(C) PLOS One This story was originally published by PLOS One and is unaltered. . . . . . . . . . . Dynamic Global Vegetation Models: Searching for the balance between demographic process representation and computational tractability [1] ['Arthur P. K. Argles', 'College Of Engineering', 'Mathematics', 'Physical Sciences', 'University Of Exeter', 'Exeter', 'United Kingdom', 'Jonathan R. Moore', 'Peter M. Cox'] Date: 2022-11 Vegetation is subject to multiple pressures in the 21st century, including changes in climate, atmospheric composition and human land-use. Changes in vegetation type, structure, and function also feed back to the climate through their impact on the surface-atmosphere fluxes of carbon and water. Dynamic Global Vegetation Models (DGVMs), are therefore key component of the latest Earth System Models (ESMs). Model projections for the future land carbon sink still span a wide range, in part due to the difficulty of representing complex ecosystem and biogeochemical processes at large scales (i.e. grid lengths ≈ 100km). The challenge for developers of DGVMs is therefore to find an optimal balance between detailed process representation and the ability to scale-up. We categorise DGVMs into four groups; Individual, Average Area, Two Dimensional Cohort and One Dimensional Cohort models. From this we review popular methods used to represent dynamic vegetation within the context of Earth System modelling. We argue that the minimum level of complexity required to effectively model changes in carbon storage under changing climate and disturbance regimes, requires a representation of tree size distributions within forests. Furthermore, we find that observed size distributions are consistent with Demographic Equilibrium Theory, suggesting that One Dimensional Cohort models with a focus on tree size, offer the best balance between computational tractability and realism for ESM applications. Funding: AA was funded by the Newton Fund through the Met Office Climate Science for Service Partnership Brazil (CSSP Brazil) and the UKRI NetZero+ project (BB/V011588/1). JM was funded by the ERC “ECCLES” project (742472), and PC was funded by the European Union’s Horizon 2020 project “Climate-Carbon Interactions in the Coming Century” (4C) (821003). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Data Availability: There is a public repository for the data used to produce Fig 1, Fig 2, and Fig 4 available on the doi: 10.5281/zenodo.6482851 . For data and methods present in Fig 3 , please refer to Moore et al., 2018, 2020. Copyright: © 2022 Argles et al. This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. To model these interactions, Dynamic Global Vegetation Models (DGVMs) are now being used within a growing number of Earth System Models (ESM). Understanding the dynamics of forest ecosystems and how the populations of plants change through time (through fecundity, growth and mortality), will inform researchers and policy makers about the risks associated with changes in land-use and climate. Climate and human influences are similarly important determinants of the geographical distribution of observed vegetation [ 4 , 5 ], which affects the land carbon storage and the hydrological cycle. A reduction in forest cover typically decreases water recycling, leading to drier conditions [ 6 ]. The additional loss of biomass is a source of carbon emissions and could lead to an acceleration of global warming [ 7 , 8 ]. Climate and land-use change could potentially cause a significant loss of regional forest ecosystems [ 9 ]. For instance, one key risk is Amazonian dieback [ 10 , 11 ]. Two models in the CMIP5 ensemble predicted dieback under the most extreme radiative forcing scenario [ 12 ], but the likelihood is increased when including deforestation, which reduces evapotranspiration and forest resilience to droughts [ 13 ]. Dieback of forest ecosystems has previously resulted in losses of biodiversity [ 14 , 15 ], social impacts [ 16 , 17 ], and the release of greenhouse gases [ 18 ]. Vegetation is a key determinant of future climate change. It is estimated that 3.4 ± 0.9 PgC yr −1 (±σ) of carbon was absorbed into vegetation and soils over the last decade [ 1 ]. This sink represents almost 30% of annual anthropogenic carbon emissions; 11.5 ± 0.9 PgC yr −1 . However, there remains a large range of uncertainty in the projections for the terrestrial sink over the coming century [ 2 , 3 ]. Dynamic Global Vegetation Models In the historical context of climate modelling, the representation of the land-surface has increased in complexity in recent decades. Initially, atmosphere-land interactions were represented in a rudimentary way, with Global Circulation Models using prescribed values for parameters such as surface albedo, surface roughness and effective root-depth [19]. Vegetation was first modelled explicitly to improve the representation of the diurnal cycle of temperature [20]. Later models would begin to prescribe vegetation types categorised by “Plant Functional Types” (PFTs): for example, C3 and C4 Grass or Broadleaf Deciduous Trees [21]. This allowed for stomatal conductance and phenological diversity to be more realistically represented [22–25]. However, even in these second generation land-surface models. the vegetation cover was prescribed and fixed, based-on either observed land-cover or “equilibrium” biogeographical models, such as the BIOME model [26, 27]. In parallel with the development of the first climate models, computational models were developed to represent community dynamics of forests, through the concept of a forest patch. Each patch would represent a community of species in both patch age and physiology, going through different “phases” of succession [28, 29]. A patch based approach was first implemented in the JABOWA [30, 31] and FORET [32] models. These are more commonly known as “gap models”, both simulating a grid of patches within an ecosystem on the scale of 0.01 to 0.1 hectares. Gaps are randomly formed by disturbances, resetting a patch at the beginning of its successional pathway. Other forest models sought to represent explicit individuals of tree species in three-dimensional space, to capture vertical heterogeneity [33, 34]. Environmental variables such as temperature and precipitation drive the photosynthesis and growth of plants. Light competition is most simply represented using Beer’s law of radiation attenuation [35, 36]. Age and size are principal variables in determining mortality and competition. The reproduction or recruitment rate is either fixed based-on species or dependent on the size of the current plant population. Gap models have proven effective in simulating the dynamics and distributions of forests in multiple site-level studies of coniferous [37], temperate [38] and tropical forests [39, 40]. They have also been successfully deployed for projecting forestry management yields [41–43]. By the 1990’s demand increased for the inclusion of vegetation dynamics within climate models [44]. This has led to the creation of a new class of ecosystem models known as Dynamic Global Vegetation Models (DGVMs) [40, 45, 46]. These DGVMs had the notable challenge of modelling scales from the size of hectares to hundred of kilometers, as illustrated by Fig 1. This led to a range of solutions either built by scaling up the dynamics of the individual or by taking a “top down” approach. PPT PowerPoint slide PNG larger image TIFF original image Download: Fig 1. Adapted from Shugart et al., 2020 [ Adapted from Shugart et al., 2020 [ 47 ], schematic showing how vegetation processes scale across space and time. Grey boxes classify the biological and community scale. Yellow text describes leaf-based processes. Blue text describes endogenous processes. Red text describes exogenous disturbances and processes. The black dotted lines illustrate the current spatial and temporal scope of Earth System Models. https://doi.org/10.1371/journal.pclm.0000068.g001 Individual Based models Individual Based DGVMs are descendants of the original gap models, but often coupled to more sophisticated representations of water and carbon fluxes. The earliest example of this method was the appropriately named, HYBRID (version 3.0) model [48, 49]. HYBRID simulates many individual trees established from a fixed density of seeds. As with gap models, crown competition for light is represented using Beer’s law. HYBRID can be driven globally using climate variables from a GCM, typically by subdividing each grid-cell into up to ten patches [50]. The SEIB-DGVM model uses a similar methodology at the global scale [51]. These individual-based models have the potential to model plant-specific processes in greater detail, resulting in more realistic modelling of site-level dynamics [52]. For practical reasons, Individual Based models use the process of “upscaling” from the scale of the stand to that of the land-surface grid-box [53]. SEIB-DGVM scales 30mx30m plots up to the order of 200kmx200km grid-boxes [51]. When coupled to the MIROC-ESM [54], SEIB-DGVM simulates in essence one random patch for one grid-box, which may lead to spurious statistical variation between grid-boxes. Nevertheless, this form of explicit upscaling has been used effectively to study the influence of soil variation on high latitude ecosystems [55], and to forecast typhoon windthrow in Japan under climate change [56]. Individual Based models have also been used to address some of the criticisms of the concept of Plant Functional Types (PFTs). Plant species have a wide and diverse continuum of traits in contrast to the discreteness offered by globally implemented PFTs [57]. It has been argued that PFTs underplay the resilience of forest ecosystems through the lack of modelled biodiversity [58, 59], and may therefore not capture the reality of environmental sorting of species and dynamics at scale [60, 61]. Recently, individual-trait based models have been developed to address this criticism [62]. aDGVM2 [61] and LPJmL‐FIT [63] use stochasticity to select the traits of individual plants (carbon allocation, fecundity, etc.) from a distribution represented with huge databases such as TRY [64]. This method has had some success in reproducing the biodiversity observed within the Amazon rainforest [65]. Individual Based models can however fail to capture some large scale dynamic processes [47, 66]. For example, the influence of forest fires can extend well past single patch scale. A single fire can eliminate an entire patch, especially in savannah, within a few minutes [67]. Therefore, to include fire, Individual Based models with a single patch must represent fire through some form of parametrisation. SEIB-DGVM relies on an end-of-year probabilistic estimation of fire based on environmental factors [68], after which the model applies an average risk of plant mortality. However, generally Individual Based models that represent multiple subgrid patches are able to explicitly simulate fire. In LPJ-GUESS in cohort and individual mode Uniquely, LPJ-GUESS has two modes “population mode”, which is an Area Averaged vegetation model (see section Area Averaged Models), and “cohort and individual mode” which simulates multiple patches with a three dimensional simulation of trees within each patch, fire mortality is applied daily across multiple patches (100 patches per 0.1 ha in grid-box size)—removing all individuals within the patch if a fire occurs [69]. Area Averaged models An alternative approach to representing global vegetation focuses on the area-averaged properties of most relevance to climate, such as biomass and leaf area index. These top-down models essentially represent the biophysical properties of a typical plant on some vegetated fraction of a grid-box. In many cases, these dynamical models were born out of the earlier equilibrium biogeography models [26]. The VECODE model [5], sought to represent dynamic vegetation fractions as a consequence of two climate variables—temperature and precipitation. Competition was simply between two PFTs—trees and grass, the fraction of which must always add up to one. VECODE successfully captured the general distribution of tree coverages [70]. However, this approach underplays how fast vegetation cover might change, due to the lack of physical processes such as disturbances or land-use change [71]. In an expansion to the VECODE approach, the DYNVEG model simulates PFTs with the inclusion of dynamic disturbances such as fire and windthrow [72]. The TRIFFID [73] and CTEM models [74] represent PFTs by the fraction of grid-box area they cover and model competition between the PFTs through a Lotka scheme [75]. In TRIFFID, vegetation tiles were simulated with area averaged productivity and carbon biomass. PFTs accumulate carbon into three pools—wood, leaves and litter. For recruitment, TRIFFID uses an LAI (Leaf-Area Index) dependent fraction of carbon mass assimilate to “spread”, or expand, the coverage. PFTs are grouped via a competitive hierarchy, where dominant PFTs (Trees) can freely-access the space occupied by sub-dominant PFTs (Shrubs and Grasses). The competition between the different tree PFTs is modelled using relative heights estimated from allometry. VECODE, TRIFFID, DIVE [76] and DYNVEG collectively provide us with one of the key large-scale simplifications for dynamic vegetation modelling; the sum of PFT/plant fractional grid-box area, ∑ν i , and the total unoccupied or bare space, ν bare , must be equal to one, the total grid box fractional area: (1) Let us call this rule the “hard exclusion approximation”, which implies that despite plants competing for many resources (light, water, and below ground nutrients, see [77]), space is the principal limiting resource [5, 78]. The “hard” element describes the absence of overlap between the spaces occupied by PFTs, whereas in reality there is a degree of tolerated overlap between plants, in terms of both physical space and ecological niches [79]. This method of area exclusion competition has had a number of criticisms [61, 80]. It has been argued that this will not realistically capture biome shifts or the species sorting that may occur under climate change [81]. Also, the abstraction of competition means that such dynamic relationships are hard to relate to field measurements [82]. Lotka approaches that use coverage and hard exclusion for competition have also been criticised in terms of biodiversity [61]. For instance, in the steady-state they cannot simulate the observed mixed communities of broadleaf or needleleaf trees seen in temperate latitudes [74, 83]. In a Lotka context, this is because having a competition coefficient equal to one can generate complete exclusion of sub-dominant PFTs. Having a competition coefficient less than one generates diversity, but this breaks the hard exclusion rule; the overall total vegetation area can then exceed the grid-box area. The CTEM achieves diversity by allowing competition coefficients to be dependent on a more complex competition-colonisation parametrisation based-on inter-PFT invasion rates [74]. The hard exclusion approximation has been used to represent area-based disturbances, from land use change to fire, along with controlling the rate of subsequent regrowth [72, 84–90]. Area based land-use change is perhaps easiest to simulate, not necessarily needing to represent plant demography. The implementation of land use change in JULES-TRIFFID is to simply replace the natural PFTs with crop PFTs at an area dependent rate [91]. However, the representation of fire or drought disturbance through purely area, may overlook important dependencies on plant size and age [92]. A second approach to Area Averaged models is to build dynamics by primarily mechanistic means. The IBIS [93], SDGVM [94, 95], MC1 [96], LPJ [97] and VEGAS [98, 99] models simulate explicit light and/or water competition of mean PFT members within the ecosystems. This is typically done by partitioning the population into different “layers”, representing either the canopy heights or root depths of vegetation. Additionally they may use an implementation of the self-thinning rule for competitive based mortality. There are benefits to this approach as, unlike the Lotka and area exclusion methods, there is a more direct link to physical processes, allowing for more model validation with empirical measurements [68, 100–102]. However, there is a cost associated with increasing the number of modelled processes, and thereby the number of uncertain parameters, which may lead to additional uncertainty at regional scales [103–105]. A stochastic trait based approach has been implemented in Area Averaged models through the JeDi-DGVM [81]. Lotka approaches are inappropriate for trait-based stochasticity, as the large number of possible PFTs implies that the competition coefficient matrix can quickly become computationally unwieldly; at rate of I2, with I being the total number of PFTs. JeDI-DGVM can generate thousands of randomly sampled growth strategies on a global scale. For grid-box scaling the model uses the “biomass-ratio” hypothesis [106], as simulated in the JeDI biogeographical equilibrium model [107, 108]. The Biomass-ratio hypothesis is where the function (i.e. dominant, sub-dominant growth strategies) of the species is dictated by the relative contribution it makes to the total community biomass. The JeDI-DGVM does not represent mechanistic competition for light or space. Similarly, the DIVE model uses the original JeDI approach to pre-select a distribution of PFT traits, but DIVE then updates the PFTs area using the hard-exclusion rule [76]. The PFTs in this latter model are prescribed using diversity outputs from the original JeDI trait model [107, 108]. [END] --- [1] Url: https://journals.plos.org/climate/article?id=10.1371/journal.pclm.0000068 Published and (C) by PLOS One Content appears here under this condition or license: Creative Commons - Attribution BY 4.0. via Magical.Fish Gopher News Feeds: gopher://magical.fish/1/feeds/news/plosone/