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

Terrestrial and oceanic carbon sinks together sequester >50% of the anthropogenic emissions, and the major uncertainty in the global carbon budget is related to the terrestrial carbon cycle. Hence, it is important to understand the major drivers of the land carbon uptake to make informed decisions on climate change mitigation policies. In this paper, we assess the major drivers of the land carbon uptake—CO2 fertilization, nitrogen deposition, climate change, and land use/land cover changes (LULCC)—from existing literature for the historical period and future scenarios, focusing on the results from fifth Coupled Models Intercomparison Project (CMIP5). The existing literature shows that the LULCC fluxes have led to a decline in the terrestrial carbon stocks during the historical period, despite positive contributions from CO2 fertilization and nitrogen deposition. However, several studies find increases in the land carbon sink in recent decades and suggest that CO2 fertilization is the primary driver (up to 85%) of this increase followed by nitrogen deposition (∼10%–20%). For the 21st century, terrestrial carbon stocks are projected to increase in the majority of CMIP5 simulations under the representative concentration pathway 2.6 (RCP2.6), RCP4.5, and RCP8.5 scenarios, mainly due to CO2 fertilization. These projections indicate that the effects of nitrogen deposition in future scenarios are small (∼2%–10%), and climate warming would lead to a loss of land carbon. The vast majority of the studies consider the effects of only one or two of the drivers, impairing comprehensive assessments of the relative contributions of the drivers. Further, the broad range in magnitudes and scenario/model dependence of the sensitivity factors pose challenges in unambiguous projections of land carbon uptake. Improved representation of processes such as LULCC, fires, nutrient limitation and permafrost thawing in the models are necessary to constrain the present-day carbon cycle and for more accurate future projections.


Introduction
Fossil fuel combustion in the industrial era has perturbed the global carbon (C) cycle by releasing a significant amount of CO 2 to the atmosphere (IPCC 2013). Terrestrial and oceanic carbon uptake offset the anthropogenic emissions of CO 2 by absorbing more than half of these emissions. The land biosphere gains carbon during photosynthesis (as measured by gross primary production (GPP) on large scales), and it releases carbon through respiration (both autotrophic and heterotrophic respiration−R h ). The land can be a sink or source of carbon to the atmosphere depending on the balance between the uptake and the net respiratory flux of carbon to the atmosphere.
The terrestrial biosphere is a sink for carbon in recent decades due to mainly the increases in net primary production (NPP=GPP − R h ), driven by the increases in atmospheric CO 2 (Friedlingstein et  identified CO 2 fertilization, nitrogen (N)-deposition, climate change, and land use/land cover changes (LULCC) as the major factors that affect the terrestrial primary production and land carbon uptake (figure 1). CO 2 fertilization increases photosynthesis (figure 1) and NPP due to the direct effect of increases in atmospheric CO 2 (Farquhar et al 1980, Kimball et al 1993, Thornton et al 2007, Kimball 2010. Availability of nutrients (mainly N and phosphorus (P)) could limit the primary production, even at higher CO 2 levels. Modeling studies have found that N is a key limiting factor for most of the temperate and boreal ecosystems (Sokolov et al 2008, Zaehle et al 2010b, Bala et al 2012, Fisher et al 2012, Goll et al 2012. Reactive forms of nitrogen (Nr)-the reduced NH x and oxidized form NO y are created mainly by human activities (e.g. Haber-Bosch process and fossil fuel combustion), and increased N-deposition stimulates vegetation productivity. Therefore, the future land carbon uptake is expected to partly depend on the changes in the N-deposition rates driven by anthropogenic activities.
Increased CO 2 in the atmosphere leads to climate change, which affects the temperature, precipitation, length of the growing season, and heterotrophic respiration. Increases in precipitation and length of the growing period can lead to increased NPP (figure 1), whereas, reduced precipitation and increased temperature (in the tropics) could limit NPP. A warmer climate increases heterotrophic respiration and reduces the net ecosystem production (NEP=NPP − R h ), and hence, weakens the land sink. The ∼5% increase in the air-borne fraction of CO 2 emissions in the last 50 years is attributed to the reduction in the land and oceanic carbon sinks, in response to climate change (Le Quéré et al 2009).
LULCC fluxes include changes in the land carbon fluxes due to deforestation, afforestation, wildfires, wood harvesting, grazing and shifting cultivation. The effects of LULCC include both biogeophysical (changes in surface albedo, evapotranspiration, surface roughness) and biogeochemical processes (Feddema 2005, Bala et al 2007, Boysen et al 2014, which can affect both regional and global climate. Cumulative LULCC contributed ∼25% to the total carbon emissions during the period 1870-2015(Le Quéré et al 2018. However, the annual contribution of LULCC to total emissions has reduced to ∼10% during recent decades (Houghton 2012, Le Quéré et al 2015, Houghton and Nassikas 2017. Atmospheric CO 2 concentration has increased by more than 40% since the pre-industrial period (∼280 to ∼410 ppm; Meinshausen et al 2011, Ciais et al 2013 and the global mean temperature has increased by about 1°C (IPCC 2013), while N-deposition rates increased from ∼30 TgN yr −1 in 1850 to ∼80 TgN yr −1 in the 2000s (Lamarque et al 2013, Figure 1. Schematic of the major four drivers (CO 2 fertilization, nitrogen (N)-deposition, climate change, and land use and land cover changes-LULCC) of land carbon uptake. Downward arrows represent the carbon flux into the ecosystem. Upward arrows represent the carbon flux from the land to the atmosphere. Positive (+) and negative (-) signs denote the effect of the factor on the land carbon uptake. Kanakidou et al 2016). During the industrial era, forests were converted to croplands and pastures to accommodate the increase in population and hence, agriculture.
Climate models constitute an important tool to estimate the contribution of the aforementioned four drivers to the land carbon uptake. The fifth phase of the Coupled Model Intercomparison Project (CMIP5, Taylor et al 2012) includes a range of simulations conducted for the historical period and future climate scenarios with Earth System Models (ESMs) that incorporate a full representation of global carbon cycle when compared to earlier CMIP phases. The projections for the 21st century from the CMIP5 models are based on four different socio-economic scenarios, termed as Representative Concentration Pathways (RCPs) (Moss et al 2010), which differ by the extent of climate mitigation policies. The majority of CMIP5 models agree with the observationally constrained estimates that the land was a source of carbon to the atmosphere during the historical period (Jones et al 2013). For the 21st century, results from CMIP5 show a large spread in land carbon fluxes and stocks (Jones et al 2013, Friedlingstein et al 2014. Although many published model-based studies assess the changes in regional and global land carbon uptake during the historical period and in the projections of future climate, only few studies have analyzed the role of the drivers of these changes. In this paper, using mainly the results from CMIP5 coupled models that include a land carbon cycle component, we synthesize the results from published literature and provide a comprehensive review of the major drivers of changes in the land carbon uptake during the historical and three future RCP scenarios-RCP2.6 (high mitigation scenario, the radiative forcing in the scenario by 2100 is 2.6 W m −2 ; Van Vuuren et al 2007), RCP4.5 (medium mitigation, radiative forcing in the scenario is 4.5 W m −2 ; Thomson et al 2011), and RCP8.5 (high emission scenario, radiative forcing in the scenario is 8.5 W m −2 ; Riahi et al 2007). The RCP2.6 and RCP8.5 scenarios are associated with extensive deforestation, increases in the N-deposition, and increases in CO 2 , while the RCP4.5 scenario includes afforestation and a reduction in N-deposition rates by 2100. A detailed overview of the climate forcings in the historical and the RCP scenarios is given in section 2.2.

Selection of literature
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, CO 2 fertilization, nitrogen deposition, LULCC, LUC, deforestation, climate warming, drivers of terrestrial land carbon' were conducted with webbased 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 (CO 2 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, Brovkin et al 2013, Devaraju et al 2016, Huntzinger et al 2017, Piao et al 2018 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 CO 2 fertilization and climate warming for these models by using the individual model's land carbon uptake sensitivity to CO 2 (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 CO 2 experiments by Arora et al (2013) (conducted by increasing the CO 2 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 CO 2 , climate, and LULCC on the net biome production (NBP) and land carbon uptake during the present-day and recent past 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 CO 2 , 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 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 (

An overview of the major drivers
For the historical period, apart from the increase in atmospheric CO 2 , climate warming, and increased nitrogen deposition rates (see Introduction), the estimated loss of primary land is ∼75 million km 2 (125 million km 2 in 1850 to 50 million km 2 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-CO 2 greenhouse gases (GHG), aerosols and ozone.
The CMIP5 historical and RCP simulations are 'concentration driven' or forced by prescribed CO 2 concentrations, rather than emissions-driven. Other forcing factors include aerosols, non-CO 2 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 CO 2 concentration is projected to increase by 41 ppmv, 159 ppmv, and 556 ppmv respectively in the RCP2.6, RCP4.5, and RCP8. 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).

Results
3.1. Contribution of major drivers to the land carbon uptake 3.1.1. Historical period  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=10 15 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

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 ( 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  increases in NPP and LAI from four global studies and one regional study. The largest contribution is from CO 2 fertilization and its percent contribution to increases in global NPP/LAI ranges from 43% (Devaraju et al 2016) to 86%  (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 CO 2 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, CO 2 fertilization and N-deposition lead to increases in the land carbon uptake in the recent past. The percent contribution of CO 2 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%.
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 CO 2 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 CO 2 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 CO 2 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 CO 2 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 CO 2 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 CO 2 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, CO 2 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 CO 2 fertilization to the intensification of the land carbon sink during the period, and they find that climate warming nearly offsets this contribution. 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 CO 2 . 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 Some studies have suggested that the projected increase in the land carbon stocks in CMIP5 models is primarily contributed by CO 2 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 (CO 2 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. CO 2 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  (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.

Future climate scenarios
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). CO 2 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 ' CO 2 experiments by Arora et al 2013) may differ for the RCP scenarios, due to the differences in the CO 2 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), CO 2 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, CO 2 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 CO 2 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), CO 2 fertilization 41.2%

Sensitivity factors
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 CO 2 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 CO 2 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  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 ( 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 intermodel spread of land carbon uptake in the future projections.

Discussions and conclusions
4.1. Major drivers 4.1.1. Consensus This review assesses the contributions of the major drivers of the land carbon uptake in the historical and future scenarios. During the period 1850-1950, LULCC has primarily led to the decline in the land Table 1. Sensitivity of global land carbon uptake to CO 2 (β in PgC ppm −1 ), climate warming (γ in PgC K −1 ) and N-deposition rate (δ in PgC/(TgN yr −1 )) from various studies for different time periods and climate scenarios.
Friedlingstein et al (  Results from the majority of the CMIP5 models indicate that CO 2 fertilization would lead to increases in the land carbon uptake in the future scenarios (Jones et al 2013). There is agreement among the CMIP5-based studies on the decline of the land carbon stocks in the RCP2.6 and RCP8.5 scenarios due to the direct effect of LULCC, i.e. removal of biomass through deforestation (Brovkin et al 2013, Tharammal et al 2019), however, in contrast to the majority of CMIP5 models, in CESM1, the larger effect of LULCC in the RCP2.6 and RCP8.5 scenarios offsets the CO 2 fertilization effect. Similarly, Müller et al (2007) through simulations using the SRES scenarios find that land use changes, especially deforestation, cause substantial changes in the future land carbon uptake, and can offset the effects of CO 2 fertilization and climate change. We infer that climate warming leads to a loss of TEC in the three future scenarios (see Simulations with CESM1 show that the increase in TEC the RCP4.5 scenario is primarily due to CO 2 fertilization and afforestation (Tharammal et al 2019). However, a lack of sensitivity studies for the RCP4.5 scenario limits the analysis of major drivers of changes in the scenario using multiple models.

Problems and challenges
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  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 CO 2 fertilization effect in models with coupled C-N biogeochemistry . 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 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 ( 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 (Topdown 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 CO 2 fertilization effect is the largest contributing factor to increases in NPP and land carbon stock, the strength of the CO 2 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 CO 2 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 CO 2 fertilization effect under varied forcings so that the land sink in the future projections is not overestimated. Large-scale Free-Air CO 2 Enrichment experiments (FACE, Norby and Zak (2011)) in nutrient/temperature/water-limited ecosystems can be used to validate the CO 2 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 CO 2 , 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 ( 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.