PaperlayerView exact revision on arXiv

Source-linked AI summary

PlacePulse: Learning Transferable Place Representations from Structural Geographic Data

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

arXiv:2610.13390v1cs.LG

TL;DR

Geospatial foundation models have largely represented places through behavioral signals or imagery, while transferable embeddings of their structural composition remain less explored. PlacePulse learns fine-resolution embeddings from structural attributes and evaluates both representation quality and downstream prediction. Its results position structural composition as a distinct, complementary modality, with performance relative to behavioral representations varying across domains and spatial supports.

  • Problem

    Geospatial foundation models have focused mainly on behavioral or visual signals, while structural composition has received comparatively little attention as a primary basis for transferable fine-resolution representations.

    Source: §1 · pp. 1–2

  • Method

    PlacePulse uses a self-supervised masked autoencoder to compress structural attributes and neighborhood aggregates into 256-dimensional embeddings for more than one million U.S. H3 resolution-7 cells, evaluated through intrinsic and downstream benchmarks.

    Source: §1 · p. 3, §7 · p. 30

  • Results

    PlacePulse outperforms a dimensionality-matched baseline on targets unrelated to its training inputs, while neither it nor PDFM consistently leads across domains and spatial supports.

    Source: §1 · p. 4

  • Takeaways & Limitations

    Structural composition alone can support a representation competitive with a dynamic-signal model, establishing structural embeddings as a distinct and complementary geospatial modality.

    Source: §7 · p. 30

    Evaluation is transductive on a fixed, fully observed grid; behavior on unseen cells, vintages, or geographies is unvalidated, and inductive generalization is not claimed.

    Source: §6.4 · p. 29

Abstract

from arXiv · show

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.

1 Introduction

Geographic analysis relies on labor-intensive, task-specific feature engineering, while transferable geospatial representations have focused mainly on behavioral or visual signals. PlacePulse learns embeddings from structural attributes and evaluates their intrinsic quality and downstream transfer, finding that structural representations add a distinct, complementary perspective on place.

1 Introduction

  • Task-specific geographic analyses require assembling and harmonizing multiple data sources, engineering variables, and selecting features, making feature construction labor-intensive.The paper situates this repeated work as a bottleneck as geospatial data grow in volume and diversity.Source: §1 · pp. 1–2
  • Structural attributes describing relatively stable demographic, socioeconomic, commercial, functional, and environmental composition have received less attention as primary signals for transferable representations than behavioral traces or imagery.The paper investigates structural information as an independent representation modality rather than only as covariates.Source: §1 · p. 2
  • The model learns 256-dimensional, unit-norm embeddings for more than one million U.S. H3 resolution-7 cells from 572 structural variables and neighborhood features.Its self-supervised masked-autoencoder objective supports national-scale nearest-neighbor search through a single matrix product.Source: §1 · p. 3, Figure 1 · p. 3
  • The benchmark separates intrinsic reconstruction and latent-space assessment from downstream transfer across targets and spatial supports, providing a common comparison basis independent of model architecture and input data.The evaluation also compares PlacePulse with PDFM.Source: §1 · pp. 3–4
  • PlacePulse outperforms a dimensionality-matched, target-blind baseline on targets unrelated to its training inputs, while comparison with PDFM shows neither representation consistently leads across domains and spatial supports.The findings suggest that structural and behavioral representations capture different aspects of place.Source: §1 · p. 4

2 Related work

The related work traces geographic representation learning from region embeddings to large-scale foundation models, highlighting behavioral and imagery signals as dominant sources. It positions PlacePulse as a national-scale investigation of structural composition as an independent, transferable modality.

2 Related work

  • The section reviews the development of geographic representation learning, self-supervised paradigms, and the place of structural representations in that literature.Source: §2 · p. 4

2.1 From region embeddings to geospatial foundation models

  • Early region embeddings learned dense vectors from heterogeneous urban data; later work expanded to multimodal and POI-based representations, while much of the literature remained city-scale.The review describes reusable representations as a consistent goal across this evolution.Source: §2.1 · pp. 4–5

2.2 Geospatial foundation models

  • Geospatial foundation models chiefly represent how places are used through behavioral signals or what they look like through imagery, leaving structural information comparatively underexplored as a primary signal.PDFM exemplifies behavioral modeling; imagery models include masked-image and contrastive approaches.Source: §2.2 · p. 5

2.3 Self-supervised learning for geospatial representations

  • Self-supervised geospatial methods reconstruct masked inputs, contrast views of locations, or predict spatial and temporal context; PlacePulse uses reconstruction on structural tabular data.Its pretraining learns transferable representations without downstream task labels.Source: §2.3 · p. 6

2.4 Structural representations of place

  • Prior structural approaches often use demographic or land-use variables as auxiliary inputs, focus on POIs, or combine demographics with mobility at city scale; PlacePulse represents broader composition without behavioral or imagery inputs.Its attributes span sociodemographic, economic, commercial, commuting, health, and environmental information.Source: §2.4 · pp. 6–7

2.5 Positioning our work

  • The paper investigates whether structural information alone supports transferable embeddings at fine national scale and compares them directly with PDFM, while leaving explicit modality fusion to future work.The comparison is framed as examining differences in information captured by structural and behavioral signals, not as replacing either modality.Source: §2.5 · p. 7

3 Data

PlacePulse is built from nationwide AGS structural data on H3 resolution-7 cells, adding neighborhood aggregates to cell attributes. Held-out targets from multiple sources and spatial supports are reserved for downstream evaluation.

3.1 Source and spatial frame

  • The 2025 AGS products provide current-year modeled estimates on a populated U.S. grid of 1,055,537 H3 resolution-7 cells, each about 5 km².The vendor-provided H3 support avoids an additional disaggregation or areal-interpolation step before training.Source: §3.1 · p. 7

3.2 Input features

  • The 572 training features are drawn from seven AGS products and include directly supplied and engineered attributes, plus an urbanity flag.The feature set is derived from 1,058 raw AGS variables; chronic-condition and health-risk variables are held out as evaluation targets.Source: §3.2 · p. 8

3.3 Spatial context features: H3 k-ring aggregates

  • For each base feature, PlacePulse adds immediate-ring mean and sample standard deviation plus second-ring mean, yielding 2,288 inputs that encode cell and neighborhood profiles.The focal cell is excluded from neighborhood summaries; the first and second rings contain at most 6 and 12 cells, respectively.Source: §3.3 · p. 9

3.4 Downstream evaluation targets

  • Downstream evaluation uses held-out targets from 12 external sources across H3 cells, ZCTAs, tracts, and voting precincts, including targets related to some AGS inputs.The Google PDFM benchmark is mixed-vintage and partially documented; the 2022 vintage is verified for CDC PLACES health indicators only.Source: §3.4 · pp. 9–11

4 Methods

PlacePulse uses a masked autoencoder to create fixed 256-dimensional representations from structural attributes and neighborhood aggregates. The methods assess both intrinsic representation properties and prediction of held-out targets, using spatially blocked evaluation across several supports.

4.1 Architecture

  • The dense encoder compresses the 2,288-dimensional input to a 256-dimensional bottleneck, with L2 normalization placing each embedding on the unit hypersphere.Spatial context enters through k-ring features rather than an explicitly spatial network architecture.Source: §4.1 · pp. 11, 13, Figure 3 · p. 12

4.2 Masked, spatially-weighted reconstruction loss

  • Training masks 30% of input positions and computes reconstruction loss only on masked positions, requiring the encoder to infer features from the remaining inputs.Center columns receive weight 2.0 and ring columns 0.667 to rebalance their loss contributions despite the 3:1 ring-to-center column ratio.Source: §4.2 · p. 13

4.3 Training

  • The encoder is trained with Adam and early stopping on a 90%/10% training-validation split, then applied to all 1,055,537 cells without a separate test set.The validation split also provides a held-out check on reconstruction quality.Source: §4.3 · pp. 13–14

4.4 Evaluation protocol

  • Evaluation pairs intrinsic checks of reconstruction and embedding structure with downstream prediction of held-out targets, since either axis alone may miss representation weaknesses.Downstream tasks use a fixed LightGBM configuration; state-blocked folds reduce leakage from spatial autocorrelation between nearby observations.Source: §4.4 · pp. 14–16
  • Intrinsic evaluation measures reconstruction fidelity, effective embedding dimensionality, and spatial coherence and autocorrelation across H3 neighborhoods.The protocol also assesses cluster structure and local hotspot patterns.Source: §4.4 · pp. 14–15
  • At native H3 resolution, PlacePulse is compared with a target-blind baseline that selects 256 representative input features once for all prediction tasks.The baseline uses agglomerative clustering and retains one medoid feature per cluster, avoiding target-label leakage.Source: §4.4 · p. 15
  • For ZCTA targets, H3 embeddings are area-weighted to ZCTAs before comparison with PDFM, which is evaluated at its native support.The paper notes that this comparison is conservative for PlacePulse; differing evaluation procedures also mean R² values are not directly comparable to PDFM's published values.Source: §4.4 · pp. 15–16
  • For other target resolutions, both models are area-weighted, but their transfers differ because their native supports and alignments with census-based units differ.PDFM transfers from coarser ZCTAs, while H3 aggregation can blend embeddings across neighboring cells at unit boundaries.Source: §4.4 · p. 16

5 Results

The internal evaluation finds that PlacePulse reconstructs structural inputs while making broad use of its embedding dimensions and preserving spatial organization across neighborhood scales.

5.1.1 Reconstruction fidelity

  • Mean reconstruction R2 is 0.853 (normalized MSE 0.147), with no feature below its mean predictor and residual error concentrated in feature-specific tails.Source: §5.1.1 · pp. 16–17, Figure 4 · p. 17

5.1.2 Embedding space health and dimensionality

  • Every embedding has unit norm, making cosine similarity equivalent to the dot product for similarity and nearest-neighbor operations.Source: §5.1.2 · p. 18
  • The decorrelation test retains 251 of 256 dimensions; median pairwise correlation is r = 0.000, with 99% of pairs below |r| = 0.47.Source: Figure 4 · p. 17, §5.1.2 · p. 18
  • The embedding raises participation ratio from 6.6 to 12.6 and retains 25.2 of 32.3 input nonlinear degrees of freedom (about 78%).Source: Figure 4 · p. 17, §5.1.2 · p. 18
  • All 256 dimensions exceed the 0.01 dead-dimension threshold, with standard deviations from 0.018 to 0.081.Source: Figure 4 · p. 17, §5.1.2 · p. 18

5.1.3 Spatial structure

  • H3-neighborhood cosine similarity is 0.980 versus 0.867 for permuted neighborhoods, and remains 0.853 after mean-centering; ZCTA similarity is 0.965 versus 0.885.Source: Figure 5 · p. 19, §5.1.3 · pp. 19–20
  • Hotspot statistics identify 55.6% of cells in significant clusters; typical k = 100 clusters draw 70.8% of cells from one k = 10 parent, versus an 18.5% chance baseline.Source: §5.1.3 · p. 20
  • PC1 has Moran’s I = 0.93 at k = 1, declining to 0.83 at k = 5; mean embedding-dimension Moran’s I falls from 0.70 to 0.50.Autocorrelation remains high beyond the model’s two-ring input radius.Source: §5.1.3 · p. 20, Figure 6 · p. 21

5.2 Downstream predictive utility

The downstream evaluation compares PlacePulse with a target-blind baseline at native H3 resolution and with PDFM at ZCTA and transferred target resolutions. Results vary by domain, with PlacePulse leading in several categories and PDFM in others.

5.2.1 Native resolution (H3 res-7)

  • Against the target-blind baseline, PlacePulse leads on Points of Interest (R2 = 0.772 versus 0.724) and Population Activity (0.801 versus 0.788), but trails health (0.707 versus 0.738); Human Activity Index R2 reaches 0.930.Source: §5.2.1 · pp. 22, 24, Figure 8 · p. 24

5.2.2 External baseline’s resolution (ZCTA)

  • PDFM leads Points of Interest (R2 = 0.906 versus 0.784), elevation (0.821 versus 0.555), and Entrepreneurship (0.822 versus 0.765), so the two model families encode complementary aspects of place.Source: §5.2.2 · pp. 25–26
  • At ZCTA support, PlacePulse averages higher than PDFM in Public Health (0.610 versus 0.573) and Economic status (0.633 versus 0.605), with Population Activity near parity (0.863 versus 0.855).Source: §5.2.2 · p. 24, Table 8 · p. 25, Figure 9 · p. 26

5.2.3 Other target resolutions (voting precinct, census tract)

  • PlacePulse leads PDFM in all three transferred categories, including presidential R2 = 0.737 versus 0.709 and social vulnerability R2 = 0.384 versus 0.269; both models fail on these indices.The indices share census socioeconomic indicators with AGS inputs, which may contribute to PlacePulse’s advantage.Source: Table 9 · p. 27, Figure 10 · p. 27, §5.2.3 · p. 27

6 Discussion

PlacePulse preserves most structural signal in a compact embedding and recovers broad spatial organization. Its downstream strengths vary by domain and spatial support, while data, transfer, and responsible-use limits bound the claims.

6.1 A compact, faithful re-encoding of structural place data

  • Average R2 is 0.853 (median 0.887) across 2,288 standardized inputs, and 251 of 256 embedding axes are mutually decorrelated.The 256-dimensional bottleneck preserves the bulk of the structural signal; no feature is reconstructed worse than its own mean.Source: §6.1 · p. 28
  • Global Moran’s I reaches 0.93 on the dominant principal component, and spatial autocorrelation persists beyond the model’s two-ring input radius.The analysis also finds significant clusters for 55.6% of cells.Source: §6.1 · p. 28
  • K-means partitions reach 94–98% of maximum size entropy, order locations along settlement density, and distinguish suburban types by income at comparable densities.As k grows, partitions refine rather than reorganize.Source: §6.1 · p. 28

6.2 Structural and dynamic representations encode complementary aspects of place

  • At ZCTA resolution, PlacePulse leads PDFM on Public Health (0.610 versus 0.573) and Economic status (0.633 versus 0.605), but trails on Points of Interest, Physical Geography, and Entrepreneurship.The comparison also reports parity on Population Activity and Social capital; relative strengths vary by domain.Source: §6.2 · p. 28
  • PlacePulse leads PDFM on 2024 presidential vote share (0.737 versus 0.709), intergenerational economic mobility (0.546 versus 0.508), and social-vulnerability and resilience scores (0.435 versus 0.302).These transfers differ in native spatial support, so the comparison reflects each representation together with its support.Source: §6.2 · pp. 28–29

6.3 Implications for practice and responsible use

  • A reusable 256-dimensional vector per H3 cell replaces per-task assembly of hundreds of variables, while unit-norm embeddings enable national-scale similarity retrieval as a matrix operation.Neighborhood coherence supports area-weighted aggregation for ZCTA-level analyses.Source: §6.3 · p. 29
  • Because embeddings reconstruct sociodemographic inputs, they encode and can proxy attributes such as race, ethnicity, and income even when those attributes are not explicit downstream.The paper intends spatial analysis, not individual-level eligibility decisions in regulated domains, and assigns deployers responsibility for applicable law and required disparate-impact testing.Source: §6.3 · p. 29

6.4 Limitations

  • Proprietary, modeled 2025 AGS estimates propagate unbounded input uncertainty and limit reproducibility; evaluation is transductive and unvalidated on unseen cells, vintages, or geographies.Target-input overlap may partly explain cross-model differences, and a controlled comparison excluding overlapping inputs remains future work.Source: §6.4 · pp. 29–30

7 Conclusions

PlacePulse compresses structural attributes into national-scale embeddings that perform competitively across downstream tasks, supporting structural composition as a complementary geospatial representation modality. The conclusion outlines future work on combining modalities, expanding temporal and geographic coverage, and studying embedding geometry and aggregation.

7 Conclusions

  • PlacePulse encodes 572 attributes, expanded to 2,288 inputs, as 256-dimensional embeddings for 1,055,537 U.S. H3 resolution-7 cells and performs competitively across 74 regression targets.It outperforms a dimensionality-matched feature-selection baseline on targets unrelated to training inputs and matches or exceeds PDFM on several outcome domains after transfer to non-native units.Source: §7 · p. 30
  • Structural composition alone can yield representations competitive with a dynamic-signal model, supporting it as a distinct, complementary modality rather than merely a substitute for behavioral or imagery data.The authors characterize the contribution as evidential rather than architectural.Source: §7 · p. 30
  • Future work should test multimodal fusion, input-family ablations, temporal representations, and portability to other countries.These directions could clarify domain contributions and input overlap, support longitudinal comparisons, and test transfer across national data regimes.Source: §7 · p. 30
  • PlacePulse uses nearly all 256 dimensions, but variance remains concentrated and embeddings occupy a narrow region of the unit sphere.Alternative encoders that omit masked inputs directly and uniformity regularizers are proposed for study; greater uniformity could either improve retrieval and prediction or remove meaningful variation.Source: §7 · pp. 30–31
  • The unit-norm geometry motivates retrieval applications such as twin-area discovery and market-whitespace analysis, alongside evaluation of embedding aggregation to policy-relevant units.The authors identify quantifying lost fine-scale structure during aggregation as a further research need.Source: §7 · p. 31

Declaration of competing interest

The authors disclose commercial interests involving both PlacePulse and its input data.

Declaration of competing interest

  • All authors are CARTO employees, and CARTO developed PlacePulse and has a commercial interest in distributing it as a data product.CARTO also funded the work and participated in study design, analysis, writing, and the submission decision.Source: Declaration of competing interest · p. 31
  • AGS licenses the input features and has a commercial partnership with CARTO for distributing its data products.AGS reviewed the description of its own data products but had no role in study design, evaluation, analysis, interpretation, or the submission decision.Source: Declaration of competing interest · p. 31
  • The authors identify these relationships as potential competing interests.Source: Declaration of competing interest · p. 31

Funding

CARTO funded the work; EuroHPC provided computing access, and no specific external-agency grant supported the research.

Funding

  • CARTO funded the work through its internal research programme.Source: Funding · p. 31
  • EuroHPC awarded two projects access to MareNostrum 5 through its AI Factory Playground Access mode.The facility is hosted by the Barcelona Supercomputing Center in Spain.Source: Funding · p. 31
  • Apart from the computing allocation, the work received no specific grant from external funding agencies.CARTO participated in study design, analysis, interpretation, writing, and the submission decision; EuroHPC had no role in these.Source: Funding · p. 32

Data availability

Access to the inputs, trained embeddings, and implementation code is restricted, while most evaluation targets remain available from their original providers.

Data availability

  • AGS input features require a provider license and cannot be redistributed by the authors.Source: Data availability · p. 32
  • The PlacePulse embeddings are a CARTO data product and are not publicly redistributable, and the model code is not publicly available.Source: Data availability · p. 32
  • Downstream evaluation targets are available from their original providers under their terms, except for the six AGS Health targets.PDFM embeddings are distributed by Google Research under separate access terms.Source: Data availability · p. 32

Author Contributions (CRediT)

The contribution statement assigns conceptualization and supervision across several authors, with García-Duarte additionally credited for data curation, analysis, methodology, software, validation, visualization, and writing.

Author Contributions (CRediT)

  • García-Duarte is credited with conceptualization, data curation, formal analysis, investigation, methodology, software, validation, visualization, and writing.She is credited with both original-draft and review-and-editing work.Source: Author Contributions (CRediT) · p. 32
  • Álvarez is credited with conceptualization, funding acquisition, project administration, supervision, and writing.He is credited with original-draft and review-and-editing work.Source: Author Contributions (CRediT) · p. 32
  • Vicente is credited with conceptualization, data curation, formal analysis, and review-and-editing work.Source: Author Contributions (CRediT) · p. 32
  • Pérez Trufero and de la Torre are credited with conceptualization and supervision; de la Torre also has funding acquisition.Source: Author Contributions (CRediT) · p. 32

A Supplemental clustering structure evaluation

Across three increases in K-means granularity, finer partitions largely preserve coarse-cluster structure while subdividing it, despite K-means imposing no hierarchical constraint.

A Supplemental clustering structure evaluation

  • Nesting matrices are strongly block-diagonal across all three transitions, with most fine clusters drawing most cells from a single coarse parent and limited mixing across parents.Together, the results indicate that increasing k refines rather than arbitrarily reorganizes the partitions, even though K-means imposes no hierarchical constraint.Source: A Supplemental clustering structure… · p. 33
  • Median fine-cluster purity is 68.4% for k = 10 →50, 70.8% for k = 10 →100, and 54.6% for k = 50 →100, with medians 3.7×, 3.8×, and 14.7× above their respective chance baselines.Purity is the largest share of a fine cluster’s cells coming from one coarse cluster; compare each median with its own baseline, since the transitions have different relative granularities.Source: A Supplemental clustering structure… · p. 33
  • Pearson correlations between coarse-cluster size and fine-cluster descendants are 0.85 for k = 10 →50, 0.88 for k = 10 →100, and 0.49 for k = 50 →100.Larger coarse clusters generally receive more subdivisions, while the weaker 50 →100 correlation suggests splitting is less determined by size alone.Source: A Supplemental clustering structure… · p. 33

B Supplemental downstream predictive evaluation

The appendix reports per-target downstream results across native and aggregated spatial supports, with evaluation details intended to show which targets drive category-level performance.

B Supplemental downstream predictive evaluation

  • The appendix gives individual-target results for 20 targets at native H3 resolution 7, 50 at ZCTA resolution, and 4 at voting-precinct and census-tract resolution.It is presented as a way to see which targets drive each category mean.Source: B Supplemental downstream predictive… · p. 33
  • Per-fold R2 is reported as mean ± standard deviation under five-fold spatially blocked cross-validation, with each state held out entirely in one fold.Both representations are scored on the same GEOIDs, the intersection of the units each covers.Source: B Supplemental downstream predictive… · p. 33, Table B.1 · p. 34
  • PlacePulse is compared with a target-blind, dimensionality-matched baseline at H3 resolution 7, while ZCTA and non-native-support tables compare it with PDFM.Tables B.1–B.3 cover H3 resolution 7, ZCTAs, and voting precincts or census tracts, respectively; their captions do not provide the individual outcome values.Source: Table B.1 · p. 34, Table B.2 · p. 35, Table B.3 · p. 38

Related papers

Loading 2610.13390v1…