Population exposure to flooding in Small Island Developing States under climate change

Estimates of coastal, fluvial, and pluvial flood hazard for all SIDS were simulated using a hydrodynamic model applied at 1 arcsecond (∼30 m at the equator) resolution, as described in Wing et al (2024). This global hydrodynamic model builds upon the framework described in Sampson et al (2015) and uses the LISFLOOD-FP hydrodynamic code to simulate in-channel and floodplain flow with a coupled 1D/2D subgrid model approach (Neal et al 2012) that solves the local inertial formulation of the shallow water equations (Bates et al 2010). In the absence of a global river bathymetry dataset, river channels are sized using a gradually varying flow solver (Neal et al 2021) to convey the 1-in-2 year channel bankfull capacity (Pickup and Warner 1976). A global 30 m forest and building removed Digital Terrain Model (FABDEM) derived from Copernicus digital elevation model (DEM) (European Space Agency and Sinergise 2021) is used as the topography input to the model (Hawker et al 2022). FABDEM is one of a number of copernicus DEM-based terrain datasets which demonstrate improved topographic representation when compared to Shuttle Topography Radar Mission-based DEMs (2022, Bielski et al 2024, Hawker et al 2024, Pronk et al 2024). More detailed information on the global flood hazard model setup can be found in supplementary text 1.1 and Wing et al (2024).

The hydraulic engine and flood hazard model used in this work has been extensively validated at the local (Neal et al 2009, Fewtrell et al 2011); national (Wing et al 2017, Bates et al 2023); continental (Wing et al 2017, 2021); and global scales (Sampson et al 2015). The latest such set of tests (Wing et al 2024) showed that the global model used here could predict flood inundation patterns at ∼30 m spatial resolution with an accuracy that approaches that of local modelling studies and satellite observations and, in addition, could simulate maximum extreme flood water levels with a root mean squared error of ∼0.6 m when compared to (uncertain) post event observations. As outlined in Wing et al (2024), the validation of this global flood models is the most comprehensive validation of a global flood model that currently exists, in thirteen locations across the US, UK, Europe and Africa. The critical success index (CSI) of flood extents are calculated against national and local scale models, as well as satellite observations, from thirteen validation data sources across a range of return periods. Water levels are also compared by calculating the mean absolute error between the global flood model water heights and state-of-the-art modelled or observed water heights. Moreover, previous work (Sampson et al 2015) has shown that flood extent over-and underprediction errors at the model native resolution largely cancel when aggregated to spatial units >1 km², which should result in unbiased estimates at whole-island scale. There is no obvious reason to believe that the skill demonstrated in existing test cases is therefore not broadly transferable to SIDS, but nonetheless, to sense check this assumption, two further validation tests were employed in this study.

DEMs are a key control on flood model performance (Hawker et al 2018)—especially for coastal hazard simulation in atolls, where the total elevation range in these locations (<2 m) has previously been smaller than the vertical error in global DEMs (2–10 m: Farr et al 2007, Rizzoli et al 2017). Firstly, we assess the performance of FABDEM in representing low-lying elevations where LiDAR data is available as a benchmark. Through a comparison of FABDEM against a LiDAR DEM for Majuro atoll in the Marshall Islands (Palaseanu-Lovejoy et al 2018), the mean error was 1.78 m, with a mean absolute error of 2.32 m, and root mean square error of 3.22 m. These errors are higher than average for low-lying areas (1.2 m: Gesch 2023), and for FABDEM built-up areas (1.22 m), and below average for FABDEM in forested areas (2.88 m) (Hawker et al 2022). The distribution of error in the histogram (see supplementary figure 5) shows that most of the error is below 2 m, which is a marked improvement over SRTM-based DEMs in a small island context (Rodríguez et al 2006). Although vegetation and urban artefacts appear to be removed in FABDEM based on qualitative inspection of the DEM, a vertical bias remains in the final product. The larger error in atolls such as the Marshall Islands has also been found in other Copernicus DEM-based DEMs such as DeltaDTM (Pronk et al 2024). This suggests that coastal flood hazard will therefore be underpredicted in these locations.

In the absence of LiDAR or ICESat data for comparison in other atolls, we qualitatively investigated the representation of topography in FABDEM in atoll islands (Kiribati and Tuvalu), as described in supplementary text 1.2 and 2.1. As in other comparison studies (Hawker et al 2022, Gesch 2023, 2024), we find that FABDEM is the most capable global DEM product (out of those we assessed) at representing low-lying elevations, but a vertical bias present in underlying Copernicus DEM reduces model performance in South Tarawa, Kiribati and Majuro atoll, Marshall Islands meaning coastal flood hazard is likely underpredicted here (supplementary text 2.1).

Secondly, the skill of the flood hazard model was assessed by calculating the Critical Success Index (CSI) against model benchmark data in two small island test cases: Ba catchment in Viti Levu, Fiji (a ∼22 km long river reach, 70 km² total area) (Archer et al 2018), and the entire island of Puerto Rico (∼9100 km²). Benchmark data serves as a baseline in which to compare model flood hazard performance, usually produced using high quality models or observations of a historical flood event. Suitable validation data to benchmark against in SIDS is currently extremely limited, meaning the only sources of appropriate validation data available for comparison were two model-based sources. Wing et al (2024) also note that availability of suitable validation data remains a key limitation in validating global scale flood models in data-poor areas such as SIDS. The CSI accounts for overprediction and underprediction between the modelled and benchmark data (0 = no match, and 1 = exact match) (Stephens et al 2014). The CSI score in the Fiji test case shows good agreement with the LIDAR model benchmark (0.69) (Archer et al 2018) (supplementary text 2.2). The CSI in the Puerto Rico test case is 0.47, which is expected considering the benchmark data and modelled data are not exactly like-for-like. This is because the benchmark data (FEMA National Flood Hazard Layer) estimates flood hazard using a 1D hydrodynamic model, which is less likely to adequately simulate flow on topographically complex floodplains when compared to the 1D/2D model used here (supplementary text 1.3.1.2 and 2.2). These results are in line with CSI scores found in previous studies (Wing et al 2017, Bates et al 2021, 2024). Hence, we find that the model has sufficient skill in SIDS to proceed with the analysis. More detail on this is available in supplementary text 1.3 and 2.2.

Present day boundary conditions for the coastal, fluvial, and pluvial models are simulated globally based on regionalised extreme value frequency distributions for each hazard, as described in supplementary text 1.4. General circulation models (GCMs) are coupled to global hydrological models and global sea level models, which are used to determine the percentage change factor in projected sea level, river flows, and precipitation compared to the present day. The climate change factor approach determines the percentage change in boundary conditions between the present day and future for a given global warming level. These percentage changes inform the change factor in the flood frequency curves for each flood hazard, where present day flood hazard maps are then interpolated to the appropriate return period for the change in frequency of this magnitude under the SSP scenario at this location (Wing et al 2024). A change factor approach was taken here, instead of directly simulating these changes through the hydrodynamic model, because it would be computationally intractable to do this at the SIDS scale for multiple climate scenarios at 30 m model resolution. The change factor approach used in this study has been applied in the UK (Bates et al 2023) and US (Bates et al 2021) and is currently the best-practice approach for estimating future flood risk in the third UK climate change risk assessment (Sayers et al 2020), which demonstrates the suitability of the change factor approach for estimating changes in flooding under climate change in this study.

Coastal flood change factors are calculated using changes in IPCC AR6 relative sea level projections at 2100 (Fox-Kemper et al 2023). Currently, the scientific consensus suggests changes in mean sea level are the dominant influence on the change in return period of coastal flood hazard under climate change, (Taherkhani et al 2020, Fox-Kemper et al 2023), including in Pacific and Indian Ocean SIDS (Jevrejeva et al 2023). Moreover, changes in storm surge and waves under climate change are uncertain (Yang et al 2020, Ewans and Jonathan 2023). Compounded by the lack of suitable multi-ensemble high resolution data on changes to storm surge and tides, this study only considers changes to future coastal flood hazard from absolute changes in mean sea level associated with sea level rise. This is a key limitation of these results which could lead to an underestimate of coastal flood hazard.

Fluvial change factors were computed using changes in median annual maximum discharge from the ISIMIP2b climate and global hydrological model ensemble (Frieler et al 2017), calculated relative to the historical baseline (1962–1992). Two of the twelve global hydrological models in the ISIMIP2B ensemble were selected to derive change factors in the global flood model (H08 and LPJmL). These were selected as these were the only ensemble members to include representative concentration pathway 8.5, use a consistent river network, and have the appropriate licensing (Wing et al 2024).

Change factors for pluvial flooding are calculated using changes in median annual maximum 1 day precipitation, relative to the same historical baseline as used for fluvial. Four of the six GCMs from the PRIMAVERA (2024) project derived from the coupled model intercomparison project 6 high resolution model intercomparison project (Haarsma et al 2016) were used: CMCC-CM2-VHR4, CNRM-CM6-1-HR, EC-Earth3P-HR, and HadGEM3-GC31-HH. These were selected because they had the best capability of representing tropical cyclones and had simulations of future climate change available (Wing et al 2024).

The WorldPop top-down, constrained population dataset with population totals adjusted to 2020 UN estimates was used to represent the population in each SIDS as the estimated number of people per 90 m grid cell (Bondarenko et al 2020). Population exposure estimates per SIDS are presented as a percentage of the total population to facilitate comparisons across all SIDS. Future population change is not considered in this study, as there are very few published datasets on future population exposure in SIDS, and none which have an adequate resolution (<100 m) for application consistently across all small islands. This has long been a limitation with the applications of global datasets in SIDS (Ranasinghe et al 2021). Moreover, 22 out of 57 SIDS do not have published information on SSP scenarios used to drive changes in population projections (Riahi et al 2017), meaning the application of future population would not be possible in many SIDS. Hence, our analysis does not account for changes in population dynamics over time meaning these estimates should be used as a guide to understand which SIDS are most exposed to flood hazard in the present day, and in the future only when considering changes in the drivers of the hazard associated with climate change. Nonetheless, changes in population are likely to be a large driver of future flood risk (Kam et al 2021) and therefore should be addressed in future studies when appropriate data are available. More information is available in supplementary text 1.5.

Vulnerability and risk indices are frequently used to identify the most vulnerable locations to climate-related hazards such as flooding, based on a set of indicators which vary depending on the index and its primary purpose (Eckstein et al 2021). Funding for adaptation measures is often granted based on a set of criteria such as these, including vulnerability to the impacts of climate change (Thomas and Theokritoff 2021, Tiedemann et al 2021). Both indicators of risk and vulnerability are used in this analysis, as well indices with different focuses such as development or climate adaptation. We calculate the Pearson's correlation coefficient between six metrics of vulnerability and risk with the projected change in population exposure to flooding under SSP5-8.5 for all available SIDS. This is important to understand because in the absence of detailed estimates of risk these generalized indicators are often used to represent climate change risk in data-sparse regions such as SIDS.

We hypothesise that despite the limitations, the highly granular estimates of flood hazard produced in this study—which have been extensively validated globally and in small islands—are more credible than the generalized indirect measures of exposure used in these indicators (see section 2.2). The metrics used are: gross domestic product (per capita), the Economic and Environmental Vulnerability Index and Human Asset Index (UN Department of Economic and Social Affairs 2024), the SIDS-focused multidimensional vulnerability index (United Nations 2023); and two global climate change risk indices developed to inform adaptation—ND-GAIN (Notre Dame Global Adaptation Initiative 2023) and INFORM climate change risk index (European Commission Joint Research Center 2024). The six vulnerability indicators were selected based on the availability of scores for the majority of SIDS. Each index only includes a subset of all SIDS due to data availability or due to the primary focus of the indicator. This means it is not possible to make a direct comparison with all—or even the same subgroup—of SIDS across each vulnerability/risk metric. Supplementary table 1 and text 1.6 summarises the six indicators used in this study, but below the metrics are described in more detail.

Figure 2 shows the flood hazard and population exposure for all SIDS for the three climate scenarios (SSP1-2.6, SSP2-4.5 and SSP5-8.5) compared to present day. We estimate that 19.5% of the population of SIDS are exposed to the 100 year return period flood hazard in the present day, or ∼8.5 million people—of which 81% is due to inland flooding (supplementary text 2.3; supplementary figure 1). Previous studies, which only focus on coastal flooding, are therefore likely to have significantly underestimated flood hazard in SIDS. 45 SIDS (out of 57) have greater than 10% of their population exposed to the 100 year flood hazard. Population exposure estimates are above 40% of the total population in six SIDS, and above 60% in three. These numbers are plausible in the context of historical flood events (Hoeke et al 2013, Cerrai et al 2020, 2021). See supplementary text 2.4 and supplementary figure 2 for other return periods. This level of exposure is disproportionately high compared to other states. The small size of SIDS means that extreme flood events affect a considerable proportion of the population at once, leading to higher damages (13% of the Caribbean's GDP, compared to 1% in larger states (Barca et al 2019)). For example, studies have suggested population exposure to the 100 year flood hazard is approximately 13% in the United States (Wing et al 2018) and 9% in the United Kingdom (UK Health Security Agency 2023), and damages are significantly lower (∼1%: Wilkinson et al 2021). Although those estimates account for flood protection, this is a reasonable comparison considering coverage of flood defenses is more widespread in these locations compared to most SIDS.

Zoom In Zoom Out Reset image size

Coastal flooding is the main driver of present-day population exposure in the top five most exposed SIDS, despite only contributing 19% to the total population flood exposure for inland and coastal flooding for all SIDS (supplementary figure 1). This suggests that in some locations, populations are much more densely located in areas of coastal flood hazard, leading to higher population exposure. For example, 100%, 76% and 55% of the population in the Bahamas, Suriname and Guyana are in the low elevation coastal zone (<10 m elevation) (Mycoo 2018).

This study finds that all SIDS except one (Dominica) experience both inland and coastal flooding to varying degrees. The proportion of population exposure due to coastal flooding ranges from 0%–92% depending on the SIDS for the 100 year return period flood hazard (Dominica only experiences inland flooding according to the 10 cm depth threshold considered—See supplementary text 2.3 and supplementary figure 1). This demonstrates the importance of considering both inland and coastal flood hazard drivers in SIDS.

We find that there is only a weak relationship between flood hazard (percentage of total land area inundated) and exposure (percentage of total population exposed to inundation), using both Kendall's rank correlation coefficient (τ = 0.252, p = 0.0057) and ordinary least squares regression (β = 0.351, p = 0.007). The right-hand side of figure 2 represents the change in rank when SIDS are ordered by percentage flooded area and population exposure. Here, we first rank the SIDS by highest to lowest percentage flooded area and population exposure, with the highest ranked 1, the second highest ranked 2 and so on. We then calculate the change in rank between the two, resulting in a rank change where a positive value (in orange) indicates a higher population exposure rank than flooded area rank, and vice versa for negative values (in blue). Small rank change numbers therefore suggest a close relationship between the two, whilst a large rank change (of either sign) suggests the reverse. With the exceptions of the Bahamas, Guyana, and Suriname, comparatively the most exposed countries are not the highest ranked in terms of hazard (as indicated by large negative numbers in blue on the right of figure 2). Although three of the five SIDS with the highest flooded area are the same as for population exposure: Bahamas (44%), Suriname (20%), Guyana (20%), the top five also includes Turks and Caicos (28%) and Cuba (20%) which rank 24 and 31 places lower for population exposure than for flood hazard. Patterns of population settlement in SIDS have been dominated by colonial and post-colonial drivers, resulting in densely populated settlements along the coastline (Cashman and Nagdee 2017, Beuermann and Schwartz 2018, Nunn and Kumar 2018). Thus, it is crucial to consider more than just the current and future flood hazard, but more critically where populations are located and how they interact with this hazard at a highly localized scale.

Figure 3 shows the percentage change in flood hazard and population exposure for all SIDS for the three climate scenarios compared to present day. We estimate that 21% of the population of SIDS would be exposed to the 100 year return period flood hazard at SSP1-2.6 compared to 19.5% under present day (+650 000 people). This would increase further to 22% (+1 025 000 people) at SSP2-4.5 and 23% at SSP5-8.5 (+1 500 000 people)—see supplementary tables 2–6. Although the absolute increases estimated here could be considered small compared to other global studies (Alfieri et al 2017, Dottori et al 2018), the percentage change in population exposure in SIDS is disproportionately large relative to their small population size and coping capacity (−4.5%–44%: SSP1-2.6) (Thomas et al 2019).

Zoom In Zoom Out Reset image size

The most exposed SIDS under the lowest emissions climate scenario (SSP1-2.6) experience the largest increases under the more extreme climate scenarios, as shown in figure 3. In these locations, mitigative action towards the lowest emission pathway is more likely to lead to a more significant difference in population exposure to flooding than in places where the difference in exposure change between SSPs is smaller. Conversely, many SIDS will experience a broadly similar exposure level regardless of the climate scenario. 20 SIDS are projected to experience a percentage change in population exposed between 5%–10%, and 20 SIDS between 10%–20%, between the lowest and highest climate scenario (SSP1-2.6 vs SSP5-8.5), meaning additional exposure to flooding is potentially 'locked in' in these locations.

The change in rank between SSP1-2.6 and SSP5-8.5 for percentage flood hazard and population exposure shown in figure 3 demonstrates that the most exposed SIDS remain so regardless of the climate scenario, as expressed using both Kendall's rank correlation coefficient (τ = 0.937, p < 0.001) and ordinary least squares regression (β = 0.978, p < 0.001). The rank change is calculated as the difference in rank when SIDS are ranked from 1 (most exposed) to 57 (least exposed) at SSP1-2.6 compared to at SSP5-8.5. A negative rank change in figure 3 indicates that the SSP5-8.5 rank is lower than at SSP1-2.6, and vice versa. The rank of the most exposed SIDS is robust to the SSP choice (supplementary text 2.5), as the top 10 most exposed SIDS at SSP1-2.6 remain in the same order and rank under SSP5-8.5 (supplementary figure 3). The top 10 SIDS with the largest population exposure change between SSP1-2.6 and SSP5-8.5 shift rank less than 4 places (figure 3). This is an important finding for the purposes of implementing proactive climate adaptation measures in SIDS (IPCC 2021, Mycoo et al 2022), because it implies that the SIDS most in need from adaptation to flooding will remain consistent regardless of the climate scenario and associated uncertainty.

Only three SIDS (Curacao, Grenada and Trinidad and Tobago) are estimated to see a reduction in flood hazard and population exposure under any of the three climate scenarios compared to present day (see figure 3). In these locations, the change factors derived from the GCM data used to drive estimated changes in inland flood hazard project a drying trend. This means that rainfall and river flow are lower in the SSP scenarios than in the present day, resulting in less inland flood hazard and population exposure in these locations. However, there is low agreement between the GCMs here meaning we do not make a robust prediction of the direction of change (supplementary figure 3).

None of the six measures of vulnerability or risk considered in this study can adequately identify the SIDS with the highest estimated percentage change in population exposure under the SSP5-8.5 climate scenario. Figure 4 shows the Pearson's correlation coefficient between each generalized vulnerability measure and our highly granular and physically robust estimates of the change in flood risk ranges between −0.09 and 0.33. Five out of six indicators have p-values that are not statistically significant at p = 0.05, and only the Economic and Environmental Vulnerability Index has a weak but statistically significant positive correlation (r = 0.33, p < 0.05). These indicators also fail to identify the SIDS with the highest present day population exposure (supplementary figure 4). The indirect measures of exposure used as a component of vulnerability in these indicators e.g. percentage of the population in low elevation coastal zones—are evidently too generic to represent highly localized hazards such as flooding, which requires the numerical simulation of flow dynamics and its interaction with the land surface at spatial scales of <10 m (Bates 2022). The sub-indicators used to represent flood hazard and population exposure in the ND-GAIN and inform climate change risk index are perhaps the most direct measures of hazard and exposure amongst the metrics included in this study, as changes in river and coastal flood hazard are taken from the Aqueduct global flood risk maps (World Resources Institute 2024) (See supplementary table 1). However, there are several key limitations of this data. Firstly, these data only include catchments >10 000 km² which is larger than the size of many SIDS, meaning most SIDS are not represented (Winsemius et al 2016). Secondly, pluvial flooding is a key driver of flooding from rainfall in SIDS (Abebe et al 2019), meaning population exposure using these maps are likely an underestimate as this flood driver is not represented. Thirdly, flood hazard is simulated using hydrological data from 1960–1999, but many SIDS have an incomplete river gauge record (Yeo et al 2007), meaning the simulated flood hazard may not adequately capture the flood magnitude or frequency in many SIDS. Hence, the application of any of these vulnerability indicators for the purposes of assessing those SIDS most in need of support to adapt to the impacts of flooding under climate change would be inadequate if used in isolation, demonstrating the importance of the inclusion of direct measures of flood hazard and exposure.

Zoom In Zoom Out Reset image size