Measuring the spatial legibility of urban indicators with a self-supervised earth-observation embedding

Self-supervised Earth observation foundation models compress decades of multi-sensor satellite data into compact embeddings, yet how much non-visible structure, such as wealth gradients and network-mediated accessibility, survives in a purely observational representation is still untested. Using 64-dimensional, 10 m embeddings (2017 to 2024) from Alpha Earth Foundations, which fuse optical, radar, thermal, elevation, and climate signals, this study measures the spatial legibility of land use, transport accessibility, environment, and relative wealth in Hanoi, Vietnam. The EO-native targets are recovered strongly (greenness R 2 = 0.90, surface temperature R 2 = 0.53, air pollution R 2 = 0.43, land cover macro-F1 = 0.56), whereas the non-visible indicators are only partly legible (accessibility R 2 = 0.40, ρ = 0.63; relative wealth R 2 = 0.20, ρ = 0.44). Critically, once skill is referenced to a coordinate-only spatial null, only greenness, land cover, and surface temperature show a positive embedding contribution (+0.89, +0.40, and + 0.20 R 2 ). Accessibility, air pollution, and relative wealth do not exceed geography alone at the 2 km blocking scale, so their apparent legibility reflects largely smooth spatial structure rather than embedding-specific information. The predicted surfaces are strongly coupled (pollution to accessibility 0.78, wealth to accessibility 0.59, reported as linear Centered Kernel Alignment (CKA) between predicted fields, which for scalar targets equals the squared Pearson correlation), recover slowly with limited labels (at 500 labels, accessibility ρ = 0.59 and wealth ρ = 0.39), and the representation changes only modestly over 2017 to 2024 (mean inter-annual velocity about 0.05).