Density is the most quoted number in urban analysis and one of the least specified. A city described as dense might have crowded dwellings and low buildings, tall buildings and small households, or a compact footprint surrounded by nothing. These are different physical situations, they have different implications for transport, service provision, and risk, and they all produce a number with the word density attached.
The confusion is not academic. Density figures drive planning decisions, infrastructure sizing, retail location models, and emergency planning. A number that is 40 percent different depending on how the boundary was drawn is a number that can support opposite conclusions, and it frequently does.
What follows is an attempt to be precise about what is being measured, where the measurements diverge, and what has to travel with the figure for it to be usable by someone who did not compute it.
At least four different things share the name
Population density is people per unit area. It is the default meaning and it depends on two quantities that are both harder than they look: a population count, which is a survey or a model, and an area, which depends entirely on the boundary chosen.
Dwelling density is housing units per unit area. It is the planner's working measure because it relates directly to what gets built and to infrastructure load. It diverges from population density wherever household size varies, which is everywhere. Two neighborhoods with identical dwelling density and household sizes of two and five are not the same place.
Floor area ratio, sometimes called plot ratio, is total built floor area divided by land area. It measures built intensity rather than occupancy, and it is the measure that regulation usually controls. A district can have a high floor area ratio and low population density if the floor area is offices, warehouses, or vacant.
Built up fraction, or impervious fraction, is the share of a given area covered by construction. It is what satellite derived layers measure most directly, and it says nothing about height or occupancy. A single story industrial estate can approach complete built up coverage while housing nobody.
These four disagree systematically rather than randomly. Central business districts rank high on floor area ratio and low on population density. Informal settlements rank very high on population density with low floor area ratio. Suburban sprawl ranks low on everything except built up fraction at the parcel level. Any statement of the form this city is denser than that one is incomplete until it says which of the four is meant, and often reverses when a different one is used.
The boundaries decide the answer
The modifiable areal unit problem is the reason two analysts can compute density correctly from the same data and disagree. It has two distinct components and both bite.
The scale effect: results change with the size of the reporting unit. Aggregate population to grid cells of 100 meters and you see the actual variation between a housing block and the park beside it. Aggregate to one kilometer and both disappear into an average. Neither is wrong, and they answer different questions, and a difference between two places measured at one scale can vanish or invert at another.
The zoning effect: results change with where the boundaries are drawn even at the same scale. Shift a grid by half a cell and every cell value changes. Redraw administrative boundaries and the density of every unit changes without a single person moving. This is not a small perturbation. Where boundaries follow a river, a rail line, or a historical parish edge, they can place a dense settlement in the middle of a unit or split it across three.
The consequence for practice is that a density figure is a property of a measurement scheme applied to a place, not a property of the place. Comparisons are only valid between figures computed with the same scheme, and the scheme has to be part of the record. When someone asks for the density of a city, the honest first response is to ask which boundary, because the answer for the municipality, the built up area, and the metropolitan region will differ by multiples.
There is a useful defensive technique: compute the figure at several scales and report the sensitivity. If a ranking between two areas holds at 100 meters, 500 meters, and one kilometer, it is probably real. If it flips, the ranking was an artifact of the unit and should not be reported at all.
Administrative units make cross city comparison meaningless
City boundaries are political and historical artifacts. Some cities annexed their suburbs and some did not. Some municipal boundaries enclose large rural hinterlands, forests, water bodies, and airports. Some are tightly drawn around the historical core with the majority of the urban population outside them.
Divide population by administrative area and you get a number dominated by how the boundary was drawn a century ago. A city that includes its farmland reports a low density. A city that stops at its old walls reports a very high one. The population of the metropolitan region might be identical. This is why administrative density rankings of world cities rearrange dramatically depending on the source, and why they are close to useless as evidence for anything.
Three alternatives are more defensible.
- Built up area density. Restrict the denominator to land that is actually built on, using a settlement or built up layer rather than the administrative polygon. This removes the airport and the forest and makes the comparison closer to like for like. It has its own sensitivity: what counts as built up is a threshold on a continuous measure, and the threshold changes the answer.
- Consistent functional definitions. Use a definition of urban extent applied identically everywhere, such as a contiguity and threshold rule on gridded population. Definitions of this kind exist in the open and are used precisely because they are reproducible across countries.
- Population weighted density. Instead of dividing total population by total area, compute the density of each small unit and then average those densities weighted by the population in each. This answers a different and usually more interesting question: what density does the average resident experience. A city with a dense core and empty periphery has a low arithmetic density and a high population weighted density, and the second figure describes life there better.
Population weighted density is the single most useful correction available for cross city work, and it is still sensitive to the size of the small units used, so the unit has to be stated.
Gridded population is a model, not a census
Global gridded population products solve the boundary problem by distributing census counts onto a regular grid. They are extremely convenient and they carry assumptions that get forgotten as soon as the data is loaded.
Every gridded product starts from census or administrative counts and redistributes them within the reporting unit using some ancillary information. The simplest approach spreads population uniformly across the unit, which is transparently wrong but at least honest. More sophisticated products weight the distribution by built up area, by building footprints, by land cover, by road density, or by a trained model combining several. The more sophisticated the redistribution, the better the result usually looks and the harder it is to reason about what it assumes.
Three consequences matter for anyone using these grids.
The output cannot be more accurate than the input census unit. If the underlying count came from a district of several hundred square kilometers, a 100 meter grid cell inside it is an interpolation, not a measurement, no matter how convincing the texture looks. The apparent resolution is a property of the redistribution model.
Redistribution weighted by built up area produces circular reasoning if you then use the grid to study the relationship between built up area and population. The correlation was assumed into existence. This trips up more analyses than it should.
Census vintage propagates. A grid published this year may redistribute counts collected several years ago, projected forward by growth rates that are themselves estimates. In rapidly growing peri urban areas the underlying count can be badly out of date while the grid looks current.
The practical rule is to read the method note of whichever product is in use and to record which product and which version produced any figure. Different products disagree with each other in ways that are systematic by region, and swapping one for another mid project silently changes results.
Residential and daytime population are different populations
Almost all population data is residential: it counts people where they sleep. A large fraction of interesting questions concern where people are during the day.
The gap is enormous in specific places. Central business districts can hold many times their residential population during working hours. Industrial zones with almost no residents hold a workforce. University districts, hospitals, transport interchanges, markets, and stadiums all concentrate people who are counted somewhere else. Conversely, dormitory suburbs empty out.
For anything involving service demand, exposure to a daytime hazard, emergency planning, retail catchment, or infrastructure peak load, residential density is the wrong denominator and using it produces errors that are large and directional rather than noisy.
Estimating daytime population from open data is possible and imprecise. Workplace counts from economic censuses where published, employment by sector attached to land use, transport ridership, and points of interest density all contribute. Mobile derived presence data is the direct measurement and it is generally not open, and where aggregated versions are published they come with their own coverage biases toward device ownership.
The minimum honest practice is to say which population is being counted. A figure labelled simply population density, used in a context where daytime presence is what matters, is a category error whatever its numerical accuracy.
Estimating density directly from buildings
Open building footprint data, combined with height estimates, allows a bottom up alternative that bypasses census units entirely.
The chain is: footprints give ground floor area, height gives an estimated story count, footprint area times stories gives gross floor area, floor area divided by land area gives a floor area ratio, and floor area times a use mix and an occupancy assumption gives an estimated population. Each step introduces an assumption and they compound.
The strengths are real. The output is at building resolution, so it respects actual boundaries rather than administrative ones. It updates as imagery updates, which is far faster than a census cycle. And it directly measures built intensity, which is what floor area ratio regulation and infrastructure sizing care about.
The weaknesses are equally real. Height estimation from surface models depends on a good terrain model underneath, and error in the terrain propagates directly into height. Floor to floor height varies by building type and region, so a story count derived from a fixed divisor is systematically wrong for warehouses, atriums, and traditional construction. Use mix is usually unknown, and residential occupancy per square meter varies more between countries and income groups than almost any other parameter in the chain.
Used as a measure of built form, the building based approach is strong. Used as a population estimate, it should be calibrated against census figures wherever they exist and reported with an interval wherever they do not. The failure to avoid is presenting a chain of four assumptions as a measurement because the last step produced an integer.
What has to travel with the number
A density figure is usable by someone else only if it arrives with the things that determine it.
- The quantity: people, dwellings, floor area, or built up fraction.
- The denominator: administrative polygon, built up area, grid cell, or parcel, and its size.
- The aggregation: arithmetic or population weighted, and at what unit size.
- The source and vintage: which population product or census, which building dataset, which imagery date.
- The method for anything modelled: the redistribution scheme, the occupancy assumptions, the height to story conversion.
That list looks bureaucratic until the first time two teams produce different densities for the same district and spend a week finding out why. Attaching the scheme to the figure turns that week into a diff.
There is a stronger version of the same discipline for anything published through an API or consumed by an automated system. The metadata should not be in a document beside the number, it should be in the record with the number, because the consumer that most needs it is the one least able to go and look it up.
The comparison to refuse
The request that arrives most often is a ranking: which of these areas is densest, so we can prioritize. It is answerable when the areas share a measurement scheme and it is not answerable across cities in different countries with different census geographies, different built up definitions, and different data vintages.
The right response is not to decline the question. It is to answer a narrower one and say so: these areas ranked by population weighted density at 250 meter resolution, using one specified gridded product, with the ranking stable across three resolutions, and with two of the areas excluded because their underlying census vintage differs by enough to make the comparison unsafe.
That answer is less satisfying and it survives being checked. The alternative, a clean ranking with a single number per city, is the one that gets quoted for years after the person who computed it has forgotten which boundary they used.