Hindcasted and forecasted deep-water wave data from WaveWatchIII (Tolman 1997, 1999, 2009) simulations forced from four Intergovernmental Panel on Climate Change (IPCC; https://www.ipcc.ch/) Sixth Assessment Report (AR6) Coupled Model Intercomparison Project, Phase 6 (CMIP6; https://wcrp-cmip.org/cmip-phase-6-cmip6/) GCMs were produced for 31 years (2020–2050) by Erikson and others (2022) for the Hawaiian, Mariana, and American Samoan Islands. Similarly, hindcasted and forecasted tide and storm surge data from the Global Tide and Surge Model (GTSM; Verlaan and others, 2015; Muis and others, 2016, 2020) simulations were forced using the same four GCMs for the same 31 years (2020–2050) by Muis and others (2022) for the Hawaiian, Mariana, and American Samoan Islands. The CMIP6 models are from the HighResMIP project (Haarsma and others, 2016) and are used for both the WaveWatchIII and GTSM simulations: GFDL-CM4C192-highresSST (Guo and others, 2018), CMCC-CM2-VHR4 (Scoccimarro and others, 2017), HadGEM3-GC-31-HM_highres-future (Roberts, 2019a), and HadGEM3-GC-31-HM_highresSST-future (Roberts, 2019b). The future simulations (2020–2050) used IPCC-AR6 Shared Socioeconomic Pathway 8.5 (Lee and others, 2021), which results in a year 2100 radiative forcing level similar to the IPCC 5th Assessment Report’s Relative Concentration Pathway 8.5 climate scenario.

Following the methodology of Camus and others (2011), more than 270,000 hourly data on wave climate parameters were propagated to the nearshore using a hybrid downscaling approach. The offshore wave climate data were synthesized into 999 combinations of sea states (wave height, wave periods, and wave directions) that best represented the range of conditions from the Erikson and others (2022) database. These selected sea states were then propagated to the coast using the physics-based Simulating Waves Nearshore (SWAN) spectral wave model (Booij and others, 1999; Ris and others, 1999; SWAN, 2016), which simulates wave transformations nearshore by solving the spectral action balance equation. Wave propagation around reef-lined islands has been accurately simulated using SWAN (Hoeke and others, 2011; Taebi and Pattiaratchi, 2014; Storlazzi and others, 2015). Standard SWAN settings were used (for example, Hoeke and others, 2011; Storlazzi and others, 2015), except that the directional spectrum was refined to 5-degree bins (72 total) to better simulate refraction and diffraction in and amongst the islands (appendix 1).

To accurately model from the scale of the island groups or large sections of coastline (on the order of tens of kilometers) down to local scales (on the order of hundreds of meters), a series of dynamically downscaled nested, rectilinear grids were used. The coarse (5-kilometer [km] or 1-km resolution) SWAN grids provided spatially varying boundary conditions for finer-scale (1-km or 200-meter [m] resolution) SWAN grids, with the finest resolution (200-m) grids used for the rest of the modeling infrastructure (fig. 1, appendix 2). The bathymetry for the SWAN grids were generated by grid-cell averaging various topobathymetric digital elevation models (appendix 3). The shallow-water wave conditions from 999 sea-state combination simulations in the finest SWAN grids were extracted at 100-m intervals along the coastline, at a water depth of 30 m, and then reconstructed into hourly time series using multidimensional interpolation techniques (Camus and others, 2011).

Benthic habitat maps defining coral reef spatial extent and percent coral cover (appendix 6) were used to delineate the location of nearshore coral reefs and their relative coral abundance along the reef-lined shorelines (fig. 2). Cross-shore transects were created every 100 m alongshore (appendix 4) using the Digital Shoreline Analysis System software version 4.3 in ArcGIS version 10.3 (Thieler and others, 2009). Transects were cast in both landward and seaward directions using the Smoothed Baseline Cast (SBC) method with a 500-m smoothing distance, perpendicular to a baseline generated from coastlines digitized from USGS 1:24,000 quadrangle maps and smoothed in ArcGIS using the Polynomial Approximation with Exponential Kernal algorithm and a 5,000- m smoothing tolerance. Transects had a cross-shore resolution of 1 m and varied in absolute length to ensure each intersected the −30 m and +20 m elevation contours relative to mean sea level. The bathymetric (appendix 3) and coral cover (appendix 6) data were extracted along these shore-normal transects and assigned to the closest transect grid cells.

The nearshore wave time series (hourly data from 2020 to 2050) at the 30-m isobath were fit to a general extreme value distribution (Méndez and others, 2006; Menéndez and Woodworth, 2010) to obtain the significant wave heights associated with the annual (1-year), 20-year, and 100-year storm return periods in the SWAN grid cells at the end (or nearest the end) of each transect. The corresponding annual (1-year), 20-year, and 100-year storm return period tide and storm surge water levels for the location were taken from the nearest GTSM output point nearest to the offshore end of the each transect.

The return value significant wave heights and associated mean peak periods from SWAN were then propagated over the coral reefs with corresponding (static) return value water levels from GTSM along 100-m spaced shore-normal transects (appendix 4) using the numerical model XBeach (Roelvink and others, 2009; XBeach, 2016), as demonstrated in figure 2. XBeach generated forcing wave time series for each modeled storm return period, which were reused as inputs for modeling different sea level rise scenarios under the same return period. XBeach solves for water level variations up to the scale of long (infragravity) waves using the depth-averaged, nonlinear shallow water equations. The forcing is provided by a coupled wave action balance, in which the spatial and temporal variations of wave energy owing to the incident-period wave groups are solved. The radiation stress gradients derived from these variations result in a wave force that is included in the nonlinear shallow water equations and generates long waves and water level setup within the model. Although XBeach was originally derived for gently sloping sandy beaches, with some additional formulations, it has been applied in reef environments (Pomeroy and others, 2012; van Dongeren and others, 2013; Quataert and others, 2015; Storlazzi and others, 2018) and proved to accurately predict the key reef hydrodynamics.

XBeach was run for 9,000 seconds (s) in one-dimensional hydrostatic mode along the cross-shore transects, at a varying resolution between 10 m seawards and 1 m landwards (resolution varies depending on water depth); the runs generally stabilized after 1,500 s (spin-up time) and thus generate good statistics on waves and wave-driven water levels for more than 2.5 hours (appendix 5). The application of a one-dimensional model neglects some of the dynamics that occur on natural reefs and shorelines, such as lateral flow. Thus, the flooding is likely underrepresented around promontories where wave-energy convergence would cause increased wave-driven flooding that is not captured with one-dimensional models. However, it does represent a conservative estimate for infragravity generation and wave runup, as the forcing is shore normal (for example, van Dongeren and others, 2013; Quataert and others, 2015). Moreover, to reduce the overprediction of infragravity wave energy, the short-wave group variance at the boundary was reduced by 45 percent (wbcEvarreduce = 0.55), per de Goede and others (2020). The choice of a one-dimensional model is warranted in this case because the offshore waves (that is, wave propagation modeled with SWAN) were generally near-normal at the offshore end of the XBeach transects.

The additional formulations that incorporate the effect of higher bottom roughness on incident wave decay through the incident wave friction coefficient (fw) and the current and infragravity wave friction coefficient (cf), as outlined by van Dongeren and others (2013), were applied. The friction induced by corals was parameterized based on the spatially varying coral coverage data and results from a metaanalysis of wave-breaking studies over various reef configurations and friction coefficients for the different coral coverages (for example, van Dongeren and others, 2013; Quataert and others, 2015). Coral coverage classes, as established by the benthic habitat maps, were assigned fw and cf (table 1) over the spatial extent of the reef along the profile as defined from the benthic habitat maps (appendix 6). The future wave and storm surge conditions for each storm return interval were then propagated using the XBeach models over the same 100-m spaced shore-normal transects but modified to account for the different sea-level rise scenarios (fig. 3).

The Deltares 2-dimensional Super-Fast Inundation of CoastS coastal flooding model (SFINCS) is a super-fast flooding model that dynamically calculates two-dimensional compound flooding maps in coastal areas (Leijnse and others, 2021), which is a vast improvement from interpolating between adjacent XBeach transects per Storlazzi and others (2019). The model uses simplified mass and momentum equations to compute flooding based on water levels and boundary conditions, such as waves, precipitation, and river discharges. Usually, SFINCS ignores the advection term, except for conditions of supercritical flow or when including waves as boundary conditions. In this case, SFINCS was forced with water level and infragravity wave time series; thus, the advection term is included (appendix 7).

The SFINCS boundary conditions were determined from XBeach still water level (tides and surge). The XBeach time series outputs were extracted at the intersection between the transect and the 0.5-m bathymetric contour (below mean sea level), which was below the still water levels during the simulations owing to tides and storm surge; these time series form the basis to force the constant 10-square-meter resolution SFINCS grids (appendix 8), which extended from the 0.5-m bathymetric contour to the 10-m contour. The water level time series boundary conditions were built with a slow ramp-up to avoid initial bathtub-type flooding. The ramp-up goes from mean sea level to the average incoming water level calculated from XBeach. The wave time series were computed as a random signal with a random phase, which were generated from the spectrum calculated from the XBeach incoming water levels (Roelvink and others, 2009; van Dongeren and others, 2013). Both boundary conditions were smoothed in an alongshore direction between adjacent output points to represent a two-dimensional environment. SFINCS was run for 3 hours after the water level ramp-up. A Manning coefficient, which represents friction applied to the flow by the seafloor roughness, of 0.035 was used to account for infragravity wave friction. SFINCS was run for the 3 storm return intervals (fig. 4) and 7 sea-level rise scenarios (fig. 5). SFINCS output flood depth raster data were exported to a geographic information system; the depth rasters were then exported as geotiffs and the flood extent polygons were exported in shapefile format.

Numerical flood modeling errors were estimated to be ±0.5 m. This value is greater than the root-mean-square and absolute errors computed between model results and measurements (van Dongeren and others, 2013; Quataert and others, 2015) but was used to compensate for the limited number of storms tested and the large geographic scope compared to regions where validation measurements are available. Uncertainties associated with the baseline digital elevation model varied based on input data; see references listed in appendix 3. Other limitations and assumptions pertaining to flood extents include the following: