Simulating Interactions Between Land Use and Land Cover Changes for Prospective Scenarios with FORESCEM

Abstract

Anticipating the socio-environmental impacts of spatial planning strategies is a prerequisite for sustainable development pathways. Land change models are increasingly employed to evaluate the impacts of spatial planning on land use and land cover, and their subsequent effects on ecosystem services and environmental resources. Nevertheless, modelling land use and land cover changes, and their interactions, at a fine scale to preserve future landscape patterns has been identified as a key challenge in the land change science community. This paper presents an innovative process-based model—the FORecasting landscapE SCEnarios Model (FORESCEM)—designed to spatially simulate fine-scale future land use and land cover changes (LUCC) based on narratives developed through participatory or expert-driven approaches. By clearly distinguishing land covers and land uses as two different but related inputs, its conception and architecture enable the assessment of interactions among LUCC within human-managed landscapes. It relies on conventional functions and properties of LUCC models, and aims at completing the existing land change models. Applied on a French case study, the validation results demonstrate the model’s capability to replicate LUCC dynamics, effectively simulating trend-based and trend-breaking LUCC trajectories under contrasting scenarios. More broadly, this paper questions and discusses the validation of land change models used for simulating future LUCC.

1. Introduction

Both land use and land cover changes (LUCC) critically affect land–atmosphere exchanges and exert significant influences on climate regulation, water and soil quality, biogeochemical cycles, biodiversity, and ecosystem services [1,2,3]. Evaluating potential future LUCC and their associated impacts has been recognized as a key priority in land change science, given their central role in global environmental change [2]. At the same time, effectively integrating considerations of climate, water and soil quality, biogeochemical processes, biodiversity, and ecosystem services into spatial planning remains a complex and persistent challenge [4,5,6,7].

A key challenge of resilient and adaptive planning strategies is the exploration of LUCC–ecosystem services interactions in the medium- to long-term. Indeed, the integration of dynamic, spatially explicit models with scenario-based approaches has emerged as a widely adopted and effective method for envisioning future LUCC [8,9,10], supporting decision-making processes for stakeholders. Yet, these approaches frequently remain short-sighted, evaluating implications of current or short-term land management plans and policies. An overarching, yet underestimated issue for medium- to long-term planning is to acknowledge the intertwining of land use and land cover. The two concepts differ in fundamental ways that are critical when analyzing, modelling, and simulating socio-ecological systems. Two key differences are that “[land cover] refers to the composition of the features of the Earth’s surface, while [land use] refers to the type of human activity taking place at or near the Earth’s surface” [11] and the relationship between land use and land cover is not a one-to-one relationship. A persistent challenge in LUCC modelling lies in adequately capturing the interactions between land use and land cover changes [12]. Three types of interactions can be considered: horizontal, vertical and temporal. Horizontal interactions rely on neighbouring and spatial network influences which are commonly well considered with cellular automaton (CA) or agent-based models [13,14]. Vertical interactions rely on land use and land cover interdependencies: some land covers (e.g., crops like maize or wheat) may depend on one land use (agriculture) while others (broadleaf forest, for instance) may belong to different land uses (e.g., silviculture or abandonment). Inversely, one land use may specifically influence land cover changes: silviculture may lead to forest harvesting and plantations cycles while abandonment can be assimilated to natural vegetation growth. Temporal interactions rely on legacies and path dependencies [15]: legacies can consider historical land uses that influence current soil quality, for instance, while path dependencies account for land system stationarity and may impact the weight of land change drivers over time [16]. However, if interactions between land use and land cover changes are prime features of the land change model [17,18,19], most of the literature considers land uses and land covers as a whole, within one single land classification map, limiting the consideration of all their interactions.

Simulating both land use and land cover changes and their interactions to explore the future requires the integration of scenario-based frameworks in modelling, such as the Story-And-Simulation (SAS) approach [20]. Scenarios are developed through participatory processes or expert knowledge and are formalized as narratives that specify where and how LUCC may occur in the future, driven by anticipated policies or land management strategies [21]. These narratives, once formalized and translated into quantitative variables, serve as inputs to parameterize land change models. The resulting outputs—LUCC maps—not only visualize the narratives but also quantify the diversity of potential future landscapes, enabling the illustration of imaginative yet data-informed scenarios [22], and relevant for ecological assessments and land use planning applications [23]. This approach aligns with the foresight paradigm [24], which emphasizes the exploration of multiple plausible futures rather than a single deterministic projection.

Furthermore, when coupled with scenario-based approaches, the integration of land use and land cover interactions helps address critical questions regarding the validation of prospective spatial planning (see the review by [25]). Commonly employed validation techniques—such as assessments of location accuracy and spatial pattern conformity—are particularly challenging to apply in land systems that are heavily influenced by human activity. These methods generally assume that land systems evolve along current trends, particularly in terms of spatial patterns, surface extents, and underlying drivers. While such techniques are effective when applied to retrospective simulations, they are inherently limited in the context of mid- to long-term future projections due to the deep uncertainty associated with future developments. In such cases, evaluating the accuracy of modelled outcomes becomes virtually impossible, as there is no empirical reference against which to compare them. Moreover, no LUCC model can perfectly reproduce reality in terms of landscape patterns and processes [26]. With integrated LUCC modelling, the validation focuses on more plausible spatial processes to assess whether the modelled landscapes exhibit spatial patterns and dynamics that are coherent, credible, and consistent with the underlying scenarios, as suggested by [22].

In this paper, we introduce an innovative and fine-scale LUCC model designed to simulate future, spatially explicit and contrasting scenarios accounting for land use and land cover interactions on the one hand, yet also able to function as conventional LUCC models on the other hand. The model is thus compared to the existing approaches before being applied on a French case study in France. Its evaluation is based on a two-pronged assessment approach: (1) a series of unitary tests to verify rule compliance, and (2) a retrospective simulation of a past period to assess the model’s capability to reproduce observed LUCC dynamics.

2. State of the Art

Simulating LUCC at fine spatial resolutions remains a challenge to preserve landscape patterns [27,28], due to data availability, computational demands, and the complexity of processes operating at multiple scales [29]. To introduce the FORESCEM (the FORecasting landscapE SCEnarios Model) framework, we compare it with some of the main cited and used LUCC models: CA_Markov [30], LCM [30], Dinamica [31], CLUE-S [32]) and PLUS ([33,34]). A comparison is made using functionalities and properties related to scenario modelling (listed in italic hereafter as reported in Table 1):

• Land demand. Two criteria are considered on how the land demand is set. First, land demand can stem from an internal module (like Markov chains), or from external estimates (e.g., expert or model-based approaches). Second, land demand depends on the model’s capability to explore a wide diversity of LUCC dynamics. Among available models, some only deal with slight variations (increases or decreases) in the past quantity of LUCC (for, e.g., [35]), while others can handle tipping points and breaking trends (non-linear variations) of future land demand.
• Probability maps. Every model can use an unlimited number of drivers to set where future LUCC are more likely to occur. Various approaches can be used to generate these maps: logistic regressions (LogReg), Weight of evidence (WoE) or deep learning/artificial neural nets (DP-ANN). These drivers can be incorporated in two main ways: some models only rely on the input map to estimate where land cover types are more likely to be located (suitability maps—SuitMap), whereas others use observed transitions between two dates to estimate where future LUCCs may occur (transition potential map—TPM) [36]. Although TPM-based approaches can outperform suitability mapping to predict LUCC, data availability and dependency on past trends may limit the ability of models to explore contrasting futures [37].
• Excluded/Constrains maps. Excluded maps are used to forbid some LUCC in some predefined areas using Boolean maps (Hard), while soft maps (ranging from 0 to 100%) can be used to prioritize or weight the influence of drivers. Their update along the simulation is also important to account for changes in land planning policies or non-path dependencies.
• Landscape patterns simulation. Simulating transparent and realistic spatiotemporal LUCC patterns is crucial to inform land policy-makers [17,29,38]. The most commonly used LUCC models are based on cellular automata (CA), relying on pixel-based or object-based approaches [33], which strongly impacts the realism of simulated anthropogenic landscape patterns and dynamics representation. Some pixel-based CA models incorporate dedicated functions to improve the patchiness of LUCC simulations [39,40], while others indirectly achieve this through the spatial rendering of probability maps [27]. By contrast, using object-based CA models may inherently favour the simulation of realistic landscape patterns, especially when the minimum mapping units (MMUs) correspond to the scale at which LUCC actually occurs. With excessively large MMUs, such models may fail to consistently simulate both anthropogenic and natural vegetation changes [33]. Lastly, integrating sub-regions that differ in land demand, probability and constraint maps can further contribute to the generation of realistic patterns.
• Land cover/Land use distinction. Accounting for interactions between land uses and land covers require a clear conceptual and operational distinction between them. In the literature, attempts have been made to differentiate the two within a single land use or land cover input map distinguishing different classes (e.g., Managed forest vs. Natural forest [41]). However, none of the above-cited models explicitly sets two different land cover and land use input maps [23] (Different maps), a fundamental distinction that underlies the FORESCEM. Lastly, while accounting for horizontal interactions is quasi-implicit with CA models, temporal interactions (legacies) require consideration of LUCC occurrences at the level of each MMU.

Table 1. Comparison of LUCC models. Definitions of criteria and abbreviations are given and detailed in the State o the Art section.

Function/Property	Criteria 1	CA_Markov [30]	LCM [30]	Dinamica [31]	CLUE-s [32]	PLUS [33]	FORESCEM (This Paper)
Land demand	Internal/external	Internal	Internal	Both	Both	Both	Both
	Breaking trends LUCC quantity	No	No, only slight variations in past LUCCs	Both	Both	Both	Both
Probability maps	LogReg/WoE/ DP-ANN	LogReg	LogReg/ DP-ANN	All	LogReg	All	LogReg
	SuitMap/TPM	SuitMap	TPM	TPM	SuitMap	A data mining combination of both [42]	SuitMap
Excluded/constraint maps	Hard/soft	Hard	Hard	Both	Both	Both	Both 2
	Updates	Not easily	Yes	Yes	Yes	Yes	Yes
Landscape patterns simulation	Pixel/object-based	Pixel-based	Pixel-based	Both	Pixel-based	Pixel-based	Both
	Related functions	No	No	No/Yes [40] 2	No	Yes	No/Yes 3
	Sub-regions	No	No	No	Yes	No	No
Land cover/land use distinction	Land covers/uses input data	Single map	Single map	Single map	Single map	Single map	Different maps
	Legacy	No	No	Yes	No/Yes 4	No	Yes

¹ Criteria are explained and listed in italic in the previous section. ² Both are mentioned and can be considered by using policy or event maps (see details in Section 3.1.2). ³ Both are mentioned because related functions depend on the MMU defined in model settings (pixel- or object-based). ⁴ CLUE-S does not fully support “Legacy” effects. While it relies on a static elasticity parameter to define conversion resistance, conversion rules account for land cover duration.

Table 1. Comparison of LUCC models. Definitions of criteria and abbreviations are given and detailed in the State o the Art section.

Function/Property	Criteria 1	CA_Markov [30]	LCM [30]	Dinamica [31]	CLUE-s [32]	PLUS [33]	FORESCEM (This Paper)
Land demand	Internal/external	Internal	Internal	Both	Both	Both	Both
	Breaking trends LUCC quantity	No	No, only slight variations in past LUCCs	Both	Both	Both	Both
Probability maps	LogReg/WoE/ DP-ANN	LogReg	LogReg/ DP-ANN	All	LogReg	All	LogReg
	SuitMap/TPM	SuitMap	TPM	TPM	SuitMap	A data mining combination of both [42]	SuitMap
Excluded/constraint maps	Hard/soft	Hard	Hard	Both	Both	Both	Both 2
	Updates	Not easily	Yes	Yes	Yes	Yes	Yes
Landscape patterns simulation	Pixel/object-based	Pixel-based	Pixel-based	Both	Pixel-based	Pixel-based	Both
	Related functions	No	No	No/Yes [40] 2	No	Yes	No/Yes 3
	Sub-regions	No	No	No	Yes	No	No
Land cover/land use distinction	Land covers/uses input data	Single map	Single map	Single map	Single map	Single map	Different maps
	Legacy	No	No	Yes	No/Yes 4	No	Yes

¹ Criteria are explained and listed in italic in the previous section. ² Both are mentioned and can be considered by using policy or event maps (see details in Section 3.1.2). ³ Both are mentioned because related functions depend on the MMU defined in model settings (pixel- or object-based). ⁴ CLUE-S does not fully support “Legacy” effects. While it relies on a static elasticity parameter to define conversion resistance, conversion rules account for land cover duration.

3. Materials and Methods

3.1. Model Design

3.1.1. Principles

The FORecasting landscapE SCEnarios Model (FORESCEM) is grounded on the following principles and definitions, described in Figure 1.

• Each portion of the study area is represented by coherent MMUs (a pixel or a group of pixels—referred to as parcels in the FORESCEM) to capture landscape patterns while reflecting the scale at which land changes commonly occur. Each parcel is characterized by a single land cover and a single land use. These MMUs can be derived from existing spatial datasets (e.g., cadastral maps or the LPIS database for European agricultural land) or generated through landscape segmentation.
• Land cover refers to a specific type of surface cover (e.g., meadows, maize, or forest), and its changes depend on the land use to which it depends (e.g., agriculture or forestry). Crop successions illustrate land cover changes linked to agricultural land use at the scale of an individual plot. In the absence of human activity, considered as a distinct land use category (“no land use” or “abandonment”), the associated land cover changes correspond to natural vegetation dynamics. In some cases, a given land cover (e.g., deciduous forest) may be associated with different land uses (e.g., “no land use” or silviculture).
• Land uses are generic categories (e.g., agriculture and silviculture) that describe how and for what purposes humans exploit the land. As such, artificialized areas are land use driven by human decisions (e.g., settlement development, land use planning) and may exhibit various types of urban fabrics or a single urban land cover class, or even vegetated areas corresponding, for instance, to urban parks.

These principles highlight subtle distinctions and are grounded in the definition provided in the Introduction, which differentiates land cover from land use, although this distinction is not always consistently made in the literature or in the input model datasets. For instance, the urban class is commonly treated either as a land cover or as a land use category for modelling simplification purposes. Land use and land cover classes are not predefined in the FORESCEM, allowing for generic and multiple possible study cases.

4.

The FORESCEM framework adopts an imbricated nomenclature, similar to those used in the Corine Land Cover and Urban Atlas classifications (which refer exclusively to nested land cover classes). The key difference is that the FORESCEM requires both land use classes and their associated land cover classes.

5.

Land cover and land use changes are modelled independently. This enables explicit consideration and control of natural vegetation dynamics and anthropogenic decisions. However, because they influence one another, their interactions are incorporated into the modelling framework. A land use change induces corresponding land cover changes, and conversely, land cover changes may trigger a land use change, for instance, in cases of abandonment.

While the first five principles constitute the foundational basis of the FORESCEM to simulate land change processes, modelling LUCC based on narratives co-constructed with stakeholders introduces two additional principles that underpin its conceptual and methodological design:

6.

Narratives provide information on the transition rules governing both LUCC, as well as on the interactions between them, enabling the simulation of complex and interdependent LUCC processes, as described in principle (5).

7.

Narratives may incorporate temporal variations in LUCC, including shifts in rates, directions, or durations, either in future land demand or in their spatial allocation’s drivers. These variations may diverge from historical trends (e.g., a slowdown in urban expansion), involve abrupt disruptions (e.g., the development of a new housing estate following planning decisions, or a spatially random land-use shift resulting from deforestation or reforestation), or arise from unexpected events (e.g., wildfires). In FORESCEM, deviations from LUCC trends are operationalized through three distinct mechanisms: events, policies and a scenario-based future land demand. Events represent sudden, discrete land changes. Policies, by contrast, represent broader spatial constraints that facilitate or restrict the allocation of specific land uses or land covers over the long term. The land demand is set a priori, may result from external (expert-based or modelling) approaches, and drives trend-breaking LUCC simulation. Such an approach is quite common nowadays, while the majority of published articles still use Markov-based trend land demand to simulate future LUCC as illustrated by [26].

3.1.2. Architecture and Functioning

The FORESCEM is built on a modular architecture and developed in Python to ensure flexibility and adaptability to multiple research contexts, and to future modelling needs. It integrates all the processing components necessary to perform LUCC simulations, including transition, calibration, and simulation modules (Figure 2). It is freely available for non-profit uses (https://gitlab.in2p3.fr/letg/forescem/forescem), accessed on 14 April 2026.

The transition module infers all land cover transitions based on two reference maps corresponding to different time points. When external information is not explicitly provided, future land cover demand can be estimated using a Markov chain approach. The transition module also assists users in identifying unrealistic or implausible transitions and in preparing the transition matrix used for simulation. For example, a direct transition from meadows to forest—observed over a long interval—may be flagged if it omits intermediate stages, such as shrub encroachment, which are required for ecological plausibility. The calibration module aims at providing suitability maps. There is no predefined limit on the number of drivers that can be considered. Logistic regression is employed to generate statistical models, as used by [26,32]. This method is widely adopted in well-established land change models [25,43] and facilitates the interpretation of the relative influence of the drivers. Finally, the simulation module loads all required inputs, including spatial masks, probability maps and policy and event layers, and executes the simulation. In this study, land cover allocation is performed by selecting, for each parcel, the land cover type that maximizes the computed probability while respecting the annual land demand and transition rules.

The simulation begins with an initialization phase during which all necessary data and configuration parameters are loaded, including land cover, land use, parcel age, policy settings, events and demand. The simulation then proceeds iteratively. At each iteration, target land areas are first identified based on current demand and the previous system state. A set of constraint masks is then applied to adjust the suitability maps, i.e., where future LUCC may or may not occur. The age mask enforces persistence or change according to parcel age. Land cover and land use masks constrain possible transformations based on preset transition matrices. Policy masks weight the probabilities to reflect policy influences, while event masks restrict potential changes to specific land covers and modulate demand accordingly. Policies may operate as soft constraints—for instance, by promoting the gradual relocation of certain crops or influencing settlement preferences—or as hard constraints, such as zoning regulations prohibiting urbanization outside designated areas. Events impose a specific land cover or land use on a targeted area and, where necessary, update the corresponding land cover or land use accordingly. For example, a silvicultural project (event) implemented on agricultural land forces the current land cover to shift to one associated with silvicultural use. However, if the same project is carried out in an abandoned area where the existing land cover (e.g., natural encroachment) is already compatible with silviculture, only the land use is updated. After probability adjustments, land cover allocation is required to satisfy the land demand. Allocation begins with non-competitive situations, in which only one land cover is possible, and proceeds to competitive cases involving multiple candidate land covers. Land uses are then assigned. For crop successions, for instance, a given crop cannot occur more than a predefined number of years for agronomic reasons (e.g., maintaining soil quality or preventing invasive species). Thus, a minimum and maximum duration (called “age”) is set for each land cover, following an approach similar to CLUE-S conversion rules [32]. Finally, parcel ages are updated: they are reset to one if a change has occurred or are incremented by one for unchanged land cover.

3.1.3. List of Input Data for FORESCEM

The FORESCEM requires a minimal input dataset to perform LUCC simulations effectively (

Table 2. Mandatory (M) and Optional (O) data in FORESCEM.

Data	Type	Mandatory (M)/Optional (O)	Description
Land cover at time t-n	Map	M	Land cover map used to compute past transitions, past trends, and the calibration.
Land cover at time t	Map	M	Land cover map used to compute past transitions, past trends, and the calibration. It is also the simulation’s starting landscape.
Land use at time t-n	Map	M	Land use map used to calibrate the model.
Land use at time t	Map	M	Land use map as starting date of the simulation.
Set of drivers	Maps	M	Maps used for the calibration and the simulation. Driver maps at time t + x can be included if drivers evolve during the simulation.
Cover duration table	Table	M	Duration for each cover that FORESCEM will intend to respect. In some situations, cover duration can be exceeded to find a solution.
Cover type	Table	M	Table defining for each land cover if demand is a constraint or not.
Drivers	Table	M	Table defining which driver needs to be used for each land cover’s calibration.
Demand	Table	M	Table used during the simulation and gathering annual demand for each land cover.
Parcels	Map	O	Simulation is processed at pixel unit if this map is not provided.
Land cover’s age at time t	Map	O	If no map is provided, a random age map respecting land cover duration is generated.
Policy	Map	O	Map used to influence land allocation according to a narrative. It can be a hard (binary) or soft map. Policy is used from the year for which it is defined. For example, urban sprawl policy.
Event	Map	O	Map used to create unexpected changes such as an urban project (housing estate, new lake), a forest project (new forest). It occurs from the date it is defined.

and Figure 2): two land cover maps (at time t-n and t0), two land use maps (at time t-n and t0), a set of driver maps, an initial age map for each land cover unit, a land cover duration table, a table linking drivers to land cover, a table linking land covers to land uses, a land cover priority table, and the annual land cover demand. To support scenario customization and enhance simulation realism, several optional inputs can also be integrated. These include updated drivers during the simulation period, event and policy maps for land uses and land covers, and a parcels map. Additionally, other optional data for customized land allocation functions can be added.

3.2. Study Site

The application and validation of the FORESCEM were conducted on the Couesnon catchment in Brittany, France. This case study covers an area of 1711 km², encompassing the catchment itself along with a 3 km buffer zone (Figure 3). This region represents a highly anthropogenic landscape dominated by intensive agricultural land uses, particularly dairy and livestock production, which occupy approximately 80% of the study area. The predominant land covers include grasslands, maize, and wheat, arranged within a characteristic bocage landscape [44]. Maize and wheat are typically found on flat or gently sloping areas with large plots and deep soils, whereas grasslands are more commonly situated on granitic plateaus, slopes with shallow soils, valley bottoms, or are integrated into crop rotations. Agricultural intensity varies spatially: areas with higher densities of hedgerows, smaller plots, and a greater prevalence of grasslands generally exhibit lower-intensity agricultural practices. In the northern part of the catchment, adjacent to the coast, a polder zone is primarily used for cash crops and vegetable production.

Forestry activities are limited (covering ~8% of the area) and are mostly confined to public forests. The principal urban centre, Fougères, is a medium-sized town located in the southeast of the catchment, with a population exceeding 50,000 inhabitants. Between 2006 and 2018, key land use trends included a significant expansion of urban areas (+2300 ha) and a slight reduction in agricultural land (−850 ha). Agricultural intensification was observed, marked by an increase in crop-dominated areas (+3100 ha, including +2300 ha of maize) at the expense of meadows (−4000 ha). The Couesnon site offers a diverse set of conditions well-suited for testing the capabilities of the FORESCEM. These include highly heterogeneous land cover mosaics, a wide range of land cover durations—from very short (1–2 years) to very long (>100 years)—and varying spatial structures. These characteristics provide a robust context to assess the model’s ability to represent complex landscape dynamics and simulate realistic land cover trajectories.

3.3. Dataset Description

Land cover maps were derived from remote sensing data. Multitemporal Sentinel-2 imagery was used to produce the 2018 land cover maps for both study sites. For historical land cover data, multitemporal Landsat archives were employed: 1990 and 2006. Supervised random forest classification was applied to enhance vegetation discrimination, achieving overall accuracies (OAs) exceeding 85%. Subsequent post-processing, informed by local expert knowledge, was conducted to correct minor misclassifications—particularly those resulting from inconsistent transitions between land covers caused by spatial and temporal resolution differences between the Sentinel-2 and Landsat datasets, both resampled at 10 m. Land use maps were generated using validated land cover maps, supplemented by local expert input and ancillary geographical data, such as areas of exploited forests.

The LUCC drivers incorporated into the model include elevation, slope, aspect, geological characteristics, travel time to major cities, and land planning maps of the Blue and Green Infrastructure Network (BGIN). Elevation, slope, and aspect were derived from Digital Elevation Models (DEMs), specifically the national BD ALTI 50 m dataset for France. All drivers were resampled to align with data consistency (10 × 10 m). Geological data were sourced from institutional databases and simplified to reduce the number of geological classes. Travel time to major urban centres was calculated using OpenStreetMap road network data. The BGIN maps, provided by local land planners, represent ecological corridors with continuous values ranging from 0 (absence of corridor) to 1 (highest corridor importance). Appendix A summarizes key model inputs and parameters: minimum and maximum land cover duration, drivers used for each land cover regression model, drivers to be used for each land use regression model, annual land cover demands, and possible transitions from one land cover to another.

Landscape patterns are incorporated through the use of rasterized cadastral parcels for urban and forested areas, and the Land Parcel Identification System (LPIS) parcels (data.europa.eu) for agricultural lands. These raster-based parcels aggregate contiguous 10 m resolution pixels that evolve as cohesive units within the model. Employing parcels instead of individual pixels enhances the representation of realistic landscape patterns in land change simulations [27,45]. The parcel count is approximately 190,000 for the Couesnon case study. All spatial data are pre-processed by aggregating median values within each parcel.

3.4. Validation Tests

Five distinct tests are therefore conducted to demonstrate the model’s capability to accurately reproduce historical LUCCs and to incorporate policies and trend-breaking assumptions present in future scenarios. The resulting simulations are benchmarked against a trend-based simulation calibrated on observed LUCC data from 2006 to 2018. The validation aims at emphasizing the FORESCEM’s adaptive capability in simulating diverse LUCC trajectories.

• Land cover duration. This feature is particularly important for land covers characterized by specific temporal dynamics, such as agricultural crop rotations. One crop may not occur more than a predefined number of years for agronomic reasons (maintaining soil quality, avoiding invasive species…). Thus, a minimum and maximum duration is set for each land cover, similarly to one of the decision rules of CLUE-S to avoid inconsistent changes [32]. Most commonly used land change models are unable to accurately simulate these types of land cover dynamics [29]. The FORESCEM’s capability to respect such constraints enables the simulation of realistic agricultural practices. For example, in the French case study with an intensive agricultural system, wheat cultivation is restricted to a maximum of two consecutive years on the same plot. Maize cultivation (1) generally does not exceed three consecutive years and (2) typically follows wheat. Temporary meadows have durations ranging from one to seven years. However, under certain conditions, the full set of constraints—including land use duration—may result in no feasible solution that satisfies all constraints simultaneously. In these cases, the FORESCEM is designed to override the duration constraints to ensure finding a consistent solution.
• Landscape dynamics. This test assesses the model’s ability to reproduce historical LUCC dynamics. A simulation is conducted over the period 2006–2018. The simulated land cover map is then compared to the observed classification derived from remote sensing data. Model performance is quantified using the figure of merit [46] and kappa indices, as defined by [25]. Kappa, kappa_histogram, kappa_location, kappa_simulation, kappa_transition, and kappa_transloc, respectively, indicate the similarity of the two compared maps, the quantitative similarity of the two compared maps, the similarity between both compared maps’ categories and their spatial allocation, the similarity between simulated and real transitions, the quantitative similarity between simulated and real transitions, and the similarity of spatial allocation between simulated and real transitions. To supplement these spatial metrics, the Shannon Diversity Index (H′) and its associated evenness are calculated to assess the model’s ability to replicate the thematic complexity and structural organization of the landscape [47]. This allows for an evaluation of the landscape’s compositional realism beyond pixel-to-pixel comparison. Analysis is conducted for all land cover classes separately, and for all crops reclassified as a single class, “agricultural land”, as crops rotations show a high degree of randomness influenced by individual farmers’ decisions.
• Unexpected land cover transitions resulting from land use changes. Land development projects, such as afforestation initiatives, can drive land cover changes that are not predicted by historical trends. To simulate such scenarios, a new silvicultural zone will be introduced in 2028, without prior information on expected land cover changes. The FORESCEM is designed to enforce these unexpected land cover transitions triggered by land use changes set as events in stakeholder narratives, while striving to maintain the overall proportions of initial land cover demand. For instance, transitions initially prohibited—such as from wheat land cover to recolonization land cover—are allowed to occur. Subsequently, the shrubland land cover changes to forest cover after 10 years, as defined by the configured land cover duration parameters.
• Variations in land cover demands. Compared to historical trends, a substantial shift in crop demand is anticipated following a reform of the Common Agricultural Policy. Based on a narrative simulating this policy shift in 2026, which triggers agricultural intensification (see [23] for details), this test examines a marked increase in wheat and maize at the expense of grasslands and other crops. During the calibration period, the respective annual increases for wheat and maize were +25.6 ha and +192.5 ha. Post-2026, these rates are projected to rise to +43.2 ha and +496.5 ha per year, respectively.
• Trend-breaking policies. Urban planning policies evolve over time. For instance, French urban planning instruments defined at the municipal level are expected to become more restrictive in the future to limit artificialization. This test assesses how the FORESCEM framework responds when areas previously designated as suitable for urban development become temporarily unavailable due to policy restrictions, and are later reopened following local policy changes. Specifically, this fifth test assesses how the model handles annual urban land demand when no suitable areas are available during a certain period, but new areas become accessible later.

4. Results

4.1. Test 1: Land Cover Duration and Crop Occurrences

We focus here on agricultural land covers—wheat, maize, and other crops, and permanent and temporary meadows—which exhibit specific temporal dynamics, including varying durations and frequencies of occurrence. The results demonstrate that the FORESCEM successfully respects land cover duration in most cases (>98% for all agricultural land covers; Figure 4, left column). Less than 2.5% of the total agricultural area exceeds the maximum duration threshold, while still meeting overall land cover demand. This level of non-compliance is particularly limited for temporary meadows (under 0.03% of their total area), wheat (less than 1%), and other crops (less than 2%). For maize, which accounts for more than 30% of the agricultural land use at the initial time step and over 35% by the final year, 3–6% of the area surpasses the three-year duration limit. However, in practice, maize can be cultivated on the same plot annually for 10 to 15 years in certain locations. This observed deviation highlights the influence of diverse and localized agricultural practices, which are challenging to model precisely. In essence, the generalization and simplification of land cover change rules in the FORESCEM can explain some localized discrepancies. With a shorter maximal duration allowed, it is increasingly difficult for the model to reconcile this constraint with annual agricultural land cover demands. For other crops, non-compliance is mainly concentrated in a specific region—namely, the polder areas north of Couesnon catchment—where pedological conditions favour more frequent transitions between wheat and other crops such as vegetables and orchards. When the duration threshold is relaxed by one year (e.g., allowing wheat to be maintained for up to four years), overall non-compliance drops to just 0.1% of the total agricultural area. The highest rate of non-compliance under this scenario is observed for wheat, reaching only 0.4% in the year 2043 (Figure 4, top right). This indicates that most violations stem from minor extensions in crop duration to maintain optimal land use efficiency on the most suitable soils.

4.2. Test 2: Landscape Dynamics

The simulated map exhibits some differences compared with the observed map derived from remote sensing (Appendix B). The figure of merit (FoM), defined as the proportion of correctly predicted changes relative to all observed and predicted changes (Figure 5), reaches 17.1%, which falls within the range of 1% to 59% reported in a meta-analysis of multiple land change models [46]. However, the figure of merit can fail to capture certain transitions, particularly those associated with crop rotations occurring within the simulation period. For instance, a sequence such as wheat–maize–wheat may be interpreted as persistence because the initial and final states are identical, despite intermediate changes. As a result, the FoM underestimates the actual level of change. When agricultural land covers involved in crop rotations are not treated as persistence, the FoM (FoM_rotation) increases up to 27.73%.

This diagnostic is further supported by the kappa indices (

Table 3. Kappa indices considering all land cover class (first column) and reclassified land cover (agricultural, forest, urban, and water classes).

Index	Meaning	Land Cover	Reclassified Land Cover
Kappa	Similarity of the two compared maps	0.38	0.85
Kappahistogram	Quantitative similarity of the two compared maps	0.97	0.97
Kappalocation	Similarity between both compared maps’ categories and their spatial allocation	0.38	0.87
Kappasimulation	Similarity between simulated and real transitions	0.06	0.19
Kappatransition	Quantitative similarity between simulated and real transitions	0.87	0.61
Kappatransloc	Similarity of spatial allocation between simulated and real transitions	0.08	0.31

). When all land cover classes are considered independently, the global correspondence remains low (kappa = 0.38), reflecting the noise introduced by the aforementioned crop rotations (

Table 4. Kappa index per land cover.

Land Cover	Kappa		Reclassified Land Cover
Broadleaf forest	1.00		Broadleaf forest
Mixed forest	0.71		Mixed forest
Shrubland	0.00		Shrubland
Water	0.92		Water
Urban	0.82		Urban
Wheat	0.08	0.86	Land covers as part of crop rotations
Maize	0.17
Other crops	0.09
Temporary meadow	0.14
Permanent meadow	0.76		Permanent meadow

). However, once these agricultural covers are grouped into a single class, the overall similarity increases significantly (kappa = 0.85). This improvement is driven by very high scores in both quantitative distribution (kappa_histogram = 0.97) and spatial allocation (kappa_location = 0.87). While the transition-specific metrics, such as kappa_simulation (0.19), appear more modest, they remain consistent with the FoM results. They highlight that while exact pixel-to-pixel transition modelling is challenging, the model effectively captures the broader spatial patterns and processes of the land system once rotational cycles are filtered out.

Beyond spatial allocation accuracy, the model’s ability to replicate the structural complexity of the landscape was assessed using the Shannon Diversity Index (H′). The observed map exhibits a Shannon index of 2.018 with an evenness of 0.876 across 10 land cover classes. The simulation replicates this systemic organization with remarkable fidelity, yielding an H′ of 1.990 and an evenness of 0.906. The negligible global difference (ΔH′ = −0.029) demonstrates that despite the challenges in fine-scale pixel allocation—as indicated by the 17.1% figure of merit—the model generates a landscape composition whose thematic richness and surface distribution are highly consistent with observed reality. This suggests that the model is able to simulate high compositional realism even though spatial allocation of crops remains highly uncertain.

4.3. Test 3: Unexpected Land Cover Transitions Resulting from Land Use Changes—The Influence of an Event

Figure 6 illustrates a scenario-based silvicultural project initiated in 2028. Prior to this intervention, urban sprawl and agricultural land cover changes follow their usual patterns. However, once the silvicultural land use project begins, urban expansion halts, and selected agricultural land covers are converted to shrubland—representing areas planted with young trees. In this scenario, a specific event triggers a local shift in land cover dynamics: for example, maize cultivated in 2027 transitions directly to shrubland in 2028, bypassing typical transitions such as grassland or wheat. This illustrates the FORESCEM’s ability to override predefined constraints in the transition matrix when necessary to accommodate scenario-based land use changes. The model adapts to event-driven shifts in land cover by prioritizing narrative-defined land use objectives, demonstrating its flexibility and responsiveness to local planning decisions.

4.4. Test 4: Variations in Land Cover Demands

The evolution of agricultural land covers closely aligns with changes introduced in 2026 (Figure 7). Areas cultivated with maize increase markedly, while wheat expands at a more moderate rate. By contrast, areas covered by other crops and permanent meadows experience a notable decline. The model adheres closely to the prescribed demand, with a maximum deviation of only 246 hectares—equivalent to 1.13% of the total study area—between simulated and expected land cover surfaces. These residual discrepancies are largely attributable to crop rotation constraints. At the beginning of the simulation, the initial age of all land covers is arbitrarily set to one year. During the first seven years, corresponding to the maximum duration allowed for permanent meadows, land cover options remain limited due to uniform age distributions and competition between covers. Beyond initialisation, increasing heterogeneity in parcel ages progressively relaxes these constraints, enabling greater flexibility to better match land cover demand.

4.5. Test 5: Trend-Breaking Policies

The FORESCEM adequately models the trickledown impacts of changes in a spatial planning policy (as developed in [23]), exemplified here with an evolving urbanization policy. In the control (reference) scenario, purple and blue areas are, respectively, available for urban development from 2018 to 2050 and from 2030 to 2050 (Figure 8). In the trend-breaking scenario, which reflects changes in urban policy over time, the purple areas likewise remain open to urbanization from 2018 to 2050. However, the blue areas become restricted after 2030, triggering further urbanization in new, more remote red areas. These red areas are located in municipalities that have become more attractive due to proximity to a major transportation axis. This test demonstrates the FORESCEM’s ability to manage spatial policy changes (opening and closing accessible areas) and to resolve resulting land use conflicts. The model ensures that new urban development does not occur in unauthorized areas while distinguishing between pre-existing urban land cover and potential expansion areas, as urban sprawl typically occurs adjacent to existing urban zones. Here, the FORESCEM’s ability to account for complex policy-driven dynamics relies on its unique consideration of the land use–land cover transition (Figure 2), and on the model-driven determination of land cover transitions based on current land use and on each parcel’s policy context.

5. Discussion

The transition from descriptive land cover mapping to prospective land system modelling requires a fundamental rethinking of how human–environment interactions are simulated. The FORESCEM is considered not merely as a spatial simulator, but as a conceptual bridge intended to reconnect the biophysical state of the land with the socio-political intent behind its management. The following discussion evaluates how this architecture addresses the current bottlenecks in LUCC science and why a departure from traditional validation metrics is essential for the credibility of future-oriented land narratives.

5.1. Addressing Current Challenges in LUCC Modelling: The FORESCEM Contribution

As highlighted by [48], land system science lacks an integrated theoretical framework, and middle-range theories offer a promising pathway to synthesize causal mechanisms across diverse land use change processes. The FORESCEM addresses this by moving away from purely probabilistic or historical trend-based transitions toward a logic-based architecture. While the FORESCEM aims to complete what is currently offered by LUCC models, and not to replace them, it emphasizes a transparent and traceable translation of stakeholder-informed narratives into model inputs.

Bridging the “Policy–Modelling” divide. A major challenge in spatial planning is the “translation” of fuzzy policy narratives into rigid spatial rules. By explicitly decoupling land use (human intent, management intensity, and socio-economic function) from land cover (biophysical state), the FORESCEM facilitates the integration of public policies as direct model inputs. This architecture responds to the need for “policy-screening” tools described by [49], allowing stakeholders to test “what-if” scenarios without being constrained by the historical path-dependency of combined LUCC classes. In our results, this is exemplified by the simulated transition from agricultural to silvicultural zones. While many traditional models rely on extrapolating historical trends where such a shift might be modelled as a slow marginalization process, the FORESCEM’s architecture is specifically designed to treat this as a deliberate land change event. By assigning a silvicultural “intent” to previously agricultural parcels, the model captures the abruptness of land system ruptures, such as large-scale afforestation for carbon sequestration or ecological restoration, which are often underestimated by models relying on incremental change [50]. Beyond its technical architecture, the FORESCEM serves as a strategic tool for bridging the persistent “research–action” gap in territorial planning. By providing a platform where researchers and practitioners can co-construct spatial rules, the model transforms from a closed simulation engine into an adaptable interface for transdisciplinary collaboration. This approach addresses a major bottleneck in geographical decision-support systems: the need for tools that are sophisticated enough to handle multi-scale trade-offs, yet transparent enough to be integrated into democratic planning processes. Through its “object-oriented” realism and the use of familiar spatial units like cadastral plots, the FORESCEM offers a visual and conceptual grammar that planning decision-makers can intuitively grasp.

Capturing Human–Environment Feedback through Object-Oriented Realism. Most LUCC models struggle to capture the non-linear “shocks” of human decisions, such as sudden administrative reclassifications or abrupt Nature-based Solution implementations. The FORESCEM’s ability to handle disruptive changes—such as rapid afforestation mandates or urban planning zoning—addresses the need for models that can simulate “regime shifts” rather than just “incremental” change. The use of parcel-based elementary units (rather than pixels) is a deliberate choice to align with the “object-oriented” turn in geography. By simulating changes on units that correspond to actual decision-making levels (cadastral plots and farm holdings), the model maintains structural coherence—preserving field shapes, road networks, and riparian buffers—which is recognizable to local actors. This spatial realism is not merely esthetic; it is a pillar of salience and legitimacy, ensuring that the model acts as an effective interface between science and policy [51].

Limitations: Local land uses vs. exogenous pressures. While the FORESCEM excels at representing local land uses and parcel-level transitions, it is important to acknowledge the influence of exogenous drivers. Local land use decisions are frequently constrained by external pressures—such as global commodity prices or European-wide directives like the Common Agricultural Policy (CAP). While the FORESCEM allows for downscaling these drivers into local rules, the feedback loop remains a challenge. Future iterations could explore how localized “territorial ruptures” (like the ones simulated in our study) might aggregate to influence broader system dynamics, moving toward a more integrated multi-scale framework, another challenge identified by [17].

5.2. Shifting the Validation Paradigm: From Predictive Accuracy to “Process-Based” Credibility

Building upon the two-pronged assessment presented in our results, we argue that the integration of exploratory narratives necessitates a broader shift in how we evaluate LUCC models. Traditional validation, relying on “goodness-of-fit” metrics like the figure of merit (FoM), kappa indices, or ROC curves, is increasingly recognized as an epistemological mismatch for prospective modelling. As argued by [52], these indices often conflict with different types of errors (quantity vs. location) and, crucially, are rooted in the assumption that the past is a reliable proxy for the future.

The “stationarity trap” and the fallacy of historical fit. When a model is optimized to maximize historical fit (hindcasting), it inherently reinforces path dependency. A model that perfectly replicates past trends is often structurally incapable of simulating “structural breaks”, the very essence of disruptive scenarios [25]. This is because traditional validation metrics reward models that maintain the status quo rather than those capable of representing regime shifts. By moving away from a strict pixel-to-pixel comparison, the FORESCEM avoids this “stationarity trap.” We argue that for prospective narratives, predictive accuracy is an insufficient and potentially misleading benchmark. A model that “fails” to replicate past LUCC might be capable of considering a future where innovative and transformative policies are enforced.

Validation as “multi-dimensional confidence building”. We propose a validation framework based on three pillars of credibility, moving from statistical to heuristic validity:

• Structural and logic validation. Instead of comparing outcomes, this approach assesses whether the model’s internal mechanism (e.g., land cover duration, soil suitability, and policy constraints) correctly represents the causal drivers of change. This aligns with pattern-oriented modelling [53], where the goal is to reproduce the internal logic of the system rather than a single historical trajectory.
• Spatial syntax and landscape heterogeneity. At fine scales, the exact location of a simulated change is often stochastic and almost impossible to predict. However, spatial configuration and compositional diversity must remain realistic to ensure ecological and functional coherence. By utilizing Shannon’s Diversity Index (SHDI), we validate that the FORESCEM preserves the landscape’s heterogeneity and evenness. A stable SHDI across simulations demonstrates that the model does not lead to an artificial homogenization of the territory but instead maintains the complex “spatial grammar” inherent to the study area. To complement this validation, grounded in landscape ecology metrics, the integration of Intensity Analysis [54] allows us to verify if the “pace” of these diversity shifts is consistent with known transition regimes, providing a dual validation of both landscape pattern and process.
• Heuristic and Social Validation. In the context of participatory modelling, the ultimate test of a model is its face validity among stakeholders. As [55] emphasizes, the model acts as a “boundary object.” This confirms the effectiveness of the Story-And-Simulation (SAS) approach outlined in our Introduction, where the model succeeds not by “predicting” the future, but by providing plausible and consistent trajectories of future land systems. The fact that stakeholders in our study challenged the underlying assumptions (e.g., urban demand) rather than the spatial patterns indicates that the FORESCEM possesses high legitimacy [23]. Moreover, this social validation was a prerequisite for considering LUCC scenarios as potential cognitive, instrumental or political outcomes that may lead to decision-making [56,57].

Addressing the data–model gap. We acknowledge that the sensitivity of fine-scale models to data availability (e.g., high-resolution cadastral data and farm-level management records) remains a significant hurdle. In many regions, the “human agency” layer is obscured by data privacy or a lack of digital registries. However, as [15] points out, the frontier of LUCC modelling lies in capturing these organizational levels (municipalities and farm-holdings). The lower scores on standard statistical indicators are often not a reflection of model failure, but of the inherent aleatory uncertainty of human-driven systems at fine scales. Moving forward, the validation of such models should prioritize their ability to explore the uncertainty space—through global sensitivity analysis—rather than converging on a single, “accurate” but potentially biased future.