Delivering More from Land: A Review of Integrated Land Use Modelling for Sustainable Food Provision

Abstract

We conduct a literature review on integrated land use modelling to guide policy on sustainable food provisioning. Modelling approaches are discussed in the spatial, temporal, and human dimensions, as well as environmental and socio-economic considerations. Many studies have focused on model development over their application to specific policy objectives, often relying on top-down approaches. While ecosystem services are a frequent focus, indicators for their assessment are inconsistently quantified. Socio-economic considerations are coarse in scale compared to biophysical ones, limiting their use in nuanced policy development. Recommendations are made such as standardising data collection and sharing to streamline modelling processes and collaboration. Comprehensive ecosystem services frameworks would benefit from a more uniform classification of values. More bottom-up modelling, connected with top-down models, could account for the heterogeneity between smaller units of analysis such as the landscape, farms, or people. This could reveal further insights into the local contexts and decision-making responses essential for effective land use policy. These advancements would help to design policies that address the complexities of sustainable food provisioning. By connecting local and global perspectives, future models can support more resilient and adaptive food systems, ensuring sustainability in the face of environmental and socio-economic challenges.

1. Introduction

Integrated land use modelling provides a holistic framework for assessing, planning, and implementing more sustainable land use practices. Given the global challenges from land use in delivering more food and other private goods, while, at the same, time delivering more public goods in relation to water, air, biodiversity, and soils, etc., the multi-dimensional analyses provided by these models are ever more important. The complexity of interactions across these multiple dimensions and diverse policy options place more pressures on these models to deliver useful analysis at the relevant resolution and scale. In this paper, we undertake a literature review of land use models in the food space to highlight opportunities to enhance usability for policy design.

Understanding the relationships between food production, agricultural sustainability, climate change, and the environment have become critical as the population and global demand increase [1,2]. Land use practices aimed at food provision have had significant impacts on the environment, such as agricultural intensification [3], deforestation [4], agrochemical use [5], GHG emissions [6], and the generation of waste, causing burdens on ecosystems [7]. Food provision depends on the environment and, as climate change progresses, will have impacts at multiple scales [3]. Ensuring sustainability and resilience in food production involves adapting practices to changing environmental conditions while maintaining productivity [8,9]. For example, climate-resilient crops, water management, conservation, and sustainable land use [10,11].

Integrated land use modelling can provide evidence for decision making in the social, economic, and environmental dimensions [12]. Incorporating inputs from stakeholders, local communities, developers, and environmental organisations, such modelling can reflect diverse perspectives and priorities [13]. Projections can be developed to explore potential patterns and their impacts [14]. These approaches have expanded into a variety of sectors [4,15,16]. Agriculture, forestry, and other land use (AFOLUs) models focus on better resource management (e.g., suitability analysis [17], nutrient loading [5], and crop yield projections [18]) and forestry, allocation, and zoning [19], timber [20], watershed analysis [4], and harvest planning [21].

Agarwal [22] presents a theoretical framework to assess models in the context of space, time, and human decision making. He highlights data availability constraints and that independent models addressing similar problems are proliferating and running in parallel rather than being connected. He concluded that land use and land cover change are driven by human actions and drive changes that impact humans. So, to answer policy questions, key variables of interest need to be identified at an appropriate scale of analysis for the scenarios they anticipate. Pattern- and process-based methods are compared in Brown et al. [23]. They recommend finer-resolution data and better integration with socio-economic and biophysical data.

Economic and biophysical models are reviewed in Plantinga [24], as well as models that evaluate the provision of ecosystem services from private lands. They discuss the need to combine modelling approaches to highlight the performance of land use policies. Difficulties in comprehensive land use systems analyses are discussed in Van Vliet et al. [25], where context-specific case studies dominate the literature. Standardising experimental design is urged for a better comparability.

Trends over the past decade are covered in Verburg et al. [26], as well as opportunities to better represent multi-scalar dynamics, human agency, and demand–supply relations. In the context of sustainability and policy planning, they discuss the prevalence of downscaling from the global to local level, but feedback from the local to global level is poorly captured. These may highlight displacement effects, the governance structures of land use, adaptative learning, and any non-rational human behavioural drivers that underpin decision making. The impacts of land use land cover change (LULCC) on ecosystem services (ESs) are reviewed in Gomes et al. [27]. Provisioning and regulating services are prevalent, with a lack of cultural services. They also criticise that most of the work did not have any external validation.

Common gaps can be summarised from these reviews. There is a lack of reliable, up-to-date land use data at an appropriate resolution. In food production, human and economic processes are a challenge to integrate. Differing methodologies and terms make collaboration difficult. Modelling is primarily quantitative, while behaviours, cultural norms, or policy impacts need qualitative description and have non-linear relationships with land. Incorporating these may not align with technical modelling approaches. Biophysical, economic, and social processes occur at different scales, which may show different outcomes. However, coordination between different scales in processes, data, analysis, and policy planning may not be feasible [22]. Despite being a key component of food, sustainability, and resilience in land use, ecosystem services (ESs) are still underrepresented [27].

For actionable food policy, land use modelling needs to overcome these challenges. While integrated land use modelling is well-established to explore climate mitigation scenarios from a biophysical process perspective, there is still a lack of detail representing the critical connections between policy levers and the transition to more sustainable and resilient food production systems.

2. Conceptual Framework

In order to define the information to be collected in the review, a conceptual framework was developed. The wide scope of integrated land use modelling for land use–food production interactions is highlighted in Figure 1. Land is a system where interactions occur between land cover, land function, and land use type, including food production.

• Land cover being the physical and biological features already present and the layers of soils and biomass, as well as natural vegetation, crops, and human structures [28].
• Land function is tied to land cover in that different land cover types contribute to distinct ecological, hydrological, and biogeochemical functions [29]. It represents the roles that land plays in supporting biodiversity, regulating water flow, and the provision of other ecosystem services [29].
• Land use types interact heavily with land cover and land function. They represent the human activities on land such as agriculture, forestry, infrastructure development, or conservation. Human decisions regarding land use have significant implications for land cover change and can alter the function of the land [22]. Agricultural practices may alter land cover through deforestation while affecting land function by introducing changes in soil composition [30].

The interactions between these three will determine the availability and functionality of ecosystem services (ESs). Water provision, pollination, and climate regulation are critical for human well-being [31]. ESs can be defined as an ecosystem’s contribution, either directly or indirectly, to human well-being [32]. They are broadly categorised into provisioning, regulating, and cultural services.

• Provisioning being the economic activities linked to goods provided by the system, such as crops or timber.
• Regulating services are indirect benefits provided by the environment.
• Cultural services are the immaterial experiences of people in the natural environment [33].

Consumers benefit from ES provision in the form of food, water, and materials, with their level and quality dependent on the land systems. Understanding and valuing the relationship between ESs and consumers is, therefore, a key element of land-based sustainability [31]. Policies, education, and conservation efforts aimed at preserving ESs can contribute to the well-being of both natural systems and the socio-economy. Therefore, integrating ES into modelling has become increasingly important.

The value chain represents the entire process of bringing the product from the land to the consumer, encompassing various stages of production, processing, and distribution and acting as a key connector between land use and consumers [34]. The choice of crop or livestock rests on the farmer. The income of the farm, rights to the land, type of tenure system, and the ability of the farmer to access funds, technology, and the market all affect decisions [35]. Different land use practices require varying combinations of inputs, while their frequency and intensity will fluctuate depending on land use practices and market conditions [36]. Consumer preferences will influence market dynamics, having an effect via the value chain [37].

As the optimal functioning of land use systems depends on policy [38], when modelling for sustainability, a balance must be reached between development needs, environmental conservation, and ensuring equitability [39]. While environmental, agricultural, and forest policy will have direct impacts on this system, there are other regulations and guidelines that will indirectly govern land use. Zoning laws, development programmes, and tax incentives may not directly target land or the value chain, but can still influence production patterns and land use [40].

The significance of these relationships will change across scales [22], resolutions, and the granularity of the analysis, with different implications for targeted policy [41]. Analysis can be conducted at a global, national, or regional level, with global regions as national comparative units down to 10 m × 10 m plots. The time dimension refers to the specific timing of the assessment, the smallest temporal unit of analysis, and the duration of the analysis [41]. Climate data can be recorded daily, while agricultural data can be yearly crop or animal yields. Forestry activity involves longer life cycles. Economic data are often recorded yearly or quarterly. Behind each element of this conceptual framework, there are people, livelihoods, sectors, and organisational structures whose decisions and priorities will influence the applicability of model outputs.

The areas of interest for this review are the biophysical, economic, and human processes seen in Figure 1, the significance of the relationships in between, and the depth analyses taken when modelling these elements.

3. Methodology

In land use modelling, the breadth of the literature and gathering information from a wide range of sources are considered to be important for a comprehensive coverage. Due to the range of disciplines and varying terminology, research approaches, and scales of the systems across the land use literature, a hybrid of scoping and bibliometric analysis was deemed appropriate, with the aim of identifying the qualitative characteristics and patterns in modelling. A qualitative description of these characteristics facilitates this interdisciplinary field. It may provide additional insights not yet captured through rigid systematic analysis, often having inflexible criteria. In this review, human decision making, policy implications, and governance structures are factors of interest (Figure 1), factors which influence land use but may not be explicit in modelling across journals. As the key connector in land and food production, the extent of value chain integration into the models may be better investigated. This narrative approach aims to draw further connections, complement traditional quantitative reviews, and provide a fuller picture of modelling trends.

Scoping reviews have been an established strategy for analysing fields with a high heterogeneity and can provide a broad overview of the key themes, consensus, and challenges across the literature landscape [42,43]. They address broad research questions that aim to explore the extent of the existing literature, rather than test specific hypotheses [44]. Bibliometrics can additionally highlight authors, institutions, and locations, as well as key journals. This review applies a five-step process that involves the following: (1) identifying the research question; (2) identifying relevant studies; (3) study selection; (4) charting the data; and (5) collating information. The aim of this review is to explore land use models in this space. We investigate the following questions: (1) What dimensions are most commonly considered in the land use modelling literature, and how are they defined? (2) What are the critical relationships and priorities between these dimensions in the context of sustainable land use? (3) How do methodological choices regarding space, time, and human decision making shape land use models and their applicability?

A search of keywords was constructed. The Boolean operators OR and AND were included and the following search string was used:

• “land use model” OR “land use modelling” OR “land use modeling”.

AND

• “agriculture” OR “agricultural” OR “food” OR “environment” OR “environmental” OR “climate change” OR “sustainability” OR “economic” OR “social” OR “policy”.

Only studies from 2013 to 2023 were considered, including any relevant reports, presentation papers, and manuals. As models used in different regions were a factor of interest, there was no geographical limit. The exclusion criteria were as follows:

• Literature reviews on land use modelling, as specific case studies were needed where dimensions were identifiable.
• Land use modelling that deals with urban development, transport, energy, or water only, as these models are not involved in the production of food or materials.
• Research not in the English language.

Spatial and temporal characteristics were noted as well as input data and biophysical, economic, and human characteristics. Where mentioned, policy and ecosystem services integration were examined. To highlight commonalities across papers, the selection of dimensions listed in this review were defined and grouped based on the literature gathered.

4. Results

A search for papers published between 2013 and 2023 yielded 14,000 results on Google Scholar. A focused selection of 60 papers was taken for analysis after screening (see Supplementary Materials for full list). In total, 60 papers represented an appropriate level of thematic saturation with a manageable timeframe for a detailed evaluation of each paper. A total of 95% were peer-reviewed articles (

Table 1. Literature selected by type.

Literature Type	No. of Papers	Share
Article	57	0.95
Report	1	0.02
Other	2	0.03
Total	60	1.0

). In additional informal searching, coupled with the level of their relevance to agriculture and food production, citations and methodological balance were also considered to ensure a mix of large, influential works and smaller specific case studies. As there were overlapping dimensions in the literature, a percentage share of 60 papers was included to represent prevalence (see Supplementary Materials for full tables).

4.1. Location

Location in this case refers to the geographical context of the study (Figure 2 and Figure 3). Global studies involved a worldwide analysis. Over half of the studies were based in Europe, Africa, and the Middle East. Most [8] studies from North America came from the US. In Europe, most studies were based in Austria (four papers) and the Netherlands [2].

Global papers included Asia-Pacific Integrated Modelling (AIM/CGE), a computational general equilibrium (CGE) model widely used in climate change studies [45]. The spatial distribution of emissions was downscaled and coupled with a land use allocation function [46]. Global sustainability objectives were explored through Shared Socio-economic Pathways (SSPs) with IMAGE [47]. The Model of Agricultural Production and its Impact on the Environment (MAgPIE) was used to forecast the effects of technological change across global regions [48]. Land system changes and intensification were mapped under the OECD Environmental Outlook [49]. The NEXUS partial equilibrium model was assessed for its accuracy against historical agricultural data [50]. The Global Change Analysis Model (GCAM) appeared frequently in US-based studies, even though this model is designed for global scope [51]. One US paper used Google Earth Engine with the cropland data layer to create a crop rotation and crop frequency map [52]. Another US paper used an EPA model to map hydrological and water quality changes [53].

Of the European studies, considerable diversity could be observed where country- or region-specific models were developed to accommodate differing datasets, compared to the more standardised analyses seen in the US. European study models included LUISA [54], Metronamica [55], the cropping system model EPIC, the agricultural sector model PASMA [56], the farm-level model FAMOS [57], the market conditions model CAPRI [58], an income maximization model [59], the interactive European model PRELUDE [60], the emissions model BioBam [61], the socio-ecological land use model SECLAND [62], and the machine learning model BLUMAP [63]. The agricultural production model ROMFARMS and discrete choice model PLUS were combined in the Netherlands [64]. The French crop model STICS was also used [65]. Short-rotation forestry in Germany was looked at with INCLUDE [66].

Figure 2. Number of papers by geographical location.

Figure 2. Number of papers by geographical location.

Figure 3. Distribution of papers by country.

Figure 3. Distribution of papers by country.

4.2. Space and Time

Figure 4 describes the spatial scale (resolution), where “Global” refers to a worldwide scale, “National” is countrywide, “Regional” is defined as an area between >1 km² and <10,000 km² within a country, or a defined spatial unitor agro-ecological zone, etc., and “Granular” has a spatial resolution of anything under 1 km². Some papers involved a multi-scale analysis [54,59,67]. The regional and granular spatial resolutions were the most represented.

The scales of the models themselves were separated and compared with the spatial resolution of the study (Figure 5). The spatial scale of a study focuses on the land area and the level of detail at which the land use is analysed, while the scale of a model refers to the model’s design and resolution in terms of its ability and scope to simulate land use change processes. Global models have a global scope and are used in cross-country analysis. Macro models focus on national or large regional systems or aggregated sectors and micro models consider the smallest units of analysis, i.e., an individual, household, or farm. In terms of spatial scale, there was a large representation of the granular resolution. However, most models were found to be macro, indicating a mismatch between the spatial scale of the study area and the level of detail in the modelling carried out.

Figure 6 represents the period of analysis or time intervals used against the model scale. Time steps of 1 year were the most common. Although some data like climate and other natural cycles are available as frequent as daily, a significant amount of land cover or land use data are collected and reported on an annual basis. Land use changes, especially agricultural, occur annually. Policy planning and economic cycles involve data collection planning processes, development projects, and budget allocations and occur in year increments. In the context of data availability, compatibility, and ease of interpretation, scenario projections would naturally follow these cycles. Due to a longer life cycle, papers with a forestry focus tended to involve 5- or 10-year time steps in their modelling frameworks.

The length of analysis represents the time scale of the modelling or how long these time steps were run (Figure 7). There was more heterogeneity in the length of analysis compared to the time intervals. The most common were 20–30-year long analyses. For example, one study projected the environment, households, and their impact on land use in response to a global policy [68]. Overall, 17 studies had no temporal dimension. Capturing temporal dynamics highlighted seasonal changes in growth cycles, harvesting, yields, and market demand [69], factors that are often key performance indicators for food policy [70]. Some static studies compared a baseline year with a projection year, with no time intervals shown in between. Longer analysis lengths incorporated historical data and were exploratory. One paper modelled the impacts of historical settlement patterns on land use during the Roman period [64]. Another study had an analysis length of 124 years [71] to map forestry patterns. These occurred in 20-year intervals. Modelling aimed at policymaking or mapping socio-economic drivers mainly occurred in the 1–30 years time length.

4.3. Application

The applications of models by model scale are included in Figure 8, with most papers having more than one focus. Agricultural-based activity such as crops, livestock, timber, or farm and commercial land can be considered as private goods. Public goods involve biodiversity, conservation, carbon sequestration, or other ecosystem services that impact the public. Public and private priorities were a factor of interest, as both are valuable in food-sustainability-related objectives [72]. Climate change impact monitoring or mitigation represented the main focus of papers for public good rather than food specifically. For example, the spatial distribution of carbon stock density was used for assessing climate change impacts [46], the drivers of global land use change in relation to climate change [47], and water quality risk assessment [56].

Agricultural management, forest management, and planning represented the majority of the modelling for private goods. The effects of technological innovation and change policies on land use were investigated [48], as well as agricultural land abandonment [54], intensification versus expansion and its effects on goods [49], and the economic potential of crops such as soy and switchgrass [73]. Overall, a “top-down” modelling approach was more prevalent in papers, where resulting land use changes were modelled from large climatic, economic, or policy shifts. A “bottom-up” approach was less common, where these shifts would be determined by emergent changes at the landscape level.

A large proportion of papers aimed to monitor overall land use land cover change (LULCC) trends, with neither a private nor public good focus, often demonstrating methodological improvements or attempting to downscale previous work to a finer spatial resolution. For example, a toolkit of functions was used to conduct the agricultural modelling of the US cropland data layer [52]. Another paper showcased a participatory land use modelling approach in which key stakeholders were involved in the parameterisation and evaluation of outcomes [55]. The downscaling of global land cover projections was carried out to match land cover data at a finer spatial resolution for climate mitigation pathways [74]. Another paper downscaled the Global Change Assessment Model (GCAM) land allocations for more region-based assessments [51]. The focus remained on methodological improvements rather than modelling specific policy scenarios.

4.4. Methodological Approach

In Figure 9, the models are categorised by common repeated methodologies and compared with the scale of the model from Figure 5. Shares in this case are expressed against the total papers found for each model scale, with some having more than one approach. There were individual models with distinct functions considered as modules or parts of a larger framework, where outputs from one model would feed into another. In total, 25 papers simulated future land use changes through spatial analyses. Some papers only focused on spatial analysis or developing a spatial decision support algorithm [67]. A spatial analysis, coupled with an empirical function to inform a land use allocation mechanism, was the most common combination in global and macro models. Spatial elements were coupled in land use allocation in [75], where urbanisation patterns were projected to show the risk areas in nearby protected wetlands. Optimisation approaches involve finding an optimum spatial arrangement of land uses under scenarios such as maximum productivity, economic returns, or conservation. For example, an optimisation approach was used to estimate the mitigation potential of incentives for carbon stock conservation [76]. Equilibrium approaches looked at the balance between supply and consumption demand. Land expropriation price and regulation scenarios were mapped from a CGE model in a province in China under a sustainable development policy [77].

Micro analytical techniques can supplement macro spatial and temporal information with enhanced human dimensions, where data may be weak regarding the spatial and temporal scope of a model. In cellular automata (CA), land area was represented as a grid of cells with designated land use types or states, influenced by neighbouring cells and, under certain transition rules, would iteratively update the states. CA was nearly always coupled with GIS due to the ease of representation through raster analysis and could account for the shortcomings of GIS, such as its static spatial variables and non-linear features. CA is well-established in modelling urban systems [78], as is agent-based modelling (ABM). ABM has featured prominently in land use studies due to its ability to pick up more discrete drivers, iteratively inform, and account for some heterogeneity [79]. It involves individuals, households, farms, or businesses as units of analysis, simulating their actions and interactions under certain conditions.

One paper simulated household behaviour as a driver of land use change using a variant of the Land-Use DynAmic Simulator (LUDAS) model [68]. A feedback loop between household and landscape agents was set up, where land use choices were determined by altering the crop yields, perceptions, and tenure relations of households. The socio-economic and climatic effects on a farm agent’s decision making were explored [62], where changing land uses and intensity depended on the income and workload of the farm. Though ABM can pick up a smaller-scale, nuanced understanding and the heterogeneity of land use changes, individual agent data are needed for effective analysis, which can either be absent or too labour-intensive to compile. CA approaches are stronger in the analysis of spatial dynamics over larger study areas and operate with simpler rules. However, they may only show aggregate effects of policy.

Not all ABM used micro units. The conditions of the agents were aggregated over sectors or large areas and behaviour rules were uniform. The focus of this ABM was on more global economic responses and overall climate effects, while studies that used micro units focused on how farmer and consumer choices affected the landscape. Microsimulation modelling could account for more local and emergent impacts on policy and has been used in urban [80,81], transport [82], and health systems [83]. However, only three papers in the food production scope were found.

Statistical approaches often accompany a spatial- or process-based model to quantify the driving factors of land use change [24]. Logistic Regression was identified in 10 papers and was nearly always coupled with CA, with land use as the dependent variable and assumed drivers shaping changes. A Markov Chain was combined with CA in three papers to map transition probabilities. Hedonic calculations were used in two papers. For example, the net returns of crops and livestock were modelled based on averaged climate data, population density, irrigation rates, and government payments [69].

Discrete choice experiments were found in the following two models: LUISA in an EU-wide scope [54] and PLUS in the Netherlands with historical settlement patterns [64], used to map preferences for behavioural responses to changes. A few of the 60 papers had a statistical-only approach. One paper focused on comparing four statistical methods with the same dataset for their accuracy in predicting land use change [84]. The merits of statistical methods are in their ease of implementation and ability to provide a picture of land use change with limited resources and data. Due to a longer history of use, they are well-established and could be an easier option for institutions compared to process-based approaches, where modelling is more likely to be tied to the developer [85]. Due to their assumptions, they can be weaker in mapping the explicit processes associated with human decision making, whereas process-based models would better account for this [86].

Machine learning approaches appeared in six papers. Random forest was linked with CA in [87] to map transition potentials in different socio-economic scenarios. A cross-scale analysis was successful, though the authors stated that their results may not be reproducible in other regions. Bayesian Networks appeared in [63], where the drivers of three land use sectors and dependency pathways were identified. Targeted knowledge from questionnaires could also be updated into these networks, meaning local actors’ influence could be integrated, though the authors stressed that BN probability tables can be difficult to interpret for non-experts. Questionnaire data were also incorporated in [88], where the MaxEnt model examined household characteristics’ effects on land use. Artificial Neural Networks (ANNs) were coupled with vegetation maps and aerial photographs in [71] to predict future forest patterns, while [89] used both an ANN and Random Forest on plant diversity. Both studies had a climate change focus. Though there was sparse representation of machine learning in the literature gathered, it showed promise in handling scarce, qualitative, or mismatched data, with some ability to handle multi-scale analysis and more local drivers of land use change. Machine learning methods in land use have been increasing in recent years due to the recent advancements in earth observation technologies and the increasing availability of data from remote sensing, as well as progress in automated classification techniques [90,91]. There are difficulties interpreting their decision-making process, which could discourage stakeholders. They are also prone to overfitting with limited or noisy data. However, their potential to pick up hidden trends and handle complex systems make machine learning methods an increasingly favourable approach for sustainability research.

4.5. Data

Where available, the input data used for the modelling were recorded and categorised by type (Figure 10). Most papers used maps or spatial attributes. A relationship could be seen between the comprehensiveness and variety offered by a single data source and the scale of the modelling. Data that were open source, offered variety, and in multiple formats, not only in statistics, but in maps, imagery, raster, or shape files, were the most utilised. The UN Food and Agriculture Organisation (FAO) maps and statistical databases were heavily relied upon as input data across the global and macro models. The data sources of global models saw repetitions of FAO, The Inter-Sectoral Impact Model Intercomparison Project (ISIMIP), World Bank, History Database of the Global Environment (HYDE), Global Trade Analysis Project (GTAP), and other global model outputs such as the Lund–Potsdam–Jena managed Land (LPJmL) model and the Global Change Assessment model (GCAM). The data sources used for LULCC monitoring included Google Earth, NASA’s LandSat and MODIS imagery, and the National Land Cover Database, which were then coupled with one of the above global model data sources to map long term projections. US papers that had a more specific interest in climate vs. net returns, farmer adaptation, or a change in economic conditions incorporated more econometrics and made additional use of the National Resource Inventory, US Forest Inventory Analysis, and the US census of agriculture.

European studies drew data from the European Commission’s EUROSTAT and IASC, European Spatial Agency, European Environmental Network (including CORINE land cover), European Space Agency’s Land Cover Project, European Environmental Agency’s Copernicus land cover, and EUNIS database of habitat types and protected sites. In many papers, the scale of the data available did not match the objective of the study. Data constraints clearly drove everything from the scale of the model chosen to objectives. One paper carried out a study at the landscape level, but the main data source was from IASC, where production values are aggregated [57]. Studies from countries that had census data were also frequently utilised in macro models. Population distribution was pulled from the Canadian census [84]. Land use type and animal numbers were collected from the agricultural census in Wales [92], though the lack of a comprehensive database meant that product price had to be constructed through averages of historical government data, highlighting the compromises made for these constraints. Countries that did not use census data drew from other available statistical agencies or government databases. Where available, Input–Output tables, statistical yearbooks, and financial yearbooks were drawn upon for macroeconomic data [77]. In general, however, even where human dimensions in relation to activity were incorporated into a model, they were at an aggregated level such as a hectarage of activity or at the greatest intensity, without sufficiently nuanced information to model policy or market interactions.

“Other” data, in this case, refer to observational or qualitative data such as interviews, questionnaires, panel discussions, or other sources collected through a participatory approach. This type of data occurred in 13 papers and was indicative of small-scale, heavily localised analyses found in micro or theoretical papers. Some of these papers were the only examples of attempting to map the local drivers of land use change and to integrate local knowledge into the modelling. In total, 17 papers carried out an analysis based on previously undertaken fieldwork or built upon other institutional projects and were location-specific. A low usage of these data relates to the difficulty in integrating modelling and the lack of time or resources in collection. But, when available, they are valuable in the context of human decision making and discrete local interactions at the landscape level.

4.6. Dimensions of Land Use

4.6.1. Biophysical Dimension

Food production is, in general, heavily dependent upon the characteristics of the land and the environment, from the soil to the topography, climate, and spatial access to markets (Figure 11). Given this relationship, all land use models have a biophysical dimension. Topography, including elevation, slope, aspect, etc., was the most prevalent dimension linked to spatial analysis. The frequent use of similar open-source land cover maps, soil databases, and climate data was observed. Some studies had already established Land Use Land Cover (LULC) maps and adopted a purely biophysical angle on land use change projections. Land Use Type and Land Use Suitability were listed as biophysical dimensions made up of varying combinations of topographical, climatic, and vegetation variables. Other papers had these separated into distinct biophysical dimensions. Although biophysical variables made up the majority of land use dimensions, biodiversity was underrepresented, despite it being a key indicator of land function [93]; it is difficult to model due to the diversity of genotypes, functional groups, and complicated interactions with land systems [94]. The prevalence of global climate models such as AIM, IMAGE, and CLUE accounted for the water/hydrology and biochemical cycles [46,47,49]. The biophysical dimensions represented the most consistent and granular level of analysis compared to modelling the economic and human dimensions, benefitting from more rigorous and standardised analysis and less objectivity compared to the economic and human dimensions.

4.6.2. Economic Dimensions

Food is one of the most traded commodities globally, with significantly different practices in different spatial locations, tied to land function in our conceptual framework. Figure 12 encompasses the economic considerations of the modelling. Specific economic variables in the modelling were sometimes hard to identify due to varying definitions across papers or not being disclosed in the methodology or appendices. In total, 10 papers had no identifiable economic variable, with a focus only on biophysical processes. The overall trend monitoring global and macro studies assumed income as an average through GDP. More localised studies used more scale-appropriate farm survey or census metrics. Commodity prices were global or national averages rather than local inventories. Trade and market dynamics were also downscaled from global indicators such as in FAO and fitted into regional studies. When considering a targeted policy for sustainable land use practices, a cost–benefit analysis based on aggregated economic outputs may not be accurate for specific areas and may not accurately reflect a policy outcome. Crop yield and crop prices were incorporated more frequently than livestock. Livestock was poorly integrated into the modelling in general. Nine papers included livestock density as a variable, but did not analyse differences between animal cohorts. Different cohorts and life stages of the animal will bring about different implications in the analysis. Just four papers included livestock efficiency, despite its importance in the sustainability of animal production [95]. The average GDP per land area was used as an indicator of socio-economic development efficiency, even in regional- or granular-resolution studies. This is a contrast to the fine spatial resolution mapping and biophysical dimensions with which they were coupled.

4.6.3. Human Dimensions

Converting land quality and potential to ecosystem services, particularly food-related, often requires significant human engagement, with implications running in both directions between humans and their environment. Figure 13 shows the ways in which human decision making was considered in the modelling. Eight papers had no human dimension, despite it making up a critical component of food production (Figure 1). Demographics was the most prominent dimension, with the most common parameter being population, either in density or distribution. “Accessibility”, in this case, refers to the distance of the individual, farm, or household to resources, including transportation networks, utilities, and public services. Some studies considered accessibility to be a physical (presence/absence) variable, while others considered it to be social (behavioural), indicating ambiguity with definitions.

Farming and land use decisions involved measuring implications based on a change in planting, crop type, or intensity. These were indicative of the CA and ABM approaches. Income and workload were used as measures of farmer satisfaction in two papers [62,89]. Semi-structured qualitative interviews with farmers were integrated into regional ABM to define farming styles. Consumption patterns were a regular variable of global models, where equilibrium and optimisation approaches would allocate land based on these changes [96,97]. Most papers incorporated consumer decisions as an overall change in demand, adjusting productivity levels and how that would affect the market, rather than considering individual behaviours and patterns of the consumers themselves. While the incorporation of human economic and social factors has improved over time, this lacks the power of more spatially aggregated models, reflecting the weak availability of spatially explicit and comparable human activity data.

4.7. Policy Choices

Food production is one of the most policy-driven economic sectors market and trade policies and agreements on farm income support policies for environmental and food safety regulations [98]. However, a deep policy analysis was limited in many studies (Figure 14). In total, 30% of papers had no policy dimension in their modelling. Applications varied from as broad as being “policy relevant” to more specific subsidising, zoning, or regulations imposed on certain areas of production. Papers carrying out only a policy-relevant analysis had more loose definitions such as “climate mitigation” or “land management”. Of the papers that did define and incorporate a policy, a graduation of analysis could be observed. Any measurable changes in environmental, economic, or social factors were considered to be the drivers that would shape the policy and give relevance and direction to its development. Papers with LULCC monitoring applications represented this type. Or papers had policy as a variable, for example, afforestation policy. Any changes in land use patterns or other related variables were evaluated as the policy outcome. These papers contained objectives in the form of identifiable metrics. Volumes of emissions, percentage yields, or the areas of land allocated to certain land use types were considered as key performance indicators. Given the weakness of rich human socio-economic data or model components, it is unsurprising that land use models have a relatively limited policy impact assessment capacity.

One paper assessed low-emission development strategies in the AFOLU sector [99]. Policy targets were decided through meetings with government officials and private sector representatives. These targets were assessed on their feasibility through changes in GHG emissions, carbon stock density, yields, prices, and land allocations. The authors concluded that policies investing in efficiency and productivity with less allocation of pasture would have more of an effect, rather than policies only aimed at preventing deforestation. Another study investigated policies on carbon pricing and pricing on nitrogen leaching against land use, production, farm management, and environmental impacts [100]. Combining national- and regional-level policies was more successful in meeting targets with the least loss in net revenue, rather than one. Despite representing useful evidence for food policy, there were fewer papers of these types. This could be due to the scope of the research and the high country- or region-specific nature of policy, or indicate some priority of “all-encompassing” models in the publishing sphere rather than connecting and leveraging the strengths of different models at different scales. There could be difficulties in defining what values would determine a successful policy outcome, as both public and private stakeholders would need to agree on targets.

4.8. Ecosystem Services

Many papers included ecosystem services (ESs) as the motivation and stressed the need for conserving these services. However, only 21 of 60 papers attempted to quantify these for modelling, highlighting the difficulty in assigning values to ESs. Of the 21 papers that included ESs in their modelling, these are further broken down in Figure 15. Provisioning services include food. Of the regulating services, habitat, nutrient cycling (C, N, and CH₄), and soil quality are included. Regulating services are tied to land function and provide an indication of land health and its ability to provide food and goods [31]. There was a prevalence of provisioning services, despite them having an antagonistic relationship with the other two services [101]. The integration of cultural services in the land use modelling literature was the sparsest. Studies that considered recreation and aesthetics were normally tied to a protected area or close to an urban area, with conservation as the main objective. Cultural services provide added value beyond the economic or ecological functions of land and shape how communities value and interact with the landscape [31].

One study integrated four ESs integrated into ABM in [102], including the production of meat and cereals (food), timber (goods), and the provision of recreation (cultural). Ecosystem services were quantified in the form of (1) capital occurring from natural productivity and human effects on productivity (such as the availability of finance or infrastructure) and (2) land manager agents, institutional agents, and competition. This approach accounted for human behaviour, multifunctional ES provision, and heterogeneity in land use intensities. However, there were many assumptions. “Nature value” was assumed to increase with a higher elevation through less footfall. The economic and social conditions of the simulations were left constant. The land manager “agents” were not able to switch land use, only their level of intensity. The multiple ownership of cells by one agent was not supported, highlighting the computational trade-offs when connecting human and ES elements into the modelling system observed in Figure 1.

Another paper developed a CA and hydrological model framework to create a spatial decision support tool for planning [103]. Although the study was not applied to any policy decisions, the initial results indicated that population growth, poor management, and communal land tenure systems were the main drivers associated with the intensive land uses responsible for declining ES function. The impact of the relationship between land use and market forces on carbon storage, forest biomass, crop production, and wildlife habitat was observed in [104]. Three national policy incentives were tested in another paper [104], a forest incentive, a tax on land converted from forest to crop or pasture, and the prohibition of conversion to urban land. Although the policy scenarios had some marginal effect on land use patterns, they did not significantly alter the trajectory of ES provision. Market and the structure of government programmes dictated trends to an extent that any policy interventions would need to be more aggressive for any notable change, indicating that organisational structures are another important human element not well-captured in land use modelling. Some ES provisions increase at the expense of others [105], where forest incentives policy saw an increase in timber production and biomass carbon, but a decline in food production.

5. Discussion and Conclusions

In this review, 60 land use modelling papers were examined in detail in terms of their usefulness for policy to drive sustainable food production. Applications were profiled, as well as approaches, spatial and temporal resolutions, and the choices made in the environmental, economic, and human dimensions.

The applications of these land use models concerning public goods provision had climate change mitigation and adaptation as the main objective, with a limited focus on food system dynamics. Productivity gains and expansion were prevalent in models concerning private goods. Even though productivity is important for adequate food provision, it relies on preserving resilience and land functionality, which was not emphasised in these studies. Other papers sought to improve techniques and downscale or enhance usability for future research, where the modelling itself was the focal point. Reflected in the word cloud of paper key words in Figure 16, land use modelling spans a diversity of disciplines, which results in a broad set of considerations ranging from the spatio-temporal to biophysical, socio-economic, and policy dimensions. While extensive and complex, integrated land use modelling’s usefulness in the multifaceted nature of sustainability research can be observed.

The spatial and temporal dimensions of the models were investigated. Many global or macro models were used for local-scale studies, while microsimulation in land use was rare, despite being the appropriate scale for the study, with the potential to uncover the drivers of land use changes and the benefits towards end consumers. Scenarios from macro models need to be downscaled to be applied locally. As macro models made up the majority of models found in this review, further analysis of these land use dimensions across a uniform number of each model scale is recommended. This could further highlight our gaps in understanding the land system and value chain relationship outlined in Figure 1. In terms of time, models commonly used one-year intervals, likely to match data availability and policy timelines, providing a balance between trend observation and computational efficiency. Few studies addressed regional or farm-level variability, although understanding this heterogeneity could clarify the production, economic, and environmental dynamics of food. The majority of the 60 studies aimed at broad trends, with few exploring underlying causes. Combining various models to leverage strengths could be observed, though data scarcity still drove many modelling choices, often using easily accessible, regularly updated datasets. Limitations in detailed local data led to averaging socio-economic figures, questioning the definition of a “fine spatial resolution.” Multiple linked models were often used to address these data gaps, highlighting the merits of working to connect existing methodologies. Observational data were scarce, but provide valuable insights for biodiversity and human behaviour studies, an area still challenging to predict in land use models. Improved standardisation and data sharing could enhance policy alignment and model reliability. Biophysical dimensions made up the most granular level of dimensions, which the economic and human dimensions did not align with. Improved data collection, standardisation, and improving the connectivity between the analytical processes of different models would also ameliorate this imbalance.

Ecosystem service analysis was included in 21 studies, focusing on provisioning services like food production over longer-term benefits like soil health. Monetary values are clearer for provisioning services, aiding in policy prioritisation. It is clear that analyses of land function, needed for our continued rate of food provision, is still lacking, and needs to be pushed in future. Highly cited urban, transport, and energy models benefit from longstanding data and incentives, which land use models in the food and climate sectors could emulate with increased investment and urgency for sustainable practices.

This paper adopts a food production lens for integrated land use modelling from the perspective of using integrated land use models for policy formulation. Although some models were explicitly focused on food or agricultural production [36,48,54,104], the level of detail related to food production, and in particular the wider value chain, was quite limited. As the demands upon the food system for multiple ecosystem services increase, and given the extensive policy levers used in driving these outcomes, there is an increasing need for analytical tools that incorporate the three dimensions of Space, Time, and Human Choice defined by Agarwal [22]. The availability of data has seen an explosion in models that contain the first or the first two, however, the human dimension has been limited. Participatory modelling techniques, incorporating local knowledge and stakeholder engagement, were sparse in this review, but showed promise in addressing this limitation. Though often confined to specific case studies, the level of behavioural economics and social science insights gathered completes this comprehensive picture in Figure 1 and is needed in future models for effective policy.

In model development, there is always a trade-off between dimensionality as greater dimensions increase in complexity in an exponential way, with consequences for validation and interpretability. Incorporating all dimensions at a global level is impossible given the differences in data and dimensions. There are few policy frameworks with real implementation power at this scale, albeit there are many global concerns and targets.

At present, even on continents with significant data analytical power and resources such as the European Union, integrated land use models are stronger for observable physical data such as crop areas, but weaker for human dimensions. Even where there are excellent human data at the national level, these data may not be geo-referenced or geo-referencing released to researchers. This inhibits the capacity to downscale, as in the case of the Farm Accountancy Data Network. Ironically, data available in low- and middle-income countries such as the World Bank’s Living Standards Measurement Study often contain very good food production information with geo-referencing available. Geo-referencing is essential to establish a detailed understanding of the interaction between food production, land use, and the environment, and, thus, needs further improvement.

A more effective approach to achieving the three-dimensional goal to support food planning is to start with the spatial scope of relevance to the specific policy framework, which may be the nation or more, as in the case of the Common Agricultural Policy. This spatial scope is likely to be an area that can generate resources required for implementation and is likely to have stronger coordination between analytical teams. Improving the coordination of definitions in the three dimensions, data release protocols, and consistent data collection approaches can increase the likelihood of take-up. Pilot national or economic regional initiatives can facilitate proof of concept and build confidence in the scalability of these tools, and can make it easier for other countries or regions to adopt these protocols, which, in time, will allow for greater possibilities for comparative analysis. Independently built autonomous models, but using common protocols, definitions, and methodologies, can, thus, be easier to build, resource, and eventually scale.

Rather than focusing on representing dimensions at all scales by one model, coupling existing models with already proven capabilities at defined scales leaves room for bottom-up modelling approaches to provide the valuable local information in national or regional policy frameworks. Bottom-up approaches in this review were less common and poorly integrated with other models, but still have potential for a better connectivity. Using these approaches might allow for an improved coordination of global food-related public and private goods. Globally, we have a reasonable understanding of what we need to do to deliver these public and private goods, but are often challenged in how to deliver them. Deepening our understanding of how human behaviour and policy- and market-related coordinating frameworks function within space- and time-related land use models can assist in designing the levels that can help us to achieve these goals.