Mangroves are among Earth's most threatened ecosystems. In order to monitor them and make informed policy decisions, national and global maps of mangrove extent, change, and carbon (including AGB) are critical. The accuracy of these datasets is of great importance as they are used to shape local policies.
Multiple global mangrove datasets have become available within the past decade, with multiple maps for mangrove extent, change, biomass, and carbon. Global datasets are often derived from a combination of existing datasets. However, each dataset includes a quantity of error.
Limitations of datasets are often not well communicated for end-users and are sometimes ignored when ingested into new data products. It is therefore critical to review existing datasets and provide transparency on the strengths and weaknesses of mangrove mapping.
We conducted a review of the current literature and selected the primary global scale mangrove forest datasets that map mangrove extent/cover, change in extent/cover, structure (height and biomass) and carbon (above and below ground). We synthesize the selected datasets below:
World Atlas of Mangroves, Spalding et al 2010
The World Atlas of Mangroves (WAM) was the first to define the global distribution of mangroves in 1997. This was subsequently updated in 2010 and a digital version of the dataset was released. It was produced from a joint initiative which included the International Tropical Timber Organization (ITTO), International Society for Mangrove Ecosystems (ISME), Food and Agriculture Organization of the United Nations (FAO), UN Environment World Conservation Monitoring Centre (UNEP-WCMC), United Nations Educational, Scientific and Cultural Organization’s Man and the Biosphere Programme (UNESCO-MAB), United Nations University Institute for Water, Environment and Health (UNU-INWEH) and The Nature Conservancy (TNC). The WAM was composed of a range of existing mangrove datasets including national and regional scale information covering a period of 1999-2003. Classification of Earth observation data was conducted to map some regions, using unsupervised classification algorithms. Quantitative accuracy assessments were not conducted but a qualitative assessment of quality was undertaken. In addition, mangroves were also digitized using manual interpretation to complete the mapping, with remaining gaps filled using maps from Giri et al (2011). While this provided the first atlas of mangrove distribution, there is a distinct lack of quality assessment other than manual qualitative reviewing. As imagery was not cloud masked prior to being classified using comparatively naive algorithms as compared to the current standard, errors in mapping are surely expected. Without a robust assessment of dataset accuracy, the quantity and distribution of this error is unknown. The dataset is available on the United Nations Environment Program World Conservation Monitoring Center (UNEP-WCMC) Ocean Data Viewer: https://data.unep-wcmc.org/datasets/5.
Status and Distribution of Mangrove Forests of the World using Earth Observation Satellite Data, Giri et al 2011
The Mangrove Forests of the World (MFW) dataset was the first to be generated entirely from satellite imagery, mapping mangrove forest extent at 30 m resolution across 18 countries. A total mangrove extent of 137,760 km2 was mapped for the nominal year 2000. The Landsat data used spanned the period 1997-2000 and was composed of over 1000 scenes that make up the Global Land Survey (GLS) archive, supplemented by imagery from the USGS archive if GLS scenes were degraded by cloud cover. Imagery was corrected to Top-of-Atmosphere (TOA) and classified using a combination of supervised and unsupervised approaches. Imagery was subset to what is now commonly termed a ‘mangrove habitat zone’, whereby data where mangroves are not expected to occur are excluded, leaving only low-lying coastal areas and inter-tidal zones. Water was mapped using a supervised classification and the land pixels were classified using an unsupervised ISODATA clustering algorithm, generating 50-150 clusters at the 99% convergence level. These clusters were then manually assigned to a landcover class using reference to field data and high-resolution satellite imagery. The accuracy of the map was not assessed quantitatively but a qualitative quality assessment was carried out and obvious errors were removed. Cloud cover inhibited the classification of some data and mangroves have been omitted or poorly mapped in some regions. Imagery was geo-corrected prior to classification to an RMSE of half a pixel but substantial geo-location error is prominent in the dataset across whole continents. Furthermore, a raster and vector dataset is available but significantly vary in extent., with the vector dataset containing 22,326 km2 more mangrove. The MFW dataset has been the most widely used mangrove forest dataset and has formed the base map of a wide range of successive products. These errors and uncertainties will propagate through the following products, unless accounted for.
Global Mangrove Watch (GMW), Bunting et al 2018
The Global Mangrove Watch (GMW) is an international collaboration originally founded under the Japan Aerospace Exploration Agency (JAXA) Kyoto and Carbon (K&C) initiative. The consortium is composed of members from Aberystwyth University (U.K.), sols Earth Observation (soloEO; Japan), Wetlands International, the Institute for International Water management (IWMI), the World Conservation Monitoring Centre (UNEP-WCMC). The consortium has been supported by JAXA through the provision of L-band Synthetic Aperture Radar (SAR) data from JERS-1, ALOS PALSAR and ALOS-2 PALSAR-2. The GMW maps global mangrove extent for the nominal year 2010, mapping 137, 600 km2 of mangrove extent, a value almost identical to Giri et al 2011 despite a decade between datasets. The GMW do this by classifying a combination of Landsat 5 and 7 imagery in combination with JAXA Advanced Land Observation Satellite (ALOS) Phased Array L-band Synthetic Aperture Radar (PALSAR) imagery. As in Giri et al (2011) a mangrove habitat region is defined based upon measures such as elevation and distance to water. An initial mangrove area was classified using the ALOS PALSAR radar imagery but due to an inability to accurately characterize the mangrove boundary, optical data was used to refine the mangrove extent. These two classifications steps utilized an extremely randomized tree classifier. The results underwent a manual quality assessment where major issues were remedied and a thorough accuracy assessment was conducted using in excess of 50,000 stratified random points. The accuracy of the mangrove map was 93.6-94.5%. The Landsat imagery is a source of error in the GMW maps, with cloud cover and Landsat SLC-off errors contributing to poor quality mosaics in some regions, particularly in West Africa. The GMW is mapped at 25 m pixel resolution but recommends using the data at a 1 ha minimum mapping unit due to difficulties in mapping fine mangrove fringes. In addition to a baseline extent, the GMW has also produced annual changes in mangrove extent for years that coincide with available JAXA radar data, namely 1996, 2007, 2008, 2009, 2014, 2015, 2016 and 2017, with 2018 expected in the future. These datasets are also available at the baseline resolution of 25 m and are publicly available, but are not yet published. A new baseline is released for each period by using a ‘map-to-image’ method.
Continuous Global Mangrove Forest Cover for 21st Century (CGMFC–21), Hamilton and Casey (2016)
The CGMFC-21 database attempts to map changes in mangrove cover from 2000-2012 by using an intersection of currently available global forest products. Primarily, the Giri et al (2011) mangrove baseline is combined with the Hansen et al (2013) Global Forest Cover (GFC) dataset. This latter dataset maps global mangrove forest cover, whereby each pixel is represented by a canopy cover percent value which is mapped annually. The GFC for 2000 was intersected with the Giri et al (2011) mangrove map to achieve mangrove cover per pixel. This cover percentage was multiplied by the mangrove pixel area, which varies with latitude, to derive an estimation of global mangrove forest area that accounts for a variation in mangrove extent associated with mangrove structure. This process was repeated for every year of the GFC, so that losses and gains in mangrove cover were mapped annually over a decade. This dataset, unlike other presence/absence maps has the benefit of accounting for mangrove canopy cover so that pixels that contain 100% canopy cover or 50% canopy cover are not treated equally. This allows the map to account for the variation that occurs in mangrove cover where tall mangrove forests with high canopy covers can thrive adjacent to short shrub mangroves with low canopy cover. The CGMFC-21 dataset estimated a total mangrove area of 83,495 km2, which is far below the similar estimates of Giri et al (2011) and Bunting et al (2018). The CGMFC-21 datasets mapped a loss of 1646 km2, equitable to 1.97% over the observation period (0.16% per year) with regional maximums of 3.11% in Indonesia. A second change analysis was carried out confined to the larger mangrove biome, defined by the Terrestrial Ecoregions of the World (TEOW) mangrove definition, in which a gain in mangrove extent could also be mapped. A disadvantage of mapping mangrove cover is that it is a planimetric estimate of cover which may not accurately represent the true cover of the mangrove forest within the pixel. Furthermore, within the true mangrove extent, changes in mangrove cover are excluded if they exist outside of the Giri et al (2011) baseline, while changes within the mangrove biome may not be mangrove forest pixels. Gains in mangrove extent are also only available within the TEOW extent, which may also be misclassified due to terrestrial forest pixels.
Tropical and Subtropical Wetlands Distribution version2, Gumbricht et al 2017
The Gumbricht et al 2017 map, known as the SWAMP CIFOR global wetlands map is not accompanied by a specific publication, but is linked to a publication by Gumbricht et al., 2017. It is assumed that this methodology is applicable to the SWAMP CIFOR map. In this map mangroves are represented by a class among a number of additional wetland classes, including but not limited to fens, marshes and open water. The classification of the wetland classes are derived from expert rules that are applied to three primary variables; 1) Interannual Water Balance – wetland Topographic Convergence Indices (wTCI): wTCI is a hydrological model that simulates surface run-off, groundwater flow and flooding volumes 2) Soil Wetness Phenology – Transformed Wetness Index (TWI) : TWI is an algorithm which estimates surface soil moisture content from optical satellite data, specifically MODIS in this study, including the determination of periods of inundation and water saturation and periods of soil water content that exceed specified thresholds 3) Hydro-geomorphical maps and indices. These included generated maps of landscape geomorphological elements, which included plains, valleys and ridges. This was composed of existing geomorphological maps combined with the wTCI hydrological model. MODIS imagery was collected for the year 2011, spanning 2012-2012 where required to fill gaps omitted by cloud cover. Additional evapotranspiraton data (New et al 2002), precipitation data (Hijmans et al 2005) and 250 m hydrologically corrected SRTM data (Brown, Sarbandi and Pierce, 2005) were used. To define the mangrove class, rules were put on these data ad models, including permanently wet but allowing for tidal variation in the soil wetness data derived from MODIS and occurrence of pixels within 5 km of the sea or estuary, not including channels nor peaks at a maximum elevation of 45 m.
Mangrove Soil C
Global patterns in mangrove soil carbon stocks and losses, Atwood et al 2017
Atwood et al (2017) compiled national country estimates of soil carbon by conducting a large grey-literature search to collect a large body of data on mangrove soil C stocks. The soil C data were scored (values 0-3) across seven criteria. Five of these were individual point data metrics that included quality of C stock data, quality of down-core stock data to 1m, quality of percentage of organic carbon data, quality of bulk density data and quality of the publication. Two additional global data set metrics included extent of genera covered and extent of marine eco-regions covered. All soil C samples were recorded or extrapolated to 1 m depth. To generate country level statistics, soil C data was averaged within a country and then multiplied by the area of mangrove forest in 2014 calculated by Hamilton and Casey (2016). While this does provide an up to date baseline from which to calculate mangrove soil C, the area estimate calculated by Hamilton and Casey (2016) is an estimate of planimetric mangrove cover which cannot be attributed to underlying soil area. Furthermore, it is unknown if soil C density varies with mangrove cover, particularly at the global scale. In this instance, the soil C stock represents only that of the mangrove cover and not the extent of the mangrove ecosystem and is subsequently assumed to be underestimated. In the 57 mangrove holding countries where no soil C data could be found, a global average of 283 +/- 194 Mg C ha-1. In addition, the authors try to estimate CO2e emissions by compiling data on the effects of different types of disturbance on soil C stocks or C content, specifically the change in mangrove Soil C after disturbance. The authors found no significant difference between the soil C loss based on the driver of disturbance and used the average loss of C (43%) stocks down to 1 m for each country. This was combined with annual mangrove loss statistics per country derived from Hamilton and Casey (2016). The authors state that it is unlikely that 43% of soil C will be lost in each case but their estimate follows IPPC protocols that all C is emitted as CO2e during the year of the land conversion (IPPC coastal wetlands supplement).
A global predictive model of carbon in mangrove soils, Jardine and Siikamaki 2014
Jardine and Siikamaki (2014) used over 900 soil carbon measurements, from 28 countries collected by 61 independent studies, to develop a global predictive model for mangrove soil carbon. Using climatological and locational data as predictors a random forest machine-learning algorithm was used to predict mangrove biomass globally. This results in a global map of soil C with a spatial resolution of approximately 1 km (5 arc minutes), totalling a global mangrove soil carbon stock of 5.00 ± 0.94 Pg C (assuming a 1 meter soil depth). The soil C dataset was generated using predictive variables that explain the carbon concentration in mangrove soils, including the distance of the observation’s sampling location from the equator, several variables describing climate conditions at the sampling locations, and regional indicators. Mean annual temperature (Bioclim 1), mean temperature in the coldest quarter (Bioclim 11), total annual precipitation (Bioclim 12), and seasonality in precipitation (Bioclim 15) werethe climatic variables used to build the model. Regional indicators were drwan from the ten biogeographic regions for mangroves developed by Spalding et al (2010). The amount of carbon per hectare in the world’s most carbon-rich mangroves (approximately 703 ± 38 Mg C ha−1) is roughly a 2.6 ± 0.14 times the amount of carbon per hectare in the world’s most carbon-poor mangroves (approximately 272 ± 49 Mg C ha−1). Considerable within country variation in mangrove soil carbon also exists. In Indonesia, the country with the largest mangrove soil carbon stock, we estimate that the most carbon-rich mangroves contain 1.5 ± 0.12 times as much carbon per hectare as the most carbon-poor mangroves.
A global map of mangrove forest soil carbon at 30 m spatial resolution, Sanderman et al 2018
The authors developed a machine learning based model to derive mangrove soil C stocks at 30 m resolution, the highest resolution available for mangrove soil C data. A harmonized globally representative database of mangrove soil C data was collected from grey-literature and contributions of unpublished data. A machine learning-based model of organic carbon density (OCD) which models OCD as a function of depth, an initial estimate of organic carbon stock (OCS) from global SoilGrids 250 m (Hengl et al (2017)) and 20 covariate layers that include: vegetation characteristics (forest percent cover and four band Landsat imagery from Hansen et al (2013)), Shuttle Elevation Topography Mission (SRTM) elevation data, sea surface temperature (SST), tidal elevation amplitude, averaged (2003-2011) monthly total suspended matter and a mangrove typology map. OCD was spatially modeled in three dimensions, using all soil horizons at varying depths which enabled a single model to derive OCD at any required depth. OCD was predicted at depths of 0, 30, 100 and 200 cm using a random forest machine learning model to derive soil C stocks by summing the OCS to depths of 1 and 2 m for every 30 m pixel within Giri et al 2011.Sanderman et al were able to model the variation in mangrove soil C when compared to reference data using a Random Forest model. The initial estimate of organic carbon stock (OCS) 250 m SoilGrids data was the most important variable in the model and improved model performance by approximately 50%. Soil depth, seasonal total suspended matter (TSM1, TSM2, TSM3) and tree cover completed the top 5 most important variables. The projection of the model to mangrove holding nations revealed that Bangladesh had the lowest per ha soil C stocks (127 Mg C ha-1). The highest per ha soil C stocks were reported to be in the pacific islands (505 Mg C ha-1) but the exact location is not reported. An improvement from this work on other soil C maps is the high spatial resolution which is capable of mapping the variation within mangrove forests and not only between mangroves that are at largely different latitudes. Global Forest Change data from Hansen et al (2013) were used to quantify the loss of mangrove soil C due to mangrove forest loss, but this dynamic component of the soil C stocks were not replicated or reviewed in this study.
Global controls on carbon storage in mangrove soils, Rovai et al., 2018
Rovai et al. investigate the influence that coastal environmental settings (CES) have on the global variability in mangrove soil C:N:P stoichiometry and soil organic carbon (SOC) stocks. Using data from 36 sites across the neotropics, the sites were categorized into coastal environmental settings, using an independent classification. These classes included deltas (river dominated), estuaries (tide dominated), lagoons (wave dominated), composite (river and wave dominated) bedrock (drowned bedrock valley) and carbonate settings. The study was confined to the neotropics due to a lack of soil data from old world mangroves, although the authors maintain that their in situ data represented the CES that mangroves were able to thrive in and the range of environmental conditions that influence them. Climate variables from the WorldClim database (1950-2000) In order to test the influence of climatic and geophysical drivers on mangrove nutrient density and SOC stocks. The climatic variables used included minimum temperature of the coldest month, and minimum precipitation of the driest month, which represent extreme or limiting environmental factors. In addition, PET data was derived from the Moderate Resolution Imaging Spectrometer (2000-2012), tidal range was calculated from the tidal atlas of finite element solutions and global river discharge was extracted from the Global Runoff Data center archive. Using the network of soil core data, a wide range of statistical tests and multivariate analysis were used to determine that mangrove SOC, total nitrogen and total phosphorus varied significantly across CES within mangrove forests. Using this knowledge, the authors used the CES framework to fit climate-geophysical models to estimate SOC density in all mangroves globally, using an expanded set of field sampled data from using a literature search. A linear equation using tidal range and minimum temperature, which accounted for 83% and 17% of the relative contributions to the global variability explained by the model, respectively. Rovai et al. predicted global mangrove soil C at a resolution of 0.25o, estimating a total global budget of 2.26 Pg, using the mangrove forest area of Hamilton and Casey (2016) within each grid cell. Beyond a global prediction of mangrove soil C, the authors were able to derive that soil C is underestimated in carbonate settings by up to 50% and overestimated in deltaic coastlines by up to 86%.
Predicting Global Patterns in Mangrove forest Biomass, Hutchinson et al 2013
Hutchinson et al. provided the first geospatial estimate of mangrove forest biomass, derived from a climatic model. This followed earlier work by (Twilley) that used latitude-based models but advanced on this by building a model using climatic variables and providing a global estimate of AGB as a global raster, with a resolution of approximately 1km. Estimates of field data were derived from a literature search provided that the location of the sample could be located to within 0.01 degrees latitude and longitude using coordinates, published maps or place names. AGB estimates within 10 km of one another were averaged resulting in 52 points. For each point temperature and precipitation information was extracted from the WorldClim Bioclim 30 arc-second dataset, forming two sets of variables from which to derive two separate models. Mean temperature and total precipitation, standard deviation of monthly temperature and coefficient of variation of monthly precipitation composed the first model. The mean temperature of the warmest and coldest quarters and the precipitation of the wettest and driest quarters composed the second model. Fir each set of variables, linear models were fitted with AGB as the response and the second model outperformed the first, based on the AIC and R2. The global model predicts a total global of 2.83 Pg with a global mean of 184.8 Mg/ha. The global biomass data created was for the whole globe and the mangrove map of Spalding was used to calculate the mangrove specific values. Hutchinson et al. state that their map is not appropriate for use at the local level due to its dependence on global model which have little relevance at the local level.
Mangrove canopy height globally related to precipitation, temperature and cyclone frequency, Simard et al 2019
Simard et al. used a 30 m global DEM to provide the first direct estimate of mangrove height, from which biomass was estimated using regionally calibrated allometric equations. Shuttle Radar Topography Mission (SRTM) data (30 m) was extracted for all mangrove regions using the map of Giri et al (2011). Pixels identified as mangrove but with an elevation value greater than 55 m were set at the maximum of 55 m. To calibrate the elevation values, GLAS lidar altimetry data were collected between 2003 and 2009 in order to remove the bias introduced by partial penetration of the SRTM radar into the canopy. Low-quality measurements were omitted, leaving 57,369 calibration points using the lidar relative height of the 100% percentile of the waveform (RH100). Mangrove basel area weighted height was also calculated which models height as a function of basel area, yielding a lower r.m.s.e than maximum height. To derive mangrove biomass, 331 field plots were acquired across Asia, Africa and the Americas in order to derive stand-level allometry between AGB and basel area weighted height (Hba) for individual regions including Americas, East Africa and South Asia. An additional global model was derived for countries outside of these regions where a published algorithm was not available. The height to biomass equation was applied to each calibrated elevation pixel, providing the first global map of mangrove biomass derived from a direct measurement of mangrove forest structure. Simard et al. estimate total global AGB to be 1.75 Pg with a mean AGB density of 129.2 Mg/ha, with large variation both between and within mangrove stands.