A review of the major drivers of the terrestrial carbon uptake: model-based assessments, consensus, and uncertainties

The studies included in this review are selected based on the relevancy of the topic, period of time and climate scenarios assessed, geographical scale (mostly global), and importantly whether they analyze one or more of the four major drivers. We selected recent (mostly from 2010 to present) peer-reviewed publications on the land carbon sink and drivers of land carbon uptake based mainly on the authors' experience in the subject. The web-based list of CMIP5 publications (https://cmip-publications.llnl.gov/) is utilized to select some of the literature. Further, literature searches with the keywords and combinations of 'recent trends in land carbon uptake, TRENDY historical land carbon, CMIP5 RCP land carbon drivers, CO₂ fertilization, nitrogen deposition, LULCC, LUC, deforestation, climate warming, drivers of terrestrial land carbon' were conducted with web-based search engines such as Google Scholar and Mendeley with cross-referencing from other journal websites.

The studies discussed here are dominantly based on the CMIP5 experiments and sensitivity experiments performed using dynamic vegetation models (DGVMs). We discuss the results from fully coupled CMIP5 historical simulations from the year 1850 to 2005 and simulations using RCP scenarios for the 21st century projections. Further, we include land model studies that analyze the changes in the land carbon sink in recent decades (in the past ∼30 years). The reason to choose the three RCP scenarios (RCP2.6, RCP4.5, and RCP8.5) is the availability of literature on scenario-based projections of carbon cycle changes. Table S1, available online at stacks.iop.org/ERL/14/093005/mmedia, provides a list of studies (with brief descriptions of the models used and the features/limitations of the studies) that analyzed the drivers of land carbon uptake in the historical and future scenarios. We select the studies that have analyzed the contributions of one or more of the four drivers (CO₂ fertilization, N-deposition, climate change, and LULCC) to the net carbon uptake for the historical period and the RCP scenarios (e.g. Piao et al 2009, 2012, Brovkin et al 2013, Devaraju et al 2016, Huntzinger et al 2017, Piao et al 2018, Tharammal et al 2019; see table S1).

Only two of the models among the CMIP5 models—NCAR CESM and NorESM—include the C–N dynamics, and both these models share the same land model-Community Land Model-CLM4 (Oleson et al 2010). Published studies that estimate the relative contributions of the major factors to the changes in land carbon uptake in the future scenarios are limited. Brovkin et al (2013), as a part of the Land-Use and Climate, Identification of Robust Impacts project (LUCID, Boysen et al 2014) assessed the effects of LULCC on the total ecosystem carbon (TEC) in the RCP2.6 and RCP8.5 scenarios using five CMIP5 models. We select four of these five LUCID models (CanESM2, MIROC ESM, HADGEM2, and IPSL-CM5A-LR) to infer the contributions of CO₂ fertilization and climate warming for these models by using the individual model's land carbon uptake sensitivity to CO₂ (β in PgC ppm⁻¹) and climate warming (γ in PgC K⁻¹) taken from Arora et al (2013). We assume that the sensitivity factors for the RCP scenarios are comparable to those estimated from the quadrupled CO₂ experiments by Arora et al (2013) (conducted by increasing the CO₂ concentration by 1% yr⁻¹ from the pre-industrial concentration of 285 ppm until concentration has quadrupled by 140 years). These four models lack N-limitation in their land carbon cycle model. The net change in the TEC in the RCP2.6 and RCP8.5 scenarios for the LUCID models are estimated from Jones et al (2013).

Apart from the fully coupled CMIP5 simulations, the results from TRENDY ('Trends and drivers of the regional scale sources and sinks of carbon dioxide', Sitch et al 2015; an ensemble of nine DGVMs forced by a common/observed climate forcing data) provide the factor-wise contribution to the historical land carbon uptake (years 1901–2010). TRENDY simulations have been used to assess contributions of CO₂, climate, and LULCC on the net biome production (NBP) and land carbon uptake during the present-day and recent past (Piao et al 2013, Schimel et al 2015a, Sitch et al 2015, Li et al 2017, Le Quéré et al 2018). Similarly, Multi-scale Synthesis and Model Intercomparison Project (MsTMIP, Huntzinger et al 2013, 2017) assesses the major drivers of changes of the land C uptake during 1959–2010, using twelve land biosphere models (five of which include coupled C–N cycle). MsTMIP models are forced by the same climate records, N-deposition, and land use data (Huntzinger et al 2013). Further, recent studies on the assessments of historical and present-day land carbon budget (Le Quéré et al 2018), NPP, and leaf area index (LAI) values using satellite-based observations, land models, and bookkeeping methods (e.g. Zhang et al 2016, Zhu et al 2016, Mahowald et al 2016) are discussed in the paper. In general, for each period, we review the changes in global land carbon uptake from the modeling studies and observations, followed by discussions of the major drivers of these changes.

Besides, we review the sensitivity of the land carbon uptake to CO₂, climate warming, and N-deposition using available model results in the published peer-reviewed literature. As a part of phase 3 of the Coupled Model Intercomparison Project (CMIP3), the Coupled Carbon Cycle Climate-Model Intercomparison Project (C4MIP, Friedlingstein et al 2006) provided a comparison of sensitivity factors for CO₂ fertilization and climate change from eleven coupled climate-carbon cycle models. C4MIP models were forced by historical emissions and the IPCC Special Report on Emissions Scenarios (SRES) A2 scenario for the years 1850–2100. Another set of C4MIP studies accompanies CMIP5, and the sensitivity factors estimated from this project are analyzed by Jones et al (2013) and Arora et al (2013), which are briefly discussed in this review. Further, Bala et al (2012) provide a comprehensive analysis of the sensitivity factors.

This review focuses on the following variables for the land carbon cycle: NPP, LAI, NEP (NEP = NPP − R_h), NBP (NBP = NEP − fire-land use), and TEC. Increases in NPP or LAI ('greening') do not necessarily imply increases in the land carbon stock as the latter is also affected by changes in other variables such as R_h and LULCC. Our discussion will focus on global changes, and regional values will be discussed whenever available, and relevant to the context. We use a sign convention where the fluxes into the ecosystem are positive (land sink), and a –ve sign for the fluxes denotes land source. As the studies assessed in this review estimate the changes in land carbon fluxes and storage through different variables such as NBP, TEC, and NEP, we primarily discuss the contributions of major drivers in terms of the percent contribution of the factors to changes in the respective variables. As the sign of the changes could be positive or negative, following Zhu et al (2016) and Huntzinger et al (2017), we calculate the percent contribution of the individual factors from the literature as:

Percent contribution of individual factor = 100 × [absolute contribution of the individual factor/sum of absolute contributions of the factors].

For the historical period, apart from the increase in atmospheric CO₂, climate warming, and increased nitrogen deposition rates (see Introduction), the estimated loss of primary land is ∼75 million km² (125 million km² in 1850 to 50 million km² by 2005, see Hurtt et al 2011). The large deforestation for agriculture and wood harvest took place in high latitudes, specifically in central and eastern N. America, Europe, and SE Asia (Hurtt et al 2011). The global mean surface temperature increased by ∼0.6 °C–0.8 °C during the historical period (IPCC 2013) because of the net result of all the forcings including non-CO₂ greenhouse gases (GHG), aerosols and ozone.

The CMIP5 historical and RCP simulations are 'concentration driven' or forced by prescribed CO₂ concentrations, rather than emissions-driven. Other forcing factors include aerosols, non-CO₂ GHG, LULCC, and N-deposition for the models with C–N biogeochemistry. The duration of the future transient RCP simulations is from 2005 to 2100. Atmospheric CO₂ concentration is projected to increase by 41 ppmv, 159 ppmv, and 556 ppmv respectively in the RCP2.6, RCP4.5, and RCP8.5 scenarios, from 379 ppm in the year 2005 (Taylor et al 2012). The rate of N-deposition increases in the RCP2.6 and RCP8.5 scenarios (by 18% and 25% in 2100 respectively in the input datasets to CESM model; Tharammal et al 2019), whereas, it decreases in the RCP4.5 scenario (by 15% in the CESM model; Tharammal et al 2019), relative to the deposition rates in 2005 (Lamarque et al 2013).

The RCP8.5 and RCP2.6 scenarios project reductions in global forest cover in the developing countries (in Africa, SE Asia and South America) due to the increase in agriculture driven by projected increase in population (Hurtt et al 2011). However, the RCP4.5 scenario assumes land-use management as a strategy to curb the global carbon emissions and the scenario projects afforestation globally (Hurtt et al 2011). The RCP8.5 and RCP2.6 scenarios project increases in cropland due to increases in agriculture and increased cultivation of biofuel crops, whereas, the RCP4.5 scenario has reduced cropland because of afforestation and better management of food production and distribution in the 21st century. Pastureland increases in the RCP8.5 scenario due to intensive grazing, whereas, the RCP4.5 and RCP2.6 scenarios project reductions. All three future scenarios project increased wood harvesting in 2100 (Hurtt et al 2011). The land use changes projected in these three future scenarios are smaller and are more stabilized compared to the historical period. Further, unlike the historical period, the LULCC in the future scenarios are mostly located in the tropics and subtropics (Hurtt et al 2011). As a result of all these forcings, the CMIP5 models project increases in global mean surface temperatures in the ranges of 0.3 °C–1.7 °C for the RCP2.6 scenario, 1.1 °C–2.6 °C for the RCP4.5 scenario, and 2.6 °C–4.8 °C for the RCP8.5 scenario (IPCC 2013).

3.1.1. Historical period (1850–2005)

There is a consensus that the land was a source of carbon in the historical period, however, the quantitative estimates have large uncertainty. For instance, based on observations, Arora et al (2011) estimate the cumulative land carbon uptake for the historical period (1850–2005) that includes the LULCC effect as −11 ± 47 PgC (land is a net source of carbon; 1 PgC = 1 Peta gram C = 10¹⁵ g). For the same period, the cumulative land carbon uptake estimated by thirteen CMIP5 models (with land use changes) shows a multi-model mean of −19 PgC (from nine out of the thirteen models), however, the values range from a land source of −124 PgC to a land sink of +50 PgC in 2005 (Jones et al 2013). Similarly, for the years 1850–2005 Friedlingstein et al (2014) estimate the mean of cumulative land carbon uptake as ∼−26 ± 32 PgC from eight of the CMIP5 models.

Devaraju et al (2016) provide an assessment of the major drivers of the land carbon uptake for the 1850–2005 period using sensitivity simulations performed with the NCAR Community Earth System Model (CESM1). CESM1 simulates an increase in global NPP by 4 PgC yr⁻¹ and a cumulative land carbon source by 2005 (reduction of global TEC by 45 PgC in 2005 with respect to 1850, Lawrence et al 2012, Devaraju et al 2016). The larger value of land source estimated by CESM1 compared to the ensemble mean of the CMIP5 models and observations (see Arora et al 2011, Friedlingstein et al 2014), is due to the N-limitation in CESM1 that weakens the CO₂ fertilization effect in the model (Thornton et al 2009, Bonan and Levis 2010). CO₂ fertilization (2.3 PgC yr⁻¹, 43.22% contribution) and N-deposition (2.0 PgC yr⁻¹, 37.10% contribution) cause most of the increase in global NPP between 2005 and 1850 in the CESM1 (Devaraju et al 2016; figure 2). Further, Devaraju et al (2016) find that the effects of CO₂ fertilization (+55.39 PgC, 26.61% contribution) and N-deposition (+26.11 PgC, 12.54% contribution; figure 2) are positive to the net land sink during 1850–2005, but a larger negative contribution of LULCC (−111.57 PgC, 53.6% contribution) augmented by climate warming (−15.05 PgC, 7.23% contribution) lead to a net decline in the land carbon sink during 1850–2005. Similarly, Eglin et al (2010) using the process-based global ecosystem model ORCHIDEE find that the effect of LULCC is dominant on the reduction of global soil carbon and biomass during the period 1901–1960, compared to the positive effects of combined CO₂ and climate. Further, CMIP5 models that include time-varying historical LULCC simulate a decline in the land carbon stocks from 1850 to ∼1950, primarily due to deforestation (Jones et al 2013).

Zoom In Zoom Out Reset image size

3.1.2. Recent decades

There is broad agreement among several studies that the land had turned into a sink in recent decades. This is based on assessments that show an increasing trend of NPP, LAI, and land carbon sink over the globe in recent decades (Nemani et al 2003, Raupach et al 2008, Friedlingstein et al 2010, Ballantyne et al 2012, Anav et al 2015, Piao et al 2015, 2018, Sitch et al 2015, Zhu et al 2016, Campbell et al 2017, Fernández-Martínez et al 2017, 2019, Keenan and Williams, 2018). The global carbon budget project (GCP, Le Quéré et al 2009) estimates a residual land carbon sink of 0.36 ± 1 PgC yr⁻¹ for the years 1960–2005 (calculated as the sum of fossil fuel flux and LULCC flux minus the sum of ocean sink and atmospheric CO₂ growth). Latest estimates of GCP using DGVMs and residual sink method find that the land carbon sink almost doubled in 2007–2016 when compared to the 1960s (from 1.4 ± 0.7 to 3 ± 0.8 PgC yr⁻¹, Le Quéré et al 2018). Further, the estimates of the land carbon uptake for the recent decades from the CMIP5 models show an increasing land sink trend during 1960–2005 (ensemble mean of land-atmosphere flux of 0.7 ± 0.6 PgC yr⁻¹, Anav et al 2013).

A number of modeling studies have assessed the drivers of this recent increases in the NPP and land carbon sink (see table S1). Figure 2 and table S2 show the percentage contributions of major drivers to recent (1976–2012) increases in NPP and LAI from four global studies and one regional study. The largest contribution is from CO₂ fertilization and its percent contribution to increases in global NPP/LAI ranges from 43% (Devaraju et al 2016) to 86% (Sitch et al 2015), with a mean of 62.5 ± 19.5%. The percent contribution of N-deposition ranges from 8% (Zhu et al 2016) to 33% (Devaraju et al 2016), with a mean of 21.2 ± 17.5%. The percent contribution of climate change varies from 8% (Zhu et al 2016, positive effect) to 50% (Piao et al 2018, negative effect) (mean 20.2 ± 20.0%). It is unclear whether climate change contributes positively or negatively to the increases in NPP or LAI as two of the global studies (Devaraju et al 2016, Piao et al 2018) show negative contribution of climate change to the NPP trends, while two other studies (Sitch et al 2015, Zhu et al 2016) show positive contributions. Only two studies (Zhu et al 2016 and Devaraju et al 2016) have analyzed the contribution of LULCC on LAI and NPP. Zhu et al (2016) find that the percent contribution of LULCC to the increase in global LAI is ∼4%, whereas, Devaraju et al (2016) estimates the percent contribution of LULCC to the NPP change to be ∼13% (negative effect on NPP). The reasons for the discrepancies between the models in the effects of climate change and LULCC on the increases in NPP/LAI may be the differences in both the climate simulated by the models and the response of models' carbon cycle to climate change, and differences in implementation of LULCC in the models (Peng et al 2017). A regional study over China (Piao et al 2015) for the period 1982–2009 using satellite observations and TRENDY climate models (which do not account for the LULCC) finds that the CO₂ fertilization is the primary driver (∼57% contribution) of change in satellite-observed LAI in the region, followed by increased N-deposition (27.6% contribution). According to this study, the contribution of climate warming is small (15.2% contribution) but negative to the greening trend because of the droughts in the region during the period.

Figure 3 shows the percentage contribution of the four factors to changes in land carbon fluxes (NBP, NEP), and land carbon stocks (TEC) during recent decades, and actual numbers including the net change from the studies are given in table S3. In general, CO₂ fertilization and N-deposition lead to increases in the land carbon uptake in the recent past. The percent contribution of CO₂ fertilization ranges from 33% (Devaraju et al 2016) to 85% (TRENDY study by Sitch et al 2015) with a mean of 55.1 ± 23.2%, while the percent contribution of N-deposition ranges from 10% to 24% with a mean of 15.8 ± 7.5% among the studies. Climate change and LULCC lead to a decline of land carbon stocks in the majority of the studies, and the mean of the relative contributions of these two factors respectively are 13.7 ± 11.8% and 39.1 ± 20.2%.

Zoom In Zoom Out Reset image size

The study by Huntzinger et al (2017) that assessed the major drivers of land carbon uptake during 1959–2010 as a part of the MsTMIP project shows large spread in model predictions due to differences in models' representations of ecosystem processes and structure, even though all the models are forced by same meteorology (see section 2.2). They find large differences in the net global land carbon uptake for the period 1959–2010 simulated by the models with the C–N cycle (net uptake of 31 ± 62.6 PgC) and the models without C–N cycle (net uptake of 93.3 ± 86.4 PgC). The models without coupled C–N cycle simulate a larger role for CO₂ fertilization during 1959–2010 (multi-model mean of percent contribution 60.2 ± 23.1%), and other two drivers-LULCC (21.2 ± 7.9%) and climate change (18.5 ± 17.3%) lead to land carbon loss. Their study shows that CO₂ fertilization still is the primary driver among the models with coupled C–N cycle (multi-model mean 43.0 ± 18.0% contribution), followed by LULCC (33.3 ± 23.4%, C loss), N-deposition (17.6 ± 12.1%, C gain), and climate change (6.1 ± 5.3%, C gain in majority of the models). Two of these models−CLM4 and CLM4-VN−however, simulate LULCC as the primary driver (56.3% and 51.1%, respectively, C loss). This is in agreement with the CESM1 result of Devaraju et al (2016), which uses CLM4 as its land model and finds that LULCC is the largest driver (43%) that leads to a reduction in the land carbon stocks during 1976–2005. However, in these three models (CLM4, CLM4-VN, and CESM1) the net land carbon uptake increases during the recent past, by combined effects of CO₂ fertilization and N-deposition. Both Devaraju et al (2016) and Huntzinger et al (2017) find that climate warming leads to a small gain of carbon in the models with the coupled C–N cycle because of increased N-mineralization due to warming.

Similar to the global studies, the regional modeling studies also find increases in the land carbon sink in recent decades and show that CO₂ fertilization is the dominant driver. Piao et al (2012) using the TRENDY models for the SE Asian region find that all factors except for climate change contributed positively to a land sink in the region during 1990–2009. They attribute ∼70% contribution from the combined effects of climate change and CO₂ fertilization, ∼26% from the increased N-deposition, and 3% due to LULCC. Similarly, Felzer and Jiang (2018) find that the increased land carbon sink since the 1950s in the US is due to CO₂ fertilization and N-deposition.

Recent studies that use DGVMs disentangle the LULCC effects into land use emissions and plant regrowth after the LULCC, and they recognize that plant regrowth and land management practices are important factors in mitigating the LULCC fluxes to the atmosphere and in increasing the land carbon sink in the past few decades. For instance, a recent study by Kondo et al (2018) using the TRENDY results finds that plant regrowth in eastern North America, southern-eastern Europe, and southeastern temperate Eurasia was an important factor for the increases in the global land sink during recent past. According to their results, the approximate global net carbon flux difference between 2000–2009 and 1960–1999 is +1.27 ± 0.34 PgC yr⁻¹ (land as a sink), and the percentage contributions of the factors are, CO₂ fertilization ∼64.86% (+1.2 ± 0.25 PgC yr⁻¹), climate warming ∼13.51% (−0.25 ± 0.11 PgC yr⁻¹), LULCC 2.35% (−0.04 ± 0.23 PgC yr⁻¹), and plant regrowth 19% (+0.33 ± 0.10 PgC yr⁻¹). Piao et al (2018) using the TRENDY models find that the reduced LULCC fluxes due to reduced deforestation in the tropics (in SE Asia and South America) and afforestation in northern hemisphere temperate regions (net LULCC contribution ∼69%) led to an increased global land sink during the slow warming period 1998–2012. They attribute only about 15% contribution to CO₂ fertilization to the intensification of the land carbon sink during the period, and they find that climate warming nearly offsets this contribution.

3.1.3. Future climate scenarios

Majority of the CMIP5 models (e.g. Jones et al 2013, Friend et al 2014, Arora and Boer, 2014, Mahowald et al 2016) simulate increases in LAI, NPP, and cumulative land carbon uptake by the end of 21st century in all three future scenarios considered in the present review (RCP2.6, RCP4.5, and RCP8.5), although estimates of the land carbon uptake by 2100 show large inter-model spread (Jones et al 2013). The tropics and northern latitudes become major carbon sinks in most of the CMIP5 models by 2100. The ranges of cumulative land carbon uptake simulated by the CMIP5 models are, −45 PgC to +220 PgC for RCP2.6, +55 PgC to +460 PgC for RCP4.5, and −190 PgC to +490 PgC for RCP8.5 (Jones et al 2013). The models with coupled C–N cycle (CESM and NorESM, which share the same land model) simulate relatively less cumulative land carbon uptake in the future scenarios compared to other CMIP5 models due to N-limitation that weakens these models' response to increases in CO₂. For instance, CESM1 simulates a cumulative land source in the RCP2.6 (−21 PgC between 2100 and 2005) and RCP8.5 (−27 PgC) scenarios, and a weaker land sink in the RCP4.5 scenario (+55 PgC) compared to many of the CMIP5 models (Lawrence et al 2012, Tharammal et al 2019).

Some studies have suggested that the projected increase in the land carbon stocks in CMIP5 models is primarily contributed by CO₂ fertilization (Jones et al 2013, Brovkin et al 2013, Friend et al 2014), despite climate warming in the future scenarios and deforestation in the RCP2.6 and RCP8.5 scenarios. The inferred contributions of three major drivers (CO₂ fertilization, climate change, and LULCC) to the mean change in global TEC in the RCP2.6 and RCP8.5 scenarios for the four CMIP5 models from the LUCID project (Brovkin et al 2013; the selection criteria and calculation of contributions for the LUCID models are explained in section 2.1) are shown in figure 4 and table S4. CO₂ fertilization effect is positive for all the models in both the scenarios. The sum of the effects of three drivers and the net change for the individual models for the RCP2.6 scenario agree in sign for CanESM and MIROC ESM, both of which, simulate a net decline in the land carbon stock in the 21st century due to climate warming and LULCC. Net change in HADGEM2 for the RCP2.6 scenario (an increase in land C uptake in 21st century) disagrees in sign with the calculated sum of three drivers, and IPSL-CM5A-LR does not include the LULCC effect in the RCP2.6 scenario. The LUCID models agree on the negative contribution of LULCC to the land carbon uptake in the RCP2.6 and RCP8.5 scenarios, despite substantial spread in the values (see Brovkin et al 2013; table S4). Similarly, Quesada et al (2018) used five of the LUCID-CMIP5 models to assess the effects of LULCC in the RCP8.5 scenario, and they find that the projected increases in global greening and the land carbon stocks in the scenario are dampened in the models by approximately 22% and 24%, respectively due to LULCC.

Zoom In Zoom Out Reset image size

The sum of effects of three drivers agrees in sign with the net change in the 21st century in TEC for all the four LUCID models in the RCP8.5 scenario (figure 4 bottom panel). CO₂ fertilization offsets the decline in the land carbon stocks caused by climate warming and LULCC in the models except for MIROC ESM, for which, the larger climate warming effect causes a net decline. The large spread in the net changes of TEC among the LUCID models is because of the differences in the parameterizations of the processes such as photosynthesis, carbon allocation, and different implementations of the wood harvest/grazing in the models (Brovkin et al 2013, Boysen et al 2014). The mismatch in magnitudes between the net changes of TEC in the 21st century and the sum of three factors for both the scenarios (figure 4) is likely caused by the sensitivity factors used for the calculation. The sensitivity factors for individual models used in the calculations here (from the '1% yr⁻¹' CO₂ experiments by Arora et al 2013) may differ for the RCP scenarios, due to the differences in the CO₂ concentrations and LULCC in respective simulations. A detailed discussion of the sensitivity factors is given in section 3.2.

Tharammal et al (2019), using CESM1 estimated the contributions of the four major drivers of changes in land carbon uptake in the RCP2.6, RCP4.5, and RCP8.5 scenarios. The percent contributions of major drivers to the changes in TEC in these scenarios estimated from Tharammal et al (2019) are shown in figure 5 and table S5. They find that LULCC (tropical deforestation) is the primary driver of the decline in the land carbon stocks in the low-emission RCP2.6 scenario. The percent contributions of the drivers to the net TEC change in this scenario (changes in TEC in the year 2100 relative to 2005 in brackets) are LULCC 55.6% (−38.76 PgC), CO₂ fertilization 28.8% (+20.08 PgC), climate change 9.9% (−6.9 PgC), and N-deposition 5.5% (3.9 PgC). In the RCP4.5 scenario, CO₂ fertilization is the major driver of the increase in the global land carbon uptake [∼49% (+52.05 PgC)], followed by the positive contribution of LULCC to the land carbon sink [due to afforestation, 27% (+28.71 PgC)]. Climate change causes a reduction in land carbon uptake [∼14% contribution (−14.85 PgC)] and reduced N-deposition rates in the RCP4.5 scenario in 2100 with respect to 2005 cause a reduction in the land carbon uptake [10% contribution (−10.70 PgC)]. In the RCP8.5 scenario, CESM1 simulates a land source by 2100, mainly due to the effects of LULCC (deforestation). Weakened CO₂ fertilization in CESM1 due to N-limitation is likely the cause for the differences in the net land carbon uptake between CESM1 and other CMIP5 models. The percent contribution of each of the factors in the RCP8.5 scenario is: LULCC 44% (−93.14 PgC), CO₂ fertilization 41.2% (+87.21 PgC), climate change 12.4% (−26.20 PgC), and N-deposition 2.3% (4.92 PgC). The contribution of N-deposition in future scenarios is smaller when compared to the historical period (2%–10% versus ∼20%, see Devaraju et al 2016, Tharammal et al 2019).

Zoom In Zoom Out Reset image size

Following Friedlingstein et al (2003), various studies on the land carbon uptake have estimated the sensitivity of the change in the land carbon stocks to (i) the increase in CO₂ concentration (β) and (ii) to climate warming (γ). Bala et al (2012) define another sensitivity factor—change in the land carbon stocks due to N-deposition (δ). In this section, we review the model-simulated sensitivity factors for the historical and future climate scenarios, along with those estimated from idealized simulations.

β is calculated as the change in the land carbon stocks per unit increase in atmospheric CO₂ when other drivers are kept constant, and similarly, γ is calculated as the changes in the land carbon stocks per unit change in global mean surface temperature, while, δ is calculated as the changes in land carbon stocks for a unit increase in the rate of N-deposition. Table 1 summarizes these sensitivities from various studies. Several values from the C4MIP studies and equilibrium simulations are adapted from Friedlingstein et al (2006) and Bala et al (2012). The concentration-carbon sensitivity factor β is positive in sign irrespective of the time period, climate scenarios, and different models used. However, the β values show a broad range (0.17 PgC ppm⁻¹ in the RCP8.5 scenario in Tharammal et al (2019) to 3.38 PgC ppm⁻¹ in the historical period from Huntzinger et al (2017), see table 1). Friedlingstein et al (2006) find that different model responses to NPP cause the inter-model differences in the β values. Another notable feature is that the β values vary with the emission scenarios and the state of the climate system (Boer and Arora 2009, 2013, Bala et al 2012, Arora et al 2013, Hajima et al 2014). The β values calculated from transient simulations are smaller compared to near-equilibrium simulations, due to the lags in ecosystem response (Bala et al 2012, 2013). A slower rate of CO₂ increase in a scenario leads to larger land CO₂ accumulation, and higher values of β (Hajima et al 2014). The models with coupled C–N biogeochemistry simulate smaller β values compared to models without the C–N biogeochemistry (e.g. Thornton et al 2007, Arora et al 2013, Devaraju et al 2016, Huntzinger et al 2017, Tharammal et al 2019, see table 1), as the N-limitation in these models weakens the CO₂ fertilization effect. Similarly, the β values in the coupled C–N models decrease with increasing atmospheric CO₂ in the RCP scenarios (see table 1, Tharammal et al 2019) due to N- limitation with the increases in CO₂ fertilization (Thornton et al 2007).

The climate-carbon sensitivity factor (γ) is negative for the historical period and future scenarios and for other transient or equilibrium simulations (see table 1), which suggests that the land loses carbon with increasing temperature. The reason for the decline in carbon sink with the temperature is primarily the acceleration of the decaying process and the consequent increases in the heterotrophic respiration. Further, climate warming reduces NPP especially in the tropical latitudes in the historical period and the future scenarios (e.g. Devaraju et al 2016, Tharammal et al 2019). The magnitudes of γ values are dependent on the models used and the state of the system, and they vary between 0.2 PgC K⁻¹ in Bonan and Levis (2010) and ∼180 PgC K⁻¹ in Bala et al (2013). The magnitudes of γ values are generally smaller in the coupled C–N cycle models compared to the models without the C–N cycle as the N-mineralization due to warming and the consequent increase in productivity counteracts the carbon loss due to increased heterotrophic respiration in these models (Thornton et al 2007, Bonan and Levis 2010, Arora et al 2013, Devaraju et al 2016, Huntzinger et al 2017, Tharammal et al 2019).

The sensitivity of the land carbon uptake to the N-deposition (δ) is positive in sign (see table 1), which implies N-deposition increases the productivity and land carbon sink. However, like the other two sensitivity factors, the values show a considerable range (0.37 PgC/TgN yr⁻¹) in the transient RCP2.6 simulation by Tharammal et al (2019) to 3.41 PgC/TgN yr⁻¹ in an equilibrium simulation by Bala et al (2013). The interaction with other factors influences the δ values; for instance, Bala et al (2013) find that the δ values increase with CO₂ fertilization, and decrease with the amount of climate warming. Similarly, Tharammal et al (2019) attribute the larger δ values simulated in the RCP4.5 scenario compared to their RCP2.6 and RCP8.5 simulations (see table 1) to the interaction of N-deposition with LULCC (afforestation in the RCP4.5 scenario).

Studies that define a sensitivity factor for the land C uptake due to net LULCC are not available in the literature, as it is difficult to uniquely quantify the LULCC, which include conversion to/from different plant functional types, harvest, grazing, and land use fluxes. Recently, Jones et al (2018) decomposed the γ and β values to non-human mediated and human mediated effects, and the latter is further divided into concentration-land cover (ε), climate-land cover (η), and land cover-carbon (μ) effects. They describe the human-induced effect of land cover change in terms of changes in cropland area. The sensitivity factor μ reflects the amount of land carbon change associated with the land conversion for agriculture; and μ is estimated as −16 PgC Mkm⁻² of cropland for the RCP8.5 LULCC. Further efforts to define a sensitivity factor for LULCC will help to better understand the inter-model spread of land carbon uptake in the future projections.

4.1.1. Consensus

Large uncertainties and disagreements are found among the modeling studies regarding the relative contributions of the drivers, and even in the sign of changes in land carbon uptake in simulations for the same climate scenarios. Further, many studies that assessed the CMIP5 earth system models' performance in simulating the present-day global land carbon cycle (Anav et al 2013, Jones et al 2013, Todd-Brown et al 2013, Koven et al 2015, Schurgers et al 2018) find broad range in the present-day vegetation carbon stock and soil carbon stocks. Lovenduski and Bonan (2017) find that ∼80% of the uncertainties in the projected terrestrial carbon uptake in the CMIP5 models are driven by differences in model structure. In addition, different definitions and approaches used in climate models to estimate the LULCC might have added to the uncertainties (Poulter et al 2011, Peng et al 2017). The offline TRENDY model results and MsTMIP models (tables S1, S3) show that there is large intermodel spread even when forced by same climate forcings. The importance of incorporating nutrient limitation, especially the C–N dynamics to the global land models has been studied extensively and it is highly likely that N-limitation substantially reduces the CO₂ fertilization effect in models with coupled C–N biogeochemistry (Sokolov et al 2008, Jain et al 2009, Thornton et al 2009, Bonan and Levis 2010, Zaehle et al 2010a, Goll et al 2012, Zaehle et al 2013, 2015, Brovkin and Goll, 2015, Wieder et al 2015). However, many of the models discussed in the present review including the CMIP5 models (Jones et al 2013) do not represent the N- and phosphorus limitation in their land carbon cycle, and hence are likely to overestimate the future land carbon uptake (Wieder et al 2015, Zaehle et al 2015). Other reasons for these inter-model differences are dependence of the simulated carbon stocks on factors such as initial carbon pool size (Brovkin et al 2013), and carbon cycle responses to climate biases in the model (Ahlström et al 2017). Biases in the climate forcing data can also lead to large uncertainties (Schimel et al 2015b, Wu et al 2017, 2018). Hence, it is crucial to reduce the uncertainties from the climate data for a better estimation of the land carbon uptake.

The main limitation of many of the studies discussed in the present review (see table S1) is that they study the role of only one or two of the major factors. Further, a large number of the studies that analyze the drivers of historical land carbon uptake use offline land models (see table S1; e.g. MsTMIP models (Huntzinger et al 2017), TRENDY model studies—Piao et al 2012, Piao et al 2015, Sitch et al 2015, Zhu et al 2016, Kondo et al 2018, Piao et al 2018), which lack atmospheric feedbacks. Uncertainties in the meteorological forcing (Ahlström et al 2017) used to force these offline models can lead to large biases in the simulated land carbon cycle. Further, the first version of the TRENDY studies (Sitch et al 2015), and some other studies we considered (Piao et al 2015, Zhang et al 2016) do not account for the land use changes during the historical period. Several regional and global studies (Piao et al 2012, Chang et al 2016, Kondo et al 2018, Piao et al 2018, see table S3) suggest that reduced land use fluxes through land management contribute positively to the net land carbon uptake in recent decades, while a few others show negative contributions of LULCC (Eglin et al 2010, Devaraju et al 2016, Huntzinger et al 2017), hence overlooking this factor might have caused under/over-estimation of recent land carbon uptake in the studies that omit LULCC.

Aggressive tuning and validation of the models to reproduce the past carbon cycle may reduce the uncertainties, however, large uncertainties in the observational data might still lead to the spread in the predictions. Lovenduski and Bonan (2017) devised a weighing scheme for the CMIP5 models based on their ability to reproduce the observed land carbon uptake and they find that this constraint led to a reduction in the uncertainty in the future projections by ∼20%. One of the major challenges for the modeling studies is constraining present-day carbon cycle simulated by the models with observations, as we lack direct measurements of the land carbon fluxes, and available observations rely primarily on the in-situ measurements or indirect satellite estimates (Ciais et al 2014, Schimel et al 2015b, Sellers et al 2018). Interestingly, emergent constraints derived from the relationships between the observable variables (e.g. interannual variability in CO₂, temperature) and present-day sensitivity of the land carbon stocks to climate variables have been used recently to reduce the uncertainties in the future projections of GPP, NBP, and the sensitivity to long term climate change (Cox et al 2013, Wenzel et al 2014, 2016, Mystakidis et al 2016, Hall et al 2019, Winkler et al 2019).

Ecosystem processes that are missing in climate models can cause further uncertainties to projections of future land carbon uptake. Although all CMIP5 models predict reduced land carbon uptake with climate warming, recent modeling studies that incorporated the plant acclimation parameterizations (photosynthetic plus root respiratory) at higher temperatures show that there could be an enhancement in land carbon uptake in the future warming scenarios (Arneth et al 2012, Lombardozzi et al 2015, Mercado et al 2018). However, the extent to which these processes will affect the future carbon uptake is uncertain as these parameterizations are in the early stages of validation/implementation in climate models. Similarly, permafrost thawing, a process yet to be represented in most earth system models can cause a release of carbon from higher latitudes in a warming climate and may enhance the positive carbon-climate feedback (Burke et al 2013, Koven et al 2013). Additionally, the prescribed LULCC for future scenarios in the CMIP5 models do not allow climate-induced land cover changes (CILCC). Davies-Barnard et al (2015) using HadGEM2-ES with the DGVM TRIFFID (Top-down Representation of Interactive Foliage and Flora Including Dynamics) find that the forest expansion due to CILCC could partly offset the effects of deforestation in the RCP8.5 scenario by 2100.

Although the modeling studies that we reviewed here indicate CO₂ fertilization effect is the largest contributing factor to increases in NPP and land carbon stock, the strength of the CO₂ fertilization effect is debated (Ahlström et al 2013, Todd-Brown et al 2013, Wieder et al 2015, Smith et al 2016). While the CMIP5 simulations show an increase in the land carbon sink in the future, Green et al (2019) warn that the increasing trend in the uptake may not be sustained past the middle of the century due to trends in soil moisture. Similarly, a recent observational study by Giguère-Croteau et al (2019) finds that even with increased CO₂ fertilization and increased water use efficiency, factors such as nutrient availability, carbon allocation strategies, and stomatal density might limit the increase in above-ground biomass, and the sink capacity of forests. These findings demand adequate assessments of the CO₂ fertilization effect under varied forcings so that the land sink in the future projections is not overestimated. Large-scale Free‐Air CO₂ Enrichment experiments (FACE, Norby and Zak (2011)) in nutrient/temperature/water-limited ecosystems can be used to validate the CO₂ fertilization effect simulated by the models.

There is large inter-model and inter-scenario spread in the sensitivity of the land carbon uptake to an increase in CO₂, climate change, and N-deposition (β, γ, and δ values) in the historical and future scenarios. These three sensitivities depend on the model's structure and climate scenario as discussed in the review by Hajima et al (2014). This uncertainty in the estimation of the sensitivity factors would lead to uncertainties in future land carbon uptake. Further, a few of the studies considered here use offline land models (Thornton et al 2007, Bonan and Levis 2010, Zaehle et al 2010a, Bala et al 2013, Huntzinger et al 2017), and therefore, the atmosphere-land interactions are absent from the computations of their γ and β values.

Although advances in remote sensing can ease the measurements of column CO₂ and GPP (via solar-induced fluorescence, Frankenberg et al 2011), more observed data products for carbon cycle variables, both in-situ and from space, are required for improving the model predictions (Ciais et al 2014, Pugh et al 2016, Rogers et al 2017, Macbean et al 2018, Sellers et al 2018). Further, it would be desirable to conduct multi-model studies with factorial experiments to assess the major drivers of the land carbon uptake with high confidence. It is expected that many models in the forthcoming C4MIP project, which is a part of the sixth phase of CMIP (CMIP6) to include improved representations of ecological processes, such as coupled C–N cycle, permafrost soil carbon dynamics, and improved spatial resolution (Jones et al 2018). These models are expected to provide a better-constrained estimate of the land carbon stocks in the future.