AlphaEarth Distinguishes Cities but Compresses Urban Variation

Author: Andrew Renninger

Submitted: 24 September 2026

arXiv: 2609.30356 [cs.CV]

Subjects: Computer Vision and Pattern Recognition (cs.CV); Physics and Society (physics.soc-ph)

DOI: 10.48550/arXiv.2609.30356

Abstract

Cities differ in built form, land cover, and development history, which complicates comparison across places and time. Satellite foundation models map Earth's surface onto common numerical representations. Yet the tasks and targets used to shape them typically do not focus on cities: globally consistent labels for urban function do not exist, and many datasets—especially land cover and land use classifications—collapse the built environment into a few classes.

Here we audit the representation, focusing on AlphaEarth but with broader applicability to other Earth embeddings, by probing the geometry and geography of embeddings for 1,000 urban areas in 162 countries. We find that cities occupy a shifted but overlapping region on the hypersphere, 62.7° from the global mean direction, and continent and climate predict 24.3% of variation among the mean directions of urban centres in excluded countries. Inside cities, degrees of urbanisation carry 8.9% of the variation, and what they leave holds shared directions whose local orientation varies, not one universal axis of urbanisation.

Retained variation is itself unequal: dispersion within urban centres is 14.1% greater per standard deviation of national development, even after adjusting for population, land area, and continent. Further controls suggest cities in developing countries present less contrast in vegetation and texture, and dispersion follows that contrast: full adjustment for it leaves at most 6.4% of the gradient. Annually, a city's representation moves nearly eight times more than redrawing its own pixels explains, and contracts where the 2022 loss of Sentinel-1B removed a pass direction. AlphaEarth's representations therefore support comparison across regions, while the differences between its annual layers are not yet validated for comparison over time.

Key Findings

• Cities cluster distinctly but overlap. Urban areas occupy a shifted yet overlapping region on the embedding hypersphere, sitting 62.7° from the global mean direction.
• Continent and climate drive variation. These two factors explain 24.3% of the variation among mean directions of urban centres in excluded countries.
• Urbanisation signals are partial. Within cities, degrees of urbanisation account for 8.9% of variation; the remainder consists of shared directions with locally varying orientation rather than a single universal urbanisation axis.
• Inequality in representation. Dispersion within urban centres is 14.1% greater per standard deviation of national development, even after controlling for population, land area, and continent.
• Vegetation and texture contrast explain much of the gap. Cities in developing countries show less contrast in these features; full adjustment for this contrast leaves at most 6.4% of the development gradient.
• Temporal embeddings are unstable. A city's representation shifts annually by nearly eight times more than pixel-level changes would predict, and contracts where the 2022 loss of Sentinel-1B removed a pass direction. Annual layers are therefore not yet validated for temporal comparison.

Significance

This audit demonstrates that AlphaEarth—and by extension other satellite foundation models—supports robust cross-regional comparison of urban areas. However, the annual differences between its embedding layers should not yet be interpreted as reliable indicators of urban change over time. These findings carry implications for downstream applications in urban planning, climate resilience, and socioeconomic mapping, where temporal consistency is essential.

Citation

@article{renninger2026alphaearth,
  title={AlphaEarth distinguishes cities but compresses urban variation},
  author={Renninger, Andrew},
  journal={arXiv preprint arXiv:2609.30356},
  year={2026}
}

via ArXiv CV