LEON: Location Embeddings from OSM Neighborhoods via Hexagonal Graph Masked Autoencoders

Szymon Soltysiak; Radoslaw Malek; Jedrzej Kusnierz; Milosz Chojecki; Piotr Szymanski; Aleksandra Kawala-Sterniuk

cs.LGOct 4, 2026

LEON: Location Embeddings from OSM Neighborhoods via Hexagonal Graph Masked Autoencoders

Authors: Szymon Soltysiak, Radoslaw Malek, Jedrzej Kusnierz, Milosz Chojecki, Piotr Szymanski, Aleksandra Kawala-Sterniuk

Organizations: Department of Artificial Intelligence, Wroclaw University of Science and Technology, Wroclaw, Poland

Abstract

Geographic information systems increasingly rely on sophisticated spatial representation learning techniques to extract meaningful patterns from complex geospatial data. This paper introduces LEON, a novel self-supervised framework that adapts Graph Masked Autoencoders (GraphMAE) for geospatial region representation learning. Our method leverages the inherent spatial structure of geographic data by constructing hexagonal grid graphs using H3 indexing and applying masked autoencoding techniques to learn robust spatial embeddings from OpenStreetMap (OSM) amenity distribution patterns. We evaluate LEON on multiple real-world datasets including EuroSAT satellite imagery classification and various geographic prediction tasks (housing prices, crime prediction, and urban analytics). Experimental results demonstrate that LEON achieves significant improvements in spatial understanding, with up to 1.87% accuracy improvement on EuroSAT classification and consistent performance gains across geographic prediction benchmarks. The learned embeddings exhibit highly structured and distinct properties, making them particularly suitable for downstream spatial analysis tasks. Our findings suggest that self-supervised learning provides an effective paradigm for geospatial region representation learning using widely available crowdsourced data.

Figures & tables

Metric / Class	LEON	Baseline	Impact
Global Summary
Accuracy	92.74%	90.87%	+1.87%
Test Loss	0.2278	0.2748	-0.0470
Weighted F1	0.9274	0.9063	+0.0211
Class-Specific F1
AnnualCrop	0.9248	0.8935	+0.0313

Table 1 : Comprehensive EuroSAT Classification Results: LEON vs. Baseline

Benchmark	Res.	Model	Method	$R^{2}$	MAE	RMSE	RMSLE	Features
Airbnb	9	DeepRegressor	LEON	0.3475	50.16	71.12	0.4696	146
		DeepRegressor	Baseline	0.3064	53.48	73.33	0.4980	18
		LinearRegression	LEON	0.2523	54.81	76.13	0.5329	146
		LinearRegression	Baseline	0.2140	57.88	78.06	0.5401	18
King County	9	DeepRegressor	LEON	0.7723	57,485	78,540	0.1619	218
		DeepRegressor	Baseline	0.7793	55,863	77,326	0.1573	90

Table 2 : Complete Geographic Prediction Benchmark Results. The Res. column indicates the H3 resolution used for graph construction.

Ablation Group	Setting	$R^{2}$ Score
Masking Ratio ( $\rho$ )	0.2	0.3951
	0.5 (Default)	0.3962
	0.7	0.3959
Decoder Type	MLP	0.3968
Decoder Type	GAT (Default)	0.3968

Table 3 : Hyperparameter Ablation Results. Model performance exhibits high stability across varied configurations.

Figure 2 : Eigenvalue spectrum analysis showing the isotropy characteristics of learned embeddings. The distribution reveals the effective dimensionality and structure of the representation space.

Temperature ( $t$ )	Uniformity Score
0.5	-0.6218
1.0	-1.1799
2.0	-2.1036
5.0	-3.7312
10.0	-4.8699

Table 4 : Uniformity Analysis across different temperature parameters (

t

). Higher temperatures probe local separability, while lower temperatures reflect global distribution.

Scale Pair	N	Mean	Std Dev	Min	Max
7-8	900	0.9695	0.0579	0.4940	1.0000
7-9	900	0.9666	0.0621	0.4786	1.0000
8-9	900	0.9955	0.0136	0.8766	1.0000
Random	899	0.5471	0.2774	-0.0594	1.0000

Table 5 : Scale Similarity Analysis

Figure 3 : Correlation between geographic distance and embedding similarity. The negative correlation ( $-0.116$ Pearson, $-0.134$ Spearman) demonstrates that spatially proximate locations have more similar representations, confirming spatial autocorrelation in learned embeddings.

Panel of Figure 4 — Figure 4 : Unsupervised Semantic Mapping. 3D PCA components of embeddings are visualized as RGB colors. The resulting patterns reveal spontaneously discovered urban structures, including transit corridors and industrial zones, where color similarity denotes semantic similarity.

Figure 5 : Spatial Projection of Universal Semantic Clusters. Embeddings for diverse urban morphologies are partitioned into $K=12$ functional zones. Mapping uses a shared cluster assignment to demonstrate semantic alignment across geographies.

Explore similar work

OSMGraphCLIP: Learning Global Location Representations from OpenStreetMap Graphs
Jun 6, 2026Dimitrios Michail, Eleni Saka, Ioannis Giannopoulos +1Graph Representation Learning Geospatial Foundation Models
Harokopio University of Athens, Athens, Greece. · National Technical University of Athens, Athens, Greece. · Vienna University of Technology, Vienna, Austria. +1
MIND the Gap: A Geographic Implicit Neural Representation with Adjustable Spatial Scale
Sep 21, 2026Isaac Corley, Arjun Rao, Esther Rolf +7Geospatial Representation Learning Hierarchical Representation Learning
Taylor Geospatial · University of British Columbia · Vector Institute +7
What's in an Earth Embedding? An Explainability Analysis of Location Encoders
Jun 23, 2026Livia Betti, Sebastian Ricke, Ivica Obadic +2Implicit Neural Representations Geospatial Representation Learning
University of Colorado Boulder · Technical University of Munich · Munich Center for Machine Learning

Jun 6, 2026, cs.AIJ/K move · Enter open · S save

OSMGraphCLIP: Learning Global Location Representations from OpenStreetMap Graphs

Dimitrios Michail, Eleni Saka, Ioannis Giannopoulos, Ioannis Papoutsis

Harokopio University of Athens, Athens, Greece. · National Technical University of Athens, Athens, Greece. · Vienna University of Technology, Vienna, Austria.+1

Graph Representation Learning Geospatial Foundation Models Geospatial Representation Learning Heterogeneous GNNs Geospatial Reasoning

Open paper AskPDF arXiv

We present OSMGraphCLIP, a CLIP-style geospatial representation model that learns global location embeddings from freely available OpenStreetMap (OSM) data. OSMGraphCLIP represents geographic environments as heterogeneous graphs of typed OSM features, preserving the topological and semantic relationships among roads, buildings, land-use regions, and points of interest. A multi-scale graph encoder captures both fine-grained local structure and broader landscape composition, and supervises a spherical-harmonics location encoder through a contrastive alignment objective. We evaluate OSMGraphCLIP across a diverse suite of downstream geospatial regression and classification tasks spanning climate, ecology, socioeconomic indicators, public health, land cover, biodiversity, and wildfire forecasting, and show that structured OSM data alone supports strong global location representations across domains. OSMGraphCLIP matches or exceeds satellite-based baselines on the majority of benchmarks, with the most pronounced advantage on socioeconomic and public-health tasks, where OSM's explicit semantic annotation of the built environment encodes patterns of human activity that satellite pixels can only capture indirectly. On ecological and environmental tasks, the model remains closely competitive with imagery-based methods despite using no Earth observation data. Qualitative analysis confirms that the learned embeddings organize geographic space coherently, recovering biome boundaries, urban gradients, and tropical--temperate distinctions from map topology alone.

Abstract

Explore similar work

OSMGraphCLIP: Learning Global Location Representations from OpenStreetMap Graphs

MIND the Gap: A Geographic Implicit Neural Representation with Adjustable Spatial Scale

What's in an Earth Embedding? An Explainability Analysis of Location Encoders