Skip to content
Back to search results

Land-Use Change Modeling

Introduced by
Peter H. Verburg and colleagues (CLUE-S); broader land-change-science community
Year
2002
Type
Family of spatially explicit models simulating land-use and land-cover change
Family
Land-use simulation / land-change science
Font size

What does it do?

Land-use change modeling is the umbrella family of methods that simulate how the land surface is converted between uses — forest to farmland, farmland to city — by combining where change is likely with how much change is demanded.

A typical model statistically relates observed change to spatial drivers such as slope, roads, and population, sets future demand for each land-use class from scenarios, and then allocates that demand across space to the most suitable cells, iterating until supply meets demand. The CLUE-S model of Verburg and colleagues, alongside the Land Change Modeler and SLEUTH, exemplifies this demand-plus-allocation architecture that underpins much of land-change science.

The intuition behind the method

Think of land-use change as solving two questions at once: how much and where. The 'how much' is set from the top down — a region might need ten thousand new hectares of urban land this decade because its population is growing. The 'where' is decided from the bottom up — each location has a suitability for becoming urban based on its slope, its distance to roads, and what surrounds it. The model behaves like an auction: it ranks all locations by suitability for each competing use and hands the demanded quantity to the best candidates, adjusting prices when several uses want the same prime land, until every parcel finds a use and every demand is satisfied. The result is a plausible future map, and by changing the demand or the rules you can explore alternative futures.

History

Land-use change modeling matured in the 1990s as land-change science emerged at the intersection of remote sensing, GIS, and the human dimensions of global environmental change, organized internationally through the LUCC research programme. Distinct traditions converged: cellular automata and the SLEUTH urban growth model from Keith Clarke, Markov transition approaches, and statistical-allocation models, most influentially the CLUE and CLUE-S models of Peter Verburg and colleagues, whose 2002 paper formalized the demand-plus-suitability-allocation architecture for regional land use. Commercial and open tools followed — the Land Change Modeler in IDRISI/TerrSet, GEOMOD, Dinamica EGO, and FUTURES — and the community established rigorous validation norms, notably Robert Gilmore Pontius's figure-of-merit. The field now increasingly incorporates machine learning and agent-based components while retaining the core separation of demand and allocation.

When to use it

Use land-use change modeling when you need spatially explicit projections of how a landscape will be reconfigured under different scenarios, and you have historical land-use maps plus driver layers to calibrate suitability. It is the right tool for assessing land-use policy, biodiversity and carbon impacts, and the downstream effects of urbanization or agricultural expansion, and for translating socio-economic scenarios into maps. It is less appropriate when only aggregate areas are needed (a Markov model suffices), when the change process is dominated by individual decision-makers and markets better captured by agent-based or economic models, or when no historical change data exist to calibrate and validate the suitability and allocation rules. The approach also assumes the drivers of past change will continue to operate, so abrupt regime shifts challenge it.

Limitations and considerations

  • Assumes the statistical relationships between drivers and change are stationary, so regime shifts and new drivers break it.
  • Demand is exogenous, so the realism of any projection is only as good as the scenario assumptions fed into it.
  • Calibration and validation demand high-quality, consistently classified multi-date maps that are often unavailable.
  • Aggregating to a single resolution invites the modifiable areal unit problem and can obscure fine-scale processes.

Source conditions do not replace independent scientific review or validation by an analysis engine.

Strengths

  • Cleanly separates how much change (demand) from where (allocation), making scenario analysis flexible and transparent.
  • Spatially explicit and driver-based, linking land-use outcomes to interpretable factors like roads, slope, and policy.
  • A mature, well-tooled family (CLUE-S, Land Change Modeler, SLEUTH) with established calibration and validation practice.
  • Integrates readily with downstream environmental, hydrological, carbon, and biodiversity assessment models.

Applications

  • Projecting urban expansion and agricultural frontier change under socio-economic scenarios.
  • Assessing deforestation, habitat loss, and biodiversity impacts of alternative land policies.
  • Estimating land-use contributions to carbon emissions and ecosystem-service change.
  • Informing regional and national spatial planning and conservation prioritization.
  • Supplying land-use trajectories as inputs to climate, hydrological, and integrated assessment models.

How it works

How it works

1

The first step quantifies where each land use tends to occur by relating the observed presence of a class to spatial driver variables, most commonly through logistic regression in which the log-odds of a cell belonging to a class is a linear function of predictors X such as elevation, slope, distance to roads, soil quality, and population density. The fitted coefficients β reveal which factors attract or repel each land use, and the model produces a suitability or probability surface for every class. These surfaces are the spatial backbone that tells the allocation step which locations are most fit for each use.

Formula
\operatorname{logit}(p_i) = \beta_0 + \sum_k \beta_k X_{ik}
2

Independently of where change can happen, the model needs to know how much of each land-use class the region will require at each future time step. This demand is set exogenously — from trend extrapolation, economic and demographic projections, or explicit policy scenarios such as a biofuel mandate or a conservation target. Separating demand from allocation is the architectural hallmark of this model family, because it lets analysts swap in different futures while holding the spatial logic fixed.

Formula
D_k(t) = \text{area demanded for class } k \text{ from scenario or extrapolation}
3

The allocation procedure reconciles the top-down demand with the bottom-up suitability by iteratively assigning land uses to cells so that each demand is met. Cells are awarded to the use for which they are most suitable, subject to neighbourhood effects, conversion rules, and hard constraints such as protected areas, and a competition mechanism resolves cases where several uses covet the same cell. The procedure repeats, adjusting per-class competitive advantage, until the allocated areas match the demanded areas across the whole region, producing the simulated land-use map for that step.

Formula
\text{allocate } D_k \text{ to cells maximizing suitability} + \text{neighbourhood} + \text{constraints}
4

Because the model produces a predicted map, validation compares it against an independently observed map for the same date. Common metrics include the ROC/AUC for the suitability surfaces, the kappa coefficient for overall agreement, and the figure-of-merit, which focuses on the agreement among the cells that actually changed and is preferred precisely because it is not inflated by the large stable background. A null model of persistence is the honest benchmark: the simulation is only useful if it predicts change better than assuming nothing changes.

Formula
\text{ROC/AUC},\ \text{figure-of-merit},\ \kappa \;\text{(simulated vs observed)}

Common pitfalls

  • Reporting overall kappa instead of a change-focused figure-of-merit, overstating accuracy because most cells never change.
  • Failing to benchmark against a no-change null model, so a mediocre simulation looks impressive in absolute terms.
  • Calibrating drivers and demand on the same period used for validation, inflating apparent predictive skill.
  • Treating a single scenario projection as a forecast rather than one of several plausible, assumption-laden futures.

Frequently asked questions

How does land-use change modeling relate to cellular automata and Markov land-use models?

Cellular automata and Markov models are specific techniques within the broader land-use change modeling family. A Markov model supplies the 'how much' through transition probabilities, a cellular automaton supplies the 'where' through neighbourhood rules, and many operational models combine both. The statistical-allocation models such as CLUE-S take a different route to the same two questions, using logistic regression for suitability and an iterative competitive allocation for placement. They are alternative engines that share the demand-plus-allocation logic that defines the family.

Why is the figure-of-merit preferred over overall accuracy or kappa?

In most landscapes the vast majority of cells do not change between dates, so a model that simply predicts persistence everywhere scores a high overall accuracy and kappa while predicting nothing useful. The figure-of-merit isolates the cells that actually changed and measures the overlap between predicted and observed change there, so it is not inflated by the stable background. Reporting it alongside the components of agreement and disagreement, as Pontius advocates, gives an honest picture of whether the model genuinely anticipates change.

What is the difference between demand and allocation in these models?

Demand is the exogenous quantity of each land-use class the region must contain at each time step, derived from scenarios, projections, or extrapolation, and it answers how much change occurs. Allocation is the spatial procedure that decides which specific cells take on each use to satisfy that demand, driven by suitability, neighbourhood effects, and constraints, and it answers where change occurs. Keeping the two separate is what lets analysts hold the spatial logic constant while exploring many different demand scenarios, which is the central strength of the approach.

Related methods

Sources

  1. Verburg, P. H., Soepboer, W., Veldkamp, A., Limpiada, R., Espaldon, V., & Mastura, S. S. A. (2002). Modeling the spatial dynamics of regional land use: the CLUE-S model. Environmental Management, 30(3), 391–405.

    Open DOI reference
    Copied.