Content uploaded by Wang-chun WOO
Author content
All content in this area was uploaded by Wang-chun WOO on Feb 27, 2017
Content may be subject to copyright.
Atmosphere 2017, 8, 48; doi:10.3390/atmos8030048 www.mdpi.com/journal/atmosphere
Article
Operational Application of Optical Flow Techniques
to Radar-Based Rainfall Nowcasting
Wang-chun Woo * and Wai-kin Wong
Forecast Development Division, Hong Kong Observatory, Hong Kong 999077, China; wkwong@hko.gov.hk
* Correspondence: wcwoo@hko.gov.hk; Tel.: +852-2926-8453
Academic Editor: Guifu Zhang
Received: 11 November 2016; Accepted: 22 February 2017; Published: 25 February 2017
Abstract: Hong Kong Observatory has been operating an in-house developed rainfall nowcasting
system called “Short-range Warning of Intense Rainstorms in Localized Systems (SWIRLS)” to
support rainstorm warning and rainfall nowcasting services. A crucial step in rainfall nowcasting
is the tracking of radar echoes to generate motion fields for extrapolation of rainfall areas in the
following few hours. SWIRLS adopted a correlation-based method in its first operational version in
1999, which was subsequently replaced by optical flow algorithm in 2010 and further enhanced in
2013. The latest optical flow algorithm employs a transformation function to enhance a selected
range of reflectivity for feature tracking. It also adopts variational optical flow computation that
takes advantage of the Horn–Schunck approach and the Lucas–Kanade method. This paper details
the three radar echo tracking algorithms, examines their performances in several significant
rainstorm cases and summaries verification results of multi-year performances. The limitations of
the current approach are discussed. Developments underway along with future research areas are
also presented.
Keywords: rainfall; nowcast; motion tracking; optical flow
1. Introduction
1.1. Background
Located at the southern coast of China (Figure 1), Hong Kong is affected by heavy rain and
severe weather every year. In the early flood season from March to May, surface frontal system and
monsoonal trough, in tandem with active mid-tropospheric westerly jets, induce squall lines that
lead to heavy rain, severe convective gusts and hail to Hong Kong and the Pearl River Delta (PRD)
region. In summertime, heavy rain is typically associated with southwest monsoon, tropical
cyclones, as well as the wind convergence zone between a low pressure system near Beibu Wan and
a subtropical ridge extending from the western Pacific. Tropical cyclone activities may extend well
into September and October, while land-sea breeze convergence and frontal system may also bring
occasional excessive rain to the area. A more detailed and quantitative account of the climatological
rainfall can be found in [1,2].
Over seven million people reside in this metropolitan with a landmass of just over 1100 km2, of
which 40% has been designated as country parks or natural reserves. The urban areas and satellite
towns where most people reside are very densely populated. With such a setting, even highly
localized downpour could easily affect a significant portion of the population. The potential impact
of rainstorm is complicated owing to the fact that over 70% of the landmass in Hong Kong is of hilly
terrain. Numerous buildings that are built on or next to man-made or natural slopes run into higher
risk brought by landslides. There is no dearth of events in which heavy rain and the associated
weather have ignited a chain reaction of flooding, landslides, infrastructure failures, fallen trees,
Atmosphere 2017, 8, 48 2 of 20
traffic congestions, accidents and the like that gravely affected livelihood of the general public. In
1972, in torrential rain a slope in Hong Kong failed, claiming 67 lives, injuring 20 people and
destroying three buildings nearby [3].
Figure 1. A map of Hong Kong and its vicinity.
1.2. Rainfall Nowcast
To mitigate the potential impacts of rainstorms, the Hong Kong Observatory (HKO) introduced
a color-coded rainstorm warning system in 1992, and subsequently revised it to the present
three-tier form (amber, red and black signals) in 1998 [4]. The amber signal gives alert on potential
heavy rain (30 mm in an hour generally over HK) that may develop into a red signal (50 mm in an
hour) or a black signal (70 mm in an hour) situation. Key Government departments and major utility
companies are put on alert. The red and black signals warn the public of heavy rain which is likely to
bring about serious road flooding and traffic congestion. They will trigger response actions by
Government departments and major transport operators and utility companies. Details of this
rainstorm warning system can be found in [5].
To support the operation of the rainstorm warning system, a radar-based nowcasting system
named “Short-range Warnings of Intense Rainstorms in Localized Systems” (SWIRLS) was
developed in 1999 and has since been in operational use. A brief history of and an introduction to
SWIRLS are given in [6].
Other meteorological services and research institutes around the world have also developed
similar observationally based nowcasting systems, some of which with outputs from numerical
weather prediction (NWP) models blended into nowcast. Notable operational systems in active
service include the followings:
a) Integrated Nowcasting through Comprehensive Analysis (INCA) by Zentralanstalt für
Meteorologie und Geodynamik (ZAMG);
b) Nimrod by Met Office, United Kingdom (UKMO);
c) Spectral Prognosis (S-PROG) by (Australian) Bureau of Meteorology (BoM);
d) Short-Term Ensemble Prediction System (STEPS) by UKMO/BoM;
e) Auto Nowcasting System (ANC) by National Center for Atmospheric Research (NCAR);
f) McGill Algorithm for Precipitation Nowcasting Using Semi-Lagrangian Extrapolation
(MAPLE) by (Canadian) McGill University etc.
A comprehensive review of nowcasting systems was conducted by Royal Meteorological
Institute of Belgium back in 2008 [7], which includes verification results for those available.
As meteorological services normally run only one nowcasting system in operation, cross system
comparison is relatively rare. Under the auspices of World Meteorological Organization (WMO)
World Weather Research Programme (WWRP), Forecast Demonstration Projects (FDP) were held as
parts of the Olympic games in Sydney 2000 (S2000) and in Beijing 2008 (B08FDP), which offered
unique opportunities to demonstrate the benefits and quantify the benefits of a real-time nowcast
service. Keenan et al. [8,9] documents the experience gained in S2000, with verification results given
in [10]. Wang et al. [11] reports on B08FDP with verification included.
Atmosphere 2017, 8, 48 3 of 20
SWIRLS is one of the first operational nowcasting systems with urban-scale applications. It
continues to be updated with innovative approaches, such as optical flow, from time to time. There
are other nowcasting systems that have experimented with or adopted optical flow: Peura and Hohti
[12] demonstrates the potential use of optical flow on nowcasting, while Bowler et al. [13] documents
a nowcasting algorithm based primarily on Horn and Schunck algorithm. SWIRLS currently adopts
“ROVER”, an algorithm to be elaborated in this paper that combines the global Horn and Schunck
approach and the Lucas–Kanade method. To the best of the knowledge of the authors, it is the first if
not the only nowcasting system doing so. Due to sheer amount of data collected over the years,
SWIRLS is also amongst the few nowcasting systems that have undergone long-term verification
with publishable results.
This paper elaborates on the correlation-based algorithm and two optical-flow-based
algorithms adopted for radar echo tracking in Section 2, and examines their performances through
case studies as well as seasonal and multi-year systematic verifications in Section 3. Section 4
discusses the limitations of these tracking algorithms and highlights certain development work
being made, while Section 5 concludes the paper.
2. Operational Experiment Setup
2.1. Nowcasting System
The main function of SWIRLS is to provide quantitative precipitation estimation (QPE) and
quantitative precipitation forecast (QPF) to support rainstorm warning services. QPE is computed
based on radar and rain gauge data analyzed through a variety of techniques including Barnes
analysis, co-kriging analysis and a blending of both. The application of co-kriging analysis with
performance assessment is given at [14]. QPF is computed using an algorithm based on backward
semi-Lagrangian advection scheme. For both QPE and QPF, SWIRLS employs the Z = aRb (Z − R)
relationship with the parameters a and b dynamically calibrated by comparing radar reflectivity with
rain gauge measurements in near real-time. The detailed implementation is elaborated in [15].
Quality control on QPE is implemented as described in [16].
SWIRLS uses data primarily from the weather radar in Tai Mo Shan (TMS), which is sited on a
hilltop at 968 m above mean sea level with unobstructed view in almost all azimuths. The radar is an
S-band radar operating at a frequency of 2.82 GHz, pulse widths of 1.0 and 2.0 µs, and antenna beam
width of 0.9°. Another radar, at about 583 m above mean sea level on top of another hilltip named
Tate’s Cairn, serves as a backup. The Tate’s Cairn radar has a transmitter frequency of 2.92 GHz,
pulse widths of 1.0 and 2.0 µs, and antenna beam width of slightly less than 1.0°. Basic radar data
processing, including removal of ground clutters, are handled by the IRIS software [17]. Both radars
complete a volume scan every 6 min. Radar data and imagery are generated based on a composite of
the two radars with Tai Mo Shan radar being the primary source. SWIRLS is configured to process
radar data immediately upon reception of radar data, and thus also runs once every 6 min.
The typical operational products of SWIRLS are based on 2 km CAPPI (Constant Altitude Plan
Projection Indicator) radar data covering a domain of 256 km in radius, embedded in a rectangular
grid of 480 × 480 pixels. Each pixel represents an area of approximately 1.1 km × 1.1 km. The output
QPF products maintain a grid of 480 × 480 pixels but the spacing between grids is halved, so the
output domain represents a rectangular area of 256 km × 256 km. The configuration is illustrated in
Figure 2.
Atmosphere 2017, 8, 48 4 of 20
Figure 2. The circle depicts the 256 km range of a radar scan, while the inner gray square shows the
product output domain. The crosshair denotes the location of Tai Mo Shan (TMS) radar.
SWIRLS’s products are generated to aid weather forecasters in the preparation of weather
forecasts and the operation of the rainstorm warning system. These products include hourly forecast
rainfall map up to 6 h, a time series showing an ensemble of estimated rainfall in the next 6 h, a plot
of estimated number of rain gauges reaching user-selected thresholds, etc. (Figure 3).
Figure 3. An integrated product of SWIRLS to support rainstorm warning system.
QPF from SWIRLS is also used to support the operation of the landslip warning in Hong Kong
[18]. In addition, SWIRLS is also capable of generating forecasts for thunderstorms [19], squalls [20]
and hail. Areas with potential threats are identified through a combination of statistical means and
physical modeling. These weather phenomena are extrapolated along the motion field for up to half
an hour to generate forecasts.
In recent years, SWIRLS was extended to provide enhanced nowcasting services to the public.
For instance, an Internet web page named “Two-hour Rainfall Nowcast for Hong Kong and Pearl
River Delta” [21] that provides animation of the evolution of radar echoes for the next two hours has
been in operational use since 2007. In late 2012, a location-specific rainfall nowcast service delivered
through mobile platform was also launched, as elaborated in [22–24].
Atmosphere 2017, 8, 48 5 of 20
Since 2008, SWIRLS had been deployed to several international events, including the Beijing
Olympic Games in 2008 [25], the Shanghai Expo 2010 [26], the India Commonwealth Games in 2010
[27], and the Shenzhen Universiade 2013. As part of regional and international co-operations,
SWIRLS has been provided to meteorological services in Mainland China, India, Macao, the
Philippines, Vietnam and Thailand for research and development in radar nowcasting techniques. In
2015, a community version of SWIRLS was developed and made available to interested National
Meteorological and Hydrological Services. Malaysian Meteorological Department and South Africa
Weather Service have successfully adopted SWIRLS for nationwide operational use and for aviation
nowcast research, respectively. Zhuhai Meteorological Bureau is also using SWIRLS operationally.
2.2. Radar Echo Tracking Algorithms
2.2.1. Tracking of Radar Echoes by Correlation (TREC)
In the first operational version of SWIRLS, the motion field of radar echoes was retrieved by a
correlation-based technique named as “Tracking Radar Echoes by Correlation (TREC)” that is
described in [15,28]. This method essentially compares two successive radar reflectivity images and
from which identifies a motion vector for each divided block of pixels through maximizing the
correlation coefficient R, defined as follows:
1
() () () ()
12 1 2
22
22
() ()
1122
Z
kZk Zk Z k
N
kkk
R
Z
kNZ ZkNZ
kk
(1)
where Z1 and Z2 are the array of pixels of reflectivity in the first field and in the second field
respectively, and N denotes the number of data points within an array.
In the implementation of SWIRLS, 93 × 93 overlapping blocks, each of size 19 × 19 pixels (i.e.,
N = 19 × 19) with the centers of two successive blocks separated by five pixels, are extracted from the
480 × 480 pixel reflectivity field, with each pixel representing an area of approximately 1.1 km × 1.1
km. For each block at time T, the correlation coefficients between itself and every other block within
a searching radius of 19 pixels recorded 6 min earlier (i.e., T − 6 min) are computed. The block at time
(T − 6 min) that yields the highest correlation coefficient with the block at time T is chosen, and the
distance between the centers of these two blocks is then used to calculate the motion vector that
represent the motion of the block at (T − 6 min). Repeating the procedure for each block, a motion
field is obtained for all blocks with echoes. Cressman analysis [29] is subsequently performed to
assign motion vectors to those blocks without echoes.
While TREC is successful in tracking motions of individual radar echoes, in practice it generally
captures the direction of individual rain cell instead of larger scale motion of weather system, e.g., a
squall line containing the rain cell in it. As a result, it occasionally produces a faster movement of
convective system such as squall lines in southern China in spring, leading to weaker intensity in
QPF. In fact, in typical implementation of TREC, the block size and resolution are pre-determined,
making it impractical to track motion of variable scales. To address the issues, there emerged several
alternative approaches, some are also correlation-based such as “COTREC” in [30] and “MTREC” in [31]
while others adopt variational optical flow technique, to be described in the following paragraphs.
2.2.2. Multi-Scale Optical-Flow by Variational Analysis (MOVA)
To overcome the limitation of TREC, another radar echo tracking method utilizing variational
optical flow technique, named as “Multi-scale Optical-flow by Variational Analysis (MOVA)”, was
developed in 2009 [32], then replacing TREC for operational use in SWIRLS in 2010.
Defined as the apparent motion of brightness patterns in an image, optical flow would ideally
equalize motion field barring lighting changes. Assuming that the brightness of a point of an object
would remain unchanged along its path (i.e., “brightness constancy assumption”), the optical flow
can be derived by solving the optical flow constraint (OFC) below:
Atmosphere 2017, 8, 48 6 of 20
0
III
uv
txy
(2)
where I represents the brightness of a point projected on a two-dimensional plane.
It can be seen that with two unknowns but just one equation, at least one more constraint is
required to arrive at a unique solution. There are several approaches to tackle the issue. One of them
is proposed by [33], which applies variational method to solving OFC by minimizing the following
cost function:
oHS
JJ J
(3)
where
2
oIII
J
u v dxdy
txy
(4)
and JHS is a global constraint on the smoothness of the gradient of the optical flow field:
22
HS
J
u v dxdy
(5)
By assuming Lagrangian persistence for echoes over their paths, Germann and Zawadzki [34]
adopts a proposal by Wahba and Wendelberger [35] and developed an algorithm named as the
variational echo tracking (VET) [36], using a smoothness penalty function:
22 2
22 2
2
WW 22
22 2
22 2
2
22
uu u
J
dxdy
xy
xy
vv v
dxdy
xy
xy
(6)
In MOVA, seven cascading levels of various box sizes that range from 256 km down to 1.6 km
are used together with the smoothness constraint. The smoothness constants for all level is
controlled by seven parameters tuned for optimal performance, as given in Table 1:
Table 1. Parameters adopted in MOVA
No. of Boxes 1 5 10 20 40 80 160
Box Size (km) 256 51.2 25.6 12.8 6.4 3.2 1.6
Smoothing constant γ 0.0005 0.01 0.1 0.1 1 50 100
When calculating the cost function, three consecutive radar reflectivity images are used to
improve continuity of the retrieved motion field, i.e.,
( to )( to )
o o 1 2 min 6 min o 6 mi n 0
J
JT T JT T (7)
2.2.3. Real-Time Optical Flow by Variational Methods for Echoes of Radar (ROVER)
MOVA was put into operational use in 2010. As it will be shown, MOVA achieved improved
overall performance over TREC, though it has also been observed in operation that MOVA has a
tendency to underestimate the speed of echo motion vectors, in particular at large scale, due to
sub-optimal retrieval of echo motion in the minimization procedure. As a result, excessive rainfall
was forecast at times. With a view to addressing the issue, another variational optical flow scheme,
named as “Real-time Optical flow by Variational methods for Echoes of Radar (ROVER)”, was used
instead. In brief, ROVER is similar to MOVA but with two major enhancements: (1) having a
pre-processing step to spatially smooth the radar reflectivity images; and (2) adoption of a variant
type of optical flow technique.
Atmosphere 2017, 8, 48 7 of 20
In convective systems, which dominate precipitation types in summer monsoon in Hong Kong,
rain echoes could be so jumpy in reflectivity that their partial derivatives can hardly be accurately
calculated. Stable estimates of the derivatives are only possible after the radar rain-rate fields are
highlighted. In ROVER, the reflectivity fields are transformed with the following function:
1
()tan
Z
Zc
GZ
(8)
where Zc and ξ control the point of inflection and its sharpness.
In SWIRLS, Zc and ξ are empirically determined to be 33 dBZ and 4 dBZ, respectively [37] based
on a number of rainstorm cases in spring and summer. The transformation effectively enhances the
contrast for reflectivity values in the range of 20–40 dBZ through suppressing the most and the least
intense parts of rain bands.
The other major enhancement is the adoption of the algorithm proposed in [38,39] in lieu of
VET. This algorithm combines the global Horn and Schunck approach and the Lucas–Kanade
method [40], thus taking the advantages of both the dense flow fields of the former and the noise
robustness of the latter. Through adopting advanced numerical methods, the algorithm further
speeds up the computation such that it can be employed for real-time operational use. An open
source project named “VarFlow” provides the source codes used to implement the algorithm [41].
The algorithm requires users to specify six parameters to control the Gaussian convolutions that
spatially smooth radar images and motion fields to ensure tracking of the most critical features and
to reduce jumpiness, to regularize smoothness constraints, to define the finest and coarsest levels,
and to set the time intervals. The meanings of the parameters and the empirically determined
parameters in ROVER are listed in Table 2.
Table 2. Optical flow parameters adopted in ROVER.
Parameter Significance Value in ROVER
σ Gaussian convolution for image smoothing 9
ρ Gaussian convolution for local vector field smoothing 1.5
α Regularization parameters in the energy function 2000
Lf The finest spatial scale 1 pixel
Lc The coarsest spatial scale 7 pixels
Tr The time interval for tracking radar echoes 6 min
As the choice of parameters would affect the quality of the motion field, the performance of the
variational optical flow algorithm naturally depends on the parameters adopted. While it might be
possible to derive a set of optimal parameters for a particular geographical setting, such a work
would require a rich data archive and tremendous computational power to experiment with and
explore all the dimensional space. It renders optimisation of parameters difficult if not impractical in
practice. In our four years of operational experience, this set of parameters worked reasonably well
in most of the time. Readers interested in implementing ROVER may consider adopting this set of
parameters for initial trial, and fine tune the parameters as and when situation warrants.
3. Results
3.1. Case Analyses
The performance of the three radar echo tracking algorithms are compared using four
rainstorms in Hong Kong during 2013 and 2014. The latter two involved torrential rain with rainfall
intensity exceeding 70 mm·h−1 generally over Hong Kong.
In the following discussion, the motion fields computed with the three radar echo motion
tracking algorithms together with forecast reflectivity images are shown in order. The actual radar
reflectivity images are provided at the bottom row for comparison.
Atmosphere 2017, 8, 48 8 of 20
3.1.1. Rainstorm on 5 April 2013
The first case is illustrated in Figure 4. In the morning of 5 April 2013, a surface trough lingered
over the southern coast of China. Coupled with a moist and active westerly jet in the lower
troposphere, a squall line located to the west of Hong Kong steadily approached the territory at
around 07:00 HKT (Hong Kong Time; HKT = UTC + 8 h). While the individual echo tracked largely
northeastward along the squall line, the squall line itself traversed primarily eastward towards
Hong Kong. The correlation-based algorithm TREC captured the motion of individual echo well but
not that of the squall line, the latter is what is required to project accurately the future rainfall
condition. As the motion vectors computed by TREC pointed to the northeast, after two hours of
extrapolation the squall line was predicted to move to Shenzhen (Figure 4) to the north of Hong
Kong. There was hardly any echo predicted within the territory of Hong Kong, and the
corresponding QPF was rather inaccurate as a result.
Suggested from its name, MOVA is capable of capturing larger scale motion. The direction of
the motion vectors was correct in terms of predicting eastward movement, but the speed was a bit
slow. Consequently, it predicted that the squall line would be way to the west of Hong Kong.
In comparison, ROVER managed to capture both the direction and speed of motion of the
squall line relatively well, and was successful in predicting that the squall line would have reached
the territory in two hours, though the squall line still lagged behind the actual slightly. It
demonstrated a significant improvement of ROVER over both TREC and MOVA.
(a) (b)
(c) (d)
(e) (f)
Atmosphere 2017, 8, 48 9 of 20
(g) (h)
Figure 4. Reflectivity and motion fields based at 07:00 (HKT/UTC + 8) on 5 April 2013 (base time). (a)
Reflectivity and motion field of TREC at base time; (b) Forecast reflectivity of TREC at 2 h from the
base time; (c) Reflectivity and motion field of MOVA at base time; (d) Forecast reflectivity of MOVA
at 2 h from the base time; (e) Reflectivity and motion field of ROVER at base time; (f) Forecast
reflectivity of ROVER at 2 h from the base time; (g) Actual radar reflectivity at base time; and (h)
Actual radar reflectivity at 2 h from the base time.
3.1.2. Rainstorm on 26 July 2013
On 26 July 2013, active southerlies in a broad trough of low pressure affected the south China
coast. At 07:30 HKT, a line of nearly north–south oriented intense echoes moved across Hong Kong
from the south.
As shown in Figure 5, unlike the previous case the motion of individual echoes and that of the
convective cell were largely aligned towards the predominately northward direction. Both TREC and
ROVER predicted that the line of intense echoes would move to the north of Hong Kong, which was
largely correct in direction but slightly slower in terms of speed. MOVA exhibited even slower speed of
movement and as a result incorrectly forecast that the line of echoes would still affect Hong Kong.
(a) (b)
(c) (d)
Atmosphere 2017, 8, 48 10 of 20
(e) (f)
(g) (h)
Figure 5. Reflectivity and motion fields based at 07:30 (HKT/UTC + 8) on 26 July 2013 (base time). (a)
Reflectivity and motion field of TREC at base time; (b) Forecast reflectivity of TREC at 2 h from the
base time; (c) Reflectivity and motion field of MOVA at base time; (d) Forecast reflectivity of MOVA
at 2 h from the base time; (e) Reflectivity and motion field of ROVER at base time; (f) Forecast
reflectivity of ROVER at 2 h from the base time; (g) Actual radar reflectivity at base time; and (h)
Actual radar reflectivity at 2 h from the base time.
As it turned out, two other patches (marked A and B) of intense echoes came from the south,
one of which traceable to a relatively small patch of echoes about 120 km away (marked A), bringing
rain to the Hong Kong Islands and the eastern part of the New Territories by 09:30 HKT. This
underlines a major limitation of nowcasting system that works by extrapolating radar echoes, in
dealing with growth and decay of storms. No radar echo tracking algorithm could have produced a
perfect forecast unless the storm cell remains unchanged in shape, size and intensity, which is
usually not the case.
3.1.3. Rainstorm on 22 May 2013
Starting from the small hours of 22 May 2013, a band of heavy rain spread from west to east
across the Pearl River Estuary. Weather in Hong Kong deteriorated with torrential rain and intense
thunderstorms. More than 150 mm of rain were recorded in many places over the territory. A few
districts even recorded over 200 mm. The heavy downpour caused a number of landslides and
flooding in Hong Kong.
In this heavy rainfall case, we take a look at one hour forecast. As shown in Figure 6, the motion
field and reflectivity field from TREC and ROVER are quite similar, and both correctly predicted
heavy rain over the whole territory. MOVA again exhibited a slow bias and as a result forecast
intense rainfall covering the western part of the territory only.
Atmosphere 2017, 8, 48 11 of 20
(a) (b)
(c) (d)
(e) (f)
(g) (h)
Figure 6. Reflectivity and motion fields based at 02:00 (HKT/UTC + 8) on 22 May 2013 (base time). (a)
Reflectivity and motion field of TREC at base time; (b) Forecast reflectivity of TREC at 1 h from the
base time; (c) Reflectivity and motion field of MOVA at base time; (d) Forecast reflectivity of MOVA
at 1 h from the base time; (e) Reflectivity and motion field of ROVER at base time; (f) Forecast
reflectivity of ROVER at 1 h from the base time; (g) Actual radar reflectivity at base time; and (h)
Actual radar reflectivity at 1 h from the base time.
3.1.4. Rainstorm on 30 March 2014
An episode of heavy rain came on 30 March 2014. Over the course of 3 to 4 h that evening, more
than 100 mm of rainfall were recorded generally in the territory and in certain districts precipitation
even exceeded 150 mm. The rainstorm was accompanied by intense thunderstorms, widespread hail,
severe squalls and floods, resulting in injuries and inflicting critical damages to a crucial
cross-border railway link under construction.
Atmosphere 2017, 8, 48 12 of 20
Figure 7 shows a highly organized northeast-southwest oriented line of echoes edging towards
the territory from the west at 1900 HKT. In this case, TREC tracked the smaller scale motion of
echoes and gave a predominately northeastward motion field. The intense echoes were predicted to
move to Shenzhen to the north of Hong Kong, hence grossly underestimating the precipitation in
Hong Kong. MOVA generated a motion field that was relatively more accurate in direction but was
again slow in speed, resulting in a rain band lying over the western part of the territory and still
failed to capture the heaviest rainfall areas. ROVER was, amongst the three, the most accurate
algorithm in this case.
(a) (b)
(c) (d)
(e) (f)
(g) (h)
Figure 7. Reflectivity and motion field based at 19:00 (HKT/UTC + 8) on 30 March 2014 (base time). (a)
Reflectivity and motion field of TREC at base time; (b) Forecast reflectivity of TREC at 1 h from the
base time; (c) Reflectivity and motion field of MOVA at base time; (d) Forecast reflectivity of MOVA
at 1 h from the base time; (e) Reflectivity and motion field of ROVER at base time; (f) Forecast
reflectivity of ROVER at 1 h from the base time; (g) Actual radar reflectivity at base time; and (h)
Actual radar reflectivity at 1 h from the base time.
Atmosphere 2017, 8, 48 13 of 20
3.2. Comparison of Tracking Algorithms
The performance of ROVER for springtime squall lines in Hong Kong is described in [42]. A
system to compare the performances of the three algorithms has been developed for continuous
verification. In this system, the QPF at each grid point is verified against radar QPE.
The ROVER algorithm started operational trial in mid-May 2012. In this verification, data over a
period of two full years, i.e., from 1 June 2012 to 31 May 2014 inclusive, are considered. A total of 2.1
billion data pairs are available for each radar echo tracking algorithm. Each forecast is paired with
the corresponding actual value and the outcome is determined with reference to Table 3.
Table 3. Contingency matrix for systematic verification.
Forecast Observation
Yes No
Yes Hit False alarm
No Miss Correct negative
The performance is then benchmarked in terms of critical success index (CSI), also known as
Threat Score (TS), Probability of Detection (POD) and False Alarm Ratio (FAR). CSI(TS), POD and
FAR are defined as follows:
hit
CSI(TS) = hit + miss + false alarm
(9)
hit
POD = hit + miss
(10)
false alarm
FAR = hit + false alarm
(11)
The intensity thresholds of 0.5 mm·h
−1
, 5 mm·h
−1
and 30 mm·h
−1
, generally representing light,
medium and heavy rainfall, respectively, within the forecast range of 1 to 6 h are examined in the
following discussion. Forecast rainfall is accumulated hourly in the following results. High intensity
rainfall is usually associated with tropical cyclones or convective development in highly unstable
atmosphere, which is typically on smaller scale and thus less predictable than low intensity rainfall.
The performances of the various algorithms at 0.5 mm·h
−1
threshold, a threshold that is adopted
to distinguish between rain and no rain situations, are compared in Figure 8a. ROVER showed
minor improvement over TREC for the first two hours of forecast. The improvement becomes more
prominent beyond the third hour. For forecast hours of 4–6 h, the performance of ROVER and
MOVA is comparable, both outperforming TREC by a significant margin.
(a) (b) (c)
Atmosphere 2017, 8, 48 14 of 20
(d) (e) (f)
(g) (h) (i)
Figure 8. A comparison of the performance of various radar echo tracking algorithms at various
thresholds in 1–6 h forecasts (x-axis). (a) CSI at 0.5 mm·h
−1
threshold; (b) CSI at 5 mm·h
−1
threshold;
(c) CSI at 30 mm·h
−1
threshold; (d) POD at 0.5 mm·h
−1
threshold; (e) POD at 5 mm·h
−1
threshold; (f)
POD at 30 mm·h
−1
threshold; (g) FAR at 0.5 mm·h
−1
threshold; (h) FAR at 5 mm·h
−1
threshold; and (i)
FAR at 30 mm·h
−1
threshold.
At higher rainfall threshold, say, 5 mm·h
−1
, the performance advantage of ROVER is more
obvious even for the first couple of forecast hours, as shown from Figure 8b. Both ROVER and
MOVA performed considerably better than TREC for the forecast range of 4–6 h, for which MOVA
performs slightly better than ROVER.
Figure 8c further compares the performance of the algorithms in predicting rainstorm as
benchmarked against 30 mm·h
−1
rainfall threshold, chosen for it being the threshold for the Amber
rainstorm warning in Hong Kong. ROVER performs better than MOVA, which in turns outperforms
TREC for the first three hour of forecast. Beyond the fourth hour, MOVA performed better than
ROVER and both MOVA and ROVER outperformed TREC, though their respective CSIs are rather low,
suggesting limiting skill of extrapolation-based QPF in predicting heavy rain at these forecast hours.
A seasonal comparison is also conducted. Figure 9a shows the seasonal results for sixth hour
nowcast (i.e., from T + 5 h to T + 6 h) with 0.5 mm·h
−1
threshold. In this case, MOVA outperforms
ROVER and both perform substantially better than TREC for all seasons.
(a) (b)
Atmosphere 2017, 8, 48 15 of 20
(c)
Figure 9. A seasonal comparison of the performance of various radar echo tracking algorithms at: (a)
0.5 mm·h
−1
threshold for 6 h forecast; (b) 5 mm·h
−1
threshold for 2 h forecast; and (c) 30 mm·h
−1
threshold for 1 h forecast.
Figure 9b reveals the performance at 5 mm·h
−1
threshold. Here, the second hour forecasts (i.e.,
from T + 1 h to T + 2 h) are compared due to lower skill at higher threshold that renders the sixth
hour forecast at 5 mm·h
−1
not having much utility in operation. As shown, the performance
advantage of ROVER over the other two algorithms is the most significant in springtime, a season in
which precipitation is dominated by persistent squall lines induced by mid-tropospheric westerly
waves. In summer, showers are often brought by monsoon troughs, active southwest monsoon and
occasionally tropical cyclones. The first two types of precipitation are intensely convective in nature,
often undergoing rapid growth and decay and even back-building occasionally, which goes beyond
the limit of radar-based nowcasting system that works by extrapolating radar echoes. In autumn,
precipitation from tropical cyclones, northeast monsoon and land sea breeze fronts are more
predictable and the skills recovers, though still lower than that in spring. In winter, the occurrence of
5 mm·h
−1
is relatively rare and so is the number of hits, while at the same time there are still
occasional false alarms due to radar noise caused by land clutter, sea clutter, anomalous propagation,
etc., which explains the lower CSI compared with other seasons.
Figure 9c further compares the schemes for first hour nowcast with 30 mm·h
−1
threshold, which
is critical for operating the rainstorm warning in Hong Kong. The result shows that ROVER is more
skillful than MOVA, which in turn outperforms TREC. All schemes achieve the highest CSI in spring
with the advantage of ROVER over the others the most obvious, as explained earlier.
3.3. Operational Verification of ROVER
In view of the superior performance of ROVER, SWIRLS adopted it as the operational echo
tracking algorithm from June 2013. The performance of SWIRLS based on ROVER continues to be
verified operationally, as given below.
The same methodology as described in Section 3.2 is adopted for verification. For the year 2016,
data are up to 30 September. Performance in terms of CSI is shown in Figure 10.
(a) (b)
Atmosphere 2017, 8, 48 16 of 20
Figure 10. Across year comparison of the performance of ROVER at: (a) 0.5 mm·h
−1
; and (b) 5 mm·h
−1
thresholds for forecast range of 1–6 h.
The verification shows that, while performance vary across years, possibly as a result of
different rainfall types and climatological anomalies, the general performances remains largely
unchanged. The CSI for the first hour forecast ranges from 49% to 57% at 0.5 mm·h
−1
threshold, and
remain around 40% at 5 mm·h
−1
threshold. This illustrates the robustness of ROVER in the first hour
of forecast.
Forecast skills are also measured in terms of POD and FAR. The POD and FAR of ROVER are
examined based on four full years of data from 1 October 2012 to 30 September 2016, both dates
inclusive and are given in Figure 11.
(a) (b)
Figure 11. POD and FAR of ROVER at: (a) 0.5 mm·h
−1
; and (b) 5 mm·h
−1
thresholds based on four
years of data.
Generally speaking, a forecast is considered operationally useful when POD exceeds FAR. This
can be said for up to 2 h at 0.5 mm·h
−1
threshold, and slightly longer than 1 h for 5 mm·h
−1
threshold.
Besides, the Heidke Skill Score (HSS) is also used to gauge the measure of skill in forecasts. A
positive HSS indicates that a forecast is better than a random based forecast. The formula for
calculating HSS is given in [43], copied below:
2 ( hit correct negative - miss false alarm)
HSS = 22
miss + false.alarm + 2 hit correct n egativ e + ( m iss + false alarm ) ( hit + correct ne gative )
(12)
As shown in Figure 12, verification with multi-year data shows that ROVER provides skillful
forecasts for thresholds at 0.5 mm·h
−1
and 5 mm·h
−1
. In both cases, the skill decreases with forecast hours.
(a) (b)
Figure 12. HSS of ROVER at: (a) 0.5 mm·h
−1
; and (b) mm·h
−1
thresholds based on four years of data.
Atmosphere 2017, 8, 48 17 of 20
4. Discussion
4.1. Limitations
In the current implementation of SWIRLS, the motion field is generated from the two latest
radar images and is taken to remain the same throughout the six-hour forecast period. In practice,
the motion field could vary according to changes in the synoptic situation, such as passage of
westerly wave or passage of tropical cyclones. Hence, errors occur due to the assumed invariability
of the motion field. Woo et al. [44] describes a method to separate the motion of a tropical cyclone
(TC) in the computation of motion field, and found it capable of better preserving TC rain bands and
achieving improvement in skills. The same concept may also apply to other weather systems, such
as mid-tropospheric troughs, that traverse Hong Kong at known and constant speed. A suitable
blending of numerical weather prediction models may also help mitigate the problem. The STEPS
nowcasting system by UKMO demonstrates the benefits of blending radar-based forecasts with
NWP forecasts [45].
The parameters adopted in ROVER are currently empirically determined. It may be possible to
further improve its performance through optimizing these parameters based on climatological data,
ideally specific to the synoptic pattern of the precipitation, or alternatively dynamically determined
based on the latest parameter fit. An experimental project to generate probabilistic rainfall nowcast
through perturbing optical flow parameters is being pursued.
Despite improvements in the radar echo tracking algorithm, the SWIRLS nowcasting system
still bears the common limitation of radar-based nowcasting system in failing to predict growth and
decay of radar echoes. Nor does SWIRLS take into account the effect orographic enhancements of
precipitation. Mesoscale analyses, application of statistical methods and advanced techniques in
blending with NWP models may prove useful in this development area, as assessed in [46,47]. The
detection of convective initiation, a precursor to the intensification of precipitation, using the next
generation satellite Himawari-8 data, is covered in [48]. The potential to predict convective initiation
and growth based on real-time reanalysis data is demonstrated in [49].
4.2. Developments in Progress
Due to uncertainties inherent in observation, radar-based advection and the limitations
discussed above, the latest development of SWIRLS is put on moving from deterministic to
probabilistic nowcast. A SWIRLS Ensemble Rainfall Nowcast (SERN) based on 36 members with
slightly perturbed optical flow parameters was setup for experimental trial. Preliminary results
given in [50,51] show that such probabilistic rainfall nowcast products are generally reliable but tend
to be over-confident for high intensity thresholds.
Rainfall nowcast really boils down to the prediction of the evolution of radar images given a
sequence of recent past images. A novel approach to tackle the problem with convolutional
Long-Short-Term-Memory (LSTM) network, a machine learning method, was developed and
verified to provide better performance at 0.5 mm·h−1 threshold [52]. Experiments are being conducted
to assess its performance at higher intensity thresholds and its suitability for operational use.
Forecasts are typically verified against gridded QPE or rain gauge records. It is equally, if not
more, important that the final products and guidance provided to forecasters are rigorously verified
as well. This is, however, a challenging task given the much smaller sample and the fact that
rainstorm warnings are subject to random errors. A progressive verification scheme to verify
rainstorm warnings had since been developed. It was employed to verify automatic warning
guidance given by SWIRLS. Based on five years of back-tested data, about 70% of amber rainstorm
guidance warnings automatically generated by SWIRLS were considered effective [53].
5. Conclusions
ROVER, a variational optical flow technique to track radar echo motion, has been employed in
the SWIRLS nowcasting system of HKO. Case analysis and systematic verification with two-year
Atmosphere 2017, 8, 48 18 of 20
data demonstrate its advantages over the correlation-based algorithm TREC and the previous
optical flow algorithm MOVA in general. Its limitations are discussed and potential future
development opportunities are presented.
Acknowledgments: The authors would like to thank Hon-yin Yeung and Tsz-lo Cheng who played a
significant role in the development, as well as Ngo-hin Chan, Chin-ming Lo and Hoi-lam Yeung for assistance
in maintaining and developing various components of the systems. The authors would also like to thank the
three anonymous reviewers, whose comments and suggestions have significantly improved this paper.
Author Contributions: Wang-chun Woo managed the operation and development of SWIRLS in HKO,
undertook the verification, and wrote the majority of this paper. Wai-kin Wong developed MOVA and
managed the overall completion of this project. Both the authors have read and approved the final manuscript.
Conflicts of Interest: The authors declare no conflict of interest.
References
1. Peterson, P. The Rainfall of Hong Kong. 1964. Available online: http://www.hko.gov.hk/
publica/tn/tn017.pdf (accessed on 2 February 2017).
2. Lee, T.C.; Leung, W.; Chan, K. Climatological Normals for Hong Kong, 1971–2000. Available online:
http://www.weather.gov.hk/publica/tnl/tnl083.pdf (accessed on 2 February 2017).
3. Yang, T.L.; Mackey, S.; Cumine, E. Final Report of the Commission of Inquiry into the Rainstorm Disasters
1972, GEO Report No. 229. Available online: http://www.cedd.gov.hk/eng/publications/geo_reports/
doc/er229/er229links.pdf (accessed on 24 February 2017).
4. Ho, P. Modernisation of the City. In Weather the Storm; Hong Kong University Press: Hong Kong, China,
2003; pp. 318–319.
5. Hong Kong Observatory Rainstorm Warning System. Available online: http://www.hko.gov.hk/
wservice/warning/rainstor.htm (accessed on 2 February 2017).
6. Li, P.W.; Wong, W.K.; Cheung, P.; Yeung, H.Y. An overview of nowcasting development, applications,
and services in the Hong Kong Observatory. J. Meteorol. Res. 2014, 28, 859–876.
7. Reyniers, M. Quantitative Precipitation Forecasts Based on Radar Observations: Principles, Algorithms
and Operational Systems. Royal Meteorological Institute of Belgium. Available online: http://www.
meteo.be/meteo/download/fr/3040165/pdf/rmi_scpub-1261.pdf (accessed on 2 February 2017).
8. Keenan, T.; Wilson, J.; Joe, P.; Collier, C.G.; Golding, B.; Burgess, D.; Carbone, R.; Seed, A.; May, P.; Berry,
L.; et al. The World Weather Research Programme (WWRP) Sydney 2000 Forecast Demonstration Project:
Overview. In Proceedings of the 30th International Conference on Radar Meteorology, Munich, Germany,
19–24 July 2001.
9. Keenan, T.; Joe, P.; Wilson, J.; Collier, C.; Golding, B.; Burgess, D.; May, P.; Pierce, C.; Bally, J.; Crook, A.; et
al. The Sydney 2000 World Weather Research Programme Forecast Demonstration Project: Overview and
Current Status. Bull. Am. Meteorol. Soc. 2003, 84, 1041–1054.
10. Ebert, E.E.; Wilson, L.J.; Brown, B.G.; Nurmi, P.; Brooks, H.E.; Bally, J.; Jaeneke, M. Verification of
Nowcasts from the WWRP Sydney 2000 Forecast Demonstration Project. Weather Forecast. 2004, 19, 73–96.
11. Wang, J.; Keenan, T.; Joe, P.; Wilson, J.; Lai, E.S.T.; Liang, F.; Wang, Y.; Ebert, B.; Ye, Q.; Bally, J.; et al.
Overview of the Beijing 2008 Olympics Project. Part I: Forecast Demonstration Project. Available online:
https://www.wmo.int/pages/prog/arep/wwrp/new/documents/B08FDP_finalreport.pdf (accessed on 2
February 2017).
12. Peura, M.; Hohti, H. Optical Flow in Radar Images. Proceedings of ERAD (2004), Vol. 2004 of, Copernicus
GmbH, 454–458. Available online: http://copernicus.org/erad/2004/online/ERAD04_P_454.pdf (accessed
on 9 February 2017).
13. Bowler, N.E.H.; Pierce, C.E.; Seed, A. Development of a precipitation nowcasting algorithm based upon
optical flow techniques. J. Hydrol. 2004, 288, 74–91.
14. Yeung, H.Y.; Man, C.; Chan, S.T.; Seed, A. Application of Radar-Raingauge Co-Kriging to Improve QPE
and Quality Control of Real-time Rainfall Data. Available online: http://www.hko.gov.hk/publica/
reprint/r968.pdf (accessed on 6 February 2017).
15. Li, P.W.; Lai, E.S.T. Short-range quantitative precipitation forecasting in Hong Kong. J. Hydrol. 2004, 288,
189–209.
Atmosphere 2017, 8, 48 19 of 20
16. Li, P.W.; Wong, W.K.; Chan, K.Y.; Lai, E.S.T. SWIRLS—An Evolving Nowcasting System. Hong Kong
Observatory. Available online: http://www.weather.gov.hk/publica/tn/tn100.pdf (accessed on 6 February
2017).
17. User Guide of IRIS Radar. Vaisala Oyj. 2016. Available online: ftp://ftp.sigmet.com/outgoing/
manuals/IRIS_Radar_Manual.pdf (accessed on 9 February 2017).
18. Cheung, P.; Wong, M.; Yeung, H. Application of Rainstorm Nowcast to Real-Time Warning of Landslide
Hazards in Hong Kong. Available online: http://www.hko.gov.hk/publica/reprint/r673.pdf (accessed on 2
February 2017).
19. Yeung, H.Y.; Lai, E.; Chiu, S. Lightning Initiation and Intensity Nowcasting Based on Isothermal Radar
Reflectivity-A Conceptual Model. In Proceedings of the 33rd International Conference on Radar
Meteorology, Cairns, Australia, 6–10 August 2007.
20. Yeung, H.Y.; Lai, S.T.; Chan, K.Y. Thunderstorm Downburst and Radar-Based Nowcasting of Squalls. In
Proceedings of the Fifth European Conference on Radar in Meteorology and Hydrology, Helsinki, Finland,
30 June–4 July 2008.
21. Two-Hour Rainfall Nowcast for Hong Kong and Pearl River Delta. Available online:
http://www.hko.gov.hk/nowcast/prd/api/index_ue.htm (accessed on 2 February 2017).
22. Yeung, H.Y. Recent developments and applications of the SWIRLS nowcasting system in Hong Kong. In
Proceedings of the 3rd WMO International Symposium on Nowcasting and Very Short-Range Forecasting
(WSN12), Rio de Janeiro, Brazil, 6–10 August 2012.
23. Woo, W.C. Location-Based Rainfall Nowcasting Service for Public. In Proceedings of the 2013 European
Geosciences Union General Assembly, Vienna, Austria, 7–12 April 2013.
24. Woo, W.C. Rain forecast on MyObservatory—A location-based rainfall nowcasting service for the public.
In Proceedings of the 14th Workshop on Meteorological Operational Systems, ECMWF, Reading, UK,
18–20 November 2013.
25. Yeung, H.Y.; Wong, W.K.; Chan, K.Y.; Lai, E. Applications of the Hong Kong Observatory nowcasting
system SWIRLS-2 in support of the 2008 Beijing Olympic Games. In Proceedings of the WMO Symposium
on Nowcasting, Whistler, BC, Canada, 30 August–4 September 2009.
26. Cheng, T.L.; Tai, K.; Yeung, H.Y. Application of the nowcasting system SWIRLS in 2010 Shanghai Expo. In
Proceedings of the 28th Annual Conference of Chinese Meteorological Society 2011, Xiamen, China, 2–4
November 2011.
27. Srivastava, K.; Lau, S.; Yeung, H.Y.; Cheng, T.L.; Hardwaj, R.; Kannan, A.; Bhowmik, S.; Singh, H. Use of
SWIRLS nowcasting system for quantitative precipitation forecast using Indian DWR data. MAUSAM Q. J.
Meteorol. Hydrol. Geophys. 2012, 1, 1–16.
28. Tuttle, J.D.; Foote, G.B. Determination of the Boundary Layer Airflow from a Single Doppler Radar. J.
Atmos. Ocean. Technol. 1990, 7, 218–232.
29. Cressman, G.P. An operational objective analysis system. Mon. Weather Rev. 1959, 367–374.
30. Novak, P. The Czech Hydrometeorological Institute’s severe storm nowcasting system. Atmos. Res. 2007,
83, 450–457.
31. Wang, G.L.; Wong, W.K.; Liu, L.P.; Wang, H. Application of multi-scale tracking radar echoes scheme in
quantitative precipitation nowcasting. Adv. Atmos. Sci. 2013, 30, 448–460.
32. Wong, W.K.; Yeung, H.Y.; Wang, Y.C.; Chen, M. Towards the Blending of NWP with
Nowcast—Operation Experience in B08FDP. In Proceedings of the WMO Symposium on Nowcasting,
Whistler, BC, Canada, 30 August–4 September 2009.
33. Horn, B.K.P.; Schunck, B.G. Determining optical flow. Artif. Intell. 1981, 17, 185–203.
34. Germann, U.; Zawadzki, I. Scale-Dependence of the Predictability of Precipitation from Continental Radar
Images. Part I: Description of the Methodology. Mon. Weather Rev. 2002, 130, 2859–2873.
35. Wahba, G.; Wendelberger, J. Some New Mathematical Methods for Variational Objective Analysis Using
Splines and Cross Validation. Mon. Weather Rev. 1980, 108, 1122–1143.
36. Bellon, A.; Zawadzki, I.; Kilambi, A.; Lee, H.C.; Lee, Y.H.; Lee, G. McGill algorithm for precipitation
nowcasting by lagrangian extrapolation (MAPLE) applied to the South Korean radar network. Part I:
Sensitivity studies of the Variational Echo Tracking (VET) technique. Asia-Pac. J. Atmos. Sci. 2010, 46, 369.
37. Cheung, P.; Yeung, H.Y. Application of optical-flow technique to significant convection nowcast for
terminal areas in Hong Kong. In Proceedings of the 3rd WMO International Symposium on Nowcasting
and Very Short-Range Forecasting (WSN12), Rio de Janeiro, Brazil, 6–10 August 2012.
Atmosphere 2017, 8, 48 20 of 20
38. Bruhn, A.; Weickert, J.; Feddern, C.; Kohlberger, T.; Schnörr, C. Real-Time Optic Flow Computation with
Variational Methods. In Computer Analysis of Images and Patterns; Lecture Notes in Computer Science;
Petkov, N., Westenberg, M.A., Eds.; Springer: Berlin/Heidelberg, Germany, 2003; pp. 222–229.
39. Bruhn, A.; Weickert, J.; Feddern, C.; Kohlberger, T.; Schnorr, C. Variational optical flow computation in
real time. IEEE Trans. Image Process. 2005, 14, 608–615.
40. Lucas, B.D.; Kanade, T. An Iterative Image Registration Technique with an Application to Stereo Vision. In
Proceedings of the 7th International Joint Conference on Artificial Intelligence—Volume 2; IJCAI’81; Morgan
Kaufmann Publishers Inc.: San Francisco, CA, USA, 1981; pp. 674–679.
41. Harmat, A. Variational Optical Flow. Available online: http://sourceforge.net/projects/varflow/ (accessed
on 2 February 2017).
42. Woo, W.C. Enhancement of optical flow technique for Nowcasting of springtime squall lines over Hong
Kong. In Proceedings of the Fifth International Workshop on Monsoons, Macao, China, 28 October–1
November 2013.
43. Doswell, C.A.; Davies-Jones, R.; Keller, D.L. On Summary Measures of Skill in Rare Event Forecasting
Based on Contingency Tables. Weather Forecast. 1990, 5, 576–585.
44. Woo, W.C.; Li, K.K.; Bala, M. An Algorithm to Enhance Nowcast of Rainfall Brought by Tropical Cyclones
through Separation of Motions. Trop. Cyclone Res. Rev. 2014, 3, 111–121.
45. Bowler, N.E.; Pierce, C.E.; Seed, A.W. STEPS: A probabilistic precipitation forecasting scheme which
merges an extrapolation nowcast with downscaled NWP. Q. J. R. Meteorol. Soc. 2006, 132, 2127–2155.
46. Liguori, S.; Rico-Ramirez, M.A. Quantitative assessment of short-term rainfall forecasts from radar
nowcasts and MM5 forecasts. Hydrol. Process. 2012, 26, 3842–3857.
47. Schellart, A.; Liguori, S.; Krämer, S.; Saul, A.; Rico-Ramirez, M.A. Comparing quantitative precipitation
forecast methods for prediction of sewer flows in a small urban area. Hydrol. Sci. J. 2014, 59, 1418–1436.
48. Lee, S.; Han, H.; Im, J.; Jang, E.; Lee, M.-I. Detection of deterministic and probabilistic convective initiation
using Himawari-8 Advanced Himawari Imager data. Atmos. Meas. Tech. Discuss. 2016, 1–38,
doi:10.5194/amt-2016-308.
49. Han, L.; Sun, J.; Zhang, W.; Xiu, Y.; Feng, H.; Lin, Y. A Machine Learning Nowcasting Method based on
Real-time Reanalysis Data. arXiv 2016, arXiv:1609.04103.
50. Woo, W.C.; Lok, K.M. Experimental Location Specific Probabilistic Rainfall Nowcast. In Proceedings of the
29th Guangdong-Hong Kong-Macao Seminar on Meteorological Science and Technology, Macao, China,
20–22 January 2015.
51. Woo, W.C.; Lok, K.M.; Wong, W.K. Probabilistic Quantitative Precipitation Forecast for Tropical Cyclone
Rainfall. In Proceedings of the Third International Workshop on Tropical Cyclone Landfall Processes
(IWTCLP-III), Jeju, Korea, 2–10 December 2014.
52. Shi, X.; Chen, Z.; Wang, H.; Yeung, D.-Y.; Wong, W.K.; Woo, W.C. Convolutional LSTM Network: A
Machine Learning Approach for Precipitation Nowcasting. arXiv2015, arXiv:1506.04214.
53. Woo, W.C.; Wong, W.K. Progressive Verification of Rainstorm Warnings. In Proceedings of the WMO
WWRP 4th International Symposium on Nowcasting and Very-short-range Forecast 2016 (WSN16), Hong
Kong, China, 25–29 July 2016.
© 2017 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).
