CartoVec: General-Purpose Location Embeddings from Open Map Data

Furkan Gursoy

cs.LGOct 12, 2026

CartoVec: General-Purpose Location Embeddings from Open Map Data

Authors: Furkan Gursoy

Abstract

Geographic coordinates are common inputs in applied machine learning, but latitude and longitude describe position rather than the surrounding environment. Image-grounded location encoders learn this context from satellite or ground-level imagery; structured open map data offers another source of supervision whose utility for continuous coordinate encoders remains less explored. We introduce CartoVec, a coordinate encoder trained to predict Overture-derived land cover, land use, transportation, place and connectivity targets at multiple spatial scales, without image inputs or downstream-task labels. We evaluate frozen embeddings on Census outcomes, Airbnb prices and Philadelphia crime in the US Northeast using spatial block cross-validation. CartoVec has the highest mean performance on most Census outcomes and the highest average across the full suite, with several comparisons close across folds. The findings support open map data as a practical source of reusable location representations with a coordinate-to-vector interface that requires no map lookup at inference. The open-weight model is available at https://huggingface.co/gursoyfurkan/CartoVec.

Figures & tables

Figure 1: CartoVec architecture. The inference path converts latitude and longitude in degrees to a spherical unit vector, adds Fourier positional features, and produces a unit-normalized 64-dimensional embedding through a multilayer perceptron with residual blocks. The dashed region contains the training-only decoder and prediction heads. Map-derived targets and regularization train the encoder; the decoder and heads are discarded at inference.

Group	Overture theme	Type	Raw Columns	Aggregated Targets	Retained Targets
Land	Base	binary	57	57	44
Water	Base	binary	49	49	32
Land cover	Base	binary	11	11	7
Land use	Base	binary	135	135	116
Transportation segment type	Transportation	binary	52	52	52
Transportation connectors	Transportation	count	2	2	2

Table 1: Training-target heads. For the Places heads, “raw” is the number of per-category columns before category grouping, “aggregated” is the number after Basic-Category grouping within scale, and “retained” is the number satisfying the occurrence filter. Other heads are not aggregated.

Figure 2: Spatial distribution of training coordinates across the US Northeast. For display, coordinates are counted in raster cells approximately 0.9 km $\times$ 0.9 km; these cells summarize sampling density. The color scale is logarithmic with a mild gamma stretch. Brighter amber-to-white colors indicate denser sampling; dark gray marks areas outside the study mask. Dense urban cores, rural road corridors and the background grid reveal the adaptive, feature-driven sampling. The outlined windows locate the Boston and Syracuse views in Figure 3 ( 3(a) ) and ( 3(b) ), respectively.

Panel of Figure 3 — Figure 3: Sampled training coordinates at local scale. (a) Boston, its harbor and nearby islands: a 16.5 $\times$ 16.5 km window with 230,966 coordinates (853 per km 2 ). (b) Syracuse and Onondaga Lake: a 20.3 $\times$ 20.3 km window with 101,586 coordinates (246 per km 2 ). Road-aligned samples, building/place clusters and a background lattice show how map features guide sampling. Syracuse’s lower sampling density makes roads and the background grid more evident outside the city core; lake coverage is sparse. The panels contrast a metropolitan center with a less densely built area and have different geographic extents. Red circles identify landmarks for orientation.

Figure 4: Spatial block cross-validation for the Census tasks. Left: all 48,087 Census block groups in the evaluation region, colored by their assigned fold. Right: a closer view around Philadelphia, with the 0.05 ∘ tile grid in dark lines and block-group boundaries in white. Each fold is held out for testing in turn; the remaining tiles are divided between training and validation. Block groups whose representative coordinates fall in the same tile share its fold, although their polygon boundaries can cross tile edges.

Architecture	Hidden layers
linear	none (a linear map on the embedding)
small	$64\rightarrow 32$
base	$256\rightarrow 128$
wide	$1024\rightarrow 512$
geoclip-shape	2048 (single layer)

Table 2: Downstream probe architectures. Hidden-layer entries give successive widths; every probe has a linear regression output.

Method	Pop. density	Educ. attainment	Median age	Median household income	Median home value	Median gross rent	Airbnb price	Phila. crime	Macro- average
CartoVec-64	0.883	0.492	0.158	0.429	0.732	0.434	0.234	0.451	0.477
Raw lat/lon	0.441	0.095	0.035	0.063	0.488	0.327	-0.004	-0.104	0.168
RFF-256	0.631	0.188	0.057	0.152	0.610	0.390	0.219	0.343	0.324
CountEmb-725	0.836	0.318	0.142	0.292	0.429	0.256	0.209	0.594	0.384
Space2Vec FMoW	0.625	0.146	0.060	0.139	0.586	0.382	0.187	0.056	0.273
Space2Vec YFCC	0.686	0.347	0.095	0.219	0.657	0.413	0.205	0.143	0.346

Table 3: Downstream prediction. Mean test

R^{2}

over 5 spatial folds. Each method is reported at the probe architecture selected for it on validation (Section 4.3 ). Bold marks the highest mean in each column. The macro-average is the unweighted mean across the eight tasks.

R^{2}

for transformed targets is measured on the transformed scale (Section 4.2 ).

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Method	Probe	Pop. density	Educ. attain.	Median age	Household income	Home value	Gross rent	Airbnb price	Phila. crime	Macro mean	Macro median
CartoVec-64	Linear	0.793	0.178	0.104	0.249	0.509	0.334	0.206	0.413	0.348	0.345
	Small	0.864	0.456	0.158	0.388	0.687	0.417	0.233	0.445	0.456	0.454
	Base	0.880	0.492	0.149	0.429	0.724	0.434	0.234	0.452	0.474	0.475
	Wide	0.883	0.480	0.116	0.415	0.732	0.419	0.222	0.449	0.464	0.463
	2048	0.876	0.481	0.121	0.379	0.718	0.355	0.208	0.451	0.449	0.445
Raw lat/lon	Linear	-0.013	-0.012	-0.003	-0.122	-0.095	-0.128	-0.049	-0.038	-0.057	-0.053

Appendix

Table A1 : Complete probe comparison. Task entries are five-fold mean test

R^{2}

; underlining marks validation-selected architectures, not the highest test scores. Macro mean and macro median average the eight tasks’ fold means and fold medians, respectively. “2048” denotes the single-layer geoclip-shape probe (Table 2 ).

Geospatial Representation Learning

Geospatial representation learning builds reusable embeddings of places, imagery, maps, mobility, and Earth data for spatial prediction and reasoning.

See all 126 papers

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
Do Location Encoders Capture Spatial Effects? A GeoShapley Benchmark Across Scales
Jun 22, 2026Daniel Kiv, Shaowen WangGeospatial Representation Learning Shapley Value Attribution
Siebel School of Computing & Data Science, University of Illinois Urbana-Champaign Urbana, IL, USA · Department of Geography & GIS and Siebel School of Computing & Data Science, University of Illinois Urbana-Champaign Urbana, IL, USA
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, 2026J/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

Geospatial Representation Learning

Explore similar work

OSMGraphCLIP: Learning Global Location Representations from OpenStreetMap Graphs

Do Location Encoders Capture Spatial Effects? A GeoShapley Benchmark Across Scales

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