Application of weather radar data in urban hydrological applications has evolved significantly during the past decade as an alternative to traditional rainfall observations with rain gauges. Advances in radar hardware, data processing, numerical models, and emerging fields within urban hydrology necessitate an updated review of the state of the art in such radar rainfall data and applications. Three key areas with significant advances over the past decade have been identified: (1) temporal and spatial resolution of rainfall data required for different types of hydrological applications, (2) rainfall estimation, radar data adjustment and data quality, and (3) nowcasting of radar rainfall and real-time applications. Based on these three fields of research, the paper provides recommendations based on an updated overview of shortcomings, gains, and novel developments in relation to urban hydrological applications. The paper also reviews how the focus in urban hydrology research has shifted over the last decade to fields such as climate change impacts, resilience of urban areas to hydrological extremes, and online prediction/warning systems. It is discussed how radar rainfall data can add value to the aforementioned emerging fields in current and future applications, but also to the analysis of integrated water systems.
In 2003 the International Group on Urban Rainfall (IGUR) under the IWA/IAHR
Joint Committee on Urban Drainage initiated joint work on the status and
development on using radar rainfall data within the context of urban
drainage. This led to a review paper entitled “Towards a roadmap for use of
radar rainfall data in urban drainage” which was published in
However, the foundation upon which the original paper was based has significantly changed during the past 1.5 decades. This is partly due to the rapid developments in radar hardware, signal and data processing; the development of new methods for data processing and analysis; advancements in urban runoff modelling incorporating more complex hydrological processes, etc.
The purpose of the current paper remains the same as the one by Einfalt et
al. (2004), i.e. to bridge the gap between current precipitation measurements
and suitable precipitation information for operation and design of urban
drainage systems. Schilling (1991) and Einfalt et al. (2004) summarized these
needs as follows: at least 20 years of recordings without data gaps, a
volumetric accuracy of less than 3 %, and a spatio-temporal resolution of
1 km
Scopus and Web of Science documents under search strings
“
The scientific interest within the field is evident from the number of
publications with specific search strings in different scientific databases.
Figure 1 shows the number of publications registered under the keywords
“
Most cited papers within radar and urban drainage.
Table 1 presents the three most cited papers according to
Climate change and consequently increase in extreme rainfall have been a significant catalyst for the development in urban hydrological models over the past decade. There is a need to be able to simulate current and future loads on drainage systems and to fully utilize the capacity of drainage systems in order to accommodate climate change. Furthermore, integrated hydrological models (e.g. integrated urban drainage, river, and inundation models) have become standard tools, e.g. to simulate inundation risks in urban areas. The use of more detailed and distributed models increased the demand for good quality, high resolution inputs, which promotes the use of radar rainfall data in urban hydrology.
In addition to the higher demand for precise local rainfall data, technological developments in hardware as well as data processing and quality have changed significantly since the publication of the papers in Table 1. These form the starting basis of this paper, which aims to provide a review of the major technical developments during the past decade, with particular focus on the most ground-breaking applications and cases, from where updated recommendations are distilled for the applications of radar rainfall data in urban hydrology. Given that many of the new developments are still within innovation projects by research communities, this review also aims to expand the new knowledge to the industry and water companies. Confidence in radar data is provided and possibilities of applications in urban hydrology and urban drainage are being mapped.
We structure the review based on the following three key areas of research
that are identified as being central in a large majority of the publications
within the field of radar rainfall application in urban hydrology.
Temporal and spatial resolution of radar data (Sect. 2.1) Rainfall estimation, radar data adjustment, and quality (Sect. 2.2) Nowcasting of radar rainfall (Sect. 2.3)
We approach these key areas from two sides. Initially, in Sect. 3, we review
the state of the art within the three key areas, respectively, focusing on
the radar and radar rainfall related issues. Secondly, we will end each
sub-section by reviewing the impacts of each of the key areas on applications
within urban hydrological modelling. The second part of the paper focuses on
the value of applying radar rainfall in urban hydrology, giving examples of
applications (Sect. 3). Finally, in Sect. 4, we present our subjective views
of what is needed and what can be recommended for current and future
applications of radar rainfall data in urban hydrology.
Urban hydrology is characterized by fast runoff and short response times on impervious surfaces, and thus small timescales and space scales compared to rural hydrology. Rainfall data for urban hydrology are therefore required to resolve these spatial and temporal scales sufficiently. However, following Willems (2001), Thorndahl et al. (2008), Schellart et al. (2012b), and others, the errors in such rainfall input data are one of the most important sources of uncertainty in (urban) hydrological models. For example, for a sewer system model in Belgium, it was shown by Willems and Berlamont (1999) that about 20 % of the total uncertainty in the downstream sewer throughflow discharges could be explained by the spatial variability of the rainfall and about 20–25 % by the rainfall measurement errors, consisting in their case of rain gauge calibration errors, rainfall intensity resolution errors and errors by wind and local disturbances. For extreme events, e.g. flash flooding, uncertainties related to spatial variability and rainfall measurement errors are expected to be even larger (e.g. Berne et al., 2004; Hossain et al., 2004; Brauer et al., 2016). Hence there is a need for high quality and high resolution rainfall inputs into urban hydrological models in order to reduce uncertainty in hydrological responses. Radar rainfall data are ideal in that respect.
Since the first use of radars for precipitation measurement, there has been rapid development and improvement of weather radar hardware, signal processing, software, etc., but the fundamental principles of applying weather radar for precipitation measurements have not changed significantly. We therefore refer to the existing literature on the fundamentals of radar and atmosphere physics, e.g. antennas, frequencies, bandwidths, polarization, and data correction, e.g. attenuation, clutter removal, and reflectivity–rainfall conversion. These fundamentals are indeed crucial for the quality of rainfall estimation and should definitely not be disregarded by users of radar rainfall, but they are omitted from the paper since they have been discussed in depth in primers such as Doviak and Zrnić (1993), Collier (1996), Bringi and Chandrasekar (2001), Meischner (2004), Michaelides (2008), and Rinehart (2010). Furthermore, there are pioneering and significant journal papers such as Marshall and Palmer (1945), Austin and Austin (1974), Wilson and Brandes (1979), Smith and Krajewski (1991), Krajewski and Smith (2002), Einfalt et al. (2004), Delrieu et al. (2009), Krajewski et al. (2010), Villarini and Krajewski (2010), and Berne and Krajewski (2013) which also provide general information on specifications and applications of radar rainfall. Also, VDI (2014) and ISO (2017) have produced a standard on precipitation measurement by radar. In the following, we focus on new developments in applications of radar in urban hydrology, and start the discussion from the temporal and spatial resolution needs.
The temporal resolution of radar data is governed by the scanning strategy of the radar. A radar scanning the atmosphere in different elevations to generate a full azimuthal volume scan can take up to several minutes depending on rotational speed and the number of scanning elevations. Radar collects instantaneous samples of rain rates (estimated from reflectivities), unlike rain gauges, which accumulate rainfall over a given time interval. Some radars operate with intermediate dedicated Doppler scans for each volume scan, hence doubling the time between two consecutive reflectivity scans. Operational meteorological S-,C-, and X-band radars usually provide reflectivity scans with a temporal resolution of 5–15 min (Table 2), whereas research radars dedicated to high resolution rainfall monitoring in specific areas and specific elevations are reported to provide data resolutions down to 15 s (e.g. van de Beek et al., 2010; Mishra et al., 2016).
Typical operating resolutions and maximum ranges for different types of weather radars used in hydrological applications.
The main strength of radars for rainfall estimation is their capability to provide spatially distributed rainfall information. The spatial resolution of radar rainfall data is basically determined by the hardware and physics. The radial resolution (or range resolution) is a function of the pulse length (in principle the range resolution is equal to half the pulse length, Battan, 1973) and can thus be very small for all radar bandwidths. However, for operational radars the radial resolution is often an indirect function of the maximum unambiguous radar range due to storage and data transmission restrictions. Thus, each radar scanline is subdivided into a fixed/selected number of range bins, which eventually determines the range resolution of the data. For pulsed radars, the number of range bins is determined by the ratio of the maximum unambiguous range and the range resolution (i.e. half the pulse length). X-band radars with a shorter range than C- and S-band radars are therefore typically operated with a finer radial resolution, e.g. down to a minimum of 500 m. Radial resolutions between 3 and 100 m have been documented by e.g. Leijnse et al. (2010), van de Beek et al. (2010), Lengfeld et al. (2014), and Mishra et al. (2016).
The spatial resolution also depends on the azimuthal (or angular) horizontal
resolution, which is a function of the beam width determined by the size and
design of the antenna. In contrast to the radial resolution, the azimuthal
resolution decreases as a function of the radial distance from the radar.
Most operational weather radars use parabolic dish antennas with a beam width of approx. 1
Typical values for spatial resolutions and maximum ranges for operational
radars are provided in Table 2. Generally, operational X-band radars function
with both higher spatial and temporal resolutions than C- and S-band radars.
This is typically because X-band radars require smaller antennas than C- and
S-band radars to achieve the same angular resolution. There are, however,
examples of configurations of C- and S-band radars where high resolution data
are derived. Such
Example of radar reflectivity at four different Cartesian spatial resolutions over Aalborg, Denmark (lat: 57.05; lon: 9.92). The radar data are acquired with a Furuno WR-2100 dual-polarimetric X-band radar (Nielsen et al., 2015) at 1 min temporal resolution at 16:20:00 UTC on 25 July 2016. Black circles are rain gauges of the Danish Water Pollution Committee network.
Examples of radar reflectivity with four different spatial resolutions
covering an area (
Many applications of radar data in urban hydrology favour projected Cartesian (gridded) over polar data with decreasing resolution as a function of range. The limit for generating high resolution Cartesian data is mainly related to the azimuthal resolution and thus range. Two common methods of data projection are (1) the Constant Altitude Plan Position Indicator (CAPPI) in which scans at different elevation angles (comprising one volume scan) are merged in order to generate a radar product with altitude independent of range (however, with an inhomogeneous zone with changing from one elevation to the next one), and (2) the Plan Position Indicator (PPI) that applies one scan elevation only, and that thus has an increasing altitude as a function of range.
Due to the curvature of the earth, the refraction of the radar beam through the atmosphere, and wind drift of the raindrops, assigning a radar measurement to a specific point on the surface can be quite challenging. This should also be considered when working with high spatial resolution radar data, since it is not certain that the rain can be allocated with the same accuracy at ground level as at a specific elevation.
In order to increase the temporal resolution of operational meteorological radar data, especially for urban hydrological applications, some authors have developed methods to interpolate between radar images (Fabry et al., 1994; Atencia et al., 2011; Jasper-Tönnies and Jessen, 2014; Nielsen et al., 2014a; Thorndahl et al., 2014b; Wang et al., 2015a). The governing principle in these downscaling methods is to apply the advection field of the rain, and, by resampling in space, convert the spatial resolution into temporal resolution. The methods have been proven to give better local peak estimates of rainfall intensities as well as more accurate accumulated quantitative precipitation estimates in comparison with point ground observations. Jasper-Tönnies and Jessen (2014), Nielsen et al. (2014a), Seo and Krajewski (2015), and Wang et al. (2015a) have successfully converted data with a 5 or 10 min resolution into a product with 1 min resolution for use in urban hydrological modelling. The concept of advection interpolation works if the raw radar data are instantaneous. If radar data are averaged (by multiple scans) over a time period, advection interpolation will not be favourable and temporal resolution cannot be increased. In relation to urban hydrological modelling, where very fine temporal resolution indeed is needed for some applications (e.g. down to 1 min), the radar data based on instantaneous sampling are therefore preferable.
Considering the advective nature of rain, it is also clear that advection interpolation yields a better estimate of the area precipitation. Accumulation of instantaneous radar data with e.g. a 10 min sampling rate might result in a “fishbone” pattern consisting of periodical variability in rainfall accumulations. This is a result of the advection of rainfall between consecutive radar scans.
Commercial radar rainfall products (see Sect. 2.2.3) often provide data that are temporally accumulated or averaged; hence, a coarser temporal resolution of data can be found in these products.
In the literature, the impact of spatial and temporal radar data resolutions
on hydrological model responses has been studied intensively (Quirmbach and
Schultz, 2002; Berne et al., 2004; Villarini et al., 2010; Emmanuel et al.,
2012b; Gires et al., 2012; Liguori et al., 2012; Nielsen et al., 2012;
Schellart et al., 2012b; Vieux and Imgarten, 2012; Gires et al., 2013;
Lobligeois et al., 2014; Bruni et al., 2015; Gires et al., 2014b;
Ochoa-Rodriguez et al., 2015; Rafieeinasab et al., 2015; Wang et al., 2015a;
Thorndahl et al., 2016). Other than different spatial and temporal
resolutions of radar rainfall input data, these studies represent a vast
variety of different types, severity and number of events, radar types,
catchment sizes, shapes and slopes, catchment imperviousness, models, model
scales and resolutions (fully distributed, semi-distributed or lumped), model
outputs (e.g. peak flows, water levels, volumes, combined sewer overflow
(CSO) discharges, volumes, frequencies, inundation levels), objective
functions for evaluating and comparing results, etc. For these reasons, it is
hard to formulate general conclusions on the impacts of the spatial and
temporal data resolutions, since they largely depend on the studied set-up.
However, three significant findings could be identified related to the
requirements for spatial and temporal resolutions in runoff response
modelling.
Schilling (1991) and Einfalt et al. (2004) recommended a minimum temporal
resolution of 1–5 min and a minimum spatial resolution of 1 km for the
application of radar data in urban hydrology in general. Berne et al. (2004)
detailed this to The concentration time of the urban catchment or to a point of interest in
the system are of importance and affected by many factors. According to the
rational method (Kuichling, 1889), increasing concentration times will lead
to greater critical rainfall aggregation levels (in this case coarser
temporal resolution). Due to the dependence between temporal and spatial
resolutions described above, increasing concentration times will reduce the
demands for high spatial resolution. Thus, high space–time resolution is
required for the simulation of peak runoff responses (surcharge, local
flooding, etc.) upstream in an urban system. However, for the simulation of
total catchment runoff or basin storage, the requirements on resolution may
be reduced (Berne et al., 2004; Bruni et al., 2015; Rafieeinasab et al.,
2015). The ability to resolve rainfall adequately in time and space for urban
hydrological applications depends on the velocity of rainfall fields. By
studies of variograms at different temporal aggregation levels and analysing
runoff responses, Ochoa-Rodriguez et al. (2015) found a strong interaction
between the temporal and spatial resolutions and the impacts on urban runoff
response. Berne et al. (2004) suggested a relation between the temporal (
The type and severity of a storm might also set requirements to the
space–time resolution. A high-intensity convective thunderstorm with small
spatial extent will need a higher resolution in both space and time to be
resolved, in contrast to a stratiform long-duration storm. This is again
related to the runoff response of the system in question. Germann and
Joss (2001), Berne et al. (2004), Bruni et al. (2015), and Ochoa-Rodriguez et
al. (2015) suggested applying climatological variograms to characterize the
spatial structure of rainfall fields and investigating the spatial resolution
requirements (given a specific temporal resolution) in order to resolve the
spatial structure of rainfall fields in a sufficient way for urban
hydrological applications.
The use of radar data implies that the data are of good quality. There are numerous items such as radar hardware calibration, clutter removal, and overshooting/vertical profile correction (Michelson et al., 2005; Villarini and Krajewski, 2010) which may play a role before radar reflectivity data can be converted into reliable rainfall intensities. A thorough quality check and potential correction are therefore required. Disturbances for a good radar measurement may be undesired reflections off mountains or high towers, airplanes, ships, wind turbines, attenuation by heavy rain or hail, snow or melting snow instead of rainfall, anomalous propagation conditions, and others. Methods to test for these problems exist, and they are partly reduced by dual-polarization information from new-generation radars. The preprocessing of radar data by meteorological services usually only covers a part of the above points.
Observed radar reflectivity can be converted into rain rates (intensities),
but in order to produce valid
Radar reflectivity,
Many different methods have emerged in the last decade for adjusting rain rates estimated from reflectivities, and several profound review papers on different adjustment/merging techniques related to hydrological applications exist (e.g. Goudenhoofdt and Delobbe, 2009; Wang et al., 2013; McKee and Binns, 2016). For specific details we refer to these. Below, we present some of the most widely applied methods.
One of the simplest methods of adjusting radar rainfall data has been
proposed by Smith and Krajewski (1991), who introduced the concept of
The optimal temporal integration period or spatial aggregation level is to a large extent dependent on the representativeness of the gauges (gauge network density) to capture the temporal and spatial variability of the rain (e.g. Gires et al., 2014a). It is difficult to recommend specific gauge network densities for radar rainfall adjustment since the optimal value will depend on storm type, homogeneity of the rain gauge network, orographic features of the rain, adjustment methods, etc. Generally you will need a rain gauge network with a higher density for smaller aggregation levels or, in other words, the density of the rain gauge network will determine the temporal aggregation level of the radar rain gauge adjustment. McKee and Binns (2016) suggest conducting a sensitivity analysis in order to identify the effect of gauge density on rainfall estimation.
For annual precipitation measurements the WMO (2008) recommends 1 per
5750 km
MFB adjustment has an implicit range adjustment feature, in that, at least for storms that do not cover a large portion of the radar coverage, the gauges reporting positive rain will be within a close distance to each other and at a similar distance from the radar, and thus the computed MFB will in some sense compensate for range dependent bias.
As an extension of the MFB adjustment, the concept of
Spatial variability adjustment approaches and geostatistical merging of radar
and rain gauge data are developed to account for range dependence issues as
well as heterogeneous DSDs. They represent another range of methods which are
widely applied for QPEs. The concept here is to merge the spatial variability
of the radar rainfall fields into the interpolated rain gauge precipitation
fields in order to increase the spatial resolution of this product. The
interpolation can be performed by many different spatial interpolation
methods, e.g. variations of Kriging (Krajewski, 1987; Todini, 2001; Sinclair
and Pegram, 2005; Haberlandt, 2007; Goudenhoofdt and Delobbe, 2009;
Velasco-Forero et al., 2009; He et al., 2011; Berndt et al., 2014; Rabiei and
Haberlandt, 2015), or by
Other methods such as the singularity approach (Wang et al., 2015b) have been proposed in order to overcome problems with spatial smoothing as a result of the variograms in the Kriging-based methods. Geostatistical merging and spatially distributed bias adjustment is mostly applied for radar composites or in mountainous areas with orographic rainfall effects (e.g. Germann et al., 2006; Sideris et al., 2014). Merged rainfall products are described in Sect 2.2.3.
Another alternative to the optimization and sensitivity approaches of the radar gauge adjustment described above is to model errors and thereby acknowledge uncertainties in rainfall estimates (e.g. Ciach et al., 2007; Gires et al., 2012; Pegram et al., 2011; Villarini et al., 2014; Rico-Ramirez et al., 2015). It is expected that these uncertainty-based methods and development of rainfall ensembles for hydrological applications will gain more impact in future applications, concurrently with development in probabilistic/ensemble models for urban hydrology.
Today, most national meteorological services produce radar rainfall products
consisting of radar composites from national radar networks. They provide
state of the art corrected
These QPE products are often provided in a fixed Cartesian grid with data summarized over a fixed time period. In some cases only historical data in hourly or daily precipitation maps but in other cases also fine temporal resolution data are available.
Generating radar composites merging data from two or more radars might be subject to inconsistencies in radar data due to merging of data from different elevations, with different scanning strategies, and using different merging techniques. In application of commercial QPE products, it is important to be aware of these inconsistencies.
Operational real-time continuous adjustment of radar rainfall against rain gauges constitutes a challenge in comparison to event-based or discontinuous adjustment based on historical data (offline mode). Nonetheless, for real-time operation of urban hydrological systems, it is crucial to be able to produce valid rainfall estimates in an online mode. The real-time adjustment is especially difficult in the beginning of rainfall events with no prior rain gauge data recordings or in situations with large spatial rainfall variability. In these cases where rain gauge observations might be sparse and thus subject to domain sampling errors, bias adjustment might be dominated by random factors and can easily result in a erroneous adjustment (Seo et al., 1999; Krajewski and Smith, 2002; Nielsen et al., 2014a). The accuracy of a real-time bias adjustment is thus dependent on the temporal aggregation scale at which the adjustment is performed. The shorter the aggregation scale (e.g. hourly or sub-hourly) the larger the risk of erroneous adjustment due to sampling errors and the larger the aggregation scale (e.g. daily or super-daily) the larger the risk of errors due to changes in DSD and bias over the aggregation interval. Several authors apply MFB adjustment rather than area-based adjustment in real-time operation due to the fact that the latter is more vulnerable to rain gauge sampling errors (Seo et al., 1999; Borga et al., 2000). In order to avoid abrupt changes in bias several authors have suggested applying algorithms to smooth the bias in time, e.g. using Kalman filtering (Chumchean et al., 2006) or exponential smoothing (Seo and Breidenbach, 2002).
It is evident that for hydrological modelling, accurate rainfall estimates at
ground level are desired. Different adjustment methods and their impacts have
been investigated in recent studies, e.g. Quirmbach and Schultz (2002),
Tilford et al. (2002), Vieux and Bedient (2004a), Emmanuel et al. (2012a),
Gires et al. (2012), Goormans and Willems (2013), Wang et al. (2013),
Leonhardt et al. (2014), and Rico-Ramirez et al. (2015). It is difficult to
recommend one method of adjustment over another, since it to a large extent
depends on the application considered. Instead we have identified some of the
key issues related to the requirements of radar rainfall adjustment or
radar–rain gauge merging for runoff response modelling in urban areas.
The choice of the adjustment method depends on the required accuracy of the
spatially distributed rainfall in the application and the radar rainfall
product available. For catchments and spatially homogeneous rainfall events,
an adjustment using rain gauges inside or outside the catchment and a fixed
MFB adjustment might be sufficient to represent rainfall variability. For a
large catchment potentially covered by multiple radars, geostatistical
merging techniques are required to represent the variability in DSD within
the study domain, and thus more sophisticated techniques might be preferred
(see e.g. Wang et al., 2013, 2015b). Urban hydrological model outputs are subject to uncertainties associated with
rainfall inputs as well as representation of hydrological and hydraulic
processes, expressed in parameter and model structure uncertainties (e.g.
Freni et al., 2008; Thorndahl and Willems, 2008; Thorndahl et al., 2008;
Willems, 2008; Dotto et al., 2012). In cases where parameter uncertainty
estimation of such processes dominates runoff response, the rainfall input to
urban hydrological models may become less important. Instead of adjusting the
radar rainfall individually, some authors have therefore calibrated or
optimized hydrological models directly to match runoff response observations
without specific adjustment of the rainfall input (Krämer et al., 2005;
Ahm et al., 2013; Thorndahl and Rasmussen, 2013; Löwe et al., 2014).
However, this is recommended only if parameter or model uncertainties are
high and/or radar rainfall data adjustment is not possible, because it may
lead to error compensation with undesired consequences for prediction. It is of utmost importance that real-time adjustment of radar data reflects
the potential changes in DSD. In the case of rapid changes e.g. between
convective and stratiform precipitation, a bias shift might occur. The
aggregation time on which a bias (either mean field of spatially varying)
adjustment is performed should therefore be able to reflect these changes.
This will to a large degree also depend on the density of rain gauges
available for adjustment. Required gauge density for an unambiguous
adjustment will thus depend on the aggregation level on which the adjustment
is performed as well as the storm extent and homogeneity of the storm.
Next to the interpolation for urban design, control or scenario simulation applications, temporal extrapolation of radar rainfall fields forms the basis of real-time forecasting and control (e.g. Austin and Bellon, 1974; Einfalt et al., 1990; Sharif et al., 2006; Smith et al., 2007; Javier et al., 2007; Achleitner et al., 2009; Einfalt et al., 2009; Liguori et al., 2012; Schellart et al., 2012a; Wang et al., 2012; Thorndahl et al., 2013; Ntegeka et al., 2015). Due to the short response time of the urban drainage system and the short lifetime and small spatial size of convective rain cells, urban rainfall forecasts are only reliable for very short lead times (Achleitner et al., 2009; Foresti et al., 2016). Short-term forecasts are called nowcasts and provide input for real-time warning and/or control of urban floods or CSO pollution.
Several generic methods have been developed to nowcast radar data, based on
deterministic approaches, e.g. TREC (Rinehart and Garvey, 1978), CO-TREC (Li
et al., 1995), SCIT (Johnson et al., 1998; Mecklenburg et al., 2000), and
SCOUT (Einfalt et al., 1990), or stochastic approaches, e.g. MAPLE (Turner et
al., 2004), SBMcast (Berenguer et al., 2005, 2011), and STEPS (Bowler et al.,
2006). We refer to the individual papers for detailed descriptions of the
methods and focus instead on the application of nowcasts within urban
hydrological applications here. We have identified three issues which
constitute the current major challenges.
Despite development of the aforementioned methods, rainfall nowcasting for
urban drainage applications is still in its infancy. Although rain cells can
be extrapolated by radar image extrapolation (e.g. Thorndahl et al., 2013;
Löwe et al., 2014) or by applying cell tracking (e.g. Sharif et al.,
2006; Einfalt et al., 2009; Muñoz et al., 2015), this is often of limited
value given the limited duration of rain cells, especially during convective
conditions. The quality of an extrapolation-based nowcast depends on the
radar range, possible merging of radar networks, resolution, climate zone,
and rainfall type. For a standard deterministic nowcast, the lead time varies
between less than 30 min (small convective cells) and more than 2 h
(large-scale slow moving systems). As a rule of thumb, extrapolation is more
difficult with small rainfall cells and for small target areas, and less
difficult with large rain fields and large target areas. The most promising alternative to simple extrapolation of radar rainfall data
is to perturb the deterministic radar extrapolation with stochastic noise to
account for the unpredictable rainfall growth and decay processes (Bowler et
al., 2006; Germann et al., 2009; Liguori and Rico-Ramirez, 2013). The
stochastic noise model aims to describe the nowcast error together with its
spatial and temporal correlations. In the Short-Term Ensemble Prediction
System (STEPS), this is done by adding stochastic perturbations to the
deterministic Lagrangian extrapolation of radar images (Liguori and
Rico-Ramirez, 2013). The perturbations moreover aim to reproduce the dynamic
scaling of precipitation fields, i.e. the observation that large-scale
rainfall structures are more persistent and predictable than small-scale
convective cells. STEPS was originally co-developed by the UK Met Office and
the Australian Bureau of Meteorology, and is currently further customized for
urban applications, e.g. in the UK (Liguori et al., 2012; Liguori and
Rico-Ramirez, 2012), STEPS-BE for the Belgian version (Foresti et al., 2016).
It provides probabilistic ensemble nowcasts. So far, however, these
nowcasting systems have relied on radar data that are too coarse for urban
applications (e.g. 1 km resolution C-band radar data for STEPS-BE). Future developments will likely involve the use of higher resolution X-band
radar data. These are currently only available at experimental sites (e.g.
Application fields for radar rainfall in urban hydrology. Applications that have emerged significantly since Einfalt et al. (2004) are marked with bold. Numbers in parentheses indicate which sub-section discusses the particular application.
Present climate
Severe rainfall warning
Insurance claims
Nowcasting
Real-time hydrological models with data assimilation
Design of basins and pipes
The field of urban hydrology has over the last decade expanded the focus from analysis, design, and operation of urban stormwater systems and wastewater treatment plants. Today, the key drivers of research include urban city resilience to hydrological extremes, water and resource recovery, climate change impacts and adaption, as well as integration with other city planning and management disciplines, including urban development. This has led to a need for new and more diverse precipitation inputs, both to address the challenges mentioned above and also because urban hydrology is becoming more complex with implementation of sustainable stormwater management infrastructure. This increased complexity often implies that the spatial distribution of precipitation becomes even more important in both planning and operation of urban systems, and therefore urban hydrology will require better resolved rainfall products in the future. The current main application fields for radar rainfall in urban hydrology are shown in Table 3. As shown in the table, several new application fields have emerged over the last decade. Radar measurements can provide important contributions to these new fields. The improvements discussed in the previous section have also enhanced the possibility of using radar data in the existing application fields.
Long-term analysis of precipitation using single or multiple rain gauges has
been applied for several decades. They provide valuable information to
decision-making within design and analysis of urban water infrastructure,
both for quantifying uncertainty and for studying non-stationary behaviour
(e.g. Ntegeka and Willems, 2008; Madsen et al., 2009; Willems, 2013a, b;
Gregersen et al., 2014). When using single site rain gauges for large
catchments, simple areal reduction factors (ARFs) can be applied to account
for the spatial distribution of extremes (e.g. Sivapalan and Blöschl,
1998; Vaes et al., 2005; and Wright et al., 2014a). While this is sufficient
within a given catchment, it is not adequate to validate spatial rainfall as
modelled by
As radar data quality improves, it can also be directly used to estimate
precipitation extremes, for example in the form of traditional
Re-analysis of extreme events was mentioned in Einfalt et al. (2004) as an important field of application of radar rainfall and a field where good approaches had been developed. The continued development of radars has enabled very accurate re-analyses of historical events (Jessen et al., 2005; Smith et al., 2013; Yang et al., 2013; Thorndahl et al., 2014a, b; Wright et al., 2014a, b).
The field of distributed 1-D–2-D hydraulic/hydrological models for urban flood simulations has matured and standard methods have been developed (e.g. Zhou et al., 2012; Henonin et al., 2013). The state of the art described in these papers use point rainfall statistics as opposed to spatial rainfall inputs. Since there is large uncertainty in estimating volume estimates for high return periods, it is argued that the error not including spatial variability of rainfall within the catchment is minor.
However, recent studies have also partitioned the contribution of spatial and temporal variability in rainfall to urban flow quantiles, and shown that spatial rainfall variability does matter, especially for high return periods (e.g. Peleg et al., 2016a).
The paradigm of using point rainfall data from rain gauges at very high temporal resolution, assuming them to be representative of an entire urban catchment, is challenged by several factors. First, rainfall data from high resolution radar have shown high spatial variability at the intra-urban scale. Moreover, many cities experience substantial development in the form of urban sprawl. This leads to very large cities, where uniform precipitation cannot be assumed, because the catchment size is larger than the spatial representativeness of point precipitation. Hence, there is a far more complex hydrological response from large urban and peri-urban areas compared to smaller urban areas.
Another driver is the climate change adaptation needs of larger cities. Many countries and regions explicitly mention nature-based solutions or sustainable urban drainage systems as a very important component in this adaptation, including countries and regions such as China, the EU, and Australia. These wetlands, rain gardens, soakaways etc. are making the hydrological response of cities more complex. Hence, there is a need to generate spatially distributed rainfall series at high resolution in space and time. As mentioned in Sect. 3.1, such series are becoming available in a few places based on radar measurements. Means to develop artificial series based on stochastic properties are being investigated (e.g. Raut et al., 2012; Sørup et al., 2016), but there is a long way to go before standard procedures are identified. Over time, these procedures will most certainly be based on spatially distributed rainfall observations such as radar rainfall observations.
With a higher risk of damage due to heavy rainfall in urban areas as a consequence of climate change and increased urbanization, there is a motivation to develop reliable warning systems which have a higher level of detail regarding urban hydrology than traditional numerical weather prediction model forecasts of heavy rainfall, cloud bursts, hurricanes, etc. The evolution in computational power and models enables operational weather models to provide finer resolutions than just a few years ago. However, neither temporal resolution nor spatial resolution is currently fine enough to resolve rainfall sufficiently for many urban hydrological applications (e.g. Thorndahl et al., 2016). Furthermore, numerical weather prediction models may still have offsets of tens of kilometers in terms of predicting the exact location of a rain cell. This constitutes a significant problem in applying weather model data for urban hydrological systems, where the location of heavy rainfall is key. For short lead times this problem can to some extent be solved by assimilating radar nowcasts into numerical weather prediction models in order to improve initial conditions of the latter (e.g. Stephan et al., 2008; Dixon et al., 2009; Jensen et al., 2015). Operational systems with assimilation of radar data are rare, so in order to issue valid urban hydrological warnings, it can be beneficial to have (1) online rainfall estimates at high temporal and spatial resolution from radars and potentially also nowcasted data, as well as (2) online information on the current state of the hydrological system, e.g. baseflow, soil saturation, and residual storage capacity.
Examples of operational warning systems based on radar data are local flood warning systems, systems for emergency planning in case of flooding, warning systems for capacity of receiving waters, etc. Operational warning systems based on radar observations have potential in rainfall warnings if radar rainfall estimates exceed a specified threshold (e.g. Einfalt and Luers, 2015) or as hydrological warnings where radar observations (or nowcasts of radar data) are applied as input to an online hydrological model as described above. With regards to the latter there are still rather few applications of operational online distributed 2-D or 1-D–2-D flood warning models, since they tend to be too computationally expensive to run in real time. Instead simplified lumped models or 1-D models are often applied (Bell and Moore, 1998; Sharif et al., 2006; Javier et al., 2007; Smith et al., 2007; Fang et al., 2008; Einfalt et al., 2009; Duncan et al., 2013; Wolfs and Willems, 2017).
In the literature, there are various examples of real-time operation of urban drainage models, which are applied to warn if flow, water level, CSO volume, storage filling, etc., exceed certain thresholds, e.g. Yuan et al. (1999), Vieux and Bedient (2004a, b), Vieux et al. (2008), Achleitner et al. (2009), Liguori et al. (2012), Liguori and Rico-Ramirez (2012), Schellart et al. (2012a), Dirckx (2013), Thorndahl et al. (2013), Thorndahl and Rasmussen (2013), Löwe et al. (2014), Schellart et al. (2014), and Löwe et al. (2016). Several of these are pre-operational and have studied the potentials of applying radar data (with or without nowcasting) in real-time prediction of sewer system states.
Simulation of the probabilistic urban rainfall nowcasts in urban drainage models allows probabilistic nowcasts to be obtained of the inundation hazards and risks in urban areas. Ntegeka et al. (2015) have shown how probabilistic urban inundation risk maps can be obtained by combining STEPS-based rainfall nowcasts with a nested 1-D–2-D sewer hydraulic and surface inundation model, and a model to assess the damages and social consequences of the urban inundations (Van Ootegem et al., 2017). Such a system, however, only becomes useful for operational management when the uncertainties in the inundation risks can be communicated in a compact and clear way, and when these are informative and manageable by decision makers or the wider public.
Examples of runoff forecast prediction in an urban drainage system
in Frejlev, Denmark, using a radar ensemble nowcast algorithm (Jensen et al.,
2017). An ensemble of 300 nowcasts, a deterministic nowcast, and observed
radar data are applied as inputs to an urban drainage model (Thorndahl et
al., 2006; Thorndahl and Rasmussen, 2013) covering an area of 0.8 km
An example of a runoff forecasting system is provided in Fig. 3. The figure illustrates differences between runoff simulation with radar observations, a deterministic radar nowcast as well as a probabilistic nowcast with 300 ensemble members for forecast lead times of 10, 30, and 60 min.
Model-based real-time control of urban drainage systems has evolved significantly during the past decade. Many model-based real-time control methods were developed for applications with online in-sewer instrumentation or rain gauges for local systems (e.g. Schütze et al., 2004). With advances in estimating spatially distributed rainfall with radars, it is possible to implement real-time control on a much larger scale, e.g. a whole city. By exploiting the spatial variability of rain and successive unequal local loading of the hydrological systems, novel developed methods aim at utilizing spare capacity systems in order to reduce spills, overflows, flooding, etc. (e.g. Faure and Auchet, 1999; Pfister and Cassar, 1999; Mounce et al., 2014).
Other real-time control applications have been used to estimate the loads on waste water treatment plants in order to reduce spills of untreated waste and stormwater and to optimize treatment processes during rain (Quirmbach and Schultz, 1999; Fuchs and Beeneken, 2005; Thorndahl et al., 2013; Vezzaro and Grum, 2014; Kroll et al., 2016). With large linked hydrological systems, centralization of treatment plants in urban areas, advances in model predictions and data, there seems to be a large potential for global predictive control of hydrological systems in cities, which is not yet fully exploited.
This paper summarized literature findings from the last decade in three key
research areas:
In the following, we summarize emerging developments and applications of
radar rainfall in urban hydrology that were identified in this review and
provide recommendations for future research as well as practical
recommendations for the application of radar rainfall in urban hydrology.
A recent and promising development is the installation of X-band polarimetric
radars in urban areas, providing high resolution rainfall estimates,
typically at or below 1 min and 100 m, but with a shorter range than C- and
S-band radars. While X-band radar is sensitive to attenuation due to its
frequency band, the use of polarimetric signals provides additional
parameters insensitive to attenuation, thus solving an important problem
associated with X-band radars. While dual-polarimetric radars are capable of
providing an independent rainfall product, single-polarimetric X-band radars
on the other hand require extensive post-processing incorporating data from
additional sensors to obtain reliable, high resolution rainfall estimates. In
S- and C-band radar networks, high resolution products are starting to be
developed, based on for instance compressed pulse lengths. This reveals a
transition from use of primarily research radars with high resolution to more
operational products from meteorological services focusing also on high
resolution for urban hydrological application. Where high resolution radar rainfall products are not available, spatial and
temporal downscaling (advection interpolation) is applied to obtain higher
resolution rainfall estimates, starting from coarse resolution radar
products. Downscaling can be based on physical processes or on stochastic
principles, the latter being more flexible for including uncertainty and
being less computationally intensive, but also having more difficulty in
reproducing the natural, physical structure of storms. Radar rainfall estimates suffer from uncertainties associated with
variability in drop size distribution, partial beam filling, overshooting,
and signal attenuation. One way to reduce these uncertainties is by using
polarimetric signals, another way is by reducing distance to the radar, by
increasing the density of the radar network. Both require significant
investments and in many situations are not foreseen in the near future. This
implies that radar data adjustment based on a network of rain gauges will
still be required to reduce radar rainfall uncertainty. The quality of radar
data adjustment in turn depends on the density and quality of the rain gauge
network. The optimal temporal integration period or spatial aggregation level
for radar adjustment is directly related to the ability of the rain gauge
network to capture the temporal and spatial rainfall variability. It is
difficult to recommend specific gauge network densities for radar rainfall
adjustment since the optimal value will depend on storm type, homogeneity of
the rain gauge network, orographic features, adjustment methods, etc. as well
as the specifications of the urban hydrological application. In many studies
simple mean field bias adjustment between radar and rain gauges has proven
sufficient and robust which is probably also the reason that this method is
applied in many operational systems. At present, the more advanced
geostatistical approaches to bias adjustment are mostly applied within the
research community. Whereas numerical weather forecast models have too coarse a spatial and
temporal resolution for reliable forecasts in urban hydrological
applications, the use of short-term forecasting (nowcasting) of radar
rainfall shows potential in many online urban hydrological applications with
warning systems or real-time control of urban hydrological systems.
Currently, there are some drawbacks with pure radar extrapolation methods in
terms of predicting convective rainfall with rapidly evolving storm structure
evolution. In order to overcome these problems, stochastic blending of radar
rainfall observations/extrapolations with numerical weather prediction models
ensembles shows potential for fast hydrological response systems. Given the
high nowcasting uncertainties, the explicit consideration of these
uncertainties, e.g. by means of stochastic modelling approaches, is
important. Pluvial flood warning for small urban catchments based on critical rainfall
thresholds or pluvial flood warning based on real-time urban hydrological
modelling are expected to be developed significantly in forthcoming years in
order to adapt to climate changes and increased urbanization.
For urban hydrological applications in general, higher resolution and higher accuracy rainfall estimates are beneficial for a better understanding of the hydrological response. Higher accuracy comes with required investments in equipment (X-band radar, polarimetric capability or dense rain gauge networks for adjustment) that need to be justifiable from either a research or societal perspective (higher efficiency of operational control, more accurate early warning). Some general recommendations can be derived from the recent literature as to requirements for radar rainfall resolution: studies have shown that the sensitivity of the hydrological response and thus the added value of higher resolution rainfall data input increases for smaller catchment size, larger catchment spatial variability, smaller storm size, larger storm variability, and higher storm movement velocity. An important consideration here is that the accuracy of rainfall estimates generally decreases for higher resolution: the accuracy of storm total rainfall is typically much higher than for 5–15 min rainfall estimates; the same applies for spatial aggregation levels. Conversely, higher rainfall measurement resolution results in higher accuracy of rainfall estimates than if rainfall estimates are derived from coarse resolutions. In applications, a balance will always be needed between the benefit of higher accuracy and the required investment obtaining such accuracy. In a nowcasting and near-real-time context, challenges are even higher, because data correction and adjustment windows are typically short, while false early warnings can have large societal impacts.
Whereas the initial word of the title of the Einfalt et al. (2004) paper
(“
S. Thorndahl coordinated the joint collaboration and developed the greater part of the manuscript with partial contributions from other co-authors on radar uncertainties and data quality (T. Einfalt); nowcasting, real-time applications and uncertainties (P. Willems); spatial and temporal resolutions (J. E. Nielsen); X-band polarimetry, summary and recommendations (M.-C. ten Veldhuis); offline applications and future outlooks (K. Arnbjerg-Nielsen); technical radar specifications (M. R. Rasmussen); and radar rainfall extremes and proofreading (P. Molnar).
The authors declare that they have no conflict of interest.
We acknowledge the International Working Group on Urban Rainfall (IGUR) under the IWA/IAHR Joint Committee on Urban Drainage for providing the network supporting the collaboration in writing this paper. S. Thorndahl acknowledges Damian Murla Tuyls from the Department of Civil Engineering, Aalborg University, for constructive ideas and proofreading as well as David Getreuer Jensen, Envidan A/S, for contributions to the example in Fig. 3. Edited by: C. Onof Reviewed by: R. Uijlenhoet and D. Wright