Quantifying the Effects of Historical Land Cover Conversion Uncertainty on Global Carbon and Climate Estimates

Previous studies have examined land use change as a driver of global change, but the translation of land use change into land cover conversion has been largely unconstrained. Here we quantify the effects of land cover conversion uncertainty on the global carbon and climate system using the integrated Earth System Model. Our experiments use identical land use change data and vary land cover conversions to quantify associated uncertainty in carbon and climate estimates. Land cover conversion uncertainty is large, constitutes a 5 ppmv range in estimated atmospheric CO2 in 2004, and generates carbon uncertainty that is equivalent to 80% of the net effects of CO2 and climate and 124% of the effects of nitrogen deposition during 1850–2004. Additionally, land cover uncertainty generates differences in local surface temperature of over 1°C. We conclude that future studies addressing land use, carbon, and climate need to constrain and reduce land cover conversion uncertainties.


Introduction
Global socioeconomic and Earth system modeling efforts, such as phase 5 of the Coupled Model Intercomparison Project (CMIP5)  for the Fifth Assessment Report (AR5) of the Intergovernmental Panel on Climate Change, aim to provide understanding of potential climate change given scenarios of human economic and agricultural activity. The Representative Concentration Pathways (RCPs)  prescribe the amounts of anthropogenic emissions and land use change used by Earth System Models (ESMs) to estimate atmospheric CO 2 concentration and climate change (Intergovernmental Panel on Climate Change, 2013). Recent advances have improved communication between these modeling communities through dataset harmonization for common and consistent anthropogenic forcing of ESMs (Hurtt et al., 2011;Lamarque et al., 2010;Van Vuuren et al., 2011). However, land use change is uniquely implemented in each ESM with differences in land cover representation, definitions, conversion processes, and assumptions Pitman et al., 2009). Furthermore, although land use was harmonized for CMIP5 (Hurtt et al., 2011), each ESM used its own land cover distribution and conversion approach because the ESMs were structured to apply exogenous land use to endogenous land cover implementations. As Land Use and Land Cover Change (LULCC) has both biophysical (e.g., Brovkin et al., 2013;A. D. Jones et al., 2013;Pitman et al., 2009) and biogeochemical (e.g., Arora & Boer, 2010 different implementations of the same land use scenario can constitute vastly different ESM LULCC scenarios, with corresponding differences in regional and global carbon (e.g., Di Vittorio et al., 2014) and climate (e.g., A. D.  projections. The contribution of LULCC uncertainty to carbon and climate projections is important for understanding potential global change impacts, as many climate mitigation and adaptation strategies rely on local LULCC (e.g., Rose et al., 2012;Smith & Rothwell, 2013;Van Vuuren et al., 2011) with corresponding effects on carbon (e.g., Jain & Yang, 2005) and climate (e.g., Bright et al., 2017). Assessment of such strategies can be confounded if LULCC uncertainty is comparable to intrascenario or interscenario differences. For example, Peng et al. (2017) estimated a 1990 forest area uncertainty (2.9 M km 2 ) due to historical land conversion uncertainty that is~161% of the estimated increase in RCP2.6 forest area from 2005 to 2100 and~242% of the difference in estimated 2100 forest area between RCPs 2.6 and 6.0 (Hurtt et al., 2011). The uncertainties in net LULCC emissions (Ciais et al., 2013, Figure 6.10;Houghton et al., 2012;Le Quéré et al., 2015) and residual land-atmosphere CO 2 flux (Ciais et al., 2013, Figure 6.16) are already large without accounting for land cover conversion uncertainty. Accounting for this uncertainty increases emissions uncertainty (Peng et al., 2017) and affects the significance of land use strategies aiming to reduce emissions. Therefore, increasing accuracy and evaluation of LULCC, and in particular land cover change, is paramount for understanding global change.
Given the significant influence of LULCC on carbon and climate, a primary question remains largely unexplored: How large are uncertainties associated with the translation of land use change information into land cover change, in terms of global carbon and climate? A primary obstacle to exploring this question has been the limited LULCC flexibility in Earth system models (e.g., Brovkin et al., 2013;Pitman et al., 2009). However, the integrated Earth System Model (iESM) Collins et al., 2015;Di Vittorio et al., 2014;A. D. Jones et al., 2013;Thornton et al., 2017) provides a unique structure for addressing this question.
Here we use this model to quantify the envelope of uncertainties associated with land conversion assumptions and their effects on the global carbon-climate system during 1850-2004. We compare these uncertainties to those of CO 2 fertilization, climate change, and nitrogen deposition, and also analyze effects on local climate.
The iESM translates GLM land use change into iESM land cover change each year using an LUT with adjustable land cover conversion assumptions (Di Vittorio et al., 2014;Lawrence et al., 2012). Historical land use is provided globally at half degree grid cell fractional resolution and includes cropland, pasture, urban, secondary, and primary vegetation, annual transitions between these categories, and both area and amounts of wood harvest (Hurtt et al., 2011). The LUT uses only annual crop and pasture and harvested area information to coincide with iESM land model implementation (Community Land Model v4, CLM) (Lawrence et al., 2011). Given user-specified land cover conversion assumptions, the LUT converts GLM land use change to changes in CLM land cover, which lacks pasture and comprises 16 Plant Functional Types (PFTs): bare ground, eight trees, three grasses, three shrubs, and one crop. The initial GLM pasture area is assigned first to grass PFTs, then to shrubs, and finally to trees, as needed. All new pasture is added as grass. A unique feature of the LUT is that it can track existing pasture in relation to CLM PFTs throughout a simulation. The LUT can be configured with different reference years for calculating LULCC and with land cover conversion assumptions ranging continuously between maximizing and minimizing forest area. There are separate parameters for expansion and contraction of agriculture, and each of these parameters defines the relative amounts of Geophysical Research Letters

10.1002/2017GL075124
forest versus grass/shrub PFTs to convert. The center of this assumption range converts land cover proportionally to existing (for agricultural expansion) or potential (for agricultural contraction) PFT coverage.

Historical LULCC Simulations 2.2.1. Land Model Only Simulations
We performed eight, half-degree, global, land only simulations to separate the effects of LULCC and atmospheric inputs on carbon (Table 1). These simulations varied in reference year, land cover conversion assumptions, and atmospheric forcing. We used two standard LULCC configurations and designed three more to span a maximum range of land cover conversion. These configurations are based on different reference years and land cover conversion assumptions, but with identical land use data and the same initial PFT distribution in 1850. "No LULCC" with no wood harvest is a standard reference configuration for estimating net LULCC emissions and ecosystem carbon changes due to LULCC. These estimates are based on the difference between another simulation and this No LULCC case, with net emissions constituted by net ecosystem exchange minus wildfire emissions. The "Default" case is the standard LULCC configuration for CESM, and it uses year 2000 as the reference for calculating each year's land use/cover distribution (Lawrence et al., 2012). The three other LULCC configurations use the previous year for reference (chronological), and are used to quantify maximum uncertainty ranges associated with land cover conversion: "Max Forest" preferentially converts grass and shrubs upon agricultural expansion and expands forest to its potential limit upon agricultural contraction, "Min Forest" preferentially converts forest upon agricultural expansion and expands grass and shrubs to their potential limits upon agricultural contraction, and "Proportional" removes or adds PFT area proportionally to existing or potential PFT coverage for agricultural expansion and contraction, respectively. These three chronological LULCC configurations also account for existing pasture when calculating PFT distribution, which means that new cropland or pasture cannot replace PFTs that are already assigned to pasture. In addition to the simulations corresponding to the LULCC configurations, we ran three Proportional simulations with specific atmospheric forcings held constant. These constant forcing simulations are used to quantify the effects of CO 2 , climate, and N deposition on the terrestrial carbon budget. Two additional, intermediate LULCC configurations are presented in Text S1 in the supporting information.
We used the land use change module (Lawrence et al., 2012) with carbon-nitrogen biogeochemistry (Thornton et al., 2007) for all cases. The CRU-NCEP data (Wei et al., 2013; CRU-NCEP data available at https://www.earthsystemgrid.org/dataset/ucar.cgd.ccsm4.CRUNCEP.v4.html, 2017) were the meteorological drivers and years 1901-1920 were cycled prior to 1901. The "Constant climate and CO 2 " case continued to cycle these years after 1920. The simulations also used transient CO 2 and aerosol concentrations and nitrogen deposition, following CMIP5 protocols, except for constant forcing cases. The "Constant climate and CO 2 " and "Constant CO 2 " cases held CO 2 concentration at the 1850 level, and the "Constant N deposition" case held nitrogen deposition at the 1850 level.

Global Carbon Cycle
Forest area is a primary driver of global carbon uncertainty due to land cover conversion assumptions. Shifting from a year 2000 reference to chronological LULCC and accounting for existing pasture reduces forest and shrub areas and increases grass area because additional pasture requires land to be cleared ( Figure 1). This causes the chronological cases to deviate from the Default PFT distribution by 2005, with a 5.1 M km 2 difference in forest area between the Max and Min Forest cases, mostly compensated for by grass. Also, the Max Forest case has a similar global forest area trajectory to the Default case, with a final value of 42 M km 2 . The 1°, fully coupled simulations have nearly identical PFT distributions to the half-degree simulations, with the exception of the Max Forest case having~1 M km 2 less forest and more grass by 2005 due to resolution-dependent limits to adding forest area. This results in a 3.9 M km 2 difference in forest area between the Max and Min Forest cases for the fully coupled analyses.
The land-only, chronological cases enable us to directly quantify and compare the effects of land cover conversion uncertainty, CO 2 concentration, climate, and nitrogen deposition on net LULCC emissions. Land cover change leading to a final forest area difference of 5.1 M km 2 constitutes uncertainty in the global carbon cycle comparable to the combined effects of CO 2 concentration and climate on LULCC carbon emissions, and greater than those of nitrogen deposition. The chronological cases generally have higher net direct annual LULCC emissions than the Default, and the annual CO 2 and nitrogen deposition effects do not exceed the Max to Min Forest range until after 1950 ( Figure 2). Cumulatively, the 59 Pg C Max to Min Forest range of emissions from 1850 to 2004 is greater than the individual effects of increasing CO 2 (À55 PgC) and nitrogen deposition (À27 Pg C). Climate change has a negligible effect on the cumulative emissions (+2 Pg C). For comparison, the range between Min Forest and Default for years 1850-1990 is 61 Pg C, which is less than  Land cover conversion uncertainty also generates large uncertainty in land and atmosphere carbon stocks.
The 33 Pg C Max to Min Forest range of terrestrial ecosystem carbon lost to LULCC by 2005 is 80% of the corresponding net effects of increasing CO 2 plus climate change (41 Pg C) ( Figure 2). As expected, the intermediate cases give intermediate results with an ecosystem carbon range that is 46% of this net CO 2 plus climate effect (Text S1 and Figure S2). Furthermore, the regional distribution of this uncertainty depends on forest area difference and ecosystem carbon content ( Figures S3 and S4). Climate change increases terrestrial carbon loss by 11 Pg C, likely through reduction of productivity on abandoned land, while CO 2 and nitrogen deposition decrease loss by 52 and 27 Pg C, respectively, likely due to fertilization effects. Based on the fully coupled simulations, the Proportional case increases the 15 ppmv Default case bias in atmospheric CO 2 to 21 ppmv, and the Max to Min Forest range is 5 ppmv. The Max Forest case has similar global forest area to the Default case, but an additional 9 Pg C of carbon is lost in the Max Forest case due to shrub loss, increasing the atmospheric bias to 20 ppmv. These differences in ecosystem carbon and CO

Local Climate
Earth system model simulations demonstrate that relatively small uncertainties in land cover lead to significant differences in regional climate through biophysical effects. The Max minus Min Forest difference in forest cover ranges from À8 to 31 percent of the grid cell, with per cell surface temperature differences ranging from À0.87 to 1.62°C (Figures 3 and S4). These values are greater than the LULCC effects on land surface temperature for RCPs 2.6 and 8.5 estimated by Brovkin et al. (2013). Our per-cell uncertainty range for June-July-August is À0.75 to 1.37°C, which is comparable to historical LULCC effects on land surface temperature estimated by Pitman et al. (2009). While albedo generally decreases with increasing tree cover, thus increasing shortwave radiation absorbed by the surface, the local surface temperature both increases and decreases with increasing tree cover due to compensating effects of latent and sensible heating. Sensible heating is more sensitive than latent heating to changes in forest cover at the grid cell level ( Figure S5), which contributes to the Max Forest case having a global average temperature (1985)(1986)(1987)(1988)(1989)(1990)(1991)(1992)(1993)(1994)(1995)(1996)(1997)(1998)(1999)(2000)(2001)(2002)(2003)(2004)) that is 0.1°C greater than that of the Min Forest case.

Discussion
Land cover conversion assumptions and uncertainties significantly affect carbon and climate projections. These uncertainties drive global carbon cycle uncertainty that is comparable to the net effects of CO 2 and climate on the global carbon cycle from 1850 to 2004, and greater than the effects of nitrogen deposition. Climate change has little effect on net LULCC emissions, but it does increase the amount of terrestrial carbon lost to LULCC. Relatively small differences (<10% of grid cell) in forest cover can generate differences in local surface temperature of over 1°C, which is comparable to estimated effects of LULCC on temperature (

Geophysical Research Letters
10.1002/2017GL075124 response to local and distributed effects of land cover change, combined with differences in the general circulation associated with different land surface trajectories. Our results are conservative, in that we focus on uncertainty in land cover conversion assumptions. Additional sources of uncertainty include the land use forcing data, the initial and present-day land distributions, and model implementation of LULCC.
Our results suggest that the initial, transient, and final CLM land cover distributions may not reflect actual distributions. Basing LULCC on changes from the previous year and accounting for existing pasture moves the iESM farther from current land cover and carbon cycle estimates and requires forest maximization assumptions to bring it back to default CESM carbon cycle behavior. However, it is unlikely that a single conversion assumption adequately represents the entire globe (Prestele et al., 2017). Nonetheless, the extreme assumptions in this study are not far from other assumptions used in ESMs (Peng et al., 2017;Prestele et al., 2017) and reliably represent a maximum uncertainty envelope. Developing more realistic conversion assumptions will require further exploration of LULCC methods and initial and final states.
The final global forest area of the chronological cases is more consistent with estimates from other land cover studies, although still high, depending on forest definition. In the iESM forest area is based on PFTs, which correspond more directly with tree cover than with a broad range of forest canopy cover. In a PFT-focused effort, Meiyappan and Jain (2012) use the International Geosphere-Biosphere Programme (IGBP) definition of forest (>60% tree cover) and a spatial-coherence method for splitting mixed forest pixels. They also use three different land use data sources and estimate 2005 global forest area between 28.1 and 30 M km 2 , which is over 7 M km 2 less than the 37.0 M km 2 in our "Min Forest" case. Their 7.1-14.2 M km 2 estimate of savanna refers to tropical grassland, which would not make up the difference in forest area, as the IGBP definitions of savanna (10-30% tree cover) and woody savanna (30-60% tree cover) do not have enough trees (Friedl et al., 2002). Similarly, Friedl et al. (2010)  Furthermore, a recent study shows that uncertainty in present-day land cover contributes substantial uncertainty to albedo, evapotranspiration, and gross primary productivity in three land surface models (Hartley et al., 2017), and variability across LULCC trajectories directly contributes to high variability across terrestrial carbon estimates (Di Vittorio et al., 2014). This indicates that assuming a single LULCC trajectory for global modeling and analysis ignores considerable uncertainty that can have dramatic effects on carbon and climate projections.
The estimated effects of land cover uncertainty on temperature include local and regionally distributed effects of LULCC in addition to changes in general circulation due to different land surface states. While new methods aim to isolate the local effects of LULCC in model outputs to improve understanding and comparisons with observations (Lejeune et al., 2017;Winckler et al., 2017), model uncertainty quantification needs to include all relevant components in order to capture the entire error range associated with projections. In this context, increases in forest cover drive regionally dependent increases or decreases in temperature, even in places with no difference in forest cover (Figures 3 and S4). This is consistent with Swann et al. (2012), who report that large differences in forest area could shift general circulation patterns, affecting both precipitation and temperature beyond the extent of forest cover change. Furthermore, our results include changes in general circulation influenced by ocean responses to differences in land cover. As such, our uncertainty estimates are comprehensive with respect to fully coupled climate projections that provide inputs to impact analyses, which rely heavily on local and regional estimates (Field et al., 2014). Overall, land cover conversion uncertainty is a substantial and important component of local climate uncertainty that becomes even more critical when augmented by the related data and methodological uncertainties discussed above.

10.1002/2017GL075124
These results demonstrate the importance of accurate LULCC implementation and reliable LULCC uncertainty characterization when assessing climate mitigation and adaptation strategies and impacts through scenario-based modeling. LULCC uncertainty can completely change the location and type of prescribed land conversion, which affects local to global carbon and climate. For example, our final forest area uncertainty is 61% of the 8.3 M km 2 of forest lost from 2005 to 2100 in RCP8.5 and 78% of the 6.5 M km 2 of forest gained in RCP4.5 (Hurtt et al., 2011, Figure 9). Given the variability in land implementation among ESMs  and the resulting potential range of effects (e.g., Di Vittorio et al., 2014), LULCC uncertainty significantly contributes to model disagreement within an RCP. For land carbon projections in particular, LULCC uncertainty plays a central role in keeping the RCPs from diverging (C. Figures 2 and 3) when they should represent differences in land-based climate mitigation strategies. While other factors also contribute to model disagreement and scenario overlap, evaluation of climate mitigation and adaptation strategies is not possible if different scenarios are not distinguishable from each other.
We conclude that improving LULCC characterization and implementation can increase understanding and improve carbon and climate projections. Current efforts include adding forest area to CMIP6 land use scenarios (Lawrence et al., 2016). Such efforts facilitate a needed increase in consistency, accuracy, and uncertainty characterization of land cover data and implementation across models. Overall, it is critical to integrate land use and land cover analysis to provide better initial, transient, present-day, and future land use and land cover distributions, improve implementations of LULCC in earth system models, and enable models to be more faithful to historical and projected LULCC.