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

Technical Summary: Geospatial Foundation-Model Embeddings for Population Estimation

Problem Statement
Reliable subnational population estimates are critical for infrastructure planning, service provision, and disaster risk reduction. However, these estimates are often difficult to generate in regions where census data are sparse, outdated, or spatially coarse. Traditional population-mapping workflows rely on "hand-built" geospatial covariates (e.g., settlement extent, night-time lights, environmental conditions) that must be manually assembled and harmonized across different scales and geographies. These approaches are limited by the need to engineer fixed sets of interpretable predictors, which may fail to capture the broader spatial context and complex human-activity signals that distinguish places. While geospatial artificial intelligence (GeoAI) offers learned representations of place, it remains unclear whether, how, and to what extent these foundation-model embeddings improve population modeling relative to established geospatial covariates, and in which specific settings they offer the greatest benefit.

Methodology
The study benchmarks the Population Dynamics Foundation Model (PDFM) embeddings against a harmonized set of 23 standard geospatial covariates for subnational population estimation. The analysis covers three countries with diverse data environments: Brazil, Nigeria, and the United States.

• Data Sources:

  • PDFM Embeddings: 330-dimensional feature vectors learned via a self-supervised graph neural network from multimodal data, including aggregated search trends, maps/busyness indicators, weather, and air-quality signals. These were generated for specific administrative units (districts in Brazil, LGAs in Nigeria, ZCTAs in the US) using data time-matched to October 2023.
  • Geospatial Covariates: A standardized set of 23 raster-based variables representing built environment, accessibility, land cover, topography, and night-time lights, aggregated to match the spatial resolution of the PDFM units.
  • Population Data: Administrative unit population shares derived from official sources (IBGE for Brazil, National Population Commission for Nigeria, and ACS for the US).

• Experimental Design:

  • Benchmarking: Three regression models (Random Forest, XGBoost, Elastic Net) were trained separately using either PDFM embeddings or geospatial covariates as predictors. Performance was evaluated using the coefficient of determination ($R^2$) and Kullback-Leibler (KL) divergence.
  • Validation Strategy: A geographically structured spatial cross-validation was employed (e.g., leave-one-state-out) to assess transferability and prevent information leakage between neighboring units.
  • Analysis of Gains: The study examined spatial heterogeneity in performance gains ($\Delta R^2$ and $\Delta KL$) and correlated these with structural regional characteristics (e.g., area size, night-time lights intensity, building density).
  • Sufficiency and Complementarity: Ablation studies were conducted to determine the marginal contribution of adding features incrementally and to test whether the two predictor sets provide complementary information.
  • Scale Sensitivity: A sensitivity analysis in the US tested how performance changed when data were aggregated from ZCTAs to the county level.

Key Results

1. Overall Performance Improvement: PDFM embeddings consistently outperformed geospatial covariates across all countries and model types. The median improvement in predictive fit ($R^2$) was 20.1% (reduction in unexplained variance), and KL divergence was reduced by a median of 23.2%.
2. Spatial Heterogeneity of Gains: The benefits of PDFM were not uniform.
   • Context Dependence: PDFM provided the most significant gains in larger, less-developed subnational areas where geospatial covariates (like night-time lights and built-up volume) weakly characterized the settlement context.
   • Developed Regions: In highly developed areas with strong conventional signatures (high night-time lights, dense road networks), the performance gap narrowed or reversed, as traditional covariates already captured the relevant spatial variation effectively.
3. Scale Coupling: A critical limitation was identified regarding spatial scale. While PDFM excelled at the operational scale of its generation (e.g., ZCTAs in the US), its performance degraded when aggregated to coarser administrative units (counties). In contrast, geospatial covariates remained robust at coarser scales, likely because they can be more directly recalculated or summarized to match new target units.
4. Feature Sufficiency:
   • PDFM: A small subset of PDFM dimensions (particularly those related to "Maps & Busyness") was sufficient to recover near-maximal predictive accuracy. Adding the full set of geospatial covariates to a model already containing full PDFM embeddings yielded negligible gains.
   • Geospatial Covariates: Conversely, adding PDFM embeddings to a model with all geospatial covariates resulted in clear performance improvements, suggesting PDFM captures latent signals (e.g., behavioral patterns, regional interests) not fully represented by physical proxies.

Significance and Claims
The paper claims that geospatial foundation-model representations of place can significantly improve population estimation in data-poor settings by encoding complex, multidimensional interactions between human behavior and environmental context that traditional covariates miss. However, the authors emphasize that these gains are conditional rather than universal.

The study concludes that PDFM embeddings should not be viewed as a universal replacement for established geospatial covariates. Instead, they represent a complementary latent representation whose value depends heavily on geography, spatial scale, and the transfer setting. The findings reveal a fundamental limitation of current geospatial AI: while powerful at their native resolution, precomputed embeddings are tightly coupled to their spatial support and do not transfer cleanly across scales. The authors advocate for hybrid modeling frameworks that integrate embedding-based representations with physically grounded settlement indicators to achieve the most robust and transferable population modeling.