Urban compactness and carbon emissions: global evidence over the period 1975–2020

Abstract

Urban areas account for a large and growing share of global carbon emissions, yet the role of urban form in shaping these emissions remains insufficiently understood. While existing studies have primarily focused on urbanization rates or population density, less attention has been devoted to the geometric configuration of cities and its relationship with sector-specific emissions over time. In this paper, we examine the association between urban compactness and carbon emissions using a global panel of more than 11,000 urban centres over the period 1975–2020. We combine harmonized definitions of urban areas with high-resolution data on residential and on-road transport emissions and multiple geometric indicators of urban form. We find that more compact urban configurations are associated with lower per capita emissions in both sectors. However, the strength of this relationship varies substantially across regions, being stronger and more robust in low- and middle-income countries, and weaker or not statistically significant in high-income regions. These results are robust across alternative specifications but should be interpreted as conditional associations rather than causal effects. Overall, the findings highlight the relevance of urban form as a dimension of environmental performance, while underscoring the importance of economic and institutional context.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-54461-9.

Introduction

Urban areas account for a large and growing share of global carbon emissions, with recent estimates suggesting that cities are responsible for more than 70% of total emissions¹. As urban population continues to expand, understanding how different dimensions of urbanization affect environmental outcomes has become a central issue in urban and environmental economics. A substantial body of literature has examined the relationship between urbanization, income, and emissions, often focusing on aggregate measures such as urban population shares or city size^2–4. These studies typically find that urbanization is associated with higher emissions, although this relationship may weaken or reverse at higher income levels due to technological progress and structural change^3–6.

More recently, attention has shifted from urbanization per se to the internal structure of cities. In particular, a growing empirical literature has explored how population density influences energy use and emissions, generally finding that denser urban environments are associated with lower per capita emissions, especially in the transport sector^7,8. However, density-based measures provide only a partial characterization of urban spatial structure, as they do not capture the geometric configuration of urban areas or the spatial distribution of built-up land within city boundaries.

In this paper, we move beyond density and focus on urban form, defined as the geometric and spatial configuration of urban settlements. We ask whether changes in urban compactness rather than changes in population size or density are systematically associated with sector-specific carbon emissions over time. Addressing this question is important because urban form is directly shaped by planning policies and land-use regulation, and therefore represents a potentially actionable lever for climate mitigation.

Our contribution is threefold. First, we construct a novel global panel dataset covering more than 11,000 cities over the period 1975–2020, combining harmonized definitions of urban centres with time-varying, high-resolution data on sectoral CO₂ emissions and urban morphology. Second, we introduce multiple geometric indicators of urban compactness, allowing us to distinguish the effects of urban form from those of density and overall city size^9–12. Third, we provide large-scale evidence on how the relationship between urban form and emissions varies across sectors (residential and on-road transport) and across regions and income groups.

We find that more compact urban forms are consistently associated with lower per capita emissions in both the residential and transport sectors, although the magnitude of this association varies substantially across regions. In particular, the relationship is stronger and more robust in low- and middle-income countries, whereas it appears weak or insignificant in high-income regions such as Europe and North America. These findings suggest that the environmental benefits of compact urban development are context-dependent and may be most relevant in rapidly urbanizing regions.

We interpret these results as conditional associations rather than causal effects. While our empirical strategy controls for time-invariant city characteristics and a set of key time-varying factors, potential endogeneity between urban form and emissions cannot be fully ruled out. Nevertheless, the consistency of the results across specifications and robustness checks provides suggestive evidence of a meaningful link between urban morphology and environmental performance.

Data on urban centres and CO₂ emissions

Recently, the literature on the environmental impact of cities has also focused on the availability of geo-localised data for studying their carbon footprint. However, obtaining data on emissions at an adequate resolution and quality has historically been complex: while different sources exist, their validity is often undermined by inconsistencies and endogeneity issues. Self-reported emission accounts from urban administrations, for instance, are generally obtained following different methods and rely on administrative definitions of “city” or “municipality”, which are themselves influenced by historical, economic and political factors. Recent progress has been made through the effort of organisations such as the Carbon Disclosure Project, which aims at harmonizing the cross-country administrative collection of data on emissions of urban centres, by providing guidelines for reporting and collecting the resulting estimates in yearly databases¹³. These data have already been employed to produce a scientific dataset of urban emissions and ancillary variables¹⁴.

Information from Environmentally Extended Input-Output (EEIO) Tables has been used to trace the carbon footprint of 13,000 cities all around the globe in the Gridded Global Model of City Footprints (GGMCF)¹⁵. The carbon footprint is a more complex concept with respect to just CO₂ emissions, as it entails an assessment of the emissions of the production processes leading to a certain product, and therefore an additional level of detail. To this end, the EEIO Tables data are used to estimate consumption flows and related carbon emissions across countries, calculating total imported and exported emissions. These are then attributed to regions using subnational CF models, and eventually to cities using data on rural versus urban consumption patterns. In our study, however, we only focus on direct emissions produced in urban centres for two sectors that have the highest potential to be affected by urban form: the residential energy sector (heating and cooling for buildings) and the on-road transport sector.

Finally, a different method to estimate emissions relies on satellite information on CO₂ atmospheric concentration. This can be used to trace emissions via combination with a set of other geographical, climate and socio-economic variables in spatial models. A database relying on this methodology has provided information on CO₂ emissions of 20 cities across the world¹⁶. More recently, atmospheric data have been used to reach a sample size of 1,236 cities across multiple continents from 2014 to 2020¹⁷. The results have also been used to investigate links between population density and the estimated emissions, finding a negative correlation between the two¹⁷. While promising, this type of data would still restrict our sample and prevent the sectorial differentiation of emissions data, which is an important feature of our analysis.

In general, and apart from these sources, the review of existing literature highlights a lack of global comparable evidence with time varying data on both emissions and urban forms. We attempt to partially bridge this gap with the current study, as we assemble a dataset using time-varying, high resolution spatial data on the degree of urbanization, sectorial carbon emissions, urban form indicators and ancillary variables (temperature, precipitations and GDP) for cities across the world. The database covers the 5-years intervals from 1975 to 2020. Urban boundaries are obtained starting from raster data at a 1 km resolution on the degree of urbanization from the GHS-SMOD dataset, released as part of the European Commission’s Global Human Settlement Layer (GHSL) project¹⁸. We only select entities qualified as “urban centres”, i.e. with a population of more than 50,000 inhabitants and a population density above 1,500 inhabitants per km², resulting in a database of 11,435 cities in 2020 (our reference year).

We assign to each “urban centre” information on CO₂ emissions drawn from the EDGAR Database^19,20. These are obtained for all world countries, using sector-specific indicators of human activity, technology, fuel mix and abatement percentage to estimate total emissions (The final estimates of CO₂ emissions in EDGAR are disaggregated by type of fuel, with a distinction between non-short-cycle-organic and short-cycle-organic fuels. These two categories are aggregated for the purpose of this study.). The country-sectorial results are down-scaled using spatial proxy data on the location of energy and industrial facilities, residential and agricultural areas, etc. to obtain the final resolution of 0.1 degrees^19,20. In addition to EDGAR data, we include ancillary information from other sources on the population²¹ and built environment of cities²², their climate²³ and GDP²⁴, and we compute urban form indicators. A comprehensive, step-by-step explanation of the data construction process is provided in the Methods section and further detailed in the Supplementary Information. We believe our data ensure consistency both in the definition of “urban centre”, which is solely based on population criteria and therefore allows us to draw comparisons across continents and income groups, and in the methods used to estimate emissions, which do not resent from the sample bias and methodological heterogeneity of public emission accounts from administrative self-reporting. At the same time, using the newly released GHS-SMOD data on the degree of urbanization across the globe¹⁸ gives us the advantage of obtaining time-varying information on evolving urban boundaries at a relatively high frequency. It allows us to build a fixed effects model and ignore city-level time invariant characteristics, such as factors linked to their geographic location or institutional setting, that may endogenously bias our results.

The use of harmonized urban-centre boundaries improves comparability in the delineation of cities, but it does not imply that urban systems are structurally equivalent across regions. Urban centres in Asia, Europe, North America, or Africa may differ substantially in terms of suburbanization, polycentricity, commuting patterns, and the spatial distribution of population within and around the urban core. For this reason, our empirical strategy should be understood as providing a consistent measurement framework rather than assuming that all urban centres represent the same type of settlement. These differences are explicitly considered in the heterogeneity analysis by continent and income group, and they are central to the interpretation of the results.

Changing urban form and the geometry of cities

Figure 1 shows the share of the population and the share of total carbon emissions produced in urban centres at the global level and in each continent. The two clearly follow parallel trends and have increased over time at the global level. In our study, however, we only look at emissions directly produced in urban centres, which are less than the total consumption-based carbon footprint of cities. The lower value compared to the urban population share thus indicates not only that efficiency dynamics may be at play, but also that cities out-source a significant share of their emissions.

Fig. 1.

Changes in the share of urban population and urban CO₂ emissions over time.

Previous literature has focused on limited samples of cities, often for specific countries, and has made use of population density as a proxy variable for urban form. In this study, we aim to provide more compelling evidence on the impact of urban compactness on carbon emissions by taking a wider perspective, using comparable data on 11,435 cities across the world and focusing on their form, that is the shape of the built area, under the hypothesis that more compact cities are more efficient and allow for lower energy consumption, especially regarding transportation. In this respect, Fig. 2 shows the evolution of the ratios of total urban area and total urban residential built-up surface to urban population from 1975 onwards. The ratios have not varied to a large extent at the global level, but with important regional differences: in Europe and Asia, urban area per citizen has increased, while it has decreased for America, Oceania and especially Africa, whose urban centres have become more dense.

Fig. 2.

Time trends in urban land use and urban population. Note: the ratios are normalised to 100 in 1975, the first year of data availability, and should be interpreted in relation to that reference year.

In our empirical analysis, we include total urban area as an indicator of urban sprawl: fixing population by always looking at emissions in per capita terms, a larger area may be expected ex-ante to increase travel distances and be conducive to sprawl. In addition to this, we construct more elaborate geometric urban form indicators including the Compactness Index (CI), the Range Index (RI), and the Sprawl Index (SI). Figure 3 highlights the data and geometric figures involved in the computation of the CI, the RI and the SI for a random city in the sample, Milan in Italy. A more detailed explanation of the formulas and computational methods for each is provided in the Methods section.

Fig. 3.

Urban form indicators for a sample urban centre. Note: the maps refer to a random city in the sample, Milan in Italy, for the year 2020.

The CI was first designed by Li & Yeh²⁵ in their study on the evolution of land use patterns in the Chinese Pearl River Delta during the period of fast economic growth of the ‘80s and ‘90s, and has often been used to describe the compactness of cities^11,26,27. It is a unit-less measure obtained by comparing the perimeter of the land patch of interest and the perimeter of a circle with equivalent area. The circle represents an “optimally” compact settlement, where the distance between each point is as small as possible for a given area, and the CI represents distance of the urban centre’s form from this idealized shape. Higher values of the index correspond to more compact cities. The RI^3,28 describes compactness in a slightly different way, comparing the diameters of a circle with area equivalent to the urban centre, and that of the smallest circle circumscribing it. Once again, the indicator correlates positively with compactness, but it is more sensitive to irregularities in shape that may strongly increase the diameter of the minimum enclosing circle, and is therefore less stable than other measures.

Finally, the SI was developed in seminal work on the determinants of sprawl²⁹ and further discussed for its compactness properties and modified in later research^28,30. The index draws on built-up classification raster data to measure the extent of urban sprawl in a given territory. For each pixel of residential built-up surface, it is obtained by computing the share of undeveloped cells (“open space”) in a given area around the pixel, and then averaging the resulting value over all residential pixels within the territory. Thus, the SI describes the extent to which the surroundings of residential areas are also exploited and occupied by other built-up surface (for residential or non-residential use). Clearly, the surface of an urban centre should not be completely full, i.e. entirely occupied by residential built-up surface, as this would remove amenities such as parks, urban forests, waterways, etc. At the same time, however, the leapfrogging of residential development over large open areas is a symptom of urban sprawl. In urban centres, greater sprawl may reduce energy efficiency resulting from agglomeration, increase travel distances between locations and therefore have an impact on per capita CO₂ emissions.

Empirical evidence

Methodology

To assess the relationship between urban form and carbon emissions, we estimate panel regressions with city-specific fixed effects. This approach allows us to control for time-invariant characteristics of cities, such as geographic location or institutional factors, that may influence emission levels. The baseline specification relates per capita CO₂ emissions in sector s (residential or on-road transport) in city i at time t to a set of urban form indicators, including total urban area and measures of compactness, while controlling for GDP per capita and climate variables.

All variables are expressed in logarithmic form, so that estimated coefficients can be interpreted as elasticities. We include year fixed effects to account for global shocks and cluster standard errors at the country level to allow for correlated shocks across cities within the same country. In additional specifications, we incorporate country-specific time trends and dynamic models using lagged dependent variables to assess the robustness of the results.

While this empirical strategy helps mitigate concerns related to omitted variables, the results should be interpreted as conditional associations rather than causal effects.

The baseline specification is:

where is the natural logarithm of per capita CO₂ emissions of sector S (where S can be either the residential or the on-road transport sector) in city  i at time  t. is the logarithm of urban form indicator  F, corresponding to the area or to one of the three urban form indicators computed (), and is the corresponding coefficient. The remaining controls include , i.e. log GDP per capita, with corresponding coefficient , and , the vector of log environmental controls (namely, temperature and precipitations) with corresponding vector of coefficients . Finally, represents year specific fixed effects and is the error term, clustered at the country level. It should be noted that the use of natural logarithms for both the dependent and the independent variables of the regression allows for the interpretation of estimated coefficients as elasticities. The coefficient , for instance, should be interpreted as the percentage change in per capita emissions associated with a 1% change in the urban form indicator used.

While a wide range of factors may influence urban carbon emissions, including industrial structure, energy systems, and technological change, data limitations at the global scale constrain the inclusion of such variables in a consistent manner across cities and over time. In this context, GDP per capita is commonly used as a proxy capturing broader economic development, technological progress, and structural transformation in cross-country analyses. Moreover, the use of city fixed effects absorbs time-invariant characteristics related to industrial composition and local economic structure, while the inclusion of time effects and additional specifications with country-specific trends helps account for global and regional dynamics. Taken together, these elements allow us to control for a substantial share of confounding factors, while preserving the comparability and coverage of the dataset.

Figure 4 displays residential CO2 emissions per capita (in log scale, to smooth extreme values) against each of the three urban compactness indicators in levels for the 2020 cross-section of cities, consisting of over 11,000 units. Each panel displays trends in the data, estimated using a Locally Estimated Scatterplot Smoothing (LOESS) technique (in red, with 95% confidence intervals) and a linear OLS fit (blue). Across all three indicators, the two curves are very similar: the LOESS confidence band largely overlaps with the OLS line, with no systematic departure that would signal meaningful non-linearity. This appears to confirm the choice of maintaining simplicity in modeling the relation between emissions and urban form in our panel regressions with city fixed effects. Finally, the compactness index, as in our panel regressions, shows the strongest correlation with residential emissions also in the plot.

Fig. 4.

Urban compactness and residential emissions.

Baseline estimates

Table 4 presents the coefficients on the urban form indicators for the per capita CO₂ emissions of the residential sector (heating and cooling for buildings), while Table 5 focuses on emissions from on-road transport. For the residential sector, the values of the are smaller than for the transport sector, for which the indicator is consistently high, indicating a high predictive power of the models employed. This suggests that the dynamics behind the emissions of on-road transport are more easily explained by urban form than those for heating and cooling buildings. Furthermore, GDP per capita fares better at explaining the emissions of the transport sector than the residential, as evidenced by the larger magnitude of its coefficients in Table 2 compared to Table 1.

Table 2.

Associations between urban form indicators and on-road transport CO₂ emissions.

	(1)	(2)	(3)	(4)
GDP p.c.	0.525***	0.564***	0.570***	0.571***
	(0.155)	(0.158)	(0.157)	(0.156)
Area	0.223***
	(0.062)
CI		− 0.338***
		(0.077)
RI			− 0.211***
			(0.050)
SI				− 0.120
				(0.247)
Cons.	− 7.483***	− 7.165***	− 7.139***	− 7.126***
	(1.153)	(1.279)	(1.300)	(1.219)
Env. controls	Yes	Yes	Yes	Yes
Observations	20,000	20,000	20,000	19,998
R 2	0.755	0.747	0.746	0.745

Standard errors in parentheses. * p < 0.1, ** p < 0.05, *** p < 0.01. All variables are taken in natural log terms.

Table 1.

Associations between urban form indicators and residential CO₂ emissions.

	(1)	(2)	(3)	(4)
GDP p.c.	0.145***	0.171***	0.175***	0.176***
	(0.047)	(0.050)	(0.050)	(0.049)
Area	0.188***
	(0.030)
CI		− 0.244***
		(0.035)
RI			− 0.123***
			(0.023)
SI				0.197***
				(0.071)
Cons.	− 4.271***	− 3.991***	− 3.963***	− 3.789***
	(0.422)	(0.425)	(0.426)	(0.424)
Env. controls	Yes	Yes	Yes	Yes
Observations	53,134	53,134	53,134	53,098
R 2	0.099	0.081	0.079	0.082

Standard errors in parentheses. * p < 0.1, ** p < 0.05, *** p < 0.01. All variables are taken in natural log terms.

Regarding urban form indicators, almost all coefficients are significant and with the expected sign: cities with a smaller area and more compact urban forms, proxied by higher values of the compactness and range index, tend to emit less CO₂ per citizen in the residential (heating and cooling for buildings) and the on-road transport sectors. Looking at the compactness index, a 10% decrease in the indicator (such as that experienced, for instance, by the cities of Hamburg in Germany or Bangalore in India in our sample from 2010 to 2015) is associated with an increase of about 2.4% in residential emissions, and 3.4% in on-road transport emissions. For the average city in our sample in 2010, with a population of about 290,000 inhabitants, this translates into a decrease in CO₂ emissions for heating and cooling buildings of about 5 kg per inhabitant, for a total of 1.4 thousand tonnes. For on-road transport, while the marginal effect is stronger in magnitude, the overall increase in emissions of the average city is entirely comparable, again with about 5 more kg per citizen and 1.4 thousand tonnes for an average urban centre. Still, results are heterogeneous across countries and corresponding income groups, an aspect that we investigate in greater detail later.

For both sectors investigated, the estimated coefficient on the RI is smaller than that on the CI, despite the indexes capturing the same broad notion of compactness. This may be due to the lower stability of the RI, which is altered more strongly by small variations in city shape due to the way it is constructed. Nevertheless, the coefficient on the RI is still negative and statistically significant. For buildings, a relevant channel behind the association of these indicators with emissions may be the sharing of carbon intensive goods linked to the provision of heating and cooling: compact city development is characterised by the construction of buildings, rather than single dwelling units, which consume less energy per person and thus emit less per capita CO₂^10,31,32. For on-road transport, on the other hand, a larger area is associated with greater average distances within the urban centre, and subsequently higher emissions per person. The CI and RI capture this aspect by looking at geometric differences in urban boundaries that may alter the average distance between two random points within the city, while keeping area constant. Finally, the coefficient on the sprawl index is positive and significant for the residential sector, while it becomes negative, although insignificant, for emissions of the on-road transport sector.

To increase the robustness of our estimates, Extended Data Tables 1, 2, 3 and 4 provide a set of additional specifications for our analysis. As a first step, we re-estimate the two main regressions after balancing our panel of cities. We only retain urban centres characterised as such for the entire period from 1995 to 2015 (the period for which our regressions including GDP are estimated), thus removing gaps in the data. Extended Data Table 1 shows the results of the fixed effects model estimated for the balanced panel on emissions from the residential and the transport sectors respectively: the coefficients retain the same sign and significance level in all cases. Extended Data Table 2, on the other hand, shows the results of the main regression estimated on total, rather than sectorial, per capita urban carbon emissions. The findings for the residential and transport sector appear to be transferred also to the total emission count: GDP per capita is positively and significantly associated with total emissions, while the compactness indicators appear to be beneficial towards a reduction in total emissions. The effect of the sprawl index is negative, but close to 0 and not significant, which mimics the results obtained for the on-road transport sector.

Extended Data Table 3 includes country-specific linear time trends in the model specification. The levels of carbon emissions in urban centres may indeed also depend on factors changing over the time span of our sample, such as ongoing technological progress or the results of long-term policy agendas aimed at fighting climate change by reducing emissions over time. We model this by including a generic 5-year trend in our main equation and interacting it with country-level dummies to control for specific trends in cities belonging to the same country. The general trend is negative and significant for the residential per capita CO₂ emissions, whose levels have slightly decreased globally over time, while it is not significant for per capita emissions in the on-road transport sector, whose emissions haven’t followed a clear pattern and have remained quite constant at the global level, despite differences across continents. As for the coefficients on GDP and urban form indicators, they remain substantially unaltered.

To reduce endogeneity, Extended Data Table 4 provides the results of transforming our main specification into a dynamic model including a lagged value of the dependent variable and using lagged values of the independent variables as instruments, employing GMM-IV estimations following Arellano and Bond³³. Such a method has already been adopted with similar data on urban centres by Castells-Quintana et al.³⁴. The results maintain the direction and significance of the main specification, confirming that the relation between urban form and emissions is not spurious for both sectors.

Heterogeneity across income and continents

Figure 5 plots the marginal effect of area and the urban form indicators, conditional on location in each continent. Extended Data Fig. 1 repeats the process, but looking at heterogeneity across income classes. The marginal effects are obtained from a model including interaction terms between urban form and the continent and income dummies (see Methods). They shed light on the geographical differences in the association between urban form and the environmental impact of cities. For area, the compactness and the range index, the effect appears to vary significantly by geographic region: the associations with residential and transport emissions are generally not significant for North America and Europe, whereas they become negative and significant for Asia and South America. In general, and with the exception of the sprawl index, the marginal effects for low and lower middle income countries are more pronounced than for high income countries. This is a promising aspect, as these countries also have higher rates of urbanization and thus greater potential to curb emissions in the future, via careful urban planning.

Fig. 5.

Heterogeneous effects of urban form indicators across continents. Note: the bands represent 95% confidence intervals around the coefficients.

In Africa, for instance, a 10% increase in the compactness index is associated with a 4% decrease in emissions for heating and cooling buildings, about double the association detected at the global level. For the average urban centre in the continent in 2010, with a population of 230,000, this translates into a decrease of 3 kg per citizen and about 700 tonnes in total in the city. This is a far smaller change compared to the global effect in absolute terms due to the ex-ante lower levels of emissions of African cities, but the higher marginal effect can be expected to play a more important role in the future, as African cities grow bigger and more polluting. In Asia, a 10% increase in the compactness index of the average urban centre, whose population is of about 300,000 inhabitants, is associated with a 2% decrease in carbon emissions from on-road transport, corresponding to around 2 kg per citizen and 600 tonnes in total.

The sprawl index appears to follow a different pattern, as the corresponding coefficients have a stronger magnitude for cities in Europe and North America. As the SI was designed focusing on the US context²⁹ and captures more complex internal structures in the urban built environment, describing interrelations between residential built-up surface and the surrounding open space, perhaps it is better suited to the analysis of cities in these continents, and especially North America. In the continent, a decrease in about 5% in the index, such as that experienced by San Diego in the United States or Quebec City in Canada between 2010 and 2015, is associated with a 3.4% decrease in residential emissions. For the average North American urban centre, this translates into 24 more kg per citizen, and 9 thousand tonnes in total at the city level.

Before concluding, heterogeneity in our sample of cities is also explored following the distribution of the dependent variable. Extended Data Table 5 shows the coefficients on our urban form indicators in a set of separate regressions (using our main specification) for each quintile in the distribution of either residential CO₂ emissions (panel a) or on-road transport CO₂ emissions (panel b), in 1990. Despite the sample sizes being approximately the same in each of the regressions, the coefficients on the indicators are generally stronger in magnitude and significance in the lower quintiles, while significance fades away in the upper 5th quintile of the emissions distribution.

Additional exploratory analyses were conducted to assess potential nonlinearities and temporal variation in the relationship between urban form and emissions. In particular, we estimated alternative specifications including nonlinear terms for the main urban form indicators and segmented the sample into sub-periods to examine whether the estimated associations varied over time. These extensions, however, did not yield consistent or robust patterns across specifications, and the core results remained largely unchanged. While these findings suggest that the relationship between compactness and emissions is broadly stable and approximately monotonic at the global scale, they also highlight the difficulty of capturing more complex dynamics in a highly heterogeneous international sample. For this reason, we focus on the baseline specification, which provides a more parsimonious and interpretable characterization of the data.

Discussion

There is a multitude of reasons for the significant impact of urban compactness on the environmental performance of cities, including economies of scale. In the United States, research using gridded population, land use and CO₂ emissions data finds that population density is negatively associated to on-road emissions in cities, while the relation is positive for urban sprawl³⁵. In Japan, a study using a cross section of 50 cities finds evidence in favour of a negative association between compactness and CO₂ emissions¹¹. At the same time, in China, where urbanization has grown considerably in recent decades, Ou et al.¹² exploit a panel of cities that expanded rapidly in the period from 1990 to 2010 to find negative associations between different compactness metrics and emissions records, and a positive link between compactness and quality of the urban road infrastructure (More studies in China and Japan yield similar conclusions^36–38, while at least one other study³⁹ finds conflicting results by uncovering a positive link between emissions and urban density while using cross-sectional data on Chinese urban centres in 2013.). A study for EU member states looks at the dynamics of compact cities using country-level data from 2000 to 2012, by computing the weighted average of a compactness metric for urban centres and investigating its effect on carbon dioxide emissions²⁶. While a beneficial effect emerges of both physical compactness and population density on reducing emissions, the effect of density is found to prevail over physical compactness.

This paper provides global evidence on the association between urban form and sectoral carbon emissions, showing that more compact cities tend to exhibit lower per capita emissions in both the residential and on-road transport sectors. While this pattern is broadly consistent with prior findings based on density measures^35,17, our results extend the literature by focusing on geometric indicators of urban form and by exploiting a long time dimension across a large global sample.

A key insight emerging from the analysis is that the relationship between compactness and emissions is not uniform across contexts. The estimated associations are substantially stronger in low- and middle-income countries, particularly in Africa and Asia, whereas they are weak or statistically insignificant in Europe and North America. This heterogeneity suggests that urban form interacts with broader structural conditions, such as infrastructure provision, transport systems, and patterns of urban growth, in shaping environmental outcomes.

From a conceptual perspective, two main mechanisms may explain the observed associations. First, in the transport sector, more compact urban forms reduce average travel distances and may facilitate the use of more efficient transport modes, thereby lowering per capita emissions. Second, in the residential sector, compactness is often associated with building typologies that allow for the sharing of energy-intensive services (e.g. heating and cooling), resulting in lower energy consumption per capita^10,31,32. These mechanisms are consistent with the idea that spatial proximity enhances efficiency through both reduced mobility needs and economies of scale in energy use.

At the same time, the weaker associations observed in high-income regions suggest that the environmental benefits of compactness may diminish once a certain level of economic development and technological advancement is reached. In these contexts, factors such as energy efficiency standards, fuel composition, and transport technologies may play a more dominant role than urban form itself. This interpretation is consistent with previous evidence highlighting the role of income, technology, and structural change in moderating the urbanization–emissions relationship^40–43. Furthermore, the weaker associations observed in Europe and North America may partly reflect the fact that harmonized urban-centre boundaries capture dense urban cores but may not fully represent wider metropolitan commuting zones or suburbanized settlement patterns. Conversely, in many rapidly urbanizing regions, especially in Asia and Africa, urban-centre boundaries may correspond more closely to the areas where future population growth, infrastructure investment, and land-use decisions are concentrated. This helps explain why the association between compactness and emissions appears stronger in low- and middle-income contexts.

The results should be interpreted with caution in light of several limitations. First, although the use of city fixed effects and a range of controls helps mitigate omitted variable bias, the analysis does not establish causal relationships. Urban form and emissions may be jointly determined, and endogeneity cannot be ruled out. Second, the construction of the dataset relies on the spatial aggregation of raster data, which may in principle introduce measurement error, particularly for variables such as emissions and GDP that are unevenly distributed within grid cells. Third, the analysis focuses on residential and on-road transport emissions, which are the sectors most directly affected by urban form, but does not account for emissions embedded in industrial production or consumption-based carbon footprints. Another important limitation of the analysis relates to the role of industrial structure. Industrial activities, particularly energy-intensive sectors, are a major determinant of carbon emissions and may differ substantially across cities and countries. Due to the lack of globally consistent and time-varying data at the city level, it is not possible to explicitly control for differences in industrial composition within the empirical framework. Consequently, part of the cross-sectional and regional heterogeneity observed in the relationship between urban form and emissions may be driven by variation in industrial activity rather than by differences in urban morphology alone. This limitation does not invalidate the observed associations but suggests caution in attributing them solely to urban form.

Despite these limitations, the consistency of the results across specifications and robustness checks suggests that urban form is an important correlate of environmental performance, particularly in rapidly urbanizing regions. From a policy perspective, this implies that urban planning strategies aimed at promoting compact development may contribute to emission mitigation, but their effectiveness is likely to depend on local conditions. In low- and middle-income countries, where urban expansion is ongoing and infrastructure systems are still evolving, spatial planning may play a more decisive role. In contrast, in high-income contexts, complementary policies targeting energy systems, building efficiency, and transport technologies may be required to achieve significant emission reductions.

The empirical results provide several insights that are relevant for urban planners and policymakers, although they should be interpreted in light of the non-causal nature of the analysis. In rapidly urbanizing low- and middle-income countries, where urban expansion is ongoing and infrastructure systems are still being developed, the results suggest that spatial planning may play a particularly important role. Policies aimed at promoting contiguous urban expansion, limiting leapfrog development, and coordinating land-use and transport planning may help reduce future emissions by shortening travel distances and supporting more efficient mobility systems. In these contexts, urban form represents a forward-looking policy lever, as current development patterns will shape long-term infrastructure and energy use. For example, in rapidly growing cities in Sub-Saharan Africa or South Asia, limiting fragmented urban expansion and promoting more contiguous development may reduce future transport emissions by decreasing commuting distances and enabling more efficient public transport systems. In contrast, in European cities where urban form is already relatively compact, further emission reductions are more likely to depend on building retrofitting and the decarbonization of energy supply rather than additional changes in spatial structure.

In high-income contexts, where urban form is relatively stable and the estimated associations between compactness and emissions are weaker, the results suggest that changes in spatial structure alone are unlikely to generate substantial emission reductions. Instead, policies targeting energy systems, building efficiency, and transport technologies, such as electrification, public transport investments, and regulatory standards, are likely to be more effective.

Finally, the results highlight that urban compactness should not be interpreted as a universal solution. Its effectiveness depends on complementary factors, including infrastructure provision, institutional capacity, and technological conditions. As such, compact development policies are best understood as part of a broader and context-specific policy mix rather than as a standalone strategy. These policy implications are consistent with the heterogeneous effects observed in the empirical analysis, where the association between compactness and emissions is stronger in low- and middle-income countries and weaker in high-income regions.

Several limitations should be considered when interpreting the findings of this study. First, the empirical analysis does not establish causal relationships. Although the use of city fixed effects, time controls, and dynamic specifications helps mitigate concerns related to omitted variables, potential endogeneity between urban form and emissions cannot be fully ruled out.

Second, the analysis does not explicitly model spatial dependence across cities. Emission dynamics and urban development patterns may be spatially correlated due to shared infrastructure systems, regional economic linkages, or environmental conditions. While clustering standard errors at the country level and including country-specific trends partially address these concerns, future research could extend the framework to incorporate spatial econometric approaches, not simply based on spatial distance.

Third, the construction of the dataset relies on the spatial aggregation of raster data, which assumes a uniform distribution of variables within grid cells. This assumption may introduce measurement error at margin, particularly for variables such as emissions and economic activity that are unevenly distributed across space. However, the use of harmonized global datasets ensures consistency and comparability across cities and over time.

Finally, the analysis focuses on residential and on-road transport emissions, which are the sectors most directly affected by urban form. Other sources of emissions, such as industrial production or electricity generation, are not explicitly considered, as they are more strongly driven by production location and national energy systems. In addition, the compactness indicators employed capture the geometric dimension of urban form and do not account for socio-spatial heterogeneity within cities. Despite these limitations, this approach allows for a consistent and globally comparable measurement of urban morphology.

Overall, while these limitations suggest caution in the interpretation of the results, the consistency of the findings across specifications supports the relevance of urban form as an important factor influencing urban carbon emissions. While these limitations warrant caution in causal interpretation, they do not undermine the central empirical regularity emerging from the analysis. Across a wide range of specifications, including alternative model formulations, balanced panels, and dynamic estimations, the association between urban compactness and lower emissions remains consistent in sign and broadly stable in magnitude. This robustness suggests that the observed relationship is not driven by a specific modelling choice or sample composition, but reflects a systematic pattern in the data. Accordingly, the contribution of this paper lies in documenting a strong and globally consistent association between urban form and emissions, rather than in establishing causal effects. From an empirical perspective, identifying such regularities is an important step toward understanding the role of urban structure in shaping environmental outcomes, and provides a foundation for future research aimed at causal identification.

Conclusion

This paper examines the relationship between urban form and sectoral carbon emissions using a global panel of more than 11,000 cities over the period 1975–2020. We show that more compact urban configurations are associated with lower per capita emissions in both the residential and transport sectors, with stronger effects observed in low- and middle-income countries.

By focusing on geometric measures of urban form rather than traditional indicators such as density or city size, the analysis provides new evidence on the role of spatial structure in shaping environmental outcomes. The results highlight that the effectiveness of compact urban development as a mitigation strategy is context-dependent and varies across regions and stages of development. From a policy perspective, the results suggest that the effectiveness of compact urban development strategies depends strongly on the stage of economic development and the structure of urban systems. In rapidly urbanizing regions, particularly in low- and middle-income countries, policies that promote more compact urban growth, such as land-use regulation aimed at limiting urban sprawl, coordinated transport and land-use planning, and the provision of high-density residential infrastructure, may contribute to reducing emissions by shortening travel distances and improving energy efficiency in buildings. In contrast, in high-income regions where urban form appears to play a more limited role, emission reductions are likely to depend more on technological and regulatory interventions, including improvements in energy efficiency standards, decarbonization of energy systems, and the promotion of low-emission transport modes. These findings indicate that urban compactness should not be viewed as a universal solution, but rather as one component of a broader policy mix that must be adapted to local economic and institutional conditions.

While the findings should not be interpreted as causal, they suggest that urban form represents a relevant dimension of climate policy, particularly in rapidly urbanizing contexts where spatial development patterns are still evolving. The findings suggest that urban form can be an effective lever for emission mitigation primarily in contexts where urban expansion is still ongoing, while in more mature urban systems, technological and regulatory interventions are likely to play a more dominant role.

Future research could further explore causal mechanisms and extend the analysis to additional sectors and broader measures of urban sustainability.

Supplementary Information

Below is the link to the electronic supplementary material.

Appendix: methods

Data on urban centres

A statistical analysis of the characteristics of cities across the world and their association with CO₂ emissions requires a consistent and internationally comparable definition of “urban centre”, as the characterization of a “city” may vary across countries¹⁶. The issue of international comparability of the information on urban centres was recently brought to the fore by the United Nations Statistical Commission, which endorsed a new, global definition of urban centres aimed at overcoming the lack of consistency in their identification across countries⁴⁴(The definition was designed in partnership with other international organisations, including: the European Union (EU), the Organization for Economic Cooperation and Development (OECD), the World Bank, UN-Habitat, the Food and Agriculture Organization (FAO) and the International Labour Organization (ILO).). As anticipated, the new definition disregards the influence of social, economic or geographical factors varying at the country level, and defines cities in terms of their overall population and density. It characterizes a “city” as a human settlement with population above 50,000 inhabitants and average population density above 1,500 inhabitants per square km⁴⁵.

In line with these developments in official statistics, our dataset is assembled starting from information on settlement classification made available by the European Commission as part of the Global Human Settlement Layer (GHSL) project: the GHS Settlement Model Grid (GHS-SMOD). GHS-SMOD provides spatial information from 1975 to 2020, at 5 years intervals and at a 1 km resolution (Each raster pixel covers an area of 1 km².), on the types of settlements on the Earth’s surface, ranging from “rural clusters” to “urban centres” following the Degree of Urbanization definition¹⁸. To obtain urban boundaries, we isolate contiguous pixels classified as “urban” in the SMOD data. We obtain information for 6,371 urban centres in 1975, almost doubling to 11,435 in 2020. The boundaries we identify do not coincide with the urban boundaries defined at the administrative level, as the criteria used in the SMOD database to classify pixels solely rely on total population and population density. To form our panel dataset, we use the most recent time period available, 2020, as reference year and assign the same panel unit ID to cities based on whether the polygons representing their boundaries in different years were overlapping with the 2020 boundaries (By taking 2020 as reference year, we include in the sample all cities that met the Degree of Urbanization threshold that year. Gaps are allowed, e.g. there may be urban centres that met the threshold in 1975 and 1980, fell short in the following decades and met again the threshold in 2020. In exceptional cases where two or more separate urban centres merged into a single city before 2020, we only include in the sample the city with the largest overlapping surface with the 2020 urban centre. This time flexibility in the entities considered to be “urban centres” is made possible by the use of a definition, the Degree of Urbanization, that is entirely population-based and, as anticipated, does not take into account administrative or other country-level criteria.). The Supplementary Information displays graphically the raster polygonization process, mapping the original GHS-SMOD data, their transformation into the “urban centres” boundaries and the evolving boundaries over time.

Urban form indicators

Having obtained spatial information on urban boundaries, we can construct measures of urban form and structure, to determine the degree of compactness of cities. First, we easily compute total urban area for each year of data availability. In addition, we consider a set of compactness indicators: the compactness index (CI), the range index (RI), and the sprawl index (SI). The CI of a land patch is computed as the ratio between the perimeter of a circle having the same area as the land patch of interest and the area of the land patch itself^25,30, according to the following equation:

where and represent, respectively, the perimeters of urban centre  i and of the equivalent-area circle, while is the area of the urban centre. The index takes values between 0 and 1, where the maximum value of one would be reached in the extreme case where the city’s shape is a perfect circle. The RI is similar, but it takes the ratio of the diameter of a circle having the same area as the urban centre of interest, and the diameter of the smallest circle circumscribing it (When the polygon has only 2 vertices intersecting the smallest circle circumscribed about it, this diameter is equivalent to the distance between the two most distant points in the polygon.). The equation:

describes the index, where and represent, respectively, the diameters of the equivalent area circle and of the minimum enclosing circle, and is once again the urban centre’s surface.

To compute the SI, we draw on raster data on residential and non-residential built-up surface at a 100 m resolution from the GHS-BUILT dataset²², part of the GHSL project. We determine a pixel to be “developed” if 20% of its 10,000 m² surface is built, and “residential” if 20% of its 10,000 m² surface is residential. For each pixel, we focus on the 810,000 m² square centred around it (i.e. containing 80 additional pixels each with a size of 100 m) and compute the share of pixels classified as “open space” in this area, excluding water surfaces from the count (To include water surfaces, we use raster data from the GHS-LAND dataset at a 100 m resolution⁴⁶, and qualify a cell as “water surface” if at least 20% of its surface is covered by water.). The process is automatically repeated for each “residential” pixel, and then the average within the urban boundaries is taken to obtain the sprawl index.

CO₂ emissions and ancillary variables

To complete the database, we include information on our main variable of interest, carbon dioxide emissions. We complement this with data on population, GDP, average temperature and precipitations. This wealth of information is drawn from a variety of high-resolution raster data sources, listed in Supplementary Information Table 1. To merge the data together, the urban centres boundaries from the GHS-SMOD dataset are first re-projected to the same Coordinate Reference System (CRS) as the raster data. Then, raster pixels falling within the urban boundaries are assigned to the corresponding urban centre, while those falling only partially within urban boundaries are weighted based on the share of surface overlapping the urban centre. Then, a weighted sum is computed for CO₂ emissions, population and GDP, and a weighted average for temperature and precipitations. The process of merging different spatial datasets is further illustrated in the Supplementary Information.

Data on CO₂ emissions is drawn from the European Commission’s Emissions Database for Global Atmospheric Research (EDGAR), providing yearly estimates of human-induced CO₂ emissions (in tonnes) from 1970 to 2022 at a 0.1 degrees resolution, approximately 11.1 km at the equator (16; 17). The estimates are available at the sectoral level: in our analysis, we include emissions from the residential sector (heating and cooling for buildings) and the on-road transport sector, as they have the highest potential to be affected by alterations in urban form and structure of the built environment (The sectorial classification follows the IPCC 1996 guidelines⁴⁷). Information on gridded GDP over the territory of the urban centre is drawn from an economic atlas²⁴ using national statistics from the World Bank and the CIA’s World Factbook, down-scaled at a 5 arcmin resolution (approximately 9 km at the equator) via triangulation with administrative sub-national data on GDP per capita⁴⁸ and population data from the HYDE database.

Data on variables such as temperature and precipitations are also important, as climate patterns may be relevant drivers of emission levels. Low temperatures, for instance, may require increased heating in buildings and therefore increased energy consumption and emissions. For this reason, we draw on information on temperature (in °C) and precipitations (in mm) from the CRU TS v4.06 dataset, where it is available at a resolution of 0.5 degrees (approximately 56 km at the equator) as an yearly historical series from 1901 to 2021²³. The data are estimated via interpolation of real measurement spatial information from weather station observations across the world. Information on population at a resolution of 1 km is merged using the GHS-POP dataset, which relies on spatial census data, down-scaled using processed satellite information on the distribution, density and classification of built-up surface²¹ (The population data used also come from the GHSL project and are available at the same resolution (1 km) and Coordinate reference System (CRS) as the GHS-SMOD data. Since they perfectly overlap the urban boundaries defined by aggregating contiguous raster pixels classified as “urban”, the weighted sum is equivalent to a simple sum of population values over the urban centre in this case.). The total urban population is then used to transform carbon emissions and GDP in per capita terms.

Finally, cities are assigned to their country, continent and World Bank income group using information on administrative boundaries from the Global Administrative Boundaries dataset (GADM)⁴⁹. In this case, the merged data are spatial polygons representing the boundaries of administrative entities across the world at different levels (municipalities, provinces, regions, countries and continents). We attach to each urban centre the information of the municipality overlapping its surface, and the corresponding country and continent. In cases where a city is intersecting multiple municipalities, we assign the name of the municipality (and corresponding information on country and continent) with the largest overlapping surface. We use this additional information to cluster errors at the country level in our regressions, to implement country specific time trends, for our analysis of heterogeneity by income and continent and for descriptive purposes.

Description of the data

To provide an overview of the urban centres in the database, Extended Data Table 6 displays descriptive statistics on the average values of key variable in cities of different continents, whereas Extended Data Fig. 2 plots them as points on a world map, on a colour scale based on the corresponding quintile in the distribution of CO₂ emissions in the residential sector. As expected, cities in developed regions such as North America and Europe have higher values of per capita emissions compared to their counterparts in Africa and most of Asia. Regarding the geographical distribution of urban centres, a majority of the world’s urban population lives in Africa and Asia (and within the Asian continent, as evidenced by Extended Data Fig. 2, in China and India, two countries that have experienced a booming urbanization rate in recent decades). Most cities in the panel are located either in lower middle income countries (40%) or in upper middle income countries (37%), and only a minority is located in low income (9%) or high income countries (14%). Urban form indicators are quite evenly distributed, even though some structural differences exist across continents (we look into this further below).

Extended Data Fig. 3 zooms in on total area and the urban form indicators, the main independent variables in our analysis. Specifically, it looks at changes in a simple average of the indicators over time and by continent. In terms of total area, there has been a slight convergence in the size of the average city across continents, as mean area has slightly increased for Asia, Africa and at the global level, while it has decreased in other regions. Nevertheless, urban centres in North America and Oceania remain far bigger on average than their counterparts in other continents. The CI and RI are persistent, and do not vary by a large amount over the time period considered (1975 to 2020). Their average values are quite similar (with the RI being slightly higher) reflecting the fact that they capture the same concept, urban form, from slightly different angles. In terms of geographical differences, urban centres in North America and Oceania appear to have smaller values of both the CI and RI, indicating less compact cities. African and European cities are instead more compact on average, as indicated by both the RI and CI. The CI and RI of Asian cities have both dropped from 1975 to 2020, suggesting that a deterioration in the compactness of these urban centres occurred in the time period considered. Finally, the sprawl index has decreased at the global level over time, and is lowest for Europe, South and North America.

Supplementary Information Fig. 4 plots time trends in average population, the share of residential built-up surface, GDP and CO₂ emissions per capita over the period of data availability (which varies depending on the variable) for urban centres in each continent. All cities have experienced a rise in the share of residential built-up surface, with the world average rising from 15% in 1975 to almost 20% in 2020. This suggests that as urbanization increased across the world, the greater need for residential buildings in already existing urban centres led to an increase in the surface devoted to residential use within their borders. It also reflects the decrease in the sprawl index across continents already displayed in Extended Data Fig. 3. In terms of emissions and GDP per capita, results are in line with expectations: urban centres in wealthier regions tend to emit more CO₂ on average, although they fare better now than they did in the past decades. The opposite happens in Asia, where urban emissions are lower but have increased over time, whereas Africa remains the continent with the smallest contribution to global CO₂ emissions per capita. Finally, urban GDP per capita of European, North American and Oceanian cities is above the world average, whereas Asian, African and South American cities are below.

Statistical information

In order to assess the impact of urban form on the emissions of CO₂, we estimate regressions with city-specific fixed effects that allow us to control for time-invariant factors, such as the geographic or institutional characteristics that may influence the emission levels of the city over time. The baseline specification is:

where is the natural logarithm of per capita CO₂ emissions of sector S (where S can be either the residential or the on-road transport sector) in city  i at time  t. is the logarithm of urban form indicator  F, corresponding to the area or to one of the three urban form indicators computed (), and is the corresponding coefficient. The remaining controls include , i.e. log GDP per capita, with corresponding coefficient , and , the vector of log environmental controls (namely, temperature and precipitations) with corresponding vector of coefficients . Finally, represents year specific fixed effects and is the error term, clustered at the country level. It should be noted that the use of natural logarithms for both the dependent and the independent variables of the regression allows for the interpretation of estimated coefficients as elasticities. The coefficient , for instance, should be interpreted as the percentage change in per capita emissions associated with a 1% change in the urban form indicator used.

To further explore the links between urban form and emissions, while exploiting the heterogeneity in the data, we also include interaction terms between compactness and the continent and income class of the country where the city is located, to assess whether the association detected is altered in groups of cities sharing more homogeneous characteristics. The regression equation takes the form:

where all variables are the same as in the previous fixed effects model, and is the natural logarithm of per capita CO₂ emissions of sector  S in city  i, and time  t. This time, the urban form indicator is interacted with a set of indicator functions taking value 1 if country c where the urban centre is located belongs to group G. The group can refer either to one of the World Bank income classes (with ), or to one of the 6 continents already employed for the computation of descriptive statistics (in that case, ). Omitting from the regression one of the continents or income groups, the coefficient on the urban form indicator represents the effect of urban form in the omitted reference group, whereas the coefficients convey the additional effect of urban form in the remaining groups. From such a model, we can estimate the marginal effect of each of the urban form indicators on per capita emissions, conditional on the city being located in a specific income class or continent, an approach allowing us to investigate heterogeneous effects.

Several methodological considerations should be noted when interpreting the results. First, although the use of city fixed effects and dynamic specifications helps mitigate concerns related to omitted variables, the empirical framework does not establish causal relationships, and reverse causality between urban form and emissions cannot be fully ruled out. Second, spatial dependence across cities may influence emission patterns, as geographically proximate urban areas often share infrastructure systems, economic linkages, and environmental conditions. While our global framework limits the feasibility of fully specified spatial econometric models, we partially address this issue by clustering standard errors at the country level and by incorporating country-specific time trends. Third, the construction of the dataset relies on the spatial aggregation of raster data, which assumes a uniform distribution of variables within grid cells and may introduce measurement error, particularly for emissions and economic activity. Finally, our analysis focuses on residential and on-road transport emissions, as these sectors are most directly affected by urban form, whereas industrial emissions and electricity generation are more strongly determined by production location and national energy systems. Moreover, the compactness indicators employed capture the geometric dimension of urban form and do not account for socio-spatial heterogeneity within cities; however, this approach allows for consistent and comparable measurement across a large global sample.

Author contributions

Marco Percoco has contributed to the conceptualisation of the research, to the design of the empirical analysis and to the final writing, whereas Giorgio Musto has contributed to the empirical analysis and to the final writing.

Funding

Musto and Percoco gratefully acknowledge financial support from Fondazione Invernizzi.

Data availability

Both the code used to perform the spatial join of the data (in the R Markdown language) and the final database are to be made publicly available. We are in the process of final cleaning of the code.

Declarations

Competing interests

The authors declare no competing interests.