Figure 1. An example of visualisation of outputs of a forest model. Visualisation of species diversity (crown colours) of a tropical forest simulated by the FORMIND model (Fischer et al.2016) in the 3D visualisation center of UFZ – Helmholtz-Centre for Environmental Research, Leipzig, Germany.
Box 1. Model inter-comparison
Comparing the outputs from different models that are run under comparable or even identical conditions of driving variables offers valuable insights beyond single model simulations. Model comparisons in environmental sciences typically have two main objectives. The first is to understand differences between models by relating the simulated pattern of each model to its underlying processes and hence to understand how different model processes influence model behaviour and to pinpoint model structural uncertainties. The second objective is to provide ensemble simulations that allow for a quantitative assessment of uncertainties related to the actual predictions of the different models.
Model comparisons have a long history in ecology and environmental sciences. Prominent examples are comparisons of different forest gap models (Bugmann et al. 1996, 2019), forest landscape models, stand-based ecophysiological models (Kramer et al. 2002; Morales et al. 2005), dynamic global vegetation models (Cramer et al. 2001; Sitch et al. 2008) and species distribution models (Araújo & New 2007). More recently, the focus of model comparisons has expanded to also compare models across different model types (e.g. including both DGVMs and SDMs, Cheaib et al. 2012) and even across a wide range of sectors such as vegetation, water, agriculture or biodiversity to study the interaction of these under climate change (Frieler et al. 2017).
Beside the development of the study design, another challenge of model comparisons is the development of the simulation framework and the standardisation of both model inputs and outputs. Moreover, when complex process-based models are involved, whose uncertainties can not simply be attributed to individual processes, a major challenge is to interpret the ensemble runs and to understand which model processes actually explain the differences between models. To address all these issues, transparent model documentations and intensive exchange between modellers is needed accompanied by systematic tests of models and their components.
Box 2. Coupling of models
Each model has its own aim, history and therefore specific advantages and limitations. The coupling of a vegetation model with other types of models can be a valuable approach to take advantage of model complementarity or expand the initial scope of model applications.
For instance, several stand-scale forest models, including IBMs, have been coupled to models of emissions of biogenic volatile organic compounds, revealing that tree species composition and species-specific emission potentials were important drivers of the feedbacks between climate change and air quality (Keenan et al. 2009a, b; Wang et al. 2018). Similarly, a forest demographic model has been coupled to models of soil microbe-mediated biogeochemistry and competition for nutrients, revealing that spatial variation in soil properties can drive a large variation of forest biomass and composition (Medvigy et al. 2019; see also Sato et al. 2007 for another example of coupling between a forest model and more detailed soil modules). SDMs have been coupled to models of habitat colonization in order to take into account dispersal limitation in species distribution projections (e.g. Iverson et al. 2004; Nobis & Normand 2014; see also Franklin 2010). Fire disturbance models have also been implemented in several DGVMs (Yue et al. 2014, 2015; Lasslop et al. 2014; Schaphoff et al. 2018), but also in forest IBMs for a long time (Shugart & Noble 1981; Pausas 1999; Knapp et al. 2018), helping to explore different modelling approaches on the interaction between vegetation dynamics and fire (Hantson et al. 2016; Forkel et al. 2019) to explain the declining trend in global burnt area (Andela et al. 2017). More generally, forest models have been coupled to models of disturbances, such as wind storms (Seidl et al. 2011; Thom et al. 2017). Other examples include the coupling of a DGVM to a global economy model to dynamically include technical and societal changes in simulating future vegetation dynamics (e.g. Dietrich et al. 2019), allowing to investigate the possible trade-offs between bio-energy production and several sustainable development goals (Humpenöder et al. 2018).
Model development can also take advantage of the complementarity of different vegetation model types by coupling their different approaches into one model (McMahon et al. 2011). As an illustration, the gap model approach was implemented into a DGVM framework to better account for demographic processes and diversity in regional- to continental-scale studies (e.g. Smith et al. 2001; Sakschewski et al. 2015). Similarly, the seed dispersal (Lischke et al. 2006), which originates from individual-based forest models (Urban et al. 1991; Groeneveld et al. 2009), can be integrated into large scale forest models (Scherstjanoi et al. 2014; Lehsten et al. 2019) to account for dispersal limitation in predictions of species distribution changes under climate change.
Model coupling is usually challenging, since, in most models, some processes are hidden in parameters or strongly simplified functions and the model is usually balanced by fitting these parameters. If the simplified process or the parameter is replaced by a more complex submodel for the process, often the balance can be lost. Additionally, error propagation among models can also prove difficult (Dunford et al. 2015). Several model systems and software frameworks have been developed to facilitate multi-model coupling in a systematic way, and they even allow for switching between different models during a simulation (Haas et al. 2013).