🤖 AI

OSMGraphCLIP: Learning Global Location Representations from OpenStreetMap Graphs

OSMGraphCLIP is a novel CLIP-style model that learns robust global location representations by encoding OpenStreetMap's heterogeneous graph structures, achieving performance comparable to or exceeding satellite-based methods across diverse geospatial tasks, particularly in socioeconomic and public health domains where explicit semantic annotations of the built environment provide superior insights.

Original authors: Dimitrios Michail, Eleni Saka, Ioannis Giannopoulos, Ioannis Papoutsis

Published 2026-06-09

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Dimitrios Michail, Eleni Saka, Ioannis Giannopoulos, Ioannis Papoutsis

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to describe a city to a friend who has never been there. You could take a photo (satellite imagery) and say, "Look, I see a lot of gray squares and green patches." Or, you could hand them a detailed map and say, "Here, this gray square is a school, that green patch is a park, and this line is a highway connecting to a hospital."

The paper introduces OSMGraphCLIP, a new AI model that learns to understand "where" a place is by reading that detailed map (OpenStreetMap) instead of looking at photos.

Here is the breakdown of how it works and what it found, using simple analogies:

1. The Problem: Photos vs. Maps

Most AI models that try to understand geography rely on satellite photos. They look at the Earth like a camera does: "That's a forest," "That's a city," "That's water." This is great for seeing what things look like.

However, the authors argue that to truly understand a place, you need to know what things do and how they connect. A photo might show a building, but it can't tell you if it's a hospital, a bakery, or a factory. It also can't easily see that a road connects a school to a park.

OSMGraphCLIP decides to skip the camera entirely. Instead, it learns from OpenStreetMap (OSM), which is like a giant, community-built digital map where every road, building, and park is tagged with a label (e.g., "residential," "restaurant," "highway").

2. The Solution: Turning Maps into "Social Networks"

The model treats the map not as a picture, but as a social network of objects.

The Nodes (The People): Every road, building, and park is a "person" in this network.
The Edges (The Handshakes): The model looks at how these objects relate. A road touches a building; a park is inside a neighborhood; a river crosses under a bridge.
The Conversation: The AI uses a special "graph encoder" to let these objects "talk" to each other. It learns that a cluster of "restaurants" near a "highway" means something different than a cluster of "factories" near a "river."

It also looks at the "neighborhood" in layers. It looks at the immediate street (2km), the wider district (10km), and the region (20km), kind of like zooming out from a single house to the whole city.

3. The Test: Can a Map Replace a Photo?

The researchers trained this model on 180,000 different locations around the world using only map data. Then, they tested it on 24 different tasks to see if it could guess things about those locations, just like models that use satellite photos do.

The Results:

The "Human" Tasks (Where the Map Wins): When the task was about human life—like predicting median income, public health (e.g., diabetes rates, obesity), or housing prices—the map-based model was often the best or tied for best.
- Why? Because human activity leaves a specific "signature" on a map. A wealthy neighborhood has specific types of roads, schools, and shops. A sickly neighborhood might lack certain amenities. The map explicitly tells the AI these facts, whereas a photo has to guess them indirectly.
The "Nature" Tasks (Where Photos Still Shine): When the task was about pure nature—like predicting bird species or wildfire risk—the satellite photo models were still better.
- Why? Because a map might say "forest," but it can't tell you if the trees are tall, dense, or dry (which matters for fire). A photo sees the actual texture of the leaves and the dryness of the ground.
The Surprising Middle Ground: Even on nature tasks, the map model was surprisingly competitive. It could distinguish between a tropical rainforest and a temperate forest just by looking at the types of roads and land-use tags, without ever seeing a single pixel of the forest.

4. The Big Takeaway

The paper concludes that OpenStreetMap data is powerful enough on its own to create a strong "sense of place" for an AI.

Satellite Imagery is like seeing a person's face: it tells you what they look like right now.
OSM Graphs are like reading a person's biography and social connections: it tells you who they are, what they do, and how they fit into the community.

The model, OSMGraphCLIP, proves that if you want to understand the human side of geography (cities, economies, health), reading the map is often better than taking a picture. It creates a "global location brain" that understands the world through its structure and labels, not just its colors.

Technical Summary: OSMGraphCLIP

Problem Statement

The paper addresses the challenge of learning globally transferable location representations in geospatial machine learning. Current dominant paradigms, such as GeoCLIP and SatCLIP, rely on contrastive alignment between geographic coordinates and visual context (ground-level or satellite imagery). While effective, these methods depend on Earth observation data, which captures spectral and physical characteristics but may only indirectly encode the functional and semantic organization of the built environment. The authors posit that structured, collaboratively curated map data—specifically OpenStreetMap (OSM)—offers a complementary perspective. OSM provides explicit semantic annotations (e.g., highway:motorway, landuse:residential) and native topological relationships (intersections, containment, adjacency) that define human activity and urban structure. The core research question is whether structured OSM data, devoid of any Earth observation imagery, is sufficient to learn strong global location representations that can transfer to diverse downstream tasks.

Methodology

1. Data Curation and Location Selection

The authors constructed a large-scale pre-training corpus of approximately 200,000 geographically diverse locations. This set comprises:

100,000 coordinates from the SatCLIP benchmark to ensure global coverage and enable direct comparison.
100,000 coordinates generated via a density-stratified H3 sampling strategy. This approach prioritizes OSM-rich regions (cities, transport corridors) while maintaining global coverage, using a hierarchical hexagonal index to prevent clustering and ensure diversity from urban cores to wilderness.
After preprocessing, the final training set contains ~180,000 locations, with ~20,000 held out for validation.

2. Graph Construction

For each location, the system retrieves OSM features (points, linestrings, polygons) within a bounding box and constructs a heterogeneous graph:

Nodes: Categorized into three types: points (amenities, POIs), linestrings (roads, waterways), and polygons (buildings, land-use zones).
Node Features: A concatenation of:
- Semantic Embeddings: All key-value tags for a feature are combined into a text description and encoded using a pre-trained Sentence-BERT model.
- Geometric Descriptors: Normalized coordinates, centroids, endpoints, or interior sample points.
Edges: Defined by topological relationships using the DE-9IM spatial predicate framework (e.g., touches, overlaps, within). For point-to-point pairs lacking topological predicates, Delaunay triangulation establishes connectivity.

3. Model Architecture

OSMGraphCLIP follows a CLIP-style contrastive alignment framework with two main encoder branches:

Context Encoder (OSM Graph):
- Graph Encoder: Utilizes a Heterogeneous Graph Attention Network (GAT) to process node features across different edge types, producing relation-aware embeddings.
- Pooling: Set2Set pooling aggregates node embeddings, followed by a learned attention mechanism to combine type-specific vectors into a fixed-dimensional graph embedding ( $g$ ).
- Multi-Scale Variants: The "Multiscale" (MS) variant supplements the local graph (fixed at 1km) with a Band Attention Encoder. This Transformer-based module aggregates spatial statistics and semantic embeddings from concentric radial bands at 2km, 10km, and 20km radii, capturing broader landscape context.
- Adaptive Variants: The "Adaptive" (A) variant dynamically selects the optimal bounding box scale (from 200m to 20km) based on a composite richness score of semantic content.
Location Encoder:
- Adopts the SatCLIP architecture: Geographic coordinates are mapped to the unit sphere, encoded via spherical harmonics (up to Legendre degree $L=10$ or $L=40$ ), and processed by a SIREN (Sinusoidal Representation Network) to produce the location embedding ( $c$ ).
Training Objective:
- The model minimizes a symmetric contrastive loss (InfoNCE) between the context embedding ( $g$ or $z_{fused}$ ) and the location embedding ( $c$ ) over batches of paired data.

Key Contributions

OSMGraphCLIP Model: A novel geospatial representation model that learns globally transferable embeddings using exclusively structured OSM graphs, eliminating the need for satellite imagery. It combines a heterogeneous graph encoder with multi-scale band encoders.
Scalable Pre-training Corpus: Development of a planetary-scale extraction pipeline involving a density-stratified H3 sampling strategy and topology-aware graph construction via DE-9IM relations, resulting in a dataset of ~180k diverse locations.
Comprehensive Evaluation: Demonstration that structured OSM graphs alone yield competitive global representations across 24 downstream tasks, including climate, ecology, land cover, biodiversity, socioeconomic indicators, public health, and wildfire forecasting.

Results

The authors evaluated four model variants (Adaptive and Multiscale, with $L=10$ and $L=40$ ) against five baselines (GeoCLIP, AlphaEarth, SatCLIP, GT-Loc, Copernicus-FM) across 24 tasks.

Socioeconomic and Public Health: OSMGraphCLIP demonstrates its strongest advantage in these domains. The MS-L40 variant achieved the best performance on 7 out of 12 CDC PLACES public health outcomes and outperformed or matched satellite-based baselines on median income and housing price predictions. The authors attribute this to OSM's explicit encoding of human activity patterns (road hierarchies, amenity categories) which are only indirectly captured by satellite pixels.
Environmental and Ecological Tasks: The model remains competitive with imagery-based methods on tasks like country classification, biome/ecoregion prediction, and land cover (reBEN), despite using no Earth observation data. However, it underperforms on tasks heavily dependent on vegetation structure and microclimate, such as bird species distribution (SatBird) and wildfire forecasting, where satellite data provides direct spectral signals.
Architecture Performance: Multiscale variants consistently outperformed adaptive variants, and higher spherical harmonic resolution ( $L=40$ ) generally yielded better results, highlighting the value of broader contextual aggregation.
Qualitative Analysis: PCA visualizations of the embeddings revealed coherent spatial organization, successfully recovering biome boundaries, urban gradients, and tropical-temperate distinctions solely from map topology. Cosine similarity analysis confirmed that the model groups geographically distant but semantically similar regions (e.g., equatorial forests or dense urban centers) effectively.

Significance and Claims

The paper claims that structured, collaboratively curated map data constitutes a powerful and viable supervisory modality for global location representation learning, particularly for tasks where the semantic organization of the built environment is the dominant predictive signal.

The authors position OSMGraphCLIP not as a replacement for Earth observation models, but as a complementary semantic-topological representation. They argue that while satellite imagery excels at capturing physical and spectral properties (vegetation, terrain), OSM graphs excel at capturing functional and relational properties (infrastructure, land use, human activity). The results suggest that for many socioeconomic and public health applications, explicit semantic annotation is more informative than spectral imagery. The work underscores that the geographic structure of OSM features carries substantial environmental signal, even in the absence of Earth observation data.

The authors conclude that the most impactful future direction involves tri-modal contrastive training to align OSM graphs, satellite imagery, and coordinates within a shared embedding space, and leveraging OSM's temporal versioning to learn representations of urban change.