Future Change in Urban Flooding Using New Convection‐Permitting Climate Projections

Rainfall intensity in the United Kingdom is projected to increase under climate change with significant implications for rainfall‐driven (combined pluvial and fluvial) flooding. In the UK, the current recommended best practice for estimating changes in pluvial flood hazard under climate change involves applying a simple percentage uplift to spatially uniform catchment rainfall, despite the known importance of the spatial and temporal characteristics of rainfall in the generation of pluvial floods. The UKCP Local Convective Permitting Model (CPM) has for the first time provided the capacity to assess changes in flood hazard using hourly, 2.2 km CPM precipitation data that varies in space and time. Here, we use an event set of ∼13,500 precipitation events across the three UKCP Local epochs (1981–2000, 2021–2040, and 2061–2080) to simulate rainfall‐driven flooding using the LISFLOOD‐FP hydrodynamic model at 20 m resolution over a 750 km2 area of Bristol and Bath, UK. We find that both the event set and uplift approaches indicate an increase in flood hazard under near‐term (2021–2040) and future (2061–2080) climate change. However, the event set produces markedly higher estimates of flood hazard when compared to the uplift approach, ranging from 19% to 49% higher depending on the return period. This suggests including the full spatiotemporal rainfall variability and its future change in rainfall‐driven flood modeling is critical for future flood risk assessment.

these projections-coupled with the significant flood risk already facing the UK (Bates et al., 2023)-plausible predictions of flood hazard under future climatic conditions are essential for adaptation (Skougaard Kaspersen et al., 2017;Zhou et al., 2012).
To assess how changes in precipitation may impact flooding, it is essential to understand what we mean by "flooding" in this study.Floods are commonly categorized by their driving process (e.g., fluvial, pluvial, coastal, and groundwater), with each of these having a different sensitivity to climate change.The two types of flooding most sensitive to rainfall are fluvial and pluvial flooding.Fluvial (or river) flooding is strongly controlled by both catchment and meteorological conditions including precipitation (Stein et al., 2021).Pluvial flooding on the other hand occurs as a result of intense localized rainfall falling directly onto the land surface.This is a particular hazard in urban areas, where surface water runoff can accumulate in low lying areas, and interact with sewer systems, resulting in damage to property and risk to life (Falconer et al., 2009;Ferguson & Ashley, 2017;Schroeder et al., 2016).This means pluvial flooding is particularly sensitive to changes in rainfall characteristics, and as a result changes in rainfall are often used as a direct proxy for changes in pluvial flooding, despite the limitations.
Flooding is modulated through many mechanisms, from the type of rainfall and its characteristics (i.e., duration, intensity, spatial pattern, magnitude), as well as through the interaction of rainfall with the topography and characteristics of a catchment (i.e., soil moisture, topography, hydraulic characteristics, human influence).Thus, just using changes in rainfall as proxy for flood hazard negates the complex interactions that modulate a flood response to a given rainfall input (Sharma et al., 2018).Yet, the exact definition of pluvial flooding in the literature differs, defining pluvial flooding as including small channels (Wing et al., 2018), or distinctly separating flooding from surface runoff with flooding from rivers and streams (Hankin et al., 2008;Rosenzweig et al., 2018).Moreover, research focusing on inland flooding driven by rainfall often models one type of flood hazard in isolation i.e., just pluvial or fluvial, despite a well-established understanding that these phenomena often interact and compound (Apel et al., 2016;Chen et al., 2010).As a result, in this paper we address this gap by modeling urban flooding from surface runoff, channels and rivers in a continuum, which we term "rainfall-driven flooding." As we categorize rainfall-driven flooding to comprise both pluvial and small-scale fluvial flooding, it is crucial to understand how pluvial flooding is currently modeled in practice.The methodology to estimate flooding driven directly from rainfall typically refers to pluvial flooding, thus here we briefly outline the generally accepted "best practice" for pluvial flood modeling in the UK (DEFRA, 2021;Environment Agency, 2020a;Natural Resources Wales, 2023).The principles outlined are also generally applied globally meaning the findings may be applicable elsewhere (Sampson et al., 2015).Pluvial flood modeling typically utilizes design rainfall events built by developing Intensity-Duration-Frequency (IDF) curves.This probabilistic method fits extreme value distributions to observed annual maxima of precipitation time series, producing curves of return periods of rainfall events with corresponding intensities and duration (Courty et al., 2019).These design events have been used to drive hydrodynamic models, from "rain on grid" models often used over large areas (Bates et al., 2023) to models also representing linkages with sewer networks which are often applied over small urban domains (Chen et al., 2010).Some studies also stochastically generate design rainfall events for a particular return period using a weather generator approach, allowing the spatiotemporal features of the IDF data to be modeled (Burton et al., 2008;Kilsby et al., 2007;Thorndahl & Andersen, 2021).Model outputs typically provide estimates of inundation depth and extent based on the rainfall return period as opposed to the return period of a given flood frequency.
To utilize this approach to model future pluvial flooding under climate change, uplifts derived from current and future predictions of rainfall from regional and/or global climate models (spatial resolution typically 10-100 km) are commonly used.Climate change uplifts are the percentage changes to rainfall between the present and future climate.These projections are then applied to IDF-derived design rainfall events as a percentage increase/ decrease to the original rainfall and either used to change depth-duration curves (Sayers et al., 2020), or ideally propagated through a hydraulic model to assess future pluvial flood risk (Environment Agency, 2020a).The benefit of this approach is that this is a relatively straight forward way to add updated climate change information to existing model set-ups.However, there are several limitations to this approach.First, the physical processes (i.e., convection) that drive the high intensity, short duration and often highly localized rainfall responsible for most pluvial flooding cannot be resolved using coarse regional and/or global climate models (Fosser et al., 2020;Kendon et al., 2019).Thus, uplifts generated using these models are unable to capture the impact of changes in local convective processes, potentially mischaracterizing future rainfall at local scales (Sayers et al., 2020).Second, the spatially uniform IDF uplift approach does not model the space-time variability of rainfall and how this interacts with catchment characteristics to create a real-world flood response.Third, the output flood hazard 10.1029/2023WR035533 3 of 17 represents the flood hazard in relation to the rainfall return period as opposed to the return period of a given flood frequency.Put simply, rainfall return period is not the same as the probability a given part of the floodplain will be inundated because of the non-linear dynamics of floodplain inundation over complex topography.Consequently, it is plausible that this approach leads to incomplete representations of rainfall, which are then propagated through to give erroneous estimates of future pluvial flood hazard.
Recently, Convection Permitting Models (CPMs)-a class of climate models with sufficient resolution to resolve convective processes at least partially-have allowed credible estimates of short-duration rainfall extremes relevant to rainfall-driven flooding to be generated (Kendon, Prein, et al., 2021).This provides promise for the development of rainfall-driven flood models which capture the inundation response to high intensity localized rainfall caused by convective processes, such as summer thunderstorms in the UK (Fowler et al., 2021).UKCP Local is such a data set which uses a CPM to produce hourly time series of plausible events at 2.2 km resolution for 12 ensemble members over the recent past (1981)(1982)(1983)(1984)(1985)(1986)(1987)(1988)(1989)(1990)(1991)(1992)(1993)(1994)(1995)(1996)(1997)(1998)(1999)(2000) and for future epochs (2021-2040 and 2061-2080) (Kendon, Short, et al., 2021).This data set captures changes in the future timing, frequency, and spatial patterns of rainfall at 2.2 km resolution (Kendon et al., 2023), which is important for the representation of pluvial flooding.As a result, UKCP Local has the potential to be applied as a full spatiotemporal time series in rainfall-driven flood models which: (a) considers the full spatiotemporal variability of rainfall and its drivers in flood-generating processes; and (b) accounts for localized convection rainfall processes.Furthermore, multiple ensemble realizations allow an uncertainty assessment in future change for a given Representative Concentration Pathway (RCP) and allows us to determine the return period of the flood hazard as opposed to the flood hazard generated by a given return period rainfall event.RCPs are global warming trajectories used by the IPCC to represent climate change through four different pathways of radiative forcing at 2100 (2.6, 4.5, 6, and 8.5 W/m 2 ) (IPCC, 2021).
Consequently, the release of UKCP Local across the UK provides the opportunity to assess how the application of CPM-driven spatiotemporal rainfall into a hydrodynamic model compares to the traditional IDF uplift approach when modeling rainfall-driven flooding under climate change (Sayers et al., 2020).To our knowledge, this is the first study to estimate rainfall-driven inundation risk modeling using a full spatiotemporal series of events selected from CPM-derived future rainfall ensembles, and the first to compare the flood hazard derived with estimates using an IDF uplift approach.This study addresses three questions: 1. To what extent does using UKCP Local as input into a rainfall-driven hydrodynamic model lead to different estimates of future rainfall-driven flood hazard under climate change relative to traditional IDF uplift approaches? 2. What is the cause of any differences between the two modeling approaches? 3. What are the implications of our findings for future rainfall-driven flood risk assessment?

Method Overview
We hypothesize that explicit modeling of spatiotemporal rainfall characteristics using events selected from ensembles of spatially varying rainfall times series (in this study UKCP Local) will result in significantly different rainfall-driven inundation patterns compared to using traditional IDF + climate uplift approaches.We address these research questions by using three rainfall inputs to drive LISFLOOD-FP-a flood inundation model which simulates flood flows by solving an inertial form of the shallow water equations over a two-dimensional regular grid (Bates et al., 2010).In this scheme, rivers and drainage channels are represented using the one-dimensional subgrid approach allowing streams smaller than the model resolution to be resolved (Neal et al., 2012).This study is executed on a 750 km 2 region of the UK, covering the cities of Bristol and Bath (see Figure 1), at ∼20 m spatial resolution.
Three rainfall inputs are used to drive LISFLOOD-FP, and produce return period maps of rainfall-driven inundation depth and extent.The three driving rainfall data sets are: 1. UKCP Local Event Set-Rainfall event sets derived from the space-time varying future rainfall ensembles from the UKCP Local 2.2 km (CPM-derived) rainfall data set (Kendon, Short, et al., 2021;Met Office Hadley Centre, 2019), described in Section 2.2.1.2. CEH-GEAR IDF-IDFs derived from CEH-GEAR observed historical rainfall (Lewis et al., 2019)  Figure 2 below presents an overview of the modeling steps.The CEH-GEAR IDF data set represents the current "best practice" approach for pluvial flood modeling in the UK under current and future climate change.The UKCP IDF input data set provides another version of this approach, using the UKCP Local data set to generate a different set of climate uplifts.Both IDF uplift approaches are compared to the event set approach whereby the rainfall event sets for both the historical and future epochs are derived from the spatiotemporal UKCP Local CPM-derived time series.The hydrodynamic flood model set up is explained in Section 2.5.The method for comparing the inundation maps from these three rainfall inputs is described in Section 2.6.
The full spatiotemporal time series from all 12 ensemble members are run in the event set approach (Section 2.3).UKCP Local is also used to derive IDFs for a rainfall uplift approach (Section 2.4).

CEH-GEAR
The Centre for Ecology & Hydrology-Gridded Estimates of Areal Rainfall (CEH-GEAR) data set provides estimates of hourly precipitation over 1 km grids across Great Britain and Northern Ireland from 1890 to 2019 (Lewis et al., 2019).The hourly estimates are derived from nearest neighbor interpolation of historical rainfall observations.Generally, rain gauge data is considered to underestimate the intensity of heavy rainfall due to measurement errors, meaning quality control is required (Lewis et al., 2018).The data set is widely applied in research, industry and the public sector in the UK.
CEH-GEAR is used in this work to derive a historical IDF data set, to which climate change percentage uplifts are applied as described below.

Climate Change Percentage Uplifts
We use climate change percentage uplift factors derived from Chan et al. (2021Chan et al. ( , 2022)).These were derived from the UKCP Local ensemble projections (Kendon, Short, et al., 2021;Met Office Hadley Centre, 2019) at 5 km resolution.An extreme value statistical model which combines a Point Process and Extreme Values Distribution approach is used to derive probability distributions and return periods for each 5 km grid point (Chan et al., 2022;Youngman, 2022).This method was applied to each of the 12 UKCP Local ensemble members to obtain return period rainfall intensity climate change percentage uplifts for 1-24-hr precipitation extremes for future 20-year epochs centered on 2030 and 2070 under RCP8.5 compared to a 1990 precipitation baseline (Chan et al., 2021).Estimates provide low (5th percentile), central (50th percentile), and high (95th percentile) projections using the method detailed in Fosser et al. (2020).These uplifts are now used to assist in urban drainage design and surface water modeling under climate change and provide climate change uplifts at 5 km resolution across the entire UK (Environment Agency, 2020a).This is a significant improvement over the default single uplift factors for the whole UK used previously (Chan et al., 2022).The central (50th) and high (95th) uplifts are used in this work to uplift IDFs derived from CEH-GEAR and UKCP Local data to provide two current practice benchmark rainfall input data sets for comparison with the event set method.

Rainfall Event Set Approach
We select rainfall events from the UKCP Local time series to simulate in the hydraulic model, rather than driving LISFLOOD-FP with the entire time series which would be computationally intractable and uneconomical given that much of the time series contains negligible rainfall.To select events, we defined an "upper threshold" (10 mm hr −1 ) above which events are considered large enough to be worth modeling.We also set a "lower threshold" (1 mm hr −1 ), below which events are considered to start and stop.However, to model flood recession, an additional "drain off time" was defined (6 hr), which is added to the stop time.In this location, 6 hr was long enough for a flood wave peak to reach the catchment outlet, which was determined through sensitivity testing of the drain off time.This demonstrates that we are modeling rainfall-driven flood events that respond quickly to rainfall and could be considered "flash floods" (Penning-Rowsell & Korndewal, 2019;Rosenzweig et al., 2018).Events are terminated once there is no subsequent rainfall above the minimum threshold during the "drain off time."This makes it unlikely that events caused by the same storm system (such as back-building convective storms) would be considered as independent events (Schumacher, 2009), as there would need to be a 6-hr gap with no rainfall across the entire model domain before an event was considered as completed.Furthermore, a pre-rain time period is added before the start time (1 hr).These time periods and thresholds were chosen through trial and error and manual inspection of the resulting events.This setup is illustrated in Figure 3.
This event selection process was applied to each of the 12 UKCP Local ensemble members, for the three epochs, resulting in 36 catalogs of rainfall events.Across the 12 ensemble members, there were a total of 4,056 events selected for the historical epoch, 4,765 events for the 2021-2040 epoch, and 4,746 events for the 2061-2080 epoch.In total, the entire catalog of 13,567 rainfall events were used as input to drive a LISFLOOD-FP model configured to model spatially and temporally varying rainfall fields (see Section 2.5.1 below).As LISFLOOD-FP includes a river network not all rainfall events will cause flooding because rainfall will be contained within the river channels.To our knowledge, this is the first time this number of realizations has been computed using a regional hydrodynamic model of such fine spatial resolution.
Rainfall intensities for five durations (1, 3, 6, 12, and 24 hr) and three return periods (2,30,100 years) were used to estimate the change in flooding under climate change for a given return period during the historical, 2021-2040 and 2061-2080 epochs.These time periods allow direct comparison with the event set epochs.Durations and return periods were determined based on the availability of corresponding climate change uplift estimates (Chan et al., 2021).As there too few samples for the 100-year return period in the event set, just the 2-and 30-year return period are used for the comparison.
The rainfall data used to derive the two IDF input data sets have different spatial resolutions.The UKCP Local data has a resolution of 2.2 km, whereas the CEH-GEAR IDFs are derived using a 1 km gridded data set.Moreover, the uplifts that are output have a 5 km resolution.However, in the standard uplift approach, the rainfall data for a given return level are averaged across the 750 km 2 hydrodynamic model domain to get a single rainfall intensity value for a set duration and return period.This means the sensitivity of output rainfall intensities to these differences in resolution should be negligible.Figure S1 in Supporting Information S1 compares an IDF derived using the CEH-GEAR data at 1 km resolution with an IDF derived from CEH-GEAR data aggregated to 5 km and shows no significant difference in rainfall intensity across all durations for the 30-year return period event.
The final rainfall intensities produced in both the UKCP Local IDF data set and CEH-GEAR IDF data set used to drive LISFLOOD-FP are the historical baseline, plus the historical baseline with a 50th and 95th percentile climate change uplift applied to produce 2030 and 2070 design flood events.These are summarized in Tables S1-S3 in Supporting Information S1.

CEH-GEAR IDF Approach
The CEH-GEAR IDFs were calculated for the specified durations and return periods by extracting annual maxima using a rolling mean, with a window size based on duration.A Generalized Extreme Value Distribution was then fitted to the annual maxima using L-moments, to obtain return period intensities for each duration.
The IDFs are represented as grids of rainfall intensity for each duration and return period at the same projection and resolution as the native 1 km CEH-GEAR grid.These calculated return period rainfall intensities were averaged across the model domain in order to generate spatially uniform rainfall intensity values for a set of return periods and durations.This is used as input to LISFLOOD-FP's standard rain-on-grid module and is the method used in the standard approach.It was deemed suitable to take the mean rainfall intensity from the IDF rainfall intensity grids, as there was only small variation in the rainfall intensity values for a given return level across this model region (<30 mm across the 750 km 2 model domain for 100-year return period).

UKCP Local IDF Approach
The UKCP Local IDF approach serves as another benchmark to compare the event set model against, highlighting the differences between the IDF and event set approach using the same data set.Furthermore, this model is compared against the uplifted CEH-GEAR IDF approach, to identify differences between models using the same approach.
IDF curves were produced based on all 12 concatenated UKCP Local ensembles to give 240 years of data for each 20-year epoch, using the same procedure to that used for the historical CEH-GEAR IDFs described above.
Uplifted rainfall intensities for the future epochs were similarly derived for 2021-2040 and 2061-2080 epochs by applying the same central and high uplifts (Chan et al., 2021).As with the CEH-GEAR historical IDFs, the UKCP Local historical IDF fields were averaged after the IDF calculation to produce spatially uniform rainfall intensities as input to the hydrodynamic model.

Hydrodynamic Model Set Up
The hydrodynamic model LISFLOOD-FP is used in this study.The model has been deployed successfully over local, regional, and global scales to estimate flood inundation from fluvial, pluvial, and coastal sources (Bates et al., 2021;Sampson et al., 2015;Wing et al., 2017), including for the United Kingdom (Bates et al., 2023).
Given its capability to model efficiently at fine resolutions over large regions, it is considered an appropriate model for this work (Bates et al., 2023).Key elements of the hydrodynamic model are a Digital Elevation Model (DEM), a stream network, and estimates of channel width, riverbank heights, and channel bed elevation (see Section 2.5.2).The base DEM is a composite of LIDAR and UK Ordnance Survey Data resampled to 1 arc second (∼20-25 m at this latitude) spatial resolution.The stream network was constructed using this DEM, using flow direction, and accumulation operations (Bates et al., 2023).Channels are assumed to be rectangular, with widths derived from empirical relationships of width with upstream catchment area (Sampson et al., 2015).Channel friction was applied as per the standard values across the UK and varied according to land use.Infiltration and sewer network information is not included in this model set up.This is considered an acceptable trade-off, as we only simulate extreme rainfall events (>10 mm hr −1 ) meaning it is likely the soil is saturated and the sewers over-capacity (Falconer et al., 2009;Guerreiro et al., 2017;Wehner & Sampson, 2021).However, this limits the possibility to investigate future changes in soil moisture on rainfall-driven flooding.
For the present work we have isolated the Bristol and Bath urban region from the UK model (Bates et al., 2023) and incorporated additional features so that it is better suited to the event set approach in this study.These are described below.

Spatiotemporal Rainfall
LISFLOOD-FP models rainfall-driven flooding as "rain on grid"-where rainfall inputs are defined across space in each grid cell of the domain.The input rainfall volumes are then routed over the land surface using the local inertial form of the shallow water equations or in steeper areas a simple routing model (Sampson et al., 2013).In order to use the spatially and temporally varying rainfall fields across the domain from the UKCP Local event sets rainfall approach, a new routine was developed in LISFLOOD-FP to read in these rainfall fields in NetCDF file format (Shaw et al., 2021).This replaced the previous routine which could only represent time varying spatially uniform precipitation.This provides functionality to represent rainfall intensity spatially and temporally-two characteristics that are important controls on the generation of pluvial flooding (Aldridge et al., 2020;Bernet et al., 2019;Guerreiro et al., 2017;Schaller et al., 2020).Accordingly, this is the first direct application of high-resolution CPM data into the LISFLOOD-FP hydrodynamic model.

River Channel Bed Elevations
In this study, channel bed elevations serve as a bias correction parameter to tune channel capacity to the input rainfall data (UKCP Local).To produce realistic flooding, it is important to set the channel capacity such that in-channel flow goes out of bank at a plausible return period.For undefended reaches, we adopt the standard assumption that the average river contains the historical 1-in-2-year flow return period (Pickup & Warner, 1976;Williams, 1978;Wolman & Miller, 1960).For defended reaches, we assume that flows go out-of-bank at a return period equal to their standard of protection.Flood defense locations and standards of protection were obtained from the Environment Agency AIMs database (Environment Agency, 2023a).
We define the riverbed elevations as follows using a two-step process.First, the channel bed elevations are set to be unrealistically low using a constant depth (set to 10 m for our relatively small rivers) from the bank heights, such that the bed slope equals the local bank slope and no fluvial flooding from channels can occur as it is all contained within the (deliberately over-large) channels.All 4,056 events extracted from the 12 UKCP Local historical ensemble members were run through the hydrodynamic model with these deep channels, yielding an estimate of maximum water depth at every pixel in the river network for each event.The return period of water depth in these deep channels can be computed empirically from this event set.This value (at the appropriate return period; 2 years for undefended channels, standard of protection for defended reaches) then becomes the channel depth which is subtracted from the bank heights at every pixel in the model network to provide an estimate of channel bed elevation across the study area.For more information on defining the riverbed elevations, please see Text S1 in Supporting Information S1.

Comparing the Event Set and IDF Approaches
To derive return period maps for the IDF approaches, the maximum flood extent from the five simulated durations (1, 3, 6, 12, and 24 hr) was taken as the maximum flood hazard for that return period.This approach assumes, as is standard in such studies, that a given return period rainfall directly translates into a given return period flood inundation.This is a key limitation of standard approaches.
However, in the event set approach, a more nuanced understanding of the return period for a given flood hazard extent and depth is possible.Here, the return period flood inundation is not necessarily equivalent to the return period rainfall.Probabilistic return period maps can be derived based on the number of times each pixel in the domain is considered flooded in the event set for a given flood depth.The return period flood extent for each depth threshold is calculated by determining the number of times we would expect a given pixel to flood based on the number of samples in the event set.For each epoch, flood maps for all return periods were derived, whereby in each event the pixel is considered flooded if the water depth is above the depth threshold specified (i.e., 10, 50, and 100 cm).For each pixel in the domain, the number of times each pixel is flooded above this threshold is calculated across all events in each epoch.These depth thresholds were selected based on guidance from the Environment Agency regarding levels of dangerous flood hazard (Environment Agency, 2019).With 10 cm of surface water cars can be damaged if driven through and water is likely to begin exceeding street curb height in some places (Environment Agency, 2020b).At 50 cm depth, property flooding of the ground floor is likely to occur, and at 100 cm the effectiveness of any property flood resilience measures is likely to be exceeded (Environment Agency, 2019).

Results
Across the three epochs in the event set, 13,567 events were simulated.To compare changes in rainfall-driven flood hazard, change in total flooded area (km 2 ) for three different depth thresholds (10, 50, and 100 cm) was used as an indicator to determine changes in flood hazard extent for the UKCP Local event set and the IDF approaches.
Comparisons here explore the 2-and 30-year return period flood hazard.There are not enough samples with the 100-year return period magnitude in the event set to produce a confident estimate at this recurrence interval (∼2 samples).This is because the event set total record length is 240 years, meaning a 100-year return period event is only statistically likely to have occurred twice.An event set with a longer time series would be required to estimate changes in the 100-year return period flood hazard using this event set approach.

Comparison Between UKCP Local Event Set Epochs
First, we investigated the differences between epochs in the event set, assessing the impact of climate change on flood hazard across the different time periods.Figure 4 shows that for all return periods in the event set, flooded area is higher under the two future epochs compared to the historical baseline.Flood hazard increases between the historical baseline and the 2021-2040 epoch, increasing further again at the 2061-2080 epoch.For example, for the 30-year return period (10 cm depth), flooded area increases by 7.68% in the 2021-2040 epoch, and 15.76% in the 2061-2080 epoch, showing a similar magnitude change between the three epochs at the 10 cm depth threshold.Thus, although change continues between epochs, we see no evidence of an increasing rate of change in flood hazard.The change in the slope of the flooded area at the 20-and 30-year return periods in Figure 4 corresponds with the associated sharp increase in locations that are flooded between the 5-and 20-year return period and the 30-and 100-year return periods, with a relative smaller slope change in the additional locations that are flooded between the 20-and 30-year return period.This is likely a result of the topography of the model domain.
Although at all return periods and all depth thresholds flood hazard increases under climate change across the epochs, there are differences in the magnitude of change between the return period and depth threshold selected.First, as the return period magnitude increases, the increase in flooded area across the epochs is greater.For example, percentage increases in flooded area are larger for the 30-year return period with a 40.23% increase (100 cm depth, 2061-2080) in comparison to the 2-year return period with a 26.5% increase (100 cm depth, 2061-2080) (see Table S4 in Supporting Information S1).Second, changes in flooded area between epochs are greater as depth increases.For example, at a 10 cm depth, the increase in flooded area for the 2-year return period for 2021-2040 is 6.85%, increasing to 11.81% for the 100 cm depth.This demonstrates that not only does the most hazardous flooding increase but does so at a greater relative magnitude than the less hazardous flooding.

Comparison Between UKCP Local Event Set and CEH-GEAR IDFs
Next, we compared the changes in flooded area in the UKCP Local event set and the CEH-GEAR IDF uplift approach (Figure 5).For all three epochs, flooded area is greater in the event set when compared to the corresponding flooded area in the CEH-GEAR IDFs, ranging from 19.11% to 49.37% higher in the event set depending on the return period, water depth and climate scenario considered (see Table S5 in Supporting Information S1).
Overall, flooded area also increases in the near (2030) and far (2070) IDF uplifted climate scenarios in comparison to the historical baseline in the CEH-GEAR IDF approach.These differences are in part due to model biases and part due to differences in approach-we disentangle these in Section 4.1.For the 30-year return period the 2030 95th percentile flooded area is higher than the 2070 50th percentile flooded area.This suggests that the spread across the ensemble members in this case is comparable to the climate change signal from 2030 to 2070.
Figure 6 illustrates how the flooded area simulated using the UKCP Local event set and CEH-GEAR IDFs differs across central Bristol for the 30-year return period flood event (10 cm depth threshold).Across all three epochs and depth thresholds, locations that are never flooded in the CEH-GEAR 30-year return period event are flooded in the UKCP Local 30-year flood hazard return period map.However, all locations that flood in the CEH-GEAR 30-year return period event map are flooded in the UKCP Local 30-year flood hazard return period map.For the historical epochs, at 10 cm, 18 km 2 are additionally flooded in the UKCP Local 30-year flood hazard return period than the CEH-GEAR 30-year return period event map, with 6 and 3.5 km 2 additionally flooded at the 50 and 100 cm depth thresholds.This is 23% of the total flooded area in the UKCP Local 30-year flood hazard return period map for the 10, 50, and 100 cm depth thresholds respectively.
Flood hazard maps for the 100-year return period are available through the Environment Agency (2018).These maps were not considered a like-for-like comparison to benchmark against as there is a mismatch of return periods between the data in this study and the Environment Agency maps, and the interaction between flood hazards is treated differently.However, the hydraulic model used in this study has been validated against benchmark data, with Critical Success Index scores of 0.62-0.84when like-for-like data is available for comparison (Bates et al., 2023), meaning we are confident that the flood inundation results from this study should perform in line with these findings.

Comparison Between UKCP Local Event Set and UKCP Local IDFs
Figure 7 shows changes in flooded area across epochs for the 30-year return period.The relative difference in flooded area between the UKCP Local event set and the UKCP Local IDFs is smaller in comparison to the CEH-GEAR IDFs, ranging from 5.62% to 16.97% depending on the return period, water depth and time period considered (see Table S6 in Supporting Information S1).This demonstrates that the reason for the differences between the event set and IDF approach are not only due to the differences between the input rainfall data and their biases, but also due to the approach selected, because the effect of the model biases are removed here.
Moreover, the UKCP Local IDF comparison with the UKCP Local event set demonstrate the same overall pattern across the 2-and 30-year return period as in the CEH-GEAR IDFs.Flooded area for the same climate scenario is higher in the event set than using the UKCP Local IDF approach, although the difference is smaller between the event set and the UKCP Local IDFs in comparison to the CEH-GEAR IDFs.This demonstrates that despite the same underlying precipitation data, the resulting flood hazard is different when using an event set approach in comparison to an IDF approach (see Figure S2 in Supporting Information S1).This has important implications for the choice of methodology when assessing changes in flood hazard under climate change.

Comparison Between UKCP Local IDFs and CEH-GEAR IDFs
UKCP Local IDFs produce a larger flooded area in comparison to the CEH-GEAR IDFs, ranging from 13.76% to 18.19% higher in the 2-year return period and 22.06%-32.54%higher for the 30-year return period, depending on depth threshold and epoch (see Figure 8 and Figure S3 in Supporting Information S1).This demonstrates three key findings.First, the UKCP Local IDFs produce higher relative flooded areas compared to the CEH-GEAR IDFs.Second, the differences in flooded area between the two input data sets is greater than the differences between approaches (see Section 3.3), hence the differences in input data dominate in comparison to the choice of approach.Third, the difference between flooded area across climate scenarios increases as return period magnitude increases.Overall, this comparison demonstrates how the rainfall differs between the UKCP Local event set and the CEH-GEAR observations and hence the bias between the two, since the same approach is used to drive the rainfall through to pluvial flooding.Other research has noted the difference between observations and the UKCP Local.Precipitation in UKCP Local is more intense than in the observations because convection is not fully resolved at 2.2 km resolution (Kendon, Prein, et al., 2021), whereas observations can be underestimates due to measurement errors (Lewis et al., 2018).

Discussion
To our knowledge, this is the first published work that applies high resolution sub-daily continuous climate data from a CPM into a high-resolution hydrodynamic model (∼20 m), simulating many thousands of rainfall-driven flood hazard events.once again demonstrates the importance of considering the impacts of climate change on rainfall-driven flood hazard when designing current and future flood mitigation and adaptation measures (Bates et al., 2021;Zhou et al., 2012).

Comparing Estimates of Flood Hazard in the Event Set and IDF Approaches
Although rainfall-driven flood hazard is expected to increase in line with increasing future precipitation extremes, this study has reinforced the importance of using hydrodynamic models to assess changes in flood hazard, as opposed to using changes in precipitation as a proxy (Sharma et al., 2018).We demonstrate that when the temporal and spatial characteristics of rainfall are represented using high resolution CPM data, the estimates of flood hazard are higher overall than the IDF uplift approach.Numerous studies have identified the importance of representing spatial and temporal characteristics of rainfall in the estimation of pluvial flooding (Aldridge et al., 2020;Bernet et al., 2019;Guerreiro et al., 2017;Schaller et al., 2020).Thus, because the temporal and spatial interaction of the rainfall with the landscape is considered in the event set-as opposed to just the relationship with the rainfall return period in the IDF approach-it is likely that the flooded area for a given return period in the event set provides a more realistic representation of flood hazard in comparison to the IDF approach (Johnson et al., 2016).As a result, as CPM data becomes more widely available, this study suggests that utilizing the full spatiotemporal data set will be important for accurately translating this data through to information on flood risk.
Moreover, this study also shows the nuanced relationship between changes in rainfall and changes in flooding.This is shown by the difference when comparing the IDF uplift and event set approaches when the same underlying precipitation data is used (see Figure 7).For example, for the 30-year return period rainfall, UKCP Local projects a 22.5%-40% increase in rainfall (corresponding to central-high uplifts) by 2070 (Chan et al., 2021), whereas the 30-year flood hazard increases by 16%-40% using the full event set approach (depending on the depth threshold).Thus, the percentage change in flooded area is not equivalent to changes in precipitation.
This research also highlights changes in magnitude of flood hazard under climate change.Although no evidence of an increasing rate of change in flood hazard is clear when comparing near (2021-2040) or future epochs (2061)(2062)(2063)(2064)(2065)(2066)(2067)(2068)(2069)(2070)(2071)(2072)(2073)(2074)(2075)(2076)(2077)(2078)(2079)(2080).In all approaches, our results suggest that as return period increases, the change in flooded area between epochs increases.This has important implications for many sectors (i.e., flood defense planning which often is undertaken to protect against the largest events), as not only does the most hazardous flooding increase but does so at a greater relative magnitude than less hazardous flooding.These findings support other recent studies investigating changes in flood hazard and the relationship between precipitation and flooding, who find the largest increase in hazard for greater return periods (Brunner et al., 2021;Swain et al., 2020).Supporting the premise that rainfall is not equivalent to flooding, in the UKCP Local precipitation data changes in hourly precipitation did not show a significant variation across return periods (Kendon, Short, et al., 2021).It is worth noting that this finding could also be due to a lower number of samples at higher return periods (>30-year return period-see Figure 4).
Our results demonstrate that the event set produces larger estimates of flooded area in comparison to estimates produced using an IDF uplift approach.Furthermore, the difference in flooded area between the UKCP Local event set and the CEH-GEAR IDFs is larger than the difference in comparison to the UKCP Local IDFs.This is likely because rainfall intensity is lower in the CEH-GEAR data set (Kendon et al., 2023).This may reflect a wet bias in UKCP Local (Kendon, Short, et al., 2021), but there is also a tendency for gauges to miss intense localized showers and there is also uncertainty due to the length of the observed rainfall gauge data (Guerreiro et al., 2017;Rajczak & Schär, 2017).The resolution of UKCP Local (2.2 km) is also much smaller than the average distance between rain gauges in the UK.However, despite the differences in absolute rainfall between the UKCP Local IDFs and CEH-GEAR IDFs, there is still an-albeit lesser-difference in flood hazard between the UKCP Local event set and UKCP Local IDF approach.Hence, this clearly suggests that the N-year rainfall simulated in the IDF approach does not necessarily equate to the N-year flood inundation.The difference being for example, that the 30-year flood hazard in the event set is representative of the probability of inundation, as it is derived from the number of times each pixel is flooded above the specified water depth.The flood hazard in the IDF approach is a signature of the 30-year IDF rainfall.Thus, the flooding as a result of the 30-year rainfall and the 30-year flood hazard are not identical.This highlights the usefulness of using an event set approach, as here we can identify the flooded area for the N-year flood hazard inundation as opposed to just the flooded area associated with the N-year rainfall.This provides important information regarding the choice of methodology when assessing the impacts of climate change on current and future flooding (Ranasinghe et al., 2021;Seneviratne et al., 2021).

Limitations and Future Work
There are two key limitations of this current work in which future work will directly build upon.First, this study only documents change in one small area (750 km 2 ) of the UK.It will be important to conduct this work in other regions of the UK, as well as for other parts of the world when and where high-resolution CPM rainfall data becomes available, to understand how generalizable these findings are to other locations.Although this was beyond the scope of this study, the CEH-GEAR IDFs could also be used to simulate design rainfall events that could be compared to the UKCP Local event set and IDFs using a stochastic weather generator approach.This could provide further understanding into specifically how the exclusion of spatiotemporal information in the CEH-GEAR uplift approach influences flood hazard, although it is likely this will support the key conclusions here regarding the importance of including spatiotemporal rainfall variability.Second, information on soil infiltration and sewer network capacity are not currently included in the hydrodynamic model set up.This means that the models may miss any future changes in soil moisture conditions that might amplify or dampen surface runoff.Currently, this is an acceptable limitation, since we are exclusively modeling extreme rainfall events where the capacity of the soil and sewer network is likely to be exceeded in many cases (Falconer et al., 2009;Guerreiro et al., 2017;Wehner & Sampson, 2021).Moreover, as we are directly comparing both the event set and IDF approach without the inclusion of infiltration, the relative differences in flooded area between the two approaches are likely to be similar even if infiltration is later included and the absolute values change.In future work, we plan to explicitly introduce infiltration into the modeling framework, as well as applying this method at the UK scale.

Conclusion
To the authors' knowledge, this research demonstrates the first known application of CPM data into a high-resolution event-based hydrodynamic model (∼20 m), simulating ∼13,500 rainfall-driven flood hazard events.Despite absolute differences in flood hazard estimated using the UKCP Local event set and IDF uplift approach, our results show that total flooded area for a given return period increases under climate change in both the near (2021-2040) and future (2061-2080) climate scenarios in comparison to the historical baseline.This highlights the importance of considering changes in precipitation under climate change when investigating current and future rainfall-driven flood risk.Within this analysis, we also show that estimates of flood hazard are higher for the Bristol/Bath area in the UKCP Local event set when compared to the IDF uplift approach using both UKCP Local and CEH-GEAR precipitation data.We conclude that the flood hazard estimated using the event set approach is likely to be more representative of the flood inundation return period in this study area in comparison to the IDF uplift approach which is traditionally used to estimate changes in flooding under climate change.As a result, we find that there are important considerations when choosing both the input data set and methodology for flood hazard assessment.This work therefore has significant implications for future flood hazard estimation in the UK following the release of the UKCP Local data.

Figure 2 .
Figure 2. Diagram showing the model set up for the UKCP Local event set approach, and the two IDF uplift approaches (CEH-GEAR and UKCP Local).

Figure 1 .
Figure 1.Map showing the study model domain, a 750 km 2 region of the UK covering the urban areas of Bristol and Bath and the surrounding areas.

Figure 3 .
Figure 3. Example schematic demonstrating how rainfall events were selected from the UKCP Local time series.

Figure 4 .
Figure 4. Figure showing the change in flooded area across the UKCP Local event set for the historical (1981-2000) and future epochs (2021-2040 and 2061-2080), for three flood depth thresholds (10, 50, and 100 cm).The dashed line represents the number of samples in the event set with a given return period.

Figure 5 .
Figure 5. Diagram showing the comparison in flooded area between the UKCP Local event set and the CEH-GEAR IDFs for the historical (1981-2000) and future epochs (2021-2040 and 2061-2080), for three flood depth thresholds (10, 50, and 100 cm) for the 30-year return period.

Figure 6 .
Figure 6.Map showing the flooded area (10 cm depth threshold) across central Bristol in both the CEH-GEAR IDF and UKCP Local event set (blue) in the historical (1981-2000) and future epochs (2021-2040 and 2061-2080).The red represents the additional areas flooded in the UKCP Local event set in addition to the flooding shown in the blue.The red thus represents the areas that are not flooded in the CEH-GEAR IDF approach but are flooded in the UKCP Local event set.