Land use-induced soil carbon loss in the dry tropics nearly offsets gains in northern lands

Abstract

Soil carbon changes are difficult to measure globally, and global models are poorly constrained. Here, we propose a framework to map annual changes in soil carbon and litter (SOCL) as the difference between the net land CO₂ flux from atmospheric inversions and satellite-based maps of biomass changes. We show that SOCL accumulated globally at a rate of about 0.34 ± 0.30 ( ± 1 sigma) billion tonnes of carbon per year (PgC yr⁻¹) during 2011-2020. The largest SOCL sink is found in boreal regions (0.93 ± 0.45 PgC yr⁻¹ in total) particularly in undisturbed peatlands and managed forests. The largest losses occur in the dry tropics (−0.50 ± 0.47 PgC yr⁻¹) and correspond with agricultural expansion from land use change, cropland management and grazing. By contrast, forests in the wet tropics act as a net soil carbon sink (0.32 ± 0.35 PgC yr⁻¹). Our findings highlight the large mitigation opportunities in the dry tropics to restore agricultural soil carbon.

Soil carbon loss from human agriculture-related land use in the dry tropics largely offsets gains in northern lands, leading to a small net sink of 0.34 ± 0.30 PgC yr⁻¹ at a global scale.

Introduction

Global changes in fossil fuel and cement CO₂ emissions, the atmospheric CO₂ growth rate and net land carbon fluxes have been reasonably quantified in global carbon budget analyses^1–3. However, more accurate partitioning of the global land carbon sink into biomass and soil carbon (SOC) changes is still needed⁴. Soils represent the largest terrestrial carbon pool⁵, but a substantial fraction of it, like deep peat deposits⁶, does not actively exchange carbon with the atmosphere. Given the size of the soil carbon pool, even if a small fraction is lost, it could result in a strong decline of the current land carbon sink⁷. Earth System Model projections of soil carbon changes are very uncertain, ranging from a large source to a sustained sink in the future⁸. Mapping current soil carbon changes and attributing them to land use versus other environmental drivers is needed to evaluate these models for the current period and to reduce the uncertainty of their projections.

Soil carbon pools are no longer in a steady state due to climate change affecting microbial decomposition and human land use and land management activities⁹. Despite many local field experiments, regional inventories and meta-analysis measuring the impact of soil warming and land management on soil carbon^10–12, there is no consensus on the magnitude of regional and global SOC changes. Further, fine litter and coarse woody litter are important carbon pools that are not reported in carbon budget assessments. The balance of carbon in soil and litter pools is determined by gains from plant inputs, losses and gains from erosion and re-deposition, losses from microbial decomposition, losses from combustion, and leaching of soil carbon to rivers^13–15. The dynamic global vegetation models (DGVMs) used for the global carbon budget¹ show a large spread in global SOC stock changes, ranging from −0.58 PgC yr⁻¹ to 1.63 PgC yr⁻¹ during 2011–2020. This spread arises from model limitations and lack of validation data for simulating the processes affecting soil carbon inputs and decomposition, i.e., the vertical distribution of plant inputs, microbial decomposition processes, stabilization of organic matter on soil minerals, and the complex impacts of land-use and land management changes^5,16–18. Further, models are not validated in terms of temporal changes of soil carbon due to a lack of observed time series¹⁹. On the other hand, regional inventories^20,21 and in situ observations^22,23 cannot easily be scaled up to a global budget of soil carbon change, given the huge spatial heterogeneity of SOC. This study represents a first attempt to map soil and fine litter carbon changes using an observation-based framework that incorporates atmospheric inversions, satellite-based biomass changes and an empirical model of carbon changes in coarse woody debris, which have separate dynamics from soil carbon. With uncertainties being estimated by using multi-model estimates, our observation-based approach provides key data to constrain models and regional carbon budget assessments.

We infer changes in SOC and fine litter, hereafter referred to as ∆SOCL, using a mass balance method²⁴ based on the net carbon exchanges between land and the atmosphere from atmospheric data and additional constraints. This framework measures SOCL changes across the entire soil profile, from surface to bedrock. That is, ∆SOCL is calculated by the difference between maps of the net land-atmosphere CO₂ flux from atmospheric inversions minus maps of CO₂ fluxes from lateral processes that do not lead to carbon storage, minus satellite-based maps of biomass changes (ΔB), minus maps of modeled carbon changes in coarse woody debris (∆CWD) (Fig. 1) as given by Eq. 1.

$$\mathrm{\Delta }SOCL=NE{E}_{\text{inversion}}-L-\mathrm{\Delta }B-\mathrm{\Delta }CWD=\mathrm{\Delta }{C}_{tot}-\mathrm{\Delta }B-\mathrm{\Delta }CWD$$ 1

where NEE _inversion is the net land-atmosphere CO₂ flux excluding fossil emissions with fire-related fluxes being part of the total land-atmosphere flux embodied in NEE _inversion, following the Global Carbon Project definition used for the RECCAP2 regional carbon budget project²⁴. The sign convention of NEE and other terms in Eq. 1 is that a carbon flux is positive if CO₂ is removed from the atmosphere or carbon accumulates in biomass, CWD or SOCL pools. L is the land-atmosphere net CO₂ flux from lateral carbon displacement processes that do not lead to a change in terrestrial carbon stocks, arising from the transfer of carbon in rivers and inland waters¹⁵ and the harvest, displacement and oxidation of crop and wood products²⁵, ΔB is the carbon stock changes in above- and belowground biomass, ΔCWD is the carbon stock changes in coarse woody debris defined as dead woody material with diameters larger than 10 cm²⁶ (see “Methods”), and ΔC_tot is the land carbon stock change once CO₂ fluxes from lateral processes are subtracted from inversions (spatial maps of NEE and lateral fluxes are shown in Supplementary Fig. 1).

Fig. 1. Carbon flux components of the terrestrial ecosystem.

Overview of carbon flux components (ΔSOCL: changes in soil carbon and litter, ΔC_tot: carbon stock changes in land carbon uptake corrected by CO₂ fluxes from lateral processes, ΔB: changes in above- and belowground biomass, ΔCWD: changes in coarse woody debris).

Five long-term atmospheric inversion models assimilating atmospheric stations CO₂ concentration data from ref. ¹ provided NEE _inversion (see “Methods”). Satellite-based passive microwave vegetation optical depth (L-VOD)²⁷ provided maps of ΔB (see “Methods”). L-VOD is characterized by a high penetration through the forest canopy and a low saturation in the dense forests^28,29, and has proven valuable to monitor interannual changes in aboveground biomass carbon stocks^30–34. ΔB was calculated from L-VOD aboveground biomass using below-to-above ground biomass ratios³⁵, which has proven to be consistent with other ΔB products⁴. Another map of ΔB obtained from multiple satellite datasets³⁶ was used as a sensitivity test. Changes in coarse woody debris, ΔCWD, result from inputs by tree mortality, which lays dead stems and branches to the ground and outputs from the slow decomposition of CWD by insects and microbes²⁶. Maps of carbon change from CWD were obtained from an empirical model simulating mortality-induced CWD inputs^3,34 using L-VOD biomass changes and woody productivity, CWD decomposition rates and fire-induced CWD combustion (see “Methods”). Uncertainties (1-sigma) of ΔSOCL in each grid cell of 1° × 1° were calculated by error propagation of the different terms of Eq. 1 with an ensemble of 45 combinations (see “Methods”). Given the substantial inter-model differences among atmospheric inversion, we used the median of the 45 combinations as the most robust indicator of the sign (direction) and magnitude of ΔSOCL. This approach assumes that while individual model predictions may vary due to differences in assumptions and input data, the median provides a central estimate that is less sensitive to outliers. In addition, a Monte-Carlo approach was used to quantify the relative contribution to uncertainty and identify key drivers of the ∆SOCL variability by using the standardized regression coefficient (SRC) method (Supplementary Table 1). The Monte-Carlo method randomly samples (n = 100,000) different estimates of each component flux, assuming equal probability for each estimate, and incorporating their Gaussian uncertainties. The resulting global map of ΔSOCL and its uncertainty at 1° by 1° is shown in Fig. 2a (median of the 45 combinations). We assessed regional carbon budgets and used additional datasets to attribute SOCL gains and losses to land use changes and environmental factors.

Fig. 2. Spatial pattern of time-averaged terrestrial carbon fluxes (2011–2020).

a Carbon stock changes in soil carbon and litter, ΔSOCL. b Changes in biomass, ΔB. c Changes in coarse woody debris, ΔCWD. d Changes in land carbon uptake, ΔC_tot. Sign convention: a negative sign indicates carbon loss to the atmosphere, while a positive sign indicates a carbon sink. The black dots in (a–d) mean high agreement (>60%) across diverse estimates regarding the sign. The latitudinal fluxes were smoothed by a 10° smoothing window, and the shading suggests the standard deviation among diverse estimates. The numbers in (a–c) suggest the latitudinal Spearman correlation (after smoothing) between each carbon flux and ΔC_tot, ***indicates P < 0.001. The boundary of the wet tropics is drawn in green and the boundary of the dry tropics is drawn in black in (a). The boundaries of wet and dry tropics are defined from the Koppen-Geiger climate classification maps¹⁰⁰, i.e., Af and Am for the wet tropics and Aw for the dry tropics. Areas of desert and no observations of ΔB due to radio frequency interference effects on microwave remote sensing retrievals have been masked in the analysis throughout the study, and the distinction of missing pixels (11.2% of all the pixels) and areas of deserts can be seen in Supplementary Fig. 10d.

Results

Global budget of changes in soil carbon and litter

We found a global SOCL sink of 0.34 ± 0.30 PgC yr⁻¹ (±1 sigma) during 2011–2020 (Fig. 2 and Table 1) with atmospheric inversion fluxes being the dominant source of uncertainty based on our Monte-Carlo analysis (Supplementary Table 1). This soil carbon sink is smaller than the median of the DGVMs used for the global carbon budget (TRENDY-v11) (0.44 ± 0.53 PgC yr⁻¹)¹, and has a smaller uncertainty. It is similar to the predictions of a machine learning-based simulation of soil carbon changes trained using local measurements (0.27 PgC yr⁻¹)³⁷ (Supplementary Table 2). Using the alternative ΔB map from ref. ³⁶ instead of the one derived from L-VOD gives a comparable ΔSOCL sink (0.35 ± 0.30 PgC yr⁻¹; Supplementary Table 2), and a consistent spatial pattern with our map (R² = 0.47, P < 0.001; see Supplementary Fig. 2). Our estimates of changes in SOCL in managed forests show a strong spatial correlation (R² = 0.8, P < 0.001) with national forest inventories reports across countries, yet with a higher soil carbon sink than estimated by inventories (note that some countries report soil carbon changes of zero, by default, see “Methods”; Supplementary Fig. 3). Across European regions²⁰, both our results and inventories show larger soil carbon sinks in grassland than in cropland (Supplementary Fig. 4), aligning with previous evidence that croplands show carbon losses or marginal gains³⁸. However, results from the machine learning model of ref. ³⁷ and DGVMs (Supplementary Fig. 4) show stronger carbon sinks in croplands, possibly because these approaches ignore the effect of cropland management practices and some DGVM models wrongly assume that biomass harvested from cropland returns to the soil³⁹. In the DGVMs¹, changes of litter include CWD, which are modeled as part of a total litter pool and represent only a small sink of 0.07 ± 0.12 PgC yr⁻¹ (Supplementary Fig. 5), contributing about 15% of changes in SOCL (Supplementary Table 3). In contrast, forest inventories suggest that litter is a sink that contributes about 26% of the changes in SOCL (Supplementary Table 3)⁴⁰. In our calculations, we found that CWD changes are of smaller magnitude than SOCL (Table 1).

Table 1.

Carbon budgets of different regions during 2011–2020 (PgC yr⁻¹)

Regions	ΔSOCL	ΔB	ΔCWD	ΔCtot	ΔSOCL/ΔCtot
Global	0.34 ± 0.3	0.53 ± 0.16	0.05 ± 0.03	0.92 ± 0.31	37.0% ± 34.9%
North America	0.08 ± 0.37	0.20 ± 0.01	−0.01 ± 0.01	0.27 ± 0.42	29.6% ± 144.6%
South America	−0.35 ± 0.42	0.08 ± 0.04	0.01 ± 0.01	−0.26 ± 0.47	134.6% ± 292.1%
Europe	0.11 ± 0.07	0.02 ± 0	0 ± 0	0.13 ± 0.08	84.6% ± 74.9%
Africa	−0.34 ± 0.14	0.14 ± 0.05	−0.01 ± 0.01	−0.21 ± 0.15	161.9% ± 133.5%
Russia	0.60 ± 0.14	−0.08 ± 0.05	0.06 ± 0.03	0.58 ± 0.15	103.4% ± 36.0%
West Asia	−0.02 ± 0.06	0.04 ± 0.01	0 ± 0	0.02 ± 0.06	−100.0% ± −424.3%
East Asia	−0.01 ± 0.13	0.11 ± 0.01	0 ± 0	0.1 ± 0.14	−10.0% ± −130.8%
South Asia	−0.03 ± 0.17	0.08 ± 0.02	0 ± 0	0.05 ± 0.2	−60.0% ± −416.2%
Southeast Asia	0.11 ± 0.28	0.03 ± 0.01	0.02 ± 0	0.16 ± 0.33	68.8% ± 225.2%
Australasia	0.19 ± 0.18	−0.09 ± 0.02	−0.02 ± 0	0.08 ± 0.2	237.5% ± 635.0%

The regions are defined as Regions from the REgional Carbon Cycle Assessment and Process (RECCAP2) project, see Supplementary Fig. 10a for the boundary⁹⁹. A 0 in the table means that the carbon fluxes are negligible in that region. The uncertainty is estimated by the standard deviation across diverse estimates. ΔSOCL carbon stock changes in soil carbon and litter, ΔB changes in living biomass, ΔCWD changes in coarse woody debris, ΔCtot net changes in land carbon stocks. ΔSOCL/ΔCtot indicates net contribution of ΔSOCL to ΔCtot.

The spatial distribution of ΔSOCL (Fig. 2a) shows a strong spatial covariance with ΔC_tot (Fig. 2d). Globally, ΔSOCL (0.34 ± 0.3 PgC yr⁻¹) accounts for 37% to ΔCtot (0.92 ± 0.31 PgC yr⁻¹, Table 1). At the regional scale, the contribution of ΔSOCL to ΔCtot ranges from −100% to 238%, with large percentage values being in regions where ΔCtot is close to zero. The largest SOCL sink is observed in northern ecosystems (Fig. 2a, Supplementary Fig. 6), over Russia (0.60 ± 0.14 PgC yr⁻¹), Europe (0.11 ± 0.07 PgC yr⁻¹) and eastern North America (Table 1). In these ecosystems, soils have a larger adsorption capacity of fine-scale organic matter on mineral particles, and the rates of respiration are small in cold regions, favoring carbon accumulation^41,42. Further, northern plant productivity has been increasing over time⁴³, sustaining increased leaf and root litter carbon inputs to soils. Locally, SOCL gains are larger (0.23 ± 0.08PgC yr⁻¹) where the boreal and arctic peatland fraction is higher than 50% (Supplementary Fig. 7, see peatland maps in Supplementary Fig. 8), as plant inputs exceed decomposition due to anoxic conditions and low PH in undisturbed peatlands⁴⁴. In particular, large rates of soil C accumulation in Fig. 2a are found in peatland-dominated regions of the lowland Hudson Bay (0.08 ± 0.05 PgC yr⁻¹), the Ob River (0.11 ± 0.06 PgC yr⁻¹), and the Lena River (0.18 ± 0.05 PgC yr⁻¹) (Fig. 2a, Supplementary Fig. 9).

In tropical regions, soils in rainforests of the Amazon and central Africa are a sink but soils of drier ecosystems, including the Brazilian Cerrado, African drylands and savannas are a carbon source (Fig. 2a). Africa and South America show large SOCL losses of 0.34 ± 0.14 PgC yr⁻¹ and 0.35 ± 0.42 PgC yr⁻¹, respectively, which offset the carbon sinks of Russia and Europe (Table 1). Our estimate for total gross tropical soil carbon loss of 1.03 ± 0.38 PgC yr⁻¹ (Fig. 2a, Supplementary Table 4) can be compared with bottom-up estimates documenting different carbon loss processes. Grazing in tropical rangelands is estimated to account for a loss of 0.39 PgC yr⁻¹ (34.5%)⁴⁵; tropical peat fires in Equatorial Asia for a loss of 0.07 PgC yr⁻¹ (6.2%)⁴⁶; farmland erosion for a loss of 0.02 and 0.04 PgC yr⁻¹ in Africa and South America, respectively⁴⁷, explaining 5.3% of our SOCL loss in the tropics. In addition, forest fires, which are underestimated by current coarse resolution burned area datasets^48,49, may contribute to additional losses which remain unquantified. At face value, declining burned area in African savanna regions may result in a small soil carbon sink of 0.038 PgC yr⁻¹ according to ref. ⁵⁰. More insights on the contribution of land use change are given in the following section.

Attribution of SOCL changes to land use changes vs. other environmental drivers

In order to evaluate the impacts of land use change (LUC) on ΔSOCL (Fig. 3a), we matched high-resolution satellite maps of ref. ⁵¹ of forest area loss attributed to commodity-driven deforestation and smallholder agriculture (Supplementary Fig. 10b) with our ΔSOCL maps. We attributed ΔSOCL to either commodity-driven or smallholder farmland expansion (LUC) if their respective fractions exceeded 50% in each 1° grid cell. We found that the LUC-attributed soil carbon losses in the dry tropics are 0.25 ± 0.32 PgC yr⁻¹ and that they explain 50% of the net SOCL losses in this region (Fig. 3a). The grid cells dominated by the expansion of smallholder agriculture are associated with larger SOCL losses (−0.21 ± 0.29 PgC yr⁻¹) than those dominated by commodity-driven deforestation (−0.04 ± 0.14 PgC yr⁻¹, Fig. 3a). Larger soil carbon losses related to deforestation from smallholder agriculture may be explained by slash and burn clearing practices followed by cultivation of low productivity crops which reduces soil carbon inputs from root and litter^52,53. Moreover, more soluble material and lower C/N ratios of crop residues compared to forest litter have a lower degree of humification and higher decomposition rates, thus favoring soil carbon losses^54,55. Thirdly, previous reports have shown that agricultural cultivation practices such as plowing disrupt soil structure and reduce the physical protection of carbon in soil aggregates⁵⁵. As detailed spatial data on tillage and plowing practices were not available at the global scale, these effects could not be included in our analysis.

Fig. 3. Attribution of changes in soil C and litter (SOCL) in the tropics.

a The SOCL changes related to land use changes and other environmental drivers. b Spatial pattern of ΔSOCL caused by land use changes (LUC) estimated from the Bookkeeping of Land-Use Emissions (BLUE) model. The black error bars in a indicate the uncertainty (standard deviation) across diverse estimates. The boundary of the wet tropics is drawn in green, and the boundary of the dry tropics is drawn in black in (b). For the statistics on grazing-induced soil carbon loss, the pixels defined as ‘forest remaining forest’, and as affected by smallholder agriculture and commodity-driven deforestation, have been masked to avoid double counting. For the attribution of carbon emissions from peat drainage/fires, the pixels defined as ‘forest remaining forest’, and as impacted by smallholder agriculture, commodity-driven deforestation and grazing have been masked to avoid double counting.

We further compared our estimates of ∆SOCL from LUC to the results of the Bookkeeping of Land-Use Emissions (BLUE) model⁵⁶ (see “Methods”), which gives a LUC soil carbon source of −0.27 PgC yr⁻¹ over the tropics (Fig. 3b), with a dry tropics dominance (−0.17 PgC yr⁻¹, Fig. 3a, b) similar to our estimated carbon source of 0.25 ± 0.32 PgC yr⁻¹ from LUC in the dry tropics. However, the BLUE SOCL loss of 0.1 PgC yr⁻¹ in the wet tropics differs from our estimates, especially over Southeast Asia (Figs. 2a, 3b), where we found only a limited LUC-induced soil carbon loss. Possibly this is because BLUE does not explicitly consider plantations⁵⁷, a dominant land use change after forest loss in Southeast Asia and Africa, which is associated with smaller soil carbon loss than clearing to agriculture⁵⁸.

In summary, for the dry tropics, losses from cropland expansion (LUC), farmland erosion⁴⁷, grazing⁴⁵ (Supplementary Fig. 11), peat drainage/fires^46,59,60, and small gains from decreasing savannas burned areas⁵⁰, altogether explain 83.4% of our ‘top-down’ observation-based net ∆SOCL loss (Fig. 3a), the rest being tentatively attributed to other processes such as warming-induced increases of decomposition rates⁶¹, grassland erosion from overgrazing⁶², and possibly increased forest fire emissions⁴⁸. For the wet tropics, our estimated SOCL changes from intact forests, land use change, and loss from peatlands altogether explain 61% of the net carbon gains (Fig. 3a).

For the temperate and arid regions (map in Supplementary Fig. 10c), we found that land use change results in small SOCL emissions of 0.09 ± 0.03 PgC yr⁻¹ and 0.05 ± 0.02 PgC yr⁻¹, respectively (Supplementary Fig. 7, Supplementary Table 5). For boreal regions, land use change does not contribute to soil carbon emissions (Supplementary Fig. 7, Supplementary Table 5). By contrast, as shown in Supplementary Fig. 7, forest loss and gain from ‘forestry’ are defined as in ref. ⁵¹, from high-resolution satellite imagery and corresponding to harvest operations within managed forests and tree plantations, is associated with carbon accumulation in managed boreal (0.29 ± 0.22 PgC yr⁻¹) and temperate forests (0.10 ± 0.10 PgC yr⁻¹). Fire-disturbed boreal forests are also associated with an overall accumulation (0.31 ± 0.07 PgC yr⁻¹, Supplementary Fig. 7), which may be related to increased productivity and plant inputs after fire disturbance⁴³. Attributing soil carbon losses to fires with our data is clearly hindered by the scale mismatch between our 1° × 1° maps of ∆SOCL and the small fraction of a grid cell which burns annually⁶³ (on the order of 1% per year or less, Supplementary Fig. 12). While recently burned soils may lose carbon, the rest of the grid is dominated by re-growing or intact forests accumulating carbon in soils from litterfall⁶⁴.

Although soil properties (cation exchange capacity, texture, bulk density) play an important role in regulating mean soil carbon stocks (Supplementary Fig. 13), these factors did not correlate at 1° spatial resolution with our maps of ∆SOCL. This result suggests that recent SOCL changes are mainly driven by disturbances such as land use changes, grazing and forestry, and environmental factors, rather than by soil properties.

Annual variability of ΔSOCL

Finally, we investigated the global interannual variability (IAV) of soil C and litter changes (Fig. 4a) based on the approach of ref. ⁶⁵ (see “Methods”). Global modeling studies have suggested that the interannual variability of the net land CO₂ sink is dominated by semi-arid ecosystems^3,65. For year-to-year changes of SOCL fluxes, we found that boreal ecosystems dominate the IAV (40.7%, Fig. 4b). This is possibly because boreal and Arctic soils hold large amounts of soil carbon (~1700 PgC)⁶⁶, which is sensitive to the thawing period duration and intensity, so that climate anomalies in warm years result in large positive CO₂ flux anomalies. Semi-arid ecosystems (arid and dry tropics) contribute 33.2% of the global IAV of ΔSOCL, possibly linked to high turnover rates and coordinated climatic variability from El Niño-Southern Oscillation (ENSO) across the pan-tropics that promote rapid SOC changes in these ecosystems⁶⁷. While semi-arid regions generally store less SOC than northern regions (Supplementary Fig. 6), their strong climate variability and wide spatial extent make them major contributors to year-to-year variability in ΔSOCL.

Fig. 4. The interannual variations of changes in soil carbon and litter (ΔSOCL), and its regional contributions.

a Detrended interannual variations of ΔSOCL, and −1 × CO₂ growth rate (CGR). b Regional contributions to the interannual variations of global ΔSOCL (multi-estimate medians). The shadings of each curve show the uncertainty of CGR and the standard deviation of ΔSOCL across diverse estimates.

The IAV of ΔSOCL shows low correlation with the IAV of the CO₂ growth rate (CGR). For example, in the 2011 record land carbon sink⁶⁷, we estimate that ΔSOCL was a net source (Fig. 4a), possibly related to priming effects⁶⁸ and moister conditions. During the extreme El Niño drought of 2015/16, ΔSOCL was an abnormal sink even though that year was a record land carbon source⁶⁹ mainly due to biomass loss (Fig. 4a). This may be attributed to increased litter inputs resulting from drought-induced tree mortality and the increase of CWD stocks (Supplementary Fig. 14). Additionally, drought conditions may inhibit the diffusion of soluble organic carbon substrates and extracellular enzymes due to the reduced thickness of soil water films¹⁷, thereby suppressing microbial activity and decomposition rates. These effects likely contributed to short-term carbon retention in soil and litter under drought. However, given the inherent inertia of soil processes, these responses may represent short-term accumulation rather than sustained carbon sequestration.

Discussion

We acknowledge the uncertainty of ∆SOCL in our maps, but our results are derived from observations, and the uncertainty is smaller than the spread of DGVM models. The Monte-Carlo-based attribution and sensitivity analyses show that the large uncertainties in our estimated SOCL changes arise mainly from the spread of atmospheric inversions (Supplementary Table 1). While atmospheric inversion provides valuable constraints on surface fluxes at very large scales, its accuracy is limited by the observational system capabilities and methodological limitations^70–72: observational limitations (e.g., sparse monitoring networks), model structural errors (e.g., transport inaccuracies and prior flux biases), algorithmic differences (e.g., batch vs. variational methods and simplified error covariance), and spatiotemporal constraints (coarse resolution of prior error covariances). In future work, improved atmospheric inversion products with higher spatial resolution and better observational constraints will be essential to reduce uncertainty. Furthermore, our mass balance approach using atmospheric CO₂ inversion systems cannot partition fluxes between litter and SOC layers. Our biomass change-driven framework tracks total non-living carbon pool inputs and coarse woody debris. As litter decomposition and SOC formation are part of a continuum, separating them would require assumptions about litter-to-SOC transfer rates. Future studies combining explicit litter mass and turnover could help to resolve these pools.

Beyond reducing the uncertainty, our modeling framework enhances the capacity to detect the regional heterogeneity of SOC changes and attribute them to specific drivers. For instance, it highlights the importance of land use-induced SOC emissions, as evidenced by the correlation between our regional ΔSOCL estimates and LUC-driven SOCL changes simulated by the bookkeeping model BLUE (Supplementary Fig. 15). In contrast, ΔSOCL estimates from DGVMs and from machine learning models show negative or weak correlations with LUC-induced ΔSOCL from BLUE, suggesting that these models may underestimate the land-use effects. These findings reveal the limitations of current SOC models’ ability in capturing land-use dynamics, pointing to an important direction for future model improvement.

Our proposed top-down mass balance approach, integrating multi-model atmospheric inversion estimates and satellite-based biomass change data, offers a spatially explicit, observation-constrained approach to quantify changes in SOCL. Our findings underscore the pressing need to better integrate land-use and natural disturbance data into SOC models and to prioritize protection and restoration in vulnerable regions, particularly agriculture lands in the dry tropics. Notably, the net soil carbon emissions in the dry tropics, mainly from deforestation from smallholder agriculture and grazing, offset about 53.8% of the boreal sink, leading to a near-neutral global soil carbon change. Although not directly included in our analysis, various management practices, e.g., through the use of reduced tillage or no-till conservation agriculture as an alternative to traditional agriculture to reduce disturbance to the soil aggregates⁷³, the application of cover crops⁷⁴, and the interaction of trees and crops, i.e., agroforestry⁷⁵, have been recognized as effective strategies to minimize erosion C loss and to achieve sustainable agriculture.

Methods

SMOS-IC L-VOD

We take advantage of microwave remote sensing low-frequency L-band (1.4 GHz) vegetation optical depth (VOD) products, derived from Soil Moisture and Ocean Salinity (SMOS) satellite observations using the SMOS-IC algorithm, to estimate carbon stock changes in aboveground biomass^3,27. The SMOS-IC product provides global daily VOD data from both the descending (18:00 observations) and ascending (6:00 observations) orbits covering the period from January 2010 to February 2021 at a spatial resolution of 25 km. The SMOS-IC L-VOD data are simultaneously retrieved with soil moisture from a two-parameter inversion of the L-band microwave emission of the biosphere (L-MEB) model based on the multi-angular and dual-polarized SMOS observations²⁷. This proxy can probe total vegetation water content (VWC) due to its high penetration through the canopy and hence low saturation in dense forests. VWC scales with dry vegetation biomass and vegetation relative water content, that is, moisture content per unit of dry biomass⁷⁶. Assuming that the yearly averages or the differences between consecutive wet seasons of relative water content are relatively constant between years, the yearly VOD values are typically strongly correlated with biomass over time³³. Notably, even at the seasonal scale, biomass emerges as the largest contributor to VOD variations in 90% of the global terrestrial ecosystems, except for tropical forests⁷⁷. Previous analysis of the estimates of long-term biomass changes from L-VOD, both with and without the removal of relative water content changes, generated similar results^78,79. The above assumption is supported by many results reporting that (1) there is a strong and almost linear relationship between L-VOD and biomass and (2) the interannual variations of VOD can capture the interannual variations of biomass, even for tropical forests^31,80. A detailed review of the evaluation of L-VOD-derived AGB changes is given in ref. ³².

Benchmark maps of aboveground biomass carbon density

Three static aboveground biomass benchmark maps were used to calibrate L-VOD and retrieve carbon stocks in aboveground biomass (AGB) in order to reduce the dependence of estimating results on the accuracy of a single benchmark biomass map. The three maps are from ref. ⁸¹, European Space Agency – Climate Change Initiative (ESA-CCI, https://climate.esa.int/en/projects/biomass/) and ref. ⁸², hereafter referred to as the ‘Saatchi’, ‘CCI’ and ‘GlobBiomass’, respectively. All benchmark AGB maps were aggregated to 25-km pixels to be consistent with the spatial resolution of the L-VOD products. We multiplied the original values from the benchmark maps by a factor of 0.5 to convert the original units of aboveground biomass density (Mg ha⁻¹) to aboveground carbon density (MgC ha⁻¹), as done in ref. ⁸³. The 0.5 factor originates from decades of tropical forest research, most notably formalized in ref. ⁸⁴, reflecting that woody biomass typically contains approximately 50% carbon by dry mass^85,86.

Carbon stock changes in biomass retrieved from L-VOD

We applied a new filtering and reconstruction method for L-VOD³ to mask the pixels with unreliable retrievals and reconstruct the long-term smoothed data based on ascending and descending combined L-VOD retrievals. We applied the reconstructed long-term trend in VOD to calculate the annual values of aboveground biomass changes (ΔAGB). To mitigate potential phenological biases in vegetation optical depth (VOD) data arising from hemispheric seasonality differences, we have calculated the biomass changes using January-centered (November, December, January, February) VOD and July-centered (May, June, July, and August) VOD datasets (Supplementary Fig. 16). We further generated a composite biomass change map by averaging the January-centered biomass estimates and July-centered biomass estimates, effectively accounting for hemispheric phenological variations. For the interannual variations of AGB, we used the monthly reconstructed long-term smoothed data from the January-centered averages (November, December, January, February) to calculate the yearly net changes in biomass, as the differences between the January-centered averaged aboveground biomass fluxes between year n and year n + 1 make the fluxes more comparable with the fluxes of the yearly −1 × NEE or −1 × CGR, as done in ref. ³. We ranked the yearly L-VOD data from low to high and pooled the VOD values into bins of 250 grid cells. The mean values of the corresponding benchmark AGB were calculated for each L-VOD bin. Aboveground carbon stocks were retrieved from L-VOD based on empirical spatial relationships regressing L-VOD against the benchmark AGB datasets^30,34. The VOD-AGB relationship in the tropics was fitted using a three-parameter function (Eq. 2), and the VOD-AGB relationship in the Southern Hemisphere and the Northern Hemisphere was fitted using Eq. 3. Our spatial calibration approach divides the globe into three zones: Northern Hemisphere (30° N–90° N), tropics (30° S–30° N) and Southern Hemisphere (90° S–30° S), following a previous study³. We then fitted three spatial calibration functions of VOD vs. AGB based on three reference biomass maps for three different regions of the world. Note that the calibrated relationships between L-VOD and the benchmark maps were consistent in the entire tropical region and in tropical America and Africa (see Supplementary Fig. 17), suggesting that the statistical relationship between L-VOD and biomass carbon stocks is robust in different regions of interest. Thus, specific regions, e.g., Asia, are not included as regions of interest.

$$AGB=-\frac{a}{1+e^{b\times c}}+\frac{a}{1+e^{\left(-b\times \left(-c+VOD\right)\right)}}$$ 2

$$AGB=a\times VO{D}^b$$ 3

where a, b and c are three best-fit parameters. Nine sets of calibration parameters based on the three reference AGB maps (‘Saatchi’, ‘CCI’ and ‘GlobBiomass’) were used for the AGB simulations (Supplementary Table 6, Supplementary Fig. 17), of which nine sets were for the tropics by using three reference AGB maps and three reference regions (the whole tropics, tropical America, tropical Africa, see Supplementary Fig. 17), three sets for the Southern Hemisphere and three sets for the Northern Hemisphere based on three reference AGB maps. These AGB maps were then combined into nine global AGB datasets by merging the tropical products and products of the Southern/Northern Hemisphere. Because L-VOD may be underestimated if the observation footprint includes sizable open water bodies, we masked the pixels with the areal fraction of water bodies higher than 10%, based on a global wetland map⁸⁷.

Notably, the L-VOD retrievals cannot fully cover all days in 2010, so to minimize the error due to incomplete observations, we only calculated the AGB changes during 2011–2020. To match the spatial resolution of the NEE (1°), we calculated the sum of the total carbon fluxes of the 25 km pixels for each 1° pixel based on the latitude and longitude information. We then used a ratio of aboveground biomass to belowground biomass³⁵ to convert the ΔAGB to total biomass changes (ΔB). The spatial distribution of ΔB ensemble medians is shown in Fig. 2b. The interannual variations of ΔB (Supplementary Fig. 18) are generally consistent with those estimated in ref. ³.

Note that ‘no observations’ of ΔB due to radio frequency interference effects on microwave remote sensing retrievals accounts for 11.2% of the whole pixels (Supplementary Fig. 10d). To assess the impact of missing pixels on global soil carbon and litter changes, we used multiple datasets without missing pixels, including machine learning-based SOC changes, dynamic global vegetation models (DGVMs), the BLUE model, and SOC changes derived from an independent biomass product (Supplementary Table 7). Our analysis revealed that the contributions of carbon fluxes in areas of missing pixels to global carbon fluxes ranged from 2.7% to 17.9%, with a median value of 8.7% across these datasets. Given the relatively modest contribution of the missing pixels to global fluxes, their overall impact can be considered limited.

Atmospheric inversions

We used five air-sample-driven CO₂ atmospheric inversions from the Global Carbon Budget (2022)¹: CarbonTracker Europe CTE⁷⁰, the Jena CarboScope sEXToc-NEET_v2022⁸⁸, the inversion from the Copernicus Atmosphere Monitoring Service (CAMS v21r1)⁷¹, the inversion from the University of Edinburgh (UoE)⁷² and the NICAM-based Inverse Simulation for Monitoring CO₂ (NISMON-CO₂ ver2022.1)⁸⁹. The five atmospheric inversions provide global-scale NEE estimates, compiled by the Global Carbon Budget 2022 through a multi-model ensemble approach. The final dataset represents the median of these inversions, with inter-model spread quantified by the standard deviation. All models underwent standardized data assimilation using ground-based CO₂ observations, and regional uncertainties may arise from sparse observing networks. Different from National Greenhouse Gas Inventories (NGHGIs), which only include changes in carbon stock over managed land, inversion models solve for all CO₂ fluxes over both managed land and unmanaged land. Here, net ecosystem exchange in these atmospheric inversion models is defined as all non-fossil CO₂ exchange fluxes between terrestrial ecosystems and the atmosphere²⁴. The five selected NEE atmospheric inversion models show high agreement with the CO₂ growth rate (Pearson’s temporal R = 0.87, P = 0, Supplementary Fig. 19).

We further used CO₂ fluxes from lateral processes, which do not lead to carbon storage, to correct NEE from atmospheric inversions. The maps of laterally-displaced carbon are an update of ref. ²⁵, extended to the recent years and disaggregated at the pixel level using high-resolution proxy data (satellite-derived net primary production, population and livestock maps). Notably, for the process-focused interpretation of the ΔSOCL map, the coarse resolution of atmospheric inversions may lead to pixel-based errors. However, since we performed the interpretation on a multiple-pixel scale (i.e., areas with a certain disturbance type), the error should be smaller due to the error offsets in space.

Outputs from DGVMs

We use the ‘cSoil’ variable (including belowground litter) of 14 dynamic global vegetation models (DGVMs) from the TRENDY v11 project¹ to compare with our results: CABLE-POP, CLASSIC, CLM5.0, DLEM, ISAM, ISBA-CTRIP, JSBACH, LPJ-GUESS, LPX-Bern, ORCHIDEE, ORCHIDEE, SDGVM, YIB and VISIT. Trendy performs three factorial simulations, and we used the variable ‘cSoil’ of the S3 simulation to indicate the model-simulated soil carbon fluxes. For the S3 simulation, the models are forced by time-varying climate, atmospheric CO₂ concentration and land-use datasets. The climate forcing data are from Climate Research Unit (CRU) and 6-hourly Japanese 55-year Reanalysis (JRA55), the land-use change data stem from the Land Use Harmonization version2 (LUHv2) reconstruction, and the atmospheric CO₂ forcing data are derived from ice-core CO₂ data merged with National Oceanic and Atmospheric Administration (NOAA) annual data on atmospheric CO₂ concentrations from 1958 onward. Moreover, to make a better comparison between DGVM ensembles and our ΔSOCL estimates, we used the variable ‘cLitter’ (includes both CWD and litter) from the 15 DGVMs to calculate the changes in the soil and litter pools, namely changes in the sum of ‘cSoil’ and ‘cLitter’, referred to as ΔSOCL__DGVM in the main text. We further quantified the contribution of global litter changes (‘cLitter’) to global soil C and litter changes (‘cSoil’ + ‘cLitter’).

Soil carbon changes from machine learning

We also compared our ΔSOCL estimates with a machine learning-based topsoil soil organic carbon dataset³⁷ (SOC__ML, with 0.04° spatial resolution, covering the period of 1981–2018, downloaded from 10.5281/zenodo.5040379), which applied a space-for-time digital soil mapping model to simulate soil organic carbon stocks using soil organic carbon density data from the World Soil Information Service (WoSIS) soil profile database as training datasets, and considered climate, terrain, lithology, ecoregions, land cover, vegetation index, and cropland management datasets as predictors. We used the differences between soil organic carbon stock in year n and soil organic carbon stock in year n-1 to indicate soil carbon fluxes in year n. We aggregated the ΔSOC__ML to 1° grid cells by summing all the carbon stock changes in 0.04° grid cells within the 1° grid cell.

Bookkeeping of land-use emissions model

The Bookkeeping of Land-Use Emissions (BLUE) model⁵⁶ is one of the three bookkeeping models (OSCAR, BLUE, H&C2023) used in the Global Carbon Budget 2022¹ to quantify the net CO₂ flux from land use change (LUC) across diverse carbon pools, i.e., biomass, soil carbon, harvested wood products, and the atmosphere. OSCAR and H&C2023 provide data at the country level, whereas BLUE delivers spatially explicit data at 0.25° resolution. The BLUE model⁵⁶ is based on the original bookkeeping approach of Houghton (2003), which tracks spatially explicit LUC-induced carbon stock changes in several discrete pools for each combination of land cover type, land-use transition type, carbon pool type and plant functional type. After a land-use change event, changes in a certain carbon pool follow an exponential function for carbon decomposition and a saturation function for carbon accumulation (growth) in BLUE⁵⁶, which are used to estimate the temporal evolution of carbon gains or losses following transitions between different natural vegetation types, croplands, and pastures. The exponential decomposition response curves in BLUE separately describe the decay of vegetation biomass and soil carbon. This makes it possible to consider carbon stock changes caused by multiple LUC histories within a pixel. The BLUE model simulates soil carbon dynamics with rapid and slow relaxation processes induced by different LUC events, including vegetation degradation, (wood) harvest, agricultural abandonment, clearing and other transitions (i.e., the transition types between non-woody land covers). An updated version of the LUH (LUH2-GCB2022) dataset is used as input for land-use transitions⁶⁰. Here, we used the BLUE model to quantify the yearly pixel-based LUC-induced net soil carbon and litter fluxes of both the rapid and slow soil pool during 2011–2020 at a spatial resolution of 0.25°. We summed the changes in both the rapid and slow soil pools to indicate the soil carbon changes caused by land use change. Notably, BLUE cannot separate the litter and SOC pools, and the model outputs are therefore referred to as ΔSOCL fluxes.

Drivers of forest loss

We defined the pixel-based drivers of forest loss using the dataset from ref. ⁵¹, which identifies the dominant driver of forest loss at 10 × 10 km spatial resolution for 2001–2015. This dataset was generated by a decision-tree model that predicts the most likely cause of forest disturbance at each 10 × 10 km grid cell based on high-resolution Google Earth imagery to visually classify nearly 5000 training sample grid cells⁵¹. Five disturbance drivers of forest loss are contained in the dataset, including forestry, wildfire, urbanization, commodity-driven deforestation, defined as the large-scale, permanent and long-term clearing of forests to non-forest lands, and shifting agriculture, defined as small or medium scale clearing of forests to agriculture, we thereby term shifting agriculture to smallholder agriculture. We calculated the areal fraction of each disturbance type at 1 × 1° grid cells based on the forest disturbance classification map at each 10 km grid cell (Supplementary Fig. 20). We further defined the different disturbed forests based on the areal fractions of the five disturbance drivers. A given pixel is assigned a dominant disturbance agent if the fraction of this agent is higher than 50%. Note that, as pixels dominated by urbanization are very limited, empirical statistics on these pixels may induce large uncertainty. Therefore, the impact of urbanization is not considered in this study (Supplementary Fig. 20). The spatial boundaries of diverse disturbances, identified from satellite images during 2001–2015 can be seen in Supplementary Fig. 10b. We calculated the total SOCL fluxes in areas with different disturbance types to indicate the carbon stock changes caused by a certain type of disturbance, as done in previous studies^3,90. This allowed us to identify areas with repeated or prolonged disturbances, where soil carbon changes observed during the study period (2011–2020) may partially reflect delayed responses to prior disturbances.

While the attribution method based on disturbance maps overlap provides insights into potential drivers of ΔSOCL, it is subject to uncertainties from scale mismatches. Spatially, the use of 1° × 1° grid with a dominant-disturbance classification (applied when >50% of the cell is affected) does not account for sub-grid intact ecosystems. To improve the large-scale assessments of soil carbon dynamics, future work would benefit from atmospheric inversion products with enhanced spatial resolution, which can reduce attribution errors due to grid-level aggregation. Moreover, long-term, site-specific monitoring networks are essential to capture delayed soil carbon responses and to derive more precise post-disturbance carbon response curves. Such empirical data could significantly improve model parameterization and benchmarking at regional and global scales.

The carbon stock changes in coarse woody debris

Tree mortality does not always result in direct CO₂ emissions to the atmosphere, as dead stems and branches form CWD, defined as dead woody material longer than 1 m and dendrometer greater than 10 cm²⁶. CWD is the largest share of the dead wood carbon pool generated by major disturbances²⁶, i.e., fires and droughts, and by natural mortality. The CWD loss is mainly driven by decomposition and combustion by fires. We thus split the CWD pool changes into three parts to quantify the net changes of carbon stocks in CWD (Eq. 4): a. the decomposition of the initial CWD pool for years prior to 2011 (CWD_initial), b. CWD changes caused by stand-replacing fires (CWD_{stand-replacing}), and c. CWD increase due to tree mortality caused by annual background mortality (CWD_mortality)³⁴. We estimated the CWD changes induced by background mortality using annual net primary productivity (NPP) with a spatial resolution of 500 × 500 m from MODIS (MOD17A3HGF), annual live biomass changes calculated by L-VOD, leaf litter (LL) and the carbon loss to the atmosphere resulting from fires (Eq. 5). Note that the dynamics of CWD from woody mortality are driven by a combination of NPP input (plant trunks, and branches) and carbon loss from decomposition. In our framework, NPP serves as the initial carbon source term for CWD accumulation from woody mortality, while decomposition rates determine its turnover. The fire areal fraction, combustion ratio and aboveground carbon density were used to calculate the fire-induced changes in the CWD pool (Supplementary Text 1). We used a first-order kinetics model²⁶ (Eq. 6) to simulate carbon emissions from the lagged decomposition of CWD from dead woody biomass^26,91. The CWD decomposition rate k was used as the regional mean decomposition rate based on experimental observations⁹² across boreal (k = 0.040), temperate/arid (0.058) and tropical (0.268) biomes. Notably, CWD changes mainly occurred in unmanaged woody plants without woody harvest, including forest, shrubland, and savanna, and we therefore only kept these plant functional types in the estimates of CWD changes. We further masked the pixels with tree cover smaller than 10%, as determined by MODIS44B.006 data. Additionally, the areas covered by harvested forests were also masked using the boundaries defined in ref. ⁹³. To reach a span-up, we iteratively run the CWD-related models for 300 times to find the initial value of the CWD carbon pool at which we assumed the incoming flux was in equilibrium with the outcoming flux (Supplementary Text 1). We note that the simplified decomposition and combustion processes modeled by using a constant combustion and decomposition rate may induce some errors in the CWD estimates. However, CWD are important only in unmanaged forests without wood harvest, so the SOCL changes in woodlands, grasslands and croplands are still robust.

$$CW{D}_{\mathrm{tot}}=CW{D}_{initial}+CW{D}_{\mathrm{stand}-\mathrm{replacing}}+CW{D}_{\mathrm{background}\_\mathrm{mortality}}$$ 4

$$Mortality=NPP-\mathrm{\Delta }B-LL-F_{fire}$$ 5

$$CW{D}_t=CW{D}_{t=0}\times e^{-kt}$$ 6

where CWD_t is the CWD carbon stocks at year t, CWD_t=0 is the initial value of CWD, namely in the year 2011. NPP is the net primary product, ΔB is the carbon stock changes in total biomass, LL is the leaf litter calculated from the annual leaf area index (LAI) (ESA-CCI) and the specific leaf area⁹⁴, as done in ref. ⁷⁸. F_fire is the flux of carbon loss directly to the atmosphere from aboveground biomass burning related to wildfires. We used nine sets of ΔB as input to the CWD modeling and obtained nine sets of ΔCWD accordingly. The details of CWD_tot calculation are given in Supplementary Text 1.

Calculation of global budgets of changes in soil C and litter

Net ecosystem exchange (NEE) is the net exchange of CO₂ with the atmosphere, including all the vertical CO₂ fluxes from the ecosystem to the atmosphere²⁴. NEE can be decomposed into separate terms using the following equations (Eqs. 7, 8). We removed from the NEE given by the inversions, the gridded surface-atmosphere CO₂ fluxes induced by lateral transport processes from rivers and crop/wood trade based on the updated estimates from ref. ²⁵ to get the net carbon storage change in terrestrial ecosystems. This net carbon storage change was separated into the contribution of living biomass and non-living carbon stock changes. The living biomass loss after mortality is not immediately decomposed as CO₂ fluxes, and there is a lagged decomposition process caused by coarse woody debris (CWD). We further removed the CWD pool changes from the non-living carbon pool changes to infer soil carbon changes (including litter). Thus, hereafter, we used ∆SOCL to indicate the net carbon balance from both the soil and litter layers. The percentage contribution of each term to ΔSOCL, both on a global scale and across regions, is quantified in Supplementary Table 8. The relative contributions of component fluxes to ΔSOCL show high variability, indicating: (1) compensatory effects among component fluxes, and (2) the existence of near-zero denominators in calculations (where small ΔSOCL values amplify the contribution).

$$NE{E}_{\text{inversion}}-L=\mathrm{\Delta }B+\mathrm{\Delta }CWD+\mathrm{\Delta }SOCL$$ 7

$$L=F_{crop/woodtrade}+F_{river}$$ 8

where L means the CO₂ fluxes from lateral processes, F_crop/wood trade, F_river means lateral fluxes from crop/wood trade and river. The uncertainty of the lateral fluxes was set to 20% in the Monte-Carlo simulations, as done in ref. ²⁴. The total fluxes of each lateral process are listed in Supplementary Table 9. Notably, as the estimates of wood trade have large uncertainties, we only included the lateral processes from crop trade and river in the calculations. Using five atmospheric inversions, nine ΔAGB maps and nine corresponding ΔCWD maps as inputs in Eq. 7, we obtained a total of 45 ΔSOCL maps, and the uncertainty was measured by the standard deviation among the 45 combinations. Note that from an ecosystem perspective, the carbon uptake of NEE is equivalent to net ecosystem production (NEP). As we concentrate on the carbon exchange between the atmosphere and land, NEE with an atmospheric convention is used here, rather than NEP.

Partitioning of interannual variations

Partitioning of IAV to climatic biomes follows the definition of Eq. 9⁶⁵. The contribution of the IAV of a climatic biome to the global ΔSOCL IAV is defined as:

$$f_j=\frac{{\sum }_{t=1}^n\frac{x_{jt}∣X_t∣}{X_t}}{{\sum }_{t=1}^n∣X_t∣}$$ 9

where x_jt is the flux anomaly (detrended) for climate biome j at time t (in years), and n is the number of years (n = 10 in this study). The X_t is the global flux anomaly of ΔSOCL, and f_j is the average relative anomaly x_jt/X_t for climate biome j, weighted by the absolute global ΔSOCL anomaly |X_t| . The definition makes sure that the sum of f_j is equal to 1, allowing a certain f_j to fall outside the range (0,1) if the global anomaly X_t results from the partial cancellation of the contribution x_jt from different climatic regions.

Comparison with inventories

We compared our estimates of ΔSOCL and ΔB in managed forests with national GHG inventories by the United Nations Framework Convention on Climate Change (UNFCCC) across 43 Annex I countries. The pixels of managed forests were considered by masking the pixels of intact forests defined as no sign of significant human activity, as did in refs. ^95,96 based on the Intact Forest Landscapes dataset (https://intactforests.org/)⁹⁷. The results show strong spatial correlations for both SOCL and biomass changes between inventories and our estimates, with R² exceeding 0.43 (P < 0.001, Supplementary Fig. 3), yet with much smaller SOCL changes in GHG inventories. We consider the underestimation of SOCL changes in inventories likely due to incomplete spatial coverage (not all countries report on all land uses) and the difficulty to measure the carbon changes in deep soil. In this regard, the IPCC guidelines recommend sampling to 30 cm depth to measure changes in soil carbon stocks, even if some countries go deeper. In addition, our SOCL changes in grassland, ranging from −97.6 to 85.9 gC m⁻² yr⁻¹ with a median of −22.24 ± 39.04 gC m⁻² yr⁻¹ (Supplementary Fig. 21), are in the range of the grassland SOC inventory datasets, compiled in ref. ⁹⁸, ranging from −150 gC m⁻² yr⁻¹ to 100 ± 25 gC m⁻² yr⁻¹.

Author contributions

P.C. and H.W. conceived and designed the overall study. H.Y. contributed to the calculations of coarse woody debris. P.S., G.G., C.S., P.P., Y.B.-O. contributed to the interpretation of the results. J.-P.W., H.Y., S.H., J.C., C.A., L.F., K.W., L.L. and F.F. provided comments and feedback on the discussion. J.-P.W. and X.L. provided the SMOS-IC L-VOD product. S.S. provided the Trendy data. P.P. provided the national inventory data of soil carbon changes. F.C. and P.I.P. provided atmospheric inversions data. H.W. and P.C. drafted the first version of the manuscript and received inputs from all co-authors. The final version was approved by all authors before submission.

Peer review

Peer review information

Nature Communications thanks Katharina H. M. Meurer and the other anonymous reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Data availability

The SMOS-IC L-VOD from this study can be freely downloaded from the INRAE Bordeaux remote sensing lab website (https://ib.remote-sensing.inrae.fr/). The five atmospheric inversions used in this study are acquired from the LSCE server. The CO₂ fluxes from lateral processes are downloaded from https://thredds-su.ipsl.fr/thredds/fileServer/tgcc_thredds/work/p24cheva/LateralFluxes/lateralfluxes_v4.1.tar. The Saatchi biomass map is available upon request from Dr. Sassan S. Saatchi. The data of land use-induced SOCL changes from BLUE are available upon request from Dr. Clemens Schwingshackl. The gridded carbon emissions from peat drainage/fires used in Fig. 3 are available from https://data.apps.fao.org/ and https://www.geo.vu.nl/~gwerf/GFED/GFED4/, respectively. The forest disturbance classification map is freely available from https://www.science.org/doi/10.1126/science.aau3445. The simulations from TRENDY DGVMs are available at https://sites.exeter.ac.uk/trendy. The datasets of global topsoil soil organic carbon changes from machine learning are downloaded from https://zenodo.org/records/5040380. The annual carbon stock changes in SOCL estimated in this study can be downloaded from 10.6084/m9.figshare.29710943.v1. Additional data used in the paper are publicly available and are provided in the corresponding references. A detailed list of the data used in this study can be found in Supplementary Table 10. The source data for the figures are available at: 10.6084/m9.figshare.30175363.v1.

Code availability

The computer codes used for data processing and figure generation are available at: 10.6084/m9.figshare.29710955.v2.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-025-64929-3.