Coastline segments
For reasons of data availability and socioeconomic relevance, the analysis was limited to latitudes between 66° N and −60° S. In this area of interest, the world was divided in 1 arcmin (~2 km) grid cells. To define a logical position for the establishment of an efficient levee, the coastline location was derived from the OpenStreetMap68, moved 100 m land inward and smoothed. For every cell containing a coastline segment, coastline length and a coast-normal transect were derived at the center of segments resulting in 495.361 transects that are on average 1.1 km apart. Bootstrapping revealed that transect distances up to 2 km give very similar results. All transects stretch 4 km seaward and 4 km inland to fully capture most foreshores.
Elevation data
A global intertidal bathymetry/elevation dataset from high-resolution EO data (USGS Landsat and Copernicus Sentinel-2), the Foreshore Assessment using Space Technology (FAST) intertidal elevation map69, was produced to compliment commonly used global data products with low resolution and higher inaccuracy in intertidal zones. Global coastlines were divided over 25000 tiles of each 40 × 40 km2. For these tiles, all available images were collected for the period between 1997 and 2017. Surface water was identified, using normalized difference spectral indices (NDSI, here SWIR1 and Green band) for all images (median of 317 images per tile) covering various tidal conditions, and the per pixel mean calculated to derive time-ensemble average (TEA) NDSI images. We developed a new technique to transform TEA images to intertidal elevation independently of in situ calibration data. TEA-NDSI images were normalized by the spatially averaged NDSI values of regions identified (using global elevation datasets) as land and water, respectively. This resulted in a single image per tile that represented the inundation probability for each pixel in the intertidal zone. The inundation probability represents the long-term average tidal inundation, because it was derived from a collection of images that span a time period similar to the tidal epoch (period of 19 years). Pixels having a probability of 1 represent permanent water, and have elevations less than or equal to the lowest astronomical tide (LAT), whereas land (p = 0) represents elevations higher than or equal to the highest astronomical tide (HAT). By deduction, p = 0.5 is equivalent to local mean sea level (LMSL). Tidal statistics from the global tide model FES2012 were used to couple the derived inundation probability to an elevation. The main source of bed level data originates from this map and has a 20 m horizontal resolution and typically a 30–50 cm vertical accuracy (RMSE = 0.52 m, MAE 0.42 m, as assessed at a number of sites with high quality elevation data (Supplementary Fig. 7)). Bathymetry data (GEBCO35; 30 arc-second horizontally, tens of metres vertically) and topography data (MERIT36; 3 arc-seconds, 2 m vertically) were merged to create a continuous bathymetry-elevation map by changing the vertical datum of MERIT from EGM96 to MSL by assuming 0 m +MSL at the OSM coastline. Global bathymetry datasets (e.g. GEBCO) and elevation datasets (e.g. SRTM and MERIT) lack accuracy (especially nearshore), but are commonly used17,18,23,34. The final bed level was constructed using FAST intertidal data where sufficient valid data points were available, complemented by the merged GEBCO-MERIT data where these points were lacking.
Vegetation extent
The FAST coastal vegetation map69 was based on Landsat-8 and Sentinel-2 satellite images collected between 2013 and 2017. The map provides actual vegetation presence at 10 m resolution. Vegetation presence was obtained by applying an individual NDVI threshold per tile, with a total of 25,000 tiles, based on the yearly NDVI average and NDVI amplitude. The FAST coastal vegetation map is validated based on NDVI comparison with local measurements taken at Zuidgors, The Netherlands (R2 = 0.92) (Supplementary Fig. 8). If vegetation was present, the vegetation type was determined by global salt marsh32 and mangrove14 maps, complemented with Corine Land Cover30 (CLC, Europe only) and GlobCover v2.231 maps when there is no coverage. Determining global coastal vegetation extent is difficult and affected by eutrophication in coastal environments. This behaviour is observed on the coast along the Persian Gulf and the Red Sea. To improve accuracy only vegetated transects identified by the global salt marsh32 and mangrove14 map and confirmed by the FAST coastal vegetation map are included for these areas. Moreover, vegetated transects with a green belt width smaller than 250 m identified by GlobCover are excluded from the study for accuracy reasons (Supplementary Fig. 8). To avoid mixed vegetation types, the vegetation type was determined by the most dominant type. The vegetation width constituted of the sum of vegetated grid cells between the start and the end of the vegetated zone.
Water level and wave data
The design water levels were based on a combination of tide and storm surge for the selected probability of occurrence (return periods 2, 5, 10, 25, 50, 100 default, 250, 500, 1000 years) and came from the GTSR dataset34. SLR and subsidence were not taken into account because this study focuses on the present situation. Moreover, quantifying the future role of vegetated foreshores would not only require SLR scenarios but also an insight in the development of wetlands over time, which is strongly determined by local conditions such as sediment supply56,57,60. Offshore wave conditions were obtained from ERA-Interim33 re-analysis, based on data from 1979 till 2017 and reprojected to Dynamic Interactive Vulnerability Assessment (DIVA)70 points. Next, the Peak Over Threshold method was applied to construct representative values for the significant offshore wave height, Hs and the peak wave period Tp for all the return periods. The nearshore wave height was limited by the local water depth at the start of the (vegetated) foreshore using a breaker criterion (gamma = 0.55). This is a fairly low value considering the range of values cited in literature71 leading to conservative wave attenuation by vegetation results. Wave-bottom interactions in the sub-tidal zone and processes such as refraction and diffraction are not explicitly simulated. The conservative breaker criterion is chosen to implicitly account for these processes in a conservative manner. The wave period remained unchanged and the wave direction was assumed coast normal and wave growth along the transect due to wind effects was excluded. However, for the current study a more sophisticated approach to account for longshore wave variability based on topography was considered infeasible at the global scale and considered to yield limited outcome looking at the uncertainty in socioeconomic factors. The average Hs,offshore = 4.6 m (std = 2.0 m) and the average Hs,startforeshore = 0.7 m (std = 0.7 m).
Profile construction
The 8 kilometre coast-normal transects consisted of 321 gridpoints, thus a horizontal grid resolution of 25 m. We used four different methods: Foreshore method 1 (based on the FAST intertidal elevation map), Foreshore method 2–4 (based on MERIT-GEBCO). The properties of the FAST intertidal elevation map, MERIT and GEBCO are described under the header ‘Elevation data’. Foreshore method 1 produced the most accurate profiles and foreshore method 4 the least accurate profiles. The profile construction steps are described hereafter. Validity checks were performed to identify false indications of intertidal area in the FAST intertidal elevation map. Individual data points were marked invalid and removed in case: (1) MERIT points were situated above the surge level with a return period of 2 years, while data from the intertidal map indicated a lower elevation. (2) Data from the FAST intertidal map was situated at open sea. (3) Data from the FAST intertidal map along the transect dropped below a minimum range threshold of 10 cm. A fourth check was performed based on the continuity of the data. Data from the FAST intertidal map contain discontinuities along the profile. These continuities exist on pixel level due to the use of the modified normalized difference water index and in some instances cloud coverage was preventing full coverage. Lastly, discontinuities arise due to the presence of (high elevated) tidal flats and banks in coastal areas. (4) Data length was defined as the length of continuous data points along the transect. If the data length of a patch decreased below a threshold of 100 m, the points were marked invalid. Gaps between valid data patches were filled using linear interpolation if the gap was smaller than 250 m. Eventually, one, none or multiple valid data patches were found along the transects. See Supplementary Fig. 2 for example transects.
Global coastline shapes range from straight sandy coastal stretches to complex coastlines often found in estuaries. With a transect length of 8 km, the start and the end of the transects could both be situated on land, hampering an unambiguous identification of the foreshore of interest. We designed the algorithm such that the last foreshore was selected. For profiles using data from the FAST intertidal map (foreshore method 1, 50.9% of populated susceptible coastlines), the last valid patch corresponds to the last foreshore. The inclusion of tidal flats as part of the foreshore was determined based on the gap length. In case no (sufficient, thus not satisfying the minimum data length criterion of 100 m) valid data was available from the FAST intertidal map based on the four described checks, the profile was based on a merged GEBCO-MERIT set (methods 2, 3 and 4), respectively, 46.1%, 3.0% and 0.01%. For the second method, data points were selected between a minimum threshold of −2 m MSL and a maximum threshold equal to the surge level with a return period of 2 years. Next, for the selected points the direction of the slope was determined by comparing elevation between the data point concerned and the next data point. This resulted in patches of upward sloping sets of data points between the minimum and maximum threshold. Similar to foreshore method 1, the validity of the patches was checked using data length, gap length and the corresponding thresholds of 100 m and 250 m. The start and the end of the foreshore were determined by the first and last valid point of the last patch. Foreshore method 3 was used if not sufficient foreshore data were available to satisfy the minimum data length threshold (100 m). In these cases, the start of the foreshore was defined as the first upcrossing intersection with −2 m MSL along the transect. The end of the foreshore corresponded to the intersection between the elevation profile and the governing surge level with a return period of 2 years. Foreshore method 4 was used if no start and or end of the foreshore could be found. In this case the start and/or end point of the foreshore corresponded to the first and last data point, respectively.
In some cases, elevation for the end of the foreshore was missing due to several reasons. First, the upper part of the intertidal zone was sometimes missing from the FAST intertidal map, due to low frequency of inundation of the upper intertidal zone or cloud cover. Second, bed elevation in mangrove belts was hard to define based on satellite imagery, as the canopy is detected as the earth surface. These uncertainties were counteracted by consulting the mangrove and salt marsh maps. If vegetation was present in one of these maps, the derived foreshore was extended until the end of the vegetated zone. An elevation equal to the surge level with a return period of 2 years was chosen as elevation for extended foreshore points with an elevation exceeding this surge level.
Vegetation parameters
As deducting the type and size of mangrove trees and salt marshes from EO data at global scale is not possible (yet), the current modelling approach relies on field and literature observations. For the scope of this research the properties of the mangrove trees occurring at the seaward side of the mangrove belt are the most relevant. To avoid overestimation of wave attenuation in young mangrove forests, the mangrove dimensions are chosen such to be representative for young fringing pioneering mangroves up to a height of 3 m that are practically vertically uniform compared to mature trees. The modelling approach uses four parameters to represent vegetation: height, diameter, number of stems and drag coefficient. The exact characteristics are based on observations in literature8,9,72,73,74,75,76 (N = 30 m−2, d = 35 mm, h = 3.0 m).
High quality observations on wave attenuation by mangroves under storm conditions do not exist. For the drag coefficient the theoretical value, 1, of a rigid cylinder is chosen, because mangrove trunks can be considered rigid. For salt marshes a winter state representative as found in NW Europe is chosen. The values are defined based on FAST field tests (Romania, UK, Spain and the Netherlands) and literature10,24,77,78 (N = 1225 m−2, d = 1.25 mm, h = 0.30 m). A drag coefficient (CD) of 0.19 is chosen, which is the lower limit found during large-scale flume tests10. The drag coefficient depends on biophysical characters as well hydrodynamics. The drag coefficient represents drag due to skin friction and pressure differences, but also effects like swaying motion of stems24. The 1D modelling approach takes into account gaps in vegetation cover, e.g. due to the presence of channels. Zonation of vegetation types is not implemented, because this level of detail is insignificant in relation to the inaccuracies induced by the use of global datasets.
Wave attenuation model
To determine wave attenuation along the foreshore transects and the resulting significant wave heights relevant for the flood defence on a transect, we used a lookup-table approach. The lookup table was generated by combining 668,304 model output values for different combinations of foreshore slopes, vegetation covers and hydrodynamic conditions. The table contained wave heights modelled by XBeach79 in surfbeat mode (a nearshore numerical wave model that accounts for the presence of vegetation) at regular intervals along a steady slope, both with and without vegetation. XBeach uses for wave-vegetation interaction the rigid cylinder80 approach and includes an energy sink term to the wave energy balance to implement wave dampening81. We used conservative vegetation characteristics, winter state salt marshes and young pioneering mangroves. We characterized foreshores by their width and slope. The foreshore profile was the same for simulations with and without vegetation. The foreshore width was determined by calculating the distance between the start and the end of the foreshore. The slope was estimated using a linear regression. This approach has two advantages over detailed modelling of wave attenuation over all transects: it is much quicker, allowing for iterative improvements of the workflow and it does not suggest the precision one would expect from detailed models but cannot be delivered with global data. Average Hs,endforeshore,noveg = 0.6 m (std = 0.5 m) and Hs, endforeshore,veg = 0.3 m (std = 0.4 m).
Coastline susceptible to flooding, urban and rural extents and population density
To assess the need for coastal flood defences, we made a distinction between areas susceptible to coastal flooding and higher, non-susceptible areas. We determined susceptible areas based on possible inundation using coastal flood maps of 1 km resolution for a 1/1000 year surge level. These maps were created with a global geographic information system (GIS) based inundation model that is forced with a spatially varying sea level, accounting for attenuation of the water level due to land surface roughness82. A method that is more sophisticated compared to a simple ‘bathtub’ inundation method. Topographic features, as visible in MERIT, protecting the land from flooding are considered. To classify coastlines as urban or rural a distinction was made based on gridded population from the LandScan database83 using the 2UP model84. A transect is characterized ‘urban’ if it intersects at least one cell with an urban population with a minimum of 1. Populated coasts have been identified by assigning the population density of the population susceptible to flooding in the proximity of the transects. We used WorldPop201785 population data and assigned population to the transects using a buffer of 15 kilometre radius. The population density is the division of the assigned population and the total area of the assigned cells. This procedure is repeated for buffer radius of 5, 10 and 20 km, giving fairly comparable outcomes. Following this approach we found a ratio between rural and urban transects of 73/27.
Levee crest heights
The empirical EuroTop formulations47 gave the required levee heights with respect to water levels and wave heights, assuming the presence of a levee at the end of the vegetated foreshore. We hereby neglected the position and characteristics of levees present in the current situation, as no global dataset of coastal protection structures exists. The assumed levee had a standard 1:3 levee profile without berms and an allowed overtopping discharge of 1 l s−1 m−1. These parameters are representative for simple, low-cost levees in developing countries but conservative for well-constructed and maintained levees. Consequently, savings on levee heights in countries with strict protection standards are overestimated, as reduction in required levee height due to vegetation presence is likely less than predicted here. However, this may be balanced out by the fact that we calculated with an average national construction cost per kilometre and levees applying to stricter protection standards may actually be more expensive (Supplementary Fig. 5).
Costs for levee construction and crest height reduction
The calculated levee crest height reductions were monetized using a levee unit price per kilometre length per metre heightening. We used an unit investment costs of levees (metre heightening per kilometre length) of USD 7.0 million42. This estimate represents an average of construction costs in the USA and the Netherlands stated in several studies86,87,88,89. It pertains to all investments costs, including ground work, construction, engineering costs, property or land acquisition, environmental compensation, and project management. Investment costs per metre heightening are well described by a linear function without intercept90. They concluded that for large-scale studies it is sufficient to assume linear costs for each metre of heightening, including the initial costs and the 95% confidence range is between 3x and x/3, where x is the unit cost value. Subsequently we applied three unit levee investment cost prices (low: USD 2.33 million, mid: USD 7.0 million, high: USD 21 million) in line with previous studies42,90. These cost estimates were then adjusted for all other countries by applying construction index multipliers (based on civil engineering construction costs91), to account for differences in construction costs across countries92. Costs were converted to USD2005 power purchasing parity (PPP), to be consistent with the SSPs, using GDP deflators from the World Bank (https://data.worldbank.org/), and annual average market exchange rates between Euros and USD taken from the European Central Bank (unit levee cost per country = unit levee cost x construction index per country / PPP MER rate 2005 index per country). Example: mid unit levee costsUSA = 7.0 ×1 / 1 = 7.0 million USD2005 PPP km m−1. If for a country data was not available in the database, we used the average of all countries in the same World Bank income group. For the reference year 2005, this applies to Western Sahara (ESH), North-Korea (PKR) and Somalia (SOM).
Reliability
A scoring table was used to get insight in the reliability of the results of the global analysis. Results were grouped into four reliability classes ranging from “poor” to “very good”. Transects were placed in these classes based on data accuracy for three characteristics: hydrodynamics, vegetation and profile elevation. In Supplementary Fig. 6 the (sub) results of the analysis are presented. The first category, hydrodynamics, included known inaccuracies in the hydrodynamic data (GTSM and ERA-I). Data from the GTSM model was considered less reliable in areas with a low tidal range and/or with tropical storms, such as cyclones or hurricanes, as those were not included in our analyses. Also wave data from ERA-I are less reliable in these areas, because the effects of tropical storms are flattened due to the relatively coarse grid size. Hence, transects in these areas were pinpointed by linking them to NOAA data of historical hurricane tracks93. In Supplementary Fig. 6B, areas where tropical storms occur can clearly be recognized. In addition, the Mediterranean Sea, the Red Sea, the Black sea and the Caspian sea stand out in inaccuracy, because of limited tidal action.
Reliability of vegetation characteristics was determined by data source and vegetation width. For transects with extensive vegetation widths, crest height reduction was less sensitive for possible deviations of the vegetation width, due the non-linear relation between vegetation width and wave reduction. Vegetation cover proved most reliable in areas where data from the salt marsh32—and mangrove map14 were available. Hence, this resulted in a ‘good’ score (Supplementary Fig. 6C). Only in cases of extensive vegetation presence was a ‘very good’ score assigned. Transects were appointed as “very good” if vegetation extended 500 m for mangroves, and 1000 m for salt marshes. These thresholds are chosen based on our model results, which show that after ~500 m (salt marshes) and 1000 m (mangroves) maximum reduced wave transmission by foreshore vegetation is reached. Vegetation cover reliability in Europe was classified as ‘good’, due to reliable vegetation type classification based on CLC30 and the salt marsh map32 in combination accompanied by relatively small vegetation widths. The reliability of the derived vegetation characteristics is especially lacking at the east coast of Canada, at Latin America’s south coast, at Africa’s coasts facing the Mediterranean Sea, coasts along the Red Sea and the Persian Gulf, and along the coasts of China, Japan and Russia. For example, in the Persian Gulf states the vegetation presence map tends to falsely identify foreshores as vegetated.
The time-ensemble average (TEA) technique applied for the FAST intertidal elevation map relies on the availability of a reasonable number of images at different tidal stages where the differences in horizontal extent of water coverage can be identified, thus allowing a composite of inundation frequency to be derived. However, the technique is limited by the effective sensor resolution (~30 m, including uncertainty in georeferencing) relative to the horizontal extent of changes in inundation, a function of the tidal range and bed slope. Hence, changes in tidal water extent in microtidal or very high bed-slope regions tend to be too small for reliable discerning differences, leading to poor performance of the technique. However, the merged GEBCO-MERIT dataset was considered less reliable than the FAST intertidal map, based on the resolution and the merging of the two underlying datasets in the intertidal zone. In addition, MERIT tends to overestimate the elevation in mangrove areas, as it measures the canopies as the earth’s surface. Besides the elevation data, the foreshore definition method is used as a profile reliability indicator. The total score per transect is given by the sum of the sub-scores. The sub-scores are normalized to give equal weight to the scoring categories.
Validation
For validation of our method to assess vegetation presence, a comparison of 280 randomly located transects with aerial imagery was carried out. The area accessed in the global assessment was divided in tiles of 90 degrees longitude and 15 degrees latitude. From each tile 6 vegetated and 2 non-vegetated transects were selected. Next, a reference dataset was created by manually identifying vegetation presence using present imagery. Lastly, the vegetation width derived by the model and the manually derived set were compared (Supplementary Fig. 8). For this comparison we made three distinctions, based on (1) vegetation type, (2) foreshore derivation method and (3) vegetation cover source. Comparison showed that the used algorithm on global EO data performs satisfactorily (Supplementary Fig. 8), but in some cases tends to assign a vegetation cover of up to 250 m where there is none. Deviation between observation and the global assessment, is caused by methodological error in the global assessment and inaccuracy in the global datasets, e.g. different timestamps are inevitably compared. This would induce an exaggeration of the effect of vegetation. However, due to the limited dimension of the vegetation extent, the threshold for substantial crest height reduction is falsely exceeded in not more than 2.4% of the cases and the effect is largely balanced out by underestimation of the vegetation cover at larger lengths.
To validate wave reduction by vegetation calculated through our lookup table approach, we compared results with local modelling results for the South-Western part of the Netherlands for 38 vegetated transects. The numerical model SWAN94 in stationary mode was used to translate wave conditions from offshore to nearshore. The simulations were performed with a grid size of 0.01 deg and bathymetry from EMODNET95. Extreme water levels were included by a water depth correction, using data from GTSR18. Both wind and wave boundary conditions were derived from the earlier described ERA-I re-analysis. The governing wave direction was based on the average of the fifteenth highest wave events in the available wave data. The wind direction was assumed to be aligned with the wave direction. A parametric JONSWAP spectrum shape was used, using a peak enhancement factor of 3.3 and directional spreading of 20 degrees. Foreshore profiles were constructed using an approach similar to foreshore method 2 in the global study but using local high-resolution bathymetry and topography data. Vegetation width was extracted from the salt marsh map32, which was confirmed locally using aerial imagery. Foreshore wave propagation was determined using XBeach in surfbeat mode79.
Our results showed an overestimation of the water depth at the start of the vegetated zone by 0.73 m on average. In addition, the global model derived milder slopes in comparison to the local analysis for narrow vegetated transects. The largest errors were found further away from the mouth of the estuary. Here, the deviation between the wave calculated by SWAN and the depth limited approach is largest. The wave height at the start of the vegetated zone was overestimated on average by 1.12 m, due to the complex geometry and the sheltered configuration of the estuary. The algorithm approximated the wave transmission reduction (RMSE 13%) and the levee crest height reduction relative to the required crest height without vegetation presence (RMSE 19%) with reasonable accuracy (Supplementary Fig. 9).
Sensitivity analysis
A sensitivity analysis has been performed to provide insight in the uncertainty in the presented potential global levee costs savings. The analysis focused specifically on single key parameters, such as the levee unit cost, the critical overtopping discharge and the wave breaker index. High, mid and low levee unit cost scenarios are taken from previous studies42,90. A high, mid, low for the critical overtopping discharge are respectively 10, 1 and 0.1 l s−1 m−1 to incorporate the quality of the levee cover47. We chose RP10 and RP1000 for, respectively, the low and high storm return period scenario. The uncertainty spread of vegetation width is based on the 75% confidence intervals of the underestimated and overestimated vegetation widths of mangroves (+436 m, −136 m) and salt marshes (+597 m, −104 m) in the vegetation presence validation study. For the breaker index we solely chose a high scenario of 0.78, because the index of the global assessment (0.55) was already quite conservative71. For topography we applied a range corresponding to the typical vertical accuracy of the FAST intertidal elevation dataset (±50 cm). Two representative subsets of 500 transects for respectively mangroves and salt marshes have been derived using the clustering method k-means96, based on hydrodynamic conditions, vegetation cover, profile characteristics and geographical location. With these subsets, we repeated the analysis procedure of the global assessment for the sensitivity scenarios. The results point out that the largest spread is caused by the uncertainty in the unit levee cost with −66% and +200% for, respectively, the low and high scenario with respect to the global reference analysis. The other scenarios: topography (−39%, +47%), critical overtopping discharge (−40%, + 40%), storm return period (−28%, +34%), vegetation width (−28%, +39%), breaker index (+21%) (Supplementary Fig. 10). Larger water depths result in a decrease of depth-induced wave energy dissipation and more dissipation due to wave-vegetation interaction, which explains the outcomes of the topography sensitivity results. Similarly, an increase of the storm return period or the breaker index shifts the ratio of wave energy dissipation by wave-bottom interaction and wave-vegetation interaction. The coastal protection costs by vegetation are sensitive to critical overtopping discharge changes, because of the non-linear relation between the wave height in front of the levee and the overtopping discharge47.
Source: Ecology - nature.com
