PlacePulse: Learning Transferable Place Representations from Structural Geographic Data

Lucía García-Duarte; Miguel Álvarez; Diego Vicente; Javier Pérez Trufero; Javier de la Torre

cs.LGOct 9, 2026

PlacePulse: Learning Transferable Place Representations from Structural Geographic Data

Authors: Lucía García-Duarte, Miguel Álvarez, Diego Vicente, Javier Pérez Trufero, Javier de la Torre

Organizations: (CARTO)

Abstract

Geospatial machine learning often depends on task-specific feature engineering over large collections of sociodemographic, business, and environmental variables. Geospatial foundation models offer reusable representations of place, but leading approaches typically encode dynamic signals, such as mobility and search activity, or visual signals from satellite imagery. Few models represent the static structural composition of places (who lives and works there, which businesses operate there, and which risks they face) as transferable embeddings at fine spatial resolution. We introduce PlacePulse, a structural geospatial foundation model that compresses 572 attributes, expanded to 2,288 inputs through H3 neighborhood aggregates, into 256-dimensional unit-norm embeddings for 1,055,537 H3 resolution-7 cells in the United States. We also introduce a reusable two-axis benchmark that separates intrinsic representation quality from downstream predictive utility across multiple targets and spatial supports. Intrinsically, the embeddings reconstruct standardized inputs with mean $R^2 = 0.853$ , use 251 of 256 approximately decorrelated dimensions, and show strong spatial organization, with Moran's $I = 0.93$ on the leading principal component and autocorrelation persisting beyond the input radius. Extrinsically, across 74 held-out targets from 12 external sources and four spatial supports, PlacePulse outperforms a dimensionality-matched feature-selection baseline on targets unrelated to the training inputs and is competitive with Google's Population Dynamics Foundation Model, leading on health, economic, election, economic-mobility and social-vulnerability outcomes, with relative strengths that vary by domain. These findings position structural composition as a distinct and complementary modality for geospatial foundation models, not simply a fallback where behavioral or imagery data are unavailable.

Figures & tables

Figure 1: Overview of the geospatial foundation model: how the embeddings are extracted and applied to downstream tasks. A self-supervised autoencoder compresses a broad set of structural attributes per location (sociodemographic, economic, health, climate, mobility, and others) into a single 256-dimensional, L2-normalized embedding at H3 resolution 7. This representation drives two families of applications: (1) similarity-based retrieval for tasks such as white-space analysis, twin-area and look-alike discovery, cannibalization and site screening, or market segmentation; and (2) supervised prediction, where the embeddings serve as off-the-shelf features for regression and classification at the native H3 resolution or aggregated to the target’s spatial unit.

Figure 2: Visualizations of the presented embeddings across the contiguous United States (H3 resolution 7, ${\sim}$ 1.06M cells), colored by principal component analysis (PCA) of the 256-dimensional embeddings. (a) The three leading components (PC1–PC3) are min–max scaled to the red, green, and blue channels of an image; cells with similar colors have similar embeddings. (b) A bivariate map of the two leading components (PC1 and PC2), each split into terciles and combined into a $3\times 3$ color scheme, from low–low to high–high.

Figure 3: Overview of the methodology. (a) A masked autoencoder compresses 2,288 standardized features per H3 res-7 cell (572 own + 1,716 k-ring aggregates) into a 256-dimensional embedding on the unit sphere, reconstructing only the 30% of positions masked at each step, with center columns upweighted to offset the 3:1 ring majority. (b) The frozen encoder is applied to all 1,055,537 cells, and the embeddings are evaluated both intrinsically and as downstream predictors at three spatial supports (see Section 4.4 ).

Component	Specification
Input / output	,288 features (572 own + 1,716 ring aggregates)
Encoder	Dense [1024, 512, 384] $\rightarrow$ Dense 256 (linear) $\rightarrow$ L2 (unit) normalization
Decoder	Dense [384, 512, 1024] $\rightarrow$ Dense 2,288 (linear)
Embedding	-d, constrained to the unit sphere ( $\lVert e\rVert=1$ )
Per-layer block	Dense (He-normal init) $\rightarrow$ LeakyReLU ( $\alpha=0.1$ ) $\rightarrow$ BatchNorm $\rightarrow$ Dropout (0.1)

Table 1: Model configuration summary.

Setting	Value
Optimizer	Adam, learning rate $3\times 10^{-4}$ , gradient clipping (clipnorm $=1.0$ )
Batch size
Max epochs
Early stopping	on validation loss (patience 20, min- $\Delta$ $1\times 10^{-4}$ )
LR schedule	ReduceLROnPlateau (patience 10)
Train / validation	% / 10%, random seed 42

Table 2: Model training summary.

Figure 4: Internal quality diagnostics of the 256-dimensional embeddings, computed over all 1,055,537 H3 res-7 cells. (a) Distribution of per-feature reconstruction

R^{2}

across the 2,288 input features; dashed line, median. (b) Distribution of Pearson correlation coefficients across all 32,640 off-diagonal embedding-dimension pairs; dashed line, median. (c) Cumulative PCA explained variance against number of components; dashed lines mark the 80%, 90% and 99% levels. (d) Distributions across the 256 embedding dimensions of six per-dimension summary statistics, each computed over all cells: mean, standard deviation, minimum, maximum, skewness, and excess kurtosis. Red dashed line, the 0.01 dead-dimension threshold ( Section 4.4 ); orange lines,

|\text{skewness}|=2

Measure	Input (2,288-d)	Embedding (256-d)
PCA participation ratio (evenness of variance spread)	6.6	12.6
Levina–Bickel MLE (nonlinear degrees of freedom)	32.3	25.2
Variance in top-3 PCA components (coverage)	47.8%	35.7%

Table 3: Variance structure of the input vs. the embedding.

Figure 5: Neighborhood coherence of the embeddings as the density of the per-cell cosine similarity between a cell’s embedding and the mean embedding of its neighborhood. Rows are the two framings: (a, b) the PlacePulse embeddings, (c, d) the same embeddings minus the global mean vector. Columns are the two neighborhood definitions: (a, c) the cumulative H3 k-ring disk for

k=1

–5, focal cell excluded, over all 1,055,537 cells; (b, d) the cell’s own ZCTA area-weighted mean, focal cell included, over the 1,001,075 cells assigned to one of 30,283 ZCTAs. Solid step curves are own-neighborhood similarities. Dotted curves are the permuted baseline, in which the same cell is compared with a randomly reassigned neighborhood of the same type. In (a, c) color encodes the ring radius, purple to green for

k=1

to 5, and all five permuted baselines are orange. Dashed vertical lines mark the means.

Figure 6: Spatial autocorrelation of the embeddings across all 1,055,537 H3 resolution-7 cells. (a) Moran’s $I$ for the 256 original embedding dimensions, using row-standardized H3 ring-neighborhood weights for $k=1$ –5. Boxes show the median and interquartile range across dimensions, with individual dimensions shown as jittered points. (b) The same analysis for the 256 principal components; points are colored by $\log_{10}$ variance explained, and the star marks PC1. In both panels, the shaded band marks $k\leq 2$ , the radius of the encoder’s ring-aggregate inputs, and the horizontal line marks $E[I]\approx 0$ . (c) Combined Getis–Ord $G_{i}^{*}$ , as defined in Section 4.4 , over uniform-weight H3 ( $k=3$ ) ring neighborhoods. Values are binned into 11 classes, with teal negative and red positive. The inset gives class edges, cell shares, and dashed $\pm 1.96$ reference lines for a two-sided $p=0.05$ threshold. Because embedding orientation is arbitrary, the sign is not directional.

$k$	Largest cluster	Smallest cluster	Entropy $H$	$H$ / max	Effective clusters	Inertia
10	195,375 (18.5%)	36,385 (3.45%)	3.14 (max 3.32)	94.5%	8.8	102,239
50	39,340 (3.7%)	4,939 (0.47%)	5.50 (max 5.64)	97.5%	45.4	89,946
100	22,963 (2.2%)	696 (0.07%)	6.48 (max 6.64)	97.5%	89.1	84,510

Table 4: Mini-batch K-means cluster structure over the 256-dimensional embeddings.

H

is the Shannon entropy of the cluster-size shares, its maximum

\log_{2}k

attained at equal sizes;

2^{H}

is the equivalent number of equally-sized clusters; inertia is the within-cluster sum of squares.

Figure 7: K-means cluster structure over all H3 res-7 cells. Color identifies the cluster and is shared across all plots: clusters close in embedding space take close colors, and at $k=50$ and 100 each cluster takes the hue of the $k=10$ cluster holding most of its cells and varies only in lightness. (a) Partitions at $k=10$ (top), 50 (middle) and 100 (bottom). Left to right: the conterminous United States, a New York metro close-up, a Dallas–Fort Worth close-up, and the cluster sizes sorted largest to smallest on a log axis, annotated with the largest and smallest share. (b) Cell flows from the coarse to the fine partition, for $k=10\rightarrow 50$ , $10\rightarrow 100$ and $50\rightarrow 100$ (left to right). Ribbon width is proportional to the number of cells. Only the largest ribbons carrying 95% of cells are drawn, and each diagram states how many flows it omits.

Category	What it covers	#	PlacePulse $R^{2}$	Baseline $R^{2}$
Public Health	Adult health-risk prevalence: diabetes, obesity, high blood pressure, cancer, high cholesterol, stroke.	6	0.707	0.738
Points of Interest (log space)	POI counts by type: total plus retail, food & drink, healthcare, offices, finance, leisure, education, public services, tourism, transport.	11	0.772	0.724
Population Activity	Human-footprint proxies: human activity index, nighttime lights (raw VIIRS radiance), cell-tower density.	3	0.801	0.788

Table 5: Downstream performance at native H3 res-7 resolution (PlacePulse vs. the target-blind, dimensionality-matched baseline). The average per-fold

R^{2}

of the tasks in each category is reported; see Table B for the uncertainty of each individual task.

Figure 8: Downstream predictive performance at native H3 res-7 resolution. PlacePulse (256-d) vs. the target-blind, dimensionality-matched baseline, per target within each of the three native-resolution categories, under 5-fold spatially-blocked cross-validation. Each row is one target: the two dots are that target’s per-fold mean $R^{2}$ under the two representations, the connector takes the color of the leading one, and the shaded band spans $\pm 1$ standard deviation across the 5 held-out folds.

Category	What it covers	#	PlacePulse $R^{2}$	PDFM $R^{2}$
Public Health	Health indicators: chronic-disease prevalence, health behaviors, self-rated physical and mental health, preventive-care access.	21	0.610	0.573
Economic status	Economic standing and hardship: median family and household income, median home value, poverty rate, energy burden and cost, low-food-access population share.	7	0.633	0.605
Population Activity (mixed scales)	Population, population density, nighttime lights; human activity index and cell-tower density.	5	0.863	0.855
Physical Geography	Natural land signals: mean elevation and tree-canopy share.	2	0.628	0.744
Points of Interest (log space)	POI counts by type, the same 11 classes as at H3.	11	0.784	0.906
Social capital	Cross-status economic connectedness and friendship clustering.	2	0.715	0.713

Table 6: Downstream performance at ZCTA resolution (PlacePulse vs. PDFM). The average per-fold

R^{2}

of the tasks in each category is reported; see Table B for the uncertainty of each individual task.

Figure 9: Downstream predictive performance at ZCTA resolution. PlacePulse embeddings, area-weighted from H3 resolution 7 to ZCTAs, are compared with Google PDFM embeddings at native ZCTA resolution. The 50 targets are grouped into the seven ZCTA-level categories and split across two columns, under 5-fold spatially blocked cross-validation. Each row is one target: the two dots are that target’s per-fold mean $R^{2}$ under the two representations, the connector takes the color of the leading one, and the shaded band spans $\pm 1$ standard deviation across the 5 held-out folds.

Category	What it covers	#	PlacePulse $R^{2}$	PDFM $R^{2}$
Elections (precinct)	U.S. presidential result by voting precinct.	1	0.737	0.709
Economic Mobility (tract)	Opportunity Atlas intergenerational mobility for children raised at the 25th income percentile.	1	0.546	0.508
Social Vulnerability and Resilience (tract)	Social-vulnerability and community-resilience scores.	2	0.435	0.302

Table 7: Downstream performance at non-native target resolutions: voting precincts and census tracts. Both representations are area-weighted from their original resolution (H3 res-7 for PlacePulse, ZCTA for PDFM) to the target resolution. The average per-fold

R^{2}

of the tasks in each category is reported; see Table B for the uncertainty of each individual task.

Figure 10: Downstream predictive performance at non-native target resolutions. PlacePulse and Google PDFM are both area-weighted from their native spatial supports to voting precincts or census tracts, under 5-fold spatially blocked cross-validation. Each row is one target: the two dots are that target’s per-fold mean $R^{2}$ under the two representations, the connector takes the color of the leading one, and the shaded band spans $\pm 1$ standard deviation across the 5 held-out folds.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Figure A.1: Finer K-means partitions mostly refine coarser partitions rather than reorganizing them. We compare three transitions in clustering granularity: $k=10\rightarrow 50$ , $k=10\rightarrow 100$ , and $k=50\rightarrow 100$ . (a) Nesting matrices. Rows represent fine clusters and columns represent coarse clusters, ordered by coarse-cluster size. Shading indicates the share of each fine cluster’s cells that fall within each coarse cluster. Rows are sorted by their dominant coarse “parent”, and dashed lines separate the resulting parent blocks. (b) Empirical cumulative distribution of fine-cluster purity, defined as the largest share of a fine cluster’s cells coming from a single coarse cluster. Dashed vertical lines mark the median of each transition, and dotted vertical lines mark the chance baseline, the cell share of the largest cluster in the source partition. (c) Number of fine-cluster descendants against coarse-cluster size, on a logarithmic x-axis. Marker shape and color identify the transition, and the legend gives the Pearson correlation $r$ and the number of coarse clusters $n$ .

Geospatial Reasoning Benchmarks

See all 31 papers

May 3, 2026

Geospatial foundation-model embeddings improve population estimation unevenly across space and scale

Reliable subnational population estimates are essential for applications, yet remain difficult where censuses are sparse, outdated or spatially coarse. Existing population-mapping workflows rely on hand-built geospatial covariates, such as settlement extent, night-time lights, and environmental conditions, which must be assembled and harmonised across scales and geographies. Geospatial foundation models offer an alternative by learning reusable representations of place from more multifaceted and heterogeneous data sources. Here, we benchmark Population Dynamics Foundation Model (PDFM) embeddings against the harmonised geospatial covariates for subnational population estimation in Brazil, Nigeria and the United States. Under geographically structured validation, PDFM increased predictive fit by a median of 20.1% (IQR: 10.0-33.2%, across country-model comparisons) reduction in unexplained variance, and reduced Kullback-Leibler divergence by 23.2% (9.2-26.2%). However, these gains were uneven. PDFM was most advantageous where the geospatial covariates weakly characterised settlement context, such as larger and less-developed subnational areas. Moreover, PDFM performance was scale-coupled with embeddings providing less flexible transfer across spatial aggregations than geospatial covariates. These findings showed that geospatial foundation-model representations of place can improve population estimation in data poor settings, but their benefits break down predictably under spatial scale mismatch, revealing a fundamental limitation of current geospatial AI.

Wenbin Zhang, Eimear Cleary, Francisco Rowe +4

WorldPop, School of Geography and Environmental Sciences, University of Southampton, United Kingdom · Geographic Data Science Lab, Department of Geography and Planning, School of Environmental Sciences, University of Liverpool, United Kingdom

Geospatial Foundation Models Geospatial Representation Learning Spatial Generalization

Ask PDF

Sep 24, 2026

Recoverable Geographic Location Information in Earth-Observation Embeddings

Earth-observation (EO) foundation models provide reusable embeddings, yet downstream task accuracy does not reveal whether these representations encode geographic information, which may be beneficial for location-aware applications but potentially detrimental when representations invariant to geographic location are desired. We therefore evaluate the geographic coordinate robustness of Tessera v1, Tessera v1.1, and AlphaEarth by testing whether coordinates can be predicted from the embedding representations using 284 quality-verified European solar farms from 2024. We assessed geographic information content information through the association between cosine and geodesic distances and through prediction of projected coordinates in EPSG:3035. Embeddings from all three EO foundation models contain recoverable geographic information. All prediction models significantly outperform training-range uniform random sampling baselines, with AlphaEarth exhibiting the strongest distance association and lowest mean geodesic error. Both Tessera variants also yielded higher geographic distance correlations than the Sentinel-2 controls. These findings motivate geographic information content as an additional criterion for auditing EO foundation models.

Peiwen Zhang, Kristie Hu, Jovana Knezevic +2

University of Waterloo · University of Cambridge · Aalto University

Geospatial Foundation Models Geospatial Representation Learning Remote Sensing

Ask PDF

Mar 2, 2026

From Pixels to Patches: Pooling Strategies for Earth Embeddings

Geospatial foundation models increasingly expose pixel-level embedding products that can be downloaded and reused without access to the underlying encoder. In this setting, downstream tasks with patch- or region-level labels require a post-hoc aggregation step that maps dense pixel embeddings to a single representation. The default choice, mean pooling, discards within-patch variability and can underperform under spatial distribution shift. To study this setting, we introduce EuroSAT-Embed: 81,000 embedding GeoTIFFs derived from three foundation models: AlphaEarth, OlmoEarth, and Tessera. Using these fixed embedding products, we benchmark 11 training-free pooling methods and 2 train-set-fitted baselines under both random and geographically disjoint test splits. Richer pooling schemes reduce the geographic generalization gap by over 50% relative to mean pooling and improve accuracy by up to 6% on spatial splits. We recommend a three-tier strategy: (1) mean as a baseline, (2) stats pooling (min/max/mean/std) as the default at 4x the embedding dimension, and (3) covariance pooling for peak accuracy. Across all three embedding products, simple distributional statistics improve spatial-split performance over mean pooling.

Isaac Corley, Caleb Robinson, Inbal Becker-Reshef +1

Wherobots · Microsoft AI for Good Research Lab

Spatial Generalization Geospatial Foundation Models Representation Learning Remote Sensing Benchmarks

Ask PDF

Abstract

Geospatial Reasoning Benchmarks

Explore similar work

Geospatial foundation-model embeddings improve population estimation unevenly across space and scale

Recoverable Geographic Location Information in Earth-Observation Embeddings

From Pixels to Patches: Pooling Strategies for Earth Embeddings