Markov Land-Use Model
- Introduced by
- Mark R. Muller & John Middleton
- Year
- 1994
- Type
- Stochastic projection of land-use/land-cover areas using a transition probability matrix
- Family
- Markov chain / land-use simulation
What does it do?
A Markov land-use model treats land-use and land-cover change as a stochastic process in which the area in each class evolves according to fixed probabilities of transitioning from one class to another between time steps.
Estimated from two dated maps as a transition probability matrix, it projects how much of the landscape will convert from, say, forest to cropland or cropland to urban, assuming the future obeys the same transition tendencies as the recent past. Introduced to landscape ecology by Muller and Middleton in 1994, it is most powerful when coupled with a cellular automaton — the CA-Markov framework — that decides where, not just how much, change occurs.
The intuition behind the method
Picture every parcel of land as a token sitting on a square labelled with its current use, and a set of dice that tells each token its chances of staying put or hopping to another square by next year. If history shows that 5 percent of forest becomes farmland and 1 percent of farmland becomes city each decade, those rates become the dice. Roll the whole landscape forward one step and you get next decade's mix of land uses; roll it many steps and the mix drifts toward a long-run equilibrium dictated entirely by the dice. The Markov model is exactly this bookkeeping of transition odds — it answers 'how much of each use will there be?' but, on its own, stays silent about which particular parcels move.
History
Markov chains, formalized by Andrey Markov in the early twentieth century, were applied to vegetation and succession dynamics in ecology from the 1960s and 1970s. Muller and Middleton's 1994 study of the Niagara Region in Ontario is a widely cited landmark that demonstrated a Markov model of land-use change dynamics from sequential maps, establishing the transition-matrix approach in landscape ecology and land-change science. Through the late 1990s and 2000s the method was integrated with cellular automata to overcome its aspatial nature, most prominently in the CA-Markov module of the IDRISI/TerrSet GIS, which made the combined model accessible to a broad community of geographers and ecologists. It remains a standard baseline against which more elaborate driver-based and machine-learning land-change models are compared.
When to use it
Use a Markov land-use model when you have two or more consistently classified land-cover maps and want a quick, transparent projection of how much land will be in each class in the future, or to characterize the long-run equilibrium implied by recent change. It is ideal for first-cut trend extrapolation, for comparing the dynamics of different regions through their transition matrices, and, in its CA-Markov form, for producing spatial change maps. It is less appropriate when transition probabilities are clearly changing over time (non-stationary drivers), when conversions depend strongly on the location's attributes rather than just its current class, or when the underlying process is driven by individual decisions or policy shocks that a memoryless first-order chain cannot capture. In those cases driver-based statistical or agent models are preferable.
Limitations and considerations
- Assumes stationary transition probabilities, so it cannot anticipate changing drivers, policies, or feedbacks.
- First-order and memoryless: the next state depends only on the current class, ignoring history and trajectory.
- Aspatial on its own — it predicts areas but not where change occurs without an added allocation model.
- Sensitive to the length of the calibration interval and to classification or registration errors between the two maps.
Source conditions do not replace independent scientific review or validation by an analysis engine.
Strengths
- Simple, transparent, and data-light: a single pair of maps yields a full transition matrix and projections.
- Quantifies the magnitude and direction of change between every pair of land-use classes in one interpretable object.
- Yields an analytic long-run steady state, revealing the equilibrium toward which the landscape is tending.
- Couples cleanly with cellular automata (CA-Markov) to add realistic spatial allocation to the quantity projection.
Applications
- Projecting future land-use composition for regional and watershed planning from sequential satellite maps.
- Quantifying rates of deforestation, agricultural expansion, and urbanization between mapping dates.
- Generating spatial land-change scenarios via CA-Markov for environmental impact assessment.
- Characterizing and comparing landscape dynamics across regions through their transition matrices.
- Providing baseline change projections as inputs to hydrological, carbon, and habitat models.
How it works
How it works
The core of the method is a transition probability matrix P whose entry P_ij is the probability that land in class i at one date is in class j at the next date. It is estimated empirically by cross-tabulating two co-registered land-use maps and dividing each cell of the resulting change matrix by its row total, so that every row sums to one. This matrix encodes the recent dynamics of the landscape — how persistent each class is and which conversions are common — and is the single object from which all projections flow.
Formula
P_{ij} = \Pr(\text{state } j \text{ at } t+1 \mid \text{state } i \text{ at } t),\quad \sum_j P_{ij} = 1Future class proportions are obtained by multiplying the current state vector π_t, which lists the share of the landscape in each class, by the transition matrix: π_{t+1} = π_t P. Iterating this n times, equivalently raising the matrix to the n-th power (π_{t+n} = π_t P^n), projects the land-use composition arbitrarily far ahead. The projection is purely about quantities — total areas per class — and assumes the transition probabilities stay constant over the projection horizon.
Formula
\pi_{t+1} = \pi_t P,\qquad \pi_{t+n} = \pi_t P^{\,n}Because repeated multiplication by a regular transition matrix converges, the chain approaches a stationary distribution π* satisfying π* = π* P, the long-run equilibrium mix of land uses implied by the estimated dynamics. Examining this steady state reveals where the landscape is heading if current transition tendencies persist indefinitely, and how far the present composition is from that attractor. It is a useful diagnostic even when the constant-probability assumption is too strong to trust for literal long-range forecasting.
Formula
\pi^* = \pi^* P
Because the Markov chain alone is aspatial — it gives areas but not locations — it is routinely combined with a cellular automaton in the CA-Markov approach. The Markov component determines how many cells should change to each class, and the CA component allocates those changes spatially using neighbourhood contiguity and suitability surfaces, so that new urban cells, for example, cluster near existing development. This division of labour pairs the Markov model's quantitative rigor with the CA's ability to reproduce realistic spatial pattern.
Formula
\text{Markov: how much} \;\Rightarrow\; \text{CA: where (neighbourhood + suitability)}Common pitfalls
- Estimating the matrix over one interval and projecting many steps ahead, where small errors compound and the stationarity assumption fails.
- Ignoring rows that sum to slightly off one because of map misregistration, which biases every projection.
- Treating the convergent steady state as a literal forecast rather than a what-if equilibrium under constant dynamics.
- Mismatching the projection time step to the interval between the calibration maps, which silently rescales the transition rates.
Frequently asked questions
What does it mean that the Markov land-use model is first-order and memoryless?
First-order and memoryless mean that the probability of a parcel's next land use depends only on its current land use, not on how it arrived there or how long it has held that state. A forest pixel has the same transition odds whether it has been forest for one year or fifty. This Markov property keeps the model simple and estimable from just two maps, but it cannot represent processes where the trajectory or age of a land use matters — for those, higher-order or semi-Markov extensions are required.
How does the Markov model relate to the CA-Markov model and to cellular automata urban models?
The plain Markov model predicts only how much land changes to each class, with no spatial detail. CA-Markov keeps that quantity prediction and hands the spatial allocation to a cellular automaton that uses neighbourhood and suitability rules to decide which cells change, producing a map. A cellular automata urban model is the spatial engine in this pairing, specialized to cities; the Markov chain supplies its demand for how many cells of each use to allocate at each step. The two methods are thus complementary halves of the same modelling workflow.
What is the steady state of a Markov land-use model and is it a forecast?
The steady state π* is the stationary distribution satisfying π* = π* P, the long-run land-use mix the chain converges to if its transition probabilities never change. It is best read as a diagnostic — the equilibrium implied by current dynamics — rather than a literal forecast, because real transition probabilities shift as drivers, policies, and economies evolve. It usefully shows the direction and ultimate tendency of change, but trusting it as a far-future prediction requires the strong and usually unrealistic assumption of perpetual stationarity.
Related methods
Sources
Muller, M. R., & Middleton, J. (1994). A Markov model of land-use change dynamics in the Niagara Region, Ontario, Canada. Landscape Ecology, 9(2), 151–157.
Open DOI referenceCopied.