h3-Cities: Standardized Urban Analytics Frameworks
- h3-cities are urban-analytics frameworks that standardize spatial representations—using grids, tiles, and network graphs—to compare heterogeneous urban systems.
- They integrate diverse methodologies from mobile traffic analysis, morphological typology, transport intermediacy, and 3D building metrics to derive measurable city fingerprints.
- The framework reveals universal urban patterns and challenges traditional aggregates by linking local neighborhood forms, temporal signatures, and infrastructural connectivity.
“h3-cities” is best understood as an Editor’s term for urban-analytics frameworks that impose a common spatial or network representation on cities so that heterogeneous urban systems can be measured comparatively. In the cited urban literature, this logic does not appear as a single named protocol; instead, it appears as a family of standardized observational scaffolds in which cities are projected onto common raster cells, sampled through fixed local tiles, embedded in continental transport graphs, or represented through Functional Urban Areas, and local measurements are then aggregated into comparable signatures, fingerprints, centrality scores, or scaling exponents [(Grauwin et al., 2014); (Nice et al., 2019); (Curiel et al., 2021); (Zhang et al., 21 Aug 2025)].
1. Standardized urban representation as the analytic core
A defining feature of h3-cities-style analysis is the replacement of locally idiosyncratic urban boundaries with standardized observational units. In comparative mobile-traffic analysis, New York, London, and Hong Kong are all projected onto a common grid, allowing temporal activity profiles to be compared on a shared spatial basis (Grauwin et al., 2014). In comparative morphology, each city is sampled within a population-scaled circular extent centered on the city centroid, with $1000$ map images per city and nearly $1.7$ million images across $1667$ global cities; the local unit is not an administrative district but a regularly sampled tile (Nice et al., 2019). In African transport-network analysis, the city system is embedded in a continental road graph built from OpenStreetMap and Africapolis, yielding a network of cities connected through standardized road vertices and edges (Curiel et al., 2021). In three-dimensional urban scaling, the observational unit is the Functional Urban Area, which combines a high-density urban center with its commuting zone and supports a sample of $2903$ cities in $42$ countries in $2015$ (Zhang et al., 21 Aug 2025).
These representations differ in geometry and semantics, but they share the same epistemic objective: they convert cities into commensurable objects. A plausible implication is that the relevance of H3 to city analysis lies less in any particular tessellation than in this broader requirement for a common spatial scaffold. The morphology paper makes this especially explicit by noting that it does not use H3 directly, while also describing a workflow that could be adapted by replacing centroid-scaled image tiles with H3 cells (Nice et al., 2019).
The significance of this standardization is methodological rather than merely cartographic. It enables cross-city inference about which observed regularities are universal, which are local, and which depend on the chosen layer of representation—morphology, activity, accessibility, or built form. Without such common supports, city comparison collapses into juxtaposition rather than measurement.
2. Morphological typologies and city fingerprints
One of the clearest h3-cities analogues is the morphology-to-typology pipeline developed for global city comparison. The analysis begins with $1667$ global cities with populations exceeding , after excluding $1000$0 South Korean cities from $1000$1 initial candidates. For each city, a population-scaled sampling radius is defined as
$1000$2
Within this extent, $1000$3 map images are sampled. Each tile represents approximately $1000$4 m to $1000$5 m depending on latitude, with high-latitude correction used to preserve spatial consistency (Nice et al., 2019).
The morphology descriptors are deliberately low-level. White regions are flood-filled, each enclosed block or region is identified, pixel count gives block size, and the difference between the region and its smallest bounding box provides a measure of irregularity. Block size and block regularity are each represented with $1000$6 histogram bins, and the feature vector is augmented by five color-pixel proportions corresponding to roads or transport network, green space, water, public transport, and remaining space. The resulting $1000$7-dimensional vector is normalized to $1000$8 and used to train a $1000$9 self-organizing map. Training uses about $1.7$0 million sampled data points, with $1.7$1 million total iterations (Nice et al., 2019).
After training, each tile is assigned to its closest SOM node. The SOM thus becomes a two-dimensional typology space of neighborhood forms, and each city is represented by the distribution of its sampled tiles across that space. The paper defines a city fingerprint as the kernel-density-smoothed distribution of SOM-node locations used by that city. In substantive terms, the fingerprint is the weighted distribution of neighborhood typologies present in the city.
The reported correlations indicate that these morphology-based city summaries track functional and environmental quantities. City-level averages derived from the SOM-based framework are associated with moving vehicles ($1.7$2), impervious surfaces ($1.7$3), sky fraction ($1.7$4), building fraction ($1.7$5), mean aerosol optical depth ($1.7$6), and mean $1.7$7 ($1.7$8) (Nice et al., 2019). This supports a strong claim about h3-cities-style workflows: standardized local samples can recover not only form but also transport intensity and environmental burden.
A common misconception is that city comparison can be adequately conducted through citywide averages such as density or GDP. The morphology results argue otherwise. Their object of comparison is the mix, distribution, and composition of neighborhood types, not a single urban aggregate. This suggests that the principal comparative unit is often the recurrent local pattern rather than the municipality.
3. Temporal signatures, inequality, and land-use inference
A second major strand of h3-cities analysis uses standardized activity signatures rather than morphology. In comparative mobile-traffic analysis, the underlying data consist of calls, SMS, requests, and upload/download counters recorded every $1.7$9 minutes at cell level over three months. After projection onto the common $1667$0 grid, each pixel $1667$1 and activity type $1667$2 is represented by a “typical week” time series $1667$3 with $1667$4 fifteen-minute intervals. To compare temporal shape rather than volume, the activity is normalized by its weekly mean: $1667$5 These normalized “signatures” emphasize recurrent behavior and suppress event noise because they average corresponding time slots across weeks and exclude civic holidays (Grauwin et al., 2014).
The same framework supports both spatial and temporal comparison. Spatial concentration is summarized through Lorenz curves and Gini coefficients $1667$6. For request activity, the Gini coefficients are about $1667$7 in London, $1667$8 in New York, and $1667$9 in Hong Kong, with the same ordering for calls, SMS, and data: Hong Kong is the most concentrated, London is intermediate, and New York is the least concentrated of the three (Grauwin et al., 2014). The important implication is that “city structure” matters more than the type of mobile counter.
At the city scale, the temporal signatures display a near-universal rhythm: activity rises in the morning, remains relatively steady through the day, falls at night, and is higher on weekdays than weekends. Yet the departures from this common form are analytically significant. London’s request and data signatures fall earlier in the evening and dip more on weekends; the paper interprets this through Wi-Fi substitution under higher mobile-data costs. Hong Kong’s SMS signature drops earlier in the evening than other activity types, consistent with a stronger daytime SMS culture. New York shows evening SMS peaks linked to media participation such as SMS voting (Grauwin et al., 2014).
The local land-use inference is performed with K-means on a 0-dimensional feature space defined by five activity types over 1 time intervals. For a cluster 2, the cluster signature is
3
and the clustering objective is
4
with
5
Using the silhouette index as a guide, the analysis selects 6, with local maxima at 7 and 8, and interprets 9 as the most meaningful nontrivial resolution (Grauwin et al., 2014).
The main empirical conclusion is that core business districts are remarkably similar across cities, whereas residential and mixed zones remain much more city-specific. When all pixels from New York, London, and Hong Kong are pooled and clustered jointly, the financial and decision-making cores of Manhattan, the City of London, and Central in Hong Kong fall into the same cluster, while residential zones separate by city (Grauwin et al., 2014). This is one of the strongest comparative results in the literature: globalization appears to produce common temporal signatures in central business districts, while everyday routines remain path-dependent.
4. Corridor structure, degree, and urban intermediacy
In transport-network analysis, h3-cities logic is applied to roads rather than grids. The African urban network is built from OpenStreetMap primary roads, highways, and trunks using the Africa OSM dump downloaded on March 17, 2021. The raw dataset contains $2903$0 million coordinates, $2903$1 road segments, and about $2903$2 km of roads. Starting from Africapolis urban agglomerations, the analysis retains cities with more than $2903$3 inhabitants and additionally promotes a road node to city-node status when it lies within $2903$4 km of a smaller city, yielding a main network of $2903$5 cities. After connectivity repair and graph simplification, the final network contains $2903$6 nodes and $2903$7 edges, of which $2903$8 were added to ensure connectivity (Curiel et al., 2021).
The core quantity is intermediacy, defined as a city’s position in the transport network in terms of how many journeys pass through it, not merely how many roads meet there. This is distinct from degree. Degree counts network connections, so degree-$2903$9 cities are terminal ends, degree-$42$0 cities are corridor settlements, and higher-degree cities are junctions or hubs. Intermediacy instead measures the estimated number of journeys passing through a city, computed as a weighted betweenness-like quantity based on modeled origin-destination flows (Curiel et al., 2021).
Those flows are generated by a gravity model. The general form is
$42$1
and the operational form is
$42$2
where $42$3 is the shortest-path network travel time. The model sets $42$4 and $42$5, leaving $42$6 as the free parameter; calibration using foodshed and trade-distance patterns for Tamale and Ouagadougou yields $42$7. Border-crossing time enters as an added cost $42$8 on edges spanning countries, and city crossing is penalized by a term proportional to $42$9, with $2015$0 (Curiel et al., 2021).
The empirical findings show that African city importance cannot be reduced to population. Small cities have a very wide range of intermediacy, and a phase transition appears around one million inhabitants. Below roughly $2015$1 million, centrality depends strongly on both population and degree; above $2015$2 million, centrality is generally larger and is driven primarily by city size, though degree still matters (Curiel et al., 2021). The paper reports that cities with degree $2015$3 or above are about $2015$4 more central than degree-$2015$5 small cities, and that for transport nodes each additional degree raises centrality by about $2015$6 on average. Degree itself scales sublinearly with population: $2015$7 with $2015$8 and $2015$9, implying that increasing population by a factor of $1667$0 raises expected degree by only about $1667$1 (Curiel et al., 2021).
The Shashemene–Yirga Chefe contrast illustrates the substantive meaning of intermediacy. Shashemene, with about $1667$2 inhabitants, ranks only eleventh by population in Ethiopia but third in centrality, with degree $1667$3, because it sits on a strategic north-south and east-west corridor. Yirga Chefe is larger at about $1667$4 inhabitants, but has degree $1667$5 and only about $1667$6 of Shashemene’s centrality because it is far from major corridors (Curiel et al., 2021). This directly contradicts the misconception that secondary cities are necessarily peripheral in transport significance.
Regional heterogeneity is also pronounced. North Africa is described as dense and complex, especially along the Nile corridor in Egypt, whereas Central Africa is sparse and fragmented, with longer network distances and stronger border effects. In Benin and Togo, total trade falls from $1667$7 units to about $1667$8 units when border costs are two hours or more (Curiel et al., 2021). The broader implication is that intermediacy is simultaneously a measure of accessibility, vulnerability, and exposure to political fragmentation.
5. The third dimension: horizontal and vertical accommodation
Most urban comparison has historically been two-dimensional. The third-dimension study argues that this is a major omission because building height “dramatically potentiates the interior space of cities.” Using the 3D-GloBFP dataset, described as the first global three-dimensional building footprint dataset, the analysis overlays building footprints and heights with Functional Urban Areas to obtain a sample of $1667$9 cities in 0 countries in 1. For each city it computes total building footprint area 2, total building volume 3, and average building height 4 (Zhang et al., 21 Aug 2025).
The central model is a Cobb-Douglas specification of population accommodation: 5 The paper links this to the simpler scaling relations
6
from which it derives
7
and
8
Empirically, 9, which explains the negative relationship observed between $1000$00 and $1000$01 (Zhang et al., 21 Aug 2025).
The main result is asymmetry between horizontal and vertical extent. Across most countries, $1000$02, usually with $1000$03, whereas $1000$04 is usually near zero. The paper states that “Most countries exhibit $1000$05, indicating that urban building height is not associated with population size.” In some cases height is negatively associated with population; Egypt, Morocco, Bangladesh, and the Philippines are cited as examples with $1000$06, while Germany is given as a case with $1000$07 (Zhang et al., 21 Aug 2025). At country level, the reduced major axis fit is
$1000$08
This finding challenges the common intuition that high-rise development substantially increases population accommodation. The paper’s substitution analysis defines
$1000$09
with $1000$10 interpreted as no strong interaction between area and height, $1000$11 as complementarity, and $1000$12 as effective substitution. Most countries exhibit $1000$13, reinforcing the conclusion that area and height are often not strongly substitutable in a way that materially changes accommodation capacity (Zhang et al., 21 Aug 2025).
The heterogeneity analysis further embeds verticality in broader urban-system structure. Lasso regression identifies the $1000$14-exponent of the city-size distribution $1000$15 as the most influential factor for $1000$16. Higher $1000$17, meaning more uneven urban concentration in a few large cities, is positively related to $1000$18 and negatively related to $1000$19. Urban population share and per capita GDP are positively correlated with $1000$20, but the paper notes that these correlations are very weak (Zhang et al., 21 Aug 2025). Robustness checks using OpenStreetMap land-use classes in Germany show that residential buildings best explain population accommodation, yet the exponents remain in the same range as those obtained from all buildings, leading the authors to conclude that total building stock serves as a reasonable proxy. The policy implication is narrow but direct: planning should not assume that taller buildings automatically deliver efficient densification.
6. Vacancy detection, seasonal occupancy, and urban function
Another major use of standardized local units is the detection of underoccupied urban space. The ghost-city study uses Baidu positioning data and Baidu Map POI data to identify vacant housing areas in China. The positioning data contain anonymized user ID, latitude, longitude, and timestamp, at the scale of several billions of points per day from 2014/9/8 to 2015/4/22. Home locations are inferred by applying DBSCAN to positioning points collected from $1000$21 am to $1000$22 pm, with neighborhood distance $1000$23 meters and $1000$24; the cluster with the largest number of points is treated as the user’s location (Chi et al., 2015).
Residential areas are identified from POIs categorized as residential area or villa, with residential POIs within $1000$25 km of villas removed because villa areas have very low density. Since the POIs are points rather than polygons, the analysis introduces a $1000$26 grid. For each residential POI, a $1000$27 grid window is centered on the POI, the top six most populated grids are selected, and their populations are summed. Using assumptions of average floor area ratio $1000$28 and average living area $1000$29 per person, a $1000$30 grid is estimated to hold about $1000$31 people, or about $1000$32 Baidu users after adjusting for coverage. A residential area is defined as vacant if the sum of the top six most populated grids is less than $1000$33, while sums above $1000$34 are used to exclude very newly built areas with too few residents (Chi et al., 2015).
The spatial distribution of vacant housing areas is not uniform. They are concentrated mostly in second-tier and third-tier cities, more common in eastern provinces, and often located in urban peripheries, new towns, or coastal tourism-oriented zones. The study identifies $1000$35 cities with large vacant housing areas, including Rushan, Ordos or Kangbashi, Binhai New Area, Rugao, Dongying, Xinghua, Yijinhuoluoqi, and Dongling, although the paper avoids presenting a strict ranking because of possible real-estate effects (Chi et al., 2015).
A central conceptual correction is that not all apparent vacancy means the same thing. The paper distinguishes genuine city vacancies from tourism-oriented housing areas by examining whether population increases during holidays. It tests this using National Day, New Year’s Day, and International Workers’ Day dates, classifying an area as a tourism area if population increases during holiday periods (Chi et al., 2015). The contrast between Kangbashi and Rushan Yintan is illustrative. Kangbashi exhibits a clear workday/weekend structure, population drops during holidays, no strong growth trend from September 2014 to April 2015, and substantial commuting from Dongsheng to government and administrative workplaces in Kangbashi. Rushan Yintan, by contrast, shows strong seasonality, a large increase in summer and holidays, strong holiday sensitivity on National Day, weak work-related structure, and tourism-related migration patterns. This directly addresses a recurrent misconception: apparently empty urbanized space is not always evidence of structural vacancy; it may instead indicate seasonal occupancy.
7. Urban technologies, Triple Helix dynamics, and smart(er) cities
Beyond measurement, the literature also provides a systems-theoretic interpretation of why standardized urban analytics matter. In the Triple Helix model, city analysis is embedded in relations among university, industry, and government, understood not only as institutional networks but as three functionally differentiated selection environments: organized knowledge production, economic wealth creation, and reflexive control (Leydesdorff et al., 2010). The key question is not merely whether institutions are connected, but whether synergy is generated among functions in networks of relations.
The formal vocabulary is evolutionary. Two selection environments operating upon one another may generate a trajectory in a process of mutual shaping. Three selection mechanisms operating upon one another can generate complex dynamics. An additional third feedback may induce meta-stabilization or alternatively a hyper-stabilized lock-in, and a further selection upon stabilizations can lead to globalization (Leydesdorff et al., 2010). Cities are therefore not treated as passive containers of innovation but as sites where local stabilization and global development regimes intersect.
Within this perspective, urban technologies are modeled as densities in networks among the dynamics of organized knowledge production, wealth creation, and governance of civil society. The paper states that the densities of relations among the three institutional spheres allow the technologies of cities to function as key components in the organization of innovation systems (Leydesdorff et al., 2010). It also stresses that smart cities are not simply ICT-saturated cities. Their “smart-er” character lies in reflexive coordination across differentiated functions, in the capacity to avoid lock-in, and in the ability to connect local regeneration with global innovation dynamics.
This suggests that h3-cities should not be reduced to a geospatial indexing problem. Standardized grids, tiles, signatures, fingerprints, and network measures can also be interpreted as empirical observables of broader knowledge-based urban systems. In that reading, the value of h3-cities lies not only in comparability, but in enabling systematic inquiry into how morphology, mobility, accessibility, occupancy, and governance co-vary across cities and across scales.