Current and Future Tropical Cyclone Wind Risk in the Small Island Developing States

Abstract

Tropical cyclones (TCs) are amongst the costliest and deadliest natural hazards and can cause widespread havoc in tropical coastal areas. Small Island Developing States (SIDS) are particularly vulnerable to TCs, as they generally have limited financial resources to overcome past impacts and mitigate future risk. However, risk assessments for SIDS are scarce due to limited meteorological, exposure, and vulnerability data. In this study, we combine recent research advances in these three disciplines to estimate TC wind risk under past (1980–2017) and near-future (2015–2050) climate conditions. Our results show that TC risk strongly differs per region, with 91% of all risk constituted in the North Atlantic. The highest risk estimates are found for the Dominican Republic and Puerto Rico, with present-climate expected annual damages (EAD) of 1.51 billion and 1.25 billion USD, respectively. This study provides valuable insights in TC risk and its spatial distribution, and can serve as input for future studies on TC risk mitigation in the SIDS.

6.1 Introduction

Tropical cyclones (TCs) are among the deadliest and costliest natural hazards, wreaking widespread havoc when they make landfall. Small Island Developing States (SIDS) in particular have been substantially affected by TCs. The 2017 hurricanes Irma and Maria, for example, caused catastrophic damage across multiple SIDS in the Caribbean, including Antigua and Barbuda, Saint Martin, the British and U.S. Virgin Islands, Puerto Rico, and Dominica (Cangialosi et al. 2018; Pasch et al. 2019). The latter island suffered damage totals exceeding 930 million USD, amounting to approximately 226% of the country’s total Gross Domestic Product (GDP) (Government of the Commonwealth of Dominica 2017). Over the past 5 years, Fiji was hit by four major TCs, each causing major damage across the country. Almost 15% of Fiji’s total population was displaced due to TC Winston (2016), and damages were estimated at 1.38 billion USD, which is 31% of the country’s GDP (World Bank 2016).

The aforementioned historical events all demonstrate that SIDS are particularly vulnerable to, and can be disproportionately affected by, TCs. Over the last few years, authorities have been increasingly aware of this disproportionality (UNESCO 2019) and that these impacts are most likely to deteriorate even further towards the future (UNFCCC 2005). Under future climate change, TC intensity, storm surge inundation, and rainfall are projected to increase globally, thereby enhancing the associated risks (Knutson et al. 2020). SIDS, unfortunately, often face limited institutional capacity, scarce financial resources, and high vulnerability. They therefore struggle to overcome past impacts and mitigate future TC risk (United Nations 2021). This is driven by the fact that many SIDS are characterized by their dependency on a single economic sector providing employment and economic growth, such as tourism (Pathak et al. 2021). When such a sector is (substantially) disrupted after a TC event, it can take years for this sector, and thereby (the majority of) the economy of the SIDS, to overcome such impacts. It is therefore crucial to quantify their respective TC risk, to support the design of appropriate risk mitigation strategies, and thereby enhancing the resilience of these SIDS.

However, despite their exposure and vulnerability to TCs, (global) risk assessments focused on SIDS are scarce, and instead are more often performed on a regional scale or for particular TC events (e.g., Webb 2020; Stephenson and Jones 2017; Commonwealth of Australia 2013). A uniform study, assessing the economic risk for all SIDS prone to TCs, is currently still lacking. This gap can be attributed to two complicating factors: (1) scarce observational records over longer temporal scales on both hazard and impact/damage totals, and (2) the relatively small size of most SIDS, making it difficult to identify and study them in meteorological datasets, as the spatial resolutions of such datasets are often lower than the entire size of an island. Recent scientific advances, however, overcome these limitations. Global-scale synthetic TC models, for instance, overcome the limited spatial and temporal hazard information imposed by historical TC data (Bloemendaal et al. 2020b; Emanuel et al. 2006; Lee et al. 2018). These models take information, commonly spanning a 30–40 year time scale, from either historical data (Bloemendaal et al. 2020b) or global climate models (Lee et al. 2018; Emanuel et al. 2006) and resample this information to an equivalent of a few 1000 years under the same climate conditions. The resulting dataset contains a wealth of information on all theoretically possible TC events for all TC-prone locations on earth, including information on low-probability, high-impact events that might not have occurred in the historical records. At the same time, increasing efforts to collect high-resolution information on assets and economic activity at a global scale also allow the analysis of previously considered data-scarce areas, particularly in developing countries such as the SIDS (Koks et al. 2019).

In this study, we aim to improve our understanding of the current TC-induced windstorm risk in the SIDS. To do so, we make use of publicly available models and datasets. We combine state-of-the-art TC hazard modelling for the current and future climate (Bloemendaal et al. 2020b) with high-resolution exposure information obtained from both OpenStreetMap and globally-consistent gridded exposure (Eberenz et al. 2020) and vulnerability information (Eberenz et al. 2021).

The remainder of this chapter is organized as follows. Section 6.2 briefly describes the SIDS that are studied, followed by Sect. 6.3 in which the methodology is explained in detail. Section 6.4 presents the results of our analysis and puts them in a broader perspective. Finally, Sect. 6.5 provides some concluding remarks and ways forward.

6.2 Small Island Developing States

The United Nations (UN) defines SIDS as “a distinct group of 38 UN Member States and 20 Non-UN Associate Members that face unique social, economic and environmental vulnerabilities” (United Nations 2021). Of these 58 SIDS, six are found within 5°N/S of the Equator, a region generally unfavorable for TC development due to a too weak Coriolis force. In addition, Bahrein, Cabo Verde, and Guinea-Bissau are located in regions generally not considered to be prone to TCs; as such, we exclude these SIDS as well. Lastly, we exclude French Polynesia from our analysis due to absence in the exposure dataset (see Sect. 6.3). This leaves 48 SIDS included in this study, which are geographically mapped in Fig. 6.1.

Fig. 6.1

Overview of the Small Island Developing States that are part of this study

6.3 Methods

This study estimates the current and future TC windstorm risk in the 48 SIDS presented in Sect. 6.2. To do so, we take a traditional catastrophe loss modeling approach (Kron 2005; Koks et al. 2019), which is commonly applied in catastrophe modeling by academia and the (re)insurance industry. More specifically, we define risk (in terms of USD) as a function of hazard – the probability of the TC event; exposure – the value of assets subject to the TC; and vulnerability – the capacity of a society to deal with the event (Koks et al. 2015b). In this study, both the exposure and vulnerability are kept constant throughout time. More specifically, only the hazard-component is changing under climate change. This means that TC risk results solely portray the effect of climate change on TC wind speeds, as socio-economic changes are not included.

6.3.1 Hazard

6.3.1.1 Baseline Climate Conditions

At the basis of the TC hazard modelling component lies the Synthetic Tropical cyclOne geneRation Model (STORM) (Bloemendaal et al. 2020b). STORM is a fully statistical model taking information on TC track, characteristics, and environmental variables (mean sea-level pressure and sea-surface temperature; SST) as input variables. These variables are resampled in STORM to extend the temporal scale of the input dataset to 10,000 years of TC activity under the same climate conditions. The resulting STORM synthetic TC dataset contains data on the position of the eye of the TC (longitude/latitude), 10-minute 10-m average maximum sustained wind speed (in m/s), minimum pressure (in hPa), and the radius to maximum winds (in km). The STORM dataset as presented in Bloemendaal et al. (2020b) resembles the climate conditions over 1980–2017, and was generated using historical TC statistics from the International Best-Track Archive for Climate Stewardship (IBTrACS; Knapp et al. 2010). We direct readers to Bloemendaal et al. (2020b) for detailed information on STORM and the STORM baseline dataset.

Please note that STORM simulates TCs with wind speeds exceeding 18 m/s (Tropical Storm-classification on the Saffir-Simpson Hurricane Wind Scale; Simpson and Saffir 1974). SIDS can also be affected by Tropical Depressions and post-TCs; however, their accompanying wind speeds are typically too low to cause (substantial) damage to housing and infrastructure (see also Fig. 6.2 and Sect. 6.3.3).

Fig. 6.2

Vulnerability curves used in this study. The left-panel presents the four vulnerability curves (one for each basin) which are calibrated using the Total Damage Ratio (TDR). The right-panel presents the four vulnerability curves (one for each basin) which are calibrated using the Root-mean-squared fraction (RMSF)

In Bloemendaal et al. (2020a), these synthetic tracks were translated to a 2D-wind field by applying the refined 2D-parametric wind field model of Holland (1980) (Lin and Chavas 2012). In this parametric approach, asymmetry in the wind field arises from the background flow. Please note that in extratropical regions, however, asymmetry may also arise through enhanced wind shear from large-scale background flows or nearby troughs, which is not modeled here (Bloemendaal et al. 2020a; Ritchie and Elsberry 2001). Wind speed RPs were empirically derived from this set of events and at every 10-km grid cell in a basin. Here, we use this STORM RP dataset as the baseline climate RP dataset, with RPs up to 1000-year. Please refer to Bloemendaal et al. (2020a) for validation and access to the RP datasets.

6.3.1.2 Future Climate Conditions

To create a future climate synthetic TC dataset, we use TC information from four global climate models (GCMs): CMCC-CM2-VHR4 (Scoccimarro et al. 2017), CNRM-CM6-1-HR (Voldoire 2019), EC-Earth3P-HR (EC-Earth Consortium 2018), HadGEM3-GC31-HM (Roberts 2017). These GCMs are part of the HighResMIP multi-model ensemble (Roberts et al. 2020), a model intercomparison study focused on simulating climate runs at higher spatial resolution to better capture small-scale processes such as tropical cyclones. HighResMIP follows the Coupled Model Intercomparison Project Phase 6 (CMIP6) modelling protocol (Eyring et al. 2016). We use the high-resolution coupled ocean-atmosphere GCM runs for the periods 1979–2014 (resembling the GCM baseline climate conditions) and for 2015–2050 (the GCM future climate conditions). These future climate conditions are based on the high-emission Representative Concentration Pathway 8.5 (RCP8.5) with the Shared Socioeconomic Pathway 5 (SSP5) scenario (SSP585) (O’Neill et al. 2016; Van Vuuren et al. 2011). The SSP585 scenario assumes no climate change mitigation strategies are implemented, and average global surface temperatures will increase by 3.3 °C–5.7 °C in 2100 compared to the late nineteenth century (IPCC 2021). While the plausibility of this scenario is under debate given the recent developments in the energy sector, we cannot rule out this scenario and its corresponding climate features altogether. As such, the future-climate results presented in this study should be perceived as a high-end view of future risk.

TCs, and in particular TC intensity, are generally poorly captured in GCMs (Roberts et al. 2020), mostly due to the relatively coarse spatial resolution (coarser than 0.25° × 0.25°), which does not adequately resolve smaller-scale TC features such as strong wind and pressure gradients (Murakami and Sugi 2010). As STORM resamples the TC statistics found in the input dataset to an equivalent of 10,000 years, this means that this poor TC representation will be propagated through STORM, and TC intensity will therefore also be underestimated in the STORM+GCM datasets. To overcome this, we use a novel methodology based on the delta approach. Here, we provide a summary of this method; the details of this method and an extensive data analysis can be found in Bloemendaal et al. (2022). In summary, we extract information on changes in the STORM input variables from the baseline and future-climate conditions of the four aforementioned GCMs. This way, we omit the first-order model bias of GCMs, such as an underestimation in TC intensity or an erroneous genesis frequency). Next, we project these changes onto the observed TC statistics (TC genesis frequency and location, track, and intensity) from IBTrACS, which was used for the generation of the baseline STORM dataset. This way, we create a future-climate version of these observed TC statistics. These future-climate statistics then form the input for STORM, hereby creating four synthetic TC datasets for each of the four GCMs.

Lastly, we point out that STORM is run on the basin-scale, meaning that TCs are cut off once they reach the basin boundaries (see Fig. 1 in Bloemendaal et al. 2020b for an overview of the basin boundaries). While almost all of the SIDS used in this study lie well within one basin boundary, Timor-Leste is located just east of the South Indian/South Pacific border in STORM. This means that in STORM, Timor-Leste is solely affected by TCs originating in the South Indian basin, while in reality, the country is also affected by TCs forming in the South Pacific. This can potentially affect the probability and intensity of TC wind hazard estimates for this country.

6.3.2 Exposure

6.3.2.1 OpenStreetMap

To gain insights into the type of assets that are exposed to baseline and future TCs, we make use of publicly available information provided by OpenStreetMap (OSM; www.openstreetmap.org). OSM is a global geospatial database covering a wide range of objects that can be spatially represented on a map. This ranges from administrative boundaries and land uses to roads and buildings. These elements are represented as polygons (e.g., land uses or buildings), linestrings (e.g., roads and electricity lines), and points (e.g. telecommunication towers). Supplementary Table 1 provides an overview of the number of assets per SIDS that are extracted from OSM. All assets were extracted from OSM on July 19, 2021. The geolocations (latitudinal/longitudinal) will be overlaid with the STORM RP maps to assess their exposure to TC wind speeds. As no future projections for the built-up area and infrastructure development exist at this point, we use the same database for both the current and future-climate analysis (i.e., exposure and vulnerability are kept constant).

It should be noted that objects in OSM are ‘tagged’ (named) and georeferenced by its users using satellite imagery. Due to its dependency on the activity of its users, completeness substantially varies between regions. While very few global numbers exist on actual completeness of buildings (or all assets combined), Barrington-Leigh and Millard-Ball (2017) estimated that in 2016, over 80% of the world’s roads were included in OSM. For SIDS in particular, Goldblatt et al. (2020) assessed the completeness of OSM for Haiti, Dominica, and St. Lucia. They found a clear relationship between the population density and the completeness of OSM. More specifically, urban areas tend to be better mapped compared to rural areas. Here we make the implicit assumption that the relative amount of mapped assets between high and low-density built-up areas is similar and that the OSM represents the spatial distribution of assets across the SIDS correctly. To circumvent gaps in the absolute number of mapped assets in OSM, we will present the OSM-derived results in terms of percentage of assets affected by Category 3 TC wind speeds on the Saffir-Simpson Hurricane Wind Scale (Simpson and Saffir 1974). These wind speeds correspond to 10-minute wind speeds equal or greater than 43.4 m/s. We use this Category 3 threshold, as this sets the lower wind speed threshold from which major to catastrophic damage to well-built structures occurs (NOAA 2021).

Due to the lack of detailed information about the characteristics of each asset (e.g., construction materials of a building, the quality of a road, the design standards of electricity assets) and the limited availability of vulnerability curves for these individual assets, we only use the asset data extracted from OSM to better understand exposure (and how exposure may change towards the future as a result of climate change). We do not try to estimate asset-level damage and risk, as dealing with the uncertainty and unavailability of the required information to do so is out-of-scope for this study.

6.3.2.2 LitPop

As previously mentioned, the OSM dataset contains information on the location of assets but does not provide information on asset values. The “lit population” (LitPop) database does contain asset value information, aggregated at 1 km × 1 km resolution (Eberenz et al. 2020). Due to this 1 km × 1 km aggregation, we cannot use LitPop to assess an individual asset’s exposure. Hence, we do not aim to connect the asset values from LitPop with the OpenStreetMap asset information.

The LitPop database is a globally consistent high-resolution exposure dataset describing the combined value of physical assets (e.g., buildings, infrastructure) within each 1 km × 1 km grid cell. The data are produced using a combination of gridded night light intensity and population data to disaggregate physical asset stock values (capital) proportionally within each country (Eberenz et al. 2020). The LitPop dataset is available for 224 countries; all SIDS but one (French Polynesia) incorporated in this analysis are included in the LitPop database. The LitPop values (in USD) used in this study are relative to the base year 2014. We direct readers to Eberenz et al. (2020) for more information on the LitPop database.

6.3.3 Vulnerability

To estimate the potential damage for each SIDS, we use the regionally calibrated vulnerability curves developed by Eberenz et al. (2021). These vulnerability curves were derived for nine regions; for all SIDS, we use the vulnerability curve corresponding to their respective region. Following Eberenz et al. (2021), we apply a sigmoidal vulnerability function satisfying two constraints: (i) a minimum threshold for the occurrence of damage with an upper bound of 100% direct damage (Emanuel 2011); (ii) a high power-law function for the slope, describing an increase in damage with increasing wind speeds (Pielke 2007). Figure 6.2 presents the four regionally calibrated curves for the four basins that are included in this study.

To account for the uncertainty in the estimation of these vulnerability curves, we use two vulnerability curves per basin: (i) the curve calibrated through using the root-mean-squared fraction (RMSF) and (ii) the curve calibrated through using the total damage ratio (TDR). The RMSF is defined as the relative deviation between modeled and reported damage for all matched events in a region, whereas the TDR is defined as the sum of simulated damages divided by the sum of normalized reported damages (Eberenz et al. 2021). Following the recommendation by Eberenz et al. (2021), the risk estimates reported in this study will be based on the TDR calibration. The risk estimates using the RMSF calibrated curves are provided in Supplementary Materials.

We point out that Eberenz et al. (2021) calibrated their vulnerability curves using 1-minute average sustained wind speeds. As the STORM wind speeds are given as 10-minute average sustained wind speeds, we convert the STORM wind speeds to their 1-minute equivalent using a wind speed conversion factor of 0.88 (Harper et al. 2008). Furthermore, we note that the Eberenz et al. (2021) curves are calibrated on total damages: such TC damages are composed of damages caused by wind speeds, storm surge, and precipitation. As a result, aside from wind speed-inflicted damages, the damages calculated in the current study also implicitly contain damages caused by storm surge and precipitation. For the future climate study, however, we use the same Eberenz et al. (2021) vulnerability curves, as equivalent future vulnerability curves have not been developed yet. This means that the future damages listed in this study do not account for changes in the construction quality of buildings and infrastructure, or whether adaptation measures have been implemented. In addition, as the curves constitute a relationship between damage and TC hazards, the latter represented as wind speed, any future-climate changes in sea-level rise (Shepard et al. 2012) or changes in precipitation (Knutson et al. 2020) are not reflected in the damage estimates.

6.3.4 Risk

We express TC risk in terms of Expected Annual Damages (EAD, in USD). This EAD is calculated as the integrated value of expected damages over all RPs. To do so, a trapezoid approach is taken to assess the EAD, see Eq. 6.1:

$$ EAD=\sum \limits_i\frac{\left({P}_i-{P}_{i-1}\right)\bullet \left({D}_{P_{i+1}}-{D}_{P_i}\right)}{2} $$

$$ {P}_i=\left\{\frac{1}{p}|p\in \left[10,20,30,\dots, 100,200,300,\dots, 1000\right]\right\} $$

(6.1)

Here, P_i is the exceedance probability of a TC wind speed (given as the inverse of the RP), and ${D}_{P_i}$ the damage related to that event (in USD) (Koks et al. 2015a). As indicated in Eq. 6.1, the RPs are given at 10-year interval levels between the 10-year and 100-year RP level, and at 100-year interval levels between the 100-year and 1000-year RP. This is done to comply with the RP levels as used in the STORM RP datasets.

6.4 Results and Discussion

6.4.1 Future Trends in Tropical Cyclone Hazard

From the STORM baseline and future climate RP datasets, we extract the wind speed RPs (up to 1000-year) in the capital cities of each of the 48 SIDS (Fig. 6.3). First of all, we note that the changes in RP curves between the baseline and future climate conditions differ substantially per SIDS and basin. In the North Atlantic (the blue curves in Fig. 6.3), our results show that for the majority of SIDS, a minor change in RPs up until the 1000-year RP is projected under future climate change compared to the baseline scenario. In this basin, the average change in wind speed at the 1000-year RP amounts to 1.3 ± 1.6 m/s across the SIDS. This relatively minor difference is primarily driven by the number of intense (Category 3 or higher) TCs entering the Gulf of Mexico and Caribbean region staying approximately the same in the future-climate STORM synthetic datasets, despite a decrease in overall TC frequency (Bloemendaal et al., 2022). These findings are consistent with other literature (e.g., Bruyère et al. 2017).

Fig. 6.3

Empirically derived return periods of maximum 10-minute 10-m average wind speeds in the capital city for each of the 48 SIDS. The solid line represents the STORM baseline climate conditions (corresponding to the average climate conditions over 1980–2017). The shaded areas represent the range of return periods of the four STORM future climate datasets (average climate conditions over 2015–2050). Colors indicate the respective basin: North Atlantic – blue, South Indian – yellow, South Pacific – red, and Western Pacific – green

Contrary to the North Atlantic, for the other basins, we project higher wind speeds at the 1000-year RP for the 2015–2050 period compared to the baseline. The largest difference between the baseline and future-climate curves is found for the South Pacific (red curves in Fig. 6.3), amounting to an average of 7.0 ± 2.0 m/s across the SIDS. Figure 6.3 illustrates the projected substantial increase in TC intensity, visualized through the future-climate RP curves (shaded areas in Fig. 6.3) all being substantially higher compared to their baseline counterparts. This change is likely caused by the combination of two factors. First, the STORM future-climate synthetic datasets show a slight increase in TC frequency in the South Pacific compared to its baseline-climate counterpart (Bloemendaal et al., 2022). Second, the SST fields that serve as input for STORM (see Sect. 6.3), play a direct role in the STORM TC intensity modelling. In STORM, the SST fields are used to calculate the Maximum Potential Intensity (MPI; Bister and Emanuel 1998). This MPI provides a theoretical upper bound of TC intensity in a given location. As the future-climate (2015–2050) STORM simulations make use of increasing SSTs present in the four input GCMs (see Sect. 6.3), this means modeled future climate conditions are more favorable for (further) TC intensification compared to the STORM baseline climate conditions. The combination of these two components (frequency and intensity) thus leads to increased probabilities of intense TC occurrences, and hence of such systems to affect the SIDS in this basin.

A few SIDS are noteworthy to discuss individually. Guyana and Suriname, for instance, are generally affected by lower TC wind speeds compared to the other North Atlantic SIDS, while their RP curves also show larger uncertainties towards the future. These two countries are located near the southern North Atlantic basin boundary, relatively close to the Equator compared to the other SIDS in the North Atlantic. As TCs in STORM are modeled to deflect away from the Equator, TC peak intensities are generally being reached northward of Guyana and Suriname (this also explains the difference in wind speeds amongst the different RP curves in the North Atlantic). However, it should be noted that a difference exists in TC genesis frequency between 5°-10°N between the STORM baseline and future-climate datasets. This region is of particular interest here, as Guyana and Suriname are generally hit by TCs forming in this latitudinal region. In the baseline STORM dataset, the probability of TC formation in this region amounts to 0.93 ± 1.42% of all TC formations, whereas this increases to 4.6 ± 5.0% of all TC formations in the future-climate STORM datasets (aggregated over all separate GCM runs). Note that in STORM, the probability of the TC genesis in a given location is weighted by the monthly TC genesis locations as extracted from the input dataset (either historical data or GCM data). This weighted number can never be smaller than zero. However, the large standard deviation hints at a large signal-to-noise ratio, implying that we cannot label this difference as a strong signal instigated by the effects of climate change, which is also reflected by the larger uncertainty in future-climate RPs shown in the Guyana and Suriname subplots.

6.4.2 Future Trends in Exposure

Figure 6.4 presents the RP curves of the fraction of buildings affected by Category 3 (43.4 m/s) wind speeds per SIDS in the baseline and future climate analyses. A similar figure for fraction of infrastructure can be found in the Supplementary Materials (Supplementary Fig. 2). By expressing the number of buildings as a fraction of the total in a SIDS, we circumvent possible data scarcity in OSM, as well as enable homogenized comparison across the SIDS (see Sect. 6.3). We point out to the reader that changes in exposure are solely driven by changes in TC wind speeds under climate change, as no future projections for the OSM database currently exist. For most SIDS, the RP curve shows a strikingly steep increase, indicative of the RP at which the capital city (most often the location with the highest density of assets) is hit by Category 3 wind speeds (Fig. 6.4). However, we find substantial differences in the exact RP at which this happens. These differences are discussed in the following paragraphs.

Fig. 6.4

Similar to Fig. 6.3, but now shown are empirically derived return periods of the percentage of buildings affected by Category 3 (43.4 m/s) wind speeds or higher in each of the 48 SIDS. In both climate scenarios, we use the same building data from OpenStreetMap; changes in the return period curve are therefore entirely driven by changes in wind speed hazard

First of all, we observe a shift in RPs at which almost all buildings are affected under the 2015–2050 climate conditions compared to the 1980–2017 period. For all SIDS located in the North Atlantic (except Bermuda), the baseline climate RP curves lie within the range of future climate RP curves. This indicates a (very) minor change in the probabilities of a certain fraction of buildings being affected under climate change. This minor change, in turn, corresponds to a minor change in Category 3 wind speed RPs in the North Atlantic region (see Fig. 6.3 and Supplementary Fig. 1). On the contrary, in the other basins our results indicate a substantial shift towards higher probabilities of buildings being affected (lower RPs) under climate change. The most notable shifts in RP curves are found for the SIDS located in the South Pacific: all SIDS show a substantial increase in affected buildings under climate change. This is driven by a similar substantial increase in Category 3 wind speed probabilities (decrease in RP) in these regions (see Fig. 6.3 and Supplementary Fig. 1). Particularly SIDS located poleward of 20°S are projected to face a more-than-tenfold increase in Category 3 probability; this includes the Cook Islands, Tonga, and New Caledonia. For the latter island country, no buildings are affected at the 300-year RP, but this number is projected to increase to 99–100% in the future climate conditions. In the Comoros (in the South Indian), buildings are not affected by Category 3 wind speeds (with a RP below 1000-year) in the baseline climate (see also Fig. 6.3). Our results, however, project that under climate change, between 20% and 60% of all buildings will be affected at the 700-year RP. This range increases to 43–85% at the 1000-year RP.

In addition, we note the apparent difference in the shape of the baseline and future-climate (2015–2050) RP curves across the SIDS: SIDS such as Montserrat and Guam show a sudden, steep increase from 0% to 100%, while other SIDS like Guyana or the Solomon Islands show a more gradual increase. The explanation for this difference is two-fold: one hazard-related and one exposure-related. First, on the hazard side, the decay of TC wind speeds after landfall plays a prominent role. In STORM, TC decay occurs when a TC is over land for more than 9 hours (Bloemendaal et al. 2020b). This implies that for SIDS where a TC generally takes less than 9 hours to pass (e.g., the SIDS that are part of the Windward Islands, or other small islands), no wind decay is modelled. For mainland countries such as Belize, a modeled TC will typically be over land for more than 9 hours, and as such this TC will be decaying after landfall. As a result, Category 3 wind speeds in the inland part of this country will have a RP greater than 1000-year. In the case of Belize, where built-up areas are spatially distributed across the country, this means that not all assets will be exposed to Category 3 wind speeds and hence a maximum of 47% of the total number of assets will be affected by these wind speeds at the 1000-year RP in the baseline climate; In the future climate, this portion amounts to 32–40%. For Guyana and Suriname, however, most assets are found along the coastline, meaning that these have a higher probability of being affected by these wind speeds. This translates to 95% of all assets being exposed to Category 3 wind speeds at the 314-year and 857-year RP, respectively. Another illustrative example where this effect of inland decay comes into play is for Haiti and the Dominican Republic, which are both located on the island of Hispaniola. The majority of their assets (68% and 65%, respectively) are located on the southern side of the country (south of 19°N), which is the direction from which TCs typically approach Hispaniola. However, in Haiti, 95% of the buildings are affected at the 76-year RP in the baseline climate, while in the Dominican Republic, 83% of all buildings are exposed at this RP. This difference is driven by the fact that, in Haiti, the other large cities are also located along the coastlines (western and northern), while in the Dominican Republic, the second largest city of the country, Santiago de los Caballeros, is located further northward and inland. Compared to the coastal cities, the assets in Santiago de los Caballeros are less exposed to Category 3 force winds as the modeled TCs moving over the Dominican Republic will generally be decaying in STORM.

Aside from the hazard component, differences in the spatial distribution of built-up areas play an important role in clarifying the different characteristics of the RP curves. For SIDS where almost all assets are located at one part of the island (e.g., small island nations such as Saint Lucia, Montserrat, or St. Maarten), these assets will face an (approximately) equal probability of Category 3 wind speed exposure. As a result, the RP curve will show a sharp increase at the RP where this Category 3 threshold is met. Such devastating impacts with near-total destruction in small island nations was for instance seen in the aftermath of Hurricane Irma (2017)’s passage over Barbuda. Barbuda was directly hit by the Category 5 TC, and it was estimated that 95% of all structures were damaged or destroyed, leaving the island uninhabitable for the first time in 300 years (Cangialosi et al. 2018).

On the other hand, for SIDS where assets are spread over a larger area or multiple islands, we find clear spatial differences in Category 3 wind speed RP across the country. This means that not all assets share the same probability of being affected by such intense wind speeds. Examples of such SIDS include the Solomon Islands and the Bahamas. The latter archipelago was directly hit by Hurricane Dorian’s Category 5 wind speeds in 2019. The second largest city of the country, Freeport, was catastrophically impacted by the TC (Avila et al. 2020). Meanwhile, the capital city of Nassau, located 200 km southeast of Freeport, suffered little damage and served as shelter for those impacted by the TC (IDMC 2020).

6.4.3 Future Trends in Damage

By combining TC wind speed RP maps with asset values from LitPop (see Sect. 6.3), we estimate damages at RPs ranging from 10-year to 1000-year (see Fig. 6.5). We note that the data from LitPop was derived for the baseline climate conditions and does not contain future projections; as such, differences in damage RPs are entirely driven by changes in TC wind speeds under climate change.

Fig. 6.5

Similar to Figs. 6.3 and 6.4, but now shown are empirically derived return periods of total damages to assets (in USD, plotted on a logarithmic scale) in each of the 48 SIDS. Damages are calculated as the aggregated damage over all physical assets in the Litpop database, see Sect. 6.3

The graphs in Fig. 6.5 illustrate the combined effects of the wind speed RP curves (Fig. 6.3) and the applied vulnerability curve (Fig. 6.2; left panel). Unlike the other basins, the vulnerability curve for the Western Pacific does not show a steep increase towards higher wind speeds. This means that the maximum damage in this basin does not exceed 6% of the total GDP, regardless of a future increase in wind speeds. As a result, the highest total damage at the 1000-year RP in the Western Pacific is found for Guam, totaling to 60.8 million USD in the baseline scenario, and estimates for the 2015–2050 period ranging from 72.9 million to 1.01 billion USD. The other basins, however, do show a steep vulnerability curve, with maximum total damages of up to 80%, 83%, and 94% for the North Atlantic, South Indian, and South Pacific, respectively. In these three basins, highest damages at the 1000-year RP are found for the Dominican Republic (85.5 billion USD), Mauritius (23.2 billion USD), and Papua New Guinea (6.39 billion USD) in the baseline climate. The largest relative change in damage at the 1000-year RP is found for Tuvalu, amounting to approximately 419% compared to the baseline climate.

6.4.4 Future Trends in Risk

Based on our results, the EAD for the baseline situation, for all SIDS combined, is estimated at 6.2 billion USD, or 1.2% of their aggregated GDP. For the 2015–2050 time period, this combined EAD increases by 2.4% (−8.1–16.2%, depending on the global climate model) compared to the baseline situation. This increase, however, varies significantly between the different basins. While we find a decrease of −3.8% (−12.1–8.1%) for the North Atlantic basin, we find an increase of almost 153.6% (104.8–201.0%) for the South Pacific. For the South Indian and Western Pacific basins, the increases are approximately 58.0% (25.5–89.4%) and 64.5% (37.6–117.6%), respectively. As the North Atlantic basin constitutes 91.6% of the total EAD, it is clear that SIDS located in this basin are the main contributors to these changes.

Taking a closer look at the individual SIDS (Fig. 6.6), we find that the top 10 SIDS with highest EAD contains eight SIDS located in the North Atlantic, complying with the earlier finding that this basin constitutes most TC risk. Our results show that the Dominican Republic (1.47 billion USD), Puerto Rico (1.25 billion USD), and Martinique (763 million USD) experience the largest risk. For these three SIDS, these EADs slightly decrease under future climate change, with average values across the four STORM global climate model datasets amounting to 1.41 billion, 1.22 billion, and 691 million USD, respectively.

Fig. 6.6

Expected Annual Damage (EAD, in USD) for the baseline and future climate conditions using the TDR-calibration method. The error bar for the future situation represents the spread of EAD across the four separate STORM future-climate datasets (see Sect. 6.3)

We point out here that the asset value exposed to certain TC wind speeds is the key driver of EAD. This becomes visible when looking at the EAD for Guadeloupe, Dominica, and Martinique. Despite their RP curves in Figs. 6.3, 6.4 and 6.5 being approximately similar, there are substantial differences in the EAD estimates (Fig. 6.6). This difference in EAD is driven by the fact that Dominica has a total asset value of 895 million USD in LitPop, whereas this value amounts to 61 billion and 78 billion USD for Guadeloupe and Martinique, respectively.

Validation of risk estimates is an essential part of any risk assessment. As pointed out in the introduction, few studies were conducted to assess the TC risk to SIDS and were mostly executed on a case study basis. A direct validation of our results using these other studies, however, is difficult. This is because there are numerous components in the entire model chain (composed of the hazard, exposure, and vulnerability modeling) that can each impose a substantial influence on the EAD estimates, and accounting for each of them lies beyond the scope of this research. Instead, we compare our results against EADs derived in another, uniformly executed study for a selection of SIDS, and briefly list some reasons for differences in EAD estimates.

The Commonwealth of Australia (2013) presents a series of reports on TC risk for Pacific island nations. These reports also list wind speed estimates at the 100-year RP for the different island nations; in Bloemendaal et al. (2020a), these estimates were compared against their counterparts from the STORM baseline climate dataset. Results showed that for eight out of 12 locations, STORM wind speeds were within 5 m/s of those in the reports. In terms of EAD, we find that for the Cook Islands, Timor-Leste, Papua New Guinea, and Samoa, our estimates for the baseline climate are within 50% of their listed values. For Fiji and Niue, their EAD estimates are approximately a factor 20 higher than ours. This can mostly be attributed to the difference in total exposure (asset) value. For Niue, this value is a factor 20 higher in the Commonwealth of Australia case studies; for Fiji, this difference amounts to a factor 10 higher compared to our results. Furthermore, our damage estimates predominantly consist of wind damage to physical assets (buildings and infrastructure), whereas the case study reports also assess flood damage from precipitation and consider damages to agriculture.

6.4.5 Limitations and Directions for Future Research

In the previous sections, we have demonstrated and discussed the different hazard, exposure, and risk datasets for the SIDS. These datasets, however, impose some limitations, which we will briefly reflect upon here, as well as provide some directions for future research.

First, TC wind speeds after landfall are modeled using an empirical inland decay function (Kaplan and Demaria 1995; Bloemendaal et al. 2020b) and translated to a 2D wind field using the Holland parametric wind field model (Holland 1980; Lin and Chavas 2012; Bloemendaal et al. 2020a). The decay function was calibrated based on US landfalling events; as such, this function may perform less in other regions. Furthermore, this decay function is activated in STORM once the TC eye will be over land for more than 9 hours. For TCs crossing land in shorter time periods, no decay is assumed. In reality, the inland surface winds will start to decay prior to landfall in response to enhanced surface friction caused by the land mass (Done et al. 2020; Nanaji Rao et al. 2021). In addition, the 2D-parametric wind field model does not take terrain effects into account, meaning that wind speeds on the leeward side of e.g., a mountain ridge, might be overestimated compared to reality. These limitations can be overcome by applying wind field models that also account for terrain effects, such as the Done et al. (2020) model. Such models, however, also require information on variables that are currently not simulated by most synthetic models (including STORM). In addition, they are computationally more intensive, meaning that application to large synthetic datasets is challenging. Nonetheless, future developments in both the field of synthetic modeling and parametric wind field modeling can aim to overcome these limitations, thereby improving wind speed estimates from synthetic models.

In addition, while we solely focused on wind risk in this chapter, TCs can also cause substantial damage through flooding, triggered by storm surges, waves, and rainfall. These hazards, however, are not simulated in the current version of STORM. It is possible to model storm surges and waves by forcing a hydrodynamical model with synthetic TCs from STORM (as was for instance done for the baseline climate in Dullaart et al. 2021), but currently no STORM module exists that allows the modeling of TC rainfall. We plan to develop such a module in future work.

Next, while OpenStreetMap is currently the most widely used and consistent database of geospatial objects available globally, it should be noted that consistency does not equal completeness (Koks and Haer 2020). Unfortunately, very few databases are available to identify how well or how bad certain places are mapped, and comparison can often only be done case-specific. For the SIDS in the South Pacific, we can compare our building estimates with estimates from island specific case studies (Commonwealth of Australia 2013). Comparison estimates show variations between 10% of the buildings in OSM versus the respective case study (e.g., Timor-Leste and Papua New Guinea) up to 135% (Tuvalu) and 168% (Niue) of the buildings in OSM versus the case study analysis. Other islands, such as Vanuatu or Samoa, show almost identical numbers of buildings between OSM and their case study analyses. While this has no implications for our risk estimates, as the LitPop database is used to estimate those values, it could mean that the RP at which 95% of the buildings are exposed to Category 3 wind speeds may shift when more buildings are included. However, we do believe this shift will be limited as our results present relative exposure values. While it is hard to underpin the following claim with numbers, visual exploration of OSM data does indicate that high-density areas do include more buildings compared to low-density areas on each of the SIDS. This implies that our general observations and conclusions with regards to exposure will still hold, even when more buildings are included.

6.5 Concluding Remarks

This study presented in this chapter is the first to provide a uniform TC risk assessment across all SIDS prone to TCs using publicly available data. To calculate this risk, we have combined state-of-the-art TC hazard modelling for the current and near-future (2015–2050) climate (represented through four global climate models), combined with high-resolution exposure information obtained from both OpenStreetMap (OSM) and globally consistent gridded exposure and vulnerability information. The analysis is performed for 48 SIDS, subdivided across four basins (North Atlantic, South Pacific, Western Pacific, and South Indian). We presented RP curves for maximum wind speeds, affected assets (buildings and infrastructure), and damages, both for the baseline and future climate. Our results showed that the largest changes in RPs across these three elements were all found in the South Pacific basin; smallest changes were observed for SIDS located in the North Atlantic.

The aggregated expected annual damage (EAD) for the baseline situation and for all SIDS combined, is estimated at 6.2 billion USD, or 1.2% of their combined GDP, which increases by 2.4% (−8.1–16.2%) towards the future. The results show a clear change towards the future for all risk components, but with distinct differences between the four basins. The North Atlantic basin experiences overall the highest absolute risk, constituting 91.6% of all EAD across the SIDS, but also shows a slight decrease in this EAD towards the future, amounting to −3.8% (−12.1–8.1%). The South Pacific, on the other hand, experiences an increase of approximately 153.6% (104.8–201.0%) compared to the baseline climate. The risk in the Western Pacific and South Indian basins increases by 64.5% and 58.0%, respectively. In terms of absolute EAD, the Dominican Republic experiences the highest absolute risk, totaling to 1.47 billion USD in the baseline climate. Even though this estimate decreases under climate change to 1.41 billion USD, it is still the highest risk across all SIDS.