UrbanGraph: Graph-Based Urban Modeling

• UrbanGraph is a graph-based representation of urban systems that captures connectivity, semantics, and temporal dynamics.
• It models diverse urban entities including street networks, multi-view imagery, and regional dependencies through specialized graph constructions.
• Its learning paradigms integrate generative, discriminative, and physics-informed methods to enable precise urban analysis and forecasting.

UrbanGraph refers to graph-based representation and learning strategies for urban systems in which roads, regions, observations, objects, infrastructures, and environmental states are modeled through explicit relational structure rather than only as independent records or raster images. In this literature, graphs are used to represent street intersections and street segments, multi-view detections of roadside assets, region-to-region dependencies derived from mobility and semantics, hierarchical scene structure in built environments, evolving RDF city states, and dynamic physical interactions in urban climate fields (Neira et al., 2022, Nassar et al., 2020, Xin et al., 1 Oct 2025). The unifying premise is that urban phenomena are inherently relational: topology, neighborhood structure, semantic heterogeneity, and temporal evolution are often first-class objects of analysis rather than secondary annotations.

1. Conceptual foundations

A central UrbanGraph principle is that urban systems should be learned and analyzed in their native graph form. For street networks, the canonical representation is an undirected, unweighted graph $G=(V,E)$ in which nodes are intersections and edges are street segments; this preserves exact connectivity and neighborhood structure that rasterization discards (Neira et al., 2022). The same relational stance appears in generalized space-syntax work, where a geolocalized primary graph $G_p=(V,E)$ is combined with an edge-label map $c:E\rightarrow C$, transformed into a line graph, and then contracted over adjacent equal-labeled vertices so that axial analysis becomes a special case of a broader graph construction over arbitrary urban quantities (D'Autilia, 2015).

UrbanGraph also extends beyond static infrastructure. In multi-view street imagery, the task is not merely per-image detection but observation-to-object reasoning: detections from all views become graph nodes, pairwise same-instance relations are predicted on graph edges, and connected components define physical urban objects such as trees or traffic signs (Nassar et al., 2020). In urban knowledge management, a versioned graph is treated semantically as “a collection of graphs, that is one graph for each version,” reflecting not only historical states but also branching planning scenarios (Gil et al., 2024).

These formulations share a common implication: urban structure is not exhausted by Euclidean proximity. A graph may encode adjacency, same-instance identity, functional similarity, semantic compatibility, or validity across concurrent city versions. This suggests that UrbanGraph is best understood as a family of graph-native urban representations rather than a single schema.

2. Graph constructions and urban entities

UrbanGraph research spans multiple graph constructions, each tailored to a different urban substrate.

Representation	Nodes	Edges / relations	Representative paper
Street network graph	Intersections	Street segments	(Neira et al., 2022)
Region-level multi-graph	Urban regions	Mobility, POI-function, and check-in similarity	(Chen et al., 2023)
Primitive graph	Points or line segments	Connectivity or pairwise geometric relations	(Wang et al., 2022)
Hierarchical scene graph	Objects, scenes, rooms, buildings	Cross-layer spatial-semantic links	(Longo et al., 27 Jun 2025)

At the region level, ATGRL partitions a city into $N$ regions $V=\{v_1,v_2,\cdots,v_N\}$ and constructs four graph views: two mobility-context graphs $\mathcal G_O$ and $\mathcal G_D$, a function graph $\mathcal G_F$, and a semantics graph $\mathcal G_S$. The edge weights are derived from cosine similarity over origin and destination distributions, POI-category vectors, and check-in vectors, so regions can be strongly related even when they are geographically distant (Chen et al., 2023).

At the geometry-first end of the spectrum, GraphMapper defines a primitive graph

$$G=\{V,E\}, \quad V\in \mathbb{R}^{N \times d}, \quad E\in \mathbb{Z^+}^{N\times N},$$

where primitives are either points $G_p=(V,E)$0 for road mapping or line segments $G_p=(V,E)$1 for building mapping. The same formalism then supports road-network topology reconstruction and building-footprint regularization (Wang et al., 2022).

Hierarchical constructions are also prominent. Pix2G organizes built environments into object, scene, room, and building layers linked to a de-noised 2D top-down map and a structure-segmented 3D point cloud, while OpenGraph uses a five-layer hierarchy consisting of point cloud, lane graph, instance, segment, and environment layers for large-scale outdoor environments (Longo et al., 27 Jun 2025, Deng et al., 2024). In the named microclimate framework “UrbanGraph,” the city is discretized into grid-cell nodes and represented as a graph sequence

$G_p=(V,E)$2

with static node set $G_p=(V,E)$3, time-varying edge set $G_p=(V,E)$4, and relation set $G_p=(V,E)$5 (Xin et al., 1 Oct 2025).

3. Learning paradigms

UrbanGraph methods cover generative, discriminative, multimodal, and physics-informed graph learning.

A representative generative formulation appears in graph representation learning for street networks. There, the objective is to learn $G_p=(V,E)$6 while modeling

$G_p=(V,E)$7

with a Transformer node model for spatial node distributions and a VGAE for topology. The VGAE is trained through the variational lower bound

$G_p=(V,E)$8

The decoder reconstructs a probabilistic fully connected graph via inner products in latent space (Neira et al., 2022).

GeoGraph exemplifies end-to-end discriminative graph reasoning from images and coarse camera metadata. It builds a fully connected undirected graph over detector-anchor features, concatenates visual features with source-image metadata and bounding box predictions, and uses 3 GraphConv layers followed by binary edge classification to decide whether two detections correspond to the same physical object (Nassar et al., 2020). The graph formulation supports arbitrary numbers of views and unifies detection, re-identification, and geo-localization in a single pipeline.

ATGRL illustrates multi-graph region embedding. It first constructs four regional graphs, applies a soft-threshold graph cleansing layer, then uses a multi-graph aggregation module and a dual-stage fusion module with an improved linear attention mechanism. The final objective is

$G_p=(V,E)$9

combining origin-destination prediction with function and semantics preservation (Chen et al., 2023).

Urban incident prediction introduces a different multiview pattern: latent-state inference under biased observations. A neighborhood graph over NYC Census tracts is combined with sparse government inspection ratings and dense but demographically biased crowdsourced reports. The latent true incident state is predicted as

$c:E\rightarrow C$0

while the report process is modeled through a logistic head conditioned on latent state and demographic covariates (Balachandar et al., 10 Jun 2025). This reframes urban graph learning as measurement correction, not only forecasting.

The named framework “UrbanGraph” (Xin et al., 1 Oct 2025) pushes the paradigm toward physics-informed spatio-temporal dynamic heterogeneous graphs. It encodes shading, vegetation evapotranspiration, and convective diffusion directly into time-varying edge types and processes each graph snapshot with a 3-layer RGCN before temporal fusion with an LSTM. This is physics-informed through graph construction rather than through explicit PDE-residual losses.

4. Perception, mapping, and scene graph construction

A major UrbanGraph theme is the conversion of raw sensing into graph structure.

GeoGraph maps multi-view street imagery to a dense observation graph whose nodes are view-specific detections carrying appearance and coarse camera metadata. Same-instance edges are inferred by the GNN, thresholded at 0.5, and converted into connected components that represent distinct urban objects; geo-coordinates are then refined and averaged across observations within each component (Nassar et al., 2020). This is a direct graph-based pipeline for urban object inventory from publicly available street imagery.

From overhead imagery, GraphMapper casts vector map extraction as primitive graph learning. A CNN-FPN encoder predicts segmentation and keypoint maps, sampled primitives are refined through a Primitive Learning Structure with multi-head attention, and a second structure predicts pairwise primitive relations. For roads, points are connected according to learned connectivity; for buildings, line segments are merged according to learned geometric relations (Wang et al., 2022). The result is explicit vector topology rather than raster masks.

Sensor-fusion scene graphs extend this logic to hierarchical environment modeling. OpenGraph extracts open-vocabulary instances and captions from images, projects them onto LiDAR point clouds, incrementally fuses them into persistent 3D objects, and organizes them under a lane-graph-derived hierarchy for large-scale outdoor environments (Deng et al., 2024). Pix2G, by contrast, is building-scale: it generates a de-noised 2D top-down environment map, segments structural components in the BEV image domain, backprojects masks to 3D point clouds, and constructs a multi-layer graph from object level up to building level (Longo et al., 27 Jun 2025).

UrbanGraphEmbeddings extends perception-to-graph pipelines into multimodal representation learning. UGData anchors street-view images to heterogeneous urban spatial graphs, provides graph-aligned supervision through Spatial Reasoning Paths and Spatial Context Captions, and trains graph-grounded multimodal embeddings with a two-stage curriculum that first aligns images and text and then injects localized spatial graphs through a GNN-based encoder (Zhang et al., 9 Feb 2026). This makes the graph not only a mapping output but also a supervision substrate for transferable urban embeddings.

5. Analysis, visualization, and urban graph systems

UrbanGraph is also an analytic and systems paradigm.

For mobility analysis, graphs are constructed either as Region Adjacency graphs or as weighted directed Origin-Destination graphs. The corresponding toolkit includes degree, closeness, harmonic, betweenness, and PageRank centralities, Perron-induced circulation $c:E\rightarrow C$1, and multiple Laplacians for both undirected and directed cases (Márquez et al., 2023). Metro traffic modeling adds station-centric graph analysis, betweenness-based importance, $c:E\rightarrow C$2-edge augmentation for robustness, Laplacian diffusion of commuter flows, and GNN/HGNN models for rush-hour forecasting on the London Underground (Dees et al., 2021).

Visualization has produced distinct cartographic graph metaphors. GMap turns relational data into map-like regions through embedding, clustering, and Voronoi-based region generation, emphasizing countries, islands, lakes, and peninsulas as carriers of community structure (0907.2585). GraphMaps addresses large-graph browsing through zoom levels $c:E\rightarrow C$3, rail-based edge routing, and a balanced visualization objective

$c:E\rightarrow C$4

which minimizes abrupt changes in visible nodes across zoom transitions (Mondal et al., 2017).

Urban graph systems also include declarative and lifecycle infrastructures. The Urban Toolkit introduces layers, knots, maps, and plots as a grammar for urban visual analytics; its “knot” abstraction links thematic and physical layers through spatial relations such as nearest, contains, and within, followed by aggregation or custom operations (Moreira et al., 2023). Graph versioning for evolving urban data argues that ordinary RDF storage, temporal databases, and Git-like systems are insufficient when city data branch across historical and hypothetical scenarios; the proposed direction is to attach version metadata to RDF triples so that a query engine can “compute partial answers for several versions at once” (Gil et al., 2024).

6. Applications, limitations, and research directions

UrbanGraph methods already cover a wide application range. Street-network graph learning supports synthetic street generation and clustering of urban morphology (Neira et al., 2022). Graph-based roadside mapping supports municipal asset inventory, road-sign databases, and tree inventories (Nassar et al., 2020). The named UrbanGraph framework targets high-resolution urban microclimate prediction and reports up to 10.8% improvement in $c:E\rightarrow C$5 with 17.0% lower FLOPs over baselines on UMC4/12 (Xin et al., 1 Oct 2025). CureGraph uses multimodal neighborhood graphs for elderly disease-risk profiling in 15-minute living circles (Li et al., 13 Jan 2025), while the incident-prediction model uses neighborhood graphs to infer latent urban conditions from sparse ratings and biased 311 reports (Balachandar et al., 10 Jun 2025).

The literature is equally explicit about limitations. GeoGraph’s fully connected graph scales quadratically in the number of detections, depends on coarse camera metadata, and is demonstrated only for static roadside objects (Nassar et al., 2020). The street-network VGAE works on small fixed-size graph patches, has limited latent interpretability, and shows artifacts such as overproduction of degree-2 nodes (Neira et al., 2022). Pix2G is optimized for indoor or floor-based built environments rather than city-scale outdoor graphs (Longo et al., 27 Jun 2025). CureGraph relies on static images and text, does not incorporate road network data, and is evaluated with simulated disease prevalence labels (Li et al., 13 Jan 2025). The graph-versioning work remains a design agenda without benchmark-scale empirical evaluation (Gil et al., 2024). UrbanGraphEmbeddings shows strong gains in retrieval and geolocation, but fine-grained metric and directional reasoning remains difficult (Zhang et al., 9 Feb 2026).

Future directions are converging on several themes. One is sparsity and hierarchy: GeoGraph explicitly suggests learned sparse neighborhoods or hierarchical graph pooling for city-scale batches (Nassar et al., 2020). Another is adaptive structure learning: the UrbanGraph microclimate paper identifies adaptive graph learning as a way to move beyond predefined physical processes (Xin et al., 1 Oct 2025). Systems work points toward native version-aware query engines over branching urban RDF graphs (Gil et al., 2024). Multimodal urban representation learning points toward richer graph grounding across images, text, infrastructure, and neighborhood context (Zhang et al., 9 Feb 2026). Taken together, these directions suggest that UrbanGraph is evolving toward urban representations that are simultaneously geometric, semantic, temporal, and operational: graphs that can be learned from perception, queried across versions, fused across modalities, and maintained as persistent substrates for planning, analytics, and urban computation.