LETTER OPEN ACCESS … global... · titleofthework,journal citationandDOI. E-mail:[email protected] ... (Atwood et al 2017, Jardine and Siikam¨aki 2014), do not capture enough

Environmental Research Letters

LETTER bull OPEN ACCESS

A global map of mangrove forest soil carbon at 30 m spatial resolutionTo cite this article Jonathan Sanderman et al 2018 Environ Res Lett 13 055002

View the article online for updates and enhancements

This content was downloaded from IP address 201199232124 on 30042018 at 1508

Environ Res Lett 13 (2018) 055002 httpsdoiorg1010881748-9326aabe1c

LETTER

A global map of mangrove forest soil carbon at 30mspatial resolution

Jonathan Sanderman121 Tomislav Hengl2 Greg Fiske1 Kylen Solvik1 Maria Fernanda Adame3 LisaBenson56 Jacob J Bukoski7 Paul Carnell8 Miguel Cifuentes-Jara9 Daniel Donato19 Clare Duncan48Ebrahem M Eid1020 Philine zu Ermgassen1718 Carolyn J Ewers Lewis8 Peter I Macreadie8 Leah Glass5Selena Gress11 Sunny L Jardine12 Trevor G Jones513 Eugene Ndemem Nsombo14 Md Mizanur Rahman15Christian J Sanders16 Mark Spalding17 and Emily Landis17

1 Woods Hole Research Center 149 Woods Hole Road Falmouth MA 02540 United States of America2 ISRIC mdash World Soil Information Wageningen The Netherlands3 Australian Rivers Institute Griffith University Nathan QLD Australia4 Institute of Zoology Zoological Society of London Outer Circle Regentrsquos Park London NW1 4RY United Kingdom5 Blue Ventures Conservation London United Kingdom6 Centre for Environment Fisheries and Aquaculture Science Lowestoft United Kingdom7 Department of Environmental Science Policy and Management University of California Berkeley CA United States of America8 Centre for Integrative Ecology School of Life and Environmental Sciences Deakin University Burwood VIC Australia9 Forests Biodiversity and Climate Change Program CATIE 7170 Turrialba Costa Rica10 Botany Department Faculty of Science Kafr El-Sheikh University Kafr El-Sheikh Egypt11 School of Applied Sciences Edinburgh Napier University Edinburgh Scotland12 School of Marine and Environmental Affairs University of Washington Seattle WA United States of America13 Department of Forest Resources Management University of British Columbia Vancouver BC Canada14 Institute of Fisheries and Aquatic Sciences University of Douala Doula Cameroon15 Graduate School of Agriculture Kyoto University Kyoto Japan16 National Marine Science Centre Southern Cross University Coffs Harbour NSW Australia17 The Nature Conservancy Arlington VA United States of America18 School of Geosciences University of Edinburgh Edinburgh United Kingdom19 Washington State Department of Natural Resources Olympia WA United States of America20 Biology Department College of Science King Khalid University Abha 61321 Saudi Arabia21 Author to whom any correspondence should be addressed

OPEN ACCESS

RECEIVED

29 June 2017

REVISED

10 April 2018

ACCEPTED FOR PUBLICATION

13 April 2018

PUBLISHED

30 April 2018

Original content fromthis work may be usedunder the terms of theCreative CommonsAttribution 30 licence

Any further distributionof this work mustmaintain attribution tothe author(s) and thetitle of the work journalcitation and DOI

E-mail jsandermanwhrcorg

Keywords blue carbon carbon sequestration land use change machine learning

Supplementary material for this article is available online

AbstractWith the growing recognition that effective action on climate change will require a combination ofemissions reductions and carbon sequestration protecting enhancing and restoring natural carbonsinks have become political priorities Mangrove forests are considered some of the mostcarbon-dense ecosystems in the world with most of the carbon stored in the soil In order formangrove forests to be included in climate mitigation efforts knowledge of the spatial distribution ofmangrove soil carbon stocks are critical Current global estimates do not capture enough of the finerscale variability that would be required to inform local decisions on siting protection and restorationprojects To close this knowledge gap we have compiled a large georeferenced database of mangrovesoil carbon measurements and developed a novel machine-learning based statistical model of thedistribution of carbon density using spatially comprehensive data at a 30 m resolution This modelwhich included a prior estimate of soil carbon from the global SoilGrids 250 m model was able tocapture 63 of the vertical and horizontal variability in soil organic carbon density (RMSE of109 kg mminus3) Of the local variables total suspended sediment load and Landsat imagery were themost important variable explaining soil carbon density Projecting this model across the globalmangrove forest distribution for the year 2000 yielded an estimate of 64 Pg C for the top meter of soilwith an 86ndash729 Mg C haminus1 range across all pixels By utilizing remotely-sensed mangrove forest coverchange data loss of soil carbon due to mangrove habitat loss between 2000 and 2015 was 30ndash122 Tg Cwith gt75 of this loss attributable to Indonesia Malaysia and Myanmar The resulting map products

copy 2018 The Author(s) Published by IOP Publishing Ltd

Environ Res Lett 13 (2018) 055002

from this work are intended to serve nations seeking to include mangrove habitats in payment-for-ecosystem services projects and in designing effective mangrove conservation strategies

1 Introduction

Mangrove forests occupying less than 14 million ha(Giri et al 2011) just 25 of the size of the Ama-zon rainforest provide a broad array of ecosystemservices (Barbier et al 2011) Mangroves are criticalnursery habitats for fish birds and marine mammals(Mumby et al 2004 Nagelkerken et al 2008) act aseffective nutrient filters (Robertson and Phillips 1995)buffer coastal communities from storm surges (Gedanet al 2011) and support numerous rural economies(Spalding et al 2014 Temmerman et al 2013) Theseecosystem service benefits have been valued at an aver-age of 4200 US$ haminus1 yrminus1 in Southeast Asia (Branderet al 2012) Despite these ecosystem service bene-fits mangroves are highly threatened by both urbanexpansion and other lsquohigher valuersquo land uses becauseof their close proximity to major human settlementsThere are no reliable estimates of original mangrovecover but some authors have suggested that 35 ormore of original cover may have been lost and widerareas have been degraded (Valiela et al 2001 Spald-ing et al 2010) Loss rates have slowed dramaticallyin the past 10ndash20 years in most areas however theyremain considerable with rates up to 31 annu-ally in some countries (Hamilton and Casey 2016)The major drivers of loss are conversion for aquacul-ture especially shrimp farming agriculture and urbandevelopment (Alongi 2002 Valiela et al 2001 Spald-ing et al 2010 Richards and Friess 2016) but loss dueto extreme climatic events are also becoming morecommon (Duke et al 2017)

With the growing recognition that effective actionon climate change will require a combination ofemissions reductions and removals (Rockstrom et al2017) protecting enhancing and restoring natural car-bon sinks have become political priorities (Boucheret al 2016 Grassi et al 2017) Mangrove forests canplay an important role in carbon removals in additionto being some of the most carbon-dense ecosystemsin the world (Donato et al 2011) if kept undisturbedmangrove forest soils act as long-term carbon sinks(Breithaupt et al 2012) As such there is strong interestin developing policy tools to protect and restore man-groves through payment for ecosystem services (Friesset al 2016 Howard et al 2017)

Mangroves can store significant amounts of car-bon in their biomass (Hutchison et al 2014) howeverthe vast majority of the ecosystem carbon storage istypically found in the soil (Donato et al 2011 Mur-diyarso et al 2015 Sanders et al 2016) For exampleKauffman et al (2014) found that within the same estu-ary soil carbon contributed 78 of total ecosystem

carbon storage in tall mangroves but 96ndash99 oftotal ecosystem carbon in medium and low staturemangrove stands Importantly Kauffman et al (2014)found that conversion of these mangrove forests toshrimp ponds resulted in the loss of 90 of this car-bon from the top 3 m of soil (612ndash1036 Mg C haminus1)In addition to avoided emissions many mangrove for-est soils are accreting as sea level rises (Krauss et al2014) providing continual carbon sequestration onthe order of 13ndash20 Mg C haminus1 yrminus1 (Breithaupt et al2012 Chmura et al 2003) Clearly there can be amajor climate benefit to halting or even slowing the rateof mangrove conversion with a rough potential esti-mated to be 25ndash122 Tg C yrminus1 (Pendleton et al 2012Siikamaki et al 2012) For nations with large mangroveholdings protection and restoration can make majorcontributions to meeting climate mitigation targets(Herr and Landis 2016)

While many mangrove forests do accumulate largequantities of soil carbon others do not There can besignificant variability in soil carbon stocks across dif-ferent mangrove forests (Jardine and Siikamaki 2014)but also within the same mangrove forest (Adameet al 2015 Kauffman et al 2011) Understanding thedistribution of soil carbon in mangrove forests will bevery important in prioritizing protection and restora-tion efforts for climate mitigation The controls on soilcarbon stocks are diverse and are likely scale dependenthowever some generalizations can be made Man-grove forests no matter how productive will struggleto have high soil carbon stocks in the upper meterof soil if they receive large annual sediment loadsMangrove forests in river deltas such as the Sundar-bans (Banerjee et al 2012) and the Zambezi river deltain Mozambique (Stringer et al 2016) typically onlycontain a few percent organic carbon throughout thesoil profile These locations may still have very highcarbons stocks but the density of carbon is low dueto the high allocthonous input of mineral sedimentsConversely forests with moderately low productivitycan accumulate large amounts of soil carbon if theyare in an isolated hydrogeomorphic setting (Ezcurraet al 2016) Within the same mangrove forest thereare typically steep hydrogeomorphic gradients fromthe seaward to landward extent of the forest whichresults in zonation of both vegetation (Snedaker 1982)and soil carbon storage (Kauffman et al 2011 Ouyanget al 2017 Ewers Lewis et al 2018) but not necessarilyfor the same reasons Within a similar hydrogeomor-phic position forest productivity and soil edaphicconditions (eg redox potential pH salinity) drivingdecomposition rates are often the dominant controlson soil carbon density Consideration of this nested

2


hierarchy of controls will be necessary to successfullycapture the variability in soil carbon at both local andglobal scales

Accurate estimates and an understanding of thespatial distribution of mangrove soil carbon stocksare a critical first step in understanding climatic andanthropogenic impacts on mangrove carbon stor-age and in realizing the climate mitigation potentialof these ecosystems through various policy mecha-nisms (Howard et al 2017) Previous global estimates(Atwood et al 2017 Jardine and Siikamaki 2014)do not capture enough of the finer scale spatialvariability that would be required to inform local deci-sions on siting protection and restoration projectsTo close this information gap we have (1) com-piled and published a harmonized global database ofthe profile distribution of soil carbon under man-groves (2) used this database to develop a novelmachine-learning based data-driven statistical modelof the distribution of carbon density using spatiallycomprehensive data at an sim30 m resolution (3) pro-jected the model results across global mangrove habitatfor the year 2000 (Giri et al 2011) and (4) over-laid estimates of mangrove forest change between2000 and 2012 (Hamilton and Casey 2016) to esti-mate potential soil carbon emissions from recent forestconversion

2 Methods

21 Mangrove soil carbon databaseA harmonized globally representative database (avail-able at 107910DVNOCYUIT) was compiled frompeer-reviewed literature grey literature and from con-tributions of unpublished data from a number ofresearchers and organizations Details of databasedevelopment and a statistical summary of the dataare given in the supplemental information availableat stacksioporgERL13055002mmedia

22 Spatial modelling of soil organic carbonIn order to maximize the utilization of available soilcarbon data we developed a machine learning-basedmodel of organic carbon density (OCD) which modelsOCD as a function of depth (d) an initial estimate ofthe 0ndash200 cm organic carbon stock (OCS) from theglobal SoilGrids 250 m model (Hengl et al 2017) anda suite of spatially explicit covariate layers (X119901)

OCD(119909119910119889) = 119889 + OCSSG +1198831(119909119910)+1198832(119909119910) + 119883119901(119909119910)

where OCSSG is the aggregated organic carbon stockestimated for 0ndash200 cm depth using global SoilGrids250 m approach down-sampled from 250 mndash30 m res-olution and xyd are the 3D coordinates northingeasting and soil depth (measured to center of a hori-zon) Note here that we model spatial distribution

of OCD in three dimensions (soil depth used as apredictor) using all soil horizons layers at differentdepths which means that a single statistical modelcan be used to predict OCD at any arbitrary depthThis 3D approach to modeling OCD reduces the needfor making complex assumptions about the downcoretrends in OCD and maximizes the use of collecteddata

The derived spatial prediction model is then usedto predict OCD at standard depths 0 30 100 and200 cm so that the organic carbon stock (OCS) canbe derived as a cumulative sum of the layers downto the prediction depth for every 30 m pixel identi-fied as having mangrove forest in the year 2000 (Giriet al 2011) Importantly we found that there is a spatialmismatch between the global mangrove forest dis-tribution (GMFD) of Giri et al (2011) and satelliteimagery (figure S2) To best resolve this spatial mis-match we have adjusted the GMFD by growing allvectors by one pixel (sim30 m) and then filtering out anypixel that falls over water by using Landsat NIR band(see SI for more details)

Environmental covariates have been compiled torepresent the postulated major controls on OCS insoils generally (McBratney et al 2003) and specifi-cally for mangrove ecosystems (Balke and Friess 2016)Covariates included

1 Vegetation characteristics including percentforest cover (Hansen et al 2013) and Landsatbands 3 (red) 4 (near infrared) 5 (shortwaveinfrared) and 7 (shortwave infrared) for the year2000 (Hanson et al 2013) retrieved from httpearthenginepartnersappspotcomscience-2013-global-forestdownload_v13html

2 Digital elevation data which at or near sea-levelapproximately follows forest canopy height (Simardet al 2006) from the shuttle radar topography mis-sion (SRTM GL1 NASA 2013) was retrieved fromhttpslpdaacusgsgov maintained by the NASAEOSDIS Land Processes Distributed Active ArchiveCenter (LP DAAC) at the USGSEarth ResourcesObservation and Science (EROS) Center SiouxFalls South Dakota

3 Long-term averaged (1990ndash2010) monthly sea sur-face temperature (SST) averaged into four seasonswere generated in Google Earth Engine from NOAAAVHRR Pathfinder Version 52 Level 3 Collateddata (Casey et al 2010) and downscaled to 30 mresolution using bicubic resampling

4 The M2 tidal elevation amplitude product(FES2012) from a global hydrodynamic tidal modelwhich assimilates altimetry data from multiple plat-forms was used to represent tidal range at eachlocation The FES2012 product was produced byNoveltis Legos and CLS Space Oceanography Divi-sion and distributed by Aviso with support fromCnes (wwwavisoaltimetryfr)

3


5 Averaged (2003ndash2011) monthly total suspendedmatter (TSM) averaged into four seasons estimatedfrom MERIS imagery collected by the EuropeanSpace Agencyrsquos Envisat satellite Processed and vali-dated TSM data was retrieved from the GlobColourproject (httphermesacrifr)

6 A mangrove typology map delineating mangrovesinto estuaries and then either organogenic ormineralogenic based on an analysis of TSM andtidal amplitude data (Zu Ermgassen unpublisheddata)

Sea surface temperature tidal amplitude and TSMare 4 km resolution ocean products and needed tobe extrapolated to each pixel containing mangroveforest Missing values in the sea surface tempera-ture tidal amplitude and TSM were first filled-inusing spline interpolation in SAGA GIS then down-scaled to 30 m resolution using bicubic resampling inGDAL Including SoilGrids and depth there were atotal of 20 covariates used in building the mangroveOCD model

The ability of the training points to representthe entire covariate space of the global mangrovedomain was assessed by conducting a principal com-ponents analysis (PCA) on 15 000 randomly selectedpoints and the 1613 points used in the spatialmodel Spatial variables were detrended and centeredby subtracting the mean and dividing by the stan-dard deviation (sd) before entering into the PCAanalysis

Soil carbon typically varies in highly non-linearways with depth and across the landscape and assuch the ability of standard parametric models to cap-ture this variation is limited (Jardine and Siikamaki2014 Hengl et al 2017) Here we model the spatial(xyd) distribution of OCD using a machine learn-ing random forest model implemented in the rangerpackage (Wright and Ziegler 2015) in the R environ-ment for statistical computing (R Core Team 2000)Given the clustered nature of the point data we haveimplemented a spatially balanced random forest modeldesign Model performance was assessed with a 5fold (Leave-Location-Out) cross-validation procedurewhere 20 of complete locations were withheld in eachmodel refitting (Gasch et al 2015) The relative impor-tance of using SoilGrids as a covariate was assessed byimplementing the cross-validation procedure with andwithout this variable

Finally prediction error of OCS for 0ndash1 m depthwas derived forplusmn1 sd based on the quantile regressionapproach of Meinshausen (2006) and implementedin R via the ranger package This procedure is rela-tively computationally demanding so a random subsetof approximately 15 000 points were selected to cal-culate prediction errors All modeling was run onISRIC High Performance Computing servers with48 cores of 256 GB RAM

23 Data analysisSoil carbon stocks were calculated for the global extentof mangroves for the year 2000 by summing the OCS ineach pixel for 1 and 2 m depths Country level carbonstocks were also calculated for the same depths Giventhe fringing nature of mangroves a global spatial vectordata layer was built that allocated the offshore area foreach country where mangrove forests can be found Itwasderived fromtheExclusiveEconomicZone for eachcountry This layer was then dissolved with the onshoreareas for eachassociated country and subsequentlyusedto quantify mangrove OCS tonnage and areal extent

Potential loss of OCS due to mangrove habitatconversion was calculated between 2000 and 2015 bysumming the OCS in mangrove forest pixels whichwere identified to be deforested While this analysiscannot distinguish between natural and anthropogenicdisturbance human-driven land use change is believedto be by far the dominant driver of deforestationin mangrove ecosystems (Alongi 2002 Murdiyarsoet al 2015) We define deforested using the GlobalForest Change dataset (Hansen et al 2013) availableonline from httpearthenginepartnersappspotcomscience-2013-global-forest We chose to use thisapproach for estimating deforestation instead of usingthe derived mangrove tree cover loss data producedby Hamilton and Casey (2016) as used by Atwoodet al (2017) because the Hamilton and Casey (2016)analysis only considered forested area as area actuallycovered by trees (ie if a 100 ha forest has 80 treecover then it is counted as 80 ha of forest) In our opin-ion this definition mischaracterizes forest area extentNext an estimate of the soil carbon emissions associ-ated with land use conversion is needed The amountof OCS lost can be highly variable and depends onthe new land use (Kauffman et al 2014 2016b Joneset al 2015) and probably on soil properties Pendle-ton et al (2012) used a 25ndash100 loss range Donatoet al (2011) used a low estimate of 25 of the OCSin top 30 cm and 75 in top 30 cm + 35 fromdeeper layers as a high estimate Expanding on ear-lier work Kauffman et al (2017) found that on average54 of belowground carbon (soil + roots) to 3 m waslost after conversion to shrimp ponds and pasturesGiven the limited number of studies comparing soilOCS change with land use change in this work weadopt the same 25ndash100 range as used by Pendletonet al (2012) applied to the first meter of soil Finallycountry level statistics for OCS loss were calculated asdescribed above All global and country level analy-ses were performed on the 30 m resolution dataset inGoogle Earth Engine (Gorelick et al 2017)

3 Results

31 Model resultsThe random forest model was successful in captur-ing the major variation in OCD across the mangrove

4


Figure 1 Model fitting results the corresponding 3D Random Forest model for soil organic carbon density (a) with cross-validationresults in (b) and relative variable importance plot (c) TSM = total suspended matter SST = sea surface temperature (numbersfollowing TSM and SST refer to quarter of the year) NIR = Near Infrared SW1 = Short wave mid infrared

database (figure 1(a)) with an R2 of 084 and a rootmean square error (RMSE) of 69 kg mminus3 compared tothe mean OCD value of 296 kg mminus3 Cross-validationresults yield an R2 of 063 and an RMSE of 109 kg mminus3

(figure 1(b)) which is the de-facto mapping accu-racy to be expected on the field Low OCD valueswere slightly over-predicted and high OCD valueswere under-predicted (figure 1(b)) The initial OCSprediction from SoilGrids 250 m was the most impor-tant variable explaining mangrove OCD Runningthe 5 fold cross-validation without and with Soil-Grids indicated that this single variable explainedimproved model performance by nearly 50 (R2

increased from 042ndash063) Seasonal total suspendedmatter (TSM) depth of sample mangrove tree coverLandsat Red band sea surface temperature (SST) andtidal range were the next ten most important vari-ables respectively (figure 1(c)) Quantile regressionanalysis indicated that the full uncertainty (plusmn1 sd)about a mean prediction of carbon stocks to 1 m depthaveraged 404 of the mean OCS with lowest rel-ative uncertainty in the most carbon-rich mangroveforests (figure S7)

32 Mangrove soil carbon storageProjection of the mangrove OCD model to globalmangrove forests revealed the distribution of soil car-bon storage in these ecosystems (figure 2) The mean(plusmn1 sd) OCS to 1 m depth was 361plusmn 136 Mg C haminus1

with a range of 86ndash729 Mg C haminus1 At the nationallevel (table S1) Bangladesh had the lowest per hastocks averaging just 127 Mg C haminus1 followed by Chinaand the nations bordering the Persian Gulf and RedSea with an average OCS of 214 and 233 Mg C haminus1respectively The highest per ha stocks were foundin many of the pacific island nations averaging505 Mg C haminus1 with much of Southeast Asia rankingwell above the global mean

While the national level comparisons are reveal-ing by modeling at a 30 m resolution much richerdetails of potential within forest variation in OCS are

seen (figure 2) Mangrove forests dominated by sed-iment laden fluvial inputs typically have consistentlylow OCS as seen in the Sundarbans and Madagascar(figures 2(a) and (e)) In non-deltaic mangroves themodel appears to have captured the large zonal vari-ation in OCS that is often observed in field studies(figures 2(b) and (c)

33 Soil carbon loss due to habitat lossUtilizing the Hanson et al (2013) global deforesta-tion analysis (figure S8) we found that 278049 ha(167 of total) of area identified as mangrove habitatin the year 2000 was deforested resulting in the com-mitted emission of 304ndash122 Tg C (111ndash447 Tg CO2)from mangrove forest soils due to land use changebetween 2000 and 2015 (figure 3) The relative rank ofnations in terms of loss of mangrove forest area andOCS were often not the same (table S1) Indonesiaalone was responsible for 52 of this global loss withMalaysia and Myanmar representing another 25 ofthe global total loss (figure 3(c)) When visualized asa percent loss from year 2000 stocks a slightly differ-ent pattern emerged (figure 3(d)) Guatemala had thehighest percent loss of mangrove OCS (09ndash68)followed by several southeast Asian nations but highpercent losses were also found in several Caribbeanisland nations as well as the United States andseveral west African countries

4 Discussion

41 Amount and distribution of Mangrove SOCOur new estimate of global mangrove OCS of64 Pg C in the upper meter and 126 Pg C to 2 m islargely consistent with past efforts to calculate thisvalue (Donato et al 2011 Jardine and Siikamaki2014 Sanders et al 2016) However our estimate isdouble that of Atwood et al (2017) primarily due totheir use of the Hamilton and Casey (2016) estimate ofmangrove extent instead of Giri et al (2011) Impor-tantly by using an environmental covariate model

5


Figure 2 Global distribution of mangrove soil carbon stocks for the top meter of soil (hex bin area sim19 000 km2) and detailed maps(30 m resolution) for selected mangrove regions of the world (1) Sundarbans along the IndiaBangladesh border (2) Bahıa de losMuertos Pacific coast of Panama (3) southwest coast of Papua Indonesia (4) Hinchinbrook Island Queensland Australia (5)Ambaro Bay Madagascar and (6) Guinea-Bissau and Guinea along the West African coast In top panel data presented as mean stock(Mg C haminus1) for mangrove forest area only within each hex bin

we have been able to make plausible estimates forregions where no sampling has taken place instead ofrelying on global mean values (ie Atwood et al 2017)

The total amount of soil carbon was similar inour analysis and the most comparable analysis thatof Jardine and Siikamaki (2014) but the spatial distri-bution of carbon-rich versus carbon-poor mangrovesvaried substantially For example we found muchhigher OCS levels in West Africa than in East Africannations (figure 2 and table S1) but the reverse wasfound by Jardine and Siikamaki (2014) Large dis-crepancies were also found for Colombia Sri Lankaand many of the countries bordering the Red SeaThese differences were most likely driven by lack ofdata in those regions at the time of the analysis byJardine and Siikamaki (2014) given that nearly halfthe data in our database was collected after theirstudy was published Additionally our analysis sug-gested a much larger range in OCS (86ndash729 Mg C haminus1)compared to 272ndash703 Mg C haminus1 in the analysis of

Jardine and Siikamaki (2014) This difference inrange was likely due to the inclusion of more datafrom sub-tropical and temperate mangroves (figureS3)

The depth trend analysis (figure S6) and ran-dom forest variable importance (figure 1(b)) bothindicated that depth should be considered in calcu-lation of OCS For locations that were either stablepeat domes or located in estuaries receiving largeannual sediment loads a stable OCD profile distri-bution would be expected and this was found for manysites (figures S6(a) and (e)) However where man-groves are growing in a mineral matrix that is receivingonly low sediment loads a decline in OCD may beexpected as the carbon inputs from the productivemangrove forest would be concentrated in the sur-face horizons (figure S6(b)) Still in other cases (figureS6(c)) changes in hydrologicsediment regimes canlead to irregular depth patterns or even an increase inOCD with depth

6


Figure 3 Top 20 nation rankings for (a) total mangrove area lost between the years 2000 and 2012 (b) area loss as a percent of year2000 mangrove area (c) total soil organic carbon stocks (d) carbon loss as a percent loss of year 2000 soil carbon stock Range invalues for (c) and (d) come from 25ndash100 loss of carbon in upper meter of soil in pixels identified as being deforested between theyears 2000 and 2015

While total area of mangroves was a key deter-minant of total soil carbon storage amongst the top25 mangrove OCS holding nations there was a nearlyeven split between nations with smaller area of highsoil carbon density forests and those nations with lotsof low soil carbon density forests (figure 4) Indone-sia was the clear exception to this trend with thelargest mangrove holdings which also contain rich car-bon stocks resulting in Indonesia alone holding nearly25 of the worldrsquos mangrove OCS

Compared to terrestrial carbon pools mangroveforests rank low due to their limited spatial extent Forexample in the upper meter of soil permafrost affectedsoils are estimated to store 472plusmn 27 Pg C (Hugeliuset al 2014) tropical forests contain sim188 Pg C andsoils under permanent cropping contain sim150 Pg C(table 1) However on an equal area basis man-grove forests on average store more soil carbon thanmost other ecosystems (table 1) Importantly our

analysis has demonstrated mangrove soil carbon ishighly variable and many mangroves actually storefairly modest amounts of carbon in the upper oneor two meters of soil While not the focus of thisanalysis it is important to point out that while somemangrove forests store modest levels of OCS in theupper meter of soil they can have high sequestrationrates and conversely carbon-dense mangroves can havelow annual sequestration rates (Lovelock et al 2010MacKenzie et al 2016)

42 Drivers of soil carbon storageOur spatial modelling framework in which global pre-dictions were combined with local high resolutionimages was successful as the general patterns of car-bon variation from SoilGrids 250 m were maintainedwhile the spatial detail was significantly improved bymoving from 250 mndash30 m spatial resolution The ini-tial SoilGrids 250 m OCS prediction (Hengl et al 2017)

7


Table 1 Soil organic carbon stocks (mean with 5th-95th percentile in parentheses) and total storage for different terrestrial ecosystemscompared to mangrove forests

1 m soil organic carbon stock

Land cover category (IGBP code)a Area (106 ha) Pg C Mg C haminus1

Mangrove forestb 166 64 361 (94ndash628)Gelisols (permafrost soils)c 1878 4720 389 (178ndash691)Evergreen Needleleaf forest (1)d 286 600 210 (121ndash346)Evergreen Broadleaf forest (2) 1248 1884 151 (85ndash271)Deciduous Needleleaf forest (3) 116 293 253 (163ndash412)Deciduous Broadleaf forest (4) 165 221 134 (83ndash223)Mixed forest (5) 771 1528 198 (93ndash343)Closed shrublands (6) 56 62 110 (39ndash223)Open shrublands (7) 1933 3258 169 (49ndash329)Woody savannas (8) 1179 1858 158 (82ndash274)Savannas (9) 1010 1129 112 (52ndash201)Grasslands (10) 1810 2801 155 (56ndash289)Permanent wetlands (11)e 104 251 241 (114ndash474)Croplands (12) 1177 1496 127 (60ndash200)CroplandNatural veg mosaic (14) 868 1177 136 (58ndash238)

a data for MODIS-based IGBP land cover classes (Friedl et al 2010) extracted from 1 m OCS map for the year 2010 produced by Sanderman

et al (2017)b mangrove area and OCS data from this studyc permafrost area from Tarnocai et al (Tarnocai et al 2009) OCS data from Hugelius et al (2014)d some overlap between class 1 (evergreen Needleleaf forest) and gelisolse class 11 (permanent wetlands) likely has overlap with mangrove area

Figure 4 Rank of nations by mangrove area plotted againstrank by soil carbon density for all nations containinggt30 Tg C Bubble size is proportional to total carbon stock(Tg C) within each nation

was based upon machine learning algorithms using237 covariates that covered the main state factors ofsoil formation (Jenny 1994)mdashclimate relief livingorganisms (vegetation) and parent materialmdashand wasfour times more important in predicting mangroveOCD than any of the local covariates Howeverthe local covariates allowed for a much more refinedpicture of the spatial variation in OCS within man-grove forests (figure 2) which was not captured in the250 m resolution SoilGrids 250 m prediction Covari-ates related to hydrogeomorphology (TSM and tidalrange) as hypothesized were important predictors oflocal variation in OCD Both TSM and tidal rangewere strongly negatively correlated with OCD suggest-ing that locations with either high sediment loads orstrong tidal flushing do not accumulate large carbon

stocks The mangrove typology ended up being unin-formative likely because most of the data that went intothis classification was already captured in the model

Covariates related to mangrove biomass (SRTMelevation and Landsat bands) were also important inexplaining the local distribution of OCD (figure 1(b))However as pointed out by Bukoski et al (2017) it isunclear whether the importance of these vegetation-related data are causal drivers of differences in OCDor they just happen to co-vary in the same way asOCD To further explore the relationship betweenforest biomass carbon and soil carbon storage weextracted aboveground biomass data from Hutchisonet al (2014) and compared it to our OCS results (fig-ure 5) While a clear positive trend was found betweenbiomass and soil carbon storage (R2 = 026) there isclearly a lot of variance especially at lower biomass levelswhere nearly the full range in OCS can be found

Depth below the soil surface was an importantcovariate in modeling OCD distribution (figure 1(b))This finding was supported by the database depthtrend analysis (figure S6) which indicated that a flatdepth distribution of OCD would be an incorrectassumption 37ndash64 of the time These findings sug-gest that the simple scaling performed in figure S5and in several previous assessments (Bukoski et al2017 Jardine and Siikamaki 2014 Atwood et al 2017)may not be an accurate estimation of total OCSespecially when extrapolating from a surface horizonsample alone

43 Soil carbon loss due to land conversion (2000ndash2015)Our analysis suggests that mangrove soils have lostor are at least committed to losing 304ndash122 Tg C dueto the land use conversion that occurred between the

8


Figure 5 Comparison of aboveground biomass carbon (Hutchison et al 2014) with organic carbon stocks (OCS) for the top meterof soil Points were generated by casting 10 000 random points into mangrove areas and extracting values from both maps Linearregression R2 = 026

years 2000 and 2015 (figure 3) Given that at the globallevel the rate of mangrove forest lost was consistentover this time period (Hamilton and Casey 2016)we estimated an annual soil carbon emission of 20ndash81 Tg C yrminus1 This value is significantly lower thanprevious estimates (Donato et al 2011 Pendletonet al 2012) for two reasons First we use remotesensing-based measurements of actual mangrove lossinstead of applying a large range of annual conver-sion rates which are notoriously variable according totheir source (Friess and Webb 2011) Second we havesummed the actual OCS values for each of the pix-els where land conversion has taken place (ie figureS8) instead of applying a conversion rate to a meanOCS value

The three nations of Indonesia Malaysia andMyanmar contributed 77 of global mangrove OCSloss for this time period (figure 3) Despite similar arealoss (figure 3(a)) Malaysia lost approximately twice asmuch soil carbon as Myanmar due to the large dif-ferences in carbon density between these two nations(mean OCS = 485plusmn 57 versus 245plusmn 63 Mg C haminus1respectively) This comparison highlights the impor-tance of using local OCS values for estimating carbonemissions attributed to mangrove conversion

Not all land use conversions result in equal lossof OCS Conversion of mangrove forest to shrimpponds results in a rapid and near complete loss of car-bon in the upper meter of soil (Kauffman et al 2014)as well as losses deeper in the soil profile (Kauffmanet al 2017) Conversion to other agricultural uses suchas pasture for beef production (Kauffman et al 2016b)and cereal crops (Andreetta et al 2016) also appearto result in large soil carbon emissions Howevermangrove degradation and loss due to over harvest-ing for fuelwood (Jones et al 2015) or due to natural

disturbance (Cahoon et al 2003) likely leads to moremoderate emissions as decomposition and erosionexceed new plant carbon inputs

It is important to note that nearly all availabledata on OCS loss due to conversion to other landuses come from organogenic mangrove forests Ina mineral-dominated mangrove systems with only afew percent sediment OCC we would not expect thesame level of carbon loss as when peat deposits aredrained or removed In fact reclamation of deltaic sed-iments for paddy rice cultivation can lead to increasesin OCS (Kalbitz et al 2013) although methane emis-sions would be expected to increase Additionally ifmangrove habitat is lost due to deforestation with-out a change in hydrologic regime mineral-dominatedmangroves can continue to accrete carbon but ata lower rate than in a system that has additionalorganic matter inputs from the mangroves themselves(Perez et al 2017)

44 Limitations and uncertaintiesWhile we endeavored to ensure that the model inputdata was of the highest quality possible there undoubt-edly remain unknown errors in the database whichare contributing to model error Machine learningmodels are particularly sensitive to outlier valuesand extrapolation (Murphy 2012) Various researchgroups use different methods for determining theorganic carbon concentration (OCC) of a samplewith not all publications reporting whether or notresults were corrected for occurrence of inorganic car-bon or whether or not roots were excluded beforefurther processing Bulk density (BD) is a difficultparameter to measure accurately in many soils andbased on our analysis of BD versus OCC (figure S1)some reported data are unlikely to be accurate We

9


developed a procedure to correct potential BD errorsbut a pedotransfer function gives only an approxi-mation of the true value Given the importance ofdepth in our models (figure 1(b)) it was unfortu-nate that so many investigations only report OCS forlarge depth increments We suggest that future studiesusing the common practice of collecting subsampleswithin larger horizon increments (eg Kauffman andDonato 2012 Kauffman et al 2016a) report the spe-cific depth increment of the sample rather than that ofthe entire core This additional level of transparency inthe data would allow mass-preserving splines (Bishopet al 1999) to be fit through the distinct measurementintervals resulting in unbiased estimates of OCS

The largest uncertainty in the input data likelyresulted from imperfect information on plot locationWhether accidental or purposeful (ie not wantingto identify exact locations) spatially misplaced data inpublications are of limited utility in geospatial appli-cations All of the covariate data for each mangrovepoint were selected from spatial layers resulting inthe potential for a mismatch between the recordedOCS and the spatial predictors In this study we visu-ally inspected all coordinates against Google Earthimagery and our adjusted mangrove domain and thencontacted many authors to seek further informationand manually adjusted coordinates when we wereconfident that the adjustments lead to better spatiallocation In the final analysis 199 soil profiles had tobe excluded from analysis because we could not confi-dently locate these plots within the adjusted mangrovespatial domain

5 Conclusions

This work has produced three resources which wehope to be of significant value to the blue carbonresearch and management communities 1) a large har-monized database of soil carbon data from mangroveecosystems 2) high-resolution (30 m) predictionswith error of soil carbon stocks across all mangroveforests globally 3) estimates of potential soil carbonlosses due to mangrove habitat loss between 2000and 2015 By using a statistical data-driven modelwe have been able to produce credible estimatesOCS for the numerous mangrove regions where nofield data exist We found that mangrove OCS ishighly variable (86ndash729 Mg C haminus1 in the top meter)but that much of the variability could be capturedusing spatially-comprehensivepredictors in amachine-learning framework Of the 6400 Tg C in the uppermeter of soil 30ndash122 Tg have likely been lost due todeforestation since the year 2000 with 77 of thisloss attributed to Indonesia Malaysia and MyanmarThese spatially-explicit estimates of mangrove soil car-bon storage and loss will provide a practical first stepfor enabling nations to prioritize mangrove protectionas part of their climate mitigation and adaptation plans

Acknowledgments

We would like to thank A Andreetta M Osland JM Smoak A DelVecchia M E Gonneea R K Bho-mia and J Kelleway for providing additional datafrom their publications S-T Kang for translatingand extracting data from Chinese language papersR K Bhomia acknowledges CIFOR SWAMP projectUSFS International Program and USAID for fund-ing M Rahman acknowledges support from USAIDfor funding PM and CJS acknowledge support fromthe Australian Research Council (DE130101084 andLP160100242) and (DE160100443 and DP150103286)respectively CD acknowledges support from the Ruf-ford Foundation and Darwin Initiative for fundingJS TH GF and KS were supported by an anonymousgift to The Nature Conservancy MFA was sup-ported by funding from the Queensland Governmentthrough the Advance Queensland Fellowship ISRICis a non-profit organization primarily funded by theDutch government

Data availability

The mangrove soil carbon database and modeloutputs can be downloaded from Harvard dataverse athttpsdataverseharvardedudatasetxhtmlpersistentId=doi107910DVNOCYUIT

ORCID iDs

Jonathan Sanderman httpsorcidorg0000-0002-3215-1706Peter IMacreadie httpsorcidorg0000-0001-7362-0882

References

Adame M F Santini N S Tovilla C Vazquez-Lule A Castro L andGuevara M 2015 Carbon stocks and soil sequestration rates oftropical riverine wetlands Biogeosciences 12 3805ndash18

Alongi D M 2002 Present state and future of the worldrsquos mangroveforests Environ Conserv 29 331ndash49

Andreetta A Huertas A D Lotti M and Cerise S 2016 Land usechanges affecting soil organic carbon storage along amangrove swamp rice chronosequence in the Cacheu and Oioregions (northern Guinea-Bissau) Agric Ecosyst Environ 216314ndash21

Atwood T B et al 2017 Global patterns in mangrove soil carbonstocks and losses Nat Clim Change 7 523ndash8

Balke T and Friess D A 2016 Geomorphic knowledge for mangroverestoration a pan-tropical categorization Earth Surf ProcessLandforms 41 231ndash9

Banerjee K Chowdhury M R Sengupta K Sett S and Mitra A 2012Influence of anthropogenic and natural factors on themangrove soil of Indian Sundarbans wetland Arch EnvironSci 6 80ndash91

Barbier E B Hacker S D Kennedy C Koch E W Stier A C andSilliman B R 2011 The value of estuarine and coastalecosystem services Ecol Monogr 81 169ndash93

Bishop T F A McBratney A B and Laslett G M 1999 Modelling soilattribute depth functions with equal-area quadratic smoothingsplines Geoderma 91 27ndash45

10


Boucher O Bellassen V Benveniste H Ciais P Criqui P GuivarchC Le Treut H Mathy S and Seferian R 2016 Opinion in thewake of Paris Agreement scientists must embrace newdirections for climate change research Proc Natl Acad SciUSA 113 7287ndash90

Brander L Wagtendonk A Hussain S McVittie A Verburg P H deGroot R S and van der Poeg S 2012 Ecosystem service valuesfor mangroves in Southeast Asia a meta-analysis and valuetransfer application Ecosyst Serv 1 62ndash9

Breithaupt J L Smoak J M Smith T J Sanders C J and Hoare A2012 Organic carbon burial rates in mangrove sedimentsstrengthening the global budget Glob Biogeochem Cycles 26GB3011

Bukoski J J Broadhead J S Donato D C Murdiyarso D andGregoire T G 2017 The use of mixed effects models forobtaining low cost ecosystem carbon stock estimates inmangroves of the Asia Pacific PLoS ONE 12 e0169096

Cahoon D R Hensel P Rybczyk J McKee K L Proffitt C E andPerez B C 2003 Mass tree mortality leads to mangrove peatcollapse at Bay Islands Honduras after Hurricane Mitch JEcol 91 1093ndash105

Casey K S Brandon T B Cornillon P and Evans R 2010 The pastpresent and future of the AVHRR pathfinder SST programOceanography from Space (Dordrecht Springer Netherlands)pp 273ndash87

Chmura G L Anisfeld S C Cahoon D R and Lynch J C 2003 Globalcarbon sequestration in tidal saline wetland soils GlobBiogeochem Cycles 17 1111

Donato D C Kauffman J B Murdiyarso D Kurnianto S Stidham Mand Kanninen M 2011 Mangroves among the mostcarbon-rich forests in the tropics Nat Geosci 4 293ndash7

Duke N C Kovacs J M Griffiths A D Preece L Hill D J E VanOosterzee P Mackenzie J Morning H S and Burrows D 2017Large-scale dieback of mangroves in Australiarsquos Gulf ofCarpentaria a severe ecosystem response coincidental with anunusually extreme weather event Mar Freshw Res 681816ndash29

Ewers Lewis C J Sanderman P E Baldock J and Macreadie P I 2018Variability and vulnerability of coastal blue carbon stocks acase study from Southeast Australia Ecosystems 21 263ndash79

Ezcurra P Ezcurra E Garcillan P P Costa M T andAburto-Oropeza O 2016 Coastal landforms andaccumulation of mangrove peat increase carbon sequestrationand storage Proc Natl Acad Sci 113 4404ndash9

Friedl M A Sulla-Menashe D Tan B Schneider A Ramankutty NSibley A and Huang X 2010 MODIS collection 5 global landcover algorithm refinements and characterization of newdatasets Remote Sens Environ 114 168ndash82

Friess D A Thompson B S Brown B Amir A A Cameron CKoldewey H J Sasmito S D and Sidik F 2016 Policy challengesand approaches for the conservation of mangrove forests inSoutheast Asia Conserv Biol 30 933ndash49

Friess D A and Webb E L 2011 Bad data equals bad policy how totrust estimates of ecosystem loss when there is so muchuncertainty Environ Conserv 38 1ndash5

Gasch C K Hengl T Graler B Meyer H Magney T S and Brown D J2015 Spatio-temporal interpolation of soil water temperatureand electrical conductivity in 3D+ T the cook agronomy farmdata set Spat Stat 14 70ndash90

Gedan K B Kirwan M L Wolanski E Barbier E B and Silliman B R2011 The present and future role of coastal wetland vegetationin protecting shorelines answering recent challenges to theparadigm Clim Change 106 7ndash29

Giri C Ochieng E Tieszen L L Zhu Z Singh A Loveland T MasekJ and Duke N 2011 Status and distribution of mangroveforests of the world using earth observation satellite data GlobEcol Biogeogr 20 154ndash9

Gorelick N Hancher M Dixon M Ilyushchenko S Thau D andMoore R 2017 Google Earth engine planetary-scalegeospatial analysis for everyone Remote Sens Environ 20218ndash27

Grassi G House J Dentener F Federici S den Elzen M and PenmanJ 2017 The key role of forests in meeting climate targetsrequires science for credible mitigation Nat Clim Change 7220ndash6

Hamilton S E and Casey D 2016 Creation of a high spatio-temporalresolution global database of continuous mangrove forestcover for the 21st century (CGMFC-21) Glob Ecol Biogeogr25 729ndash38

Hansen M C et al 2013 High-resolution global maps of 21st centuryforest cover change Science 342 850ndash3

Hengl T et al 2017 SoilGrids 250 m Global gridded soil informationbased on machine learning PLoS ONE 12 e0169748

Herr D and Landis E 2016 Coastal blue carbon ecosystemsOpportunities for Nationally Determined Contributions PolicyBrief (Gland IUCN and Washington DC TNC)

Howard J Sutton-Grier A Herr D Kleypas J Landis E Mcleod EPidgeon E and Simpson S 2017 Clarifying the role of coastaland marine systems in climate mitigation Front Ecol Environ15 42ndash50

Hugelius G et al 2014 Estimated stocks of circumpolar permafrostcarbon with quantified uncertainty ranges and identified datagaps Biogeosciences 11 6573ndash93

Hutchison J Manica A Swetnam R Balmford A and Spalding M2014 Predicting global patterns in mangrove forest biomassConserv Lett 7 233ndash40

Jardine S L and Siikamaki J V 2014 A global predictive model ofcarbon in mangrove soils Environ Res Lett 9 104013

Jenny H 1994 Factors of soil formation a system of quantitativepedology (Courier Corporation)

Jones T et al 2015 The dynamics ecological variability andestimated carbon stocks of mangroves in Mahajamba BayMadagascar J Mar Sci Eng 3 793ndash820

Kalbitz K et al 2013 The carbon count of 2000 years of ricecultivation Glob Change Biol 19 1107ndash13

Kauffman J B Arifanti V B Basuki I Kurnianto S Novita NMurdiyarso D Donato D C and Warren M W 2016aProtocols for the measurement monitoring and reporting ofstructure biomass carbon stocks and greenhouse gasemissions in tropical peat swamp forests Working paper 221(Bogor CIFOR)

Kauffman J B Hernandez Trejo H del Carmen Jesus Garcia MHeider C and Contreras W M 2016b Carbon stocks ofmangroves and losses arising from their conversion to cattlepastures in the Pantanos de Centla Mexico Wetl EcolManage 24 203ndash16

Kauffman J B Arifanti V Trejo H del Carmen Jesus Garcıa MNorfolk J Hadriyanto D Cifuentes-Jara M Murdiyarso D andCifuentes-Jara M 2017 The jumbo carbon footprint of ashrimp carbon losses from mangrove deforestation FrontEcol Environ 15 183ndash8

Kauffman J B and Donato D 2012 Protocols for the measurementmonitoring and reporting of structure biomass and carbonstocks in mangrove forests Working paper 86 (Bogor CIFOR)

Kauffman J B Heider C Cole T G Dwire K A and Donato D C2011 Ecosystem carbon stocks of micronesian mangroveforests Wetlands 31 343ndash52

Kauffman J B Heider C Norfolk J and Payton F 2014 Carbonstocks of intact mangroves and carbon emissions arising fromtheir conversion in the Dominican Republic Ecol Appl 24518ndash27

Krauss K W McKee K L Lovelock C E Cahoon D R Saintilan NReef R and Chen L 2014 How mangrove forests adjust torising sea level New Phytol 202 19ndash34

Lovelock C E Sorrell B K Hancock N Hua Q and Swales A 2010Mangrove forest and soil development on a rapidly accretingshore in New Zealand Ecosystems 13 437ndash51

MacKenzie R A Foulk P B Klump J V Weckerly K Purbospito JMurdiyarso D Donato D C and Nam V N 2016Sedimentation and belowground carbon accumulation ratesin mangrove forests that differ in diversity and land use a taleof two mangroves Wetl Ecol Manage 24 245ndash61

11


McBratney A Mendonca Santos M and Minasny B 2003 On digitalsoil mapping Geoderma 117 3ndash52

Meinshausen N 2006 Quantile regression forests J Mach LearnRes 7 983ndash99

Mumby P J et al 2004 Mangroves enhance the biomass of coral reeffish communities in the Caribbean Nature 427 533ndash6

Murdiyarso D Purbopuspito J Kauffman J B Warren M WSasmito S D Donato D C Manuri S Krisnawati H TaberimaS and Kurnianto S 2015 The potential of Indonesianmangrove forests for global climate change mitigation NatClim Change 5 8ndash11

Murphy K P 2012 Machine Learning A Probabilistic Perspective(Cambridge MA MIT)

Nagelkerken I et al 2008 The habitat function of mangroves forterrestrial and marine fauna a review Aquat Bot 89 155ndash85

Ouyang X Lee S Y and Connolly R M 2017 Structural equationmodelling reveals factors regulating surface sediment organiccarbon content and CO2 efflux in a subtropical mangrove SciTotal Environ 578 513ndash22

Pendleton L et al 2012 Estimating global lsquoblue carbonrsquo emissionsfrom conversion and degradation of vegetated coastalecosystems PLoS ONE 7 e43542

Perez A Machado W Gutierrez D Stokes D Sanders L Smoak J MSantos I and Sanders C J 2017 Changes in organic carbonaccumulation driven by mangrove expansion anddeforestation in a New Zealand estuary Estuar Coast Shelf Sci192 108ndash16

Richards D R and Friess D A 2016 Rates and drivers of mangrovedeforestation in Southeast Asia 2000ndash2012 Proc Natl AcadSci 113 344ndash9

Robertson A I and Phillips M J 1995 Mangroves as filters of shrimppond effluent predictions and biogeochemical research needsHydrobiologia 295 311ndash21

Rockstrom J Gaffney O Rogelj J Meinshausen M Nakicenovic Nand Schellnhuber H J 2017 A roadmap for rapiddecarbonization Science 355 1269ndash71

Sanderman J Hengl T and Fiske G J 2017 Soil carbon debt of 12 000years of human land use Proc Natl Acad Sci 114 9575ndash80

Sanders C J Maher D T Tait D R Williams D Holloway C Sippo JZ and Santos I R 2016 Are global mangrove carbon stocksdriven by rainfall J Geophys Res Biogeosci 121 2600ndash9

Siikamaki J Sanchirico J N and Jardine S L 2012 Global economicpotential for reducing carbon dioxide emissions frommangrove loss Proc Natl Acad Sci 109 14369ndash74

Simard M Zhang K Rivera-Monroy V H Ross M S Ruiz P LCastaneda-Moya E Twilley R R and Rodriguez E 2006Mapping height and biomass of mangrove forests inEverglades National Park with SRTM elevation dataPhotogramm Eng Remote Sens 72 299ndash311

Snedaker S C 1982 Mangrove Species Zonation Why(Netherlands Springer) pp 111ndash25

Spalding M Kainuma M and Collins L 2010 World Atlas ofMangroves (London Earthscan)

Spalding M McIvor A Tonneijck F Tol S and van Eijk P 2014Mangroves for Coastal Defence Guidelines for CoastalManagers and Policy Makers (Wetlands International and TheNature Conservancy)

Stringer C E Trettin C C and Zarnoch S J 2016 Soil properties ofmangroves in contrasting geomorphic settings within theZambezi River Delta Mozambique Wetl Ecol Manage 24139ndash52

Tarnocai C Canadell J G Schuur E A G Kuhry P Mazhitova G andZimov S 2009 Soil organic carbon pools in the northerncircumpolar permafrost region Glob Biogeochem Cycles 23GB2023

R Core Team 2000 R Language Definition (Vienna R Found StatComput)

Temmerman S Meire P Bouma T J Herman P M J Ysebaert T andDe Vriend H J 2013 Ecosystem-based coastal defence in theface of global change Nature 504 79ndash83

Valiela I Bowen J L and York J K 2001 Mangrove forests one of theworldrsquos threatened major tropical environments BioScience 51807ndash15

Wright M N and Ziegler A 2015 ranger a fast implementation ofrandom forests for high dimensional data in C++ and R(arXiv150804409)

12

Environ Res Lett 13 (2018) 055002 httpsdoiorg1010881748-9326aabe1c

LETTER

A global map of mangrove forest soil carbon at 30mspatial resolution

Jonathan Sanderman121 Tomislav Hengl2 Greg Fiske1 Kylen Solvik1 Maria Fernanda Adame3 LisaBenson56 Jacob J Bukoski7 Paul Carnell8 Miguel Cifuentes-Jara9 Daniel Donato19 Clare Duncan48Ebrahem M Eid1020 Philine zu Ermgassen1718 Carolyn J Ewers Lewis8 Peter I Macreadie8 Leah Glass5Selena Gress11 Sunny L Jardine12 Trevor G Jones513 Eugene Ndemem Nsombo14 Md Mizanur Rahman15Christian J Sanders16 Mark Spalding17 and Emily Landis17

1 Woods Hole Research Center 149 Woods Hole Road Falmouth MA 02540 United States of America2 ISRIC mdash World Soil Information Wageningen The Netherlands3 Australian Rivers Institute Griffith University Nathan QLD Australia4 Institute of Zoology Zoological Society of London Outer Circle Regentrsquos Park London NW1 4RY United Kingdom5 Blue Ventures Conservation London United Kingdom6 Centre for Environment Fisheries and Aquaculture Science Lowestoft United Kingdom7 Department of Environmental Science Policy and Management University of California Berkeley CA United States of America8 Centre for Integrative Ecology School of Life and Environmental Sciences Deakin University Burwood VIC Australia9 Forests Biodiversity and Climate Change Program CATIE 7170 Turrialba Costa Rica10 Botany Department Faculty of Science Kafr El-Sheikh University Kafr El-Sheikh Egypt11 School of Applied Sciences Edinburgh Napier University Edinburgh Scotland12 School of Marine and Environmental Affairs University of Washington Seattle WA United States of America13 Department of Forest Resources Management University of British Columbia Vancouver BC Canada14 Institute of Fisheries and Aquatic Sciences University of Douala Doula Cameroon15 Graduate School of Agriculture Kyoto University Kyoto Japan16 National Marine Science Centre Southern Cross University Coffs Harbour NSW Australia17 The Nature Conservancy Arlington VA United States of America18 School of Geosciences University of Edinburgh Edinburgh United Kingdom19 Washington State Department of Natural Resources Olympia WA United States of America20 Biology Department College of Science King Khalid University Abha 61321 Saudi Arabia21 Author to whom any correspondence should be addressed

OPEN ACCESS

RECEIVED

29 June 2017

REVISED

10 April 2018

ACCEPTED FOR PUBLICATION

13 April 2018

PUBLISHED

30 April 2018

Original content fromthis work may be usedunder the terms of theCreative CommonsAttribution 30 licence

Any further distributionof this work mustmaintain attribution tothe author(s) and thetitle of the work journalcitation and DOI

E-mail jsandermanwhrcorg

Keywords blue carbon carbon sequestration land use change machine learning

Supplementary material for this article is available online

AbstractWith the growing recognition that effective action on climate change will require a combination ofemissions reductions and carbon sequestration protecting enhancing and restoring natural carbonsinks have become political priorities Mangrove forests are considered some of the mostcarbon-dense ecosystems in the world with most of the carbon stored in the soil In order formangrove forests to be included in climate mitigation efforts knowledge of the spatial distribution ofmangrove soil carbon stocks are critical Current global estimates do not capture enough of the finerscale variability that would be required to inform local decisions on siting protection and restorationprojects To close this knowledge gap we have compiled a large georeferenced database of mangrovesoil carbon measurements and developed a novel machine-learning based statistical model of thedistribution of carbon density using spatially comprehensive data at a 30 m resolution This modelwhich included a prior estimate of soil carbon from the global SoilGrids 250 m model was able tocapture 63 of the vertical and horizontal variability in soil organic carbon density (RMSE of109 kg mminus3) Of the local variables total suspended sediment load and Landsat imagery were themost important variable explaining soil carbon density Projecting this model across the globalmangrove forest distribution for the year 2000 yielded an estimate of 64 Pg C for the top meter of soilwith an 86ndash729 Mg C haminus1 range across all pixels By utilizing remotely-sensed mangrove forest coverchange data loss of soil carbon due to mangrove habitat loss between 2000 and 2015 was 30ndash122 Tg Cwith gt75 of this loss attributable to Indonesia Malaysia and Myanmar The resulting map products

copy 2018 The Author(s) Published by IOP Publishing Ltd



1 Introduction






2




2 Methods



OCD(119909119910119889) = 119889 + OCSSG +1198831(119909119910)+1198832(119909119910) + 119883119901(119909119910)









3










3 Results


4











4 Discussion


5







6







7














8









9




5 Conclusions


Acknowledgments


Data availability


ORCID iDs


References









10









































11



























12



1 Introduction






2




2 Methods



OCD(119909119910119889) = 119889 + OCSSG +1198831(119909119910)+1198832(119909119910) + 119883119901(119909119910)









3










3 Results


4











4 Discussion


5







6







7














8









9




5 Conclusions


Acknowledgments


Data availability


ORCID iDs


References









10









































11



























12




2 Methods



OCD(119909119910119889) = 119889 + OCSSG +1198831(119909119910)+1198832(119909119910) + 119883119901(119909119910)









3










3 Results


4











4 Discussion


5







6







7














8









9




5 Conclusions


Acknowledgments


Data availability


ORCID iDs


References









10









































11



























12










3 Results


4











4 Discussion


5







6







7














8









9




5 Conclusions


Acknowledgments


Data availability


ORCID iDs


References









10









































11



























12











4 Discussion


5







6







7














8









9




5 Conclusions


Acknowledgments


Data availability


ORCID iDs


References









10









































11



























12







6







7














8









9




5 Conclusions


Acknowledgments


Data availability


ORCID iDs


References









10









































11



























12







7














8









9




5 Conclusions


Acknowledgments


Data availability


ORCID iDs


References









10









































11



























12














8









9




5 Conclusions


Acknowledgments


Data availability


ORCID iDs


References









10









































11



























12









9




5 Conclusions


Acknowledgments


Data availability


ORCID iDs


References









10









































11



























12




5 Conclusions


Acknowledgments


Data availability


ORCID iDs


References









10









































11



























12









































11



























12



























12

Documents

LETTER OPEN ACCESS … global... · titleofthework,journal citationandDOI. E-mail:[email protected] ... (Atwood et al 2017, Jardine and Siikam¨aki 2014), do not capture enough