No Trajectories, No Satellite Imagery: 1.05 Million US Grid Cells Compressed to 256 Structural Dimensions

Geospatial machine learning has largely described a place through behavioral traces or satellite imagery, while comparatively stable structural attributes such as demographics, businesses, commuting, health and environmental risk have usually entered only as auxiliary variables. Existing geospatial foundation models fall into two families: one learns from dynamic behavioral signals such as mobility, visitation and search activity, the other from satellite or street-level imagery; structural attributes have rarely served as the sole pretraining signal. The model uses no behavioral traces and no imagery, only 572 structural attributes plus their neighborhood aggregates, covering 1,055,537 H3 resolution-7 cells in the United States.

The vectors retain most of the structural information when reconstructing their inputs, and they vary smoothly across geography, so neighboring cells carry highly similar vectors. No comparable public embedding product is natively available at H3 resolution 7, so the native-resolution comparison is limited to a 256-feature, target-blind selection drawn from the original inputs. Reconstruction of the standardized inputs reaches a mean R² of 0.853, 251 of the 256 dimensions are approximately decorrelated, and the mean cosine similarity between a cell and its neighborhood mean is 0.980 against 0.867 for a permuted-neighborhood baseline.

Across 74 external targets the vectors serve directly as features, spanning four spatial supports: H3 resolution 7, ZCTA, census tract and voting precinct. At ZCTA resolution the comparison is Google's Population Dynamics Foundation Model, published natively at that support, while PlacePulse must first be aggregated up from H3 cells by an area-weighted mean. At ZCTA, public health is 0.610 versus 0.573 and economic status 0.633 versus 0.605, while points of interest is 0.784 versus 0.906 and physical geography 0.628 versus 0.744; at precinct and tract level, elections is 0.737 versus 0.709, economic mobility 0.546 versus 0.508, and social vulnerability and resilience 0.435 versus 0.302.

The work targets researchers and engineers doing spatial analysis within the United States: inputs are structural attributes on H3 resolution-7 cells, and the output is one 256-dimensional unit-norm vector per cell that can be attached to downstream regression tasks. Because the vectors lie on the unit sphere, cosine similarity reduces to a dot product, so nearest-neighbor retrieval across the national grid becomes a single matrix product. The authors propose testing concatenations of structural and dynamic or imagery representations directly, and extending evaluation to coarser units such as counties.

The inputs are proprietary, modeled, single-vintage 2025 estimates rather than direct observations, so their own estimation error propagates into the embeddings and into every downstream result, and their license limits full reproducibility. Coverage is limited to the United States on a single national grid, leaving transferability to other countries and data regimes untested. The evaluation is transductive: embeddings are produced for a fixed, fully observed grid and evaluated as features over that grid, and the authors make no claim of inductive generalization to unseen cells, vintages or geographies. Results come from a single configuration selected from an exploratory sweep, so sensitivity to architecture and hyperparameters is only partially characterized. Some downstream targets share covariates with the inputs, so cross-model differences may partly reflect input-target overlap rather than representation quality alone.