We present a transparent and validated
climate-conditioned catastrophe flood model for the UK, that simulates
pluvial, fluvial and coastal flood risks at 1 arcsec spatial resolution
(
Flooding is the principal environmental hazard identified in the UK's National Risk Register (Cabinet Office, 2020), and past major events have resulted in substantial economic damage and loss of life. For example, the coastal floods of 1953 resulted in 307 deaths, whilst inland flooding during summer 2007 inundated around 55 000 homes and left more than 400 000 people temporarily without drinking water. Despite significant investment in river and coastal defences over the last 50 years, including a further GBP 5.2 billion to be invested from 2021–2026 (Cabinet Office, 2020), floods continue to be a problem for the UK, with major events occurring in winters 2013–2014, 2015–2016 and 2019–2020, and in summer 2021. In England alone, the Environment Agency has previously stated that over 5 million properties, or around 1 in 6 of the total building stock, have a greater than 0.1 % annual probability of either fluvial or coastal flooding or an unspecified probability of pluvial flooding (Environment Agency, 2009). Flood risks in the UK will also very likely increase in the future as a result of population growth, changes in vulnerability and anthropogenic climate change (Merz et al., 2021; Committee for Climate Change, 2021).
The reasons for this are not difficult to see. The UK lies under the
westerly track of mid-Atlantic storm systems (including extratropical
cyclones) that can cause storm surges and extreme waves on exposed coasts
(Haigh et al., 2017). On making
landfall, these storm systems encounter extensive upland areas to the west
of the country, resulting in orographic enhancement of precipitation. This
subsequently falls onto river catchments that are (in global terms at least)
relatively short and steep, and therefore prone to flooding
(Black and Law, 2004; Luca et al., 2017). Convective rainfall activity in summer can be intense (Chan et al., 2016) and may lead
to flash flooding in urban areas and small catchments
(Archer and Fowler, 2021), whilst atmospheric
rivers can cause major flood events during winter
(Lavers et al., 2011). Along the eastern
seaboard lies the shallow marine basin of the North Sea, which is
effectively closed at its southern end by the Straits of Dover and is
therefore a setting extremely conducive to the development of storm surge
flooding (Horsburgh and Wilson, 2007).
Moreover, the UK is densely populated (281 people km
Despite the threat posed by floods, the methods currently used to map national-scale flood hazard and risk in the UK are, at best, opaque. The approaches adopted by government bodies and commercial organizations are largely undocumented, either in the peer-reviewed literature or in accessible reports, and validation studies are rarely reported. Indeed, considerable detective work is required to even understand what methods and data sets underpin existing flood risk information in the UK, despite these being used to inform critical long-term planning appraisals such as the UK's Climate Change Risk Assessment (Committee for Climate Change, 2021) or the national level of investment in flood defences (Environment Agency, 2019b). This lack of transparency is, from a scientific standpoint, unhelpful and unhealthy, and likely to hinder robust decision making. In addition, data sets available for flood research either represent only a limited range of return periods, do not consider the spatial correlation in flood hazard (cf. Heffernan and Tawn, 2004; Keef et al., 2009, 2012) or do not account for climate change impacts.
The purpose of this paper is therefore to address both technical and transparency issues in the flood risk information available in the UK. The UK is not unique in this respect: official flood mapping approaches in many countries lack scientific transparency and accountability (e.g. Pralle, 2019), and the methods are rarely subject to peer review. Lessons learned through this work are therefore likely to be more widely applicable, for example for the Federal Emergency Management Agency flood mapping programme in the US and for modelling conducted in support of the European Floods Directive in the European Union. Accordingly, we describe the development and validation of a climate-conditioned catastrophe risk model for UK pluvial, fluvial and coastal flooding. Current UK flood hazard and risk data sets are reviewed in Sect. 2. The methodology underpinning the model is described in detail in Sect. 3. Validation results and projections of current and future risk are presented in Sect. 4, and conclusions drawn in Sect. 5. Details of how to obtain the data for academic use are given at the end of the paper.
UK flood hazard and risk information at national and sub-national scales can be found in five broad classes of data product. A brief description of each follows, with further information in Sect. S1 in the Supplement.
Floodplain zonation (i.e. hazard) maps for fluvial, coastal and sometimes pluvial flooding are developed separately by government bodies in the four countries of the UK (Northern Ireland, Wales, Scotland and England), and are predominantly used to inform land use planning decisions. These maps are typically constructed using a patchwork of local modelling studies commissioned from commercial engineering consultants for individual river reaches. In England alone, there are over 2000 local hydraulic models that have been developed in this way to map the 1-in-100-year return period fluvial floodplain, with inundation in unmodelled areas likely to have been filled in from historic observations or local knowledge. The process is therefore similar to that used by the Federal Emergency Management Agency (FEMA) in the US (see Wing et al., 2018) and other hazard management organizations worldwide.
The individual reach scale modelling studies from which the national map is
derived typically use 1D, 1D/2D or 2D hydraulic models with airborne lidar
floodplain elevation data, bathymetric survey information, and most commonly
represent undefended conditions. For England, a 1-in-1000-year floodplain
layer was also created by a 50 m spatial resolution national-scale hydraulic
model developed in the early 2000s (Bradbrook et al., 2005), and pluvial
flooding has also been mapped nationally. For coastal flooding, simple
GIS-based “bathtub” models, which can severely overestimate areas at risk
(Vousdoukas et al., 2016), may be used
instead of true hydrodynamic simulations. Model boundary conditions are
usually derived from either extreme value frequency analysis of long
duration river, tide or meteorological records, or from a UK standard
regionalized flood frequency approach for discharge in ungauged basins in
the case of river flows (Robson and Reed, 1999). Where
available, model outputs may be calibrated to match observations of historic
floods. The resulting flood maps therefore represent average conditions over
the period of the instrumental record in the UK (typically from the 1960s
onwards at most sites). The true present-day hazard will therefore differ
from the recent historic average because of natural climate variability,
land use change and already-observed (but modest) changes in extreme
rainfall resulting from anthropogenic climate change (Kendon et
al., 2021). In combination, these effects have already led to observed
changes in mean annual flood magnitudes in different regions of the UK of
between
Beyond this broad overview, obtaining more detailed information to properly understand how UK flood hazard mapping is conducted is extremely difficult. Some small streams are not covered, and it is clear from national water body data sets that catchment headwater areas can be missed out. Despite this, no metadata exist showing the spatial limit of the mapping, so we cannot tell the difference between areas with no flood risk and those simply with no data because they have not been modelled. No information is provided on the individual local models, the specific data sets used, the model simulation performance or when the modelling was completed. The latter metadata are important because the complete national map required the commissioning of thousands of individual local studies and therefore took a significant amount of time to complete. As a result, some of the national map's component models are likely to be outdated. The strength of these flood hazard maps is that they are often created using well established hydraulic modelling approaches by trained engineers who may have access to local data that are not included in national databases. For this reason, and because of their official status, they are typically considered the “gold standard” in flood modelling, however no systematic assessment of such hazard mapping has ever been presented.
Flood risk maps or spatially aggregated risk data (i.e. the product of flood probability, exposure and vulnerability) are also produced in the UK, and are predominantly used to inform flood defence investment policy and long-term risk planning. Risk is quantified either in terms of the number of properties exposed to flooding of a given probability or the expected annual damage (EAD). More formally, the latter quantity is the integral of the loss-exceedance probability curve for a particular hazard. Data are presented as economic losses, and so represent the current value of assets that are damaged by the flood event minus any taxation element.
Only limited information on how this is done is available publicly, but it is apparent that the four countries of the UK vary in terms of the approach adopted, the flood probabilities which are reported, and which sources of flooding are considered (see Table S1). Most information is available in England, where flood risk maps are produced by the Environment Agency as part of their National Flood Risk Assessment (NaFRA) programme (Environment Agency, 2009). A summary of what is known of the method is provided in Sect. S1.2. In Scotland and Wales, only limited information on the method used is made available, and only the total number of properties exposed to flooding is openly reported. In Northern Ireland, it appears that a simple GIS overlay of flood hazard maps and exposure has been undertaken to calculate risk, but no further details are in the public domain. Spatial correlations in flood depths (cf. Heffernan and Tawn, 2004; Keef et al., 2009, 2012; Quinn et al., 2019) are not taken into account by any of these methods, so only average annual losses can be computed and not the full loss-exceedance curve.
No public validation of these risk outputs has been undertaken by the
government agencies responsible, but
Penning-Rowsell (2015, 2021) has shown how the methods
and output from the NaFRA analysis in England have changed significantly
over time. In particular, the raw output from NaFRA 2008 indicated
implausibly high flood losses (EAD of
Changes in flood risk as a result of climate change are obviously an
important consideration for policymakers, and this requires a consistent
UK-wide analysis. Given the differences in flood risk assessment methods and
reporting between the constituent countries of the UK, outlined in Sect. 2.1 and 2.2, this is simply not possible using the data sets described above.
Instead, the Future Flood Explorer methodology of
Sayers et al. (2016) and Sayers (2017) is used in the UK's 5 yearly cycle of Climate Change Risk
Assessment (Committee for Climate Change, 2021) to address
this limitation. Future Flood Explorer is a statistical emulation approach
that attempts to fill the gaps in existing hazard and risk information
available from the responsible authorities in the UK's constituent countries
(see Sect. S1.3 for further details). This spatially consistent
information can then be extrapolated into the future, allowing for different
climate, socio-economic and adaptation scenarios. However, the method
necessarily inherits any of the errors in the underlying hazard and risk
data sets produced in each UK country, and unsurprisingly produces similar
results for EAD to these methods (see Table S1). Similar to NaFRA, the loss
calculation in the method does not use depth–damage curves, and instead
substitutes a simpler approach based on inundated area only. Spatial
correlations in flooding between locations are also not accounted for. As a
result, only an expected annual damage can be computed and not the full
loss-exceedance curve. Validation of the outputs has been undertaken for the
number of properties flooded during the 2007 summer floods in England, which
showed a 2.2
A number of catastrophe risk modelling firms produce stochastic models of UK flooding on behalf of the insurance industry and other sectors. However, the methods and data from these schemes are typically regarded as commercially sensitive, and few details are available in the public domain. To date, there are no peer-reviewed journal publications or comprehensive public validation studies for the UK instances of these approaches, and the data are not available for academic use. Anecdotally we know that, unlike the publicly available data above, these methods do take flood spatial dependence into account, and so can provide the full loss-exceedance curve. Underlying hazard data are available for a range of return periods, and climate change scenarios may also have been computed for some of the models.
Finally, the Association of British Insurers (ABI) collates data on payouts
by its members in respect of residential flooding claims. These data have
been produced annually since 1998 and are described in detail in
Penning-Rowsell (2015, 2021). Unlike the UK government risk data, the figures
represent total financial losses (i.e. the actual money paid out) due to
flooding from all sources for the whole of the UK. The figure does not
account for losses incurred by the
It is clear from the above review and further information in Sect. S1 that important details of the methods and data used to create current flood risk products for the UK are not available in the public domain. Despite this, a lack of alternatives means that these data necessarily underpin nearly all current academic studies of flood risk in the UK (e.g. Rözer and Surminski, 2021; Sayers et al., 2018). Significant inconsistencies occur in the methods adopted between the different countries of the UK, and most approaches only represent historic average conditions rather than the present day or future. Validation is limited, and those data sets that are publicly available cannot be used to answer important scientific questions about extreme UK annual flood losses and the impacts of climate change.
To address the challenge identified above, we here describe a climate-conditioned catastrophe risk model for UK pluvial, fluvial and coastal flooding for historic, current and future conditions. A detailed description is provided in Sect. S3 and a summary in Fig. 1. The method can broadly be conceptualized as: (i) the creation of hazard layers across the UK for specified return period intervals and climate scenarios; (ii) the characterization of spatial dependence in flood discharge and synthetic event catalogue generation; (iii) the creation of flood event footprints through sampling from the existing hazard layers; and finally, (iv) the intersection of exposure data with vulnerability functions and event depths to estimate loss.
Method employed to compute climate-conditioned UK flood hazard and risk maps using a catastrophe model approach. See Sect. S3.1 to S3.3 for further details. Green cells represent data sources, blue cells represent the hydraulic model and its inputs and outputs, black and orange cells represent the same for the stochastic model, whilst purple cells designate the climate change analysis. Cells in bold represent the four main modelling stages of the method.
At the heart of the method (blue cells in Fig. 1) lies a standard 1D/2D
hydraulic model. In our case, this is a variant of the LISFLOOD-FP code
(Almeida et al., 2012; Almeida and Bates, 2013; Bates et al., 2010; blue, in bold in Fig. 1), however any comparable model would give similar results (cf.
Hunter et al., 2008). The
2D component of the model is run over the whole of the UK at 1 arcsec
spatial resolution (
Blue cells in Fig. 1 also denote the boundary conditions for the hydraulic
model and its main outputs, which are a series flood hazard maps for 10
different return periods from 1 in 5 to 1 in 1000 years for historical
conditions. Boundary conditions for fluvial, pluvial and coastal floods are
obtained from, respectively, a regionalized flood frequency analysis of UK
gauged flows from the National River Flow Archive, using an index flood
method similar to the UK's Flood Estimation Handbook (FEH,
see Robson and Reed, 1999), rainfall intensity–duration–frequency curves
derived from the CEH-GEAR1h precipitation database
(Lewis et al., 2018), and the likelihood of
coastal extreme water levels derived from the UK tide gauge network
(Environment Agency, 2019a). Data inputs to this process are shown
as green cells in Fig. 1. The baseline data therefore represent the
extreme value distribution calculated over the period of the historical
record (approximately 1960 to the present day for river flow and sea level,
and 1990–2014 for rainfall), with sea level de-trended based on 2018 mean
sea level values. The historic extreme event magnitude–frequency curves
therefore represent an average over the observation period, with a mid-point
around 1985–1995, which is when the planet reached 0.6
These historic boundary conditions are then adjusted to current (2020) and
future (2030, 2050, 2070) conditions using future climate projections from
the UKCP18 12 km regional simulations under the Representative Concentration Pathway (RCP) 8.5 carbon emissions
scenario (consisting of 12 ensemble members) and sea level rise projections
from Kopp et al. (2014).
Because we consider near-future projections of flood risk, the impact of
climate scenario choice is somewhat limited because, at least until
mid-century, the differences over the UK amongst the different emissions
pathways are relatively small. However, subsequent work should extend this
“proof of concept” to consider ensembles of climate models and a wider range
of emission trajectories. The UKCP18 projections shows no significant change
in storm surge, so we assume the UK storm surge climate (as captured in tide
gauge data) persists into the future. Changes in extreme precipitation and
sea level are used directly in the modelling, but for changes in extreme
river discharge, the future rainfall is used as input to a set of
Lastly, stochastic modelling is used to generate realistic event footprints by characterizing the spatial dependence in flooding (black cells in Fig. 1). The spatial dependence is determined using a conditional exceedance statistical model (Heffernan and Tawn, 2004; Keef et al., 2009, 2012), and this information is then used to sample synthetic events from the pre-computed hazard layers (cf. Quinn et al., 2019). By combining these hazard event footprints with exposure data and vulnerability functions, we are able to compute financial losses and obtain the full loss-exceedance probability distribution. Without such a stochastic method which includes spatial dependence, it is only possible to compute expected annual damage from a set of return period hazard layers.
Whilst the hazard model is run over the whole of the UK, suitable exposure data are not publicly available over the whole of Northern Ireland, so, for now, loss computations are restricted to Great Britain (Wales, Scotland and England). To determine exposure, we use the Verisk UKBuildings (formerly “Geomni”) data set, which gives information on property type, age and use for each building in Great Britain, and for vulnerability we use a modified set of standard UK depth–damage curves (the so called Multi-Coloured Manual approach; see Penning-Rowsell et al., 2013). Finally, loss results are presented in terms of specific global warming levels to decouple these from the carbon emission pathway (RCP8.5) simulated by UKCP18 Regional. The different warming levels therefore represent different future times, but an advantage of this approach is that it gives a degree of scenario-independence. Whilst the RCP8.5 trajectory is increasingly considered unlikely, we only use this scenario to extract results at specific warming levels, and so we are making no judgements about its probability.
Outputs from the hazard model were first compared to equivalent return
period flood extent maps produced by the different national administrations
in the UK (see Sects. 2.1 and S1.1). Specifically, we compared the 1-in-100-year return period fluvial and 1-in-200-year return period coastal
hazard layers produced by the historic run of the national model to the
complete set of equivalent flood hazard maps produced by the Environment
Agency for England and Natural Resources Wales (see Sects. 2.2 and S1.1),
using (mostly) a patchwork of local models. Whilst this is a model-to-model
comparison where neither simulation represents truth, we assume that the
local models can potentially have higher skill because they have been built
manually, often using local data to supplement national sources, and are
typically calibrated to match available flood observations. Important points
to note are that the national model in this paper simulates flooding in all
catchments down to just a few km
Contingency matrix of possible cell descriptors in a binary
classification scheme. The four possible states of the matrix are denoted
We compare the complete set of local maps to the national layers using the
following standard metrics:
The relevant equations for these metrics are
For more details on each metric, see Wing et al. (2017). In Table 2, we give the aggregate performance results for England and then Wales as a whole (top two rows) and for each English region separately, whilst in Fig. 2, we compare national and local hazard maps for a variety of inland and coastal locations across the UK. Performance scores for these specific sites are also given in Table 2.
Validation metrics when comparing the 1-in-100-year return period fluvial and 1-in-200-year coastal hazard layer from the model developed in this paper to equivalent government flood maps in England and Wales. The four validation metrics are hit rate (HR), false alarm ratio (FAR), critical success index (CSI) and error bias (EB). For further details on how these are computed, see Eqs. (1)–(4) and Table 1.
Comparison of fluvial and coastal hazard layers produced by the
national model developed in this paper (left-hand panels, labelled 1, in
blue) to equivalent government flood hazard maps (right-hand panels,
labelled 2, in red). Maps are shown for:
Table 2 and Fig. 2 show a coherent match between the two modelled layers with an overall CSI of 0.65 for England and 0.76 for Wales. CSI is a challenging metric as it penalizes both over and underprediction and ignores large areas of non-floodplain that are easy to predict. The metric is also sensitive to the shoreline length to inundated area ratio (Stephens et al., 2014), so that what constitutes a “good” match varies between sites. CSI values are therefore invariably less than 1, and this results from both model uncertainties (Hocini et al., 2021) and errors in observed data (Hawker et al., 2020; Horritt et al., 2001), both of which can be significant. To put the CSI scores achieved in this paper in context, comparison of modelled flood inundation extent with observations from airborne or satellite sources for individual river reaches typically results in CSI values in the range 0.65–0.9 (Aronica et al., 2002; Horritt and Bates, 2001a, 2002), with the higher value only ever achieved for sites with very high quality input and validation data (Bates et al., 2006; Neal et al., 2009). CSI values in the range 0.7–0.8 have also been obtained on the few occasions when separate remote sensing systems have acquired simultaneous images of the same flood (Bates et al., 2006; Biggin and Blyth, 1996; Schumann et al., 2009), which indicates typical uncertainties in remote observations of flood extent. Regional validation studies tend to produce slightly lower aggregate CSI values, mostly because regional models, unlike local ones, are never calibrated or optimized to fit the observed data. Example CSI values in regional scale inundation modelling studies to date therefore include 0.36–0.43 in Ward et al. (2017), 0.56–0.67 in Sampson et al. (2015), 0.76 in Wing et al. (2017) and 0.78 in Bates et al. (2021), which also shows the general pattern of improvement over time as regional models have become more sophisticated.
The national model developed in this paper and the local ones developed by
the UK's national administrations therefore have differences similar to
those between local models and observations, or between simultaneous
observations of the same flood using different sensors. The performance of
the UK national model is also in line with that of recent regional
inundation models created for other territories (Bates
et al., 2021; Wing et al., 2017). We conclude that our model is a plausible
representation of the UK flooding system, with errors in inundation extent
likely similar to those in either observations or local models. At the four
sites examined in Fig. 2, the similarity to government hazard maps is good
for the fluvial flooding examples (panels a and d, CSI values of 0.84 and
0.68, respectively) and the coastal and river flooding in Somerset (panel c,
CSI
Next, we compared modelled historic water depths to high quality
observations of maximum water height for a major flood that occurred in the
UK city of Carlisle in 2005. Flooding occurred in the city centre and
extensively through surrounding districts, with approximately 1900 homes
inundated. Subsequent to the event, survey teams from the Environment Agency
and the University of Bristol mapped wrack and water marks using
differential GPS systems (Neal et al.,
2009), with a precision of
Parkes and Demeritt (2016) estimate the return
period of the 2005 Carlisle flood to have been
Simulating maximum water elevations in a dense urban area is not
straightforward, and the Carlisle simulation presents a difficult test for
any hydrodynamic model. Nevertheless, comparison of the national model to
observations yielded root mean square error of 0.41 m, mean error of
Estimated annual damage values for previous modelled analyses (England's National Flood Risk Analysis, NaFRA, and the UK's Third Climate Change Risk Assessment, CCRA3), observed insured losses from the Association of British Insurers (ABI) and the new model analysis conducted for this paper. These figures represent direct financial losses due to fluvial, pluvial and coastal flooding in 2020 values for residential and non-residential properties in Great Britain.
Finally, the expected annual damage produced by the catastrophe model was validated by comparison with the observations of annual insured losses compiled by the Association of British Insurers (ABI) discussed in Sect. 2.5 and given in Table S1. To convert the ABI's residential-only losses from 1998 to 2018 for the whole UK to combined residential and non-residential losses in 2020 values for GB only (as produced by the catastrophe model), we broadly follow the approach outlined in Penning-Rowsell (2021). We therefore adjust the ABI data to make allowance for underinsurance, the ABI's incomplete market share and observations of the ratio of residential to non-residential losses from past UK flooding episodes. We can also compare these adjusted EAD values to previous model analyses of flood loss undertaken by government agencies in the UK and the third UK Climate Change Risk Assessment (CCRA3, see Sects. 2.3 and S1.3). Of the four countries which comprise the UK, flood losses are only publicly reported for England as part of their NaFRA programme. Wales, Scotland and Northern Ireland have their own flood risk mapping programmes with different methodologies that only report number of properties exposed and not financial losses (see Sects. 2.2 and S1.2). To create a GB loss, we therefore scale the NaFRA result for England using the ratios reported in Penning-Rowsell (2021). These were taken from the emulation methodology used in the 2017 UK Climate Change Risk Assessment (Sayers, 2017), which determined that England accounts for 79 % of flood losses, Scotland 12 %, Wales 6 % and Northern Ireland 2 %. Next, to convert the upscaled NaFRA and CCRA3 estimates of economic loss to financial values, we need to allow for betterment and taxation and to adjust the NaFRA EAD so that it also represents losses due to pluvial flooding. All data sets are normalized for inflation and increasing gross domestic product to 2020. Further details of the corrections made to bring the data onto a consistent basis are given in Sect. S2 and the final EAD values are reported in Table 3.
Of the modelled EAD of GBP 730 million, GBP 382 million of the losses come from fluvial flooding, GBP 150 million from pluvial and GBP 198 million from coastal, although the boundary between what constitutes fluvial and pluvial flooding is somewhat arbitrary. The small deviation between our modelled EAD and the ABI data is pleasing, and likely within the range of observational error (resulting from approximations during loss adjustment, errors in reporting, etc.). However, the 20 years of ABI historical observations represent just one realization of the insured losses that could potentially occur during this period. The catastrophe model developed here simulates a 10 000-year synthetic catalogue of flooding, and therefore includes very low probability, high loss events that may not be present in the ABI's historical record simply due to chance. To get a sense of likely uncertainty in the ABI data as a result of the undersampling of very extreme events, we randomly selected 10 000 periods of 20 years from the catastrophe model and calculated the EAD for each of these. The frequency histogram for these random 20-year samples is reported in Fig. 4, along with vertical lines representing the ABI, NaFRA and CCRA3 expected annual damages.
Simulated expected annual damage from 10 000 random samples of 20-year time periods for the model developed in this paper, for 2020 conditions compared to adjusted values from previous model estimates from the Environment Agency's National Flood Risk Analysis (NaFRA), the UK's Third Climate Change Risk Assessment (CCRA3) and observations from the Association of British Insurers (ABI). The grey shading denotes the interquartile range of the ABI data.
Figure 4 makes clear the likely large impact on EAD of random sampling of
very extreme floods during any finite period of historical data.
Nevertheless, the observed ABI data sit squarely within our bootstrapped
loss distribution. Expected annual damages from previous UK model analyses
(NaFRA, CCRA3) are 3 times larger than ABI observed losses (as
previously noted by Penning-Rowsell, 2021), and lie well outside the
distribution of 20-year samples of loss from the catastrophe model reported
here. Based on our distribution of losses, one would expect values
equivalent to NaFRA and CCRA3 expected annual damages to occur every
Based on the hazard and risk validation evidence presented above, the catastrophe model developed here appears to be a reasonable representation of UK flood patterns and losses. To examine how climate change will impact flooding in Great Britain, we calculate loss-exceedance curves in 2020 monetary values from the catastrophe model for specific global warming levels above pre-industrial (1850–1900). The loss-exceedance curve shows the probability that a particular total annual loss will be exceeded in any given year. These are shown in Fig. 5, with values for the average annual and 1 % annual probability (1 in 100 year) loss given in Table 4.
Great Britain loss-exceedance curves due to flooding (in GBP billion at 2020 values) for different specific global warming levels since the pre-industrial (1850–1900).
Great Britain expected annual and 1 % annual probability flood losses in GBP billion at 2020 values for different specific global warming levels since the pre-industrial (1850–1900).
A specific global warming level of 1.1
Figure 5 shows that, according to our model, changes in flood losses over
Great Britain due to climate change alone between recent historical average
conditions and the present day have so far been minor, with differences only
really emerging for annual loss return periods greater than 70 years. The
increase in EAD between these warming levels is only 1.5 %, whilst the
1 % annual probability (1-in-100-year return period) loss with
1.1
Figure 5 also shows that Great Britain will only be able to avoid major
increases in flood risk due to climate change if all countries' current
COP26 and net zero emission reduction pledges are met in full and warming
above pre-industrial is limited to 1.8
Spatial distribution of expected annual damage (EAD) due to
flooding for historical conditions (0.6
The curves in Fig. 5 represent UK aggregate losses, however this conceals
important changes in the geography of risk (Fig. 6). Figure 6 shows
absolute expected annual damage aggregated to 10 km hexagons across Great
Britain for historical average conditions (0.6
Current assessments of flood hazard and risk in the UK lack transparency, are insufficiently validated and, with very few exceptions, are not exposed to independent peer review. Whilst the public availability of the data sets is impressive by international standards, the methods used in their creation are clouded in secrecy. Calls for proper peer review of UK national flood risk assessments have been made before (Penning-Rowsell, 2015), but have effectively been ignored. The methods are therefore not repeatable by others.
This situation slows down and hampers model review and improvement cycles
whilst restricting the number of researchers that can contribute to the
effort, thereby creating significant barriers to progress. To give an
example of this in operation, the NaFRA methodology used in England was
developed by Hall et al. (2003) with a new major version in 2008 and a further major release in 2018 (Penning-Rowsell, 2021). A wholesale revision
of the methodology (NaFRA2) is currently underway and is scheduled to come
into operation in 2024 or 2025. Major model updates are thus
This paper attempts to kick-start this process for the UK by demonstrating a first climate-conditioned catastrophe risk model for UK flooding, which shows skill at simulating both hazard and risk. A model-to-model comparison with official 1-in-100-year return period fluvial and 1-in-200-year coastal hazard maps across the whole of England and Wales gave an overall critical success index (CSI) value of 0.65 for England and 0.76 for Wales. CSI is a challenging metric to maximize, as it penalizes both under and overprediction, and ignores easy to predict dry areas, but this value is similar to that obtained in comparisons of local hydrodynamic models to remote sensing observations of flooding (Aronica et al., 2002; Horritt and Bates, 2001b, 2002). CSI values over the UK are slightly lower than those obtained in a similar study over the US (Bates et al., 2021), in part because the US has more big rivers which are, in general, easier to model. CSI values for the UK are also influenced by the different methods employed in coastal areas by the local and national models i.e. bathtub GIS models used in local studies and the hydrodynamic approach used here at national scale, which is likely to be more accurate. In fluvial areas, the similarity between the local and national models is generally better (see Fig. 2a and d), and comparison of maximum water levels predicted by the national model for the 2005 Carlisle flood gave an RMSE of 0.41 m compared to 0.36 m for a high-resolution local model that was built and calibrated with site-specific data. Most importantly, the national model was, unlike existing schemes, able to provide a good match (i.e. one that is within likely error) to observed annual flood losses from the Association of British Insurers (ABI).
Our model analysis of course comes with a number of caveats. Driving the analysis with different climate models would change the detail of local predictions, however there is agreement on the broad patterns of UK climate change, and the 12 km regional climate model used in UKCP18 is the current official estimate of UK future climate. Hydrological modelling contains significant uncertainty arising from the boundary forcing, input data, model parameters and calibration (Beven, 2006; Coxon et al., 2019), and hydrodynamic models are predominantly sensitive to DEM (digital elevation model) and forcing errors as well as the quality of nationally available flood defence information. The latter is likely a key limiting factor for any large-scale flood inundation analysis, and whilst some workarounds are possible (e.g. Wing et al., 2019), these are by no means perfect. We also use a single change factor for all event return periods, which may be an oversimplification (see for example Bertola et al., 2020). More sophisticated work could use a multi-model ensemble of climate model simulations (e.g. Cloke et al., 2013), run multiple simulations to account for uncertainty (e.g. Keef, et al., 2012), use higher model resolution over urban areas (Fewtrell et al., 2008) and take into account the probability of defence failure (Shustikova et al., 2020). We also need to find better ways to recover and assemble local and ad hoc data (e.g. on river bathymetry, flood defences and validation data) into consistent national databases and develop algorithms to replicate the decision making of skilled local modellers in automated ways. Our ultimate goal should be to create national models which have equivalent performance to local approaches. Nevertheless, the model simulations shown here do have skill and represent a significant advance on previous work, such that there can be confidence in the broad conclusions that we draw.
Expected annual damage (EAD) due to flooding in official UK data gives
values that are
Our modelling shows that, even with a more sensible loss distribution than
official UK government estimates, the COP26 2030 pledges on decarbonization
are not on their own sufficient to restrict increases in UK flood risk to
It is also clear that the UK is not well adapted to the flood risks it
currently faces, let alone any further increases in risk due climate change.
Current expected annual damages of
In summary, we have presented a plausible and sober assessment of current and future UK flood risk. The analysis contains a greater level of detail and nuance compared to previous work, and represents our current best understanding of the UK's changing flood risk landscape. Whilst we should be cautious of over-interpreting the fine-scale spatial detail of the predictions, we expect that the national-scale results and broad regional patterns can be used in framing policy. The complexity of the climate-driven change we find in UK flood risk is likely to ring universally true in other parts of the world and should cause us to question simplistic flood risk projections and policy responses.
A version of the LISFLOOD-FP model similar to that used in this work is available at
The supplement related to this article is available online at:
PDB, JS and OW conceived and wrote the paper. JS, OW, NQ, AS and CS undertook the analysis and produced figures, whilst JN contributed key methodological developments and detailed review.
The contact author has declared that none of the authors has any competing interests.
Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
The authors are extremely grateful to the two anonymous reviewers and
This research has been supported by the Natural Environment Research Council (grant nos. NE/S006079/1 and NE/S015795/1).
This paper was edited by Bruce D. Malamud and reviewed by two anonymous referees.