Process-oriented analysis of dominant sources of uncertainty in the land carbon sink

Anthropogenic CO2 emissions have risen continuously over the past 60 years, with about a quarter taken up by the terrestrial biosphere, forming a natural land sink that mitigates climate change. To project the carbon cycle and climate, it is necessary to understand the processes and timescales controlling the land sink, including photosynthesis, allocation, growth, litterfall, mortality, and soil turnover operating from days to centuries. Evidence suggests the global net land sink is concentrated in northern latitudes, while tropical regions are closer to carbon neutral due to large LULCC-related losses counteracting natural uptake. Process-based dynamic global vegetation models (DGVMs) are used to attribute observed changes to external drivers (rising CO2 and nitrogen deposition, climate variability and trends, and land-use and land-cover change) and to quantify regional sinks. However, substantial inter-model spread exists, especially at regional scales and for the partitioning between vegetation and soil stocks. This study applies a triple (3D-matrix) approach to identify the dominant external drivers, regions (tropics vs extratropics), and processes (productivity/input vs turnover/output) responsible for changes in the global net land carbon sink over 1959–2020, using 18 DGVMs from TRENDYv10/GCB2021 and a process attribution framework to separate input- and turnover-driven changes.

Previous work indicates a strong northern land sink and limited net tropical sink due to substantial LULCC emissions, with Amazonia close to carbon neutral when balancing natural uptake and LULCC sources. The DGVM multi-model mean is broadly consistent with the top-down Global Carbon Budget residual land sink, but large spread remains among models, particularly for regional attribution and internal stock partitioning. Observational constraints on long-term global vegetation and soil carbon changes are lacking, especially pre-satellite era, complicating model evaluation. Recent observation-based syntheses suggest a net forest biomass sink since 2000, supporting modeled gains but not narrowing magnitude uncertainties. The literature also highlights the importance of forest management, regrowth and degradation, nitrogen deposition synergies with CO2, and the need to better represent mortality processes, disturbance, and soil biogeochemistry (e.g., priming, microbial activity, mineral associations).

- Models and period: Analysis of 18 DGVMs from TRENDYv10 (CABLE-POP, CLASSIC, CLASSIC-N, CLM5.0, DLEM, IBIS, ISAM, ISBA-CTRIP, JSBACH, LPJ-GUESS, LPJ, LPX-Bern, OCN, ORCHIDEE, ORCHIDEEv3, SDGVM, VISIT, YIBs) over 1959–2020, the period constrained by atmospheric CO2 measurements and the Global Carbon Budget (GCB2021). - Forcing and data: Models forced with merged CRU (monthly) and JRA-55 (6-hourly) climate, prescribed atmospheric CO2, temporally varying nitrogen deposition and fertilizer, and HYDE v3.3 land-use data (some models also using LUH2-GCB2021 land-use transitions). A HYDE artifact around 1960 was corrected by replacing 1959–1961 LULCC estimates with the average of 1958 and 1962. - Simulations to isolate drivers: Four standard simulations: S0 (fixed preindustrial CO2 and land use; recycled 1901–1920 climate), S1 (transient CO2 and N deposition; recycled climate; fixed land use), S2 (transient CO2/Ndep and climate; fixed land use), S3 (transient CO2/Ndep, climate, and land use). Driver effects: CO2+Ndep = S1–S0; climate = S2–S1; LULCC = S3–S2; all drivers = S3. - Diagnostics: Global and regional (north of 30N; south of 30N) annual means of NBP (net land sink), NPP, heterotrophic respiration, vegetation carbon (Cv), soil carbon (Cs; including litter and coarse woody debris where available). - Process attribution framework: Using an analytical approximation of carbon pool dynamics, changes in vegetation and soil carbon stocks are decomposed into contributions from changes in inputs (vegetation: ΔNPP; soil: Δflux from vegetation to soil, fvs) and changes in turnover time (vegetation: Δτv; soil: Δτs), plus an interaction term. Steady-state approximations (Cv ≈ NPP·τv; Cs ≈ fvs·τs) are used to express ΔCv and ΔCs between 1959 and a given year as sums of input-driven, turnover-driven, and interaction components. Adjustments correct for non-steady-state behavior, multiple-pool structures, and linearization errors by reconciling approximated changes with actual modeled stock changes via proportional scaling. Uncertainty decomposition attributes inter-model spread in ΔCv and ΔCs to baseline input, change in inputs, baseline turnover, and change in turnover components.

- Net land sink magnitude and drivers: The GCB top-down land sink increased from 0.2 ± 0.4 PgC yr−1 (1960s) to 1.7 ± 0.6 PgC yr−1 (2011–2020). DGVMs (S3) captured this increase, from −0.1 ± 0.6 to 1.6 ± 0.5 PgC yr−1. CO2 and nitrogen deposition drove a strong increasing sink (1.2 ± 0.2 to 3.5 ± 0.8 PgC yr−1), while LULCC produced relatively constant net emissions of 1.3 ± 0.5 PgC yr−1. Long-term climate trends since the 1980s reduced the sink by −0.4 ± 0.5 PgC yr−1, with climate variability causing large interannual fluctuations (±2 PgC yr−1). - Spatial patterns: CO2+Ndep-driven sinks are concentrated in northern and tropical forests. LULCC-related losses are largest in the tropics and in hotspot regions across the northern hemisphere (e.g., China, USA, West Eurasia). Climate effects display regional gains (e.g., east Brazil, Australia, high northern latitudes) and losses (e.g., Amazon, Sahel, South Africa). Models disagree on the sign of ecosystem carbon change in many regions where CO2 and LULCC effects compete. - Vegetation carbon change (1959–2020): Global ΔCv = 28 ± 26 PgC (multi-model mean), with 2 models simulating net losses. CO2+Ndep contributions: input-driven ΔCinput,CO2 = 87 ± 26 PgC; turnover-driven ΔCoutput,CO2 = −3 ± 21 PgC; net ΔCCO2 = 84 ± 31 PgC (range ~30 to 150 PgC among models). LULCC reduces vegetation carbon: ΔCv,LULCC = −58 ± 21 PgC, split between reduced inputs (ΔCinput,LULCC = −23 ± 22 PgC) and increased turnover/losses (ΔCoutput,LULCC = −29 ± 18 PgC), including loss of additional sink capacity. Climate-driven global vegetation changes are small and uncertain (ΔCinput,CLIM = −1 ± 14 PgC; ΔCoutput,CLIM = 3 ± 9 PgC), with regional contrasts: northern productivity gains vs tropical productivity declines. - Soil carbon change (1959–2020): Global ΔCs = 21 ± 32 PgC (mean), with 3 models simulating net losses. CO2 increases soil carbon via enhanced inputs but also increases apparent turnover due to higher shares of fast pools (false-priming effect); the net CO2-driven soil change is positive but partitioning between input vs turnover is uncertain. LULCC reduces soil carbon mainly through reduced litter inputs: ΔCs,LULCC = −25 ± 29 PgC; ΔCs,input,LULCC = −36 ± 76 PgC (models range from >100 PgC losses to 45 PgC gains), with smaller, mixed turnover effects. Climate causes a net global soil carbon loss: ΔCs,CLIM = −13 ± 12 PgC, as warming-enhanced heterotrophic respiration offsets productivity-driven input increases. - Temporal acceleration: Since 2000, climate impacts accelerated: northern soil productivity-driven gains increased by 19 TgC yr−2 and tropical biomass productivity-driven losses by 9 TgC yr−2, associated with observed warming. - Dominant uncertainties: Differences in turnover (baseline and change) explain roughly 70% of the inter-model spread in both vegetation and soil carbon changes (supported by uncertainty attribution), with additional uncertainties from NPP responses, LULCC representation (e.g., wood harvest, grazing, shifting cultivation), land management, and spatial patterns of LULCC. - Regional LULCC: All models simulate net biomass losses due to LULCC, with magnitude sensitive to inclusion of management processes and regrowth dynamics; compensating regional effects (e.g., losses in Russia/USA vs regrowth in Eurasia) add uncertainty.

The TRENDYv10 DGVM ensemble reproduces the observed increase in the global net land carbon sink and attributes the long-term sink primarily to rising atmospheric CO2 and nitrogen deposition, with partial offsets from LULCC and climate change. The ensemble supports a strong northern sink and relatively neutral tropics, consistent with independent top-down assessments. However, models disagree on the partitioning of net carbon gains between vegetation and soil, and on the magnitude of stock changes. The analysis identifies turnover (residence time) processes—both their baseline values and their responses to environmental change—as the principal source of uncertainty in modeled vegetation and soil carbon changes, with additional contributions from NPP responses, allocation, demographic and mortality processes, and nutrient limitations. LULCC exerts substantial reductions in vegetation and soil carbon through both diminished inputs and increased turnover, with the loss of additional sink capacity further reducing potential storage. Climate change increasingly affects stocks, with warming-driven soil carbon losses and regionally divergent productivity effects (northern gains, tropical declines). These findings highlight the need for improved representation of internal carbon cycling (allocation, tissue lifespan, mortality, demography), land management and disturbance, and process-based soil biogeochemistry and nutrient cycling to better resolve the drivers and processes controlling the land carbon sink.

The study provides a process-oriented attribution of changes in the land carbon sink over 1959–2020 using 18 DGVMs, separating input- and turnover-driven contributions across major drivers and regions. It confirms that CO2 and nitrogen deposition underpin the land sink, while LULCC and climate reduce it, and identifies turnover processes as the dominant source of model spread in vegetation and soil carbon changes. Major contributions include quantifying driver-process-region interactions, demonstrating the partial cancellation of regional climate impacts at the global scale, and highlighting the large, uncertain LULCC impacts including loss of additional sink capacity. Future research should prioritize: improving plant allocation, tissue lifespan, and mortality/demography; incorporating detailed management and disturbance processes (wood harvest, grazing, shifting cultivation, forest degradation, pests, wind, fire dynamics); advancing soil carbon modeling with explicit microbial, priming, mineral association, aggregation, mycorrhizal interactions, and peatland processes; enhancing carbon-nutrient coupling; and evaluating individual carbon pools and fluxes to constrain turnover and input processes.

- Observational constraints: Lack of global, long-term observations of vegetation and soil carbon stocks, particularly pre-satellite era, limits evaluation of modeled stock changes and their partitioning. - Model structure: DGVMs often have simplified internal carbon cycling, limited demography, and incomplete mortality and disturbance processes (e.g., drought-induced mortality, pests, wind), and many do not simulate or accurately force fire dynamics. Soil modules typically use first-order decay with cascading pools, omitting explicit microbial, priming, mineral, aggregation, and mycorrhizal processes; peatlands are generally not represented. - Process attribution method: Uses steady-state approximations and aggregates multiple pools into single vegetation and soil pools, leading to potential biases (e.g., false-priming). Adjustments reconcile with modeled stock changes but cannot fully resolve multi-pool dynamics; treating soil as a single pool may underestimate active turnover and confound deep inactive carbon. - Forcing and LULCC uncertainties: Historical land-use maps have artifacts (partially corrected) and uncertainties in transitions and management; models differ in inclusion of wood/crop harvest, grazing, and shifting cultivation, contributing to spread in LULCC impacts. Forcing data uncertainties (climate, N deposition, fertilizer) also propagate into results. - Pool availability: Not all models provide litter and coarse woody debris pools, requiring approximations for soil carbon aggregation and potentially affecting attribution.