A global biophysical typology of mangroves and its relevance for ecosystem structure and deforestation

Mangrove forests provide many ecosystem services but are among the world’s most threatened ecosystems. Mangroves vary substantially according to their geomorphic and sedimentary setting; while several conceptual frameworks describe these settings, their spatial distribution has not been quantified. Here, we present a new global mangrove biophysical typology and show that, based on their 2016 extent, 40.5% (54,972 km2) of mangrove systems were deltaic, 27.5% (37,411 km2) were estuarine and 21.0% (28,493 km2) were open coast, with lagoonal mangroves the least abundant (11.0%, 14,993 km2). Mangroves were also classified based on their sedimentary setting, with carbonate mangroves being less abundant than terrigenous, representing just 9.6% of global coverage. Our typology provides a basis for future research to incorporate geomorphic and sedimentary setting in analyses. We present two examples of such applications. Firstly, based on change in extent between 1996 and 2016, we show while all types exhibited considerable declines in area, losses of lagoonal mangroves (− 6.9%) were nearly twice that of other types. Secondly, we quantify differences in aboveground biomass between mangroves of different types, with it being significantly lower in lagoonal mangroves. Overall, our biophysical typology provides a baseline for assessing restoration potential and for quantifying mangrove ecosystem service provision.

www.nature.com/scientificreports/ At a global scale, mangroves are of considerable value to humans 8 yet it is also recognised that the value derived from mangroves varies geographically and that this variability is as yet poorly quantified 9 . Mangroves show substantial geographic variation in structure, height 10 , and species diversity 11 , driven by factors such as climate, tidal amplitude and particularly geomorphic setting. These factors, in turn, can also influence variability in ecosystem functions and services such as carbon storage [12][13][14][15][16] , coastal protection 17 and fisheries 18 .
Despite the importance of the geomorphic setting of mangroves in determining their ecosystem service delivery 12 , their relative risk under future climate change and sea level rise 19 and in influencing optimal restoration actions 20 , many recent global analyses have assumed a spatial uniformity of mangrove forests 9,21 . This has, in part, been determined by the binary (presence/absence) nature of previously available global mangrove extent maps 11,22,23 . By contrast, recent efforts to quantify the variability in mangrove soil carbon have illustrated the utility of applying broad coastal geomorphic settings to explain levels of ecosystem service delivery [13][14][15] . However, until now a mangrove-specific global biophysical typology of geomorphic setting has not been generated. Such information would allow for tailored conservation and management strategies to be developed to protect ecosystem services provided by mangroves 24 and determine appropriate restoration actions 20 .
Here, we present a global-scale, mangrove specific, biophysical typology that integrates the main drivers of spatial heterogeneity of mangrove ecosystems into mappable units. The biophysical typology was developed by reviewing existing, largely qualitative, classifications and applying a model of key spatial attributes that could be mapped consistently at the global scale. The biophysical typology was applied to maps of global mangrove extent generated by Global Mangrove Watch (GMW) 25 and is not itself a predictor of mangrove presence or absence. This typology provides a framework for future analyses, allowing for better incorporation of the spatial heterogeneity of geomorphic and sedimentary setting. We provide two examples of such analyses: firstly, quantifying how mangrove extent change over the period 1996 to 2016 varied between different mangrove types; and secondly showing the potential application of our typology in informing global analyses of ecosystem structure using a dataset of mangrove above-ground biomass 10 .

Results and discussion
Global distribution of mangrove types. We sought to create a broad-scale biophysical typology that was parsimonious with existing theoretical classifications 12,[26][27][28][29] , with our types, deltaic, estuarine, lagoonal, and open coast mangroves, comparable to previous typological classes (Table 1). Our efforts represent the first attempt to map a mangrove biophysical typology beyond individual case study areas. To map the biophysical typology, we developed a map of coastal embayments and used a machine-learning classifier to assign each embayment with a type through reference to ten environmental covariates. The biophysical typology was framed around three of the macroscale groupings defined by Woodroffe and colleagues 29 , and Twilley and Rivera-Monroy's 28 'geomorphic types' . In addition, we derived an 'open coast' type that incorporates several of the divisions in other typologies (Table 1), such as drowned bedrock valleys 26 and carbonate mangroves found on oceanic islands 28 . The four mangrove types represent macroscale units with a resolution of kilometres 30 . Open coast and Table 1. A summary of existing mangrove typologies illustrating the relationship between previously described mangrove types and the one developed and mapped in this study. Where GEO refers to geomorphic setting, and SED refers to sedimentary setting.

This Typology
Thom 26 Woodroffe 27 Twilley and Rivera-Monroy 28 Balke and Friess 20 www.nature.com/scientificreports/ lagoonal mangroves were also assigned a second-tier sedimentary type, as either terrigenous (i.e. dominated by minerogenic sedimentation from terrestrial sources), or carbonate (i.e. dominated by calcareous sedimentation), based on sediment supply, and tidal energy 20 . Full definitions of the types are given in Supplementary Section 1. We used the most recently available high-resolution mangrove presence/absence time-series to map the biophysical typology and enable spatially explicit estimates of change in mangrove type. The GMW generated a 2010 baseline of mangrove extent 25  This total was split into 4,318 individual patches ranging in extent from 0.0005 to 6,517 km 2 . Within this maximal mangrove extent, approximately 40% (58,681 km 2 ) of the world's mangrove forest were confined to just 84 river deltas, with estuarine mangroves covering the next greatest area (n = 961 patches; 39,448 km 2 ). These two dominant types can form large individual extents of mangrove where accretion of fluvially transported terrigenous sediment 29,31,32 allows opportunistic colonization by mangroves [33][34][35] . Open coast mangroves covered an area of 30,586 km 2 and were by far the most numerous unit type (n = 2,639). Open coast mangroves were prevalent in areas with limited freshwater and terrigenous sediment inputs, such as the Middle East and the Pacific Islands 29 . Lagoons were largely restricted to high wave energy coasts; conditions that limit the potential mangrove establishment 27 . This combination of factors helps to explain the minimal global coverage of lagoonal mangroves (n = 634; 16,880 km 2 ).
In addition to geomorphic setting, the establishment and stability of mangrove forests are driven by sedimentary processes 20 . Sedimentary setting also determines the density of soil organic carbon stocks 13,15 and the optimal rehabilitation techniques 20 . We determined the sedimentary setting of mangrove typological patches based on the aquatic inorganic suspended particulate matter concentration and tidal amplitude of the site. Of the 145,595 km 2 combined GMW 1996, 2007, 2010 and 2016 mangrove extent, 14,657 km 2 (n = 1,023, 10.1%) was classified as carbonate. In these sediment-poor settings, including isolated oceanic islands in the Caribbean (Fig. 1) and the Pacific (e.g., Solomon Islands, northern Papua New Guinea, Micronesia), the Red Sea (Fig. 2b) www.nature.com/scientificreports/ and Sri Lanka (Fig. 3a), peat substrate is derived from autochthonous material 29,36 . These habitats appear particularly vulnerable to human disturbance including future sea-level rise 37 .
To examine spatial differences in the proportion of mangroves of different types, the global mangrove distribution was split into ten regions based on those identified in the World Atlas of Mangroves 11 . The proportion of deltaic mangroves was highest in West and Central Africa (56.5%) ( Table 5). These extensive deltaic mangrove areas form on highly dynamic coastlines that are subject to large inputs of terrigenous material. For instance, mudbanks of the deltaic coast of northern Brazil are rapidly prograding seaward, allowing colonization by mangrove vegetation 39 . Estuarine mangroves formed a large proportion of the mangroves of East Asia (82.0%) (Fig. 3a), Australia and New Zealand (57.9%) (Fig. 3b), and East and Southern Africa (45.6%) (Fig. 2c), with large individual patches in West Africa and Indonesia (Supplementary Table 5). Highly productive river-dominated coastal settings in West Africa and South America are home to some of the largest mangrove trees globally 10 . Conversely, in the xeric areas of the Middle East, there was an absence of estuarine mangroves (Fig. 2b), with mangrove stands characterised by low canopy heights and reduced aboveground biomass 10,40 . Open coast mangroves were more prevalent in Australia and New Zealand (36.6%) (Fig. 3b), the Middle East (69.4%) (Fig. 2b) and the Pacific Islands (42.4%), as well as there being large individual extents in Indonesia (Supplementary Table 5). Lagoonal mangroves are most common in the neotropics 27 and were largely confined in our typology to North and Central America and the Caribbean region ( Fig. 1), but also formed an important component of mangroves in the Middle East (26.9%; Fig. 2b).
Regional trends in mangrove loss by type. Over the period for which we have data on mangrove extent (1996-2016), we found that, by 2016, the total area of mangrove had been reduced to 135,870 km 2 (Table 2)   www.nature.com/scientificreports/ Lagoonal areas provide multiple ecosystem services, including tourism and fisheries enhancement 41 ; however, degradation of lagoonal environments is often linked to overexploitation of these services 42 .
Changes in area for deltaic and open coast mangroves were lower and similar to one another (− 4.3%), while estuarine mangroves experienced the smallest change in area (− 3.1%). Given that delta regions around the world support exceptionally high population densities 43,44 we expect that historic losses (prior to 1996) in deltaic mangroves through land conversion are likely to have been large. Anthropogenic impacts are also likely to disproportionately impact delta regions into the future 45 , with projected sea-level rise, upstream sediment capture by dams and subsidence increasing vulnerability to flooding 46 .
Our analysis of the sedimentary settings of different mangrove types indicated that losses of carbonate mangroves were more than double (− 8.1%) those of terrigenous areas (− 3.9%). These higher rates of change in types. Carbonate mangrove systems may be both more sensitive to natural disturbances such as cyclones, and to anthropogenic threats such as hydrological modification 47 . Disturbances have a longer-term negative impact on carbonate mangroves because they can cause rapid peat collapse and concomitant local increases in relative sea level 48 . Carbonate systems are also potentially more at risk from sea-level rise, as lower suspended sediment concentrations reduce minerogenic contributions to positive elevation change that could match sea-level rise 49 .
Rehabilitating organogenic carbonate mangrove systems requires techniques that restore and maintain surface elevation 20 , which are technically challenging (e.g. for marshes 50 ) and require monitoring and rapid intervention if restoration trajectories are not being maintained 51 . This analysis provides the first opportunity to identify these at-risk systems, which is important because avoiding peat collapse through mangrove protection is a far more efficient conservation action than attempting to implement technically demanding restoration options. Over the period 1996 to 2016, the patches that recorded the largest net losses in area (> 100 km 2 , n = 8) were deltaic and a single lagoon (Bahía de Chetumal, northern Belize and southeastern Mexico). Based on changes in the GMW dataset, the units with the largest losses were the Rakhine River Delta, Myanmar (316.2 km 2 ); the Mahakam Delta, Kalimantan, Indonesia (277.6 km 2 ); the Kayan Delta, Kalimantan, Indonesia (239.8 km 2 ); the deltaic coast of northern Brazil (170.1 km 2 ), and the Sesayap Delta, Kalimantan, Indonesia (147.4 km 2 ). Globally, the drivers of loss in deltaic mangroves vary spatially 52,53 . For instance, expansion of rice agriculture has been highlighted as the major factor in mangrove loss in Myanmar, whilst conversion to aquaculture is more prevalent in Kalimantan, Indonesia 6 and is also a proximate driver of mangrove deforestation across Latin America and the Caribbean 54 . In addition, shoreline erosion can contribute a significant amount of mangrove loss in deltas 53,55 . Quantifying ecosystem structure using the biophysical mangrove typology. The biophysical typology can also contribute to assessing the potential ecosystem structure of an area. Inorganic suspended particulate matter concentration and sediment delivery, aboveground biomass (AGB), tidal amplitude, river dominance, precipitation and substrate composition all influence the structure, species composition and health of mangrove stands and therefore the goods and services they provide. This analysis is the first to attribute global AGB to mangrove-specific types and to investigate the likely role of mangrove type on ecosystem structure. Significant differences between mangrove types were detected (F 3,3771 = 85.65, P < 0.0001); however, the Nagelkerke pseudo-R 2 = 0.059, suggested low model explanatory power. Post-hoc analysis revealed significant differences  (Fig. 4). The variation in the data around these averages is high, because typologies span climatic and precipitation gradients, which also influence mangrove biomass 56 . This supports previous plot-scale studies that have shown that estuarine/deltaic mangroves store more biomass and soil carbon than open coast mangroves 16 , and suggests that such patterns exist at multiple scales. This same pattern is not so clearly reflected in mangrove soil carbon, where, under the influence of high minerogenic sediment loads, estuaries and deltas have a much lower percentage of soil carbon per unit volume of soil compared to carbonate or lagoonal settings 13,15 . This is consistent with lower levels of biomass allocation to belowground root material and higher rates of decomposition in deltaic minerogenic settings, which have higher levels of nutrient availability compared to those in carbonate settings 57 . conclusions Applications of a global biophysical mangrove typology in ecosystem services and restoration. In this study we extend the utility of the presence/absence time series of mangrove extent by assigning mangroves into discrete types based on their geomorphic and sedimentary setting. The wider landscape context of a mangrove forest is important for identifying the drivers of ecosystem degradation and loss, determining the  20,58 , and assessing the delivery of ecosystem functions and services 12,13,15 . The biophysical typology can also help with projecting the impact of climate change and sea level rise on mangroves 19 , as geomorphic setting determines the boundary conditions affecting mangrove surfaces, and sedimentary settings determines the processes by which mangroves can increase their surface elevations to potentially keep pace with rising seas. This global mangrove biophysical typology therefore has the potential to play a significant role in understanding spatial variability in mangrove threats, ecosystem functions and service values and restoration potential.

Methods
Geomorphic setting. We first identified geomorphic features (deltas, estuaries, lagoons, bays) within the mangrove regions of the world using a high resolution coastline, and then determined which mangrove patches were associated with each feature. The first step was therefore to identify coastlines containing either deltas, estuaries, lagoons, bays, or indeed none of these coastal features. Open coast mangroves are areas associated with bays, or no coastal embayment. The other mangrove types were associated with their respective coastal feature. Classifying coastal embayment polygons. Delta CEPs were identified using two procedures. Firstly, deltas (n = 81) in mangrove areas were identified from the World Atlas of Mangroves 11 , The Major River Deltas Of The World 59 and Major World Deltas: A Perspective From Space 60 . Secondly, CEPs were assessed based on the number of drainage outlets to the ocean. Those with more than two outlets were identified and visually assessed. CEPs were classified as deltas based on polygon shape, having a large catchment area with multiple river flowlines (distributaries), and an internet search identifying reference to the river having a delta (n = 21). Delta extents were created using either those already derived in the Deltas at Risk dataset https ://www.globa ldelt arisk .net/data.html or manually using online sources and Google Earth (Google Earth Pro version 7.3.3.7699, https ://www.googl e.com/earth /). The delta extents were used to combine multiple CEPs into a single unit (further details given in Supplementary Section 2.2). www.nature.com/scientificreports/ Delta CEPs and CEPs identified visually as errors were removed before a random forest classifier was used to assign the remaining CEPs into three types ('bays' , 'estuaries' and 'lagoons'). The random forest classification was based on ten variables describing the shape of the polygons, their associated upstream hydrological catchment and the amount of precipitation entering the catchment (Supplementary Table 1). The hydrological catchment data were accessed from the (HydroSHEDS) dataset (https ://www.hydro sheds .org/). We identified HydroSHEDS river network flowlines that intersected with the CEPs, and the HydroSHEDS watershed polygons that intersected with these selected flowlines were selected and aggregated to form a single catchment extent (further details given in Supplementary Section 2.3). The amount of precipitation moving through the river network to each CEP was of the form of monthly precipitation and accessed from https ://www.earth env.org/ strea ms. The precipitation data was developed to fit alongside the HydroSHEDS framework 61 (further details given in Supplementary Section 2.3).
The random forest (randomForest package 62 ) analysis using 100,000 trees was initially run on a CEP training dataset containing 800 bays, 71 lagoons and 300 estuaries (total n = 1,171) in R (version 3.4.4 63 ), with 20% of the data randomly selected for model validation. All other parameters were left as the default. Selection of the CEPs for the training dataset was undertaken by expert annotation and was not randomised. Instead 100 bays in each of the following mangrove regions were included: North America, South America, West Africa, Southeast Africa, Middle East, Asia, Australasia, the Pacific. Estuary and lagoon CEPs were visually identified using a global typology of nearshore coastal systems 64 , from 'tidal systems' or 'lagoon' coastal types respectively.
The resulting random forest model was fitted to the remaining CEP dataset. A random sample of 500 bays and all estuary and lagoon CEPs (n = 1,271) were visually inspected at a 1:500,000 scale in ArcGIS (ArcGIS Desktop version 10.6, https ://deskt op.arcgi s.com/en/) to assess the accuracy of the model and correct misclassifications. Visual assessment was based on the size and shape of each CEP, the river catchment inputs to each feature and the wider geographical context (further details given in Supplementary Section 2.4).
Given misclassifications from the initial random forest model, the process was repeated with a further 75,000 trees on the non-visually assessed bay CEPs using the original 1,171 training points and the visually inspected and corrected CEPs from the first random forest model. The second random forest iteration was then fitted onto these remaining bay CEPs. If there was a disagreement in the predicted type between the two random forest models, the CEP was visually assessed and, where necessary, corrected (results of the Random Forest given in Supplementary Section 2.5 and limitations of the methodology in Supplementary Section 4).
Attributing mangroves to the biophysical typology. We then determined which mangrove patches were associated with the classified CEP. The mangrove extent used as a framework for the biophysical typology was the union of the GMW 1996, 2007, 2010 and 2016 maps, with mangrove patches classified into one of four types: deltaic, lagoonal, estuarine or open coast. Assigning the mangrove patches to a type and an individual CEP followed a stepwise procedure (see Supplementary Figs. 2, 3). While there were many steps, they can be broadly classified into three aims: firstly; ensuring that all existing mangrove patches could be assigned to a single CEP by splitting very large mangrove patches where appropriate or those mangrove patches intersecting two or more CEPs; secondly allocating patches that directly intersected to a single CEP and finally; assigning mangrove patches that did not directly intersect with a CEP to the appropriate typological unit using HydroSHEDS catchment boundaries and distance between CEPs and mangrove patches. Following the stepwise procedure, several rounds of visual quality assessment and corrections were carried out (further details given in Supplementary Section 2.6).

Sedimentary setting.
For the non-deltaic and estuarine patches we further sought to determine the sedimentary setting. Following Balke and Friess 20 we determined the sedimentary setting of the mangrove typological patches based on the aquatic inorganic suspended particulate matter concentration and tidal amplitude of the site. Two hundred and forty monthly inorganic suspended particulate matter concentration (g/m 3 ) global data rasters were downloaded from the Globcolor website https ://www.globc olour .info and the mean inorganic suspended particulate matter concentration for each pixel was calculated. A tidal data raster from the Finite Element Solution tide model, FES2014, was downloaded from AVISO + products (https ://www.aviso .altim etry.fr) (further details on the data sources given in Supplementary Section 3).
Training points were taken from 152 locations with known typological (riverine or non-riverine) and sedimentary (terrigenous or carbonate) status and within 10 km of the GMW maximum extent. These were identified through reference to the literature or the authors' own diverse field experiences. The tidal amplitude value and mean inorganic suspended particulate matter concentration nearest to the training location was determined in ArcGIS (ArcGIS Desktop version 10.6, https ://deskt op.arcgi s.com/en/) and then imported into R (version 3.6.2 63 ).
Estuarine or deltaic sites (n = 70) were removed from the data set and a two-sided binomial generalized linear model with a logit link was fitted to the remaining 82 sites in R (version 3.6.2 63 ), with the resulting model being used to classify the lagoonal and open coast mangroves as either carbonate or terrigenous. The pseudo-R 2 was calculated as the (null deviance -residual deviance)/ null deviance and was 46.8%. M 2 tidal amplitude was a significant predictor (z = 4.5, P = < 0.001); however mean inorganic suspended particulate matter concentration was a non-significant predictor (z = 0.95, P = 0.34), but was retained in the model. The model misclassified six (of n = 29) carbonate and nine (of n = 53) terrigenous sites. The resulting model was then mapped in ArcGIS to determine whether lagoonal and open coast mangrove patches were in a terrigenous or carbonate setting. Estuarine and deltaic mangroves were universally classed as terrigenous.
Mangrove above ground biomass. Global data on the AGB of mangroves (Mg ha −1 ) was downloaded from (https ://daac.ornl.gov/) 65  www.nature.com/scientificreports/ height combined with field measurements 10 . The AGB values were developed for the year 2000 using the Giri and colleagues 23 global mangrove distribution dataset. This resulted in a mismatch with our biophysical mangrove typology, which was based on the combined area of the GMW 1996, 2007, 2010 and 2016 timesteps. Therefore, the AGB raster dataset was converted to points that were then spatially joined to our biophysical typology dataset and the mean AGB value in each typological unit calculated (deltaic type n = 84, estuarine type n = 907, lagoonal type = 591, and open coast type = 2,193).
To determine whether there was a significant difference in the AGB between the mangrove types, a two-sided generalized least squares model was developed using the 'nlme' package in R 66 . Validation of the initial model, undertaken by creating histograms of the normalized residuals and plotting the normalized residuals against the fitted values and the covariate 67,68 , suggested issues with non-normality and heteroscedasticity. Therefore, the square root of the mean AGB was used and a variance structure for mangrove type was included 69 , with a clear improvement in residual validation plots. Post-hoc tests on the difference between the estimated marginal means of each mangrove type were computed using the 'emmeans' 70 . The Nagelkerke pseudo-R 2 was calculated using 'rcompanion' 71 .

Data availability
The global biophysical mangrove typology is available for download from the Ocean Data Viewer (https ://data. unep-wcmc.org/).