The forest, savanna, and grassland biomes, and the transitions between them, are expected to undergo major changes in the future due to global climate change. Dynamic global vegetation models (DGVMs) are very useful for understanding vegetation dynamics under the present climate, and for predicting its changes under future conditions. However, several DGVMs display high uncertainty in predicting vegetation in tropical areas. Here we perform a comparative analysis of three different DGVMs (JSBACH, LPJ-GUESS-SPITFIRE and aDGVM) with regard to their representation of the ecological mechanisms and feedbacks that determine the forest, savanna, and grassland biomes, in an attempt to bridge the knowledge gap between ecology and global modeling. The outcomes of the models, which include different mechanisms, are compared to observed tree cover along a mean annual precipitation gradient in Africa. By drawing on the large number of recent studies that have delivered new insights into the ecology of tropical ecosystems in general, and of savannas in particular, we identify two main mechanisms that need improved representation in the examined DGVMs. The first mechanism includes water limitation to tree growth, and tree–grass competition for water, which are key factors in determining savanna presence in arid and semi-arid areas. The second is a grass–fire feedback, which maintains both forest and savanna presence in mesic areas. Grasses constitute the majority of the fuel load, and at the same time benefit from the openness of the landscape after fires, since they recover faster than trees. Additionally, these two mechanisms are better represented when the models also include tree life stages (adults and seedlings), and distinguish between fire-prone and shade-tolerant forest trees, and fire-resistant and shade-intolerant savanna trees. Including these basic elements could improve the predictive ability of the DGVMs, not only under current climate conditions but also and especially under future scenarios.
Savannas cover about a fifth of the Earth's land surface, and have wide
socioeconomic importance regarding land use and biodiversity (Scholes, 2003).
Savannas are the central biome in the transition between grasslands and
forests, and they are characterized by the coexistence of two types of
vegetation: trees (i.e., woody vegetation), and grasses (i.e., grasses and
herbs). In most of the savanna ecosystems, we observe highly shade-intolerant
and fire-tolerant C
In addition to water availability, fire is an important driver of tree–grass
dynamics. C
Savannas are expected to undergo major changes in the future due to
increasing temperature and CO
Dynamic Global Vegetation Models (DGVMs) are an important tool for understanding large-scale vegetation dynamics, and they are considered important also to study the forest, savanna, and grassland biomes, and their interactions within past, current and future climates (Higgins and Scheiter, 2012; Murphy and Bowman, 2012). Some DGVMs are part of Earth System Models (ESMs), where they describe the interactive role of the Earth's land surface in the climate system. Given their global application, DGVMs necessarily keep the descriptions of vegetation dynamics simple. For example, they represent the enormous plant trait diversity of tropical regions by distinguishing only one or two plant functional types (PFTs). Nevertheless, they realistically reproduce the distribution of the majority of the world's biomes (Fisher et al., 2010; Sitch et al., 2003). However, projections of vegetation distribution by DGVMs are often uncertain, especially for the forest, savanna, and grassland biomes (Bonan et al., 2003; Cramer et al., 2001; Hely et al., 2006; Hickler et al., 2006; Sato et al., 2007; Sitch et al., 2008). This is probably a consequence of the fact that most DGVMs were not specifically designed for these tropical systems (House et al., 2003), and thus they do not include the specific internal feedbacks typical of these biomes (Moncrieff et al., 2013). Improving the DGVM representation of ecological processes under present climatic conditions is essential for projecting biome boundary shifts and climate change impacts into the future (Beerling and Osborne, 2006; Murphy and Bowman, 2012; Sitch et al., 2008).
To evaluate why DGVMs may have difficulties predicting the distribution and dynamics of savannas, we will analyze three DGVMs, with a particular emphasis on the representation of what in the following we call the “ecological interactions” between grasses and trees, i.e., the most important tree–grass competition mechanisms, and the feedbacks with their environment. While physiological processes are often included in detail in DGVMs, the ecological interactions are not represented with the same accuracy in many models, despite their potentially large influence on the DGVM outcomes (e.g., Fisher et al., 2010; Scheiter et al., 2013). Reflecting on the current ecological understandings of savannas, we will describe if and how the key mechanisms are included in current DGVMs. We chose to analyze three different DGVMs: JSBACH (Brovkin et al., 2009; Raddatz et al., 2007; Reick et al., 2013), LPJ-GUESS-SPITFIRE (Smith et al., 2001; Thonicke et al., 2010) and aDGVM (Scheiter and Higgins, 2009). JSBACH represents a DGVM as typically used in ESMs (and representative for most models included in the current IPCC coupled model inter-comparison project, CMIP5). LPJ-GUESS additionally includes the demography of PFTs, which is likely to affect competition dynamics, and it includes SPITFIRE, i.e., a new specific module to represent fire dynamics. Finally, aDGVM represents a new class of DGVMs, including functional variation within PFTs (e.g., phenology, allocation, and physiology adapt to changing environmental conditions). The aDGVM was specifically designed for African vegetation and savannas. In the following, we will focus on the African continent, where savannas occupy large areas and where all of the three models have been applied (Brovkin et al., 2009; Hickler et al., 2009; Higgins and Scheiter, 2012; Lehsten et al., 2009; Scheiter and Higgins, 2009). Focusing on one continent also has the advantage that the mechanisms driving the dynamics are more likely to be similar (Lehmann et al., 2014). We will compare the model outputs with observations from field and remote sensing data (Hirota et al., 2011; Sankaran et al., 2005; Staver et al., 2011). We attempt to bridge the knowledge gap between our ecological understanding and the representations of vegetation in global vegetation models. Our aim is to determine which mechanisms need to be included or improved in the representation of ecological interactions of existing DGVMs in the forest, savanna, and grassland biomes, to ameliorate the current vegetation model predictions, as well as their projections under future (e.g., climate change) scenarios.
DGVMs were developed to quantify transient responses of terrestrial
ecosystems to past, present and future climates, and this required the
inclusion of modeling vegetation dynamics in addition to biogeochemical
processes (Cramer et al., 2001; Pitman, 2003; Prentice et al., 2007). To
account for processes at the subgrid-scale, DGVMs often assume fractional
vegetation cover within the model grid cell (tiling or mosaic approach).
Vegetation description is based on PFTs, which aggregate and represent
species with similar functions. Biomes are then represented by a mixture of
PFTs, such as evergreen and deciduous, broadleaved and needleleaved trees,
shrubs, and C
DGVMs in general have a quite standard set of assumptions to represent plant
physiology, including photosynthesis and biomass production. Most of them
calculate gross primary production (GPP) by a coupled
photosynthesis–transpiration scheme, and estimate autotrophic respiration as
a function of temperature. Net primary production (NPP) is dependent on the
climate and CO
Models and their specifics concerning the tree–grass transition.
Continued.
For the purpose of this paper, we will focus on the description of how the
ecological processes relevant for tropical vegetation dynamics are included
in the three selected DGVMs (JSBACH, LPJ-GUESS-SPITFIRE, and aDGVM). Only the
physiological aspects relevant for the difference in PFT composition in
grasslands, savannas, and forests will be described. JSBACH is part of an
ESM and was designed to represent the interactive role of vegetation and
land surface in the climate system. While LPJ-GUESS has been included in an
ESM in several studies, LPJ-GUESS-SPITFIRE has never been used in such
context, and the same holds for aDGVM. Both models are so far used only
“offline”, i.e., they are driven by external forcing, such as climate and
CO
DYNVEG (Brovkin et al., 2009; Reick et al., 2013) is the submodel for
vegetation dynamics implemented in the land surface component JSBACH (Raddatz
et al., 2007) of the Max Planck Institute – Earth System Model (MPI-ESM,
Giorgetta et al., 2013). DYNVEG groups its various PFTs into a grass class
(C
DYNVEG includes a simple representation of fire disturbance. The fraction of burned area increases with higher amounts of litter (i.e., fuel), mostly produced by woody vegetation and decreasing air humidity (a substitute for litter moisture). As a result, savannas in northern Africa with relatively low air humidity and high productivity frequently burn. After the fire, the burned area is quickly occupied by grasses, while woody cover recovers slowly. Thus, in these transient dynamics, grasses indirectly slow down tree growth. Fire disturbance is the main process that keeps a mixture of trees and grasses in drylands.
LPJ-GUESS (Smith et al., 2001) was developed to incorporate forest age
structure into LPJ (Sitch et al., 2003), thus simulating gap model behavior
and including the competition of different age cohorts for light and water.
For each grid cell, LPJ-GUESS simulates a number of replicate patches. For
the tropical regions, LPJ-GUESS results in one type of (C
The fire module SPITFIRE (SPread and InTensity of FIRE, Lehsten et al., 2009; Thonicke et al., 2010) was coupled to LPJ-GUESS to include the role of vegetation fires. The effects of fire, simulated by SPITFIRE, vary for the different demographic stages (or height classes). For each fire, fuel load, wind speed, and a proxy for fuel moisture are used to calculate the rate of spread of a potential fire. The fuel load depends on NPP and decomposition rates, which are both related to climate. Grassy fuels are more flammable (due to their lower fuel bulk density), but trees can accumulate more fuel over years without fire, since they decompose more slowly. Hence if burned at high to medium fire frequency, grasslands provide more fuel than forests, while if forests are allowed to accumulate fuel over longer time periods, they provide higher fuel loads than grasslands. All fires remove the aboveground biomass of all grasses. Low-intensity fires can cause high mortality of all young trees, while the effects of these fires on tall trees are limited for savanna trees and more pronounced for forest trees. In general, damage to trees may be underestimated by SPITFIRE in the current parameterization. In fact, frequent fires lead to high mortality of young (and thus physically small) age cohorts, while the direct effects on old age cohorts are very limited, and only large fires can cause high mortality for highly resistant savanna trees. Further details on the implementation of fire effects on vegetation can be found in Lehsten et al. (2009).
The aDGVM (Scheiter and Higgins, 2009) is explicitly designed to study
tree–grass dynamics in savannas. While the original version of the model only
simulates savanna trees and C
Plants compete mostly for water and light. Light competition is modeled by considering the light available to grasses below and between canopies. Hence, once a vegetation stand attains a high tree leaf area index (LAI), grass growth is light limited. In addition, seedlings and small trees are shaded by grasses and by adult trees. Savanna trees suffer more from light limitation than forest trees, which are more shade tolerant. Plants extract water from different soil layers, depending on their rooting depth, which increases with the individual root biomass, until reaching maximum value, typically parameterized as being deeper for trees than for grasses. This allows trees to have exclusive access to water in deep soil layers. A simple bucket scheme is used to simulate water extraction and percolation into deeper soil layers. The extent to which soil moisture limits photosynthesis is calculated as a function of soil moisture in the layers in which the plant has roots. Hence, rooting depth, the amount of water transpired, and drought tolerance (i.e., the ability to withstand a low soil water content) determine the outcome of competition for soil moisture.
Fire intensity in the aDGVM is a function of the grass fuel load, its moisture content and wind speed (following Higgins et al., 2008). Fire spreads when the fire intensity exceeds a minimum intensity, when a fire ignition event (for example lightning strike) occurs, and when ignition probability is exceeded. Days when ignitions occur are random, the number of ignition events per year is linked to tree cover. Fire is assumed to consume a large proportion of aboveground grass biomass. Aboveground grass biomass burns as a function of the fire intensity. The aDGVM models the probability of stem biomass combustion of individual trees (so-called “topkilled” trees, which remain alive after fire and can resprout from their roots) as being a logistic function of stem height and fire intensity (following Higgins et al., 2000). This function varies with tree type. Topkill rates are higher for forest than for savanna trees, and savanna trees have higher re-sprouting rates than forest trees, which can be killed by a sequence of fires. Fire affects tree mortality only indirectly, by influencing the carbon balance of topkilled trees. The fire sub-model and the topkill model together determine if trees remain trapped in a cycle of topkill and resprouting, or if they can attain larger, fire-resistant sizes. Scheiter and Higgins (2009) illustrated that the aDGVM simulates the current distribution of African biomes well, and that it can simulate biomass observed in a long-term fire manipulation experiment in Kruger National Park, South Africa (Higgins et al., 2007).
To simulate current conditions, transient simulations were performed where
CO
Tree cover as a function of mean annual rainfall (mm yr
For the comparison between data and models, we used two different types of
tree cover observational data sets that have been recently used to study
savanna dynamics. One data set is a collection of tree cover data from savanna
field sites from Africa (Sankaran et al., 2005), while the other is derived
from remote sensing (MODIS), as used, e.g., in Hirota et al. (2011) and Staver et
al. (2011b). In both cases, we selected only the data points between
35
The data set from Sankaran et al. (2005) includes data from 854 field sites
across Africa. They gathered data from several sources, with no recent human
influence, not situated in riparian or seasonally flooded areas, and where
vegetation was sampled on a sufficiently large area (
The tree cover data set, derived from remote sensing data, was the result of
two combined databases. Tree cover data were obtained from the MODIS woody
cover product (MOD44B), developed by Hansen et al. (2003). This product used
MODIS images between Oct 2000 and Dec 2001 to calculate the fraction of tree
cover, with a spatial resolution of 500 m. To exclude areas highly
influenced by humans, we combined this data with the global land cover map
(GlobCover 2009) with a high spatial resolution (300 m). We excluded land
cover types that were classified as post-flooding or irrigated croplands;
rainfed croplands; mosaic cropland (50–70 %) and vegetation (grassland,
shrubland, forest; 20–50 %); mosaic vegetation (grassland, shrubland,
forest 50–70 %) and cropland (20–50 %); artificial surfaces and
associated area (urban areas
As for the data, for the three models we analyzed the simulated tree cover
output (i.e., all woody vegetation) as a function of the corresponding mean
annual rainfall conditions, and we select only the points in the African
continent between 35
In the models, the precipitation ranges where grasslands, savannas, and forests were simulated resulted not only from the different representations of vegetation dynamics, but also from the way climate was included. aDGVM and LPJ-GUESS-SPITFIRE were forced with (different) climate data, while JSBACH was coupled to an atmospheric model. Both the rainfall (NCEP, CRU, and TRMM) data sets and the simulated climate have inevitable biases, and are hard to compare with each other. Therefore, precipitation estimations were not totally comparable, and for this reason, we will compare the models in the parameter space (i.e., vegetation cover versus mean annual rainfall) and not in the geographical space. Also, we will not discuss the exact mean annual rainfall values at which forest, savanna, and grassland are observed, but we mostly refer to ranges of low, medium, or high mean annual rainfall. For these ranges, we will perform a qualitative comparison of the modeled and observed data in the parameter space (i.e., maximum values, spread, distribution).
In addition to mean annual rainfall, other factors, such as temperature (Higgins and Scheiter, 2012) and temporal distribution of rainfall, are known to be important as well for tropical grasslands, savannas, and forests. Rainfall heterogeneity, intermittency, and seasonality affect water availability (D'Onofrio et al., 2014) and fire return times, and are very important predictors of savanna/forest distribution (Lehmann et al., 2011), with rainfall seasonality reducing growth rates (e.g., limiting water availability, Sarmiento, 1984), influencing root–shoot biomass ratio and local cover (Yin et al., 2014a), and increasing fire frequency (Archibald et al., 2009). Nevertheless, these factors have not yet been thoroughly examined in many ecological studies, possibly, among other reasons, because of lack of accurate rainfall data sets in these areas. Therefore, in the following, we will focus only on mean annual rainfall, whose importance has extensively been studied. We separately evaluate arid and semi-arid savannas (Sect. 3.1), and humid savannas and forests (Sect. 3.2), analyzing also if and how the ecological interactions are included in the different models. Finally, we discuss the effects of expected future climatic changes on the outcome of tree–grass competition in the three models (Sect. 3.3).
In the drier African savanna regions, i.e., with mean annual precipitation
lower than a value estimated between 650 mm yr
At first glance, the relation between tree cover and mean annual rainfall
simulated by the models (Fig. 2) is similar to that observed in the data
(Fig. 1). In JSBACH output, the maximum tree cover increases between zero and
800 mm yr
In the LPJ-GUESS-SPITFIRE model output (Fig. 2b), almost no tree cover is
observed for mean annual rainfall up to about 300 mm yr
In the aDGVM output, the tree cover displays a maximum value that grows with
precipitation between zero and about 500 mm yr
Model outputs for tree cover as a function of mean annual rainfall
(mm yr
In more humid conditions, bimodality of vegetation cover below and above
60 % is observed in the MODIS data for precipitation in a range between
around 1000 and 2000 mm yr
Schematic diagram of the main ecological interactions that determine
the forest–savanna–grassland transition, according to
The role of fire in maintaining savannas in humid environments is included in
all of the models, although in different ways. At high precipitation, JSBACH
tree cover output displays a constant maximum value (above about
800 mm yr
LPJ-GUESS-SPITFIRE simulation results do not show any low tree cover value
(e.g., below 50 % cover) for rainfall higher than about
900 mm yr
In aDGVM, maximum tree cover values can reach full cover above about
500 mm yr
Finally, we note that at extremely high rainfall values, when water is not limiting and tree canopies close into a forest, both in LPJ-GUESS-SPITFIRE and aDGVM, trees exclude grasses through light competition (Fig. 2b–c). This mechanism is only implicitly included in JSBACH, and it acts along the whole precipitation gradient giving competitive advantage to trees in general.
Hereafter we discuss results from two simple conceptual experiments (namely
increasing CO
Expected increase in CO
Another consequence of climate change is a possible decrease in precipitation. This scenario also leads to different model behavior. In JSBACH and LPJ-GUESS-SPITFIRE, drier conditions lead to lower (woody) biomass productivity, but the impact on fire spread differs between these two models. JSBACH predicts no major effects on fire as drier conditions would lead to higher fuel flammability, thus compensating for the impacts of the woody biomass decrease. In LPJ-GUESS-SPITFIRE the decrease in productivity is dominant, and hence a strong decrease of fire frequency is expected (Lehsten et al., 2010). In aDGVM the strong positive feedback would lead to a magnification of the woody vegetation decrease, as lower precipitation leads to increased grass productivity (because of less competition with woody vegetation) and lower humidity, increasing the likelihood of fire occurrence.
In summary, we expect that in JSBACH, LPJ-GUESS-SPITFIRE and aDGVM, savanna systems have quite different sensitivities to climate change, and their predictions on the effects of climate change on fire occurrence diverge substantially. Given the importance of fires for estimating the global carbon budget (Le Quéré et al., 2013), this is remarkable, and it illustrates clearly how representing the ecological interactions more or less accurately can lead in some cases to similar results under present conditions (where the models have been tuned), but their predictions can diverge substantially when the models are used for future scenarios.
Up to now we considered water limitation and fires as the main drivers of grassland, savanna, and forest distribution. Several additional factors can be important for vegetation dynamics, especially at the local scale. The first factor is herbivory. Browsing (particularly by mega-herbivores in Africa) is known to have an important limiting effect on tree cover, similar to the effects of fire (e.g., Scheiter and Higgins, 2012; Staver et al., 2012), while grazing can favor trees because it limits grass expansion (e.g., Sankaran et al., 2008). However, large herbivores do not seem to be critical in determining forest and savanna distributions (Murphy and Bowman, 2012). Secondly, although it has been observed that savannas can be associated with nutrient poor soils (Lloyd et al., 2008), it is generally accepted that nutrient limitation does not explain the savanna–forest transition (Bond, 2010; Favier et al., 2012; Murphy and Bowman, 2012). For these reasons, and to avoid inconsistencies while evaluating different models, we only used DGVMs that did not include nutrient cycling. Thirdly, vegetation tends to have local spatial dynamics and to feed back to the environment at much smaller spatial scales than the DGVMs' resolution. These local spatial water–vegetation interactions are strictly connected to vegetation resilience in arid and semiarid ecosystems (e.g., Rietkerk et al., 2004), and they can also influence the coexistence of trees and grasses in the most arid savannas (Baudena and Rietkerk, 2013; Nathan et al., 2013). Although the local scale is partly taken into account in some DGVMs by including individual-based dynamics or tiling schemes (that represent different vegetation types and bare soil next to each other within the same cell), these assume a common use of soil and hydrological resources within the grid cell, thus not allowing representation of the local, sub-grid mechanisms, which are fairly difficult to scale up (Rietkerk et al., 2011). Finally, on the African continent the vast majority of fires are ignited by humans (Archibald et al., 2009; Saarnak, 2001), although their decisions on when to burn an area, as well as the fire spread and intensity, are still related to fuel composition (Govender et al., 2006). Humans maintain the grass–fire feedback, since they aim to keep the land free of woody vegetation, and also because fire spread is favored by grass presence (Ratnam et al., 2011). Changes in land use have therefore strong influences on the current and future outcomes of tree–grass competition. Also, humans are expected to change their application of fire as a land use tool, as a consequence of changed environmental conditions. These elements are partly taken into account in some DGVMs (e.g., in LPJ-GUESS-SPITFIRE), but we do not consider them here for the purpose of this paper.
Current ecological understanding identifies water limitation and grass–fire feedback as dominant mechanisms driving the forest–savanna–grassland transition in Africa. In arid and semiarid savannas, trees are water limited, and the water competition with grasses is the key factor determining savanna existence. In these conditions, grasses compete particularly fiercely with tree seedlings. In wetter areas along the climatic gradient, savannas are maintained by the presence of a positive grass–fire feedback. Fire spread is increased by grasses, which provide fuel load. Grasses re-grow faster than trees after fires, while tree recruitment is limited. Thus, trees do not close their canopies, leaving more free space for grasses. On the other hand, when trees manage to close their canopies, grasses are outcompeted because of light limitations, and because fire is suppressed. This grass–fire feedback is reinforced by the higher flammability of forest trees with respect to savanna trees. Both water limitations and fires act differently on tree adults and seedlings, which compete more directly with grasses and are the most sensitive stage in tree life.
These mechanisms are to varying extents included in the three DGVMs we
analyzed (JSBACH, LPJ-GUESS-SPITFIRE and aDGVM). Indeed, the three models
predict the main features of the current tree cover along the mean annual
rainfall gradient in Africa, as derived from ground and satellite
observations. aDGVM output matches the observations better than the other two
models. This is perhaps to be expected since this model is specifically
designed for African vegetation and it includes more detailed representations
of ecological interactions, especially the vegetation–fire feedback. For the
other two models, the main differences between observations and model outputs
are (i) JSBACH overestimates tree cover in dry areas (see also Brovkin et
al., 2013); (ii) LPJ-GUESS-SPITFIRE does not show any savanna at medium to
high annual rainfall rates; (iii) both these DGVMs do not show bimodality of
savannas and forests in humid areas. This latter point might feed the debate
over whether or not bimodality between savanna and forest cover actually
exists (see, e.g., Hanan et al., 2014). Despite their reasonably good
performances, not all the mechanisms included in JSBACH and
LPJ-GUESS-SPITFIRE are fully appropriate to represent vegetation in the
tropics and subtropics. In JSBACH, competition between trees and grasses
favors the former irrespectively of water availability, which is one of the
reasons behind JSBACH tree cover overestimation. At the same time, in this
model, fire is fostered disproportionately by woody vegetation as compared to
grasses, resulting in a negative feedback. This is responsible for observing
savannas in larger parts of the rainfall gradients, and no savannas would be
simulated without them. Although the three models display comparable outcomes
under the current climate, the negative fire–vegetation feedback in JSBACH,
the positive feedback in aDGVM, and the intermediate behavior of
LPJ-GUESS-SPITFIRE lead to different predictions of fire frequency and
effects under climate change scenarios between the three models. In JSBACH,
the initial increase in woody vegetation due to higher CO
Tree seedlings are the bottleneck stage of tree life in the forest–savanna–grassland transition (Salazar et al., 2012; Sankaran et al., 2004), and the two most important mechanisms we identified here, i.e., fires and water competition and limitation tend to affect tree seedlings particularly strongly. Thus, including tree demography, as in LPJ-GUESS and the aDGVM, improves the representation of ecological interactions in the models. Also, representing forest and savanna trees with different flammability and shade tolerances (as in LPJ-GUESS and aDGVM) is beneficial, and they reinforce the positive grass–fire feedback if included (as in aDGVM).
Having in mind that DGVMs need to be kept as simple as possible, we conclude that the most important mechanisms for better representing the forest–savanna–grassland transition are (i) how water limits tree growth and regulates tree–grass competition and (ii) the grass–fire feedback. Distinguishing between tree life stages and representing the different responses of forest and savanna trees are less important features for the models, although they can considerably ameliorate the representation of the two main mechanisms. As parts of these mechanisms are already included in most DGVMs, extensions should be relatively simple, but they can substantially improve the predictions of vegetation dynamics and carbon balance under future climate change scenarios.
This study was initiated during a TERRABITES workshop in 2010 at INIA-UAH, Madrid. We gratefully acknowledge all of the workshop participants for starting an early discussion on the topic, and V. Gayler and T. Raddatz for kindly preparing the DYNVEG-JSBACH data. We thankfully acknowledge the support of this workshop by TERRABITES COST Action ES0805, which also financed a short-term scientific mission allowing B. Cuesta to begin work on this paper. S. Scheiter acknowledges financial support by the Hessian Landes–Offensive zur Entwicklung Wissenschaftlich–ökonomischer Exzellenz (LOEWE). Edited by: K. Thonicke