JULES simulates the processes of photosynthesis, autotrophic and heterotrophic respiration, and calculates the turbulent exchange of CO2, heat, water, and momentum between the land surface and the atmosphere (Cox et al., 1998; Best et al., 2011; Clark et al., 2011). Vegetation dynamics are simulated by TRIFFID. Recently, new PFTs were added to JULES (Harper et al., 2016) (Table 1), which required updates to the TRIFFID competition scheme, described below. In this paper, we compare two versions of JULES: JULES-C1 and JULES-C2 based on JULES version 4.6. The former is a configuration of JULES with five PFTs as described in Harper et al. (2016) (called JULES5 in that paper) and as used in the TRENDY multi-DGVM synthesis project (Sitch et al., 2015). The latter (JULES-C2) is the new version, with nine PFTs and the vegetation dynamics and updates described in Sect. 2.2–2.3.

Within TRIFFID, carbon acquired through NPP is allocated to either spreading (in other words increasing fractional coverage of a PFT in a grid cell) or growth (increasing height). The time evolution of the fractional coverage of the ith PFT (νi) is calculated as follows: CVidνidt=λiΠiν*1-∑jcijνj-γνiν*CVi, where CV is the vegetation carbon (kg C m-2), Π is the accumulated NPP between calls to TRIFFID (kg C m-2 (360 d)-1), ν* is the maximum of the actual fraction and a “seeding fraction” (0.01), and γν is a PFT dependent parameter representing large-scale disturbance (360 d)-1. In the present study, TRIFFID ran on a daily time step. The fraction of NPP allocated to spreading, λ, is a function of the balanced leaf area index (LAI), Lbal, which is the seasonal maximum of LAI based on allometric relationships (Cox, 2001): λ=1forLbal>LmaxLbal-LminLmax-LminforLmin<Lbal≤Lmax0forLbal≤Lmin, and the fraction allocated to growth is (1-λ). The PFT dependent parameters Lmax and Lmin determine the balanced LAI at which plants allocate 100 % of NPP toward expanding PFT coverage (spreading: Lbal≥Lmax) or 100 % toward vertical plant growth (Lbal≤Lmin).

Competition for space in the grid cell between PFT i and the other PFTs is represented by the matrix cij, which represents a dominance hierarchy where height is the most important factor as it determines access to light. Effectively, the (1-Σcijνj) term in Eq. (1) is the space available to PFT i. In the original version of TRIFFID, trees were assumed to dominate shrubs, and shrubs were assumed to dominate grasses (Cox, 2001). Within tree (broadleaf and needle-leaf) and grass (C3 and C4) PFTs, there was co-competition and cij was calculated as a function of vegetation height for the two competing PFTs: cij=11+exp⁡20×hi-hjhi+hj.

We made two changes to the original TRIFFID: first we removed the hardwired dominance hierarchy (trees > shrubs > grasses) to allow for a generic number of PFTs. The dominancy hierarchy is now completely height-based, so that the tallest PFTs get the first opportunity to take up space in a grid cell. Second we removed co-competition, so that cij is either 1 or 0. This simplifies the equilibrium solution for vegetation coverage (Sect. 3.2). When PFT i is dominant, cij=0 and PFT i is not affected by PFT j; when type j is dominant, cij=1 and PFT i does not have access to the space occupied by PFT j (νj).

Although the version of JULES described in this paper is similar to that described previously by Harper et al. (2016), there are four differences, which are summarized in the following. Impacts of the new equations for leaf, root, and stem nitrogen are discussed in detail in the Supplement.

The distribution of PFTs was evaluated by first dividing the land surface into 8 biomes, based on the 14 World Wildlife Fund terrestrial ecoregions (Olson et al., 2001). The map of biomes (Fig. S9 in the Supplement) acted as a mask for the results to calculate biome-scale averages, and each grid cell was assumed to be 100 % composed of the biomes shown in Fig. S9. For each biome, we calculated the average fractional coverage of each PFT, average grid-box fluxes (GPP and NPP), and average grid-box carbon stocks (soils and vegetation), as well as average fractional coverage of agricultural land. These biome-scaled distributions and averages were then compared to observations. For observed PFT distribution, we used the global vegetation distribution from the European Space Agency's Land Cover Climate Change Initiative (ESA LCCCI) global vegetation distribution (Poulter et al., 2015; Hartley et al., 2017). To quantify the evaluation of PFT distribution, we calculated an error metric ε for each PFT (εi, Eq. 15) and for each biome (εB, Eq. 16). The former enables a ranking of PFTs in terms of their improved distributions and is weighted by biome areas. The latter enables a comparison between models of the vegetation distribution on a biome scale and implicitly includes an area weighting since all fractions in a biome sum to one. εi,PFT=∑B=18AB×νB,imod-νB,iobs2∑B=18ABεB,biome=∑i=1npftνB,imod-νB,iobs2npft In these equations, AB is the area of biome B, npft is the number of PFTs (in this case eight because C3 and C4 grasses are combined), and νB,i is the fractional coverage of PFT i in biome B.

To evaluate the carbon fluxes, we used gross primary productivity (GPP) from the Model Tree Ensemble (MTE; Jung et al., 2011), and MODIS NPP from the MOD17 algorithm (Zhao et al., 2005; Zhao and Running, 2010). Soil and vegetation carbon were from Carvalhais et al. (2014). In addition, we compared biomass stocks to the dataset from Ruesch and Gibbs (2008). In all evaluations, we used model years corresponding to the available observation years: 1982–2011 for GPP, 2000–2013 for NPP, and a 30-year period for soil and vegetation carbon (1980–2009). All datasets were regridded to the model resolution of 1.25∘ latitude × 1.875∘ longitude.

There are a total of six simulations: one using JULES-C1 and five using JULES-C2. Both versions of the model were run with transient climate, CO2, and land use over the historical period. The climate was from version 6 of CRUNCEP, which is a merged dataset of CRU and NCEP reanalysis from 1901 to 2015. Climate variables used were downwelling longwave and shortwave radiation, total precipitation, air temperature, specific humidity, zonal and meridional wind speeds, surface pressure, and a constant diffuse fraction of shortwave radiation of 0.4. The fraction of agriculture in each grid cell was included as the fraction of crop and pasture from the harmonized dataset based on HYDE3.2 (Hurtt et al., 2011). CO2 concentration was from Dlugokencky and Tans (2013). We ran three additional experiments with JULES-C2 to assess the contributions of climate change, land use change (LUC), and CO2 fertilization to the changes in carbon cycle components over the historical period (Table 5). Experiment SCLIM was forced with the transient climate from CRUNCEP-v6 to assess the contribution of climate change alone, while atmospheric CO2 and land use were held to preindustrial (1860) values. In experiment SLUC,CLIM, climate and land use changed, while CO2 was held constant, and in experiment SCO2,CLIM, climate and atmospheric CO2 changed, while land use was held constant. For the discussion of attributing changes to these drivers we refer to the main experiment as SALL, which has transient climate, LUC, and CO2. The impact of LUC on the present-day carbon cycle is given by SALL - SCO2,CLIM, and the impact of CO2 fertilization is given by SALL - SLUC,CLIM. A fifth simulation with JULES-C2 was done to test the model with raw climate model output without bias correction to assess sensitivity of PFT distribution to the climate. This simulation was forced with the HadGEM2-ES RCP2.6 climate and CO2. The available climate variables from HadGEM2-ES were downwelling longwave and shortwave radiation, stratiform rain, convective rain, stratiform snow, convective snow, air temperature, specific humidity, wind speed, surface air pressure, and the incoming diffuse shortwave radiation.

The simulated distribution of PFTs in TRIFFID is sensitive to the large-scale disturbance parameter γν from Eq. (1). The parameter represents several missing processes in JULES related to disturbance-induced mortality (such as fires, pests, and wind events), and provides an estimate of turnover rates for the PFTs. We developed a method for quickly estimating a global value of γν for each PFT. Updated values of γν were necessary due to new physiology, which resulted in a new NPP per PFT (Π in Eq. 1), and an expanded set of PFTs. The method is based on a formula for the equilibrium distribution of PFTs, made possible by the removal of the hardwired dominance hierarchy in TRIFFID. The equilibrium vegetation fractions are calculated by rearranging Eq. (1), meaning that for PFT i, the disturbance rate can be calculated as follows: γνi=λiΠi1-∑j=1npftcijνj×1CVi, where npft is the number of PFTs.

To estimate new values for γνi, we ran JULES for 60 years under present-day climate, CO2, and land use, solving for the equilibrium vegetation fractions (as summarized in Sect. 7 of Clark et al., 2011). We used the simulated vegetation carbon (CV), canopy height (to calculate the competition coefficients cij), and NPP for spreading (λΠ) at the end of the 60 years, together with the ESA LCCCI observed fraction of PFTs (νi) (Poulter et al., 2015), to solve for γνi in each grid cell. The result was a map of the γν (∼ disturbance rate) per PFT required to get the observed PFT distribution based on simulated carbon available. Based on global distributions of γν for each PFT in grid cells with < 50 % agriculture from 1950 to 2012, we used the median value in our simulations (Table 2). The new values of γν do not guarantee a perfect simulation of PFT distribution; this is due to the use of one value per PFT, and because the calculation was based on solving the equilibrium solution to Eq. (1). However, this method does result in a range of γν that make physical sense: there are low turnover rates for trees, high turnover rates for grasses, and moderate turnover rates for shrubs.

The vegetation fractions and soil carbon both require a long initial simulation to reach equilibrium. In a standard simulation, soil carbon spin-up needs to continue for 1000–2000 years after vegetation types have stabilized. There are two ways to speed this up: first by solving for vegetation fractions based on the equilibrium solution to Eq. (1); and second by using the “modified accelerated decomposition” technique (modified AD) (Koven et al., 2013). This results in a three-step spin-up, summarized below. Note that the first two steps used CRUNCEP-v4, which was available at the beginning of the project.

Solve for steady-state vegetation fractions in TRIFFID, increasing the time step for TRIFFID and phenology to 5 years and 10 days, respectively. Recycle the climate from the first 20 years of the simulation for a total of 60 years; in this case, CRUNCEP begins in 1900, so we recycled the 1901–1920 climate. In the simulations with HadGEM2-ES climate, the first 20 years of climate driving data is from 1860 to 1879. Specify land use and CO2 at their 1860 values.

In the last 100 years of the spin-up, soil carbon changed by -0.06 and 0.43 % with the CRUNCEP-v6 and HadGEM2-ES climates, respectively. These drifts are < 6 Pg C (100 years)-1, or 2.8 ppm (100 years)-1, which is below the C4MIP spin-up requirement for drifts of less than 10 ppm per century (Fig. S7). Therefore, 300 years is adequate for spinning up the model, but there is a benefit to using 500 years: the drift in soil carbon in the CRUNCEP-v6 climate from years 200 to 299 was -3.5 Pg C, compared to only -0.9 Pg C from years 400 to 499.

We evaluate the distribution of vegetation with two methods. First, to compare JULES-C1 and JULES-C2, we aggregate the nine PFTs into the original five. Figure 1 shows fractional coverage in each grid cell of the five vegetation types and bare soil for the models and the observations (BT  is  broadleaf trees, NT  is  needle-leaf trees, C3  is  C3 grasses, C4  is  C4 grasses, and SH  is  shrubs). Second, we calculated fractional coverage of each PFT in eight biomes based on the WWF ecoregions (Fig. 2). The eight biomes are tropical forests (TF), temperate mixed forests (MF), boreal forests (BF), tropical savannas (TS), temperate grasslands (TG), tundra (TU), Mediterranean woodland (MED), and deserts (D) (Fig. S9).

Most carbon in a tree/shrub is stored as woody biomass. Therefore, in terms of vegetation carbon content, the most important distinction between plant types is between trees, grasses, and shrubs. With the CRUNCEP-v6 climate, JULES-C2 represents the distribution of these broad vegetation types very well (Fig. 1). There are several improvements compared to JULES-C1; for example, both the amount of tropical broadleaf trees in the central tropical forests and the spatial extent of boreal forests are more realistic in JULES-C2. The boreal forests in JULES-C1 do not extend far enough across the North American and Eurasian continents. Instead, large areas of shrubs dominate at high latitudes. This bias is reduced in JULES-C2, although there is an underestimation (overestimation) in the coverage of needle-leaf trees in northeastern Eurasia (northern Europe).

Biome-scale distributions of the PFTs are shown in Fig. 2, with results from JULES-C2 with both the CRUNCEP-v6 and HadGEM2-ES climates. Differences between JULES-C2 run with different climates are typically small, with a tendency for the climate with higher precipitation to result in more trees (Fig. 3) (r2=0.66). Comparing the ESA vegetation fractions and CRUNCEP-v6 climate reveals a weaker positive relationship between tree coverage and annual rainfall (r2=0.36). JULES is also sensitive to the specific humidity (r2=0.25) but this is not supported by the ESA fractions. Coverage of needle-leaf deciduous trees ranges from 16 % with the CRUNCEP-v6 climate to 27 % with the HadGEM2-ES climate. This PFT was developed to have a competitive advantage in cold, dry environments. The annual average air temperature in the boreal forests is below freezing but precipitation is about 50 % higher in the HadGEM2-ES climate compared to the CRUNCEP-v6 climate (Fig. S8).

Agriculture is shown as a separate category since JULES can only grow C3 and C4 grasses in the agricultural fraction of grid cells. Agriculture accounts for 22–40 % of all biomes except the two high latitude biomes (boreal forests and tundra). To compare with the ESA PFT distributions, we reduce the “observed” agricultural fraction (from the HYDE3.2 dataset) on grid cells where the prescribed agricultural fraction is greater than the coverage of ESA observed grasses. This discrepancy between the observational datasets results in an apparent overestimation of agricultural fractions in some biomes. Although the agricultural fraction is prescribed, there can be bare soil on agricultural land if the JULES NPP is not sufficient to support vegetation (possibly due to the lack of irrigation in JULES). For this reason, in some biomes the agricultural fraction is underestimated (e.g., in temperate grasslands and deserts with JULES-C1).

JULES-C2 tends to overestimate the observed coverage of trees by 10–12 % in tropical forests and savannahs, and by 3–5 % in Mediterranean woodlands. The overestimation of trees in the tropical biomes is due to too many tropical broadleaf evergreen trees (BET-Tr). For example, in the tropical forest biome, 31 % of the biome is covered with BET-Tr in the observations compared to a simulated range of 40–44 % (with the HadGEM2-ES and CRUNCEP-v6 climates, respectively). The simulated coverage of broadleaf deciduous trees is very realistic in the tropical savannahs. The coverage of dominant tree types is also close to observed in the boreal and mixed forests, with needle-leaf deciduous and evergreen trees in the former and broadleaf deciduous and needle-leaf evergreen trees in the latter. However, the coverage of broadleaf deciduous trees is underestimated by 2–6 % in both biomes.

Grasses are overestimated compared to observations by up to 21 % in the boreal forests and tundra. The fractional coverage of bare soil is generally close to observed with errors < 5 % for every biome except for tundra, where it is underestimated. In this biome, JULES-C2 produces 10–13 % more shrubs and 10–21 % more grass than observed. In the temperate grasslands, JULES-C2 with HadGEM2-ES climate overestimates the grass and needle-leaf evergreen tree coverage and underestimates bare soil coverage. Precipitation is almost twice as high in this biome in HadGEM2-ES compared to CRUNCEP-v6 (Fig. S8). Shrubs in JULES-C2 tend to do best in cold environments – they are underestimated in tropical and mid-latitude biomes, very well simulated in the boreal forests, but overestimated in the tundra biome.

The total model biases based on bias per PFT are between 0.55 and 0.57 for all versions of the model (Table 4). The bias is an area-weighted fractional error per grid cell where the PFT exists (Eq. 15). The PFT biases are reduced for shrubs and grasses, but they are higher for broadleaf trees due to too many broadleaf trees in the tropics. The bias for needle-leaf trees in JULES-C2 depends on the climate: the bias is higher with the HadGEM2-ES climate compared to the CRUNCEP-v6 climate. Figure 2 also shows the bias calculated per biome for each simulation (Eq. 16). The biome biases are lowest in JULES-C2 with the HadGEM2-ES climate for five of the biomes, the exceptions being temperate grasslands, tundra, and deserts. In these biomes, the bias is lowest in JULES-C2 with the CRUNCEP-v6 climate. Comparing biomes, JULES-C2 represents vegetation distribution better in boreal and tropical forests than in mixed forests. The tropical savannahs have the highest bias.

The patterns of gross and net primary production (GPP and NPP, respectively) simulated by JULES are similar to estimates derived from observations, although JULES fluxes are slightly high (Fig. 4). From 1982 to 2011, GPP is 133 and 138 Pg C yr-1 according to JULES forced with CRUNCEP-v6 and HadGEM2-ES climate, respectively, compared to observation-based estimates from the same time period of 123 ± 8 Pg C yr-1 (1982–2011; Beer et al., 2010). JULES-C1 with the CRUNCEP-v6 climate produces a higher GPP (143 Pg C yr-1). GPP is lower in JULES-C2 compared to JULES-C1, and closer to observations, in the tropical biomes (savannahs and forests, Fig. 5a).

From 2000 to 2013, MODIS estimates an NPP of ∼ 55 Pg C yr-1, compared to 71 and 75 Pg C yr-1 in JULES with the CRUNCEP-v6 and HadGEM2-ES climates, respectively. During the same time period, JULES-C1 NPP is 66 Pg C yr-1. On average, NPP is 54 % of GPP in JULES-C2, while it is 46 % in JULES-C1. Both of these are similar to observation-based estimates that NPP should be roughly half of GPP. In JULES-C2, the largest overestimations of NPP occur in the tropical forests, savannahs, and mixed forests (Fig. 5b). JULES-C1 has high biases for GPP and NPP in tropical savannahs due to over-productive C4 grasses, and this bias is reduced in JULES-C2.

Global total vegetation carbon is 542 and 553 Pg C in JULES-C2 with the CRUNCEP-v6 and HadGEM2-ES climates, respectively, which is within the range supported by observations (400–600 Pg C, Prentice et al., 2001), and is 65 Pg C higher than the dataset from Ruesch and Gibbs (2008). The high bias mostly occurs in boreal and temperate forests and in tropical savannahs, where JULES produces more trees than observed (Fig. 5c). The spatial distribution of vegetation carbon is similar to observations (Fig. 4), but due to the extent of the broadleaf forests the total vegetation carbon in the tropical forest biome is higher than observed. However, there is large uncertainty in global biomass datasets, for example the tropical savannah biome in JULES is very comparable to the data from Carvalhais et al. (2014). JULES-C1 has lower vegetation carbon (468 Pg C), with the largest differences between the models being in the tropical forest and savannah biomes. There are two reasons for the increase in Cveg for JULES-C2. First, tropical evergreen and deciduous broadleaf trees are more prevalent in JULES-C2 (Fig. 1). Second, the low vegetation carbon was previously identified as a bias and the allometric parameters awl and aws were increased for broadleaf trees (Sect. 2.3.1).

The largest biases in JULES occur for soil carbon, which is underestimated in both the high latitudes and the tropics. Globally there is 1422 Pg C in JULES-C2 with the CRUNCEP-v6 climate and 1440 Pg C with the HadGEM2-ES climate, compared to 2420 Pg C in observations and 1362 Pg C in JULES-C1. Soil carbon is the result of centuries (or longer) of litter accumulation. Woody PFTs contribute more resistant material to the soil, while grasses turn over carbon in a more decomposable form. Therefore, relatively small differences between simulations in PFT distribution and NPP can contribute to large differences in the soil carbon. For example, in the tropics, soil carbon is higher in JULES-C2 corresponding to the presence of more broadleaf trees and fewer shrubs than in JULES-C1. In addition, due to the increased productivity simulated by JULES-C2, the amount of carbon going into the soils through litterfall is also increased.

Over the past century and according to JULES-C2, the land surface was a net sink of carbon due to an increase in soil carbon (+57 Pg C) that offset a smaller decrease in vegetation carbon (-48 Pg C) (Fig. 6). The changes in brackets are the average during 2000–2009 minus the average during 1900–1909. These changes can be attributed to climate change acting on its own, climate change plus CO2 fertilization, or climate change plus LUC. In the experiment with climate change only (SCLIM, Table 5), vegetation carbon increases by 40 Pg C, and there is a smaller increase in soil carbon since warming encourages decomposition.

The effects of CO2 fertilization and LUC on land carbon are given by the differences between experiments SALL and SLUC,CLIM, and between SALL and SCO2,CLIM, respectively. Higher levels of CO2 over the 20th century result in an additional 63 Pg C of soil carbon and 49 Pg C of vegetation carbon. This is due to larger increases in NPP and litterfall than heterotrophic soil respiration (Rh). Both NPP and Rh are 58 Pg C yr-1 in 1900 in SALL. NPP increases to ∼ 72 Pg C yr-1, while Rh increases to 70 Pg C yr-1 by the end of the simulation. Land use change results in a loss of 14 Pg C of soil carbon and 124 Pg C of vegetation carbon. The largest reductions in vegetation carbon occur in the tropics and in the eastern US and Europe (Fig. 7). The total land use source simulated by JULES (138 Pg C from 1900 to 2009) is very close to a recent estimate of total land use and land cover change emissions of 155 ± 50 Pg C from 1901 to 2012 (Li et al., 2017).

The annual sink is the net biosphere productivity (NBP), or NPP - Rh. The simulated NBP from 2000 to 2009 in JULES-C2 is 2.1 ± 1.0 Pg C yr-1. The net land sink simulated by JULES is within the range of estimates from both the Global Carbon Project (1.7 ± 0.8 Pg C yr-1 over the same period, Le Quéré et al., 2018) and the IPCC Fifth Assessment Report (AR5) (1.5 ± 0.7 Pg C yr-1) (Table 6). The JULES land sink is slightly high compared to the other two estimates, but this is not the case during the 1980s and 1990s. Excluding LUC, JULES-C2 simulates an NBP of 3.4 Pg C yr-1 in the 2000s, which is nearly double the natural NBP in the 1980s. The increase is due to a larger increase in simulated NPP in the experiment without land use change relative to the increase in Rh (Fig. 6). In SALL, the simulated NBP fluctuates around zero until the 1970s, after which it steadily increases due to the fertilizing effect of atmospheric CO2. Between 1980 and 2009, the NBP increases by 0.08 Pg C yr-1 yr-1, due to a stronger positive trend in NPP (+0.27 Pg C yr-1 yr-1) than in Rh (+0.19 Pg C yr-1 yr-1). This increase is not seen in the experiment with preindustrial CO2.