Overall simulation approach
We estimated net changes in soil organic, above-ground and below-ground living biomass carbon stocks due to mangrove-related LCC that occurred globally between 1996 and 2016, for the year 2016. We did not estimate carbon sequestration because while sequestration rates in mangrove forests are higher than in many other ecosystems, there is only limited capacity for this process to substantially impact global carbon fluxes due to the small area of mangrove forest7,32,39. In addition, global high-resolution maps of mangrove carbon sequestration rates are not available.
Uncertainty in estimates of the area of mangrove loss and gain, carbon stock density, date of deforestation or forestation, proportional carbon stock degradation due to deforestation, and proportional carbon stock accumulation due to forestation, was carried forward using a bootstrap simulation method, through which 1000 replications were used to generate median estimates and 95% confidence intervals. A summary of the sources of data, simulation parameters, and modelling of parameter variability is provided in Supplementary Table 1. For each bootstrap iteration, the area of mangrove gain and loss, carbon stock density, rt, and at values were simulated within each patch of mangrove gain or loss. The D, F, Drt, and Fat values for each patch were then calculated. The median value of the 1000 replicates was used as the estimate. The 2.5 and 97.5 percentiles of the 1000 simulation estimates of change in carbon stock were calculated, thus corresponding to bootstrap 95% confidence intervals for each of the estimated losses or gains in carbon stock within each patch. A comparable bootstrap method was applied to simulate carbon stocks in 1996 for each patch of mangrove present at that time.
In addition to the sources of uncertainty incorporated in the bootstrap simulation, we made several key methodological decisions that could be expected to impact the conclusions of the study. We conducted four sensitivity analyses to quantify the impacts of such decisions. The sensitivity analyses were conducted only for the region of Southeast Asia, because one sensitivity analysis required detailed and spatially explicit data on the replacement land uses following mangrove deforestation, which are only available with the necessary categorisation for Southeast Asia5. Methodological details and results of the sensitivity analyses are included in Supplementary Methods 1.
Mangrove areal extent
We used the Global Mangrove Watch (GMW) datasets of mangrove cover to quantify deforestation and forestation14,15,16,17. While other global mangrove map products exist for specific years42,43, GMW provides maps of extent from multiple years, allowing temporal comparison. The mangrove extent in 1996 and 2016 was mapped using the GMW data products, which are derived from ALOS PALSAR and Landsat satellite-borne sensor data14,15,16. Mangrove-related LCC was defined either as a conversion from mangrove to another form of land or water cover between the 1996 and 2016 datasets (deforestation), or vice versa (forestation). Any change in mangrove cover that was reversed before the end of the study period was therefore not captured. Areas of overlapping and nonoverlapping mangrove extent were compared between these dates to quantify mangrove present in 1996 that was not present in 2016, mangrove present in 2016 that was not previously present in 1996, and areas of no change in mangrove cover between 1996 and 2016. For each patch of mangrove cover in 1996, gain, and loss, the area was calculated under the Eckert VI equal-area projection. This projection was used to calculate the area of mangrove and mangrove change polygons only, and all other analyses were conducted using the World Geodetic System (WGS) 1984 projection17.
The GMW mapping of mangrove forest has an error rate15, leading to quantifiable uncertainty over the presence of mangroves at each location in 1996 and 2016. Accuracy statistics have not been published for each year in the GMW dataset17, so we assumed that all years had an identical accuracy to the best-documented year, which is 201015. Published error rates correspond to individual pixels in the original GMW dataset, so we modelled uncertainty at this spatial scale. For each pixel of recorded mangrove gain or loss, there is a probability that it is an erroneous, false positive example of gain or loss. For each pixel of mangrove or non-mangrove that is recorded in GMW as being the same in 2016 as in 1996, there is a similar probability of error—a false-negative case of gain or loss. While false positive gain and loss errors can be quantified fully based on the available information, we were not able to incorporate false negative errors in the simulation (Supplementary Methods 2). However, false-negative errors can be expected to impact estimates of forestation and deforestation area almost equally, while false-positive errors are biased toward a greater effect on estimates of deforestation (Supplementary Methods 2). For these reasons we incorporated only false positive classification uncertainty into the bootstrap simulation and generation of carbon stock change confidence intervals. Uncertainty in the areal extent of mangroves was not incorporated into the bootstrap simulation for the estimate of carbon stocks in 1996. As the uncertainty estimation for areal extent change does not include false negative classification errors, we do not report confidence intervals for area change statistics, and report only the median estimates from the bootstrap replicates.
For areas of mangrove loss and gain, we incorporated the probability of false positive recording of loss and gain into the simulation. For each pixel within each patch of mangrove gain, we simulated whether it was actually not mangrove in 1996 according to the misclassification error rate for non-mangrove (Supplementary Table 2), and whether it was truly mangrove in 2016 according to the misclassification error rate for mangrove (Supplementary Table 2). The simulated number of gain pixels in each patch was thus calculated as the number of pixels that were simulated to have been both not mangrove in 1996, and mangrove in 2016. Similarly, for each pixel recorded as mangrove deforestation, there is a probability that it was a false positive example of loss. For each pixel within each patch of mangrove loss, we simulated whether it was truly mangrove in 1996 according to the misclassification error rate for mangrove (Supplementary Table 2), and whether mangrove was truly absent in 2016 according to the misclassification error rate for non-mangrove (Supplementary Table 2). The simulated number of loss pixels in each mangrove patch was thus calculated as the number of pixels that were simulated to have been mangrove in 1996, and not mangrove in 2016.
Carbon stock density
Spatial patterns in mangrove carbon densities were quantified using previously published datasets of soil carbon to 1 m depth36, and above- and belowground tree biomass carbon37. Both datasets are derived from systematic reviews of the literature, so may be biased towards relatively high-quality mangrove forests, rather than those that have experienced some natural or anthropogenic disturbance37. The resulting maps of carbon stock density are therefore likely to represent an upper estimate for the potential carbon stock density at each location37. As the dates and resolutions of these mangrove carbon datasets differed from the GMW mangrove extent, per hectare carbon densities were extracted for each patch of mangrove extent in 1996, gain, and loss of mangrove. Where possible, we extracted mean carbon densities for the 0.05 degree grid cell (approximately 5 km) in which the centre of the mangrove patch coincided (Supplementary Table 3). Where data did not coincide at this resolution, 0.5° grid cells (approximately 50 km) were used (Supplementary Table 3). Any remaining data gaps were filled using the global mean carbon stock density (Supplementary Table 3).
Uncertainty in the estimate of carbon stock was modelled as a normally distributed random variable, with the mean value taken as the reported carbon stock density extracted from the published map layers36,37, and the standard deviation of the distribution taken as the reported root mean squared error (RMSE) between the model predictions and validation data36,37. The RMSE for soil carbon is reported in the study as 109 Mg per hectare36. The RMSE for aboveground biomass carbon was calculated as 104.1 Mg per hectare, based on a plot of observed versus predicted values digitised from the original study37 (Supplementary Fig. 2). The soil and biomass carbon stock densities for each patch of mangrove in 1996, patch of mangrove loss, and patch of mangrove gain were simulated, and values of less than zero were replaced with zero, to avoid negative carbon densities.
Loss and accumulation of mangrove carbon
After mangrove deforestation, carbon is lost gradually over a period of time, with biomass carbon typically depleting more rapidly than carbon stored in soils13. We modelled temporal losses of soil carbon stocks according to a previously-published meta-analysis of the proportion of the reference carbon stock (rt) lost over time13. For losses of biomass carbon stocks, we used a meta-analysis of temporal changes in the proportion of the reference tree diameter as a proxy for biomass carbon stock, because a meta-analysis of temporal changes in biomass carbon stock was not available13. The shape of these temporal rt relationships can be observed in Supplementary Fig. 3a, b. The approximate date of mangrove deforestation was quantified by cross-referencing several dates from the GMW dataset to establish the dates of presence and absence17. We cross-referenced the dates of 1996, 2007, 2010, and 2016 to identify the dates of mangrove presence and absence at each location.
Uncertainty in proportional losses of carbon due to mangrove deforestation was incorporated in two ways. First, there is uncertainty in the date of mangrove deforestation since the most recent observed date of presence. To model uncertainty in the date of deforestation we used a uniform distribution to select a date between the most recent date of observed mangrove presence and the oldest date of observed mangrove absence. Second, there is uncertainty in the relationship between the time since deforestation and proportion of mangrove carbon lost, quantified as the error present in the regression models (Supplementary Fig. 3a, b). We simulated the projected proportions of mangrove carbon remaining as a function of the length of time since deforestation (2016—date of deforestation), accounting for the error inherent in each linear model (Supplementary Methods 3).
As mangrove forests grow, they typically accumulate carbon in soil and tree biomass stocks, until reaching the value held by the reference community12,44. This process can be slow, taking from 20 to more than 50 years12,44. At a given point in time before the climax community is reached, the mangrove ecosystem contains a proportion of the value held in the climax community (at). The whole-ecosystem carbon accumulation curve for afforesting mangroves was estimated using data taken from a meta-analysis of blue carbon ecosystem restoration18, that we used to estimate the proportion of the reference ecosystem carbon accumulated following restoration (Supplementary Methods 4). The shape of the temporal relationship can be observed in Supplementary Fig. 3c. We used this general relationship describing restoration of all blue carbon ecosystems, because a mangrove forestation-specific meta-analysis is not currently available. To assess the impacts of this selection on the study findings, we also conducted a sensitivity analysis using data from two case studies of mangrove soil carbon and biomass accumulation in foresting mangroves (Supplementary Methods 2). The approximate date of mangrove forestation was quantified by cross-referencing several dates from the GMW dataset to establish the dates of presence and absence14,15,16. We cross-referenced the dates of 1996, 2007, 2010, and 2016 to identify the dates of mangrove presence and absence at each location.
Uncertainty in gains of carbon due to mangrove forestation was incorporated in two ways. First, there is uncertainty in the date of mangrove forestation since the most recent observed date of presence. To model uncertainty in the date of forestation we used a uniform distribution to select a date between the most recent date of observed mangrove absence and the oldest date of observed mangrove presence. Second, there is uncertainty in the relationship between the time since deforestation and proportion of mangrove carbon lost, quantified as the error present in the meta-analytic regression model (Supplementary Fig. 3c). We simulated the projected proportions of mangrove carbon remaining as a function of the length of time since deforestation (2016—date of deforestation), accounting for the error inherent in each linear model45 (Supplementary Methods 3).
We estimated D, F, rt, and at for each patch of mangrove gain and loss between 1996 and 2016. These data were then used to quantify four indicators of net change in mangrove carbon stocks, to evaluate the sensitivity of estimation to the inclusion or exclusion of afforestation and remnant carbon processes. The first indicator estimated the maximum carbon stock at risk of loss due to deforestation (D), following the approach used in the most recent global estimate of potential mangrove carbon emissions9. The second indicator estimated net loss of carbon assuming 100% carbon loss and gain rates (D − F). For the third indicator, we estimated the carbon stock loss due to deforestation but accounting for remnant carbon (Drt)8. Finally, the fourth indicator estimated net changes in mangrove carbon stock between 1996 and 2016 accounting for both forestation and proportional accumulation and loss rates of carbon following LCC (Drt − Fat). For mapping of spatial variability in net gains and losses of mangrove carbon stocks, we quantified the net change in mangrove carbon stock (Fat − Drt) by summarising all patches of mangrove gain and loss with their centroids located in cells across a global grid (Figs. 2 and 3).
Source: Ecology - nature.com
