Main

Over the past century, unprecedented changes in both the character and scale of the land used for food production have made agriculture by far the most prominent driver of biodiversity loss1,2. Today, the food sector is also the primary driver of nitrogen pollution, freshwater consumption and soil degradation3,4, while about one-third of all greenhouse gas (GHG) emissions are linked to the food system5. Population growth and the ongoing ‘nutrition transition’ with rising average incomes towards higher intakes of resource-intensive foods, particularly from livestock, could further escalate the environmental pressures of food production6.

Recent assessments have shown that a reduction in the intake of resource-intensive livestock-based foods in favour of a variety of plant-based foods would not only provide substantial health benefits and a reduction in food-related mortality but also largely lower the massive environmental burden of food production7,8. In terms of biodiversity conservation, several studies have also suggested substantial benefits of a shift towards healthier diets, in particular owing to the enormous ‘land-saving effect’ as a result of changing agricultural demand9,10. However, conservation benefits from dietary change depend not only on the quantity of land that is protected from conversion but more importantly on the quality and location of conserved habitats11,12,13. Moreover, dietary changes may address specific conservation aspects, such as reducing the conversion of pristine habitats for feed production14. Yet, other aspects, such as the simplification of landscapes with a long history of land use15 and associated losses in nature’s contributions to people (NCP)2, are more closely connected to broader structural trends in food production, such as specialization, intensification and the concentration of agricultural land16,17. Hence, conservation benefits from dietary change may not only vary across space but may also differ across conservation perspectives, depending on whether the focus is on the intrinsic values of nature, such as preserving intact habitats, or its instrumental values, such as the supply of important NCP in managed landscapes. Consequently, some of the principles that apply to other environmental outcomes of food production, such as an approximately linear scaling of CH4 and N2O emissions with livestock production levels18, may not necessarily extend to conservation outcomes in relation to shifts in food demand.

The aim of this study was to disentangle the impacts of dietary change across different conservation value perspectives, as illustrated in the Intergovernmental Science-Policy Platform on Biodiversity and Ecosystem Services (IPBES) Nature Futures Framework19, in particular the ‘Nature for Nature’ (intrinsic) and ‘Nature for Society’ (instrumental) perspectives. We therefore applied a quantitative and spatially explicit modelling framework to assess how a shift towards healthier diets could affect indicators across these conservation perspectives at a spatial scale and detail of habitat representation that goes well beyond previous efforts20,21 (Extended Data Table 1 and Extended Data Fig. 1). Because a shift towards healthier diets on a globally relevant scale is a complex and multifaceted challenge that is subject to various social, political and cultural barriers22, we also assessed different levels of convergence towards healthier diets and compared the results with conservation actions in line with the targets of the Kunming-Montreal Global Biodiversity Framework (GBF)23. Our analysis differs from earlier studies that did not distinguish between different conservation value perspectives14,21,24, only used simple aggregated biodiversity indicators8,9,25,26, assessed the substitution of livestock food without considering further adjustments in food consumption20,26 or lacked the integration of other environmental and socio-economic outcomes21, and thus represents an important addition to the debate on diet-related benefits for nature.

The Model of Agricultural Production and its Impact on the Environment (MAgPIE, v4.8.2), a forward-looking open-source land and food-system model, formed the core of our modelling framework. MAgPIE is continually validated against observed data27 and has frequently been used to study questions related to the sustainability of the food system18,24. It uses a wide range of socio-economic and spatially explicit biophysical information to simulate global land-system dynamics over the twenty-first century (see Methods). To map projected land-cover changes at a scale relevant for assessing fine-scale landscape and habitat change, we coupled MAgPIE with the Spatial Economic Allocation Landscape Simulator (SEALS). SEALS uses empirically calibrated adjacency relationships, physical suitability and conversion eligibility information to allocate projected land-cover changes at a spatial resolution of 10 arcsec (300 m × 300 m at the equator)28,29.

Future scenarios of dietary change and conservation action

In our scenario set, we compared the effects of a partial and complete shift towards healthier diets with a bundle of conservation actions across the ‘Nature for Nature’ and ‘Nature for Society’ perspectives. Our reference scenario (SSP2-REF) follows the ‘middle-of-the-road’ shared socio-economic pathway (SSP2) for the land-use sector30 with moderate population and income growth. In our reference scenario, we also assumed full compliance with currently implemented land policies, such as protected areas (PAs). In our alternative scenarios, we added a 50% (DIET50) and 100% (DIET100) convergence to a healthy reference diet by 2050 as outlined by the EAT-Lancet Commission8, a set of targeted conservation measures to support biosphere integrity (BIOS) and lastly a combination of a partial diet shift and targeted conservation action (DIET50-BIOS) to the reference case.

The healthy reference diet is based on minimum daily per capita intakes of legumes, fruits, vegetables and nuts and maximum intakes of sugar, oils, poultry, eggs, red meat and dairy products, complemented by a convergence to a healthy body mass index (that is, 20–25)7 (Fig. 1a). Within these limits, country-specific dietary patterns can still exhibit considerable diversity. Globally, the assumed partial (DIET50) and full (DIET100) convergence towards the healthy reference diet is associated with a reduction in average per capita calorie intake from livestock products of 33% and 66%, respectively, by 2050 (Fig. 1a). The reduction in livestock-based food intake is compensated by an increase in plant-based food intake, in particular from legumes (+173 and +346% kcal person−1 day−1) and fruits, vegetables and nuts (+57 and +114% kcal person−1 day−1).

Fig. 1: Global food consumption and agricultural demand in our dietary change scenarios.
figure 1

a, Per capita calorie intake per day across different food groups and for different levels of convergence towards recommended intake levels. In the reference scenario, calorie intake is derived from projected population growth, demographic changes and per capita income, while recommended intake levels in the two dietary change scenarios for the different food groups are based on Springmann et al.7 (Methods). b, Overall changes in agricultural demand expressed in gigatonnes of dry matter (GtDM) per year for different levels of convergence towards the healthy reference diet across different demand categories. Food crops mediated by biotic pollination include oil crops such as sunflower or rapeseed, legumes, and fruits, vegetables and nuts.

Full size image

In the BIOS scenario, we assessed land conservation measures in line with the GBF targets. These measures expand PAs to 30% of the global land surface and address key drivers of biodiversity loss outside of PAs. The assumed expansion of PAs includes unprotected land within Key Biodiversity Areas (KBAs)31, intact habitat in Biodiversity Hotspots (BHs) and ecoregions with a high β-diversity (EBDs)32, and critical connectivity areas (CCAs) as defined by Brennan et al.13. Conservation measures outside PAs include maintaining at least 20% of semi-natural habitats in agricultural landscapes33,34 by 2030 and a no-net-biodiversity-loss scheme after 2030 whereby any decrease in biodiversity intactness24 must be compensated at the biome level35 (Methods).

More effective habitat conservation through targeted action

We estimated changes in the area of habitat (AOH) of vertebrate species based on a fine-scale allocation of land-cover projections and the corresponding changes in the suitable habitat within the range of each assessed species (Extended Data Table 1). AOH projections provide valuable information on potential habitat change and species extinction risk. They can also guide conservation efforts and have been proposed as an additional indicator for the International Union for Conservation of Nature (IUCN) Red List36,37. Although the number of species assessed (27,726) represents only around 2.2% of recorded species and 0.3% of all predicted species38, the use of a large set of surrogate species across different taxa with varying traits and habitat preferences has shown to be effective in reproducing the spatial variation in biodiversity at national, regional or global scales11,39. In terms of conservation values, the AOH indicator is mostly associated with the ‘Nature for Nature’ value perspective.

In our reference scenario (SSP2-REF), we found that 17,359 (63%) of the assessed species could experience a combined global habitat loss of 22.5 Gha (Fig. 2). Of these species, 1,392 and 607 species could lose more than 20% and 40%, respectively, of their habitat by 2050, of which 60% are currently listed only as ‘least concern’ or ‘near threatened’ according to the IUCN Red List (Fig. 3). Much of this habitat loss is driven by the expansion of agricultural land in tropical Africa, Asia and Latin America (Supplementary Fig. 16) as a result of population growth and increasing per capita food supply, in particular of meat and dairy products owing to rising incomes (Extended Data Figs. 2–4). While in Latin America (LAM) and Asian countries excluding China, India and Japan (OAS) more than 80% of the habitat losses in SSP2-REF are due to further cropland expansion, nearly 40% of habitat losses in sub-Saharan Africa (SSA) are driven by pasture expansion (Fig. 2). In OAS, forest loss resulting from cropland and pasture expansion is particularly high, which especially affects forest-dwelling birds (Extended Data Figs. 5 and 6). In LAM and SSA, the loss of open ecosystems also significantly drives habitat loss in SSP2-REF, partly due to the relatively high baseline protection of forest compared with open ecosystems such as the Cerrado in Brazil (Supplementary Fig. 2). We also found that in Europe (EUR and NEU), China (CHA), as well as in Canada, Australia and New Zealand (CAZ), forest establishment in open ecosystems, for example due to land abandonment and rising timber demand, is a critical driver of habitat loss, further increasing the pressure on species adapted to open habitats that have already experienced strong population declines over recent decades40.

Fig. 2: AOH losses at global and regional levels between 2020 and 2050.
figure 2

a,b, Total sum of AOH losses at the global (a) and regional (b) levels between 2020 and 2050 caused by the expansion of aggregated land-use/land-cover (LULC) types across all assessed species.The sum of AOH losses is much higher than overall projected global and regional land-use changes. This is because one unit of expansion of each LULC type can cause loss of AOH for several species. Hence, the measure also indicates the relative importance of LULC expansions for AOH losses by putting a higher weight on areas that affect more species. The legend in a also applies to b. GLO, global; CAZ, Canada, Australia and New Zealand; CHA, China; EUR, European Union; IND, India; JPN, Japan; LAM, Latin America; MEA, Middle East and Northern Africa; NEU, non-EU European countries; OAS, other Asia; REF, reforming economies of the former Soviet Union; SSA, sub-Saharan Africa Note the different x-axis ranges across the regional plots.

Full size image
Fig. 3: Species with more than 20% estimated AOH loss by 2050.
figure 3

Number of species with more than 20% estimated AOH loss by 2050 across modelled scenarios and different groupings. ‘Species of concern’ includes species that are currently ‘vulnerable’, ‘endangered’ and ‘critically endangered’ according to the IUCN Red List. ‘Lower-risk species’ includes species that are currently listed as ‘least concern’ or ‘near threatened’.

Full size image

In our alternative scenarios, we found that compared with the SSP2-REF scenario, combined habitat loss is roughly halved in both the SSP2-DIET50 (−46%) and SSP2-DIET100 (−52%) cases (Fig. 2). This suggests only marginal benefits in terms of habitat loss are gained from full compared with partial convergence to the healthy reference diet, even though the reduction in global agricultural demand is almost twice as high in the SSP2-DIET100 scenario compared with in the SSP2-DIET50 scenario (Fig. 1). Similarly, the global number of species with a habitat loss ≥20% is also reduced by 60% across both diet change scenarios with almost no additional benefits from a complete transition to healthier diets (Fig. 3). In addition, and contrary to the global trend, we found that in SSA, the number of species with a habitat loss ≥20% and ≥40% actually increases in the SSP2-DIET100 scenario compared with in SSP2-REF (Extended Data Fig. 6) because land savings due to lower feed demand are offset by a shift from staple foods to other plant foods with a broader nutrient spectrum but a lower productivity per hectare, such as legumes or fruits, vegetables and nuts (Extended Data Fig. 4). By contrast, in LAM and OAS, the relative reductions in habitat loss (−84% and −78%, respectively) from even a partial dietary change compared with SSP2-REF are much higher than those observed at the global scale, suggesting disproportionate regional benefits in terms of habitat conservation resulting from a shift to more plant-based diets (Fig. 2).

However, our results also indicate that dietary change is not per se a prerequisite for habitat conservation as land conservation in line with the GBF exceeds the benefits of dietary change and drastically reduces habitat loss from food production. In particular, we found that the combined habitat loss in the SSP2-BIOS scenario is 77% and 51% lower compared with the SSP2-REF and SSP2-DIET100 scenarios, respectively (Fig. 2), even though agricultural demand is the same as in SSP2-REF. This also pertains to the number of species with ≥20% habitat loss, which is also 76% and 35% lower than in the SSP2-REF and SSP2-DIET100 scenarios, respectively. In SSA, we found that the conservation actions in our SSP2-BIOS scenario would reduce the number of species with ≥20% habitat loss by 70% and 71% compared with SSP2-REF and SSP2-DIET100, respectively (Extended Data Fig. 6). This indicates that habitat losses can be mitigated by stringent spatial planning that prevents land expansion in ecologically sensitive areas, even as agricultural demand increases. Despite a larger cropland area, especially for non-staple crops, we found no significant increase in the number of species with high habitat losses in the SSP2-DIET50-BIOS scenario compared with SSP2-BIOS (Extended Data Figs. 4 and 6), indicating that targeted conservation action better distributes impacts across space (Extended Data Fig. 7). Overall, these findings also hold against varying assumptions regarding socio-economic trends along the SSP1 (Sustainability) and SSP3 (Regional Rivalry) trajectories30, although we found that, especially in SSA, cropland expansion in response to biosphere protection in the SSP1-BIOS scenario is larger because of greater trade liberalization (Supplementary Figs. 5–15).

Mixed outcomes of dietary shifts on NCP supply

Our analysis also includes indicators for the supply of key regulating NCP in cultured landscapes and climate regulation, which address the ‘Nature for Society’ conservation value perspective (Extended Data Table 1). These include changes in landscape pollination sufficiency scores, which are an indicator for wild pollination supply, and configurational landscape heterogeneity. Configurational landscape heterogeneity has also shown to be positively associated with a range of other key regulating NCP, such as pest control, and is an important driver of biodiversity in managed landscapes41,42. We also used the Global Soil Erosion Modelling (GloSEM) platform to estimate changes in soil loss by water erosion, which is a dominant driver of degradation of soil-related NCP at the global scale43,44. Finally, we also report net CO2 uptake and losses from land-use change for climate regulation and show how these affect total GHG emissions in the land sector (Extended Data Table 1).

Our data indicate that in 2020, the majority of cropland (970.6 Mha) was situated in landscapes with low configurational heterogeneity and pollination sufficiency scores. By 2050, declining landscape heterogeneity in SSP2-REF further increases the total cropland area with insufficient levels of wild pollination (Fig. 4b), particularly in SSA (+44.8 Mha), LAM (+11.4 Mha) and OAS (+34.6 Mha). This is notable because in SSP2-REF, the demand for pollinator-dependent food crops with a high nutritional value, such as fruits, vegetables, nuts, legumes and oil crops, also increases strongly in SSA (+167%), the Middle East and North Africa (MEA; +68%), OAS (+52%) and India (IND; +43%; Supplementary Fig. 17). In our diet change scenarios, the demand for pollinator-dependent food crops increases even further to 343% (DIET50) and 508% (DIET100) in SSA and 158% (DIET50) and 255% (DIET100) in OAS compared with today. Yet, despite some global improvements (Fig. 4b), we still found substantial increases in cropland area with insufficient levels of wild pollination, particularly in SSA (DIET50: +41.1 Mha; DIET100: +43.5 Mha) and to some extent in OAS (DIET50: +9.8 Mha; DIET100: +12.7 Mha; Extended Data Fig. 8). Maintaining at least 20% of semi-natural habitats at the landscape scale in SSP2-BIOS and SSP2-DIET50-BIOS, however, can address the structural trends that lead to cropland concentration, landscape simplification and reduced NCP supply (Fig. 4b and Extended Data Fig. 8).

Fig. 4: Landscape pollination sufficiency and soil loss by water erosion.
figure 4

a,c, Spatial (left) and latitudinal (right) estimates of landscape pollination sufficiency scores in total cropland (a) and soil loss by water erosion (c) for the reference year 2020. Spatial and latitudinal estimates for the year 2020 are directly derived from European Space Agency’s Climate Change Initiative land-cover maps. Soil loss estimates are based on the GloSEM platform (Methods). Country polygons are adapted from ref. 100. b,d, Projected changes in landscape pollination sufficiency (b) and soil loss by water erosion (d) between 2020 and 2050 based on MAgPIE–SEALS land-cover projections. Continuous pollination sufficiency scores have been simplified into two categories of ‘insufficient’ (<1) and ‘sufficient’ (≥1) values. Basemaps in a and c adapted from World Bank Official Boundaries under a Creative Commons license CC BY 4.0.

Full size image

In terms of soil loss by water erosion, we found that dietary change can effectively mitigate the high losses observed in SSP2-REF, mainly due to the smaller agricultural area in OAS and LAM, but also due to a higher share of perennial crops (Fig. 4d and Extended Data Fig. 9). However, again, we found only limited additional benefits of a complete dietary shift compared with a partial shift. By contrast, in SSA, we found no co-benefits of dietary change for soil loss by water erosion. However, as a result of production shifts to areas less prone to erosion, land conservation measures in the SSP2-BIOS and SSP2-DIET50-BIOS scenarios show significant co-benefits for soil loss in SSA, with soil loss lower by 1.1 Pg (BIOS) and 0.7 Pg (DIET50-BIOS) in 2050 compared with 2020, despite a larger cropland area.

In line with previous work7,45, we also found significant co-benefits from dietary change both in terms of CO2 emissions from land-use change and total GHG emissions from agriculture, forestry and other land use (AFOLU). Our results show that by 2050, a partial dietary shift (DIET50) can achieve more than three-quarters and almost two-thirds of the reductions in cumulative CO2 and GHG emissions, respectively, of a full dietary shift (DIET100) compared with SSP2-REF (Fig. 5a,b). Furthermore, land conservation in our SSP2-BIOS scenario significantly increases carbon uptake on land (Fig. 5a and Supplementary Fig. 23). Reductions in total cumulative GHG emissions from land conservation are therefore similar to those achieved in SSP2-DIET50. The combination of GBF-aligned conservation with a partial dietary shift (DIET50-BIOS) even results in slightly lower cumulative GHG emissions than the SSP2-DIET100 scenario.

Fig. 5: Global projections of GHG emissions from AFOLU, land productivity and food accessibility.
figure 5

a–e, Outcomes projected by MAgPIE for cumulative CO2 emissions from land-use (LU) change (a), overall cumulative GHG emissions from AFOLU (b), expenditure on agricultural products for food use (c), average annual crop yield changes (d) and agricultural imports (e). Regional outcomes and sources of validation data (circles) are shown in Supplementary Figs. 21–30. GtCO2, gigatonnes CO2; GtCO2e, gigatonnes CO2-equivalent.

Full size image

Furthermore, a decomposition of the contributions of the conservation measures in the SSP2-BIOS scenario shows that a simultaneous implementation of measures that address different conservation objectives is crucial to mitigate potential trade-offs between interventions. The no-net-loss scheme, however, caused notable benefits across all conservation value perspectives (Supplementary Figs. 31–35).

Partial dietary shift can lower barriers to conservation

We further complemented our analysis with additional indicators that provide an integrated perspective on changes in land productivity and food accessibility across our scenarios (Extended Data Table 1).

Our results show that targeted land conservation measures most consistently improve outcomes across both the ‘Nature for Nature’ and ‘Nature for Society’ conservation perspectives. However, we also found that compared with SSP2-REF, land conservation measures would incur notable additional socio-economic costs in terms of higher food expenditures (+13%), additional requirements for yield increases (+22.5% on average) and higher total agricultural imports (+3.5% globally between 2020 and 2050) if the demand trajectory remains unchanged (Fig. 5c–e). These higher costs would pose significant barriers to conservation action, particularly in LAM and many countries in OAS and SSA (Supplementary Figs. 26, 27 and 30). Yet, we crucially found that a partial convergence to healthier diets (SSP2-DIET50-BIOS) could largely offset these added barriers to conservation and would eventually result in lower food prices, required average yield increases and total imports compared with SSP2-REF.

Discussion

Our study provides several key learnings. First, our findings indicate that, across both the ‘Nature for Nature’ and ‘Nature for Society’ value perspectives, the added conservation benefits of a full dietary transition are modest compared with only a partial shift. This is because in regions of high global importance for conservation, such as Africa or Asia, converging towards healthier diets would mainly involve reducing the calorie supply of staple foods in favour of more diverse plant-based foods with a lower per hectare productivity rather than replacing resource-intensive livestock products as in most developed regions, and thus the marginal benefits of a complete dietary transition are limited. Secondly, GBF-aligned conservation action shows the most consistent conservation benefits, underlining the critical role of targeted land-use policies in decoupling conservation outcomes from food demand. Lastly, our results show that every step towards more plant-based diets would not only create strong co-benefits for GHG emissions but would also significantly reduce important socio-economic barriers to more effective conservation, such as higher food expenditure, agricultural intensification or increased imports. In summary, we have shown that conservation across different value perspectives critically depends on proactive conservation action regardless of dietary patterns, but that the costs of conservation are significantly lower under a shift towards healthier diets.

Our analysis is an important addition to previous studies and provides further nuance to the debate on diet-related benefits across different conservation perspectives. In particular, the non-linear dynamics in our model results are not captured by static life cycle assessment approaches10, trade analyses with input–output tables9 or the use of biodiversity metrics that assign stylized biodiversity values to land-cover types25,26. Moreover, the high spatial granularity and dynamic representation of underlying socio-economic drivers, such as population and income growth or agricultural trade, were critical for understanding the consequences of dietary change across different regional contexts. In addition, our projections have shown to be robust to variations in key socio-economic assumptions, such as population and income growth, under the SSP1 and SSP3 scenarios (Supplementary Figs. 5–15).

While our study provides insights into the impacts of various land-system interventions on conservation outcomes, evaluating policy instruments to achieve the GBF targets or that support a transition to healthier diets was beyond the scope of the analysis. Achieving the GBF targets requires embedding nature conservation into cross-sector sustainable development programmes46 to ensure reliable funding for implementation and monitoring47, as well as addressing the distributional consequences of conservation actions48. The strong co-benefits of land conservation for both climate change mitigation and adaptation specifically imply an integration of policy approaches, such as redirecting carbon pricing revenues to ecosystem management programmes that prioritize the avoidance of habitat loss over offsets49,50. Although conservation action produces considerable benefits for society at large29, local communities often face high costs that necessitate adequate compensation for foregone opportunities51. Furthermore, this study shows that some of the added costs of conservation can be avoided by a shift to healthier diets. Notable examples of strategies to alter dietary patterns include price-based interventions, information-based approaches and food environment policies52. Policy approaches that emphasize both human health and environmental benefits are likely to be the most effective53. This also includes improving economic access to food by addressing inequality46, as 2.8 billion people, particularly in Africa and Asia, are currently unable to afford a healthy diet54.

A critical component in our model framework is the link between modelled land-cover changes and the associations of species with habitat and elevation55. The MAgPIE model is systematically validated against historical data at the regional and global scale using a regularly updated validation database for most model outputs27. The SEALS downscaling model, in turn, is trained by using a loss function that evaluates modelled outputs against historic land-use patterns at high resolution28,29. The habitat–land-cover model55 used here systematically reduces commission (false presence) errors and has shown to achieve high accuracies in AOH projections56. In reality, however, habitats are not only defined by biophysical land cover but by a complex interplay of various factors. Changes in habitat quality, for example through management57, extraction58 or pollution by nutrients or pesticides59, could not be considered here. Higher input use efficiencies and improved management practices must accompany any yield increases to halt habitat degradation by pollution7. Climate change could also cause changes in the quality and distribution of suitable habitats60 and may therefore interact with other drivers of habitat change61. Drastic reductions in GHG emissions and land-based carbon uptake through ecosystem restoration28 are therefore critical to prevent climate-driven disruption of biodiversity62, while adaptive conservation approaches need to consider habitat connectivity and the dispersal of species along environmental gradients13,63.

Our study represents habitat and landscape change at a high spatial resolution within a dynamic framework. Yet, some socio-economic outcomes, such as price and endogenous yield changes, could only be derived at the scale of our model regions. Spatial variations in sub-regional food prices that would provide more detailed information on the distributional consequences of conservation due to the loss of opportunity or heterogeneity in market integration and access to infrastructure64 were therefore not considered.

Our findings underline the global importance of coordinated and targeted action for nature conservation. This also involves reconciling diverse conservation value perspectives—including the ‘Nature for Culture’ perspective, which was not covered in our analysis—through active, informed deliberation among scientists, policymakers and the general public at local-to-global scales65,66.

Methods

Model descriptions

MAgPIE

MAgPIE is a global, multi-regional land-system model for assessing global land-use dynamics and their implications for sustainable development in the twenty-first century46,67. It spatially allocates agricultural commodity production to meet global demand for food, feed, bioenergy and biomass under biophysical and socio-economic constraints based on a recursive dynamic cost-optimization approach. The model considers various cost components, such as factor requirements (capital, labour and fertilizer), land conversion, transport, investment in agricultural productivity and irrigation costs. The model also accounts for depreciation and new investments in capital stocks for crop production in each time step, favouring historically cultivated locations while limiting the unrestricted relocation of production to areas with favourable climatic conditions but with a lack of required infrastructure. Trade flows are split between historical import and export patterns and comparative advantages in production, including trade costs68. Spatially explicit biophysical information is derived from the dynamic global vegetation, crop and hydrology model LPJmL at a 0.5-degree spatial resolution based on historical climate data69 (no climate change effects). Vegetation, litter and soil carbon stocks as well as water availability are taken from LPJmL4 (refs. 70,71), while crop irrigation water requirements and yield patterns (cropland and grassland) are derived from simulations with LPJmL5 with unlimited nitrogen supply72. To address computational constraints, all model inputs at 0.5-degree resolution are aggregated to spatial clusters for optimization using a k-means clustering algorithm73. Socio-economic constraints, such as trade patterns and interest rates, are defined at the scale of 12 model regions (Supplementary Fig. 1 and Supplementary Table 1), with larger economies resolved individually and smaller ones grouped together. All major crop and livestock species, non-food agricultural commodities and supply chain losses are included in MAgPIE. Land competition is modelled in terms of the cost-effectiveness between crop and livestock production, and considering the land conservation targets formulated in each scenario.

Following previous studies, several of the outcome indicators used here are directly derived from MAgPIE (Extended Data Table 1). Net CO2 emissions from land-use change are calculated on the basis of carbon losses and uptake due to the conversion of forested and non-forest ecosystems and regrowth on abandoned agricultural land, respectively, between time steps18. CH4 emissions include emissions from enteric fermentation, animal waste management, rice production and the burning of agricultural residues and are estimated from the level of ruminant livestock production, manure and area of rice production. N2O emissions are estimated according to a nitrogen budget approach74 that accounts for differences between organic and inorganic nitrogen inputs and removals through harvesting or grazing on agricultural land and also for differences between excreted and recycled manure in animal waste management. On non-agricultural land, we assume a stable state where nitrogen fixation equals surplus. Average annual crop yield changes are calculated from yield differences between time steps. Yield changes are driven by cropland relocation to areas with higher yields and the cost-effectiveness of investments into irrigation as well as research and development investments into yield-increasing technological change75 compared with other cost components, such as land conversion or transport. Food prices are derived from the shadow price (Lagrange multiplier) of producing one additional unit of food and multiplied by the per capita supply of agricultural food products (including waste) to obtain the expenditure on agricultural products76. MAgPIE is intensely validated using the R library mrvalidation tool27, which draws on a validation database that contains historical data for most output indicators (see also Supplementary Figs. 21–30).

SEALS

SEALS is used to map simulated land-cover changes in MAgPIE at a scale relevant for assessing landscape and biodiversity change28,29 (see ref. 28 for a discussion of the MAgPIE–SEALS framework). SEALS spatially allocates projected land-cover changes from MAgPIE to a resolution of 10 arcsec (field scale) using adjacency relationships, physical suitability and conversion eligibility information. The initial state of the model landscape is defined by a high-resolution LULC map from the European Space Agency’s Climate Change Initiative (ESA-CCI) for the year 2020. The 37 ESA-CCI land-cover classes are aggregated into 7 functional types, namely, cropland, grassland, forest, non-forest vegetation (for example, shrubland or herbaceous cover), urban, barren land and water. Subsequently, the spatial adjacency strength for each functional type on neighbouring land-cover pixels is quantified, considering both distance and pixel agglomeration. The physical suitability for land-cover allocation is determined by high-resolution data on soil type and quality, topographic information, climate and travel time to the nearest market. In addition, allocation constraints are applied to prevent further cropland or pasture expansion in areas such as urban land, current PAs or water bodies. In the BIOS scenario, the expansion of crop, pasture and urban land is also prohibited in conservation priority areas to meet the 30 × 30 target of the GBF (see below). The overall suitability map is formed by combining adjacency relations, physical suitability and conversion eligibility, where pixels are ranked according to their suitability. In an iterative process, land-cover changes are then allocated to the highest-ranked grid cells until all projected changes from MAgPIE between 2020 and 2050 are assigned to generate high-resolution land-cover maps for each modelled scenario. Specific coefficients for different adjacency relationships and physical suitability are empirically determined by iteratively applying the allocation algorithm on a time series (2000–2010) of LULC maps from ESA-CCI. The coefficients most predictive on withheld historical data (2011–2015) are selected during this process28,29.

Biodiversity and NCP indicators

Landscape pollination sufficiency

Changes in configurational landscape heterogeneity and pollination sufficiency were derived on the basis of fine-scale projections (300 m × 300 m at the equator) of land-cover changes using MAgPIE–SEALS coupling. Pollination sufficiency is defined by the presence of semi-natural habitat (native forest and non-forest vegetation, including native grassland) near cropland28,77. To derive pollination sufficiency scores, we ranked all cropland pixels from 0 to 1 based on the proportion of semi-natural habitat within a 2-km foraging radius typically found in wild pollinator communities. We assigned a value of 1 where the proportion of semi-natural habitat within the foraging range around cropland is >30%, while values between 0 and 1 indicate a proportion of 0–30% (refs. 28,78). Although this metric does not account for other important factors that affect wild pollinator abundance, such as pesticide applications, the presence of meta populations, or the quality of nesting habitats or floral resources, multiple global and local studies have shown that the amount of semi-natural habitat around cropland can function both as a reliable proxy for wild pollination supply and as an important driver of biodiversity and other regulating NCP in various landscape contexts41,79.

Soil loss by water erosion

Soil loss by water erosion was estimated using the GloSEM platform. GloSEM uses a global geographical information system (GIS) implementation of the Revised Universal Soil Loss Equation (RUSLE) model, which was developed and validated by Borelli et al.4,44. In contrast to process-based soil erosion models, which have limited global applicability, GloSEM uses a parsimonious yet physically plausible approach to quantify soil erosion resulting from sheet and drill erosion processes at field scale. The accuracy of soil loss estimates from RUSLE-type models such as GloSEM has demonstrated to be suitable for most practical and policy applications44. Similar to other RUSLE-type models, soil loss in GloSEM is determined by a driving force (rainfall erosivity), a resistance term (erodibility of the soil) as well as topographical and land-cover information. Global rainfall erosivity maps were derived from the Global Rainfall Erosivity Database using Gaussian process regression and covariates from the WorldClim database80, while soil erodibility was estimated using soil data from the International Soil Reference and Information Centre (ISRIC) SoilGrids database81. Topographical information was obtained by processing digital elevation model (DEM) data using a two-dimensional GIS-based approach82. The land-cover and management factor was calculated using separate approaches for cropland and non-cropland areas. For cropland, spatial cropping patterns from MAgPIE at the 0.5-degree level were used and individual land-cover factors were assigned to 20 crop functional types based on literature values44. Subsequently, an area-weighted mean between all crop groups in each 0.5-degree grid cell was calculated and matched to projected fine-scale cropland maps from MAgPIE–SEALS. In non-cropland areas, land-cover factors were estimated by combining literature values for forested and non-forested areas with potential annual vegetation and forest-cover maps based on FCOVER data83 and tree-cover data84, as detailed in von Jeetze et al.28.

Area of habitat

The AOH refers to the habitat available to a species within its geographic range37. Here, we produced vertebrate AOH maps for 6,374 amphibian, 9,124 bird, 5,351 mammal and 6,877 reptile species for the year 2020 and used land-cover projections based on the MAgPIE–SEALS framework to assess AOH changes under different land-use scenarios. Range polygons for each species were obtained from the IUCN Red List85 and BirdLife International86. In line with previous work, we retained separate range polygons for resident, breeding, non-breeding and uncertain seasonality during processing11,36, but aggregated the final results across all seasonal ranges of each species to simplify the presentation of our results. We further excluded all species ranges with unknown presence and species with no habitat information. Habitat preferences for each species were derived from the IUCN Red List85 and are reported for 12 different habitat classes based on the IUCN Habitats Classification Scheme87. To relate the IUCN habitats to spatial land-cover information we used the habitat–land cover model of Lumbierres et al.55. The model associates 12 main habitat classes of the IUCN Habitats Classification Scheme with ESA-CCI land-cover classes and provides a translation table based on 3 different thresholds of positive association. We selected the threshold with the strongest habitat–land-cover association, which in similar applications has shown to considerably reduce commission errors without compromising validation performance of the AOH estimates36,56. For artificial degraded forest and plantation habitats, we chose the second strongest association threshold because none of the association values surpassed the threshold with the strongest association. Artificial degraded forest and plantation habitats were therefore associated with ESA-CCI classes 12 (cropland, rainfed, tree or shrub cover), 20 (cropland, irrigated or post-flooding) and 190 (urban areas55; see Supplementary Table 2). Habitats classified as marginal or of no major importance in the IUCN Red List were omitted from the data. For abandoned agricultural land we followed the approach of Humpenöder et al.88 with updated parameters for the Chapman–Richards growth function from Braakhekke et al.89, which assumes the same carbon density and the same species pool as during natural succession. Thus, in the early stages of natural succession, when the carbon density is ≤20 tC ha−1, the land is classified as non-forest vegetation (native grassland or shrubland) in the ‘other land’ pool. After reaching a carbon density of >20 tC ha−1, the land is classified as forest vegetation. We maintained this approach in the MAgPIE–SEALS coupling and classified abandoned agricultural land in the early successional stage as open habitat (native grassland or shrubland) and land in the later successional stages as forest habitat.

Changes in AOH were estimated following a two-step approach involving a pre-processing step and a scenario processing step (Supplementary Fig. 3). The pre-processing step involved subtracting unsuitable elevation areas from the range37 and translating IUCN habitats to ESA-CCI land-cover classes for each species. Species elevational range data were also obtained from the IUCN Red List and areas outside the elevational limits were subtracted on the basis of WorldClim 1-km elevation data90. Some range polygons had no remaining areas within their species elevational limits and were therefore omitted from the dataset. During the scenario processing step, unsuitable areas were subtracted from the pre-processed range polygons based on habitat preferences and land-cover projections from MAgPIE–SEALS to produce the final AOH estimates. AOH estimates for the year 2020 were derived directly from the ESA-CCI land-cover map. All calculations were performed using R (ref. 91) and the terra package92. Owing to the high computational load, computations were run on a high-performance cluster and parallelized using the packages foreach93 and doParallel93.

Modelled scenarios

We used the ‘middle-of-the-road’ socio-economic pathway (SSP2) for the land-use sector30 as our reference scenario. The SSP2 pathway assumes continued trends of moderate population and income growth and includes currently implemented land-use regulations. Ensuing food demand projections are based on anthropometric and econometric approaches that incorporate country-specific information on income, body-height, age-cohort, sex and body-mass-index distributions with elasticity parameters derived from historical data6. Information on PAs in the historical period (1995–2020) was taken from the World Database on Protected Areas (WDPA)94. We selected all legally protected areas that meet the IUCN and Convention on Biological Diversity definitions of PAs. Land use within PAs was determined on the basis of ESA-CCI land-cover data and aggregated to MAgPIE land pools. In the reference scenario, land use within PAs after 2020 was held constant (that is, no cropland, pasture or urban expansion), assuming full compliance and considering mixed land use, such as in IUCN PA categories V and VI.

Our alternative scenarios differ from the reference scenario in terms of per capita food demand multiplied by the projected total population, as we assume a 50% (DIET50) and 100% (DIET100) convergence towards healthier diets, and the implementation of biodiversity conservation measures (BIOS) at the global scale.

  • DIET50 and DIET100: the healthy reference diet is based on minimum and maximum recommended daily intakes for different food groups, while allowing for considerable diversity in country-specific dietary patterns within the ranges of the recommended intake levels7. Minimum per capita intakes are formulated for legumes, fruits, vegetables and nuts. In countries where per capita intakes of these food groups do not meet the recommended levels, intakes are increased to these minimum levels. Maximum intakes are set for sugar, oils, poultry, eggs, red meat and dairy products. Where per capita intakes exceed the recommended amounts, intakes of these food groups are reduced to meet the recommendations. In addition, a convergence towards a healthy body mass index of 20 to 25 is assumed. The food-specific maximum and minimum intakes for the healthy reference diet were derived from Springmann et al.7, who have already considered country-specific dietary patterns. In the DIET100 scenario, we assume a full linear convergence towards the recommended dietary intakes by 2050. In the DIET50 scenario, we assume only a 50% convergence towards the per capita intake targets of the healthy reference diet. These are then linearly combined with the reference food intake following the SSP2 trajectory (Fig. 1). In terms of total daily calorie intake, we also assume a 50% convergence from the reference total calorie intake level to the total calorie target level, which is informed by the healthy body mass index target (20–25). However, to maintain minimum and maximum targets for the different food items of the healthy reference diet, we only adjusted the intake of staple calories to meet the total calorie targets.

  • BIOS: the BIOS scenario includes a bundle of conservation measures to support biosphere integrity. Compared with the reference scenario, the area of protected land is increased by 2,055 Mha to meet the GBF target of conserving at least 30% of the global land surface by 2030. The enlargement of protected areas is focused on areas of global importance for biodiversity conservation (Supplementary Fig. 2), including KBAs31, unprotected habitat in BHs and EBDs32, as well as the CCAs identified by Brennan et al.13. Our choice to expand protected areas in KBAs, BHs, EBDs and CCAs is informed by the current literature on conservation priority areas and seeks to balance different conservation objectives, taking into account areas of global importance as well as species migration13,32. Yet it remains just one plausible implementation of the GBF 30 × 30 target aimed at illustrating how land protection could contribute to global conservation. We assume that within these conservation priority areas, restoration targets for different land-cover types are based on pre-industrial (1750) levels derived from the LUH2 (v2) dataset95. In addition, the BIOS scenario also includes a set of measures that address key drivers of biodiversity loss outside of PAs. These include maintaining at least 20% of semi-natural vegetation (SNV) in managed landscapes by 2030 and a no-net-biodiversity-loss scheme after 2030 at the biome level35. The constraint to maintain at least 20% of SNV (native forest and non-forest vegetation) in managed landscapes is directly applied in the MAgPIE model and is formulated such that 20% of the potentially available cropland must be reserved for SNV. Because evidence suggests that the 20% target should be fulfilled at the 1 km × 1 km scale to achieve maximum benefits33,34, we also used high-resolution satellite imagery from the Copernicus Global Land Service96 to estimate how much current cropland would need to be relocated to other areas to fulfil this target. This information is then used as additional input in the model optimization. Furthermore, the constraint to maintain 20% SNV at the 1 km × 1 km level is also applied during the land allocation at high resolution with SEALS. Because SEALS uses a latitude/longitude World Geodetic System 1984 coordinate reference system (CRS), where the area covered by a single pixel varies with latitude, we used an Eckert IV equal-area CRS to identify which pixels should be conserved to maintain 20% SNV at the 1 km × 1 km scale and reprojected them to the latitude/longitude CRS. The no-net-biodiversity-loss scheme prohibits any reduction in the Biodiversity Intactness Index24 (BII) of each biome35 within a MAgPIE region and biogeographic realm. Any decrease in the area-weighted average BII at the biome level (for example, due to cropland or urban expansion) must therefore be offset by an increase in land uses with higher BII values within the same biome. The no-net-loss scheme thus acts as an additional guardrail for areas outside conservation priority areas and disincentivizes avoidable land conversions to agricultural land use. MAgPIE-specific BII coefficients for different land-use classes are based on Leclère et al.24 and are presented in Supplementary Table 3.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.