and 48d02 mean the combination of the VET
and features as seen when combining the wind of
(144 3 144, 7 km, and 48 3 48, 20 km) and the WRF wind in outputs the reanalysis ECMWF with the echo extrapolation
(3 and 9 km); (b) the tests of different pressure levels; (c) the dif- system.
ferent experimental tests (Table 4).

Brought to you by SEJONG UNIVERSITY LIBRARY | Unauthenticated | Downloaded 04/16/21 05:58 PM UTC
1110 MONTHLY WEATHER REVIEW VOLUME 148

TABLE 4. The experimental tests of different weightings on Eq. (7). of MAPLE_WRF is generally higher than the ETS
of MAPLE in all different categories of precipitation.
Experiment Echo area Nonecho area
Moreover, the ETS of MAPLE_WRF is around 0.5
Exp1 w1 5 1.0, w2 5 0.0
after a 4-h nowcast lead time for the threshold
Exp2 w1 5 0.5, w2 5 0.5
Exp3 w1 5 0.0, w2 5 1.0 above 40 mm.
Exp4 w1 5 0.7 w1 5 0.3 The entire period of Typhoon Fanapi over the Taiwan
w1 5 0.3 w1 5 0.7 area (;50 h) is validated by categorical verification and
Exp5 w1 5 0.3 w1 5 0.7 the neighborhood method. Figure 9 displays the ETS score
w1 5 0.7 w1 5 0.3
of accumulated rainfall for Typhoon Fanapi. Results show
no significant difference of changing the density of VET
(48 3 48 versus 144 3 144). In addition, the results of echo
Figure 8 shows the ETS score of MAPLE and extrapolation from MAPLE (48 3 48, 144 3 144) and
MAPLE_WRF at 0600 UTC 19 September 2010. An blending with MAPLE_WRF are in general better than
ETS score above 0.3 is marked in red and considered as the result for WRF, and the ETS scores for different
good performance. Results of accumulated rainfall in thresholds of rainfall accumulation are all above or
Fig. 8a show that MAPLE performs relatively well near 0.3 for 4-h nowcast lead time. MAPLE_WRF
at nowcasting, especially at capturing the heavy rain- performs especially well during the heavy rainfall
fall (a precipitation threshold above 30 mm h21). On accumulation up to the second hour. These results
the other hand, Fig. 8b illustrates that MAPLE_WRF indicate that blending system (MAPLE_WRF) is an
has further improved the nowcasting system. The ETS improvement for Typhoon Fanapi.

FIG. 7. The radar echo and the rain map at 0600 UTC 19 Sep 2010 of the second and fourth hour accumulated precipitation of (a) the
observation, (b) MAPLE, and (c) MAPLE_WRF.

Brought to you by SEJONG UNIVERSITY LIBRARY | Unauthenticated | Downloaded 04/16/21 05:58 PM UTC
MARCH 2020 CHUNG AND YAO 1111

and MAPLE_WRF performs better in the 2–4 h now-

cast. Figure 11 demonstrates multiple verification mea-
sures of forecast accuracy on the performance diagram
(Roebber 2009). The best forecast quality means POD,
the successive ratio (1-FAR), BIAS, and CSI all ap-
proach 1, and lie in the upper-right corner of the dia-
gram. The warm color area represents CSI . 0.6. The
dashed line represents BIAS, and the results along the
diagonal are equals to 1, meaning the forecast is unbi-
ased. Figures 11a and 11b display thresholds of 15 and
35 dBZ, respectively. Again, MAPLE with different
densities of VET (blue and green dots) shows no signifi-
cant difference in skill, while in contrast MAPLE_WRF
(red dots) moves the scores toward the upper-right corner
consistently for all the cases, indicating improvements
over the two MAPLE runs without blending. This indi-
cates some improvement by using multiple verifications
in the performance diagram.
Figure 12 compares the ETS scores of WRF,
MAPLE48, MAPLE144, and MAPLE_WRF verified
against surface rain gauges. When examining a total of
16 typhoon cases in Fig. 12a, results show that the
nowcasting systems outperform complex model fore-
casts in the 0–4 h nowcast in general. In addition,
MAPLE_WRF has very good performance (ETS .
0.3) for the heavy rainfall condition (a threshold of
100 mm accumulation) in the 2-h nowcast lead time
(Fig. 12a). When differentiating the types of typhoons
by landfall (Fig. 12b) and without landfall (Fig. 12c)
FIG. 8. The ETS score at 0600 UTC 19 Sep 2010. (a) MAPLE;
over Taiwan, nowcasts by MAPLE and MAPLE_WRF
(b) MAPLE_WRF. The red represents the ETS score above 0.3. are better than WRF in general given this particular
experimental setup. In addition, Fig. 12b illustrates that
when typhoons land over Taiwan and bring large
d. Statistical performance
amount of accumulated rainfall (threshold . 90 mm),
As the results of Typhoon Fanapi in section 4c show, MAPLE_WRF can perform better than WRF with our
the blending system (MAPLE_WRF) quantitatively experimental setup.
improves and extends nowcasting skill after 2 h. In ad- On the other hand, according to the arrival and de-
dition, significant improvement is demonstrated when parture time of typhoon issued by Central Weather
Typhoon Fanapi is inland over Taiwan. In this section, Bureau (CWB), 16 typhoon cases are further divided into
16 real typhoon cases are examined in order to evaluate three stages: T1 is the stage when a typhoon approached
the performance of the nowcasting system statistically. but its center did not land on Taiwan; T2 is the stage after
First, verification of the echo is performed. Figure 10 the typhoon center landed and before it left the island;
presents the SCC and RMSE of MAPLE48 3 48, and T3 is the stage after the typhoon left Taiwan. When
MAPLE144 3 144 and MAPLE_WRF (the blending evaluating the nowcasting system in different stages in
system). The SCC score demonstrates that MAPLE Fig. 13, the ETS score at stage T1 illustrates that nowcasts
has the best performance in the 0–2 h nowcast, and from the WRF model perform better than the others.
MAPLE_WRF outperforms the echo extrapolation in However, MAPLE and MAPLE_WRF perform better
the 2–4 h nowcast. As mentioned above, this is because than WRF at stages T2 and T3. The ETS scores of both
the nowcast of the MAPLE system presents severe MAPLE and MAPLE_WRF at stage T2 are near 0.3 in
shape distortion after 2 h of nowcast lead time. By the 0–2 h nowcast lead time (Fig. 13b), much higher than
blending the information from the WRF model, this those of WRF. For the threshold .90 mm when typhoons
issue can be alleviated. The results of RMSE also show are away from Taiwan, MAPLE_WRF in Fig. 13c shows
that MAPLE has smaller errors in the 0–2 h nowcast, that the blending system can still capture the heavy

Brought to you by SEJONG UNIVERSITY LIBRARY | Unauthenticated | Downloaded 04/16/21 05:58 PM UTC
1112 MONTHLY WEATHER REVIEW VOLUME 148

FIG. 9. The ETS score of the accumulated rainfall of Typhoon Fanapi during entire period.

rainfall situation that the other systems could not achieve good for all the systems. Considering the moderate
(with no data for computing ETS). rainfall (35 dBZ) with search distance L 5 14 km
Figure 14 presents the FSS score for 16 typhoons in (purple line), while the performance of MAPLE48 and
the entire period over Taiwan area. MAPLE and MAPLE144 can last between 2 and 3 h of nowcast lead
MAPLE_WRF perform well for light rain (Figs. 14a– time, MAPLE_WRF shows some improvement and
c). For the moderate rainfall (35 dBZ) with search extends the capability of nowcasting for more than 3 h
distance L 5 14 km (purple line), the nowcast ability of statistically at stages T1 and T2. For heavy rainfall
MAPLE is less than 3 h, while MAPLE_WRF can im-
prove and extend the nowcast capability beyond 3 h
statistically (Figs. 14d–f). For heavy rainfall events
(45 dBZ) with L 5 29 km (Figs. 14g–i), there is no
significant difference among the three nowcasting sys-
tems. MAPLE_WRF only demonstrates improvement
when a very large distance is examined (L 5 41, not
shown). The FSS score is further computed in two
groups: one group in which all typhoons landed over
Taiwan, and another in which typhoons passed by
but did not land in Taiwan. Results of FSS are
similar to Fig. 14 (not shown), and they indicate
that the capability of nowcasting is not affected by
whether or not the typhoon has landed in Taiwan.
The FSS score for typhoons landed in Taiwan is
computed at stages T1 (Fig. 15a), T2 (Fig. 15b), and T3 FIG. 10. The average SCC (solid) and RMSE (dashed) of all
(Fig. 15c). The threshold of light rain (15 dBZ) is quite real cases.

Brought to you by SEJONG UNIVERSITY LIBRARY | Unauthenticated | Downloaded 04/16/21 05:58 PM UTC
MARCH 2020 CHUNG AND YAO 1113

FIG. 11. The performance diagram of 16 typhoon cases for the reflectivity threshold of (a) 15
and (b) 35 dBZ. The colors represent the forecast results of the different model in the legend,
and the circles along a line are the scores of the echo forecast per 30 min.

events (45 dBZ), MAPLE_WRF only demonstrates than the other stages statistically. This demonstrates
slight improvement when very large distances are ex- that MAPLE_WRF can provide beneficial information
amined (L 5 41, not shown). Overall, the results of to mitigate the disastrous events after a typhoon makes
MAPLE and MAPLE_WRF at stage T2 are better landfall in Taiwan.

Brought to you by SEJONG UNIVERSITY LIBRARY | Unauthenticated | Downloaded 04/16/21 05:58 PM UTC
1114 MONTHLY WEATHER REVIEW VOLUME 148

FIG. 12. The average ETS in 4-h accumulated rainfall. In total, there were
(a) 16 cases and (b) 11 cases that landed in Taiwan. (c) There were five cases that
did not land in Taiwan.

Brought to you by SEJONG UNIVERSITY LIBRARY | Unauthenticated | Downloaded 04/16/21 05:58 PM UTC
MARCH 2020 CHUNG AND YAO 1115

FIG. 13. The average ETS of 11 cases landed Taiwan as in Fig. 12. (a) Stage T1, (b) stage
T2, and (c) stage T3.

Brought to you by SEJONG UNIVERSITY LIBRARY | Unauthenticated | Downloaded 04/16/21 05:58 PM UTC
1116 MONTHLY WEATHER REVIEW VOLUME 148

FIG. 14. The average FSS score of 16 typhoon cases during entire period. (a)–(c) 15 dBZ (;0.27 mm h21), (d)–(f) 35 dBZ
(;5.8 mm h21), and (g)–(i) 45 dBZ (;27 mm h21) during entire period of MAPLE (48 3 48), MAPLE (144 3 144), and MAPLE_WRF,
respectively. The Z–R relationship in Eq. (1) is used to estimate the hourly rainfall rate.

In summary, the blending of WRF wind information extremely heavy rainfall. In addition, when differen-
with echo tracking motion field is found to be a fea- tiating typhoons in three stages, the forecast of WRF
sible method for precipitation nowcasting. Examining performs better before typhoons land over Taiwan.
the performance of nowcasting for typhoon cases Both MAPLE and MAPLE_WRF have good perfor-
over Taiwan, we demonstrated that MAPLE is able mance when typhoons stay inland of Taiwan. The re-
to improve nowcast skill over WRF for at least 2 h. sults indicate that the nowcasting system based on the
After solving the issue of severe shape distortion by extrapolation method can nowcast (0–3 h) relatively
blending wind information from WRF, significant well even when typhoons have made landfall and are
improvement is seen in the 2–4 h nowcast, even for affected by complex terrain.

Brought to you by SEJONG UNIVERSITY LIBRARY | Unauthenticated | Downloaded 04/16/21 05:58 PM UTC
MARCH 2020 CHUNG AND YAO 1117

FIG. 15. The average FSS of 11 cases landed in Taiwan as in Fig. 14. (a) Stage T1, (b) stage T2, and (c) stage T3.

Brought to you by SEJONG UNIVERSITY LIBRARY | Unauthenticated | Downloaded 04/16/21 05:58 PM UTC
1118 MONTHLY WEATHER REVIEW VOLUME 148

5. Conclusions improvement is found after the 2-h nowcasting lead

time, especially for moderate and heavy rain events
In this study, we investigate the capability of echo
(with threshold of 35 and 45 dBZ, respectively). The
extrapolation over the Taiwan for severe precipitation
performance of MAPLE_WRF illustrates that it
nowcasting area with 16 typhoon cases. By providing
improves the nowcasting system for another hour,
additional information from reanalysis or a NWP fore-
which extends the nowcast skill up to 3 h.
cast, the blending system (MAPLE_WRF) is formed
(v) Before landing in Taiwan, the radar network could
and its performance is assessed. Through the statistical
not effectively capture the entire typhoon structure
results of 16 typhoon cases, feasibility assessments of
and the corresponding rapid growth and/or decay of
MAPLE and MAPLE_WRF are presented. The verifica-
rainbands when propagating from ocean to land.
tion is examined by using continuous verification with SCC
Therefore, the performance of nowcasting needs im-
and RMSE, categorical verification with POD, FAR, CSI,
provement during stage T1. When the entire typhoon
BIAS, and ETS and the neighborhood method with FSS.
structure is covered by the radar network, the perfor-
The results of this study are summarized as follows:
mance of nowcasting could be adequate for at least
(i) When examining the relationship of the motion field 3 h at the stage of T2 and T3, even though the
(VET) and horizontal wind, it is found that VET and structure of typhoon is affected by the complex
the reanalysis wind field of ECMWF are highly cor- terrain over Taiwan.
related, and the high correlation coefficient between
Overall, the above results show the nowcasting system
VET and the reanalysis wind field can be extended
by using echo extrapolation can provide useful nowcasting
from the low levels to high levels as mature typhoons
information in the very short-term forecast in typhoon
are, to a first order, essentially barotropic systems.
cases. By combining the VET motion with the wind field
Furthermore, the ‘‘best-case’’ results of combining
from other sources, such as data from a numerical weather
the reanalysis wind field (which is not generally
model, the nowcast can better capture the rotational
available at runtime) and the VET illustrate the
structure of the typhoon, especially in the nonecho areas.
potential improvement of the nowcasting for main-
This blending scheme is feasible for application to other
taining the rotation structure and presenting a better
radar-based nowcasting systems in general. Although the
precipitation pattern of the accumulated rainfall.
growth and decay of the weather system is still an un-
(ii) The severe distortion of the typhoon circulation is
resolved issue in the current system, the blending now-
illustrated after 2 h of the nowcasting during the
casting system extends the ability of very short-term
process of the echo extrapolation. To improve the
forecasts, and it can be used to help reduce the devastation
nowcasting system, a series of sensitivity tests are
caused by severe weather systems.
conducted to test how to properly combine the
NWP model and MAPLE. The result shows that
blending the resolution of VET ;10 km with WRF Acknowledgments. This work was supported by
horizontal wind at 3-km resolution provides the Ministry of Science and Technology of Taiwan under
best performance. For the weightings between VET Research Grant 107-2625-M-008-003. Discussions with
and the wind of the numerical model, results show Dr. Isztar Zawadzki on the MAPLE system have been of
that VET is quite reliable and it should be given great inspiration. The authors thank the Central Weather
more weight in the echo region. On the other hand, Bureau for providing QPESUMS data. The authors are
adopting the wind of the numerical model in the grateful to Dr. Jeffrey Lawrence Steward, who provided
areas without radar echo improves the nowcasting review and proofreading of the manuscript.
from the second hour of nowcast onward.
(iii) Through the statistical performance of 16 real REFERENCES
cases, the capability of MAPLE nowcasting is
shown to last around 2 h. For the heavy rain event Aksoy, A., D. C. Dowell, and C. Snyder, 2010: A multicase com-
parative assessment of the ensemble Kalman filter for assim-
of 45 dBZ, the system can last for approximately ilation of radar observations. Part II: Short-range ensemble
1 h. Overall, in typhoon cases, echo extrapolation forecasts. Mon. Wea. Rev., 138, 1273–1292, https://doi.org/
can provide beneficial information for up to 2 h in 10.1175/2009MWR3086.1.
both categorical and neighborhood verifications. Atencia, A., and I. Zawadzki, 2014: A comparison of two tech-
(iv) The results of MAPLE_WRF show the blending niques for generating nowcasting ensembles. Part I: Lagrangian
ensemble technique. Mon. Wea. Rev., 142, 4036–4052, https://
system helps to capture and maintain the rotation of the doi.org/10.1175/MWR-D-13-00117.1.
typhoon rainband structures. Compared to MAPLE, Bellon, A., I. Zawadzki, A. Kilambi, H. C. Lee, Y. H. Lee, and
ETS and FSS scores both show that the significant G. Lee, 2010: McGill algorithm for precipitation nowcasting

Brought to you by SEJONG UNIVERSITY LIBRARY | Unauthenticated | Downloaded 04/16/21 05:58 PM UTC
MARCH 2020 CHUNG AND YAO 1119

by Lagrangian extrapolation (MAPLE) applied to the South ——, and ——, 2004: Scale dependence of the predictability of
Korean radar network. Part I: Sensitivity studies of the precipitation from continental radar images. Part II: Probability
Variational Echo Tracking (VET) technique. Asia-Pac. forecasts. J. Appl. Meteor., 43, 74–89, https://doi.org/10.1175/
J. Atmos. Sci., 46, 369–381, https://doi.org/10.1007/s13143- 1520-0450(2004)043,0074:SDOTPO.2.0.CO;2.
010-1008-x. Gourley, J. J., R. A. Maddox, K. W. Howard, and D. W. Burgess, 2002:
Bowler, N. E., C. E. Pierce, and A. W. Seed, 2006: STEPS: A An exploratory multisensor technique for quantitative estimation
probabilistic precipitation forecasting scheme which merges of stratiform rainfall. J. Hydrometeor., 3, 166–180, https://doi.org/
an extrapolation nowcast with downscaled NWP. Quart. 10.1175/1525-7541(2002)003,0166:AEMTFQ.2.0.CO;2.
J. Roy. Meteor. Soc., 132, 2127–2155, https://doi.org/10.1256/ Grell, G. A., and D. Dévényi, 2002: A generalized approach to
qj.04.100. parameterizing convection combining ensemble and data as-
Chan, J. C. L., and W. M. Gray, 1982: Tropical cyclone movement similation techniques. Geophys. Res. Lett., 29, 1693, https://
and surrounding flow relationships. Mon. Wea. Rev., 110, doi.org/10.1029/2002GL015311.
1354–1374, https://doi.org/10.1175/1520-0493(1982)110,1354: Halperin, D. J., H. E. Fuelberg, R. E. Hart, and J. H. Cossuth, 2016:
TCMASF.2.0.CO;2. Verification of tropical cyclone genesis forecasts from global
Chang, P.-L., P.-F. Lin, B. Jong-Dao Jou, and J. Zhang, 2009: An numerical models: Comparisons between the North Atlantic
application of reflectivity climatology in constructing radar hy- and eastern North Pacific basins. Wea. Forecasting, 31, 947–
brid scans over complex terrain. J. Atmos. Oceanic Technol., 26, 955, https://doi.org/10.1175/WAF-D-15-0157.1.
1315–1327, https://doi.org/10.1175/2009JTECHA1162.1. Hong, S.-Y., Y. Noh, and J. Dudhia, 2006: A new vertical diffusion
Chang, W., K.-S. Chung, L. Fillion, and S.-J. Baek, 2014: Radar package with an explicit treatment of entrainment processes. Mon.
data assimilation in the Canadian high-resolution ensemble Wea. Rev., 134, 2318–2341, https://doi.org/10.1175/MWR3199.1.
Kalman filter system: Performance and verification with real Jacques, D., W. Chang, S.-J. Baek, T. Milewski, L. Fillion, K.-S.
summer cases. Mon. Wea. Rev., 142, 2118–2138, https:// Chung, and H. Ritchie, 2017: Developing a convective-scale
doi.org/10.1175/MWR-D-13-00291.1. EnKF data assimilation system for the Canadian MEOPAR
Chang, S.-F., Y.-C. Liou, J. Sun, and S.-L. Tai, 2016: The im- project. Mon. Wea. Rev., 145, 1473–1494, https://doi.org/
plementation of the ice-phase microphysical process into a 10.1175/MWR-D-16-0135.1.
four-dimensional Variational Doppler Radar Analysis System Kain, J. S., and Coauthors, 2010: Assessing advances in the assimilation
(VDRAS) and its impact on parameter retrieval and quanti- of radar data and other mesoscale observations within a collab-
tative precipitation nowcasting. J. Atmos. Sci., 73, 1015–1038, orative forecasting–research environment. Wea. Forecasting, 25,
https://doi.org/10.1175/JAS-D-15-0184.1. 1510–1521, https://doi.org/10.1175/2010WAF2222405.1.
Chen, C.-S., Y.-L. Chen, C.-L. Liu, P.-L. Lin, and W.-C. Chen, 2007: Kato, R., S. Shimizu, K.-I. Shimose, T. Maesaka, K. Iwanami, and
Statistics of heavy rainfall occurrences in Taiwan. Wea. H. Nakagaki, 2017: Predictability of meso-g-scale, localized,
Forecasting, 22, 981–1002, https://doi.org/10.1175/WAF1033.1. extreme heavy rainfall during the warm season in Japan using
Chen, J.-Y., W.-Y. Chang, and T.-C. Chen Wang, 2017: Comparison high-resolution precipitation nowcasts. Quart. J. Roy. Meteor.
of quantitative precipitation estimation in Northern Taiwan Soc., 143, 1406–1420, https://doi.org/10.1002/qj.3013.
using S- and C-band dual-polarimetric radars. Atmos. Sci., 110, Laroche, S., and I. Zawadzki, 1994: A variational analysis method for
1354–1374. retrieval of three-dimensional wind field from single-Doppler
Chung, K.-S., I. Zawadzki, M. K. Yau, and L. Fillion, 2009: Short- radar data. J. Atmos. Sci., 51, 2664–2682, https://doi.org/10.1175/
term forecasting of a midlatitude convective storm by the as- 1520-0469(1994)051,2664:AVAMFR.2.0.CO;2.
similation of single–Doppler radar observations. Mon. Wea. ——, and ——, 1995: Retrievals of horizontal winds from single-
Rev., 137, 4115–4135, https://doi.org/10.1175/2009MWR2731.1. Doppler clear-air data by methods of cross correlation and vari-
——, W. Chang, L. Fillion, and M. Tanguay, 2013: Examination of ational analysis. J. Atmos. Oceanic Technol., 12, 721–738, https://
situation-dependent background error covariances at the convec- doi.org/10.1175/1520-0426(1995)012,0721:ROHWFS.2.0.CO;2.
tive scale in the context of the ensemble Kalman filter. Mon. Wea. Lee, G., I. Zawadzki, J. Wilson, M. Xu, A. Kilambi, and J. Pinto, 2009:
Rev., 141, 3369–3387, https://doi.org/10.1175/MWR-D-12-00353.1. Improved precipitation nowcasting: Model errors and their cor-
Dixon, M., and G. Wiener, 1993: TITAN: Thunderstorm Identification, rection in operational numerical weather predictions at different
Tracking, Analysis, and Nowcasting—A radar-based methodol- scales. 34th Conf. on Radar Meteorology, Williamsburg, VA,
ogy. J. Atmos. Oceanic Technol., 10, 785–797, https://doi.org/ Amer. Meteor. Soc.
10.1175/1520-0426(1993)010,0785:TTITAA.2.0.CO;2. Lee, H. C., Y. H. Lee, J.-C. Ha, D.-E. Chang, A. Bellon,
Dudhia, J., 1989: Numerical study of convection observed during I. Zawadzki, and G. Lee, 2010: McGill algorithm for precipi-
the Winter Monsoon Experiment using a mesoscale two- tation nowcasting by Lagrangian extrapolation (MAPLE)
dimensional model. J. Atmos. Sci., 46, 3077–3107, https://doi.org/ applied to the South Korean radar network. Part II: Real-time
10.1175/1520-0469(1989)046,3077:NSOCOD.2.0.CO;2. verification for the summer season. Asia-Pac. J. Atmos. Sci.,
DuFran, Z. M., R. L. Carpenter, B. L. Shaw, G. Lee, A. Kilambi, and 46, 383–391, https://doi.org/10.1007/s13143-010-1009-9.
I. Zawadzki, 2009: Improved precipitation nowcasting algorithm Li, P. W., and E. Lai, 2004: Short-range quantitative precipitation
using a high-resolution NWP model and national radar mo- forecasting in Hong Kong. J. Hydrol., 288, 189–209, https://
saic. 34th Conf. on Radar Meteorology, Williamsburg, doi.org/10.1016/j.jhydrol.2003.11.034.
VA, Amer. Meteor. Soc., 3B.6, https://ams.confex.com/ams/ ——, W. K. Wong, and S. T. Lai, 2005: RAPIDS—A new rainstorm
34Radar/techprogram/paper_156040.htm. nowcasting system in Hong Kong. Proc. WWRP Symp. on
Germann, U., and I. Zawadzki, 2002: Scale-dependence of the Nowcasting and Very Short Range Forecasting, Hong Kong,
predictability of precipitation from continental radar images. WWRP, 7.17, http://www.meteo.fr/cic/wsn05/resumes_longs/
Part I: Description of the ethodmology. Mon. Wea. Rev., 130, 7.17-24.pdf.
2859–2873, https://doi.org/10.1175/1520-0493(2002)130,2859: Liang, Q., Y. Feng, W. Deng, S. Hu, Y. Huang, Q. Zeng, and Z. Chen,
SDOTPO.2.0.CO;2. 2010: A composite approach of radar echo extrapolation based

Brought to you by SEJONG UNIVERSITY LIBRARY | Unauthenticated | Downloaded 04/16/21 05:58 PM UTC
1120 MONTHLY WEATHER REVIEW VOLUME 148

on TREC vectors in combination with model-predicted winds. Seed, A. W., 2003: A dynamic and spatial scaling approach to advec-
Adv. Atmos. Sci., 27, 1119–1130, https://doi.org/10.1007/s00376- tion forecasting. J. Appl. Meteor., 42, 381–388, https://doi.org/
009-9093-4. 10.1175/1520-0450(2003)042,0381:ADASSA.2.0.CO;2.
Mandapaka, P. V., U. Germann, L. Panziera, and A. Hering, 2012: Shrestha, D. L., D. E. Robertson, Q. J. Wang, T. C. Pagano, and
Can Lagrangian extrapolation of radar fields be used for H. A. P. Hapuarachchi, 2013: Evaluation of numerical
precipitation nowcasting over complex alpine orography? weather prediction model precipitation forecasts for short-
Wea. Forecasting, 27, 28–49, https://doi.org/10.1175/WAF-D- term streamflow forecasting purpose. Hydrol. Earth Syst. Sci.,
11-00050.1. 17, 1913–1931, https://doi.org/10.5194/hess-17-1913-2013.
Mlawer, E. J., S. J. Taubman, P. D. Brown, M. J. Iacono, and S. A. Skamarock, W. C., and Coauthors, 2008: A description of the Advanced
Clough, 1997: Radiative transfer for inhomogeneous atmo- Research WRF version 3. NCAR Tech. Note NCAR/TN-
spheres: RRTM, a validated correlated-k model for the 4751STR, 113 pp., https://doi.org/10.5065/D68S4MVH.
longwave. J. Geophys. Res., 102, 16 663–16 682, https://doi.org/ Sokol, Z., J. Mejsnar, L. Pop, and V. Blizn ák, 2017: Probabilistic
10.1029/97JD00237. precipitation nowcasting based on an extrapolation of radar
Navon, I. M., and D. M. Legler, 1987: Conjugate-gradient methods reflectivity and an ensemble approach. Atmos. Res., 194, 245–
for large-scale minimization in meteorology. Mon. Wea. Rev., 257, https://doi.org/10.1016/j.atmosres.2017.05.003.
115, 1479–1502, https://doi.org/10.1175/1520-0493(1987)115,1479: Sun, J., and Coauthors, 2014: Use of NWP for nowcasting convective
CGMFLS.2.0.CO;2. precipitation: Recent progresses and challenges. Bull. Amer. Meteor.
Pan, J.-W., K.-S. Chung, H.-H. Lin, W. T.-C. Chen, and I.-A. Yao, Soc., 95, 409–426, https://doi.org/10.1175/BAMS-D-11-00263.1.
2018: Feasibility assessment of applying variational radar echo Tao, W. K., and Coauthors, 2003: Microphysics, radiation and
tracking method over complex terrain in Taiwan. Atmos. Sci., surface processes in the Goddard Cumulus Ensemble (GCE)
46, 1–34. model. Meteor. Atmos. Phys., 82, 97–137, https://doi.org/
Powers, J. G., and Coauthors, 2017: The Weather Research and 10.1007/s00703-001-0594-7.
Forecasting Model: Overview, system efforts, and future di- Turner, B. J., I. Zawadzki, and U. Germann, 2004: Predictability
rections. Bull. Amer. Meteor. Soc., 98, 1717–1737, https:// of precipitation from continental radar images. Part III:
doi.org/10.1175/BAMS-D-15-00308.1. Operational nowcasting implementation (MAPLE). J. Appl.
Rinehart, R. E., and E. T. Garvey, 1978: Three-dimensional storm Meteor., 43, 231–248, https://doi.org/10.1175/1520-0450(2004)
motion detection by conventional weather radar. Nature, 273, 043,0231:POPFCR.2.0.CO;2.
287–289, https://doi.org/10.1038/273287a0. Tuttle, J. D., and G. B. Foote, 1990: Determination of the boundary
Roberts, N. M., and H. W. Lean, 2008: Scale-selective verification layer airflow from a single Doppler radar. J. Atmos. Oceanic
of rainfall accumulations from high-resolution forecasts of Technol., 7, 218–232, https://doi.org/10.1175/1520-0426(1990)
convective events. Mon. Wea. Rev., 136, 78–97, https://doi.org/ 007,0218:DOTBLA.2.0.CO;2.
10.1175/2007MWR2123.1. Wu, C.-C., K.-H. Chou, P.-H. Lin, S. D. Aberson, M. S. Peng, and
Roebber, P. J., 2009: Visualizing multiple measures of forecast T. Nakazawa, 2007: The Impact of dropwindsonde data on
quality. Wea. Forecasting, 24, 601–608, https://doi.org/10.1175/ typhoon track forecasts in DOTSTAR. Wea. Forecasting, 22,
2008WAF2222159.1. 1157–1176, https://doi.org/10.1175/2007WAF2006062.1.
Ryu, S., G. Lyu, Y. Do, and G. W. Lee, 2019: Improved rainfall Zhang, J., and Coauthors, 2009: High-resolution QPE system for Taiwan.
nowcasting using Burgers’ equation. J. Hydrol., 124140, https:// Data Assimilation for Atmospheric, Oceanic and Hydrologic
doi.org/10.1016/j.jhydrol.2019.124140, in press. applications, S. K. Park and L. Xu, Eds., Springer, 147–162.

Brought to you by SEJONG UNIVERSITY LIBRARY | Unauthenticated | Downloaded 04/16/21 05:58 PM UTC