Estimating global land system impacts of timber plantations using

MAgPIE 4.3.5 Abhijeet Mishra1,2, Florian Humpenöder1, Jan Philipp Dietrich1, Benjamin Leon Bodirsky1, Brent Sohngen3, Christopher P. O. Reyer1, Hermann Lotze-Campen1,2, and Alexander Popp1 1Potsdam Institute for Climate Impact Research (PIK), Member of the Leibniz Association, P.O. Box 60 12 03, 14412 Potsdam, Germany 2Humboldt University of Berlin, Department of Agricultural Economics, Unter den Linden 6, 10099 Berlin, Germany 3The Ohio State University Department of Agricultural, Environmental and Development Economics, Columbus, Ohio, United States of America. Correspondence: Abhijeet Mishra (mishra@pik-potsdam.de)


Introduction
Forests cover 4060 million hectares (Mha) of the global land (31%) in 2020. Out of this 4060 Mha, 1110Mha are primary, 2657 Mha are secondary and 293 Mha are planted forests of which plantations cover 131 Mha and other planted forests cover 162 25 Mha, based on FAO (2020a) definitions. According to FAO (2020a), 1150Mha of forest are designated as production forests.
Plantations, as a very special forest land-use type according to FAO definitions, account for 11% of that area (and only 3% of global forest area) but likely supply more than 33% (654 Mm 3 ) of global industrial roundwood demand (1984 Mm 3 ) in 2020 based on historical trends (Jürgensen et al., 2014). This relatively large contribution compared to the area covered underlines plantations' special role in global land use dynamics. Roundwood consists of two sub-categories, industrial roundwood and 30 wood fuel.
Historical trends show a continuous increase in the share of roundwood production coming from plantations (Jürgensen et al., 2014). This trend indicates the efficacy and importance of timber plantations in meeting roundwood demand and the role of renewable forest management in natural forests (i.e. primary and secondary forests) especially in North America and Europe (Siry et al., 2018;Biber et al., 2020). The remaining share comes from other sources including harvesting of natural 35 forests or managed secondary or planted forests. Deforestation continues to occur at a large scale with wood harvesting being an important driving factor after cropland expansion (Curtis et al., 2018).
Deforestation contributes to about a third (3.8 Gt CO 2 yr -1 ) of Agriculture, Forestry and Land-Use (AFOLU) change emissions (10-12 Gt CO 2 yr -1 ) (Jia et al., 2019;Smith et al., 2014), and as it is an important driver of biodiversity loss, a better understanding of how we can produce timber using land resources efficiently is imperative. Plantation forests for timber pro-40 duction have potentially higher annual average increment per area than natural forests and managed natural forests IPCC (2006) because they are managed more intensively (fertilizer, thinning) and rely on high quality seeds and seedlings for regeneration. Because of their higher productivity as compared to natural forests (FAO, 2013;IPCC, 2006;Cubbage et al., 2007;Payn et al., 2015), timber plantations have the potential to fulfill a major portion of global roundwood demand while using a relatively small amount of land. Yet, assuming land distribution among different land-uses to be a zero-sum game, higher Model to Assess the Global Environment (IMAGE) (Stehfest et al., 2014). Forests are also included in varying degrees of representation in recursive dynamic optimization models like the Global Forest Sector Model (EFI-GTM) (Kallio et al., 2004) and the Global Biosphere Management Model (GLOBIOM) (Havlík et al., 2011) coupled with the Global Forest Model (G4M) (Kindermann et al., 2006). Timber supply and demand are also represented in the Global Timber Model (GTM) (Sohngen et al., 1999) which is an inter-temporal optimization model. A detailed review of recent developments and applications of partial 60 equilibrium models in the forest sector is provided by Latta et al. (2013). Yet, existing land-use models or forest economics models at higher spatial resolution either simulate detailed forest types and neglect competition for land or vice-versa. No existing land-use model to our knowledge combines both of these features at a global scale.
To correctly represent the competition for land and the role of different forest types in meeting growing roundwood demand, ideally, a land-use model should a) represent land resource competition while accounting for food, feed and timber demand, and, b) represent different growth rates between natural and planted forests (with the accounting of optimal rotations in timber plantations).
Yet, out of the recursive dynamic models mentioned above, partial equilibrium models like EFI-GTM and GTM do not use spatially explicit differences in forest growth rates but use aggregated forest inventory data as model inputs. Both of these models rather focus on a detailed representation of the forest and timber industry with great detail but do not model competition 70 for land between forests and agriculture at a fine spatial scale. IMAGE and GLOBIOM, both use spatially explicit differences in forest growth rates and tree species while representing competition for land between forests and agriculture but do not explicitly differentiate between natural forests and timber plantations. In IMAGE, land-use evolution for timber plantations is a model parameter and is not endogenously determined. GLOBIOM when coupled with G4M also circumvents the myopic nature of recursive dynamic models as G4M results are linked to GLOBIOM for making appropriate land-use change decisions regarding 75 wood production and forest land use. GCAM models competition between land-uses via land competition nests (Snyder et al., 2020) where land-use categories belonging to the same category in the nest (e.g. crops) are assumed to compete more directly with each other than with land-uses in other categories (e.g. forest) (van de Ven et al., 2021). Additionally, the choice of rotation lengths in plantations is an important component for managed forests that follow even-aged management systems. To the best of our knowledge, the determination of optimal rotation lengths for timber plantations has not been done in any of the 80 uncoupled global recursive dynamic models so far (Kallio et al., 2004;Calvin et al., 2019;Havlík et al., 2011).
In light of these limitations of representing timber plantations in the land-use modeling frameworks described above, tools that quantify and analyze land competition while explicitly accounting for the specifics of forest plantations within a uniform modeling framework are required. The Model of Agricultural Production and its Impact on the Environment (MAgPIE) uses both biophysical and economic drivers to simulate land-use change and its impact on the environment while accounting for 85 feed, food and livestock demand (Popp et al., 2010;Lotze-Campen et al., 2008;Dietrich et al., 2019;Bodirsky et al., 2020).
Driven by the motivation to represent coherent forest land-use dynamics within a single modeling framework, we present here an extension of the MAgPIE 4 modeling framework by timber production and associated land-use dynamics. The extension not only addresses the forestry sector modeling gaps outlined above via new MAgPIE modules that differentiate timber plantations and natural vegetation land-use, but it also includes forest age-class dynamics in a large-scale global land-use model like 90 MAgPIE for the first time. The MAgPIE modeling framework (Dietrich et al., 2019;Lotze-Campen et al., 2008) is a global multi-regional land system 95 model. The objective function of MAgPIE is to minimize the global costs to produce food, feed, bioenergy and timber throughout the 21st century in a recursive dynamic model with limited foresight. Provided the long time horizons in the establishment of new trees today, followed by harvesting such trees sometime in the future, calls for using a recursive-dynamic model for understanding how today's decisions impact tomorrow's behaviour. MAgPIE is driven by demand for agricultural commodities and roundwood, which is calculated based on population and income projections for the 21st century from the Shared 100 Socioeconomic Pathways (SSPs).
MAgPIE derives specific land-use patterns, yields and total costs of agricultural and roundwood production for each simulation cluster as described in Dietrich et al. (2019). MAgPIE's optimization is bound by spatially explicit biophysical constraints derived from the global gridded crop and hydrology model LPJmL (Bondeau et al., 2007). For this assessment, the spatially explicit (0.5°resolution) LPJmL outputs are aggregated for MAgPIE into 200 simulation units/clusters using a clustering al-105 gorithm (Dietrich et al., 2019(Dietrich et al., , 2013 as shown in fig. 1. MAgPIE is a non-linear mathematical programming model written in General Algebraic Modeling System (GAMS) (GAMS, 2021) and solved with CONOPT4 solver (Drud, 2015). CAZ (28) CHA (24) EUR (10) IND (7) JPN (3) LAM (53) MEA (17) NEU (8) OAS (22) REF (7) SSA (11) USA (10) The existing MAgPIE 4 framework (Dietrich et al., 2019) has been extended by the inclusion of timber production via forest land and timber demand, which we refer to as MAgPIE 4.3.5 in the text. Growth function for forests (Humpenöder et al., 2014) 110 are parameterized by using plantation and natural vegetation specific parameters from Braakhekke et al. (2019). Finally, the trade representation was also extended to include industrial roundwood and wood fuel trade. The extension of the MAgPIE framework from version 4 to version 4.3.5 is shown in fig. 2. Figure 2. Extended MAgPIE 4.3.5 framework. Blue color represents update to existing modules, green color represents new inclusions to Dietrich et al. (2019). See the model documentation  for a more detailed presentation of module interactions and their implementations.

Scenarios
We analyse two scenarios here namely default and forestry (Table 1). Both, default and forestry scenarios take assumptions 115 from the SSP2 storyline also known as business as usual or middle of the road scenario (Riahi et al., 2017). In the default case, we replicate assumptions from a standard MAgPIE configuration based on Dietrich et al. (2020b), where a) Timber demand is not modeled, b) No forest is harvested for timber production, c) No competition for land between agriculture and forestry, and d) Secondary forests and plantations are assumed to belong to the highest age-class during model initialization. The setup of the default scenario without wood demand, no harvest from plantations (and other forests) and no new plantation establishment The forestry scenario on the other hand accounts for a) GDP and population-driven industrial roundwood and wood fuel demand, b) Plantations and natural forests as a source of timber production, c) Endogenous competition for scarce land resources between agriculture and forestry, and d) Heterogeneous age-class structure of secondary forests and plantations during initialization. Plantation forests are initialized such that there is a higher weight provided to younger age-classes reflecting the 125 notion that replanting has continued to exceed harvests in plantations in the last decades. Secondary forests are initialized based on the land distribution among age-classes described in Poulter et al. (2019).
In terms of protected areas, both scenarios account for National Policies Implemented (

Rotation lengths
According to the maximum sustained yield rotation-period model described in Amacher et al. (2009), a forest owner's approach is to maximize the volume of timber that can be obtained from a given stand on a sustained yield basis. Such optimal time to harvest trees occurs when the timber volume increment is maximized such that the Mean Annual Increment (MAI) is equivalent 135 to the Current Annual Increment (CAI). Maximizing increment for choosing rotation lengths however results in longer rotation lengths compared to economically optimal Faustmann rotations. Additionally, in the MAgPIE framework, high rotations (ca. >100 years) affect how plantation area is initialized and result in much lower availability of plantations for timber production (see section Forest initialization). Therefore, for our implementation, we use maximization of CAI to ascertain the prescribed rotation lengths for timber plantations in MAgPIE as from a empirical point of view, this criteria is closer to economically In equation 1, f' ac is the first derivative of the the age-class (ac) specific carbon density with respect to age-classes (f ac ). Instead of using forest volume described in Amacher et al. (2009), we use carbon density as a proxy for the same. Long term average potential carbon density information for each MAgPIE cluster is obtained from LPJmL (Bondeau et al., 2007). This carbon 145 density information is fed into a Chapman-Richard's growth function to derive age-class specific carbon densities i.e. f(ac) based on Humpenöder et al. (2014) (fig. 3a). The first derivative of these carbon densities provides the marginal values with respect to age-classes ( fig. 3b). Equating first derivative of CAI to zero provides the cluster specific optimal rotation lengths ( fig. 3d) i.e., the optimal age-class at which harvest of timber plantation is allowed in each cluster. Rotation length decisions once made cannot be altered at a later time step, which is in line the recursive-dynamic optimization in MAgPIE. Natural

Forest initialization
In MAgPIE, forestry rotation lengths determine what the initial distribution of planted forest area should look like in 1995. The country-level planted forest area from FAO (2015) is downscaled to a 0.5°grid using area-weighted mean of wood removals 155 (Hurtt et al., 2018) and then upscaled to MAgPIE cluster level (Dietrich et al., 2019) for initialization of 1995 values. Distribution of this area among different age-classes i.e., the age-class structure in plantations during initialization is driven by rotation lengths. Aggregated cluster level planted forest area is distributed first between plantations and other plantation areas based on the historical share of such distinction based on FAO (2020b). Cluster level plantation area is then divided among age-classes such that there is a higher weight provided to younger age-classes reflecting the notion that plantation area establishment has 160 increased in the last decades. Figure  is assumed to exist in the highest age-class in 1995. The area allocated to secondary forests is assumed to follow the distribu-tion of forests in different age-classes based on Poulter et al. (2019). After the initialization of forest areas, the development of forest cover is modeled endogenously in the model and driven by roundwood demand, timber harvest costs, expected yields, carbon prices, demand for agricultural land, land-use change costs and land-use change constraints.

Timber demand
Demand for end-use wood products in MAgPIE is driven by changes in per capita income and population for the Shared Socioeconomic Pathway 2 (SSP2) storyline. Here we take assumptions from the SSP2 storyline to derive the timber demand.
We use a simple demand function specification from Lauri et al. (2019) Here, t is the simulation time step i.e. time and wp are different demand categories for wood products. Q is the annual in MAgPIE. By-products of end-use production activities and recycling of roundwood is also not accounted for in MAgPIE.
Wood fuel is assumed to come from two different sources: direct harvest and logging residues from harvesting for industrial roundwood.
Global industrial roundwood and wood fuel demand modeled in MAgPIE is shown in fig. 6 along with validation from historical data reported by FAO (regional numbers in fig. A4). Wood fuel enters demand calculations with a negative income 190 elasticity based on Morland et al. (2018) to be consistent with the decreasing residential sector biomass use for energy in an SSP2 world (Lauri et al., 2019;IIASA, 2018). We use the logging residue data from Oswalt et al. (2019) indicating that 30% of industrial roundwood harvest is residue. Assuming 50% of this is recovered from forests (Pokharel et al. (2017) report a range of 30-70% from available literature), we use a maximum of 15% of biomass removed during industrial roundwood production as wood residues which can contribute towards fulfilling wood fuel demand. 195 We assume that the residues are collected from the overall production system i.e., we do not explicitly differentiate if the residue comes from plantations or natural forests harvest. We do not model the decay in productivity after residue removal as at least for some plantations, fertilization would be applied to maintain productivity. The residue generation constraint in MAgPIE is an upper bound for the model which provides flexibility in deciding (based on the cost of production) if the residue should be removed or not from the part of production which comes from plantations. Historical data for validation is based on FAO (2017). The MAgPIE default scenario does not include timber demand by assumption.

Forest biomass
Biomass which can be potentially removed from natural forests is calculated based on the average long-term vegetation carbon densities in natural vegetation from LPJmL. Growth of natural vegetation in MAgPIE follows an s-shaped growth curve as described in Humpenöder et al. (2014), but with updated growth curve parameters based on Braakhekke et al. (2019). Timber plantations on the other hand are considered more productive (for a younger stand age per unit area) compared to primary 205 forests and secondary forests (FAO, 2006). To reflect this, we use a different parametrization of the timber plantation growth function as compared to natural forests based on Braakhekke et al. (2019). Harvestable biomass from forests are calculated as shown in equation 3 based on Ravindranath and Ostwald (2007) and Standard (2013).
Here, t is the simulation step i.e. time, j is the MAgPIE simulation cluster, ft is the forest type i.e., plantations or natural 210 vegetation. ac is the forest age-class, clcl is the Köppen-Geiger climate class. y is the age-class (ac) and forest type specific biomass yield in tDM/ha, C is the forest type specific carbon density in tC/ha, r is shoot-to-root ratio, cf is the carbon fraction in dry matter (IPCC, 2019), kg is the Köppen-Geiger climate classification (Rubel and Kottek, 2010) and b is the biomass expansion factor (FAO, 2013). Forest classification in MAgPIE is represented in fig. 7 and the detailed description of forest land dynamics are described in Dietrich et al. (2020a). Harvestable biomass yield (y) is different between natural forests 215 (primary and secondary forests) and plantations by virtue of differences in parametrization of underlying growth function(s).
Primary forests are assumed to exist in the highest age-class, and are therefore attributed with old-growth forest yields. Both, secondary forests and plantations yields are age-class specific but differ in growth-dynamics. The carbon density in plantations and natural forests is calibrated using a scaling factor to match the historically reported forest growing stock at regional level (FAO, 2020a). This scaling factor is calculated as the ratio between observed growing The amount of newly established timber plantations depends on current roundwood demand, the assumed future share of production coming from plantations and expected future yields. Expected future yields in plantations are calculated based on the rotation lengths. As shown in equation 4, we define a regional constraint while establishing new timber plantations.
i j,ac Here, plant is the plantation land, j is the MAgPIE simulation cluster, ac' is the age-classes to be established (usually the 230 youngest age-class that is ac0), Q i,rw is the regional annual demand for roundwood (rw) i.e., industrial roundwood and wood fuel in region i as shown in fig. 6. σ i,rw is the regional self-sufficiency ratio of roundwood (industrial roundwood and wood fuel) production (Table A3), η i is the share of production which can come from plantations based on extrapolations from Pöyry (1999). For the extrapolation of these shares, we assume (starting from last historically available data in 2000), 1% increase per annum till 2020, 0.4% increase per annum between 2020-2050 and 0.2% increase from 2050-2100 (Table A4).

235
ES i is a calibration factor to nudge the model towards historical plantation area patterns (Table A5) via establishment of new plantations.
For example, assuming industrial roundwood demand of 100 Mm3 in 2020 in region i with a self-sufficiency ratio of 0.8 and η i of 0.5, the model will need to establish plantations such that 100 * 0.8 * 0.5 = 40 Mm3 of timber can be produced from this region in the future. The model then tries to establish new plantations in the simulation step depending on expected yields.

240
Assuming this region has 2 clusters, both with an expected yield of 5 m 3 ha−1, there will be 4 Mha ((1/2)*40/5) of plantations established in each cluster i.e, 8 Mha of total new plantations in this region.

Timber harvesting
Timber plantations are harvested once they reach maturity at the specified optimal rotation lengths. After every time step, forest age-classes are shifted forward. Plantations are protected from harvest during the whole duration of time below their specified 245 rotation length. There is no such restriction on the harvest of natural vegetation based on age and maturity as natural forests are not bounded by rotational constraints. Forests in MAgPIE are harvested based on harvesting costs and associated trade-offs.
MAgPIE's objective function is to minimize global production costs and using a lower harvesting cost (per ha) for plantations than in natural forests implicitly provides a signal to the model to harvest forests with higher growing stock first.
Roundwood (industrial roundwood and wood fuel) can be produced from both natural forests (primary and secondary forests) 250 and from managed plantations (forestry), which we distinguish according to figure 7. Additionally, wood fuel can also be harvested from other land, which is defined as non-managed land that has an insufficient carbon stock (<20 tC ha -1 ) to be classified as forest. Timber production from forests is calculated based on the area harvested and the harvestable yields (3).

Land-use change emissions
Net CO 2 flux from land-use, land-use Change and Forestry (LULUCF) includes CO 2 fluxes from forest harvest (for roundwood 255 production), deforestation (clearing forest for alternative land-use), afforestation, shifting cultivation (deforestation followed by abandoning) and regrowth of forests following wood harvest or abandonment. Some of these activities lead to emissions of CO 2 to the atmosphere (burning wood fuel after harvest, conversion of forests to agricultural land), while others lead to CO 2 sinks (afforestation, regrowth, long term carbon stored in harvested wood products).
Land, in particular biomass production from vegetation, affects both the source and sinks of CO 2 . While reporting on 260 LULUCF emissions, usually the long term carbon stored in wood products is either not reported or not accounted for in models which simulate forest land-use Havlík et al., 2011;Braakhekke et al., 2019;Doelman et al., 2018Doelman et al., , 2020Humpenöder et al., 2018). As management of forests and different uses of harvested wood play a crucial role in the regulation of the concentration of atmospheric CO 2 , it is important to account for this pool while reporting LULUCF emissions (IPCC, 2019;Johnston and Radeloff, 2019;Böttcher and Reise, 2020;Zhang et al., 2020).

265
In MAgPIE we account for gross land-use change emission (i.e. land-use change emissions not including regrowth), emissions due to shifting agriculture (as part of gross land-use change emissions) based on historically observed deforestation driver rates from (Curtis et al., 2018), regrowth in forests and other land as well as long term carbon storage in wood products while also calculating the slow release of CO 2 back into the atmosphere from these wood products due to decay ( fig. 9). Carbon stored in harvested wood products (HWPs) can affect national greenhouse gas (GHG) inventories, in which the production according to the guidance provided by the Intergovernmental Panel on Climate Change (IPCC) as defined in equation 5 (IPCC, 2019). Figure 9. Concept for accounting for carbon emission and storage dynamics from forests and harvested roundwood. Wood fuel is assumed to be emitted within the optimization step in which it is harvested. Industrial roundwood enters a long term storage pool, from which slow turnover happens and is tracked via IPCC (2019) methodology described in equation 5.
Here, C is the carbon stock in industrial roundwood at the beginning of year t in Mt C. k is the decay constant of first  Table 2 (rounded to nearest zero) and fig. 10. land increases by 649 Mha at an expanse of forests as well as other land indicating that more cropland intensification takes place when timber production is included. Timber plantation area increases by 171 Mha in forestry scenario to satisfy a considerable portion of industrial roundwood and wood fuel demand from plantations, given the increasing timber demand due to income and population growth. Primary and secondary forest area declines by 422 Mha and 377 Mha respectively between 1995 and 2100 due to the expansion of cropland and timber plantations in the forestry scenario. Other land area decreases by 255 Mha 300 between 1995-2100 in the forestry scenario (as compared to 468 Mha in the default scenario).
To satisfy food and feed demand and to accommodate the land-use competition between cropland and forestry, MAgPIE estimates an agricultural yield-shift of 113% and 116% in the default and forestry scenarios respectively by 2100 relative to 1995 through investments in yield-increasing technological change. Such yield-increasing technological change is realized via agricultural land-use intensity in MAgPIE and is measured using a τ -factor developed by Dietrich et al. (2012). The global and 305 regional land-use intensity indicator τ for the forestry and default scenarios is shown in fig. A3.  As plantations compete with cropland for limited land resources, it is important to see how the inclusion of roundwood production interacts with cropland usage globally. Figure 13 shows the difference in cellular cropland area between forestry 315 and default scenarios on a 0.5°grid and Table 3 shows the regional differences for the same.

Secondary forest age-class structure
Secondary forests are initialized in MAgPIE as described in section 2.4. Once harvested (for timber production) or cleared (for cropland or plantations), secondary forests move to the youngest age-class (ac0) and are subject to natural regrowth. Primary 325 forests once harvested are re-classified as secondary forest of the youngest age-class and follow regrowth. Table 4 shows the difference in secondary forest area between 1995-2100. Development of age-class structure in secondary forests for default and forestry scenarios is also shown in fig. 15. Selection of appropriate initial age-class distribution is especially important as they have a direct relationship with AFOLU emissions (further discussed in section 3.5).   Figure 16 shows the annual amount of forest area harvested for meeting the roundwood demand globally (forestry scenario; no harvested area in the default scenario). On average, between 1995-2100, we observe 2 Mha yr -1 of plantations and 7 Mha yr -1 of natural forest harvest in the forestry scenario. In this scenario, natural forests are harvested more than timber plantations in all periods. In line with the assumptions for timber plantations establishment (increasing share of timber production from plantations in the future), the harvested area from timber plantations increases in the future. Regional details of the annual Johnston and Radeloff (2019). Regional distribution is available in fig. A6 In the default scenario, land-use change emissions decrease from 3.0 Gt CO 2 yr -1 in 2000 to 1.8 Gt CO 2 yr -1 in 2100. In the forestry scenario we observe that emissions increase from 1.2 Gt CO 2 yr -1 in 2000 to a peak of 3.1 Gt CO 2 yr -1 midcentury and then fall gradually back to -1.3 Gt CO 2 yr -1 by the end of this century. The gross land-use change emissions are comparable between the default and the forestry scenario with results from forestry scenario slightly closer to historically 345 reported numbers from Gasser et al. (2020) than in default scenario. The net land-use change emissions and removals from regrowth differ substantially between both scenarios, where, in the forestry scenario, removals from regrowth compare much better to values from the literature (Gasser et al., 2020). Overall, we present a historically consistent evolution of regrowth emissions in the forestry scenario due to accounting for timber production and age-class structure in timber plantations and natural forests.

350
Compared to the default scenario, we observe lower CO 2 emissions in the forestry scenario during the initial periods due to higher carbon uptake driven by assumptions of a heterogeneous initial age-class structure in secondary forests (carbon uptake can be interpreted as negative emissions where a mathematically lower value is higher carbon uptake). In the default scenario, carbon uptake is much lower because of two reasons: 1) During initialization, all secondary forest is assumed to exist in the highest age-class, which limits the amount of regrowth, and 2) No secondary forest is harvested for timber production in the default scenario. Without such disturbances, the age-class structure in secondary forests does not shift much towards the younger age-classes (also seen in fig. 15) where usually regrowth is faster as compared to old-age forests.
In this paper, we expanded the MAgPIE modeling framework by a detailed representation of land-use dynamics in natural forests and timber plantations while accounting for roundwood production and competition for land with agriculture. Rep-360 resenting forestry and timber production in a recursive-dynamic land-use model is a challenging issue due to complexities associated with long term planning horizons needed for roundwood production and forest management. This explains why major land-use models focus on better representation of the agricultural sector or the forestry sector, but not on the competition between both within the same model (Calvin et al., 2019;Wise et al., 2014;Stehfest et al., 2014;Kallio et al., 2004;Havlík et al., 2011;Kindermann et al., 2006;Sohngen et al., 1999). As timber, food and feed production happen simultaneously in 365 the real world, the inclusion of the forestry sector, next to the agricultural sector, substantially improves the representation of land-dynamics and GHG emissions in MAgPIE.
While including the forestry sector in MAgPIE, we present a historically consistent development of timber plantation area over time when compared to observed data (FAO, 2020b). We also present a historically consistent development of growing stocks in plantations and natural forests over time (FAO, 2020b). Our results show that the inclusion of timber production and 370 plantation establishment in the MAgPIE modeling framework competes with cropland for limited land resources. While the total global cropland is similar between the default and the forestry scenario at the global level, the spatial cropland patterns differ substantially between the two scenarios, which indicates that timber plantations compete with cropland for the same scarce land resources. The net effect is a stronger decline of natural forest in the forestry scenario as compared to the default scenario. New timber plantations might be partly established on cleared natural forests. However, considering the substantial 375 changes in spatial cropland patterns it seems likely that plantations are also established on cropland and pasture land, which causes deforestation for cropland expansion elsewhere.
Our land-related CO 2 emissions and removals match better with observed data (Houghton et al., 2012;Gasser et al., 2020;FAO, 2017;Gütschow et al., 2016;JRC and PBL, 2010) in the forestry scenario as compared to the default scenario, in particular the gross land-use change emissions, reflecting the higher deforestation for the expansion of managed land and 380 timber production, and the carbon uptake, reflecting the regrowth in natural forests and timber plantations.
Our modeling study also indicates that timber plantations are an important source of roundwood production. If timber plantations would not increase, in contrast to our forestry scenario, the projected increase in roundwood demand would need to be fulfilled by wood harvest from natural forests. Of particular importance is that plantations can produce more timber on less area, making them a candidate for reducing roundwood production pressure from natural forests. This opens up a similar 385 question with respect to the land-sharing versus land-sparing debate. Establishing high yielding plantations for roundwood production might provide the benefit of producing a large quantity of timber using a small land area but such plantations do not synergize well with biodiversity. Species richness in plantation forests is usually significantly lower than in natural forests (Phillips et al., 2017). When plantations are established after clearing natural forests, there will be a decline (or even loss) of biodiversity. On the contrary, it is also important to keep in mind that even when timber plantations embody lower species 390 richness than natural forest in comparable geographic locations, plantations, if established on degraded land, will almost always support higher species richness (Brockerhoff et al., 2008). Plantations may generally be lower in biodiversity, but eventually spare natural forests for CO 2 sequestration, biodiversity and soil preservation purposes (Moomaw et al., 2020;Waring et al., 2020;Buotte et al., 2020).
We are aware that our research may have certain limitations as extending a recursive dynamic land-use model to include a 395 dynamic forestry sector is not straightforward and includes some strong generalizations. First, we do not account for future climate change impacts in this study. In principle, the modelling framework is capable of accounting for climate change impacts. However, in this study, we deliberately chose to focus on the overall forestry implementation and the implications on land-use dynamics and GHG emissions. and are only affected by the shape of assumed growth curves (Braakhekke et al., 2019) and carbon densities (Humpenöder et al., 2014) but are unchanged by fluctuations in timber prices and interest rates, which is a simplification of reality.
Third, in forests managed for timber production, thinning is practiced by removing the smaller and poorer quality trees. This operation generates income with the sale of harvested timber and also makes sure that growth is favorable for the remaining 410 trees. This operation also results in a higher volume and quality of harvested timber, which can generate a higher income in the future as the price for such timber is higher in the market. We do not simulate this activity in our updated modeling framework and thereby underestimate the amount of roundwood production capabilities of timber plantations to some extent.
Fourth, we do not account for spatial differences in tree species as MAgPIE in its current format does have no mechanism in place to handle such information explicitly. Even though the growth curves used in MAgPIE are parametrized differently for 415 natural forests and plantations, they are not perfect proxies for differences in growth and biomass volume accumulation among different species. As a corollary, we also do not prescribe a minimum diameter constraint for harvesting as MAgPIE cannot ascertain the thickness of tree-trunks at every stage of tree growth.
Fifth, the results presented here are driven by socio-economic assumptions from the SSP2 scenario which is considered to be a "middle of the road" scenario. Inherently, our results are as uncertain as the future socio-economic drivers i.e., the 420 wide range of possible future socio-economic development in different SSPs bring a wide range of uncertainty about the future development of the forest sector (Lauri et al., 2019) and associated land-use change. On a spatial scale, there is a considerable uncertainty in spatially explicit data on plantation forest with respect to the differentiation between productive and non-productive plantations which in turn also has a bearing on the results. Additionally, management of plantations in reality also depends on other factors such as availability of workforce, investment, research and development, which are not 425 considered for plantations in MAgPIE.

Conclusions
Since the inception of MAgPIE, the modeling framework has evolved with time to include a broad range of land-use processes.
In this paper, we describe an extension of the existing MAgPIE framework by a detailed representation of timber demand and production, forest land and timber plantations. MAgPIE 4.3.5 allows land-use processes for timber production to be simulated 430 with feed, food and livestock demand simultaneously, advancing the land-use representation from previous MAgPIE versions.
Given the growing importance of timber plantations in meeting growing global timber demand, it is also imperative that timber plantation systems are modeled explicitly within forest systems in land-use modeling. Timber production has not been a part of the MAgPIE modeling framework since its inception, which means that a major driver for deforestation and land-use change emissions has been missing. With this paper, we bridge this gap and expand the coverage in the representation of the most 435 relevant land-use change drivers in MAgPIE.
Inclusion of the forestry sector in MAgPIE offers improved understanding of land resources, which plays a vital role in climate change mitigation (Doelman et al., 2018), biodiversity conservation (Gibson et al., 2011;Phillips et al., 2017) and maintaining crucial ecosystem services (Foley et al., 2005). This expanded version of MAgPIE not only provides an improved tool for comprehensive assessments of the Sustainable Development Goals (SDG) but may also contribute to other important 440 scientific processes, such as providing inputs for Earth System Models (ESMs) (Hurtt et al., 2018;Luyssaert et al., 2014;Reid et al., 2010;Bonan and Doney, 2018), Biodiversity models (Thuiller et al., 2013;Urban et al., 2016), or international networks like the Agricultural Model Inter-comparison and Improvement Project (AgMIP) (Ruane and Rosenzweig, 2018) or the Inter-Sectoral Impact Model Inter-comparison Project (ISIMIP, www.isimip.org).
Code and data availability.     Year Figure A5. Modeled contribution of timber harvest from natural forests and plantations to industrial roundwood and wood fuel production in forestry scenario (1995-2100).