Urban Growth: CA Models of City Expansion
When the first Landsat satellite began returning images of the Earth’s surface in 1972, researchers gained something that had not previously existed: a time-lapse record of how cities grow. Over the following decades, as the satellite archive accumulated, a pattern became clear that was simultaneously obvious and, in its formal implications, remarkable.
Cities expand the way cellular automata evolve.
Growth happens at the edge, where developed and undeveloped land meet. It clusters around roads and highways, following the topology of the transportation network. Islands of development appear in the countryside — a shopping center here, a subdivision there — and over time the space between them fills in. The urban boundary is fractal: rough at the scale of neighborhoods, rough at the scale of blocks, with similar texture at each scale. The density gradient from city center to exurb follows a smooth exponential function that emerges from the growth dynamics rather than from any planning rule.
None of this was obvious before the satellite imagery made it visible. And the CA connection was not obvious to urban researchers until several theorists in the late 1980s and early 1990s began asking: what rule, applied locally, produces this pattern?
Why Cities Look Like Cellular Automata
Before examining the models, it is worth being precise about why the cellular automaton framework is appropriate here, rather than merely appealing.
A city, at the scale of individual land parcels, is a grid. Each parcel has a use: residential, commercial, industrial, agricultural, open space, transportation. The use of each parcel changes over time, and the probability of a change depends heavily on the current use of neighboring parcels. A parcel adjacent to an existing residential area is more likely to become residential than a parcel surrounded by farmland. A commercial parcel adjacent to a highway is more likely to expand than one in a residential neighborhood without road access. A parcel that is already developed in multiple adjacent directions fills in at higher priority than one with a single developed neighbor.
This is the CA update structure: the future state of each cell depends on the current state of its neighborhood. The grid is geographic space. The rule is the land-use transition probability. The update is the annual or decadal change in land use across the metropolitan area. The simulation runs forward in time.
This is not an analogy. It is a direct implementation of the CA formalism in a real geographic system. The question is not whether the CA framework is appropriate but which CA rule best reproduces observed urban dynamics.
White and Engelen: Fractal Urban Form
The first rigorous CA model of urban land-use change was published by Roger White and Guy Engelen in 1993: “Cellular Automata and Fractal Urban Form: A Cellular Modelling Approach to the Evolution of Urban Land-Use Patterns,” in Environment and Planning A (volume 25, pages 1175–1199).
White and Engelen’s model worked on a grid of cells, each assigned a land-use type. The transition probability for each cell was determined by a weighted sum of the land-use types of neighboring cells at several distances, with weights that varied by land-use pair. Residential use, for example, attracted more residential use nearby, repelled industrial use, and had complex relationships with commercial use (attraction at some scales, competition at others). The model included an element of stochasticity — the transition probability was not deterministic — which allowed the simulation to produce the irregular, quasi-random boundary texture that real cities exhibit.
The key result was that the model produced fractal land-use boundaries. The fractal dimension of the simulated urban boundary matched the fractal dimension measured from real aerial photographs of actual cities — typically around 1.6 for the urban perimeter. This is not merely a qualitative similarity; it is a quantitative match. The model was producing the correct statistical structure of the boundary at multiple scales.
White and Engelen also showed that the CA model produced a hierarchical spatial structure in land use — a gradient from high-density urban core through medium-density suburban zones to low-density exurban fringe — without any explicit rule encoding that hierarchy. The hierarchy was an emergent property of the local transition probabilities. This is the CA pattern formation principle operating in urban space: local rules, global structure, no blueprint required.
Batty and Xie: From Cells to Historical Cities
Michael Batty — who would go on to direct the Centre for Advanced Spatial Analysis at University College London and become arguably the leading theoretical urbanist of his generation — and his collaborator Yichun Xie approached the urban CA problem from a different direction in their 1994 paper “From Cells to Cities,” published in Environment and Planning B (volume 21, pages 31–48).
Where White and Engelen were primarily interested in the dynamics of contemporary land-use change, Batty and Xie asked whether CA could reproduce the historical development of specific cities. Their chosen case was Savannah, Georgia — a city with an unusual and well-documented planning history.
Savannah was founded in 1733 by James Oglethorpe on a grid plan that is, literally, a cellular automaton in physical space. The plan consists of wards, each containing a central square surrounded by residential trust lots and flanked by commercial lots. As the city grew, new wards were added to the grid, each identical in structure to the others. The rule for growth was explicit: add a complete ward when the existing city has grown to the boundary of the current configuration.
Batty and Xie showed that this growth rule — in its CA formalization — reproduced the development pattern of Savannah’s urban morphology through its period of major growth in the 18th and 19th centuries. The spatial structure of the city, the relationship between wards, the density distribution across the urban area: all emerged from the local CA rule encoding Oglethorpe’s planning principle.
The Savannah case is unusual in having an explicitly designed cellular structure — it makes the CA interpretation almost too easy. But Batty and Xie also applied their model to Amherst, New York, an unplanned suburb of Buffalo, and achieved similar fidelity. The emergence of an organized urban structure from a local growth rule is not confined to planned cities; it appears to be a general feature of urban dynamics.
SLEUTH: The Model That Ran for a Century
The most ambitious and most widely applied urban CA model is SLEUTH, developed by Keith Clarke at the University of California, Santa Barbara, and Leonard Gaydos at the Western Ecological Research Center of the US Geological Survey. Their foundational paper, “Loose-Coupling a Cellular Automaton Model and GIS: Long-Term Urban Growth Prediction for San Francisco and Washington/Baltimore,” was published in the International Journal of Geographical Information Science in 1998 (volume 12, pages 699–714).
SLEUTH is an acronym for the input data layers the model requires: Slope, Land Use, Exclusion (areas protected from development), Urban Extent (the historical record of urbanized area), Transportation (road networks), and Hillshade (a topographic layer used for visualization). The model is coupled to a Geographic Information System (GIS), which means it uses real geographically-referenced data for all its inputs and produces outputs in the same geographic coordinate system.
The CA rule encodes four types of growth behavior:
Spontaneous growth. A cell randomly selected anywhere in the domain is urbanized with a probability proportional to the inverse of the local slope. This models isolated new development — a farmhouse converted to a subdivision, a rural commercial node. The equivalent in Life is a spontaneously born cell in an otherwise empty region.
New spreading center. A recently spontaneously urbanized cell grows outward, recruiting adjacent cells. This models the expansion of a new development node before it connects to the main urban fabric.
Edge growth. Cells adjacent to already-urbanized cells are urbanized with a probability that increases with the number of already-urbanized neighbors. This is the primary growth mechanism — the spreading of urban development from the existing urban frontier. It is the direct analogue of Life’s birth rule.
Road-influenced growth. Cells near transportation corridors are urbanized at higher rates, and new spontaneous growth can “search” along road networks to identify favorable distant locations. Roads act as attractors in the development landscape, channeling growth along transportation corridors.
These four mechanisms, operating simultaneously at every cell on the grid, produce the characteristic pattern of urban growth: concentrated expansion at the urban fringe, corridors of development along roads, and scattered exurban growth that gradually fills in. The five parameters controlling the relative weights of each mechanism (diffusion, breed, spread, slope resistance, road gravity) were calibrated against historical satellite imagery of each modeled city.
Clarke and Gaydos validated the model against the San Francisco Bay Area — one of the most extensively documented urban growth cases available — and against the Washington-Baltimore corridor. In both cases, the model reproduced the observed 1990 urban extent accurately when calibrated to historical data from 1900 to 1980, and then used to project from 1980 to 1990. More ambitiously, they projected San Francisco Bay Area growth forward to 2050 and 2100, producing maps of expected urbanization that have been used in regional planning ever since.
What SLEUTH Predicts and What It Gets Wrong
SLEUTH’s accuracy in reproducing historical growth patterns is genuinely impressive. Calibrated to historical data, it can reproduce the spatial pattern of urban expansion with correlations of 0.85 to 0.95 between simulated and observed urban extent. For a model with only five parameters, calibrated to only a few decades of historical data, this is a striking result.
But the model has systematic failures that reveal important limitations of the CA approach to urban modeling.
It ignores economics. Urban growth is driven by land prices, construction costs, income levels, and employment location — none of which appear in SLEUTH. The model produces the spatial pattern of growth but cannot explain why growth happens where it does in economic terms. A zone predicted by SLEUTH to remain undeveloped might develop rapidly if a major employer opens nearby; the model has no mechanism to capture this.
It ignores policy. Zoning regulations, urban growth boundaries, infrastructure investment, and public policy all shape urban development patterns significantly. SLEUTH encodes only the exclusion layer (areas legally protected from development) and has no mechanism for modeling the effect of policy changes. This limits its usefulness for evaluating the effect of proposed interventions.
It is calibrated to extrapolation, not structural change. SLEUTH learns from historical patterns and projects them forward. When the structural conditions that generated those patterns change — a new highway is built, an industry collapses, remote work becomes common — the model’s projections become unreliable. The model cannot anticipate structural breaks.
It lacks feedback. Real urban development is not a one-way process: development changes land prices, which changes the incentives for further development, which changes the pattern. SLEUTH does not model this feedback loop.
These failures are not unique to SLEUTH. They are general limitations of the CA approach to urban modeling. The CA framework captures the spatial dynamics of growth propagation well. It captures the economic, institutional, and feedback dynamics poorly or not at all. The practical approach in contemporary urban modeling is to couple CA with economic models and policy modules, using the CA component for its spatial strengths and supplementing it with mechanisms the CA framework cannot represent.
Real Cities in the Model
SLEUTH and related models have been applied to dozens of cities worldwide, and the results have been used in actual planning decisions.
San Francisco Bay Area. The original Clarke-Gaydos application. The model projected that under historical growth rates, the Bay Area would be essentially fully urbanized by 2050, with remaining open space confined to protected parkland. This projection influenced regional planning discussions about urban growth boundaries.
Washington-Baltimore corridor. The second original application. The model showed the corridor developing into a continuous megalopolis, with growth connecting along the Interstate 95 corridor.
Lisbon and Porto. The model was calibrated to the Portuguese coastal cities, showing good correspondence between simulated and observed growth patterns and demonstrating the model’s transferability to European urban forms.
Chinese cities. Applications to rapidly growing Chinese cities like Guangzhou and Shenzhen have tested the model under conditions of much faster growth than the US cases for which it was originally calibrated. The results are mixed: the model captures spatial patterns reasonably but is poorly calibrated for the pace of Chinese urbanization.
The pattern across these applications is consistent: the CA model performs well on slow to moderate organic growth in developed economies. It performs less well when growth is rapid (the calibration procedure cannot keep up), when policy drives spatial structure (the model has no policy mechanism), or when the transportation network changes fundamentally (road-influenced growth requires a stable network to extrapolate from).
After SLEUTH: Current Urban CA Research
The generation of urban CA models that followed SLEUTH addressed its limitations in several directions.
Fuzzy CA. Rather than discrete land-use types, fuzzy CA models assign each cell a continuous membership in multiple land-use categories, allowing the model to represent the mixed-use reality of contemporary urban areas more accurately.
Agent-based coupling. Coupling the CA model with an agent-based model of household location decisions, firm behavior, and developer activity adds the economic and behavioral mechanisms that pure CA models lack. The CA provides the spatial propagation dynamics; the agents provide the economic drivers.
Machine learning calibration. Training CA transition rules using machine learning (rather than the manual parameter search of SLEUTH) produces models that better capture the spatial heterogeneity of transition probabilities. Different parts of the metropolitan area have different growth dynamics; machine-learning-calibrated models can reflect this.
Climate and hazard integration. Models that incorporate flood risk, wildfire risk, and sea-level rise into the CA transition probabilities allow projections of how climate change will reshape urban growth patterns over the coming decades.
The field has moved well beyond SLEUTH, but Clarke’s 1998 model remains widely used, widely taught, and the standard reference point for new urban CA research. Its strength — simple, transparent, calibrated to real geographic data — is also its limitation. The transparency that makes it easy to understand is the same feature that prevents it from capturing the full complexity of urban dynamics.
The city is a cellular automaton. But it is a very complicated one.