You are on page 1of 18

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/335809546

Improved rainfall nowcasting using Burgers’ equation

Article  in  Journal of Hydrology · September 2019


DOI: 10.1016/j.jhydrol.2019.124140

CITATIONS READS

9 272

4 authors, including:

Soorok Ryu Geunsu Lyu


Kyungpook National University 5 PUBLICATIONS   17 CITATIONS   
10 PUBLICATIONS   64 CITATIONS   
SEE PROFILE
SEE PROFILE

Gyuwon Lee
Kyungpook National University
122 PUBLICATIONS   1,661 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Meso- to synoptic-scale severe weather analysis and modeling research View project

Warm/Cold precipitation microphysics View project

All content following this page was uploaded by Gyuwon Lee on 10 May 2020.

The user has requested enhancement of the downloaded file.


2019. 9. 16. Improved rainfall nowcasting using Burgers’ equation - ScienceDirect

Journals & Books Create account Sign in

Download Share Export

Journal of Hydrology
Available online 13 September 2019, 124140
In Press, Journal Pre-proof

Research papers

Improved rainfall nowcasting using Burgers’ equation


Soorok Ryu a, Geunsu Lyu a, Younghae Do c, GyuWon Lee a, b
Show more

https://doi.org/10.1016/j.jhydrol.2019.124140 Get rights and content

Abstract
Nowcasting of surface precipitation from radar data typically relies on algorithms that calculate advection, such as the McGill
Algorithm for Precipitation nowcasting by Lagrangian Extrapolation (MAPLE). This method offers high spatial and temporal
resolution it cannot represent the growth-decay of precipitation and non-stationary advection vector fields.

In this study, we propose some nowcasting rainfall models based on advection-diffusion equation with non-stationary motion
vectors. The diffusion term of this equation gives to smoother rainfall predictions for lead times and increased skill scores. The
motion vectors are updated in each time step by solving a system of two-dimensional (2D) Burgers’ equations. The proposed
forecasting models use the following three steps. First, an initial motion vector field is approximated using the Variational Echo
Tracking (VET) algorithm. Second, a forecast is obtained for each time step by solving a time-dependent advection or advection-
diffusion equation. In this step, the motion vectors are updated by solving Burgers’ equation. Lastly, forecasts are evaluated with
lead times from min to 3 h, and forecasts are compared with rain rate observations for six events over a km2 region
in southeastern South Korea.

To observe the effects of the diffusion term and Burgers’ equation, four variants of the proposed modeling methods are
considered, depending on the equations: advection equation (Type 1), advection and Burgers’ equations (Type 2), advection-
diffusion equation (Type 3), and combination of the advection-diffusion and Burgers’ equations (Type 4). The forecasts from the
Type 1 method are very similar to those of MAPLE. The other models (Type 2-4) yielded clearly better skill scores and
correlation on average, with up to 3 hours’ lead time. Models that use Burgers’ equation (Type 2 and Type 4) give much better
scores than other methods using fixed motion vectors when the temporal variation of the motion vectors is large.

Keywords
Nowcasting; MAPLE; Advection-diffusion; Burgers’ equation; Motion vectors

MSC 2010
00-01; 99-00

1. Introduction
Precipitation is extremely difficult to forecast because of the chaotic and non-linear nature of meteorological system. As a result,
numerical weather predictions (NWP) are very sensitive to small perturbations in initial conditions, boundary conditions, and
round-off errors. The most successful nowcasting of precipitation uses weather radar, which can provide good areal coverage
with high resolution. Many radar-based nowcasting techniques have been developed over the years (Bellon and Austin, 1978,
Browning et al., 1982, Dixon and Wiener, 1993, Handwerker, 2002, Seed, 2003, Rasmussen et al., 2003, Mueller et al., 2003,
Turner et al., 2004, Fox and Wikle, 2005, Ruzanski et al., 2011).
https://www.sciencedirect.com/science/article/pii/S0022169419308753?via%3Dihub 1/17
2019. 9. 16. Improved rainfall nowcasting using Burgers’ equation - ScienceDirect

Nowcasting of precipitation fields commonly involves Lagrangian extrapolations of advection mechanisms from radar data
(Mueller et al., 2003, Bowler et al., 2004, Li and Lai, 2004). This method can provide high resolution forecasts using a simpler
algorithm than NWP models. The Lagrangian extrapolation model is formally stated as

(1)

where is the observed precipitation field at the starting time and location , and is the forecast
precipitation field after lead time at and position .

The McGill Algorithm for Precipitation Nowcasting by Lagrangian Extrapolation (MAPLE) was developed by Germann and
Zawadzki (2002). They studied the scale dependence of predictability for spatial resolution of 4 km over a 2720 km 2720 km
domain using Fourier low-pass filtering. By assuming the Lagrangian dynamics:

Eq. (1) is obtained from the conservation equation for R:

(2)

where is the position of precipitation pattern, and is the advection vector that depends on spatial
coordinates alone. Then, the solution of (2) satisfies the persistence model in Eq. (1).

MAPLE uses a variational echo tracking (VET) algorithm that minimizes a cost function to determine the advection vector
(see Section 3). The VET algorithm provides motion vectors at multiple scales and can account for changes in the size
and intensity of precipitation over the forecasting period. However, Lagrangian persistence fails when the source-sink term or
non-stationary terms in the motion field are dominant over the advection term. Hence, better prediction require a model that
accurately predicts the source-sink term and non-stationary features of the motion field (Germann and Zawadzki, 2002).

Prediction of the source-sink term is related to predictions of growth or decay of a precipitation system. Several researchers
have attempted to consider growth and decay terms in forecasting models (Tsonis and Austin, 1981, Mueller et al., 2003,
Radhakrishna et al., 2012, Tang and Matyas, 2018). However, it is found that these are rather difficult task and the results were
not much successful. Most recent research efforts have tried to account for growth and decay of precipitation systems, using
orographic forcing (Foresti et al., 2018) or the diurnal cycle (Atencia et al., 2017). In Bechini and Chandrasekar (2017), Doppler
wind information is used to enhance the extrapolation technique.

Bowler et al. (2006) analyze the relative importance of the evolution of advection fields and concluded that uncertainties due to
the advection fields are less important than uncertainties in the evolution of the precipitation field. This conclusion is likely
only valid if the advection vectors of the precipitation field are stationary. The present study questions whether the nature of
advection vector affects the forecasts (see Section 5.4).

As for the non-stationarity of the motion vectors, which also presents challenges when using the Lagrangian persistence model,
the present research addresses this problem by generating a time-dependent advection vector that is updated at each time step
of the forecasting process. As an alternative to MAPLE, the proposed method solves advection or an advection-diffusion
equations directly without relying on the assumption of Lagrangian persistence (1). The two-dimensional (2D) Burgers’ equation
(see Section 4) and and the advection-diffusion equation are solved simultaneously to resolve non-stationary vector field.

The present paper is organized as follows. Section 2 introduces the rainfall events and data used in this study. Section 3 provides
a brief description of the MAPLE algorithm that was developed by the radar group at McGill university. The forecasting
methods developed in the present study are given in Section 4. Forecasts obtained from proposed methods are compared with
those of MAPLE in Section 5. The final section summarizes our results.

2. Data setup and evaluation methods


Recordings of six rainfall events in a southeastern km2 area of South Korea were selected for analysis (Fig. 1). The data
used in the study are rainfall intensity (R) fields converted from the composite radar reflectivity (Z) using the relationship
. All the computations discussed below deal with fields of rainfall rate with units of mm h -1, and any pixels with
magnitude mm h-1 are considered as not having rain. The computation ranges over pixels with resolution of
km. The initial motion vector is computed using VET with resolution of 10 km over km2, which is the common area
shared by three radars within the domain. In this rectangular region, bilinear interpolation or extrapolation is applied to vectors
when they are used in semi-Lagrangian schemes and our equation-based schemes.

https://www.sciencedirect.com/science/article/pii/S0022169419308753?via%3Dihub 2/17
2019. 9. 16. Improved rainfall nowcasting using Burgers’ equation - ScienceDirect

Download high-res image (641KB) Download full-size image

Fig. 1. Six precipitation events and the corresponding motion vectors estimated using VET algorithm. The verification region is
inside the heavy red line ( km2), which is the common location of the three radars within a km2 domain.

We choose continuous and categorical verification strategies to evaluate the different forecasts. The correlation coefficient and
mean absolute error (MAE) are used as continuous indicators. The correlation coefficient is calculated with the following
equation (Germann and Zawadzki, 2002):

(3)

where is the correlation coefficient at time t, is the number of pixels in the quality mask region, and and are the
observed and forecast rain rate fields (in mm h-1), respectively. The quality mask chooses all pixels with non-negative values at
time t inside the area marked with a heavy red line (250 250 km2) in Fig. 1. The number of pixels n(t) will change over time,
but and , where is the number of pixels in the quality mask at time t of a forecast calculated with
method Type (see Section 4).

To verify the forecasts delivered by the methods we tested, binary rain/no-rain events were classified by applying thresholds. All
pixels with rain-rate values R above a certain threshold were considered rainy, and all other pixels were considered to be
without rain. For each tested lead time, we applied this threshold to the forecast and observed the rain fields in the quality
mask. Using these binary forecasting images, the following symbols are defined to describe the number of forecast pixels falling
into each class: hits, misses, false alarms, correct negatives.

We also selected skill scores for evaluating the forecasting methods, including the probability of detection (POD), the false-
alarm rate (FAR), the critical success index (CSI), and the equitable threat (or Gilbert) score (ETS). They are defined using the
counters , and d as POD , FAR , CSI , and ETS , where
. For continuous verification indices, the mean absolute error (MAE) is defined as follows:

where is the number of pixels in the quality mask at time t.

3. Overview of MAPLE algorithm


Before introducing our methods, we will briefly summarize the standard MAPLE algorithm. MAPLE’s two main steps are
finding the motion vector field using VET and extrapolating radar echoes using constant a motion vector in a semi-Lagrangian

https://www.sciencedirect.com/science/article/pii/S0022169419308753?via%3Dihub 3/17
2019. 9. 16. Improved rainfall nowcasting using Burgers’ equation - ScienceDirect

scheme. Note that the our schemes and MAPLE both use the VET algorithm to determine the initial motion vector.

3.1. Variational echo tracking


The VET algorithm (Germann and Zawadzki, 2002) is used to obtain an initial motion vector from rain rate data. This technique
minimize the following cost function

(4)

The first term in (4) is the sum of the squares of the residuals of conservation Eq. (2), and is calculated using an upstream
semi-Lagrangian scheme:

where is a weighting factor that presents the quality of the data. The second term in (4) is a smoothness penalty function:

where is a constant weight. The conjugate gradient algorithm is used to find the global minimization solution of the cost
function (4), . To avoid convergence toward secondary minima, a scaling-guess procedure is used. We adopt the motion
vector fields starting with a coarse resolution of one vector (in ), then move to a resolution of vectors, and
finally vectors (spatial resolution of ).

3.2. Semi-Lagrangian scheme


We employ the semi-Lagrangian backward scheme in Germann and Zawadzki (2002) to generate forecast. The semi-Lagrangian
scheme is obviously preferable over a constant vector scheme since it allows for rotation and nearly conserves mass

The lead time is divided into N time steps of length , with , and is iteratively calculated as

where the motion vector is obtained from the VET algorithm at time . In this scheme, the origins of precipitation
echoes follow upstream streamlines, and will end up at a particular grid point in the forecast field. With longer lead times,
however, the origin may not be located at the grid point in the current rain field, or it may be outside the domain. In the former
case, bilinear interpolation is used to force the origin onto a grid point. In the latter case, we impose a a negative value on the
grid point and label the pixel as a no data point. Recall that the quality mask region is the common area inside km2
without any negative valued pixels for all lead times. Verification scores were computed with data taken from only this quality
mask region.

4. Forecast models based on partial differential equations (PDEs)


The four types of nowcasting methods that we have developed are based on time-dependent PDEs. The VET algorithm is used
to find the initial motion vector, and then the motion vector can be used in calculations of four types of PDEs: advection
equation (Type 1), advection equation and Burgers’ equation (Type 2), advection-diffusion equation (Type 3), and advection-
diffusion equation and Burgers’ equation (Type 4). The four methods share the following common steps:
1. Find the motion vector field at by VET

2. Solve time-dependent PDEs to derive forecasts

3. Compare the forecasts with observations

A more detailed explanation of four types of forecasts follows.

4.1. Type 1 - Advection equation


The first method is based on the simple advection equation. Rainfall intensities are advected by motion vector , which
is stationary in time. Instead of the Lagrangian persistence relation (1), the following PDE is solved in Type 1 with the initial
observation field and the initial vector generated by the VET algorithm at time .

(5)

To approximate the partial derivative terms (or ), the centered difference formula (Wendroff, 1968),

https://www.sciencedirect.com/science/article/pii/S0022169419308753?via%3Dihub 4/17
2019. 9. 16. Improved rainfall nowcasting using Burgers’ equation - ScienceDirect

is used in the computation domain and the conditions and are assumed for the
boundary pixels, where km.

The Type 1 method gives forecasts that are similar to those of MAPLE’s semi-Lagrangian extrapolation. This result is expected
because the semi-Lagrangian persistence Eq. (1) is the analytical solution of the advection Eq. (5) used in the Type 1 forecast.
Building on the Type 1 equation, precipitation models can be updated by adding a diffusion term or coupling the equation to
other PDEs. The Type 2, 3, and 4 forecasting schemes represent the main results of this study.

4.2. Type 2 - Moving motion vectors using Burgers’ equation


MAPLE and Type 1 assume that the motion vector field is stationary for lead times. However, real motion vector fields will
change over time. We assume that the motion vector field can be advected itself and changes over time and follows a physically
meaningful equations, including diffusion terms. That is, we assume that the motion vector field follows a fluid equation, such
as the 2D Burgers’ equation:

(6)

where, is a kinematic viscosity.

Burgers’ equation is one of the fundamental equation in fluid mechanics involving both nonlinear propagation and diffusion
effects with the pressure term removed in Navier-Stokes (N-S) equations. 2D Burgers’ equation is used for natural application
such as modeling of gas mechanics (Brio and Hunter, 1992) and shallow water wave investigation (Bartuccelli et al., 1985).

In meteorology, Burgers’ equation is primarily used for data assimilation. For instance, the Kalman filter from Burgers’
equation was used in Menard (1994) to investigate problems related to atmospheric data assimilation and meteorological
forecasting. In Stahl and Morten Aamo (2011), this equation is also used as a constraint equation to obtain the desired image
vector from image sequences for image motion estimation. This process is similar to estimate motion vector to generate
forecasting rain fields from observed precipitation images. Although this equation is not perfect for modeling the motion of
precipitation fields, but it is the simplest equation that includes the minimum conditions (advected itself, changes over time
and includes diffusion term) we want to assume.

We apply the Eq. (6) to the initial vector obtained from the VET algorithm to update the motion vector at each time
step. From the Eq. (6), the motion vector can be updated at each time , with km2
resolution (25 25 vectors). The motion vector updated at each time is interpolated and extrapolated, and is then
substituted into the motion vector term of Eq. (5) to generate the forecasting field .

The same first derivative formula from the Type 1 forecast is used and for the second derivative term in (6), the centered
difference quotient (Wendroff, 1968) is used:

where and km.

The constant coefficient in Eq. (6) is related to the smoothness of the motion vector field : the larger the value of , the
smoother the motion vector . Here, is used as a fixed value such as ( 3.3 m s-1) and is determined empirically.

4.3. Type 3 - Advection diffusion equation


In nature, fluid transport is governed by a combination of advection and diffusion. The Type 1 forecast assumes that only
advection acts on a rain cell. In Jinno et al. (1993) and Kawamura et al. (1997), they showed that 2D advection-diffusion equation
is able to describe the observed rainfall intensity at ground level. This model assumes that over a short time period, the main
forces driving a rain cell are advection and diffusion. In the Type 3 scheme, diffusion term is added as the main forcing term in
Eq. (5). If the flow has a constant diffusion coefficient and is incompressible without sink or source, the equation can be
simplified to:

(7)

https://www.sciencedirect.com/science/article/pii/S0022169419308753?via%3Dihub 5/17
2019. 9. 16. Improved rainfall nowcasting using Burgers’ equation - ScienceDirect

where is the diffusion coefficient or mass diffusivity of particle motion. In rainfall forecasts, this term plays a filtering or a
smoothing role for longer forecast lead times. In this work, the value is varied in the ranges ( 1.67 ms-1). When
, the Type 3 scheme is the same as Type 1, but if , the smoothing effect will increase some of the skill scores.

4.4. Type 4 - Advection diffusion equation with Burgers’ equation


The last scheme is combination of the Type 2 and Type 3 schemes. The non-zero diffusion coefficients and smooth the rain
intensity field and motion vector , respectively. The Type 4 scheme solves Eq. (7) using the updated motion
vector from Eq. (6). The motion vectors in the Type 4 scheme are the same as those in Type 2. That is, in the Type 4
scheme, forecasting results are smoothed by the diffusion term while using the vector updated by Burgers’ equation.

4.5. Time steps for numerical computations


Time dependent PDEs are solved with the explicit Runge-Kutta order 4 (RK4) method and the finite difference method (FDM)
throughout this paper. Note that, according to the Courant-Friedrichs-Lewy (CFL) condition (Courant et al., 1967), when solving
a PDE using an explicit time-integration scheme, one must select time steps that are small enough relative to the magnitude
of the motion vector , the space intervals , and the viscosity of the fluid. To ensure the convergence of numerical
methods for PDEs, we use “min” and “km” as units of time and distance, respectively. If no diffusion term is included in
the model Eq. (7), explicit iteration method will converge only for for (min). However, smaller is desired for lager ,
so we use (min) with the range ( ms-1) for all methods based on PDEs.

When using the time iterative algorithm (RK4), the time intervals do not need to be the same when forecasting the rain rate
R (in Eqs. (5) or (7)) and motion vector (in Eq. (6)), since the spatial resolution of km and 10 km are different. We use
different updating time steps of (min) and (min) for the precipitation field R and motion vector , respectively.

5. Forecast evaluation results

5.1. Examples of nowcasting


Fig. 2 shows forecasts generated by MAPLE, Type 1 and Type 3. The Type 1 forecasts are almost the identical to those of MAPLE.
In the Type 3 results, the shape of forecast is similar to that of Type 1, and only the smoothing effect is added.

Download high-res image (509KB) Download full-size image

https://www.sciencedirect.com/science/article/pii/S0022169419308753?via%3Dihub 6/17
2019. 9. 16. Improved rainfall nowcasting using Burgers’ equation - ScienceDirect
Fig. 2. Observation (top) and forecasts obtained by MAPLE, Type 1, and Type 3 for lead times of 1, 2, and 3 h. The starting time is
0300 LST 30 Jun 2012. For the Type 3 scheme, the diffusion coefficient, ( ms-1) is used.

The three vectors in the first row of Fig. 3 are those obtained from the VET algorithm using the three consecutive rain rate
maps , and with time step min. The bottom motion vectors are predicted by
Burgers’ Eq. (6) using the starting vector at 0300 (top left in Fig. 3) for diffusion coefficient ( ms-1). In the vectors in
the first row of the Fig. 3, the northeast arrows (in the red or green circle) move from left to right. In the bottom row, the motion
vectors move in the same direction.

Download high-res image (424KB) Download full-size image

Fig. 3. Motion vectors obtained from VET algorithm (first row), and estimated by Burgers’ Eq. (6) (second row) using the vector
of 0300 LST 30 Jun 2012. The black line is the western edge of the domain advected by motion vector.

In Fig. 4, the correlation functions show the average correlations (or cosine angles) between two vectors using (3). Note that the
correlation value of two vectors in (3) is , where is the angle formed by the two vectors. An average correlation coefficient of
means that the average angle between two vectors is about . In Fig. 4, and represent the
motion vectors obtained from the VET at time and each subsequent time t, respectively, and motion vector simulated
by Burgers’ equation at time t using the initial vector . Fig. 4(a) shows the correlations between and ,
and between and in the quality mask region (see Section 2). The same correlations are presented similarly in the
overlap region precipitation area and the quality mask region. In Fig. 4(a), the correlation coefficient of is slightly higher
than that of . In Fig. 4 (b), the difference between two correlations is much larger than that of Fig. 4(a). Thus, the
predicted motion vector ( ) from the Type 2 scheme is more similar to the vector derived from the VET algorithm that
uses the observed rainfall over time than the initial vector .

Download high-res image (113KB) Download full-size image

https://www.sciencedirect.com/science/article/pii/S0022169419308753?via%3Dihub 7/17
2019. 9. 16. Improved rainfall nowcasting using Burgers’ equation - ScienceDirect

Fig. 4. Averaged correlation coefficients (for 6 events) between (motion vector obtained from VET at each time t) and the
initial vector (black line), and between and (simulated motion vector by Burgers’ eq.) in (a) quality mask
region and (b) the intersection region of quality mask and positive rain field.

Since the predicted motion vector changes over time, the Type 2 forecasts are different from those of Type 1 or MAPLE
as shown in Fig. 5. In particular, a large difference appears at the western edge of the area, and this difference is moved east by
the motion vector . Type 4 forecasts are smoother than Type 2 forecasts, as shown in Fig. 5. In the binary verification shown in
Fig. 6, the Type 4 forecast has a larger count a (hits) and smaller b (misses) than MAPLE for the lead time of 3 h. In this event, the
Type 4 forecast shows much better performance than does MAPLE.

Download high-res image (437KB) Download full-size image

Fig. 5. Observations (top) and the forecasts obtained by MAPLE (second row), Type 2 (third row), and Type 4 (fourth row) starting
at 1900 LST 30 Jun 2012 for lead times of 1, 2, and 3 h.

https://www.sciencedirect.com/science/article/pii/S0022169419308753?via%3Dihub 8/17
2019. 9. 16. Improved rainfall nowcasting using Burgers’ equation - ScienceDirect

Download high-res image (356KB) Download full-size image

Fig. 6. The binary type verifications for mm h-1 of forecasts starting at 1900 LST 30 Jun 2012.

5.2. Average skill scores of six events


By comparing the average skill scores of the six events, we compare the performance of MAPLE against that of the four
proposed methods. In Fig. 7, one can see that the average skill scores of MAPLE such as POD, CSI and ETS are lower than those
of Type 2 and Type 4 for a lead time of 3 h and with threshold of and 2 mm h-1. For , the difference in POD
between Type 2 and Type 4 is large, but as increases, this difference is reduced. Thus, the smoothing effect increases the skill
score for predicting weak precipitation.

https://www.sciencedirect.com/science/article/pii/S0022169419308753?via%3Dihub 9/17
2019. 9. 16. Improved rainfall nowcasting using Burgers’ equation - ScienceDirect

Download high-res image (288KB) Download full-size image

Fig. 7. Averaged skill scores of six events for thresholds , and mm h-1. In Type 3 and Type 4, is used.

In Fig. 8, the averaged correlation coefficient and the MAE between the forecast and observed fields are shown. The shapes
of for all methods are not exponential functions, so the corresponding life time is not easy to find. Instead,
the integrated values are considered and they are ordered as follows:

where the numbers in parentheses are the integrating values for 3 h lead time. In Fig. 8, since Type 4 has the best correlation
function, we conclude that Burger’ equation and the non-zero smoothing coefficient improve the forecast. The MAE values are
not significantly different between the methods and are slightly reduced by the smoothing effect. As for MAEs, Types 2-4 are
lower than MAPLE for 1.7 hours lead time, but for Type 2 and Type 4, the MAEs are slightly larger than MAPLE after 1.7 hours.
This difference may have been due to the difference in (the number of pixels in the quality mask at time t) between MAPLE
and the proposed methods Type 2 or Type 4.

https://www.sciencedirect.com/science/article/pii/S0022169419308753?via%3Dihub 10/17
2019. 9. 16. Improved rainfall nowcasting using Burgers’ equation - ScienceDirect

Download high-res image (119KB) Download full-size image

Fig. 8. Correlation and MAE between the observed and forecasted rain-rate fields as a function of lead time. All lines represent
the average values of six events.

5.3. Dependence on diffusion coefficients


Turner et al. (2004) applied forecast filters that reduce the root mean square error (RMSE). Optimal forecast filters (OFF) and
operationally useful near-optimal forecast filters (NOFF) have been developed using wavelet transforms and are applied in
MAPLE. These filters grant improved precipitation forecasts in terms of correlation and RMSE. However, improvements in skill
scores such as CSI were not noticeable.

In the present work, filtering or smoothing effects are obtained by adding a diffusion term to PDEs, such as Eq. (7). This
diffusion term allows the forecasts (or solutions) to be smoothed as time progresses. The skill scores can be improved with this
term, and the intensity of smoothing can be controlled by altering the diffusion coefficients . Fig. 9 shows correlation
coefficients and CSI scores for Type 3 and Type 4 forecasts, calculated with , and . Note that, when , Type
3 reduces to Type 1, and Type 4 reduces to Type 2. For , Type 4 gives the best correlation and CSI scores. The orders of
integrating values of for 3 h lead time are:

where the superscripts and numbers in parentheses indicate the corresponding and integrating values, respectively.

https://www.sciencedirect.com/science/article/pii/S0022169419308753?via%3Dihub 11/17
2019. 9. 16. Improved rainfall nowcasting using Burgers’ equation - ScienceDirect

Download high-res image (232KB) Download full-size image

Fig. 9. Averaged values of correlation coefficients (top) and CSI (bottom) for the Type 3 and Type 4 methods for different
diffusion coefficients , and . The threshold mm h-1 is used for CSI.

Larger values may lead to better validation scores, but very large values are less realistic and offer only limited improvement
in smoothing. In the CSI score shown in Fig. 9, the score does not increase significantly even if increases from to . In
addition, the scores are affected more by the type of forecast than the value of .

5.4. Dependence on events


Note that temporal variation of vectors is critical to the accuracy of a forecast. Thus, the forecast can be improved by improving
the accuracy of the motion vectors. The difference between the prediction results of Type 2 and MAPLE depends on the kind of
event that is to be predicted. If the difference can be attributed to temporal variation of motion vectors, we need to find a
measure of the state of the motion vector. To measure the degree of non-stationarity of the vector fields , we estimate
the rate of change of the motion vector over time, as , which is the advection term in Eq. (7). The hourly averages
of spatial mean values for are listed in Table 1 for each event. For motion vectors , the VET algorithm is used
for each event up to 3 h lead times every 10 min. If the motion vector in Eq. (7) is stationary, the shape of the vector will not
change significantly over time and the magnitude of will be close to zero. In this case, the Type 2 forecasts will likely
resemble those of MAPLE and and Type 1. On the other hand, if the motion vector fields are strongly non-stationary, Type 2 will
produce different forecasts from MAPLE. Based on the mean values in Table 1, we classify the six events into two classes of
stationarity with the threshold of ms-2.

Table 1. The spatial mean of the vector length . The hourly value is the average of motion vectors for 1 h lead time. The
“Mean” indicates the average for 3 h lead times. The unit is ms-2.

Events (LST) h h h h Mean

30 Jun 2012, 0300

30 Jun 2012, 1900 0.8130 2.0881 0.7921 0.5928 1.1395

17 Sep 2012, 1000 1.9801 2.7822 1.5587 1.3370 1.8972

https://www.sciencedirect.com/science/article/pii/S0022169419308753?via%3Dihub 12/17
2019. 9. 16. Improved rainfall nowcasting using Burgers’ equation - ScienceDirect

Events (LST) h h h h Mean

17 Sep 2012, 1200 1.6197 1.3370 1.1015 0.7200 1.0827

23 Jun 2014, 1900 0.3160 0.3755 0.4741 0.4612 0.4306

25 Aug 2014, 1100 0.5962 0.6501 0.8278 0.9863 0.8096

The temporal variation of the vectors can be evaluated quantitatively by calculating the correlation coefficient with the initial
vector from VET. In the top row of Fig. 10 (when mm s-2), the correlation scores decrease faster than
the steady state cases of the bottom row. The predicted vectors (from Burgers’ equation) also have much higher
correlations with than the initial vector does. In the steady state cases (of the bottom row in Fig. 10), the gaps between
the two correlations are not much more than those of non-stationary precipitation events (the top row in Fig. 10).

Download high-res image (227KB) Download full-size image

Fig. 10. Comparisons of averaged correlation coefficients of two vectors over time for 3 events (1900 LST 30 Jun 2012, 1000 LST
17 Sep 2012, and 1200 LST 17 Sep 2012) when the mean values of are larger than mm s-2 (top row), and the other
3 events (0300 LST Jun 2012, 1900 LST 23 Jun 2014, and 1100 LST 25 Aug 2014) less than mm s-2 (bottom row).

Similarly, the upper row in Fig. 11 shows the average verification scores for events with thresholds greater than ms-2, and
the bottom row shows the average scores for the remaining three events. In this figure, the Type2 and Type 4 methods, using
Burgers’ equation, yield much better skill scores than the other methods when is relatively large, whereas the other
cases does not show significant improvement in the skill scores. This finding implies Burgers’ equation improves prediction
when the motion vector varies greatly in time

https://www.sciencedirect.com/science/article/pii/S0022169419308753?via%3Dihub 13/17
2019. 9. 16. Improved rainfall nowcasting using Burgers’ equation - ScienceDirect

Download high-res image (297KB) Download full-size image

Fig. 11. Averaged skill scores for 3 events (1900 LST 30 Jun 2012, 1000 LST 17 Sep 2012, and 1200 LST 17 Sep 2012) when the
mean of are larger than mm s-2 (top row) and the other 3 events (0300 LST Jum 2012, 1900 LST 23 Jun 2014, and
1100 LST 25 Aug 2014) less than mm s-2 (bottom row) for mm h-1 and .

5.5. Relative computational costs


The computational costs of the PDEs based methods (Type 1-4) depend mainly on the number of iteration needed to solve the
PDEs. Thus, the four methods should take longer to compute than MAPLE. If we take min, 1800 iterations should be
performed for 3 hours of forecasting. Under these conditions, Type 1 and Type 3 took 1.9 times longer than MAPLE and Type 2
or Type 4 took about 2.25 times longer than MAPLE for the 3-hour forecasting.

The PDE-based methods naturally smooth forecasting through the diffusion parameter, but they require greater computational
costs and complexity than semi-Lagrangian methods. Alternatively, a backward semi-Lagrangian scheme can be applied with
Burgers’ equation for motion vector updating. Semi-Lagrangian algorithms that use Burgers’ equation, which are not
considered in this study, could be a simple and low-cost solution, and we plan to study this approach in future work.

6. Conclusions and discussion


This paper has presented forecasting models that use advection-diffusion equations with non-stationary motion vectors. The
presented methods solve PDEs with initial conditions of the rain field and motion vectors obtained from the VET algorithm. To
test performance, all four proposed methods were tested with radar rain field data with high spatial resolution, and the
resulting evaluation scores were compared with those of the MAPLE algorithm.

The Type 1 method uses simple advection equation and its forecasts are very similar to those of MAPLE. Type 2 uses the same
advection equation as Type 1, but also solves a 2D system of Burgers’ equations and updates the motion vector in each time step.
This scheme provides distinctly different forecasts from the Type 1 and MAPLE, whose the motion vector is not updated after
the initial motion vector.

The Type 3 algorithm uses the same motion vector as Type 1, but is based on the advection-diffusion equation. Adding the
diffusion term improves both the correlation and skill scores of the forecast. The diffusion term is implemented with a filter (or
smoothing) to eliminate the increased dissipation of numerical errors and small-scale characteristics. This filtering effectively
increases the correlation coefficient and skill score at the small threshold . The smoothness of the image can also be
controlled easily by changing the diffusion coefficient .

https://www.sciencedirect.com/science/article/pii/S0022169419308753?via%3Dihub 14/17
2019. 9. 16. Improved rainfall nowcasting using Burgers’ equation - ScienceDirect

The Type 4 algorithm combines the features of Type 2 and Type 3, and it is the best method for smoothing the prediction, since
it updates the motion vectors in every time step. Type 4 forecasts are generally better than those if MAPLE and the other
methods We find that the Type 2 and Type 2 methods for updating motion vectors are superior to MAPLE, especially when the
motion vector changes dramatically in a short time. To measure the change of over time, we estimate the magnitude of
using the vector obtained from the VET algorithm.

We expect that our new forecasting method using motion vectors that can be updated every time from the initial vector will
overcome some of the weaknesses of semi-Lagrange persistence for systems in which non-stationarity is more important than
advection.

7. Author contributions
S. Ryu and G.W. Lee conceived of the models and wrote the manuscript. S. Ryu and G. Lyu performed all numerical
computations and data analysis. Y. Do aided in interpreting the results and worked on the manuscript.

Acknowledgments
This work was funded by the Korea Meteorological Administration Research and Development Program under Grant KMI2018-
06810.

Recommended articles Citing articles (0)

References
Atencia et al., 2017 A. Atencia, I. Zawadzki, M. Berenguer
Scale characterization and correction of diurnal cycle errors in maple
Journal of Applied Meteorology and Climatology, 56 (2017), pp. 2561-2575, 10.1175/JAMC-D-16-0344.1
DOI: 10.1175/JAMC-D-16-0344.1
CrossRef View Record in Scopus Google Scholar

Bartuccelli et al., 1985 M. Bartuccelli, P. Carbonaro, V. Muto


Kadomtsev-petviashvili-burgers for shallow-water waves
Lettere al Nuovo Cimento, 1971–1985 (42) (1985), pp. 279-284, 10.1007/BF02722453
DOI: 10.1007/BF02722453
View Record in Scopus Google Scholar

Bechini and Chandrasekar, 2017 R. Bechini, V. Chandrasekar


An enhanced optical flow technique for radar nowcasting of precipitation and winds
Journal of Atmospheric and Oceanic Technology, 34 (2017), pp. 2637-2658, 10.1175/JTECH-D-17-0110.1
DOI: 10.1175/JTECH-D-17-0110.1
CrossRef Google Scholar

Bellon and Austin, 1978 A. Bellon, G. Austin


The evaluation of two years of real-time operation of a short-term precipitation forecasting procedure (sharp)
Journal of Applied Meteorology, 17 (1978), pp. 1778-1787, 10.1175/1520-0450(1978) 017<1778:TEOTYO>2.0.CO;2
CrossRef View Record in Scopus Google Scholar

Bowler et al., 2004 N. Bowler, C. Pierce, A. Seed


Development of a precipitation nowcasting algorithm based on optical flow techniques
Journal of Hydrology, 288 (2004), pp. 74-91, 10.1016/j.jhydrol.2003.11.011
Article Download PDF View Record in Scopus Google Scholar

Bowler et al., 2006 N. Bowler, C. Pierce, A. Seed


Steps: A probabilistic precipitation forecasting scheme which merges an extrapolation nowcast with downscaled nwp
Quarterly Journal of the Royal Meteorological Society, 132 (2006), pp. 2127-2155
CrossRef View Record in Scopus Google Scholar

Brio and Hunter, 1992 M. Brio, J. Hunter


Mach reflection for the two-dimensional burgers equation
Physica D: Nonlinear Phenomena, 60 (1992), pp. 194-207, 10.1016/0167-2789(92)90236-G

https://www.sciencedirect.com/science/article/pii/S0022169419308753?via%3Dihub 15/17
2019. 9. 16. Improved rainfall nowcasting using Burgers’ equation - ScienceDirect

http://www.sciencedirect.com/science/article/pii/016727899290236G
Article Download PDF View Record in Scopus Google Scholar

Browning et al., 1982 K. Browning, C. Collier, P. Larke, P. Menmuir, G. Monk, R. Owens


On the forecasting of frontal rain using a weather radar network
Monthly Weather Review, 110 (1982), pp. 534-552, 10.1175/1520-0493(1982)110<0534:OTFOFR>2.0.CO;2
CrossRef View Record in Scopus Google Scholar

Courant et al., 1967 R. Courant, K. Friedrichs, H. Lewy


On the partial difference equations of mathematical physics
IBM Journal of Research and Development, 11 (1967), pp. 215-234, 10.1147/rd.112.0215
CrossRef View Record in Scopus Google Scholar

Dixon and Wiener, 1993 M. Dixon, G. Wiener


Titan: Thunderstorm identification, tracking, analysis, and nowcasting-a radar-based methodology
Journal of Atmospheric and Oceanic Technology, 10 (1993), pp. 785-797, 10.1175/1520-
0426(1993)010<0785:TTITAA>2.0.CO;2
CrossRef View Record in Scopus Google Scholar

Foresti et al., 2018 L. Foresti, I.V. Sideris, L. Panziera, D. Nerini, U. Germann


A 10-year radar-based analysis of orographic precipitation growth and decay patterns over the swiss alpine region
Quarterly Journal of the Royal Meteorological Society, 144 (2018), pp. 2277-2301
CrossRef View Record in Scopus Google Scholar

Fox and Wikle, 2005 N. Fox, C. Wikle


A bayesian quantitative precipitation nowcast scheme
Weather and Forecasting, 20 (2005), pp. 264-275, 10.1175/WAF845.1
View Record in Scopus Google Scholar

Germann and Zawadzki, 2002 U. Germann, I. Zawadzki


Scale-dependence of the predictability of precipitation from continental radar images. part i: Description of the
methodology
Monthly Weather Review, 130 (2002), pp. 2859-2873, 10.1175/1520-0493(2002) 130<2859:SDOTPO>2.0.CO;2
CrossRef View Record in Scopus Google Scholar

Handwerker, 2002 J. Handwerker


Cell tracking with trace3d-a new algorithm
Atmospheric Research, 61 (2002), pp. 15-34, 10.1016/S0169-8095(01)00100-4
Article Download PDF View Record in Scopus Google Scholar

Jinno et al., 1993 K. Jinno, A. Kawamura


R., B., Larson, M., Niemczynowicz, J., Real-time rainfall prediction at small space-time scales using a two-dimensional
stochastic advection-diffusion model
Water Resources Research, 29 (1993), pp. 1489-1504, 10.1029/92WR02849
View Record in Scopus Google Scholar

Kawamura et al., 1997 A. Kawamura, K. Jinno, R. Berndtsson, T. Furukawa


Real-time tracking of convective rainfall properties using a two-dimensional advection-diffusion model
Journal of Hydrology, 203 (1997), pp. 109-118, 10.1016/S0022-1694(97)00088-7
Article Download PDF View Record in Scopus Google Scholar

Li and Lai, 2004 P. Li, E. Lai


Short-range quantitative precipitation forecasting in hong kong
Journal of Hydrology, 288 (2004), pp. 189-209, 10.1016/j.jhydrol.2003.11.034
Article Download PDF View Record in Scopus Google Scholar

Menard, 1994 R. Menard


Kalman Filtering of Burgers’ Equation and its Application to Atmospheric Data Assimilation. Ph.D. thesis
UNIVERSITY (CANADA), MCGILL (1994)
Google Scholar

https://www.sciencedirect.com/science/article/pii/S0022169419308753?via%3Dihub 16/17
View publication stats

2019. 9. 16. Improved rainfall nowcasting using Burgers’ equation - ScienceDirect

Mueller et al., 2003 C. Mueller, T. Saxen, R. Roberts, J. Wilson, T. Betancourt, S. Dettling, N. Oien, J. Yee
Ncar auto-nowcast system
Weather and Forecasting, 18 (2003), pp. 545-561, 10.1175/1520-0434(2003)018<0545:NAS>2.0.CO;2
View Record in Scopus Google Scholar

Radhakrishna et al., 2012 B. Radhakrishna, I. Zawadzki, F. Fabry


Predictability of precipitation from continental radar images. part v: Growth and decay
Journal of the Atmospheric Sciences, 69 (2012), pp. 3336-3349, 10.1175/JAS-D-12-029.1
View Record in Scopus Google Scholar

Rasmussen et al., 2003 R. Rasmussen, M. Dixon, S. Vasiloff, F. Hage, S. Knight, J. Vivekanandan, M. Xu


Snow nowcasting using a real-time correlation of radar reflectivity with snow gauge accumulation
Journal of Applied Meteorology, 42 (2003), pp. 20-36, 10.1175/1520-0450(2003)042<0020:SNUART>2.0.CO;2
CrossRef View Record in Scopus Google Scholar

Ruzanski et al., 2011 E. Ruzanski, V. Chandrasekar, Y. Wang


The casa nowcasting system
Journal of Atmospheric and Oceanic Technology, 28 (2011), pp. 640-655, 10.1175/2011JTECHA1496.1
View Record in Scopus Google Scholar

Seed, 2003 A. Seed


A dynamic and spatial scaling approach to advection forecasting
Journal of Applied Meteorology, 42 (2003), pp. 381-388, 10.1175/1520-0450(2003)042<0381:ADASSA>2.0.CO;2
CrossRef View Record in Scopus Google Scholar

Stahl and Morten Aamo, 2011 Stahl, A., Morten Aamo, O., 2011. A New Framework for Motion Estimation in Image Sequences
Using Optimal Flow Control. volume 89. pp. 307–320. DOI: 10.1007/978-3-642-19539-6_20.
Google Scholar

Tang and Matyas, 2018 J. Tang, C. Matyas


A nowcasting model for tropical cyclone precipitation regions based on the trec motion vector retrieval with a semi-
lagrangian scheme for doppler weather radar
Atmosphere, 9 (2018), p. 200
CrossRef Google Scholar

Tsonis and Austin, 1981 A. Tsonis, G. Austin


An evaluation of extrapolation techniques for the short-term prediction of rain amounts
Atmosphere-Ocean, 19 (1981), pp. 54-65, 10.1080/07055900.1981.9649100
CrossRef View Record in Scopus Google Scholar

Turner et al., 2004 B. Turner, I. Zawadzki, U. Germann


Predictability of precipitation from continental radar images. part iii: Operational nowcasting implementation (maple)
Journal of Applied Meteorology, 43 (2004), pp. 231-248, 10.1175/1520-0450(2004) 043<0231:POPFCR>2.0.CO;2
View Record in Scopus Google Scholar

Wendroff, 1968 B. Wendroff


Difference methods for initial-value problems (robert d. richtmyer and k. w. morton)
SIAM Review, 10 (1968), pp. 381-383, 10.1137/1010073, DOI: 10.1137/1010073
CrossRef Google Scholar

© 2019 Published by Elsevier B.V.

About ScienceDirect Remote access Shopping cart Advertise Contact and support Terms and conditions Privacy policy

We use cookies to help provide and enhance our service and tailor content and ads. By continuing you agree to the use of cookies.
Copyright © 2019 Elsevier B.V. or its licensors or contributors. ScienceDirect ® is a registered trademark of Elsevier B.V.

https://www.sciencedirect.com/science/article/pii/S0022169419308753?via%3Dihub 17/17

You might also like