s41561-025-01742-z.pdf

Nature Geoscience | Volume 18 | August 2025 | 761–768 768 Article https://doi.org/10.1038/s41561-025-01742-z 48. Sardanyés, J., Ivančić, F. & Vidiella, B. Identifying regime shifts, transients and late warning signals for proactive ecosystem management. Biol. Conserv. 290, 110433 (2024). 49. Stevens‐Rumann, C. S. et al. Evidence for declining forest resilience to wildfires under climate change. Ecol. Lett. 21, 243–252 (2018). 50. Bede‐Fazekas, Á. & Somodi, I. Precipitation and temperature timings underlying bioclimatic variables rearrange under climate change globally. Glob. Change Biol. 30, e17496 (2024). 51. Munang, R. et al. Climate change and ecosystem-based adaptation: a new pragmatic approach to buffering climate change impacts. Curr. Opin. Environ. Sustain. 5, 67–71 (2013). 52. Díaz, S. et al. Assessing nature’s contributions to people. Science 359, 270–272 (2018). 53. Drake, J. E. et al. Trees tolerate an extreme heatwave via sustained transpirational cooling and increased leaf thermal tolerance. Glob. Change Biol. 24, 2390–2402 (2018). 54. Liu, H. et al. Nature‐based framework for sustainable afforestation in global drylands under changing climate. Glob. Change Biol. 28, 2202–2220 (2022). 55. Kaiser-Bunbury, C. N. et al. Ecosystem restoration strengthens pollination network resilience and function. Nature 542, 223–227 (2017). 56. Vanbergen, A. J. & Initiative, T. I. P. Threats to an ecosystem service: pressures on pollinators. Front. Ecol. Environ. 11, 251–259 (2013). 57. Settele, J., Bishop, J. & Potts, S. G. Climate change impacts on pollination. Nat. Plants 2, 16092 (2016). 58. Staude, I. R. et al. Prioritize grassland restoration to bend the curve of biodiversity loss. Restor. Ecol. 31, e13931 (2023). Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/. © The Author(s) 2025 1MTA-SZTE ‘Momentum’ Applied Ecology Research Group, University of Szeged, Szeged, Hungary. 2HUN-REN-UD Functional and Restoration Ecology Research Group, Debrecen, Hungary. 3HUN-REN Centre for Agricultural Research, Institute for Soil Sciences, Budapest, Hungary. 4Institute of Ecology, School of Sustainability, Leuphana University Lüneburg, Lüneburg, Germany. 5Avignon Univ, Aix Marseille Univ, CNRS, IRD, IMBE, Avignon, France. 6Department of Genetics, Ecology and Evolution, Federal University of Minas Gerais, Belo Horizonte, Brazil. 7Taxonomy and Macroecology, Royal Botanic Garden Edinburgh, Edinburgh, UK. 8School of GeoSciences, University of Edinburgh, Edinburgh, UK. 9University of Debrecen, Department of Ecology, Debrecen, Hungary. 10Balaton Limnological Research Institute, Tihany, Hungary. 11Department of Ecology, University of Szeged, Szeged, Hungary. 12HUN-REN Centre for Ecological Research, Institute of Ecology and Botany, Vácrátót, Hungary. 13Department of Environmental and Landscape Geography, Institute of Geography and Earth Sciences, ELTE Eötvös Loránd University, Budapest, Hungary. 14These authors contributed equally:
Csaba Tölgyesi, Nándor Csikós.  e-mail: [email protected]

Nature Geoscience Article https://doi.org/10.1038/s41561-025-01742-z Methods Model construction Ecosystem restoration can have a large variety of goals and natural target vegetation states. For modelling purposes, we narrowed these down to four broad ecosystem types: forest, shrubland, grassland and wetland. These are also listed among the major land-cover categories of Earth in the Copernicus Global Land Service–Land Cover raster and available as fractional scores in pixels of 1-ha resolution (version 2019)59 (A in Fig. 1). Other land-cover categories in the database cor - respond to either non-target ecosystem types for classical restoration (cropland and built-up areas; but see Elmqvist et al. 60 and Klaus and Kiehl61 for urban restoration initiatives) or natural ecosystem types with
little vegetation and therefore minor carbon sequestration rates
(bare/sparse vegetation, snow and ice and moss and lichen). T o construct a predictive model for the potential global cover distribution of the four selected ecosystem types, we followed mod- elling principles in Bastin et al.9 For learning sites (that is, pixels with known fractional cover scores of target ecosystems and environmental predictors), we used their 78.850 pixels (100 m × 100 m) supplemented with 20.000 more pixels with non-zero fractional cover percentages for each ecosystem type (80.000 additional pixels in total) in protected areas worldwide (except Antarctica) (B in Fig. 1). These latter pixels were selected by first assigning every pixel in protected areas that contained at least one of the target ecosystem types to a pool of each ecosystem type (pixels could be allocated to more than one pool if they contained more than one target ecosystem type), then we randomly selected (without repetition) 20.000 pixels from each pool, leading to 158.850 pixels in total (Extended Data Fig. 6). Fractional cover scores were kept for subsequent model construc- tion, so pool assignment did not mean that a pixel was fully attributed to an ecosystem type. The four pools were needed to enable similar representations of each ecosystem type in the final set of learning sites. We used the World Database on Protected Areas62 for the delineation
of protected areas. We included all listed protected areas (for exam- ple, EU Natura 2000 areas) and did not restrict pixel selection to strict nature reserves, as that procedure would exclude sustainable
coexistence of humans with nature, such as extensive grassland man- agement or silvopastoral systems, which can also lead to high bio - diversity and co-benefits for climate change mitigation/adaptation and are often practical targets of restoration efforts63. Thus, our approach includes components of land sharing and land sparing. We allocated ten environmental predictors, including five climatic (WorldClim database64; 1 km × 1 km resolution), three edaphic (Soil - Grids250m65; 250 m × 250 m) and two topographic (GMTED2010 66; 250 m × 250 m) variables (Extended Data Fig. 9) to every learning site (Supplementary Table 1) (C in Fig. 1). These environmental variables proved to be reliable predictors in Bastin et al. 9. The differences in the spatial resolution of the variables were handled with the bilinear resampling technique of ESRI ArcMap 10.8. The bilinear resampling technique involves attributing the average value of the four nearest pixels within a 2 × 2 window to the relevant output pixel. This process follows a bilinear mathematical function along both the horizontal and vertical axes and is the commonly applied method for smoothly transitioning continuous datasets lacking clear boundaries67. We used Random Forest machine learning regression models with fivefold cross-validations68 for training 500 trees of (potentially) unlimited depth (D in Fig. 1). The number of variables to possibly split at each node was set to three (that is, the default value for ten predictors). The models were evaluated by regressing observed values by predicted values (both of which were fractional) 69. Spatial distribution of the model uncertainty was assessed by the standard deviation of the predic- tions made by the five trained sub-models originated from the fivefold cross-validation. According to the regression-based evaluation, the models of forest, shrubland and grassland had high predictive power (0.70 < R2 < 0.82), although at high (above 60%) observed shrubland cover scores, the predictions tended to yield underestimations. This was due to the low prevalence of such pixels in the datasets (approxi- mately 0.1%), but because they are rare, uncertainties in their range do not affect the overall predicted amount of shrubland (Extended Data Fig. 7). The model for wetlands was fair (R2 = 0.32), with uncertainties also caused by the rarity of high empirical cover scores. Models’ uncer- tainty, estimated using standard deviations, was evenly distributed across continents and biogeographic regions (Extended Data Fig. 8). Predicting restoration targets We predicted the potential forest, shrubland, grassland and wetland cover distributions as the ensemble mean of the five trained sub-models (according to the fivefold cross-validation) of each of the four ecosys- tem models to a global grid of 1 × 1 km cells (henceforth ‘modelling sites’) (E in Fig. 1). We also made predictions for a future period to take into account the effect of changing climate on the potential cover of ecosystem types, by updating the climatic predictors but leaving the non-climatic predictors unchanged. We chose the period 2061–2080 because it is in the middle of our study period (between 2030 and 2100). From among the global climate models of the Coupled Model Intercom- parison Project Phase 6 (CMIP6)70 having predictions in the WorldClim database64, we selected EC-Earth3-Veg model71, which has a medium (that is, 4.33) effective climate sensitivity value72 and its predecessor (that is, EC-Earth) has been successfully used for carbon sequestration prediction73. Among the Shared Socioeconomic Pathways (SSPs), we considered SSP1–2.6, SSP2–4.5, SSP3–7.0 and SSP5–8.5. The differences in the spatial resolution of the variables were handled with the bilinear resampling technique of ESRI ArcMap 10.8. Due to the independent construction of the four ecosystem models, the predicted total cover could exceed 100%. In such cases, we pro - portionately reduced the cover of each ecosystem type to sum up to 100% (for example, a prediction of 60% forest and 60% grassland was reduced to 50% each). Once having the potential distribution of target ecosystems
(in percentage fractions for every modelling site), we compared them to the present cover of the target ecosystems using the Copernicus Global Land Service–Land Cover raster (version 2019)59. We used the difference to identify degraded areas globally (F in Fig. 1). However, built-up areas are rarely available for restoration pur- poses, so we removed them (their fractional cover from the modelling sites) from subsequent calculations. Likewise, restoration activities cannot be implemented in all agricultural areas due to their role in food security. Therefore, we identified intensive agricultural areas using the Land Use Systems of the World 1.1. database74 and removed all corresponding modelling sites (G in Fig. 1). Fractional removal was not possible due to the data structure of the database. We considered the following categories as intensive agriculture: rain-fed crops, crops with moderate or higher livestock density, crops with high livestock density and large-scale irrigation agriculture. It was common that a currently natural ecosystem type in a parti- cular location is not the potential type in the future predictions, indicating that state transitions are expected worldwide in the near future. These can take place spontaneously but can also be assisted by landscape and conservation management (for example, to prevent catastrophic, pyric transitions from forests to savannahs by allowing a smooth transition with management). We did not lump them with restoration but handled them separately as ‘state transitions’ , with additional (often negative) effects on carbon sequestration rates. We finally had five potential cover values (one based on current and four based on future climates) for every target ecosystem type in all modelling sites, covering the terrestrial surface of Earth that is available for restoration (or prone to state transitions). Our model did not identify specific mosaic ecosystems (for example, savannah-forest mosaics) but handled them adequately, because the ecosystem-type cover values were proportional, meaning that such a modelling site, Nature Geoscience Article https://doi.org/10.1038/s41561-025-01742-z for instance, contained certain proportions of forest, shrubland and grassland alike. These constituting ecosystem types were then used separately in subsequent carbon gain calculations. Predicting carbon gain potential As the next step, we assigned carbon sequestration rates (t ha −1 yr−1) to the modelling sites. Carbon sequestration rates of a target eco - system type can vary greatly among biomes (for example, tropical rainforest vs boreal forest). T o account for this heterogeneity, we first removed all modelling sites located in polar or arid biomes using the ‘Resolve EcoRegions2017’ database 75 (H in Fig. 1 ), because these cli - mates allow very low primary productivity and therefore low carbon sequestration rates. Then, we combined the four target ecosystem types with the remaining biomes listed in the database, resulting in 12 biome-specific ecosystem types: boreal forest, temperate forest, tropical/subtropical dry forest, tropical/subtropical moist forest, mangrove forest/shrubland, boreal/temperate shrubland (including heathland), tropical/subtropical shrubland, boreal/temperate grass- land, tropical/subtropical grassland, boreal wetland (predominantly peatland), temperate wetland (including salt marshes) and tropical/ subtropical wetland (I in Fig. 1 ). For boreal, temperate, tropical/sub- tropical dry forest and tropical/subtropical moist forests, we adopted the sequestration rates from Cook-Patton et al.76, who compiled and summarized over ten thousand records from regenerating forests. They provided ecoregion-specific rates, which we averaged across our broader biome-specific forest ecosystem types. For open ecosystems, we heavily relied on the database of the European Environmental Agency77. For biome-specific ecosystem types that were not included in any of these databases or had a poor coverage in the European Envi- ronmental Agency’s database, we searched for published literature records that considered total ecosystem carbon sequestration rates, including both above- and belowground biomass and soil organic carbon, except in tropical/subtropical grasslands, where frequent fires (and herbivory) preclude the formation of a stable aboveground carbon sink. So in their case, we used only records from belowground compartments (Extended Data Fig. 9 and J in Fig. 1). Once having the carbon sequestration rates of the 12 biome-specific ecosystem types, we calculated the carbon sequestration rate of
each modelling site after the implementation of restoration or the com- bination of restoration and state transitions in future-climate-based predictions (K in Fig. 1). T o get the carbon gain potential of modelling sites, we calculated their current, pre-restoration sequestration rate (L in Fig. 1) and subtracted this from the post-restoration rate (plus state transition, where applicable) (M in Fig. 1). We assumed 0 t C ha−1 seques- tration to non-target ecosystem types (for example, bare soil surface and snow and ice). In a few rare cases we predicted negative carbon gain based on current climates, which can happen if, for example, current forest cover is locally higher for some reason than predicted by our model. We ignored these rare cases but negative rates owing to state transitions in future-climate-based predictions were retained. We used carbon gain values based on current climates to estimate the impact of ecosystem restoration on carbon emissions projected by Shared Socioeconomic Pathways (SSP1–2.6, SSP2–4.5, SSP3–7.0, SSP5–8.5) between 2030 and 2100 (SSP Public Database 2.046). Besides the global sequestration potential until 2100, we also identified priority regions for restoration where the highest carbon gain can be achieved first. Selecting the modelling sites with the top carbon gain values is a possible option, but they may be scattered across the world and due to their relatively small size are not especially suitable for planning global restoration strategies. Therefore, we used a coarser global grid of 100 × 100 km cells (22.880 cells in total). We ranked these large grid cells according to their total potential carbon gain as the summed values from the included modelling sites and set the cut-off at that rank above which the cells contained 20% of all restorable land area, that is, the 80th percentile. This way, we identified regions where the highest carbon sequestration can be achieved until 2030, which was set as a milestone year to restore the first 20% of the available land. This selection method yielded 2,669 priority regions (corresponding to 11.7% of all grid cells). We performed prioritization using only the present-climate-based predictions, because of the very low overall carbon gains predicted when future climates were considered. Data preparation and model construction were done in ESRI
ArcMap 10.8 and R 4.2.0 statistical software, respectively. We used
the R packages ‘groupdata2’ , ‘ranger’ , ‘raster’ and ‘sf’ for model con- struction and predictions. Limitations of the model Using carbon sequestration rates instead of the carbon stocks of mature stands is a major improvement of our model compared to previ- ous ones, but we used linear functions, although deviations from this can happen during stand maturation in some ecosystem types78. The lack of global data across different biomes precluded us from taking into account these nonlinearities. Our input dataset structure allowed us to consider restoration actions that entailed a change in vegetation or ecosystem type through the restoration, although this is not a pre- requisite for ecosystem restoration and for carbon gain (for example, restoration of intensively managed grasslands to natural grasslands or tree plantations to forests, although these actions are sometimes difficult to distinguish from improved management). However, using proportional ecosystem type cover values alleviated this deficiency in mosaic ecosystems if the degradation manifested in changes in the proportions of constituting ecosystem types; so, for example, our model included the restoration of degraded savannah-forest mosaics to pristine ones with the original ecosystem-type cover proportions. Furthermore, future changes of some input data and other relevant trends (for example, changes in the area of intensive agriculture, biome boundary shifts and the effects of elevating CO2 and changing climate in general on assimilation, decomposition, fire frequency) were not taken into account due to high uncertainties, but this also gives room for future research. Lastly, the 1-km2 model resolution does not mean that the output maps are directly usable for local-scale restoration planning because the predictors are often extrapolated values that are not sensitive enough to fine-scale environmental heterogeneity (topography, soil and so on), which cannot be ignored in small-scale planning. Data availability Data used in this study are available in Extended Data Fig. 9 and Extended Data Table 1, Supplementary Data Table 1 and in the free databases cited. Processed data and models are available via Dryad at https://doi.org/10.5061/dryad.ksn02v7g4 (ref. 79). Source data are provided with this paper. Code availability R codes to prepare the models are available via Dryad at https://doi.org/ 10.5061/dryad.ksn02v7g4 (ref. 79). References 59. Buchhorn, M. et al. Copernicus Global Land Service: land
cover 100 m: collection 3: epoch 2019: globe (V3.0.1). Zenodo https://doi.org/10.5281/zenodo.3939038 (2020). 60. Elmqvist, T. et al. Benefits of restoring ecosystem services in urban areas. Current Opinion in. Environ. Sustain. 14, 101–108 (2015). 61. Klaus, V. H. & Kiehl, K. A conceptual framework for urban ecological restoration and rehabilitation. Basic Appl. Ecol. 52, 82–94 (2021). 62. UNEP-WCMC and IUCN Protected Planet: The World Database
on Protected Areas (WDPA) (UNEP-WCMC, 2019); www.protectedplanet.net Nature Geoscience Article https://doi.org/10.1038/s41561-025-01742-z 63. Lyons, K. G. et al. Challenges and opportunities for grassland restoration: a global perspective of best practices in the era of climate change. Glob. Ecol. Conserv. 46, e02612 (2023). 64. Fick, S. E. & Hijmans, R. J. WorldClim 2: new 1-km spatial resolution climate surfaces for global land areas. Int. J. Climatol. 37, 4302–4315 (2017). 65. Hengl, T. et al. SoilGrids250m: global gridded soil information based on machine learning. PLoS ONEs 12, e0169748 (2017). 66. Danielson, J. J. & Gesch, D. B. Global Multi-Resolution Terrain Elevation Data 2010 (GMTED2010) Report No. OFR 2011-1073
(US Geological Survey, 2011). 67. Baboo, S. S. & Devi, M. R. An analysis of different resampling methods in Coimbatore, District. Glob. J. Computer Sci. Technol. 10, 61–66 (2010). 68. Breiman, L. Random forests. Mach. Learn. 45, 5–32 (2001). 69. Piñeiro, G., Perelman, S., Guerschman, J. P. & Paruelo, J. M. How to evaluate models: observed vs. predicted or predicted vs. observed? Ecol. Modell. 216, 316–322 (2008). 70. Eyring, V. et al. Overview of the Coupled Model Intercomparison Project Phase 6 (CMIP6) experimental design and organization. Geosci. Model Dev. 9, 1937–1958 (2016). 71. Döscher, R. et al. The EC-Earth3 earth system model for the coupled model intercomparison project 6. Geosci. Model Dev. 15, 2973–3020 (2022). 72. Zelinka, M. D. et al. Causes of higher climate sensitivity in CMIP6 models. Geophys. Res. Lett. 47, e2019GL085782 (2020). 73. Pechanec, V. et al. Modelling of the carbon sequestration and its prediction under climate change. Ecol. Inform. 47, 50–54 (2018). 74. Land Use Systems of the World (FAO, 2010); https://data.apps.fao. org/map/catalog/srv/eng/catalog.search#/metadata/fc32c5de- 440c-46aa-9cad-81f4c8b84c6a 75. Dinerstein, E. et al. An ecoregion-based approach to protecting half the terrestrial realm. BioScience 67, 534–545 (2017). 76. Cook-Patton, S. C. et al. Mapping carbon accumulation potential from global natural forest regrowth. Nature 585, 545–550 (2020). 77. Carbon storage in EU terrestrial and marine ecosystems —
European Environment Agency (EEA, 2022); https://www.eea. europa.eu/data-and-maps/data/carbon-storage-in-global-terrestrial 78. Zhou, T. et al. Age‐dependent forest carbon sink: estimation via inverse modeling. J. Geophys. Res.: Biogeosci. 120, 2473–2492 (2015). 79. Tölgyesi, C. et al. Global ecosystem restoration has unexpectedly low potential to mitigate climate change. Dryad https://doi.org/ 10.5061/dryad.ksn02v7g4 (2025). 80. Page, S. E. et al. A record of Late Pleistocene and Holocene carbon accumulation and climate change from an equatorial peat bog (Kalimantan, Indonesia): implications for past, present and future carbon dynamics. J. Quat. Sci. 19, 625–635 (2004). 81. Saunders, M. J., Jones, M. B. & Kansiime, F. Carbon and water cycles in tropical papyrus wetlands. Wetlands Ecol. Manage. 15, 489–498 (2007). 82. Bernal, B. & Mitsch, W. J. A comparison of soil carbon pools and profiles in wetlands in Costa Rica and Ohio. Ecol. Eng. 34, 311–323 (2008). 83. Lü, X. T., Yin, J. X., Jepsen, M. R. & Tang, J. W. Ecosystem carbon storage and partitioning in a tropical seasonal forest in Southwestern China. For. Ecol. Manage. 260, 1798–1803 (2010). 84. Adame, M. F. et al. Carbon stocks and soil sequestration rates of tropical riverine wetlands. Biogeosciences 12, 3805–3818 (2015). 85. Akpa, S. I., Odeh, I. O., Bishop, T. F., Hartemink, A. E. & Amapu, I. Y.
Total soil organic carbon and carbon sequestration potential in Nigeria. Geoderma 271, 202–215 (2016). 86. Hribljan, J. A., Suárez, E., Heckman, K. A., Lilleskov, E. A. & Chimner, R. A. Peatland carbon stocks and accumulation rates in the Ecuadorian páramo. Wetland Ecol. Manage. 24, 113–127 (2016). 87. Kolka, R. K., Murdiyarso, D., Kauffman, J. B. & Birdsey, R. A. Tropical wetlands, climate, and land-use change: adaptation and mitigation opportunities. Wetland Ecol. Manage. 24, 107–112 (2016). 88. Davila, A. & Bohlen, P. J. Hydro-ecological controls on soil carbon storage in subtropical freshwater depressional wetlands. Wetlands 41, 66 (2021). 89. Sjögersten, S. et al. Coastal wetland ecosystems deliver large carbon stocks in tropical Mexico. Geoderma 403, 115173 (2021). 90. Black, T. A. et al. Annual cycles of water vapor and carbon dioxide fluxes in and above a boreal aspen forest. Glob. Change Biol. 2, 219–229 (1996). 91. Frolking, S. et al. Modelling temporal variability in the carbon balance of a spruce/moss boreal forest. Glob. Change Biol. 2, 343–366 (1996). 92. Yarie, J. & Billings, S. Carbon balance of the taiga forest within Alaska: present and future. Can. J. For. Res. 32, 757–767 (2002). 93. Röser, C. et al. Net CO2 exchange rates in three different successional stages of the ‘Dark Taiga’ of central Siberia. Tellus B: Chem. Phys. Meteorol. 54, 642–654 (2002). 94. Suni, T. et al. Vesala, Interannual variability and timing of growing‐ season CO2 exchange in a boreal forest. J. Geophys. Res.: Atmos. 108, D9 (2003). 95. Curtis, P. S. Biometric and eddy-covariance based estimates of annual carbon storage in five eastern North American deciduous forests. Agric. For. Meteorol. 113, 3–19 (2002). 96. Urrutia-Jalabert, R., Malhi, Y. & Lara, A. The oldest, slowest rainforests in the world? Massive biomass and slow carbon dynamics of Fitzroya cupressoides temperate forests in southern Chile. PLoS ONE 10, e0137569 (2015). 97. Parada, T., Lusk, C. H. & Donoso, P. J. Evidence that emergent Nothofagus dombeyi do not depress carbon sequestration rates of canopy species in an old-growth Chilean temperate forest.
N.Z. J. Bot. 56, 311–322 (2018). 98. Gough, C. M., Vogel, C. S., Harrold, K. H., George, K. & Curtis, P. S. The legacy of harvest and fire on ecosystem carbon storage in a north temperate forest. Glob. Change Biol. 13, 1935–1949 (2007). 99. Michelsen, A., Andersson, M., Jensen, M., Kjøller, A. & Gashew, M. Carbon stocks, soil respiration and microbial biomass in fire-prone tropical grassland, woodland and forest ecosystems. Soil Biol. Biochem. 36, 1707–1717 (2004). 100. Vargas, R., Allen, M. F. & Allen, E. B. Biomass and carbon accumulation in a fire chronosequence of a seasonally dry tropical forest. Glob. Change Biol. 14, 109–124 (2008). 101. Coetsee, C., Gray, E. F., Wakeling, J., Wigley, B. J. & Bond, W. J. Low gains in ecosystem carbon with woody plant encroachment in a South African savanna. J. Trop. Ecol. 29, 49–60 (2013). 102. Cao, S., Sanchez-Azofeifa, G. A., Duran, S. M. & Calvo-Rodriguez, S.
Estimation of aboveground net primary productivity in secondary tropical dry forests using the Carnegie–Ames–Stanford approach (CASA) model. Environ. Res. Lett. 11, 075004 (2016). 103. Pereira Júnior, L. R. Carbon stocks in a tropical dry forest in Brazil. Rev. Cienc. Agron. 47, 32–40 (2016). 104. Abreu et al. The biodiversity cost of carbon sequestration in tropical savanna. Sci. Adv. 3, e1701284 (2017). 105. Pelletier, J. et al. Carbon sink despite large deforestation in African tropical dry forests (miombo woodlands). Environ. Res. Lett. 13, 094017 (2018). 106. Calvo-Rodriguez, S., Sanchez-Azofeifa, G. A., Duran, S. M., Do Espirito-Santo, M. M. & Ferreira Nunes, Y. R. Dynamics of carbon accumulation in tropical dry forests under climate change extremes. Forests 12, 106 (2021). 107. Maia, V. A. et al. The carbon sink of tropical seasonal forests in southeastern Brazil can be under threat. Sci. Adv. 6, eabd4548 (2020).