9th International Conference on Urban Drainage Modelling

Belgrade 2012

Weather radar and heavy rainfall - how to estimate the

real amount of precipitation?
Thomas Einfalt1
1

hydro & meteo GmbH & Co. KG, Breite Str. 6-8, D-23552 Lbeck, Germany (einfalt@hydrometeo.de)

ABSTRACT
Extreme precipitation often occurs on relatively small areas which are not
sufficiently equipped with rain gauges to completely observe the occurred event.
Therefore, radar data are extremely helpful to localise the most intense parts of the
precipitation. However, radar data are prone to errors under extreme rainfall and
bear higher uncertainties at higher rainfall intensities due to the non-linearity of the
relationship between radar reflectivity and rainfall intensity and due to the
unknown drop-size distribution of the rain cells. A case study illustrates a practical
approach to test several assumptions on the drop-size distribution and important
radar data quality considerations for high intensity precipitation.

KEYWORDS
Extreme events, radar rainfall, rain gauges, sparse network

INTRODUCTION

Local heavy precipitation is usually not captured by traditional point rainfall gauges. Weather radar,
although less precise in rainfall volume at a point, permits a detailed view into spatial structures of
precipitation. Therefore, weather radar plays an increasingly important role for the a-posteriori
analysis of such events, in particular in presence of damage. For online data processing, different
procedures are required.
Crucial points with radar are the potential measurement errors (Michelson et al., 2005) and the
unknown drop size distribution required for a good estimation of rainfall amounts (Collier, 1989). For
practical work with radar data in the urban context, a number of quality controls and data corrections
need to be performed before a reliable result can be achieved (Einfalt et al., 2004).

RADAR DATA QUALITY CONTROL

As mentioned by Michelson et al. (2005), there are several sources of error which affect the ability of
radars to measure precipitation and which influence the accuracy of the measurements. These errors

are not always present in the radar measurement, but many of them appear only under certain
meteorological conditions:

Attenuation: reduction of the measured reflectivity due to heavy precipitation (see 2.2),

Bright band: high values in the melting layer of the atmosphere when snow melts to rain,

Anomalous propagation: measurement of ground targets due to atmospheric conditions

preventing a straight propagation of the radar beam.

Among (mostly) stationary influences on a correct radar measurement, ground clutter (see 2.1) and
(partial) blockage of the radar beam by buildings or topography are the most frequent ones.
Readers interested in more details on error sources and their effects on radar measurements should
refer to the document of Michelson et al. (2005).
2.1

Ground clutter

Ground clutter is usually strong due to the relative radar cross-section of the ground being much
greater than that from meteorological targets. Ground clutter can be minimized through intelligent
radar siting, Doppler suppression, and through the use of post-processing methods such as static clutter
maps.
2.2

Attenuation of the radar signal

Heavy rain, graupel and hail can attenuate energy, leading to strong underestimation of precipitation
intensities. Especially in hail the scattered energy can be attenuated to the point of virtual extinction of
the signal. Shorter wavelengths (X and C bands) are more seriously affected.
Attenuation can be detected by a close look into measurements from rain gauges and radar (figure 1)
and by visual inspection of the radar images (e.g. figure 2). The time series analysis shows during
which time interval radar has seen considerably less precipitation than the rain gauge (in the red
ellipse): the thin line shows the rain gauge measurement and the two bold lines radar measurement at
two neighbouring points. It becomes obvious that radar has seen much less precipitation during the
short time interval just before 18:00 hours.

Figure 1: Attenuation of the radar signal: detection through time series comparison to rain gauges

Figure 2: Attenuation to the NW of the radar

2.3

Attenuation at the radar site (radome attenuation)

In heavy rain, a thin film of water will cover the radome, causing signal attenuation. In cold
conditions, snow and ice may build up on top of the radome, also causing attenuation and limiting the
quantitative use of reflectivity measurements.
Radome attenuation can best be observed in a sequence of radar images (figure 3 - left) where the
decrease and later increase of the radar signal over the whole radar range can be observed. A counter
measure is extremely difficult because the attenuation is not necessarily uniform over the different
viewing directions of the radar. Else, a simple correction factor over the complete radar scope could
sometimes save the data this manual method in some cases nevertheless improves the data (figure 3 right).

Figure 3: Uncorrected radar measurement with radome attenuation (left) and correction attempt (right)

ESTIMATION OF THE PRECIPITATION AMOUNT

The relationship between the reflected radar signal in [dBZ] and the rainfall intensity [mm/h] depends
on the drop size distribution of the precipitation. The drop size is related to the reflectivity by a power
of 3, and to intensity by a power of 6, yielding a highly nonlinear relationship between reflectivity and
intensity (Collier, 1989). Because the drop size distribution in practice is not known, assumptions have
to be made in order to assess it: convective (summer thunderstorm) events are associated with a higher
amount of large drops, whereas stratiform events are merely associated to smaller drops. These
assumptions are important because intense precipitation corresponds to the higher reflectivities where
the resolution of the radar in terms of mm/h is worse. Figure 4 shows for two examples, the MarshallPalmer relationship for tropical wide spread rainfall and the original relationship of the German
Weather service for convective type rainfall the difference: a reflectivity value of 50 dBZ may
correspond to 50 mm/h or to 70 mm/h in intensity, depending on the weather type associated.
The large uncertainty in intensity values due to different assumptions about the unknown drop size
distribution is the main reason that for quantitative use of radar data in hydrology it is indispensable to
compare radar measurements to rain gauge data: the adjustment of radar to rain gauge values is the
consequence of this.
In practice, the most realistic relationship to convert reflectivity to intensity is selected, and the
resulting values are then fitted to the ground measurements at the rain gauge locations.
200
Marshall-Palmer

180

DWD

160

Intensity [mm/h]

140
120

100
80
60
40
20
0
0

10

20

30

40

50

60

Reflectivity [dBZ]

Figure 4: Relationship between rainfall intensity [mm/h] and radar reflectivity [dBZ]

4
4.1

PRACTICAL WORK STEPS: ANALYSIS OF A CASE STUDY

The event of 28 August 2002

The event of 28 August 2002 produced damage in the area of Eitorf in North Rhine-Westphalia where
a part of a highway was washed away. During the two hour event, rain gauges registered up to 70 mm
of precipitation. Radar images suggest that the peak of the precipitation was higher.

4.2

Radar adjustment to gauges

Since radar measurements are uncertain for estimating rainfall amounts correctly, rain gauges are used
as an anchor point for a quantitative precipitation estimation. Figure 5 shows that rain gauges are
measuring a higher amount at the rain gauge sites than radar for this case study (figure 6), regardless
of the Z-R relationship employed. Radar on the other side clearly shows that the centre of the highest
precipitation is situated between the four rain gauges with the highest amounts. Thus, a combination of
both yields the best possible precipitation estimation.

Figure 5: Precipitation amount seen by the available rain gauges (event sum, inverse distance weighted
(IDW) interpolation)

Figure 6: Rainfall amount seen by radar with a moderate Z-R relationship (left) and a convective Z-R
relationship (right). The locations of the closest rain gauges are circled in red.

In order to identify the Z-R relationship presumably closer to reality, a cross validation of the four
main rain gauges has been performed during the adjustment process: Each of the gauges has been
skipped in turn, and the results have been compared with the results from the adjustment with all
gauges. Due to the selected Brandes type adjustment scheme (Wilson / Brandes, 1979), the radar
values at the gauge locations correspond to the rain gauge amounts.
The statistics chosen to analyse the results are
The absolute difference between the gauge measurements and the adjusted radar amounts,
The mean percentage difference, allowing to detect systematic deviations,
The absolute percentage difference, giving a larger weight to gauges with smaller rainfall
amounts than the first criterion.
Table 1 shows that the rain gauge with the highest observed rainfall amount (Eitorf) is the most
difficult to reach by both Z-R relationships. Although the differences between the two selected Z-R
relationships are not very large, all statistics show that the convective Z-R relationship appears to be
more appropriate to use for this event.
Table 1: Results of the cross validation of the adjustment for the four main rain gauges and two Z-R
relationships

Gauge
Eitorf
Lascheid
Hanfmhle
Kuchenbach
Parameter Sum

Rain at
Radar
gauge
mod. conv.
[mm]
[mm] [mm]
70.3 57.42 61.99
50.4 54.65 56.74
31.3 31.77 30.21
36.5 41.23 38.42

absolute
difference
mod.
conv.
12.88
8.31
4.25
6.34
0.47
1.09
4.73
1.92
22.33 17.66

mean percentage
difference
mod.
conv.
-18.32
-11.82
8.43
12.57
1.52
-3.50
12.96
5.27
4.58
2.53

absolute percentage
difference
mod.
conv.
18.32
11.82
8.43
12.57
1.52
3.50
12.96
5.27
41.22
33.16

The use of the convective type Z-R relationship resulted in a maximum value at the point with the
highest precipitation amount of 103 mm (Figure 7), whereas the moderate relationship would have
yielded 95 mm as event sum for the maximum pixel. Figure 7 also demonstrates the necessity to adjust
the radar data: without this procedure, radar would have estimated the precipitation to be lower by
approximately a factor of 2.

.
Figure 7: Original (left) and gauge adjusted (right) rainfall amounts based on radar data with the
convective type Z-R relationship
7

CONCLUSIONS AND OUTLOOK

Radar data are a required source of information for the analysis of small scale heavy precipitation. The
measurements of both radar and rain gauges are to be checked for errors in a detailed manner for such
events. Then radar is able to give a good estimate of the precipitation between the rain gauge
locations. A method to cross compare the use of different Z-R relationships and their quality to assess
the amount of rainfall has been presented.
An automated cross-validation check of the type presented in Table 1 and using a small catalogue of
likely Z-R relationships has the capability of improving quantitative precipitation estimates.

REFERENCES

Collier, C.G., 1989: Applications of Weather Radar Systems a Guide to Uses of Radar Data in
Meteorology and Hydrology, Ellis Horwood Limited, Chichester, England.
Einfalt, T., Arnbjerg-Nielsen, K., Golz, C., Jensen, N.E., Quirmbach, M., Vaes, G., Vieux, B. (2004).
Towards a Roadmap for Use of Radar Rainfall data use in Urban Drainage. Journal of
Hydrology, 299, pp. 186-202.
Michelson D., Einfalt T., Holleman I., Gjertsen U., Friedrich K., Haase G., Lindskog M., Jurczyk A.,
2005: Weather radar data quality in Europe quality control and characterisation. Review, COST
Action 717 - Use of radar observations in hydrological and NWP models, Luxembourg.
Wilson, J.W., Brandes E.A. (1979). Radar measurement of rainfall A summary, American
Meteorological Society, 60, 1048-1058.

9th International Conference on Urban Drainage Modelling

Belgrade 2012

Radar-raingauge data combination techniques: a

revision and analysis of their suitability for urban
hydrology
Li-Pen Wang1, Susana Ochoa2, Nuno Simes3, Christian Onof4, edo
Maksimovi5
1

Imperial College London, United Kingdom, li-pen.wang08@imperial.ac.uk

Imperial College London, United Kingdom, s.ochoa-rodriguez@imperial.ac.uk
3
Imperial College London, United Kingdom & University of Coimbra, Portugal, nuno.simoes08@imperial.ac.uk
4
Imperial College London, United Kingdom, c.onof@imperial.ac.uk
5
Imperial College London, United Kingdom, c.maksimovic@imperial.ac.uk
2

ABSTRACT
The applicability of the operational radar and raingauge networks for urban
hydrology is insufficient. Radar rainfall estimates provide a good description of the
spatiotemporal variability of rainfall; however, their accuracy is in general
insufficient. It is therefore necessary to adjust radar measurements using raingauge
data, which provide accurate point rainfall information. Several gauge-based radar
rainfall adjustment techniques have been developed and mainly applied at coarser
spatial and temporal scales; however, their suitability for small-scale urban
hydrology is seldom explored. In this paper a review of gauge-based adjustment
techniques is first provided. After that, two techniques, respectively based upon the
ideas of mean bias reduction and error variance minimisation, were selected and
tested in an urban catchment (865 ha) in North-East London. The radar rainfall
estimates of four historical events (2010-2012) were adjusted and applied to the
hydraulic model of the study area. The results show that both techniques can
effectively reduce mean bias; however, only the technique based upon error
variance minimisation can correctly reproduce the spatial and temporal variability
of rainfall, which proved to have a significant impact on the associated hydraulic
outputs. This suggests that error variance minimisation methods may be more
appropriate for urban hydrological/hydraulic applications.

KEYWORDS
Gauge-based adjustment, merging/combination, pluvial flooding, radar, rainfall, urban hydrology.

INTRODUCTION

Rainfall constitutes the main input for urban pluvial flood models and the uncertainty associated to it
dominates the overall uncertainty in the modelling and forecasting of this type of flooding (Golding,
2009). The rainfall events which generate pluvial flooding are often associated with thunderstorms of
high intensity and small spatial scale ( 10 km), whose magnitude and spatial distribution are difficult
to monitor and predict (Collier, 2009; Golding, 2009; Vieux and Imgarten, 2011).
The sensors that are commonly used for estimation and prediction of rainfall at catchment scales are
raingauges and radars (Cole and Moore, 2008); however, the applicability (i.e. achievable accuracy
and resolution) of the currently operational radar and raingauge networks for urban hydrology is
insufficient. In general, raingauges provide accurate point rainfall estimates near the ground surface;
nonetheless, they cannot capture the spatial variability of rainfall which has a significant impact on the
physical processes and thus on modelling of urban pluvial flooding (Tabios and Salas, 1985; Syed et
al., 2003). Moreover, since dense raingauge networks cannot cover large areas, it is difficult to
forecast rainfall with longer lead time based on raingauge data only (Looper and Vieux, 2012). In
contrast, radars can survey large areas and can capture the spatial variability of the rainfall, thus
improving the short-term predictability of rainfall and flooding. However, the accuracy of radar
measurements is in general insufficient, particularly in the case of extreme rainfall magnitudes (Einfalt
et al., 2005; Harrison et al., 2009); this has a tremendous effect on the subsequent rainfall forecast
(which uses radar estimates as starting point) and on the associated flood forecast (whose main input is
the rainfall forecast) (Ligouri et al., 2011). The low accuracy of radar measurements is mainly due to
the fact that, unlike raingauges, which directly collect rain droplets, radar devices obtain rainfall
measurements through an indirect process, which introduces more uncertainty. This indirect process
comprises the following sub-processes: (i) noise filtering, (ii) identification of clutter and occultation,
(iii) removal of anomalous propagation, (iv) attenuation correction, (v) calibration or conversion of
reflectivity to rain rate and (vi) raingauge-based adjustment. The last two sub-processes involve
calibration or adjustment of radar estimates based on raingauge measurements; nonetheless, the scale
at which this is done cannot ensure that radar estimates capture high intensities and local conditions
accurately. The conversion of reflectivity to rain rate (i.e. Z - R conversion function) is the result of a
calibration process based on the comparison of a large number of coincidental observations of radar
and raingauges. In order to obtain statistically optimal results, the conversion function has to
compromise the capacity of deriving extreme values since the frequency of their occurrence is
relatively low; hence, the Z - R conversion performs poorly at capturing intense rainfall rates in
particular (Einfalt et al., 2004, 2005), which is vital for urban applications. In addition, the conversion
functions are in general static; i.e. they are not dynamically updated (they only change according to the
storm type; however, for each storm type the conversion function is fixed). After the conversion or
calibration process and with the purpose of further enhancing the suitability of radar estimates for
hydrological and hydraulic applications, gauge-based adjustment techniques, also referred to as
re-calibration, combination or merging in the literature (Einfalt et al., 2004), are widely used to
dynamically correct the bias between radar estimates and the coincidental raingauge measurements
(Fulton et al., 1998; Seo et al., 1999; Harrison et al., 2009). However, these adjustment techniques are
mostly applied in catchments of large area (~1000 km2) and use hourly rainfall rates, and although
they provide benefits for hydrological applications at large scales, the suitability of the resulting
rainfall estimates for urban hydrological applications is still insufficient. For example, Smith et al
(2007) re-investigated 35 rainfall events selected from the US NEXRAD (network of 159 high
resolution Doppler weather radars operated by the US National Weather Service (Fulton et al., 1998))
during the period 2003 - 2005, and compared them with the coincidental point observations recorded
by a dense network of raingauges (one raingauge per km2 in average) over a small urban area

( 14.3 km2). In this comparison large and event-varying bias were observed over the study catchment
even though NEXRAD rainfall products had been dynamically adjusted with raingauge measurements
(similar to the UK Nimrod rainfall data and adjusted based upon hourly scales).
Moreover, the feasibility of applying the existing gauge-based adjustment techniques at smaller spatial
and temporal scales (i.e. for urban applications) has not yet been fully analysed.
In this paper a review of gauge-based adjustment techniques is first provided. After that, two of these
techniques were selected and applied to a study urban catchment in North-East London, with the
purpose of assessing their ability to improve operational radar measurements for urban applications.

REVIEW OF GAUGE-BASED ADJUSTMENT TECHNIQUES

Gauged-based adjustment techniques aim at combining the advantages and overcoming the drawbacks
of radar and raingauge rainfall estimates; that is, to retain the accuracy of the point rainfall information
provided by raingauges and at the same time the broader description of the spatial and temporal
variations of rain-fields provided by radar. As previously mentioned, the final purpose of these
techniques is to enhance the suitability of rainfall estimates for hydrological and hydraulic
applications, including flood modelling and forecasting.
After reviewing different gauge-based adjustment techniques, it was noticed that, in general, they can
be classified into two types: (i) mean bias reduction techniques and (2) error variance minimisation
techniques. A review of these two types of gauge-based adjustment techniques is provided.
2.1

Mean bias reduction techniques

Mean bias is the difference between the mean radar rainfall estimates and the mean raingauge
measurements at the locations of raingauges for a given time period. In the literature it is also termed
systematic error and is thought to be the most important source of uncertainty affecting the
suitability of radar rainfall estimates for hydrological and hydraulic applications (Vieux and Bedient,
2004). Consequently, many adjustment techniques focus on reducing raingauge-radar mean bias in
order to improve radar rainfall estimates. The idea of the mean bias adjustment is to analyse the
differences between raingauge observations and the coincidental radar measurements over a given
period, and then apply this event-varying difference directly to each radar rainfall grid.
An example of this is the adjustment method implemented in the operational UK Nimrod system,
where an adjustment ratio, based on comparisons between processed radar and raingauge hourly
rainfall is applied to the entire domain of each radar site and is updated on an hourly basis (Harrison et
al., 2009). A similar adjustment technique is used in the US NEXRAD system (Seo et al., 1999). As
mentioned before, the adjustments carried out in the Nimrod and NEXRAD systems provide benefits
for large scale hydrological applications; however, the resulting rainfall estimates are not accurate
enough for urban applications (Vieux and Bedient, 2004; Smith et al., 2007).
Mean bias adjustment techniques have also been used at smaller scales in order to further improve
radar rainfall estimates (which may have already been adjusted at larger scales, as in the case of
Nimrod and NEXRAD products). For instance, in the above mentioned work by Vieux and Bedient
(2004) over a 260 km2 urban catchment. They evaluated the hydrological prediction uncertainty
caused by rainfall input errors through event re-construction. Five events were selected from
NEXRAD during the period 1998 2003. These events were re-constructed by applying a simple ratio
to reduce the mean bias between radar and the co-located raingauge observations. This simple ratio
was derived from the comparison between mean radar rainfall estimates and mean local raingauge
3

measurements at the locations of raingauges over the duration of a rainfall event. Results show that the
corresponding flow prediction could be significantly improved by using mean bias adjusted
(corrected) radar rainfall estimates as inputs. A similar work was carried out by Smith et al. (2007) but
over a relatively small catchment (14.3 km2), in which 35 rainfall events were re-investigated and
significant bias were observed between NEXRAD products and raingauge data. Similarly, the authors
reduced bias by applying a simple ratio to scale up or down the radar rainfall rates to approximate the
coincidental raingauge records. These works suggest that mean bias is the most important
uncertainty source decreasing the suitability of radar rainfall estimates for urban hydrological and
hydraulic applications. In addition, they suggest that the suitability of rainfall data for these
applications could be massively improved through locally and dynamically adjusting radar rainfall
estimates using co-located raingauge records. However, this simple mean bias adjustment was carried
out through post-event (or historical rainfall records) comparisons. It is therefore more suitable for
improving the applicability of historical rainfall events to hydrological and/or hydraulic design, rather
than for short-term real-time forecasting. If intended for real-time flood forecasting applications, this
method would require a very dense raingauge network (or a larger area) and a longer temporal
comparison basis (i.e. hourly) to obtain a more reliable ratio to scale the radar rainfall (Anagnostou
and Krajewski, 1999; Seo et al., 1999). It is therefore more suitable for coarser spatial- and temporalresolution rainfall adjustment.
A methodology for improving the local and dynamic capacity of conventional mean bias adjustment
methods was proposed by Moore et al. (1989) and further modified by Wood et al. (2000). A dynamic
calibration factor was introduced to carry out 15-min radar rainfall adjustment in real time. This factor
is based on the comparison of raingauge and radar estimates at each time step in synergy with a
positive correction value and a static calibration factor , where and are long-term derived
constants. Cole and Moore (2008) further examined the applicability of this methodology over two
UK catchments (Darwen and Kent, respectively 135.7 and 212.3 km2). In this work three types of
gauge-based adjustment techniques were used to correct radar estimates: (i) static adjustment, (ii)
standard dynamic adjustment, and (iii) dynamic adjustment including mean bias. The first one is
similar to the aforementioned simple mean-bias adjustment, but based upon a long-term
radar-raingauge comparison. The second and third techniques are respectively based upon the original
local adjustment methodology (Moore et al., 1989) and the modified one (Wood et al., 2000). Results
suggest that the applicability of radar rainfall estimates can be significantly improved by the local
adjustment methodology (i.e. the second and third techniques).
More recently, a geostatistical merging method which also focuses on reducing mean bias was
developed by Ehret et al. (2008) and applied to real-time small-scale flood forecasting in the
Goldersbach catchment, Germany ( 75 km2). At each time step, the point raingauge records are
interpolated into a rainfall field and further merged with the coincidental radar image. The Block
Kriging interpolation technique was employed to ensure that the synthetic rainfall field is unbiased. A
deviation ratio field can be then obtained by comparing the interpolated rainfall and radar rainfall at
each radar grid. This deviation field is then adjusted to ensure that its mean is equal to 1 and it is
further applied to the interpolated rainfall field (in this way it is ensured that the raingauge totals are
retained). A merged rainfall field is therefore obtained at each time step and further used as input for
flood forecasting. The quality of radar rainfall was significantly improved by this merging process
and, consequently, the accuracy of the rainfall and flood forecasts was also largely improved.
Although the mean bias adjustment methods mentioned above have proven to significantly improve
rainfall estimates and the associated flow estimates and forecasts, they have some common drawbacks.
First, the spatial structure (e.g. spatial variability) of radar rainfall fields could be altered by simply
multiplying a ratio to the rainfall estimate at each radar grid. The ability to characterise the spatial
4

variations in rainfall is however one of the most reliable features of radar sensors and should therefore
be retained. Second, the mean bias adjustment methods have difficulty in correcting the temporal and
spatial profiles of radar rainfall. For example, if the original radar rainfall estimates fail to capture the
time of the rainfall peaks, this error will not be corrected by the mean bias adjustment methods since
these focus on the correction of quantitative differences (see Figure 4 of Ehret et al. (2008)).
2.2

Error variance minimisation techniques

Another type of gauge-based adjustment techniques focuses on minimising the error variances. The
error herein represents the difference between true (or raingauge) and radar rainfall estimates. The
concept of minimising the error variances is similar to maximum likelihood approaches; therefore, in
addition to mean bias, the spatial and temporal patterns of rainfall are also taken into account in the
adjustment process (Krajewski, 1987; Todini, 2001; Mazzetti, 2004; Gerstner and Heinemann, 2008).
In general, these techniques assume that there is a true (or best estimated) rainfall field at each time
step, made up of grids whose rainfall volume is the (linear) combination of the coincidental radar and
raingauge estimates. The total rainfall of this true rainfall field is equal to the raingauge total; this
means that the raingauge records are unbiased. However, it is seldom possible to have at least one
raingauge per radar grid; therefore, some further assumptions are necessary. For example, Gerstner
and Heinemann (2008) defined that the rainfall volume at a specific grid of the true rainfall field as
follows:
K

Pa xi , yi Pr xi , yi wik Pg xk , yk Pr xi , yi

(1)

k 1

where Pa(xi, yi) and Pr(xi, yi) are, respectively, the true and radar rainfall at the grid point (xi, yi);
Pg(xk, yk) is the raingauge measurement at the raingauge position (xk, yk); wik is the to-be-determined
weight and K is the number of raingauges. Through minimising the error variances, wiks can be
estimated and their values are in general inversely proportional to the distance between raingauge and
radar grids. Results show that this weighting technique can effectively reflect the local point
information to radar rainfall. However, the study was carried out on a daily basis and its applicability
to sub-daily and sub-hourly radar rainfall adjustment is therefore unknown.
Different from Gerstner and Heinemann (2008), Krajewski (1987) and Todini (2001) employed the
(Block-) Kriging interpolation technique to generate point raingauge information at each radar grid
before merging it with radar estimates. In Krajewski (1987)s Cokriging combination technique, the
best rainfall estimate V*(x0, y0) at a specific location was defined as follows:
NG

NR

i 1

i 1

V * x0 , y0 Gi Gi xi , yi Ri Ri xi , yi

(2)

where NG and NR are the numbers of raingauges and radar grids around location (x0, y0) and Gi(xi, yi)
and Ri(xi, yi) are the associated measurements at location (xi, yi). In the process of deriving Gi and Ri,
the information of the covariances between true and raingauge rainfall CovVG and between true and
radar rainfall CovVR are required; however, it is impossible to obtain them. They are therefore
approximated by the forms CovVG = GCovGG and CovVR = RCovRR, where CovGG and CovRR are the
covariances respectively between raingauges and radar grids and G and R are two to-be-determined
constants ranging from 0 to 1. This approximation however largely decreases the applicability of the
Cokriging technique because the values of G and R are usually determined subjectively or through a
large number of simulations.

Todini (2001), different from Krajewski (1987), employed the Kalman filter algorithm to merge the
interpolated raingauge and radar rainfall field to obtain the true rainfall field at each time step. This
method (called Bayesian combination), instead of using CovVG and CovVR, uses the covariance of
errors Covt to help deriving the true rainfall field. The error t is estimated by comparing the
co-located Block-Kriged and radar rainfall estimates; a Covt matrix can be therefore constructed in
real time to update the original radar estimates to produce the true rainfall field. This Bayesian
combination method has been applied to a 1051 km2 river catchment near Bologna (Italy), where
1 km2 C-band radar images and point rainfall information recorded by a network of 26 raingauges are
available. Results show that the bias and variance between radar and observed rainfall estimates were
significantly reduced. This application however was undertaken in an hourly basis and its potential to
be used in sub-hourly rainfall adjustment needs to be further examined.

METHODOLOGY

In this work, one of each type of gauge-based adjustment techniques was selected and tested at the
urban scale, with the purpose of assessing and comparing their ability to improve operational radar
measurements for urban applications. The selected adjustment methods are next described.
3.1

Mean bias reduction method selected for testing at the urban scale

As mentioned above, the bias between raingauge and radar rainfall estimates is widely regarded as the
dominative factor in the uncertainty of the corresponding hydrological and hydraulic modelling. A
mean bias adjustment method was therefore implemented in this work to evaluate the impact of bias
reduction on the corresponding hydraulic outputs of an urban catchment.
The implemented technique is a post-event one, where the mean sample bias is defined as the ratio of
the mean raingauge accumulations to the co-located mean radar rainfall accumulations on an event
total basis, i.e.,

Bi

j 1

RGij m

(3)

R x j , y j m
j 1 i

where Bi is the sample bias for the i-th event, m is the number of raingauges, RGij is the rainfall
accumulation (in mm) for the i-th event at the j-th raingauge, (xj, yj) is the geographical location of the
j-th raingauge and Ri(xj, yj) denotes the radar rainfall accumulation (in mm) for the i-th event at
location (xj, yj).
In order to correct the mean bias, the event sample bias Bi is applied to each radar rainfall grid over the
study area for each selected rainfall event (the adjusted values will be referred to as Corrected radar
estimates).
3.2

Error variance minimisation method selected for testing at the urban

scale

In this work, the Bayesian combination method proposed by Todini (2001) was selected for testing in
an urban catchment. The reasons for selecting this method are the following:
1. It has a strong theoretical background and relatively little approximation
2. The software is available and well maintained

3. It does not require numerous simulations and historical rainfall events to determine parameters
This is a dynamic method intended for real-time applications. The first step of the method is to, for
each time step, interpolate the real-time raingauge measurements into a synthetic rainfall field using
the Block Kriging (BK) interpolation method. After that, the interpolated rainfall field is merged with
the coincidental radar image using the Kalman filter algorithm (Todini, 2001; Mazzetti, 2004).
The idea of the BK interpolation method is to synthesise a rainfall field whose semi-variogram curve
is very similar to the semi-variogram curve empirically estimated from the associated point rainfall
information, where the semi-variogram curve is a function used to characterise the degree of spatial
dependence of a spatial random field (e.g. a rainfall field). In other words, the Block Kriged rainfall
field contains not only accurate point rainfall estimates but also the spatial dependences between these
point estimates over a specific area. This information can reflect the spatial structure of rainfall right
above the ground; this can be very useful to correct the spatial structure observed by radar, which is at
a given elevation above the ground and could be horizontally shifted by wind advection. This idea of
using spatial dependences estimated from raingauge observations to improve the generation of rainfall
processes has been widely used in the hydrological field (Wheater et al., 2005; Yang et al., 2005).
The Kalman filter algorithm used for carrying out the actual merging comprises two steps: predict and
update (Kalman, 1960). In the predict step, the a priori estimates and status at the current time step
are firstly predicted based upon the estimates and status at the previous time step. These a priori
estimates and status are then updated using real-time observations and the a posteriori estimates and
status can be obtained by minimising the variance between the a priori estimates and the observations
(termed error variance). In the method proposed by Todini (2001), the radar image represents the a
priori estimates and the interpolated rainfall field constitutes the observations to update the predicted
estimates for obtaining the output field (a posteriori estimate) at each time step.

4
4.1

EXPERIMENTAL SITE AND DATASET

Cranbrook Catchment

The above mentioned gauge-based adjustment techniques were tested in the Cranbrook catchment.
This catchment is located within the London Borough of Redbridge (North-East part of Greater
London - see Figure 1). It is predominantly urbanised and has a drainage area of approximately
865 hectares; the main water course is about 5.75 km long, of which 5.69 km are piped or culverted.
This area has experienced several pluvial, fluvial and coincidental floodings in the past.
4.2

Radar (Nimrod) Data

The Cranbrook catchment is in the coverage of two radars, Chenies and Thurnham (Figure 1(a)). The
radar data are measured by the C-band radar network operated by the UK Met Office through the
British Atmospheric Data Centre (BADC) with spatial and temporal resolutions of 1 km and 5 min,
respectively. The radar (Nimrod) data have been quality-controlled by the UK Met Office following
the correction techniques proposed by Harrison et al. (2009) to account for all the errors inherent to
radar rainfall measurements.
4.3

Local Monitoring System: raingauges and level gauges

A real time accessible monitoring system is installed covering this catchment since April 2010. It
includes three tipping bucket rain gauges, one pressure sensor for monitoring water levels at the

Roding River (downstream boundary condition of the catchment), two sensors for water depth
measurement in sewers and one sensor for water depth measurement in open channels (Figure 1(b)).
4.4

Hydrological/Hydraulic Model of the Study Area

The focus of our work is on urban pluvial flooding, which, as was mentioned before, is one of the
major issues in the Cranbrook catchment. For this reason, a pluvial flood model was implemented for
the study area. The model is a dual-drainage, physically-based one and was set up in
InfoWorks CS 10.5. In this model the urban surface was modelled in 2D (2-dimensions), using a
triangular mesh. The model of the surface was coupled with a 1D (1-dimensional) model of the sewer
system (Figure 1(c)) and the interactions between the two models take place at the manholes. The
implemented model was calibrated using the rainfall and water level measurements collected through
the local monitoring system (Section 4.3).

Beal High School

Valentine Sewer Chadwell Heath Foundation School

Ursuline High School

Rain gauge
Level gauge in sewer
Level gauge in open channel
Pressure sensor for river level

(a)
(b)
(c)
Figure 1. Cranbrook catchment (a) location of the catchment in relation to radars and the Roding River
catchment; (b) monitoring system; (c) sewer network
4.5

Rainfall events selected for testing of gauge-based adjustment methods

Four rainfall events occurring between August 2010 and January 2012 were selected to test the
gauge-based adjustment methods. The dates and statistics of these events are summarised in Table 1.
In this table RG Total is the mean raingauge accumulation, Radar@RG Total is the co-located
mean radar rainfall accumulation, Radar Total is the mean radar accumulation over the whole
catchment and Peak Flow Depth corresponds to the maximum flow depth recorded in the Valentine
Sewer (located in the mid-stream section of the catchment, see Figure 1(b)) for each event.
Table 1. Statistics of rainfall events selected for testing of adjustment methods.
Date

Duration
(h)

RG Total
(mm)

Radar@RG
Total (mm)

Radar Total
(mm)

Peak Flow Depth

(m)

23/08/2010

23.53

7.29

6.80

0.633

26/05/2011

15.53

5.10

4.88

0.672

05-06/06/2011

24

20.87

9.43

9.48

0.346

03/01/2012

13

8.93

7.72

7.55

0.547

RESULTS AND DISCUSSION

As can be seen from Table 1, large and event-varying bias between raingauge and radar measurements
were observed in the four rainfall events that were selected for testing. In order to correct this, the two
8

adjustment methods described in Section 3 were applied to each rainfall event. Due to space
limitations, only the results associated with the event on 23/08/2010 will be presented. However,
similar results were obtained for all events.
In this section, the corrected radar estimates obtained with the mean bias reduction method are referred
to as Corrected Radar 1 km and the results of the error variance minimisation method are referred to
as Bayesian Radar 1 km. Raingauge measurements are usually denoted RG and the original radar
(Nimrod) estimates are referred to as Radar 1 km.
Figure 2 shows the results for the entire Cranbrook catchment for the 23/08/2010 event: i.e. it shows
mean values of raingauge, radar and adjusted radar estimates, as well as the associated hydraulic
results. Figure 3 shows the 5-min rainfall profiles and accumulations of the adjusted radar estimates,
the coincidental raingauge records and the original radar rainfall estimates at the location of one of the
raingauge sites (Chadwell Heath Foundation School) for the 23/08/2010 event. Similar results were
obtained for the other 2 raingauge sites, but these are omitted due to space limitations.
From Figures 2 and 3 it can be seen that radar rainfall rates and rainfall accumulations, both at a
specific raingauge location as well as for the entire catchment, were largely improved by both
adjustment methods. In terms of total rainfall accumulation (see Figures 2(b) and 3(b)), the Corrected
Radar 1 km produced slightly better results than the Bayesian Radar 1 km. For the rainfall profiles,
however, the Bayesian method produced significantly better results than the mean bias one. For
example, in Figures 2(c) and 3(a) some underestimation (e.g. around 00:05 - 02:05), overestimation
(e.g. around 03:05 - 04:05) and faulty timing of rainfall peaks can be observed in the rainfall profiles
of the Corrected Radar 1 km, as compared to the RG profiles. In contrast, the profiles of the
Bayesian Radar 1 km fit the RG profiles significantly better. Moreover, in Figures 2(b) and 3(b) it
can be noted that the shape of the cumulative rainfall for the Bayesian Radar 1 km is very similar to
that of the RGs, as opposed to the shape produced by the mean bias corrected estimates (Corrected
Radar 1 km). This conclusion is further strengthened by the q-q plot of Figure 1(a): it can be seen that
the Bayesian Radar 1 km estimates provide a better fit to the RG observations, particularly for high
rain rates (it can be noted that the markers of Bayesian Radar 1 km estimates are more concentrated
around and closer to the straight line with slope equal to 1, as compared to the Corrected Radar 1 km
estimates). The faulty reproduction of rainfall profiles by the mean bias adjusted estimates is due to
the fact that this adjustment method fully relies on correcting the accumulated difference between
radar and raingauge measurements over the entire rainfall event, without taking into account the
temporal variation within a storm process.
In addition to the rainfall (temporal) profiles, Figure 2(d) also demonstrates that the spatial structures
of rainfall fields are largely altered by simply multiplying a given constant (i.e. the sample mean bias)
to the original radar rainfall fields. In contrast, it can be seen that the spatial structure of the rainfall
field is preserved when the Bayesian adjustment technique is applied. As previously mentioned, the
ability to reflect the spatial variability of a rainfall field is one of the main advantages of radars and it
is desirable to retain it. After carrying out the rainfall adjustment, the different rainfall estimates were
applied to the dual-drainage model of the Cranbrook catchment. The associated flow levels in one of
the sewers (located in the mid-stream part of the catchment) are shown in Figure 2(c). As can be seen,
the hydraulic outputs obtained with the adjusted radar measurements are quantitatively much more
similar to the RG outputs and to the observed water levels, as compared to the outputs resulted from
the radar rainfall estimates before adjustment. This demonstrates the predominant role of rainfall mean
bias in hydrological and hydraulic modelling. However, it can also be observed that, as compared to
the hydraulic outputs of the Corrected Radar 1 km, the outputs of the Bayesian Radar 1 km show
better agreement with the RG outputs and with the flow level measurements, particularly regarding the
timing and magnitude of flow level peaks. It is mainly due to the better reproduction of rainfall
9

profiles (both in quantity and geometry) achieved with the Bayesian adjustment method. These results
suggest that, in addition to rainfall mean bias, the spatial and temporal variability of rainfall is also an
important factor which has a significant impact on the associated urban hydrological/hydraulic
applications.

(a)

(b)
Pipe 463.1 (Mid-stream)

1.6

1.4

15

Radar 1km

Corrected Radar 1km

0.8

Bayesian Radar 1km

Obs. 463.1(Mid-Stream)

20

0.6

Flow Depth (m)

Rain (mm/hr)

1.2
RGs

10

0.4
25

0.2

30
0005

0
0105

0205

0305

0405
0505
0605
23 August 2010 (Time, GMT)

0705

0805

0905

(c)
Spatial Variability: 23/08/2010 event
6
Radar 1km

Bayesian Radar 1km

Standard Deviation

Corrected Radar 1km

0
00:05

01:05

02:05

03:05

04:05
Time (5 min)

05:05

06:05

07:05

(d)
Figure 2. Results for the entire Cranbrook catchment for the 23/08/2010 event: (a) Quantile-quantile
(q-q) comparison of mean raingauge, radar and adjusted radar rainfall estimates (mm) per 5-min time
step; (b) q-q comparison of cumulative mean raingauge, radar and adjusted radar rainfall estimates
(mm); (c) Mean raingauge, radar and adjusted radar rainfall profiles and associated hydraulic outputs,

10

as represented by water depth at the Valentine Sewer Pipe 463.1 of the hydraulic model; (d) Spatial
variability as represented by the standard deviation of the original and adjusted radar rainfall fields.

Figure 3. Comparison of 5-min rainfall profiles and accumulations for the Chadwell Heath Foundation
School raingauge and the coincidental original and adjusted radar estimates 23/08/2010 event.

CONCLUSIONS AND FUTURE WORK

In this work, a detailed review of state-of-the-art gauged-based radar rainfall adjustment techniques
was firstly conducted with the purpose of analysing their theoretical foundation. In general, rainfall
adjustment techniques can be classified into two types: (i) mean bias reduction techniques and (ii)
error variance minimisation techniques. Moreover, the existing techniques have mainly been applied at
large scales and on hourly or daily basis; however, their suitability for smaller spatial and temporal
scales (i.e. for urban applications) has not been fully analysed.
After this review, one technique of each type was selected and tested in a small urban catchment (865
ha) in North-East London. The radar rainfall estimates of four historical events occurring between
August 2010 and January 2012 were adjusted using point rainfall measurements recorded by three in
situ raingauges. The adjusted rainfall estimates were applied to the physically-based dual-drainage
model of the catchment and the associated outputs were compared to flow level records in addition to
the outputs resulted from raingauge measurements and the original radar data.
In these case studies, the method based upon error variance minimisation performed better, as it not
only reduced mean bias, but it managed to correctly reproduce the spatial and temporal variability of
rainfall, which proved to have a significant impact on the associated hydraulic outputs. These results
suggest that error variance minimisation methods may be more appropriate for small scale urban
hydrological/hydraulic applications. However, before further conclusions can be drawn, more work
should be done to deeply understand the impact of parameters such as catchment size and geometry,
raingauge network size, density and relative location. Moreover, the feasibility of using these
techniques in real-time and its potential benefits for rainfall and flood forecast should be further
studied.

ACKNOWLEDGEMENTS

The authors would like to thank Dr Cinzia Mazzetti and Prof Ezzio Todini, from Progea and the
University of Bologna (Italy), for making freely available to us the RAINMUSIC software package for
11

meteorological data processing and data interpolation. The authors would also like to thank the UK
Met Office for providing the radar data and Innovyze for providing the InfoWorks CS software. The
first author would like to acknowledge the support of the European Unions Interreg IVB NWE
Programme to the RainGain project (PSC8-236H), of which this research is part.

REFERENCES

Anagnostou, E. N. and Krajewski, W. F. (1999). Rea-time radar rainfall estimation. Part I: algorithm
formulation. Journal of Atmospheric and Oceanic Technology, 16, 189-197.
Cole, S. J. and Moore, R. J. (2008). Hydrological modelling using raingauge- and radar-based
estimators of areal rainfall. Journal of Hydrology, 358, 159-181.
Collier, C. G. (2009). On the propagation of uncertainty in weather radar estimates of rainfall through
hydrological models. Meteorological Applications, 16(1), 35-40.
Ehret, U., Gotzinger, J., Bardossy, A. and Pegram, G. G. S. (2008). Radar based flood forecasting in
small catchments, exemplified by the Goldersbach catchment, Germany. International Journal
of River Basin Management, 6(4), 323-329.
Einfalt, T., Arnbjerg-Nielsen, K., Golz, C., Jensen, N.-E., Quirmbach, M., Vaes, G. and Vieux, B.
(2004). Towards a roadmap for use of radar rainfall data in urban drainage. Journal of
Hydrology, 299(3-4), 186-202.
Einfalt, T., Jessen, M. and Mehlig, B. (2005). Comparison of radar and raingauge measurements
during heavy rainfall. Water Science and Technology, 51(2), 195-201.
Fulton, R. A., Breidenbach, J. P., Dong-Jun, S. and Miller, D. A. (1998). The WSR-88D rainfall
algorithm. Weather and Forecasting, 13(2), 377-395.
Gerstner, E.-M. and Heinemann, G. (2008). Real-time areal precipitation determination from radar by
means of statistical objective analysis. Journal of Hydrology, 352, 296-308.
Golding, B. W. (2009). Uncertainty propagation in a London flood simulation. Journal of Flood Risk
Management, 2(1), 2-15.
Harrison, D. L., Scovell, R. W. and Kitchen, M. (2009). High-resolution precipitation estimates for
hydrological uses. Proceedings of the Institution of Civil Engineers: Water Management,
162(2), 125-135.
Kalman, R. E. (1960). A new approach to linear filtering and prediction problems. Transactions of the
ASME-Journal of Basic Engineering, 82(D), 35-45.
Krajewski, W. F. (1987). Cokriging radar-rainfall and rain gage data. Journal of Geophysical
Research, 92(D8), 9571-9580.
Liguori, S., Rico-Ramirez, M. A., Schellart, A. N. A. and Saul, A. J. (2011). Using probabilistic radar
rainfall nowcasts and NWP forecasts for flow prediction in urban catchments. Atmospheric
Research, 103, 80-95.
Looper, J. P. and Vieux, B. E. (2012) An assessment of distributed flash flood forecasting accuracy
using radar and rain gauge input for a physics-based distributed hydrologic model. Journal of
Hydrology, 412-413(0), 114-132.
Mazzetti, C. (2004). Report on performance of the new methodology. Deliverable 7.4, MUSIC Project
(Multiple Sensor Precipitation Measurements, Integration, Calibration and Flood Forecasting),
European Commission.

12

Moore, R. J., Watson, B. C., Jones, D. A. and Black, K. B. (1989). London Weather Radar Local
Calibration Study: Final Report, Contract report prepared for the National Rivers Authority
Thames Region, NERC/Institute of Hydrology, Wallingford, UK.
Seo, D.-J., Breidenbach, J. P. and Johnson, E. R. (1999). Real-time estimation of mean field bias in
radar rainfall data. Journal of Hydrology, 223, 131-147.
Smith, J. A., Baeck, M. L., Meierdiercks, K. L., Miller, A. J. and Krajewski, W. F. (2007). Radar
rainfall estimation for flash flood forecasting in small urban watersheds. Advances in Water
Resources, 30(10), 2087-2097.
Syed, K. H., Goodrich, D. C., Myers, D. E. and Sorooshian, S. (2003). Spatial characteristics of
thunderstorm rainfall fields and their relation to runoff. Journal of Hydrology, 271(1-4), 1-21.
Tabios, G. Q. and Salas, J. D. (1985). A comparative analysis of techniques for spatial interpolation of
precipitation. Journal of the American Water Resources Association, 21(3), 365-380.
Todini, E. (2001). A Bayesian technique for conditioning radar precipitation estimates to rain-gauge
measurements. Hydrology and Earth System Sciences, 5(2), 187-199.
Vieux, B. E. and Bedient, P. B. (2004). Assessing urban hydrologic prediction accuracy through event
reconstruction. Journal of Hydrology, 299(3-4), 217-236.
Vieux, B. E. and Imgarten, J. M. (2011). On the scale-dependent propagation of hydrologic
uncertainty using high-resolution X-band radar rainfall estimates. Atmospheric Research,
103(0), 96-105.
Wheater, H., Chandler, R., Onof, C., Isham, V., Bellone, E., Yang, C., Lekkas, D., Lourmas, G. and
Segond, M. L. (2005). Spatial-temporal rainfall modelling for flood risk estimation. Stochastic
Environmental Research and Risk Assessment, 19(6), 403-416.
Wood, S. J., Jones, D. A. and Moore, R. J. (2000). Static and dynamic calibration of radar data for
hydrological use. Hydrology and Earth System Sciences, 4(4), 545-554.
Yang, C., Chandler, R. E., Isham, V. S. and Wheater, H. S. (2005). Spatial-temporal rainfall
simulation using generalised linear models. Water Resources Research, 41, 1-13.
Zeng R. J., Lemaire R., Yuan Z. and Keller J. (2004). A novel wastewater treatment process:
simultaneous nitrification, denitrification and phosphorus removal. Water Science and
Technology, 50(10), 163-170.

13

9th International Conference on Urban Drainage Modelling

Belgrade 2012

Improving urban drainage modelling with path-average

rainfall from telecommunication microwave links
Martin Fencl1, Jrg Rieckermann2, David Strnsk1, Vojtch Bare1
1

Czech Technical University in Prague, Department of Sanitary and Ecological Engineering, Thkurova 7, Praha
Swiss Federal Institute of Aquatic Science and Technology (Eawag) and Swiss Federal Institute of Technology
(ETH), 8600 Dbendorf, Switzerland
2

ABSTRACT
Incomplete knowledge about spatio-temporal rainfall dynamics causes input
uncertainty in rainfallrunoff modelling. This is especially critical in urban areas,
where subcatchments are small and runoff is generated extremely fast on the
impervious areas. Networks of telecommunication microwave links (MWL) are
very dense in urban areas and therefore could provide novel rainfall information
which has the potential to reduce input uncertainty in urban drainage modelling. In
this study, we therefore investigate how the better information on spatio-temporal
rainfall variability from MWL observations improves pipe flow predictions.
Specifically, we perform numerical experiments with virtual rainfall fields and
compare the results of MWL rainfall reconstructions to those of rain gauge
observations. For a case study of a suburb in Prague, Czech Republic, we are able
to show that MWL networks in urban areas are sufficiently dense to provide good
information on spatio-temporal rainfall variability. Total rainfall volumes are
reproduced very well, with errors of about 6 %. Although peak rainfall intensities
are still systematically underestimated by approximately 30%, this clearly
outperforms rain gauge observations. Also, we find that peak flows from MWL
observations are only biased by 6 %, whereas rain gauge observations cause a bias
of 16 %. In our study, we did not include effects of uncertainties in both MWL and
rain gauge measurements, which, arguably, can be high. Nevertheless, MWL
networks could provide hundreds of rainfall sensors, which would be available at
virtually no cost. As demonstrated, MWLs can significantly reduce input
uncertainties in future rainfall-runoff modeling and thus improve discharge
predictions.

KEYWORDS
Rainfall estimation, rainfall spatial dynamics, telecommunication microwave links, urban drainage
modelling

INTRODUCTION

Commercial microwave links from telecommunication networks (MWL) are new source of rainfall
information which has a potential to improve incomplete knowledge about spatio-temporal rainfall
dynamics especially in urban areas (Messer et al., 2006). Traditional rain gauges provide data with
sufficient temporal resolution, but usually cannot reflect the spatial variability of rainfalls. Similarly,
weather radar data provides insight into the spatial distribution of precipitation, but only have limited
accuracy for urban rainfall monitoring. For example, in Czech Republic, operational radar observes
precipitation with a resolution of about 2.5 x 2.5 km2 every 5 min. Recent Local Area Weather Radars,
which can provide higher spatial resolution of rainfall, are not always available. In addition they are
potentially affected by significant sources of error (Thorndahl, 2011).
Similar to weather radars, MWL operate in the microwave range, where the rain drops are the major
source of signal attenuation. Therefore, rain-induced MWL attenuation can be used to compute pathaveraged rainfall intensities using a simple power law model (Berne and Uijlenhoet, 2007). Using
MWL in urban drainage modeling is conceptually interesting, because MWL networks i) are already
built and could provide rainfall information at virtually no additional cost, ii) observe near-surface
precipitation a few tens of meters above ground, and iii) have a high density in urban areas
(Rieckermann et al., 2009). In addition, MWL provide path-average rain intensities, which, given
typical length scales from a few hundred meters to a few kilometers, should conceptually correspond
very well to the scale of urban subcatchments.
In this manuscript, we therefore investigate for the first time how data from commercial
telecommunication networks can improve urban drainage modeling. Specifically, we analyze in how
far the better information about spatio-temporal rainfall variability improves pipe flow predictions.
Our analysis for a suburb of Prague, Czech Republic, shows that MWL networks in urban areas are
sufficiently dense to provide good information on spatio-temporal rainfall variability. Although peak
rainfall intensities are still systematically underestimated, we find that MWL rainfall estimates predict
pipe flows more accurately than point rainfall measurements.

METHODS AND MATERIAL

To assess the potential of MWL in urban drainage modeling, we investigate how runoff predictions
using MWL rainfall improve in comparison to those using rain gauges. As proposed previously, we
here present the results based on realistic numerical experiments. First, we compute reference rainfall
fields and, second, extract point intensities as measured by a rain gauge. Third, we compute pathaveraged rainfall intensities and reconstruct space-time rainfall as seen by a MWL network. These
three datasets are then propagated through a hydrodynamic model to compare the influence of
different rainfall monitoring techniques on predicted pipe flows. As we only investigate input
uncertainties due to spatial rainfall variability, we disregard any other uncertainties.
2.1

Reconstruction of space-time rainfall

Reference rainfall fields: The reference areal rainfall intensities are simulated using a virtual drop
size distribution (DSD) generator (Schleiss et al., 2012). The simulator estimates the medium and
large scale rainfall variability (1-50 km) and advection direction and velocity using radar data. The
small scale variability (0.1-1 km) of the DSD is parameterized based on disdrometer data.
Rain gauge observations: Virtual rain gauge measurements are extracted from the reference rainfall
at one particular cell of a rainfall field.

MWL rainfall reconstruction: The rain-induced MWL attenuation contains information on the pathaveraged rainfall intensity. As a typical network contains MWLs of different lengths and orientations,
the two dimensional rainfall spatial variability can be reconstructed to some extent from joint analysis
of nearby MWL data. For simplicity, we used the algorithm by Goldshtein et al., (2009) which is
straightforward to implement: Each i-th MWL is divided into Ki equal subsections of approximately
0.5 km. Each subsection is substituted by a data point Mj, located at its centre. Each link is thereafter
represented as a set of Ki data points (Figure 1). The mean rainfall intensity of all points along a link
has to correspond to the path-averaged intensity:

jMWLi

Li

= Ri

(1)

The rainfall distribution between points (Mj, Mj+1, ..., Mj+ki) representing i-th link is approximated from
data points belonging to neighboring links as follows (Figure 1):

j =

(r l
l

kMWLi

k
kMWLi

)
(2)

where j is the rainfall estimate at Mj of the i-th link, rk is a rainfall at each neighboring point Mk which
does not belong to the i-th link and lk is a distance of point Mj to the neighboring point Mk. The
distribution of rainfall intensity along the i-th link is thereafter corrected by minimizing the cost
function:

F=

(r

jMWLi

j )2

(3)

under the condition (1). Here, the procedure is applied to each link and iterated 20 times as suggested
by Goldshtein, (2009). To obtain two dimensional reconstructed rainfall field the iterated rainfall
intensities at data points Mi are transformed to the regular grid. The rainfall intensity of particular grid
point is calculated as a mean of intensities of data points weighted by square distance between grid
point and respective data point Mi, analogically to (2).
To propagate the three sets of rainfall information through a hydrodynamic model, each
subcatchments in the model requires the input of a specific rainfall time series. For our case study, the
areal rainfall intensity over each subcatchment is therefore calculated by averaging the corresponding
cells of the rainfall field proportionally to the area of intersection.
2.2

Urban rainfall runoff modelling

To investigate the potential of MWL information to better predict pipe flows, we use a standard
1D-hydrodynamic model. Details on the surface runoff and pipe flow modules are given below.
2.3

Performance assessment

To compare the different monitoring techniques, we compute relevant performance statistics of

rainfall reconstruction and pipe flow in comparison to the reference rainfall case. For each rainfall
event and input data set, we compute a) the peak intensity (Rmax), and b) the rainfall volume (RV),
considering only the rainfall cells restricted by the catchment area. From the corresponding runoff
hydrograph we compute c) outflow volume (QV) and d) peak flow at the catchment outlet (Qmax).

z
Z=0

=0

z=1

z = 20

Figure 1. Rainfall reconstruction from MWL data. Left: Initial distribution of a rainfall between data
points representing particular links. Middle: Distribution of a rainfall along links after first iteration of
a rainfall intensity at the first link. Right: The reconstructed rainfall distribution along the links after
last iteration (z).
We use the relative error to compare estimated values of respective magnitude with its reference value.
The mean value of the relative error represents the bias and its standard deviation the uncertainty due
to limited spatial information of each measuring technique.
2.4

Cases study Urban Catchment in Prague, Czech Republic

The case study area is located in a suburb of Prague. The investigated catchment has an area of
2.33 km2, with an impervious area of about 64 %, which is drained by a separate sewer system. The
Prague urban area is covered by a dense network of many hundred MWL. We selected 29 MWLs
owned by T-Mobile using Ericsson MINI-LINK platform which are located in the direct vicinity of
the catchment (see below).
The reference rainfall fields are sampled every 5 minutes and have a spatial resolution of
0.1 x 0.1 km2. The original size of a rainfall field is 40 x 40 km2. As the response of the catchment
fundamentally depends on the rainfall characteristics, we generated three convective rainfall events
with different storm velocities and intermittencies: first of advection velocity 10 m/s to the North-East
intermittency of 55 %, second of advection velocity 7.7 m/s to the South-East and intermittency of
50 %, third of advection velocity 5.5 m/s to the North East and intermittency of 70 %. The duration of
each event is 50 minutes. To eliminate the influence of positioning of the rainfall field over the study
area, the relative position of the catchment to the rainfall fields was repeatedly changed to cover 25
different locations uniformly distributed over the field. This finally resulted in a comprehensive set of
75 reference areal rainfalls of a size of 7 x 7 km2. From these, we computed rain gauge data and MWL
reconstructions as follows.
For the rain gauge information, we only used one rain gauge, as this would correspond to the
information in a typical consultant study. Also, the available hydrodynamic model was calibrated
based on a single gauge.
In the study area, we selected a network of 29 MWLs with a total length of 54.3 km (Figure 2). For the
MWL rainfall reconstruction analysis, we first chose all links which intersect with the catchment
boundary polygon. Then, some links are filtered out to obtain a more regular structure of the MWL
network and improve the coverage of the subcatchment. We found that this improves the
reconstruction, because a) links which intersected the catchment boundary polygon only by a small
4

proportion spun the mass of the rainfall outside of the polygon and b) links of a similar length and
orientation biased the iteration of a rainfall distribution along the path of links (2).(Figure 2). After
filtering, the remaining 15 MWL have a median length of about 1 km. 10 operate at 38 GHz, one pair
on 32 GHz and one pair on 26 GHz and one link at 23 GHz.
The rainfall reconstruction algorithm splits the 15 MWL into 39 data points, from which three links
are represented by a single point only. The longest link is subdivided into 5 points. The number of
iterations is set to 20. The reconstructed rainfall field has a size of 7 x 7 km2 and spatial resolution of
0.25 x 0.25 km2.

-1038000

-1036000

-1039000

-1038000
-1040000

MWL data points

rain gauge

-1037000

-1034000

MWLs used
MWLs filtered out
rain gauge

-1036000

As described above, we use a standard hydrodynamic model for the rainfall-runoff simulations. It has
been implemented in the commercial solver MIKE URBAN with computational engine MOUSE
(MOUSE, 2009). The model is owned by the municipality of Prague and was constructed for the
general drainage masterplan for Prague. The model has been carefully calibrated on data from a rain
gauge and an ultrasonic Doppler flow meter, which have been collected from April 2006 to June 2006.
The model comprises 188 subcatchments with a median area of 0.34 ha and has 517 manholes with
pipe length of around 18.2 km. Diameters ranges from 0.3 to 1.6 m. The surface runoff module uses
the simple time-area method (Type A). From this, pipe flows are computed by the dynamic
Saint-Venant equations. The routing time step varies between 5 and 10 seconds.

-739000

-737000

-735000

-733000

-738000

-737000

-736000

-735000

Figure 2. Study catchment and MWL network (left), the links displayed by the solid line were used for
the rainfall spatial reconstruction. Right: Data points representing links used in the spatial
reconstruction algorithm (right).
To guarantee a realistic evaluation, only those events were considered in performance assessment,
where the reference peak flow (Qmax,ref) exceeded 10 l/s, because those flows are relevant from an
engineering viewpoint. The events with an estimated peak flow (Qmax,est) lower than 5 l/s were also not
included, since the runoff is very sensitive to small changes in model parameters (e.g., pipe roughness
coefficients) and discharge predictions at such low flows are not robust.

RESULTS

In general, we found that the ability to predict runoff dynamics of a storm event depends on the
estimation of total rainfall volume, its spatial stratification and temporal dynamics.
Regarding the spatio-temporal characteristics of the rainfall fields, we found that the MWL
reconstruction in general correctly reproduces the location of peak rainfall intensities. However, peak
intensities are generally averaged and therefore underestimated (Figure 3). In contrast, low intensities
are overestimated. In comparison to reference rainfall volumes, RVMWL are systematically
overestimated by only 6 %, with a standard deviation of 38 percent points (Table 1). On average,
Rmax MWL is underestimated by 31 % (17 %). In our view, this is a very good result, considering that we
deliberately chose convective events with a rather high intermittency (Table 1). In contrast, RVRG are
systematically underestimated by 9 %. The lack of information about spatial variability of rainfall
shows in the peak intensity estimates, Rmax RG,, which underestimate the reference values by almost
50%.
Regarding the performance to predict sewer discharges, the threshold for evaluating the rainfall
induced flows (Qmax ref> 10 l/s and Qmax est> 5 l/s) was exceeded by 33 rain events. For these, we found
that the MWL-based flow predictions match the runoff from reference rainfall considerably better than
those from rain gauge data. (Table 2).
Table 1.
Performance statistic of rainfall reconstruction in comparison to the reference rainfall
(standard deviation is given in brackets)
RV

Rmax

mean rel. error

Mean rel. error

RG

-9 % (32 %)

48 % (25 %)

MWL

6 % (38 %)

-31 % (17 %)

10

10

R [mm/h]

R [mm/h]

0
-1034000

-738000
-736000

-734000

-1038000

-1034000

-738000

-1036000

-736000

-734000

-1036000
-1038000

Figure 3. Example of ref. rainfall (left) and MWL reconstruction (right) - Event No. 26, t = 06:35.

Table 2.

Statistical comparison of estimated peak flows and flow volumes

RG

25 % (115 %)

16 % (85 %)

MWL

-2 % (38 %)

-6 % (23 %)

50

200

1000

RG
MWL

10

50

200

QV rg ; QV mwl [m^3]

1000

RG
MWL

10

Qmax rg ; Qmax mwl [l/s]

5000

Qmax
mean rel. error

5000

QV
mean rel. error

10

50

200

1000

5000

10

50

Qmax ref [l/s]

200

1000

5000

QV ref [m^3]

Figure 4. Scatter in peak flow estimates (left) and absolute outflow volume estimates (right)

100
0

rel. error in VQ [%]

200

RG
MWL
mean RG
mean MWL

-200

-100

100
0
-100
-200

rel. error in Qmax [%]

200

RG
MWL
mean RG
mean MWL

500

1000

1500

2000

2500

Qmax ref [l/s]

1000

2000

3000

4000

VQ ref [m^3]

Figure 5. Relative peak flow error for different reference peak flows (left). Relative outflow volume
error for different outflow volumes (right)
The outflow volume from MWL rainfall measurements is biased only off by -2 %, with a standard
deviation of 38 percent points (Table 2). In contrast, the rain gauge overestimates the runoff volume
by 25 %, with a much higher standard deviation. Similarly, the peak flows are only biased by 6 %
using MWL fields, whereas rain gauge observations cause a bias of 16 %. Interestingly, the highest
relative errors occur for both MWL and rain gauge data at low reference discharges (Figure 4,
Figure 5). In summary, despite slight systematic under- and overestimation of rainfall fields, MWL
observations considerably reduce input errors in hydrodynamic modeling, which leads to more
accurate pipe flow predictions in comparison to traditional rain gauge observations (Figure 6).

MWL
RG

150
0

50

100

Q [l/s]

200

250

ref

06:30

07:00

07:30

08:00

time [hh:mm]

Figure 6. Example of the sewer hydrograph resulting from rain event No. 26 fromthe reference rainfall
(solid), MWL reconstruction (dashed) and rain gauge observations (dotted). The better spatial
resolution of the MWL data visibly reduces the input uncertainty due to incomplete rainfall
observations.

DISCUSSION

The time-area method used for rainfall-runoff transformation smoothes the temporal variability of
rainfall fallen on a particular subcatchment. This corresponds to the behavior of real-world
subcatchments, where larger subcatchments lead to comparably smoother hydrographs, because the
rainfall-runoff dynamic is averaged over the time. In addition, runoff dynamics are further smoothed
during wave routing in the drainage system. Therefore, the averaging of peaks in reconstructed rainfall
fields significantly influences the accuracy of pipe flow predictions.
Since a rain gauge measures the rainfall only at one point it easily misses the rainfall intensity peak. In
contrast to peak averaging, this systematically underestimates the total rainfall volume and,
consequently, has a much greater impact on the predicted rainfall-runoff. The uncertainty in the rain
gauge measurement is, however, enhanced in our study also by the coarse temporal sampling of the
rainfall fields. Since there is a relationship between spatial and temporal rainfall dynamics (Schleiss et
al., 2012) the monitoring techniques which reflect the spatial variability do not require such a high
sampling frequency as point measurements. This is, because discontinuities in temporal rainfall
dynamics are compensated by its spatial dynamics. Consequently, the rain gauge cannot compensate
the loss of information about temporal rainfall dynamics by rainfall spatial information.
While our results are based on virtual data, MWL rainfall estimates from real power level
measurements contain additional uncertainties. The most significant errors are caused by the baseline
determination, quantization noise, wet antenna effect and model structure deficits of the power law
rainfall-attenuation relation (Zinevich et al, 2010; Fencl, 2011). In general, the uncertainty in MWL
rainfall estimation grows with the rainfall intensity because of the rainfall-attenuation power law
relationship. This negative effect is reduced for frequencies around 30 GHz, where the attenuationrainfall relation is almost linear (Berne and Uijlenhoet, 2007). Although MWL reconstructions under
real world conditions will always be less accurate than reported here, the uncertainties could be
minimized i) by combining the MWL information with rain gauge and weather radar information ii)
by implementing a more sophisticated reconstruction algorithm.
8

Also, it is well known that traditional rain gauges are also affected by uncertainties. While they can
measure moderate rainfalls very accurately, heavy and very heavy rainfall observations are often
systematically underestimated. Especially for tipping bucket gauges, this can be partly filtered out by
proper calibration (Stransky et al., 2007).
Future research is needed to investigate how the uncertainty is related to the link length, which, as
shown, governs the averaging and location detection of peak rainfall intensities. Based on our results,
we suggest to further improve rainfall reconstruction algorithms. This should include a better selection
of data points and weights for the rainfall intensity iterations (Eq. 2), as well as better consideration of
the redundant information from parallel links and the irregular ray-like structure of a MWL network.
The rainfall reconstruction technique used in this study also reconstructs the rainfall fields separately.
More sophisticated methods should take advantage of the spatio-temporal correlation of different
observations (Zinevich et al., 2008) and use a rain field model, possibly in combination with radar
data. In addition, this approach enables short term forecasting which can find a great use in drainage
system real time control applications.

CONCLUSION

In our study, we found that the better information from Telecommunication Microwave Links about
spatio-temporal rainfall variability has the potential to improve pipe flow predictions compared to
those based on traditional rain gauge observations. Our results show, first, that MWL rainfall
reconstruction underestimates extreme rainfall peak intensities, but very well reproduces areal
averaged rainfall intensities. Second, for urban drainage applications, we found that the averaging of
rainfall peaks over a larger area from MWL observations does not influence the runoff dynamics as
significantly as missed rainfall peaks by rain gauges. Rain gauges often miss peak intensities of a
storm cell, especially for convective high-intensity events and, here, generally produced too low
runoff. These results do not yet include effects of uncertainties in both MWL and rain gauge
measurements, which can be very high. Especially for short MWL, wet antenna and quantization
effects can lead to biased and imprecise observations. On the other hand, we found that MWL
networks can be extremely dense in urban areas, where many rainfall sensors could be available at
virtually no cost. Thus, considering the redundant information from many links seems promising to
improve the accuracy of MWL observations. In the future, the MWLs have potential to complement
rain gauge point measurements with the missing spatial rainfall information. Thus, they can reduce
input uncertainties in rainfall-runoff modeling and improve discharge predictions.

ACKNOWLEDMENTS

This work was supported by the project of Czech Technical University in Prague project no.
SGS12/045/OHK1/1T/11. Further, we would like to thank T-Mobile Czech Republic a.s. for kindly
providing us with information on the MWL network. The Prazska Vodohospodarska spolecnost, a. s.
is acknowledged for providing us with their hydrodynamic model. People from Veolia Voda, a.s. were
very helpful in selecting the appropriate case study area. We would also like to thank the employees of
Hydroprojekt, a. s. and DHI, a. s. for consulting regarding the rainfall-runoff model. Last but not least
we thank Marc Schleiss, EPFL, Lausanne, for providing us with DSD fields for the numerical
experiments.

REFERENCES

Berne A., Uijlenhoet R. (2007). Path-averaged rainfall estimation using microwave links: Uncertainty
due to spatial rainfall variability. Geophys. Res. Lett. 34(7).
Fencl M. (2011). Reducing the uncertainty in rainfall-runoff modelling using commercial microwave
links. Master's Thesis. Department of Sanitary and Ecological Engineering, Czech Technical
University in Prague, Czech Republic.
Goldshtein O., Messer H. Zinevich A. (2009). Rain rate estimation using measurements from
commercial telecommunications links. Signal Processing, IEEE Transactions on 57, 1616
1625.
Messer, H., Zinevich, A., Alpert, P. (2006). Environmental Monitoring by Wireless Communication
Networks. Science 312, 713713.
MOUSE (2009). User Guide. DHI. http://www.hydroasia.org/jahia/webdav/site/hydroasia/shared/
Document_public/Project/Manuals/US/MOUSE_UserGuide.pdf (accessed 7 April 2012)
Rieckermann J., Lscher R. and Krmer S. (2009). Assessing Urban Precipitation using Radio Signals
from a Commercial Communication Network, 8th International Workshop on Precipitation in
Urban Areas, 10-13 December, 2009, St. Moritz, Switzerland..
Schleiss, M., J. Jaffrain and A. Berne (2012), Stochastic simulation of intermittent DSD fields in time,
J. Hydrometeorol., vol.13, No.2, 621-637.
Stransky D., Bares V., and Fatka P. (2007). The effect of rainfall measurement uncertainties on
rainfall-runoff processes modelling. Water science and technology, 55 (4), 103111.
Thorndahl, S., Rasmussen, M.R. (2011), Marine X-band weather radar data calibration, Atmospheric
Research, vol. 103, 33-44.
Zinevich, A., Alpert, P., Messer, H. (2008). Estimation of rainfall fields using commercial microwave
communication networks of variable density. Adv. Water Resour. 31, 14701480.
Zinevich A., Messer H. and Alpert P. (2010). Prediction of rainfall intensity measurement errors using
commercial microwave communication links. Atmospheric Measurement Techniques 3, 1385
1402.

10

9th International Conference on Urban Drainage Modelling

Belgrade 2012

State-space adjustment of radar rainfall and stochastic

flow forecasting for use in real-time control of urban
drainage systems
Roland Lwe1 , Peter Steen Mikkelsen2, Michael R. Rasmussen3, Henrik
Madsen4
1

Department of Informatics and Mathematical Modelling, Technical University of Denmark (DTU), Denmark,
rolo@imm.dtu.dk
2
Department of Environmental Engineering, Technical University of Denmark (DTU), Denmark,
psmi@env.dtu.dk
3
Department of Civil Engineering, Aalborg University, Denmark, mr@civil.aau.dk
4
Department of Informatics and Mathematical Modelling, Technical University of Denmark (DTU), Denmark,
hm@imm.dtu.dk

ABSTRACT
Merging of radar rainfall data with rain gauge measurements is a common
approach to overcome problems in deriving rain intensities from radar
measurements. We extend an existing approach for adjustment of C-band radar
data using state-space models and use the resulting rainfall intensities as input for
forecasting outflow from two catchments in the Copenhagen area. Stochastic
greybox models are applied to create the runoff forecasts, providing us with not
only a point forecast but also a quantification of the forecast uncertainty.
Evaluating the results, we can show that using the adjusted radar data improves
runoff forecasts compared to using the original radar data and that rain gauge
measurements as forecast input are also outperformed. Combining the data
merging approach with short term rainfall forecasting algorithms may result in
further improved runoff forecasts that can be used in real time control.

KEYWORDS
Flow forecast, greybox model, radar rainfall, state space model

INTRODUCTION

Radar observations are increasingly used for measuring rainfall in urban areas. The good spatial
coverage, however, comes along with problems in determining the rainfall intensity due to problems
such as beam attenuation and the drop size dependency of the relation between reflectivity and rain
intensity. Merging the radar measurements with gauge observations is a practitioners approach to this
problem.

Classically, radar rainfall measurements are adjusted with mean field bias to reflect ground
measurements as good as possible. Thorndahl et al. (2010) follow this approach in a two-step
adjustment that is used operationally within the real time control framework in the Copenhagen area
(Grum et al. (2011)). Uncertainties of the ground measurements are thereby neglected. Further,
assumptions need to be made on how to apply rain gauge point measurements to the radar rainfall
plane. Integrating gauge and radar rainfall measurements using state space models has been proposed
by several authors in the past. Chumchean et al. (2006) and Costa and Alpuim (2009) use these
techniques for temporal updating of the mean field bias. Brown et al. (2001) integrate spatial
interaction into their model via a vector autoregressive process. Similarly Grum et al. (2002) construct
a simple state space model that implicitly enables spatial interaction between the pixels and allows for
the integration of a multitude of measurement types that can be related to the rainfall process.
We adopt this last approach due to its ability to incorporate spatial interaction and various
measurement types and extend the uncertainty structure. The reconstruction of the rainfall process is
then used to create stochastic runoff forecast from a simple grey-box model. We evaluate the quality
of different forecasts using skill scores.

2
2.1

METHODOLOGY
Data and Catchments

We consider two catchments in the Copenhagen area. The Ballerup catchment has a total area of
approx. 1300 ha. It is mainly laid out as a separate system but has a small combined part. The runoff in
this area is further strongly influenced by rainfall dependent infiltration.
The Damhusen catchment is located close to Ballerup but drains to a different treatment plant. We
consider the northern part of the catchment with a total area of approx. 3000 ha. The catchment is laid
out as a combined sewer system and a multitude of CSOs are located in the area. Flow measurements
are available from both catchments in 5 min resolution.
A C-band radar is operated by the Danish Meteorological Institute (DMI) in Stevns approx. 45 km
south of the considered catchments. The spatial resolution of the radar pixels is 2x2 km. The provided
radar data are rain intensities derived using the Marshall Palmer relationship, where the coefficients
have been adjusted such that the average rainfall depth observed by the radar during the considered
period matches selected gauge measurements (Thorndahl et al. (2010)). We denote these data
unadjusted radar data. We consider an area of 9x11 pixels that covers the whole Copenhagen area
(Figure 1).
Within the catchments online rain gauge measurements are available from the Danish SVK network
(Jrgensen et al. (1998)). The gauges marked red in Figure 1 are used to adjust the radar
measurements. Only few of the available gauges are used for this purpose as one objective for using
radar rainfall data is to derive rain intensities from as few ground measurements as possible. To make
results comparable, we use the same gauges that are used for radar adjustment in a real time control
project in the Copenhagen area (Grum et al. (2011)). A reference simulation is performed where flow
forecasts are generated using rain gauge measurements as an input. The gauges for these simulations
were selected with respect to their location to the catchment as marked in Figure 1.
We have selected a 3-month period of measurements from 25/06/2010 until 29/09/2010 for this study.
The period contains several summer storms that should be relevant for control applications in urban

drainage systems. From 25/08/2010 until 14/09/2010 several extensive gaps can be observed in the
radar data also during some smaller rain events.
A modelling time step of 10 min is adopted corresponding to the resolution of the provided radar
measurements. The flow and rain gauge data are averaged to match this time step.
!

!(#*

!(

")

!(#*

!(#*")

!(

!(

#
*

!(

!
!

!(

1.500 3.000 Meters

Figure 1. Considered area with C-band radar pixels, Ballerup (left) and Damhusen (right) catchments,
rain gauges in the area (small dots), gauges used for radar adjustment (grey circles), gauges used as
input for reference simulations with gauges for Ballerup (white rectangles) and Damhusen (black
triangles) catchments
2.2

Radar Adjustment

We investigate a state-space approach that has first been described in Grum et al. (2002). We only
give a brief summary here and describe parts that differ from the previous publication. The general
setup is as follows:

Create a model to predict rainfall at the next time step for every pixel (system or state
equation)

Relate model prediction and observed rainfall values from different sources in a set of
observation equations

Determine the adjusted rainfall values by weighting between model and observation
uncertainties using a Kalman filter

In the state equations as well as in the variance matrices for model predictions and observations
parameters need to be defined. These are estimated using a maximum likelihood routine, maximizing
the probability of obtaining all the measured values included in the observation equations.

2.2.1

Model Setup

The adjusted rainfall depth in each radar pixel in the considered area is considered a state. The rain
intensity of the current time step is predicted as a weighted average of the rain intensities in the 3x3
neighbourhood of the pixel at the previous time step.
X i , j ,t

k ,l

(1)

X i k , j l ,t 1 ei , j ,t

k 1 l 1

Xi,j,t refers to the adjusted rainfall value at pixel (i,j) in the radar matrix at time step t, ei,j,t to the
corresponding gaussian prediction error with variance x and k,l to the weighting factors. We define
a

k l 0

k ,l 1 a

(2)

k 0, l 0

i.e. the sum of the weighting factors is 1 and all non-central pixels in the 3x3 neighbourhood receive
the same weighting. In matrix notation we have
(3)

X t A X t 1 et

where X is a vector containing 99 rainfall state values corresponding to 9x11 pixels, A is a weighting
matrix according to (1) and e is a vector of model errors with covariance matrix 1 with constant
variance x for all states on the diagonal and 0 on all off-diagonal elements, i.e. no correlation between
the states. States and measurements are related in the observation equation (4):
(4)

Yt C X t st

The observation vector Yt contains 99 non-adjusted measurements from all radar pixels and 8 rain
gauge measurements. The matrix C relates states and observations (see Grum et al. (2002)) and s is a
vector of observation errors with covariance matrix 2.
2.2.2

Observation Error Covariance Structures

We investigate different structures of the observation error covariance matrix 2. Model 1 (Eq. 5)
includes constant variances R and G for radar and rain gauge observations, respectively. Spatial
correlation between observations is not considered.
R

R
0

(5)

Model 2 extends the above setup by considering correlation R only between neighbouring radar pixel
observations. The correlation is found in the parameter estimation procedure.
Model 3 considers correlation for each radar pixel observation with all other pixels. The correlation is
assumed to decay as a power function of distance between the pixels according to Eq. 6, where the
distance D between pixels is defined in no. of pixels and parameters a and b are estimated from the
variogram of the radar observations and fixed during the maximum likelihood estimation of the whole
model. No correlation is considered for the rain gauge measurements.

(6)

a D b

Model 4 is equivalent to model 1. However, we do in addition introduce an error marker. If a radar or

rain gauge observation is missing, the corresponding variance is set to a large value and the correlation
values are set to 0.
2.3

Stochastic Runoff Forecasting

The estimation of rainfall forecast models does not permit a direct evaluation of the quality of the
adjusted radar data. We therefore generate runoff forecasts with different rainfall inputs and evaluate
the forecast quality. We use stochastic greybox models for generating the forecasts with focus on wet
weather periods as these are most relevant for real time control.
A lumped model consisting of a cascade of two reservoirs is applied for both catchments. The model
setup and development is described in Breinholt et al. (2011) using the (smaller) Ballerup catchment
as an example. For the (bigger) Damhusen catchment better forecasts could most likely be obtained
by applying a more elaborated model. However, here we are mainly interested in the effect of different
rainfall inputs on the forecast quality, not the best forecasting model. We consider the following
lumped model structure:
1

(S1,t )
S 1,t A P a 0 K S 1,t
d
dt
dt
1
1
S
2 ,t
(S2 ,t )
S 1,t S 2 ,t
K
K

(7)

1
S 2 ,k Dk ) ek
K

(8)

logQk log (

Similarly to the rainfall model described above, the model is laid out as a state-space model where
Eq. 7 is termed system or state equation and Eq. 8 observation equation. S1 and S2 correspond to the
storage states, A to the impervious catchment area, P to the rain intensity, a0 to the mean dry weather
flow and K to the travel time constant. The uncertainty of model predictions is captured by the Wiener
process dt with incremental variance 2. The variance depends on the current state values, so a
Lamberti transform is applied and the estimation performed with transformed states (Breinholt et al.
(2011)). In Eq. 8 Q corresponds to the observed flow values, D describes the variation of the dry
weather flow using trigonometric functions and e corresponds to the observation error with standard
deviation e.
Differently from Breinholt et al. (2011) we do not estimate the model parameters based on one-step
ahead flow forecasts. The runoff forecasts are intended to be used in a real time control setup (Grum et
al. (2011)). The relevant decision variable in the setup is expected runoff volume over the prediction
horizon. We therefore compute the expected flow values for the next 10 time steps (step length
t=10min) starting from time step k and integrate them to a predicted runoff volume (Eq. 9).
Vk (

10

k i

(9)

) t

i 1

The extended Kalman filter used in the modelling procedure also provides a variance for each
predicted flow value. Assuming normal distribution, we derive a 95% prediction interval on the flow
predictions for each horizon. Equivalent to Eq.9, we integrate the upper and lower bounds for the
different horizons to derive upper and lower prediction bounds for the expected runoff volume over
the whole prediction horizon.

Vk ,up (

10

(Q

k i

i 1

Vk ,low (

10

(Q
i 1

k i

n0.975 Q k i )) t

(10)

n0.975 Q k i )) t

(11)

In Eq. 10 and 11 indices up and low mark the upper and lower prediction bounds, respectively, n0.975
the 97.5% quantile of the standard normal distribution and Qk+i the standard deviation of the flow
prediction i steps into the future starting from time step k.
Comparing the above stochastic volume forecast to the observed runoff volume, we find the optimal
model parameters by minimizing the skill score (Sk) described in the next section. As wet weather
periods are the main focus of real time control, only the model parameters relevant to runoff (A, K,
uncertainty parameters) are estimated and dry weather periods are excluded from the evaluation of the
skill score function. The dry weather parameters (a0, D) for the two catchments are estimated
deterministically from a 14 day dry weather period at the beginning of the considered period and then
fixed during estimation of the other model parameters.
2.4

Forecast Evaluation

When evaluating stochastic flow forecasts, we need to consider the quality of prediction intervals
rather than just a mean error between prediction and observation. Criteria for forecast evaluation were
proposed by Jin et al. (2010) and Thordarson et al. (in press):

Reliability (Rel) percentage of observations not contained in a 95 % prediction interval

ARIL - average width of the 95 % prediction interval (=Sharpness Sh) relative to observation

Skill score (Sk)

Sk Sh

2
0.05 N

Li

where N is the number of wet weather observations, Sh is the average width of the 95 %
prediction interval and Ui and Li are the distances of the i-th observation from the upper /
lower prediction interval (over-/ undershoots). Ui and Li are 0 if the observation is contained in
the prediction band.
We compute these criteria for a runoff volume prediction interval as described above. Only wet
weather periods are considered in the computation of the evaluation criteria.

RESULTS

Table 1 shows the parameters derived for the radar adjustment models 1-4. The small weighting factor
a of the central pixel indicates that the model predictions include information from the whole 3x3
neighbourhood, rather than just the central pixel. The variance of the model predictions x is generally
estimated smaller than that of the observations R and G. When computing the adjusted rainfall values,
the Kalman filter will therefore show a tendency to smoothen the observed values.
We further observe that including spatial correlation between the radar observations into the model
(Models 2-3) decreases the ratio between the variances of radar and rain gauge observation errors. The
correlation term reduces the weight of the single radar observation and allows for retrieving
information from the gauges also if they are considered more uncertain. Similarly, the variance of the
model prediction errors x can be increased in this case as less weight is put on the observations.

Generally, the estimation of the state-space radar adjustment models using the described Maximum
Likelihood approach has turned out problematic in application. Similar likelihood values may be
obtained for rather different sets of parameters making identifiability of the models difficult. The
improved flow forecasts obtained with adjusted radar input as compared to e.g. rain gauge input
(Table 2) indicate that we were able to identify reasonable parameter sets. Improved estimates can
most likely be obtained if an objective function based on flow measurements is also used to find the
parameters of the rainfall adjustment models.
Table 1. Parameter values for state-space radar adjustment models
Model
a
x
R
G
R
Model 1
Model 2
Model 3
Model 4

0.20
0.20
0.20
0.20

1.2310

-4

1.5810

-4

1.5110

-4

1.2310

-4

8.1210

-4

4.9010

-4

8.1210

-4

8.1210

-4

9.1210

-4

3.3910

-1

3.8010

-1

9.1210

-4

0.22

0.614

0.384

Table 2. Forecast quality criteria for the two catchments with different rainfall inputs. Measures are
given in m3 per 100 min and values are evaluated in wet weather periods only
Model Input
Ballerup catchment
Damhusen catchment
Rel
ARIL
Sk
Rel
ARIL
Sk
Rain gauge

5%

65%

1466

4%

116%

11777

Radar no adjustment

5%

56%

1378

6%

95%

12283

Radar Model 1

5%

56%

1342

6%

90%

10975

Radar Model 2

5%

57%

1345

5%

92%

10721

Radar Model 3

5%

64%

1403

6%

93%

11265

Radar Model 4

5%

59%

1339

5%

94%

10479

Table 2 shows the results of the stochastic runoff volume forecasts generated using the different
rainfall inputs. The prediction intervals in the Damhusen catchment are generally wider than in the
Ballerup catchment, indicating a too simple model structure for this catchment. Still, the simple model
allows us to judge on the quality of different rainfall inputs for flow forecasting. Comparing the
volume forecast quality obtained with pure rain gauge and pure radar rainfall input to that obtained
with adjusted radar rainfall input, we notice skill scores improved by 3-15%. The prediction intervals
are generally narrower when using radar rainfall input compared to rain gauge input.
Including correlation into the covariance structure of the radar observations (Models 2 and 3) does not
give clear improvements of the runoff forecasts. At this stage it is not possible to conclude if
consideration of this effect actually has no significant effect on flow predictions. A better estimation
method for the radar adjustment models may be able to exploit this effect better.
Figure 2 illustrates the effect of using an error marker in the adjustment of the radar data. The radar
observations are missing for the small events between time steps 11500 and 11800. Using the error
marker in model 4, we are able to reconstruct rainfall values from the gauges.

0.20
0.10
0.00
0.20
0.10
0.00
0.20
0.10
0.00

gauge [mm/min]
radar [mm/min]
radar calib. [mm/min]

11500

11600

11700

11800

11900

12000

time step [10 min]

Figure 2. Mean area rain intensities for the Ballerup catchment from gauge measurements (top), nonadjusted radar measurements (centre) and adjusted radar measurements (bottom) without (Model 1,
full thin line) and with (Model 4, dotted bold) error marker

CONCLUSIONS

We have evaluated the possibility of adjusting radar rainfall measurements with rain gauge
measurements using state space models and evaluated the effect of the adjustment on runoff forecasts
for two catchments generated from stochastic greybox models. With the adjusted radar data as input
we obtain improved runoff forecasts as compared to using rain gauge or non-adjusted radar data as
model input. Using an error marker allows to reconstruct adjusted rainfall values also if radar or some
of the gauge observations are missing.
Despite the improved flow forecasts obtained with the adjusted radar data, we see several possibilities
for improvements of the presented approach. Estimating parameters for the radar rainfall adjustment
based on rainfall observations only has proven difficult. Better results can most likely be obtained by
including a runoff prediction into the rainfall adjustment model and comparing predicted and observed
runoff.
Further, we have considered an area of 9x11 C-band radar pixels in this study. This area is sufficient to
cover the whole of Copenhagen. However, it is too small to generate short term rainfall forecasts from
the radar. Preferably, the whole radar matrix of 240x240 pixels should be considered for this purpose
corresponding to 57600 observations. Operating on variance matrices with 57600x57600 entries in the
Kalman filtering procedure is impossible. A modified procedure that directly estimates the Kalman
gain may be a possible solution to this problem.
With respect to the runoff forecasting models, improved forecasts for the bigger catchment can very
likely be obtained by applying a more elaborated model structure that accounts e.g. for effects such as
overflows. Further, improvements could be obtained by distinguishing between dry and wet weather
situations in the physical and stochastic model parts and by modelling prediction uncertainties
8

depending on rainfall characteristics. These characteristics should aim at identifying convective events
as these imply the highest forecast uncertainties. Using these methods we aim at providing forecasts
that clearly improve decision making in real time control of sewer networks.

ACKNOWLEDGEMENTS

This research is financially supported by the Danish Council for Strategic Research, programme
Commission on Sustainable Energy and Environment through the Storm- and Wastewater Informatics
(SWI) project. Catchment and flow data were kindly provided by Avedre Wastewater Services and
Copenhagen Energy. Radar data from the Danish Meteorological Institutes (DMI) C-Band radar at
Stevns were kindly provided from a repository at Aalborg University. We thank Pjerre-Julien Trombe
for assistance and advice in handling radar data.

REFERENCES

Breinholt A., Thordarson F.., Mller J.K., Grum M., Mikkelsen P.S. and Madsen H. (2011). Greybox modeling of flow in sewer systems with state-dependent diffusion. Environmetrics, 22(8),
946-961.
Brown P.E., Diggle P.J., Lord M.E. and Young P.C. (2001). Space-time calibration of radar rainfall
data. Journal of the Royal Statistical Society. Series C (Applied Statistics), 50(2), 221-241.
Chumchean S., Seed A. and Sharma A. (2006). Correcting of real-time radar rainfall bias using a
Kalman filtering approach. Journal of Hydrology, 317, 123-137.
Costa M. and Apluim T. (2009). Adjustment of state space models in view of area rainfall estimation.
Environmetrics, 22, 530-540.
Grum M., Harremes P. and Linde J.J. (2002). Assimilating a multitude of rainfall and runoff data
using a stochastic state space modeling approach. In Proceedings of the 9th International
Conference on Urban Drainage, 811 September 2002, Portland, Oregon USA.
Grum M., Thornberg D., Christensen M.L., Shididi S.A. and Thirsing C. (2011). Full-scale real time
control demonstration project in Copenhagens largest urban drainage catchments. In
Proceedings of the 12th International Conference on Urban Drainage, 11-16 September 2011,
Porto Alegre, Brazil.
Jin X., Xu C.-Y., Zhang Q. and Singh V. P. (2010). Parameter and modeling uncertainty simulated by
GLUE and a formal Bayesian method for a conceptual hydrological model. Journal of
Hydrology, 383(3-4), 147-155.
Jrgensen H.K., Rosenrn S., Madsen H. and Mikkelsen P.S. (1998). Quality control of rain data used
for urban runoff systems. Water Science and Technology, 37(11), 113-120.
Thordarson F.., Breinholt A., Mller J.K. ,Mikkelsen P.S., Grum M. and Madsen H. (in press).
Evaluation of probabilistic flow predictions in sewer systems using grey box models and a skill
score criterion. Stochastic Environmental Research and Risk Assessment, accepted.
Thorndahl S., Rasmussen M.R., Neve S., Poulsen T.S. and Grum M. (2010). Vejrradarbaseret styring
af spildevandsanlg (Weather radar based control of wastewater systems). DCE Technical
Report No. 95, Aalborg University, Institut for Byggeri og Anlg, ISSN 1901-726X.

9th International Conference on Urban Drainage Modelling

Belgrade 2012

A spatial-temporal rainfall generator for urban drainage

design
Fiona McRobie1, Li-Pen Wang2, Christian Onof3, Stephen Kenney4,5
1

Imperial College London, United Kingdom, fiona.mcrobie@gmail.com

Imperial College London, United Kingdom, li-pen.wang08@imperial.ac.uk
3
Imperial College London, United Kingdom, c.onof@imperial.ac.uk
4
Thames Water, Reading, United Kingdom
5
MWH, Warrington, United Kingdom, steve.kenney@mwhglobal.com
2

ABSTRACT
The work presented here is a contribution to the Thames Water project of
improving the sewerage system draining the Counters Creek area in London. The
need for further improvements to the system has been highlighted by an increase in
the number of floods affecting basement flats in the area. However, the cost of
designing appropriate additional components for the system is potentially very
high, so it is important to know whether any substantial overestimation results from
using the traditional approach of 30-year spatially uniform design storm events.
This work therefore developed a simple spatial-temporal stochastic rainfall tool to
generate simulations of spatially distributed rainfall events, from which 30-year
storms can be extracted.
Storm events are modelled as clusters of Gaussian rainfall cells by extending the
method implemented by Willems (2001) to radar rainfall data. The parameters
describing the cells and their motion are sampled from probability distributions
derived from estimates of the parameters gained from 45 historical storm events
within the catchment from the past ten years. The spatial-temporal stochastic
rainfall generator combines the parameter distributions to produce a two
dimensional time series of simulated storm events from which events of given
return period can be identified.

KEYWORDS
Gaussian, radar rainfall, rainfall cell, spatial-temporal, stochastic, urban hydrology

INTRODUCTION

It has been documented by Singh (1997) that spatial variation in rainfall has a significant impact on
catchment responses in a number of urban and rural contexts. Although the volume of rainfall is
generally recognised as the most important factor determining runoff, spatial and temporal variations
in rainfall in particular the speed and direction of movement of a storm event have been proven to
1

increase the runoff peak or to explain the variability of runoff (de Lima and Singh, 2002; Pechlivanidis
et al, 2008). As such, spatial rainfall modelling is integral to accurate flooding and drainage analysis.
Typically, the results of studies regarding the level of impact the spatial properties of the rainfall field
have on generated runoff vary from one catchment to another. Pechlivanidis et al. (2008) found that in
the Upper Lee catchment in the south east UK, the spatial properties of rainfall have no greater impact
on runoff as catchment scale varies. In contrast, Syed et al (2003) found that in a catchment studied in
Arizona, the location of the storm core is more significant at larger catchment scales, as attenuation
of peak flow can be observed when storms are located further from the catchment outlet. Since the
hydrological responses are catchment-varying, there is therefore no rule of thumb for the impact of the
spatial and temporal variability of rainfall on catchment response. In order to incorporate spatial
variability into the network analysis, a programme able to generate spatially distributed rainfall must
be developed to help characterise the hydrological responses of a given catchment.
Willems (2001) produced a spatial rainfall generator for a 100 km2 urban catchment covering Antwerp
(in Belgium). Based on a network of 12 rain gauges, small mesoscale area storms of size 10 2 to 103
km2 were modelled as a cluster of rainfall cells of a bivariate Gaussian form. However, Willems
(2001)s algorithms to retrieve the generators parameters from raingauge records cannot be directly
applied to radar measurements; the latter provide more complete information of spatial and temporal
variability of rainfall fields but at the same time there is more noise in the dataset. This work therefore,
based upon the assumptions of Willems (2001), proposes a new fine resolution spatial-temporal
rainfall generator, whereby the parameters that characterise rainfall storms are derived from historical
radar data.
1.1

Experimental Catchment and Data Set

The region of interest is a 11 by 12 km2 area covering the Counters Creek catchment in north west
London, and spatially distributed rainfall fields encompassing the wider Beckton catchment are
required in order to allow for backflow. Radar data are available at 1x1 km2 resolution at 5 minute
intervals for this catchment from 2000, and were obtained from Thames Water. A set of 80 storm
events from the period 2000 to 2011 were highlighted by MWH as significant with respect to network
flooding, and of these, 45 storms were selected after being tested for quality of the radar data (MWH,
2011). In order to identify the full extent of rainfall cells, the radar data were analysed from a larger
area, fully encompassing the Beckton catchment. This region covered 8400 km2 for post November
2006 data and 5525 km2 for data prior to November 2006.

2
2.1

METHODOLOGY
Model Definition

The simulated storm event is characterised by a duration , a velocity vector (magnitude: ; angle
to the x axis: ), and a mean number of cells per km2, . At any given time, the model assumes that the
distribution of the cells in space is given by a two-dimensional Poisson process, with parameter .

wa
CATCHMENT AREA

ws
O

SIMULATION AREA

la

ls

Figure 1: Simulation area and catchment area

Following Willems (2001), the assumption of constant velocity allows us to consider the rainfall field
at time as a translation of a random simulation field moving across the catchment with velocity .
The area this process covers must be sufficient to enable the whole simulation of the event of duration
. Given the angle of the velocity vector, this area will have the shape indicated in Figure 1, with
length
and width
. In order to reduce error when translating from the simulation area to the
catchment area, the rainfall field is simulated with 100x100 m2 resolution while the resulting storm
event in the catchment has a spatial resolution of 1x1 km2.
The following steps are then undertaken in order to create the spatially-varied stochastic event within
the catchment:
1. Rainfall cells are placed within the simulation area by sampling from the Poisson distribution
to yield a number of cells. This is based on the assumption that rainfall cells are
uniformly distributed across a mesoscale rainfall event, following Willems (2001). These
cell centres are allocated positions over the simulated rectangle using the uniform property of a
given number of Poission variables. Sampling from each of
and
yields
the coordinates and of a cell centre, is then repeated times to yield the cell centres.
2. Following Willems (2001), the cells were treated as ellipses with one axis in the direction of
movement and the other perpendicular, and the rainfall intensity is represented by a Gaussian
function along the each axis, centred on
. This differs from the approach of Einfalt et al
(1990) where the main axes of inertia are used, but it sits inline with empirical observation
that cell movement is orthogonal to any cell banding (Wheater et al, 2000). Cells are then
constructed by sampling for the peak cell intensities
, which is attributed to the point
, and cell spread parameters
and , corresponding to standard deviations, which
describe the rate of decay of the rainfall intensity in the plane of motion and the transverse
plane respectively.

3. Having now simulated a static storm within the simulation area, the whole simulation area
starts to move and cross the catchment area according to the sampled velocity and angle .
For each time step t , the rainfall intensity in the catchment area can be obtained by mapping it
to the moving simulated rainfall field. This is achieved using basic geometric relationships,
and yields the 1x1 km2 grid square
within the simulation field which corresponds
2
directly to a group of 100x100 m grid points
, where
. The simulated
rainfall intensity of this grid,
, is thus obtained by averaging these 100x100 m2 grids
.
2.2
2.2.1

Model Calibration
Cell Identification

In order to characterise storms, it is necessary to isolate individual cells within radar images; a
hierarchical thresholding method (Peak & Tag, 1994) was thus employed. The hierarchical
thresholding method applies a finite set (say, five) of threshold intensities, , to the radar dataset and
identifies a rainfall cell as a connected group of pixels with values exceeding the threshold
if for
threshold , with
, only a single, smaller, rainfall cell is identified. The thresholding algorithm
was applied to each snapshot of the radar data, outputting a numbered set of cells at each five minute
interval.
2.2.2

Parameter Estimation

The parameters for which suitable distributions are to be fitted were estimated from the observed
storm events.
The estimator of the number of cells per unit area, , was the number of cells identified at each time
step. Likewise, duration, d, was estimated by taking the time when the first cells were identified until
the last cells were identified.
The speed of the storm event, , and, the direction of movement, , were estimated using spatial
correlations relative to a central pixel in the radar data set over a series of time step lags. This allowed
for an estimate of the distance travelled during the lag to be determined, and thus velocity (magnitude
and direction) could be estimated for each storm event.
The cell parameters, and and
had to be estimated for all cells in each timestep, as no suitable
method for linking the rainfall cells over time was found. Estimating the parallel and perpendicular
cell spreads relied on the assumption that rainfall cells can be modelled as displaying a Gaussian
structure along both the direction of motion (given by ), and the transverse. In the plane of these
directions, the rainfall cell is assumed to take the form:

Where is distance in the plane of the direction of motion or transverse, is cell spread, and is a
proportionality constant allowing for the total rainfall intensity to be represented. This allows and
to be estimated independently from each other. For each direction, the rainfall intensity at the point
where the cell is cut off due to the thresholding method, and the peak intensity are substituted into the
Gaussian function, and the standard deviation is solved for.
Despite the assumption that rainfall cells remain a constant shape and size over time, in reality this is
not true, and therefore these estimated values of spread will include all the fluctuating sizes of the

cells. This results in a large sample of spread estimates containing not only a useful sample of spread
estimates, but also some unrealistically small numbers that would have been averaged out over the
life-time of the cell. This was corrected for by transforming the estimates of the spread variables such
that they are distributed according to the average of the spread variables.
2.2.3

Distribution Fitting

Table 1 gives the distributions fitted to each set of parameter estimates. The distributions were fitted
by maximum likelihood estimation of the storm and cell parameters. Although the velocity parameter,
v, is fitted to a lognormal distribution, this distribution is later discarded in favour of the Weibull
distribution as per Willems (2001). This is covered in detail in the model validation section.
Table 1. Distributions Fitted to Parameter Estimates Using 45 Storm Events
Parameter Set
Fitted Distribution
Distribution Parameters
Storm Duration (d)

Gamma

Number of Cells Per

Square Kilometre ()

Normal

Velocity Magnitude (v)

Log Normal

Velocity Direction ()
Cell Spread in Direction
of Motion (s1)

a (-) = 2.4374

SEa = 0.5286

b (hrs) = 17.9238

SEb = 4.2339

-2

SE = 0.000519

-2

(km ) = 0.00324016

SE = 0.000374

(km/hr) = 78.3045

SE = 0.053533

(km/hr) = 23.6352

SE = 0.038667

(rads) = 0.5077

SE = 0.150082

(rads) = 1.1644

SE = 0.108474

= 0.700118

SE =0.00324

= 1.05799

SE = 0.00229

(km ) = 0.00991809

Log Normal
Log Normal

Standard Error

(mean = 3.52469 km; var

= 25.6279 km2)
Cell Spread in Direction
Perpendicular to Motion
(s2)

Log Normal

Maximum Intensity in
Cell (rmax)

Generalised Pareto

= 0.703913

SE = 0.00297

= 1.00079

SE = 0.0021

(mean = 3.33576 km; var

= 19.1676 km2)
k = 0.586392

SEk =0.002328

= 1.03958

SE =0.002718

= 0.2 mm

SE =0

(mean = 2.71 mm)

2.3

Model Validation

The possibilities of validating the proposed model are limited. The distributions obtained are for all the
storms together. So it is not possible to check that one is able to reproduce a particular kind of
observed storm, therefore a more limited and qualitative validation is carried out. First, it is useful to
apply the parameter estimation methods to storms which have been generated with the stochastic
model, to assess the extent to which the cells can be identified and parameters estimated. Second, in
order to evaluate the rainfall generator itself, a number of simulated storm events are generated using
the parameter distributions given above, and analysed in comparison to a sampled number of the
historically observed storms. Here, the basic statistical structure of the events is considered, such that
5

the simulated storm events take the same form as those observed. In addition, a series of storm events
was simulated using the parameter distributions proposed by Willems (2001).
Applying the parameter estimation methods to stochastically generated storms highlighted uncertainty
in the velocity estimation method, while cell spread parameter estimation appeared more effective.
Cell identification was also considered to be effective.
In order to evaluate the modelled rainfall cells, intra-storm profiles of maximum intensity, mean
intensity and spatial variance were produced. This meant taking a sample of observed storms and
using the velocity ( and ), duration ( ) and cell density ( ) parameters in order to simulate storms
which should be identical apart from the location and size of cells. Then the mean, maximum and
spread of rainfall intensity over the catchment were plotted over time for each storm, along with the
autocorrelation (temporal correlation) for the first 10 lags of 5 minutes. These were compared to the
same plots produced for the observed storm. It was observed that the rainfall generated from the model
is far too small in intensity. Looking at the cells generated, it appears that only very rarely is a large
cell intensity sampled (as expected from the shape of the Generalised Pareto Distribution). When this
happens, it does not contribute to a substantial storm however, as the rest of the cells are most likely to
be insignificant.
Comparison to storms generated using the parameter distributions proposed by Willems (2001) was
required on evaluation of the autocorrelation function for the first ten five-minute lags. All the
stochastically generated storm events exhibited low autocorrelation from the second or third lag, while
historically observed storms displayed persistent autocorrelation to much later lags. This bias in the
stochastic rainfall generator can be remedied by reducing the storm velocity, and in particular, higher
autocorrelation functions can be produced if the velocity parameter, , is sampled from the Weibull
distribution as proposed by Willems (2001), as this allows for smaller sampled values of more
frequently. The persistence of the rainfall field (as evidenced by a shallow autocorrelation function) is
likely to have a significant impact on the urban catchment, and thus it was deemed necessary to
replicate this with the rainfall generator. As a result, the lognormal distribution fitted to the velocity
parameter, , was discarded in favour of the Weibull distribution (Willems, 2001).
Finally, as noted above, the rainfall generator creates rainfall cells of sometimes highly varied
intensities, with small intensities dominating the whole picture. The rainfall intensity of cells within a
given storm were treated as independent from one another, however, from visual inspection of an
observed storm event, this is an inaccurate representation. Considering a box plot yields little
information, as the majority of cells are of low intensity, and the differences are only noticeable in the
upper quartiles. This implies that the maximum rainfall intensity of cells may be dependent on that of
other cells in the same storm, thus creating a relationship between cell intensity and the characteristics
of the storm event. This is a similar concept as that employed by Onof et al (1996) in determining a
relationship between the cell intensity distribution parameter and storm duration. As a result, the
simulation program was adapted to allow for a level of correlation between maximum cell intensities
within a storm: in this way, the model will be more likely to generate storms with many substantial
intensities. To allow for variation in the level of dependence, a percentage level K% was set such that
cell peaks were sampled as independent, identically distributed variables from the Generalised
Pareto function (as previously), while

cell peaks were sampled from a truncated

distribution based around the maximum of the independent

cell peaks.

CONCLUSIONS AND RECOMMENDATIONS FOR FURTHER WORK

The aim of this study was to provide a proof of concept of the possibility of generating a wide variety
of spatially distributed storms which emulated the observed variety of spatial storm structures. As
explained above, certain issues arose in the identification of the model parameters which it is
important to bear in mind, and could form the basis of further analysis:
1. The cell tracking algorithm, developed in order to track cells from one timestep to the next,
thus allowing for velocity parameters
and
to be estimated directed, was not able to
identify enough cells to be useful. This is largely due to the very noisy intensities of the radar
data which make any such cell tracking more difficult than it would have been with raingauge
intensities, albeit it is easier to see cells with radar data. As a result, the velocity parameters
were estimated using spatial correlations (mentioned in point 2), and the impact of this
problem upon the cell spread estimates was partly remedied by reducing the distributions of
the variances of these parameters.
2. The velocity estimation relied upon spatial correlations over time-lagged radar pictures.
Again, the noisy radar data led to some spurious maxima being identified which made the
results rather questionable. To remedy this, we assumed we could use the distribution of
velocities found for the Brussels area in Belgium (Willems, 2001). It is worth noting here that
for each storm, the cell spread estimates were based upon the velocity direction parameter
estimate, which was shown to be highly uncertain. However, as the resulting spread parameter
distribution reflected actual values effectively, it was retained.
3. It appeared clearly from the observed data for different storms that the distributions of
maximum intensity showed significant differences from storm to storm. Although this was not
apparent in terms of the bulk of the distributions, the tails i.e. starting at the 80% upper
percentile, are variable. This variability is not accounted for in the original method, hence a
variant was introduced which sampled a percentage of the cell intensity maxima from one
region of the distribution for a certain proportion of the cells in any given storm. In addition,
the use of the General Pareto function caused the peak intensities to be bounded in the upper
tail, thus discounting the occurrence of extreme events. Further work is recommended to
determine a more suitable fit for the peak intensity distribution.
4. The distribution fitting for the spread parameter estimates, and , assumed that and
likewise , were distributed identically from one storm to the next. Although the cumulative
distribution functions for each storm were visually similar, a Kolmorogov-Smirnov test,
considering the empirical cumulative distribution functions of the parameter sample gained
from pairs two different storms, was inconclusive, and for more than 95% of storm pairs, the
hypothesis that they are drawn from the same distribution cannot be accepted with 90%
confidence. This casts significant doubt over the fitted parameter distributions, which may be
due to the estimation method for and , or may be an indicator that the model is
oversimplified, and unable to accurately replicate storm heterogeneity.
5. All the issues above are related to the uncertainty and noise in radar rainfall measurements. As
is generally acknowledged, radar is a useful tool to identify the spatial structure of rainfall, but
can often be out by a factor of at least 2 when it comes to intensities (and this error is not
consistent, but spatially variable and dependent upon the actual intensity). This project did not
include a merging of radar with other sources of data. As a result, the results are affected by
these radar data problems. This means that the distributions of cell spreads, as well as of the
number of cells are likely to be the most reliable estimates obtained here. It is possible that

some scaling factor could usefully be applied to the outputted model intensities, but it is not
clear what that scaling factor should be.
The model makes a number of assumptions, some of which could be made more realistic in an
improved version of the model:
1. The cell intensities remain constant throughout, but clearly cells actually grow and decay. This
is probably not too much of a problem over a small area, but would be if the model were to be
used to simulate rainfall for an area the size of London. This issue could be addressed by
imposing a growth and decay period so that the mean remains the same.
2. The maximum cell intensity has been assumed to be independent of the cell spread in either
direction. Although we found no evidence of any significant correlation, this may be due to
the problem with the cell tracking. Assuming a log-normal distribution were to be used instead
of a Generalised Pareto distribution, it would in principle be easier to include dependence
between the maximum intensity and the two cell spread dimensions.
3. The cell arrival process is Poisson, which assumed total mutual independence between the
cells. Although this assumption may be a reasonable approximation in the case of a small
catchment area, clustering would have to be included were the model to be applied over a
larger area. This could be achieved by having cells embedded within clusters of cells. The
latter could be represented as arriving according to a Poisson process, while the cells would be
displaced from the centre using a Gaussian distribution for instance (see Wheater et al., 2000)
In conclusion, we have constructed a generator of spatially varied rainfall for the Counters Creek
catchment in the London region. This generator is based on radar rainfall data for 45 hydrologically
significant storm events from the past 11 years, which were used to calibrate the model. This study
was based on the work undertaken by Willems (2001), however estimation of characteristic
parameters had to be adapted to allow for the use of radar data.

4
4.1

REFERENCES
List of references

de Lima, J.L.M.P & Singh, V.P. (2002). The influence of the pattern of moving rainstorms on
overland flow. Advances on Water Resources. 25(7), 817-828.
Einfalt, T., Denoeux, T. & Jacquet, C. (1990). A radar rainfall forecasting method designed for
hydrological purposes. Journal of Hydrology. 114, 229-244.
MWH (2011). Counters Creek spatial rainfall study: Stage (a) How important is spatial rainfall
variation? MWH Technical Note. TN/EWI/CH11/186/01.00
Onof, C., Faulkner, D. & Wheater, H.S. (1996) Design rainfall modelling in the Thames catchment.
Hydrological Sciences Journal. 41, 715-733.
Peak, J. E. & Tag, P. M. (1994). Segmentation of satellite imagery using hierarchical thresholding and
neural networks. Journal of Applied Meteorology. 33, 605-616.
Pechlivanidis , I.G., McIntyre, N.M., Wheater, H.S., Jackson, B.M. & Orellana, B. (2008). Relation of
spatial rainfall characteristics to runoff: An analysis of observed data. BHS 10th National
Hydrology Symposium, Exeter, 2008.
Singh, V.P. (1997). Effect of spatial and temporal variability in rainfall and watershed characteristics
on stream flow hydrograph. Hydrological Processes. 11(12), 1649-1669.
8

Syed, K.H., Goodrich, D.C., Myers, D.E., & Sorooshian, S. (2003). Spatial characteristics of
thunderstorm rainfall fields and their relation to runoff. Journal of Hydrology. 312, 191-206.
Willems, P. (2001). A spatial rainfall generator for small spatial scales. Journal of Hydrology. 252,
126-144.
Wheater, H.S., Isham, V.S., Cox, D.R., Chandler, R.E., Kakou, A., Northrop, P.J., Oh, L., Onof, C. &
Rodriguez-Iturbe, I. (2000). Spatial-temporal rainfall fields: Modelling and statistical aspects.
Hydrological Earth Systems Sciences. 4(4), 581-601.