CRAFTY-EU covers the European Union 27 (EU member states that include the UK but exclude Croatia) together with Norway and Switzerland. The model operates at a 10′ (arcmin) resolution, with 23 871 grid cells in total. This resolution was selected for its consistency with input data, all of which had the same resolution, for its low computational demands, allowing multiple model runs to be carried out quickly, and because of a shortage of appropriate calibration data at finer resolutions. Nevertheless, this resolution is relatively coarse for an agent-based model application and means that modelled agents cannot be seen as representative of individual real-world land managers in most cases. Instead, they are drawn from semi-aggregated AFTs designed to represent coherent localised land use systems (Letourneau et al., 2012), with management and behavioural characteristics expressed at appropriate generality, as described below.

The agent functional types used in CRAFTY-EU were designed to provide generic coverage of the major sectoral and cross-sectoral land systems at local (10′) scale across Europe. Key distinctions were made between levels of management intensity and between the ranges of ecosystem services produced (Arneth et al., 2014; Letourneau et al., 2012; Murray-Rust et al., 2014; Paul et al., 2017). The final typology was intended to capture the primary form of land management within each grid cell, while allowing for secondary land uses and variation in local land management practices (Table 1; Appendix A). The initial distribution of these AFTs across the modelled land surface was based on the distribution of land use categories modelled by the IAP under baseline conditions (Appendix A), ensuring consistency across initial simulation conditions and comparability with subsequent scenario-based changes.

The abilities of different AFTs to utilise capitals and produce ecosystem services were defined via capital sensitivity and productive ability parameters (summarised in Table 1 with further details and exact parameterisations in Appendix A; see also Murray-Rust et al., 2014). Where possible, values were derived from simulated production data in the IAP and otherwise assumed on the basis of land management intensity and diversity. Behavioural differences between AFTs were also introduced to assess the robustness of model outcomes. These differences, described in Sect. 2.6, affected agents' ecosystem service production levels and their willingness to change land use or abandon land, and they were intended to approximate differences in productive abilities, priorities and sensitivities between and within agent types. Urban land use was not actively modelled but constrained to follow the results of the IAP, which includes advanced modelling of urban development (Terama et al., 2019).

The productive potential of each modelled grid cell was described via five capitals: natural capital (crop productivity, grassland productivity, forest productivity), human capital, social capital, manufactured capital and financial capital. Each capital was derived from the IAP as described in Appendix A. Scenario-specific changes in capital values were produced by running the IAP under each scenario in turn and repeating the derivation process. Each of the productivity capitals accounts for climate-induced changes during the period of simulation, including the effects of changes in temperature, precipitation and CO2 levels. These changes were simulated for the IAP by combinations of global and regional climate models: EC_Earth/RCA4 for RCP2.6 and HADGEM2-ES/RCA4 for RCP4.5 and 8.5 (Harrison et al., 2019). Socio-economic conditions (as defined by the SSPs; Riahi et al., 2017) affected anthropogenic capitals (human, social, manufactured and financial) as determined via a stakeholder-led elaboration of scenario narratives and a subsequent uncertainty-based quantification (Harrison et al., 2019; Pedde et al., 2019b). Because IAP outputs were only available at three time slices (2020s, 2050s and 2080s), capital values were linearly interpolated to give annual values for each grid cell over the period 2010–2086.

The CRAFTY framework is designed to account for the demand and supply of a range of ecosystem services, and we incorporate a representative group for which calibration data are available or for which assumptions related to calibrated land management can be made: timber, meat, crops, carbon sequestration, landscape diversity and recreation. Annual demand levels for each of these services were derived from IAP outputs via the conversion of simulated land cover to service production levels as described in Appendix A. All demand levels are available in Supplement 1.

Demand levels were converted via mathematical functions to “benefit” values for each agent and cell throughout the simulations, taking account of current supply–demand gaps, assumed societal valuations of each service and agent abilities to provide each service on each cell (Appendix A). In this case, linear functions of supply–demand gaps were used to calculate benefit values, and these functions were calibrated to ensure the equal relative valuation of services (i.e. the production of an equal proportion of unmet demand was assigned an equal benefit value whatever the service). This created a balanced competition between agents that was not skewed towards any particular service(s), with no benefit accruing from production when there was no unmet demand, prompting production under shortfalls but not under surpluses.

The resulting benefit values were then used as a basis for competition between agents for the ownership of cells, which occurred on a defined proportion of cells at each time step (Sect. 2.6). Agents producing combinations of services with the highest benefit values were best placed to win this competition but did not necessarily do so because existing agents could choose to persist with their existing land use rather than submit to competition. This process was controlled by “abandonment” and “competition” thresholds describing the ranges of benefit values within which agents would abandon their land, persist with their existing land use or cede their land to another agent (representing an alternative land use) (Holzhauer et al., 2019; Murray-Rust et al., 2014). These thresholds were initially fixed across agent functional types and then varied at typological and individual levels as described below. Therefore, agents did not necessarily optimise their land uses according to benefit values, and these values were not used to ensure full supply of each service. With thresholds set to zero, individual agents would choose a more beneficial land use when made aware of one through the competition process (i.e. when their cell was selected for competition by another agent type) but would not necessarily do so under non-zero thresholds. In neither case was any system-level optimisation involved. Therefore, benefit values, responding to changes in demand and supply levels, stimulated production but did not guarantee a given production level. The model therefore contains no assumptions that override the emergence of suboptimal or non-equilibrium outcomes from scenario conditions. Service production in any part of the EU contributed to satisfying demand levels, representing an assumption of free trade across the modelled area (constrained by the infrastructure and transportation networks described in the manufactured capital values). This is a reasonable assumption given that the EU is a free trade zone. A full description of the valuation and competition process is given in Appendix A.

The CRAFTY modelling framework has been extensively evaluated and applied in previous studies (e.g. Alexander et al., 2017; Blanco et al., 2017a; Brown et al., 2014b, 2018b; Holzhauer et al., 2019; Murray-Rust et al., 2014), as has the IAP upon which this application of CRAFTY is based (e.g. Brown et al., 2014a; Harrison et al., 2016; Holman et al., 2017; Kebede et al., 2015). Both sets of evaluations have included sensitivity and uncertainty analyses (Brown et al., 2014a, b, 2018b; Kebede et al., 2015; Synes et al., 2019), comparisons to empirical data and to the results of other models (Alexander et al., 2017; Blanco et al., 2017a), full descriptions of model design and functioning (Harrison et al., 2015; Murray-Rust et al., 2014), and full free access to the models themselves including interactive online systems for exploring model outputs (Holzhauer et al., 2016; IMPRESSIONS Project, 2018; https://landchange.earth/CRAFTY, last access: 3 December 2019). Both models have also been extensively used in, and informed by, a stakeholder engagement process that has occurred over several years across the EU (Kok et al., 2019).

Here, additional model evaluation focused on the behaviour, stability and interpretability of the European application of CRAFTY. These characteristics were primarily assessed through two sets of runs under static baseline conditions, starting from an unassigned (empty) land use map and from the baseline land use map derived from the IAP. The purpose of these two exercises was, respectively, (1) to check whether baseline conditions would generate a “realistic” land use configuration purely on the basis of capital levels and AFT characteristics (i.e. in the absence of any spatial information about land management) and (2) to check for divergence in outcomes from a common starting point consistent with the starting point of other scenario runs. Model dynamics were checked visually and statistically using the numbers of agents within each AFT and levels of service provision. Both evaluation exercises are described in detail in Appendix B.

CRAFTY-EU runs on annual time steps at which a proportion of cells are subject to potential abandonment, adoption or competition (Murray-Rust et al., 2014). In the first evaluation exercise, the model was run over 800 time steps, with 20 % of cells being randomly selected for potential change (i.e. the maximum number of cells that could change at each time step if required by the competition process). This arbitrary but high rate of competition allowed for rapid changes to the simulated land system, ensuring model dynamics could be clearly perceived. The period required for the model to reach a steady state was identified, and 10 further independent simulations were then run to this point using different random number generator seed values. The second evaluation exercise was performed over 100 time steps, again with a 20 % rate of cell selection. This exercise was designed to run a sufficient number of replicates to identify and understand any divergence from stationarity in model dynamics.

Following the evaluation exercises, simulations were run for 71 time steps, representing the period 2016–2086, with 5 % of cells selected for potential change at each of these time steps. As an upper limit, this rate is up to an order of magnitude greater than observed (e.g. Loveland et al., 2012) and projected land use changes (e.g. Schmitz et al., 2014), allowing for the majority of potential changes to be rejected while maintaining scope for rapid land use change under extreme scenarios. These simulations all began from the baseline land use map (Fig. A1.2) and proceeded according to scenario conditions in terms of ecosystem service demand levels and capital values (Supplement 1). Seven distinct scenarios were simulated, each of which comprised a combination of Representative Concentration Pathways (RCPs) and Shared Socioeconomic Pathways (SSPs) (O'Neill et al., 2017) as described in Table 2. Socio-economic scenarios were developed from the SSPs through a stakeholder engagement process described in detail in Kok et al. (2019). These scenarios were first run through the IAP as described above in order to produce representative levels of capitals and demands for use in CRAFTY-EU. Throughout each simulation, land use maps, numbers of agents in each AFT, ecosystem service production levels and fragmentation indices (fractal dimensions) were recorded. Fractal dimension (Mandelbrot, 1983) was used to capture the patch sizes of each land use, with higher values indicating more complex patterns, and was compared on the basis of relative change from the baseline year using the R package SDMTools (Vanderwal et al., 2014).

For each scenario, five distinct parameter sets were applied to assess the effects of variations in modelled agent behaviours (full parameterisations for each of these are given in Appendix A and a description of each behavioural parameter in Table 3). These parameter sets differed in terms of the abilities of agents to produce services and their tolerance of low benefit values and of competition, all of which varied at AFT and individual agent levels. These variations were designed to represent general behavioural effects arising from land manager decision-making, accounting for aggregation to the model's spatial resolution, but with a focus on sensitivity analysis across a large range rather than calibration at levels thought to be realistic. Parameter set 1 represented a behavioural “baseline” in which agents responded directly to benefit values with no additional individual or typological behaviour (i.e. the land use with the greatest benefit value was adopted in each case). Parameter set 2 (“increased thresholds”) introduced differences in abandonment and competition thresholds to induce the abandonment of land when benefit values fell below the abandonment threshold and persistence of an existing land use unless an alternative had an additional benefit value of at least the competition threshold. Intensive land use agents were parameterised to be less tolerant of low benefit values and more willing to switch to a land use with higher benefit values. Parameter set 3 (“individual variation”) introduced differences between individual agents in terms of their levels of production of different ecosystem services as well as their abandonment and competition thresholds. In this case, individual values were randomly assigned from a Gaussian distribution centred on the mean value for each type (Appendix A). Parameter sets 4 and 5 replicated sets 2 and 3, respectively, but with higher values for thresholds and variations (up to 10 times higher, with thresholds of the same order of magnitude as mean benefit values).

Scenario simulations showed considerable divergence between land systems in the mid-2080s under different scenario combinations, and this was not substantially altered by behavioural variations between agents (Table 2, Supplement 2). These differences were primarily driven by socio-economic scenario conditions, but also by different levels of climate change between the three climate scenarios used (Figs. 1 and 2). Broadly, where socio-economic capital levels were maintained or increased, the land system diverged from the baseline scenario by a relatively limited amount, with widespread intensive management of land and small shortfalls or surpluses of most modelled services. Conversely, where these capitals declined substantially, widespread extensification and abandonment of land occurred, and large shortfalls in service levels developed (Fig. 1, Table 2, Supplement 2). These dynamics were partly ameliorated by increases in productivity in some areas associated with high-end climate change, especially north-western Europe.

Of particular importance were manufactured and financial capitals, which increase greatly (up to 250 %) in some scenarios (e.g. SSP1) and decrease (by around 90 %) in others (e.g. SSP3) depending on scenario storylines (Fig. 1 and Table 2). These capitals are crucial in supporting intensive land management in CRAFTY-EU (Appendix A) and so determine the scope for the most productive uses of land. Where these capitals increased, surpluses of services (especially food) developed, and where they decreased, shortfalls developed, reaching 56 % of food demand in the RCP4.5–SSP3 scenario combination (Fig. 2). These capital dynamics are also linked to reduced per capita food consumption in SSP3 (Table 2), suggesting that shortfalls in this context translate more directly into hunger than they do in other scenarios in which overconsumption provides a buffer to any health impacts of underproduction.

We simulated three socio-economic scenarios in different climate scenarios and all showed notable similarities between climates. SSP1 had the clearest differences; this scenario has high demands for all services, and the difference between climate scenarios was due to increases in average crop and forest productivity capitals under RCP4.5 relative to RCP2.6. These productivity changes increased the competitiveness of intensive management enough to allow it to outcompete more extensive, multifunctional land uses and so allowed production to meet 15 %–30 % more of the demand than under RCP2.6.

The most consistent and most negative scenario was SSP3, in which economic and social challenges led to disintegration of the land system across much of Europe, with large areas being abandoned, managed extensively or fluctuating over time (Fig. 1, Table 2, Supplement 2). These dynamics were particularly pronounced in more fertile areas of Europe, where currently dominant intensive management declined dramatically during the first half of the century. As a result, similar shortfalls of almost 50 % of food demand were found between 2050 and 2080 in both RCP4.5 and RCP8.5, suggesting that changes in climate were minor in comparison to the almost complete loss of financial and manufactured capitals that undermines the productive use of land in SSP3. Nevertheless, supply levels increased markedly towards the end of the century in RCP8.5, as increased natural capitals (i.e. yield increases) offset some of the losses from declining socio-economic capitals. Conversely, in technologically advanced scenarios (e.g. SSP4), in which manufactured and financial capitals increase greatly, demands for services could be met relatively easily during most of the century, leading to a decline in intensive management because of a lack of need rather than a lack of opportunity.

Results also show some broad geographical patterns. While the most unproductive areas of Europe (e.g. mountain ranges, high latitudes) were the most resistant to change under any scenario, other areas responded differently depending on the scenario conditions. South-eastern Europe (Greece, Bulgaria, Romania and Hungary) was slightly more vulnerable to extensification and abandonment, whereby supply levels matched demands as in SSP4 (i.e. when the benefit or profit levels were low), but were more robust to low levels of capitals in SSP3. In contrast, western Europe (particularly Germany, France, England and intensively managed areas of Spain) suffered widespread abandonment in SSP3. As climate change increased in magnitude through RCP2.6, 4.5 and 8.5, land management in north-eastern Europe (Poland, Lithuania, Latvia, Estonia and southern Finland) tended towards forestry, as increases in forest productivity and decreases in crop productivity made arable agriculture less competitive.

Behavioural parameter variations had distinct effects in different scenarios and on different metrics (Table 2, Fig. 2 and Appendix C). In general, land use patterns in scenarios with less intensive management (and also lower land use fragmentation as measured by the fractal dimension) were less affected by behavioural parameter changes; these scenarios included RCP4.5–SSP3, RCP8.5–SSP3 and RCP4.5–SSP4. Conversely, scenarios with more intensive management (RCP4.5–SSP1, RCP2.6–SSP1 and RCP2.6–SSP4) were more affected, producing more fragmented land systems. These differences were not necessarily reflected in ecosystem service supply levels, however (Table 2, Fig. 2). Instead, these showed increasing effects of parameter variations over the simulation period, with the largest effects occurring in SSP3 and 4 and the smallest in SSP1. In two scenarios (RCP2.6–SSP4 and RCP4.5–SSP3), behavioural parameterisations determined whether food was oversupplied or undersupplied by the 2080s, with increasing variation between agents leading to increased food supply. These differences were also partially correlated with climatic scenarios, with land systems that were more productive under high-end scenarios proving more robust to behavioural differences (particularly RCP4.5–SSP4, RCP4.5–SSP1 and RCP8.5–SSP5, in which variations had almost no impact on food supply). Of the two forms of variation simulated, increased requirements for benefit from land management (thresholds) led to increased fragmentation within scenarios on average but also increased differences in fragmentation between them. Individual variation increased the differences in fragmentation between scenarios more than it did average levels across them, but at the higher strength these differences were reduced to approximately baseline levels, suggesting a “peak effect” under small levels of individual variation beyond which the effects of behavioural parameters were reduced (Fig. C1). Plots of the evolution in behavioural parameter values during the simulations (Fig. C2) show that this effect was partially due to a context-specific “selection pressure” towards particular parameter values, especially in the SSP3 simulations. In this context, the initially random distribution of agent parameters became skewed towards higher values of competition thresholds, lower values of abandonment thresholds and lower variation between agents, demonstrating a disproportionate persistence of agents who were relatively unlikely to respond to benefit values. We did not observe the converse persistence of more “optimising” agents in any scenario.

In all cases, the delicate balance between food and timber production highlights the sensitivity of results to demand levels for ecosystem services derived from agriculture and from forestry. In many cases, simulations resulted in the widespread adoption of multifunctional land uses that provide both sets of services to some extent, with the locations of these being scenario dependent. The levels of demand, relative valuation and production of these services therefore appear to be major determinants of the nature of European land systems in this model.

CRAFTY-EU is an explorative model and is not designed to predict (inherently unpredictable) land system changes (Brown et al., 2016a). Further, the CRAFTY framework is intended to provide relatively simple, generic methods for exploring land manager decision-making over large geographical extents (Murray-Rust et al., 2014) and is used here to represent decision-making within local land systems rather than at the level of individual managers. As such, this model application is a first step towards an improved understanding of behavioural processes within large-scale land systems.

At a general level, the results presented here are realisations of a single approach to land systems modelling, which complement alternative projections made by other models (e.g. Harrison et al., 2019; Stürck et al., 2018; Verkerk et al., 2018). In particular, conceptual or theoretical frameworks within which behavioural modelling can occur are diverse and disputed, and a universally applicable representation of the complex social processes involved in land use change is not available or even necessarily possible (Brown et al., 2016a; Huber et al., 2018; Meyfroidt et al., 2018). Even given this caveat, our exploration of behavioural parameters is illustrative rather than exhaustive, intended to reveal the implications of basic assumptions more than exact parameter values. We do not attempt to derive realistic parameter values but instead assess model sensitivity across a large range, from identical agents attempting to maximise their benefit, to highly diverse agents tolerating benefit differences of around 100 % of mean values (Table A4). In this respect, we deliberately build on earlier studies of behavioural processes in CRAFTY, including in a similar scenario context (e.g. Brown et al., 2014b, 2018b). While the CRAFTY framework is specifically designed to allow for the exploration of abstracted behaviours that do not require precise parameterisation, linking these values to empirical data is clearly desirable even where apparent effects on model outcomes are small.

In this case, more precise parameterisation might necessitate finer spatial and behavioural resolution, as is typical of many agent-based models. Nevertheless, conceptual and practical limits to behavioural parameterisation exist, and the aggregation of agents by location, land use and behaviour (or combinations such as functional type) is an established approach to large-scale modelling (Arneth et al., 2014; Valbuena et al., 2010). Identifying an appropriate balance between scale and resolution in the agent-based modelling of the European land system certainly requires further investigation of behavioural processes and data across scales, but the fact that we find clear systemic impacts of behaviours simulated at 10 arcmin resolution implies that finer resolutions may well be informative, as they have proved in previous applications of the CRAFTY framework (e.g. Blanco et al., 2017). At present, data availability is the greatest barrier to the adoption of these resolutions in European modelling.

A number of more specific considerations are also important for interpreting our findings. Most significantly, the simulations presented here form an experiment into the effects of simulating land management as the provision of multiple (but arbitrarily limited) ecosystem services, which depend upon a set of scenario-dependent capitals and which are valued equally per standardised unit of demand. This design ensures that trade-offs between services are clear but does not assume the preferential production of some services (such as food) when supply levels are equally insufficient. As a result, scenarios in which shortfalls in service provision exist might represent an artificially balanced outcome, with real-world equivalents potentially diverging towards more homogeneous land uses. In this respect, our findings suggest that the further exploration of trade-offs between service provisions, in terms of both production systems and valuation, should be a priority for land system modelling. This is especially important given potential changes in current valuation practices, for example through carbon pricing or payments for ecosystem services, which could transform the competitiveness of currently minor land uses and require models to account for the services that they produce (Kay et al., 2019).

Beyond Europe, neither CRAFTY-EU nor the IAP that is used to calibrate it explicitly represent production and trade. While scenario-specific import levels are assumed, these are likely to be overestimates in challenging scenarios with large shortfalls in service provision that imply shortages elsewhere in the world (Dellink et al., 2017; Harrison et al., 2015; Stevanović et al., 2016). This would reduce supply levels even more than simulated here. Furthermore, alternative treatments of international trade based on assumptions of economic equilibrium would be inconsistent with the supra-economic behavioural approach used in CRAFTY-EU (Arthur, 2006). The relative provision of different services is also subject to substantial uncertainty in our representation of forest growth, with assumed adaptation to changes in species suitability likely to overestimate real-world adaptation (Schelhaas et al., 2015), as the CRAFTY framework has previously been used to demonstrate (Blanco et al., 2017a, b).

Notwithstanding the above limits on the model's accuracy, the robust, cross-sectoral nature of the model, building on the established and evaluated IAP and CRAFTY framework, means that it is capable of providing well-founded and novel insight into land system dynamics. Model evaluation performed for this and earlier studies has revealed no clear biases or instabilities, with CRAFTY-EU producing realistic outcomes in the absence of information about baseline land uses. The responses of the model to the scenarios can be seen as coherent responses to a set of land system drivers that are fully interpretable in light of transparent model assumptions.

A key finding of this work is that the sensitivity of land use to social (as well as economic and climatic) conditions makes land systems vulnerable to breakdown when these conditions worsen substantially. Such worsening is a key characteristic of some future scenarios (e.g. SSP3) but one that has generally only been explored through qualitative scenario descriptions (e.g. Cradock-Henry et al., 2018; Kebede et al., 2018; Pedde et al., 2019a). In SSP3, declines in socio-economic capitals are so precipitate and substantial that the resulting breakdown of the simulated land system is highly plausible and proves almost impossible to avoid in our modelling, regardless of specific parameterisations. In other model projections of this scenario, similar outcomes are avoided only by very large increases in food prices that compensate for relatively low crop yields and stimulate food production at the expense of forest cover (Doelman et al., 2018; Fujimori et al., 2017; Harrison et al., 2019; Hasegawa et al., 2015; Popp et al., 2017). These large increases are generally enabled by an assumption of economic equilibrium and not constrained by financial or other productive limitations related to the scenario.

This implies that scenario modelling using economic equilibrium assumptions could prove misleading, and scenario conditions place potentially non-equilibrium limits on price or production levels. Substantial declines in financial and manufactured capitals, for instance, may effectively preclude the necessary economic stimuli or production responses to meet demand in SSP3. While this problem is starkly illustrated by non-equilibrium modelling such as that presented here, it is not in principle impossible to account for it in models that assume equilibrium dynamics within scenario-specific constraints (Heistermann et al., 2006). However, knock-on effects on consumption, demand and supply (and wider socio-economic systems) are obscured by the predefinition of those factors in scenario storylines. The CRAFTY-EU model therefore makes only one of two crucial connections, linking social conditions and supply levels through capital–production relationships without completing the link back to demand. Most starkly, if insufficient food is produced to maintain population levels, populations and subsequent demand would inevitably decrease – a fundamental feedback that remains absent from scenario modelling.

Such an effect is particularly pertinent to our simulations of SSP3. In the storylines of SSP3 and 4, large populations suffer low and unequally distributed economic growth, leaving per capita demand for agricultural production approximately half that in the wealthier and more equitable contexts of SSP1 and 5 (Table 2; Kok et al., 2019; Riahi et al., 2017). The large and long-lasting shortfalls in food production generated by SSP3 simulations therefore imply far greater challenges than the shortfalls generated by SSP1 simulations because the latter allows adequate (if undesirable) levels of nutrition to be maintained, while the former does not. This is clearly an important area for further research, albeit one that requires much improved treatment of food security and its consequences as well as fully internally consistent scenario simulations that account for a range of economic and social factors.

In addition to identifying very large negative impacts of some scenario combinations, we also find that these impacts differ widely across Europe. Some areas appear to face high likelihoods of substantial changes; for example, we find that eastern Europe is broadly more vulnerable to changes in demand levels (and hence benefit or price levels) and western Europe broadly more vulnerable to changes in capital levels. Many of the worst simulated outcomes have notable mirror images in history, whereby land systems gradually became more intensive, homogeneous and efficient as financial, technological and social capitals developed (e.g. Petit and Lambin, 2002). Projected declines of these capitals produce a return to fragmented, extensive production in our simulations; this is a reverse precedent that adds some credibility to model responses, while clearly not suggesting predictive accuracy. It is also notable that greater climatic change can actually ameliorate the worst outcomes in some cases (e.g. SSP3) because it allows higher yields in parts of Europe to offset losses and socio-economic difficulties elsewhere – at least in the absence of the pests, diseases and extreme weather events that may be associated with such climatic change (Donatelli et al., 2017). Similarly, technologically advanced scenarios (e.g. SSP4) allow for relative ease of production and therefore free up land, leading to some extensification and abandonment.

Within these broad findings, variations in behaviour can have substantial effects. These are generally more pronounced in low-end climate change scenarios that rely on slight competitive advantages of intensive land systems to meet service demand levels and which are therefore sensitive even to slight deviations from simple economic rationality in management decisions. The literature suggests that intensive farmers are more vulnerable to changing price levels (van Vliet et al., 2015), and this vulnerability is amplified here, as it may be in reality, by a reliance on socially mediated capitals that support farming (Sutherland and Burton, 2011). It is also notable that the behavioural effects we observe are similar at both simulated strengths of behaviour, suggesting that even small differences in land manager responses to scenario conditions can have substantial consequences. Indeed, increased behavioural differences do not necessarily persist in the face of challenging scenario conditions, under which some behavioural responses may consistently outperform others. In our SSP3 simulations, such an effect is apparent in the emergence of greater homogeneity and tenacity in the agent population and the loss of more responsive, economically rational agents (Fig. C2). Widespread evidence already exists of deviation from economically optimal behaviour amongst land managers in the real world, as in the selection of economically inferior options for social reasons or socially mediated uptake that spans long time periods (Brown et al., 2018a, 2019a; Sereke et al., 2016). Our findings further show that the effects of such behaviours on supply levels can accumulate over time (Fig. 2) when permitted to through a path-dependent and non-optimising modelling approach. The inclusion of more realistic and heterogeneous behaviours in land system models therefore appears to be a prerequisite for accurate assessments of future scenarios and so for effective land management planning and policy making.