The limits of ecological decodability of a geospatial foundation-model embedding: a registered evaluation of AlphaEarth over shallow Great Barrier Reef habitats

Adam Cropp 

Annual satellite embedding fields compress a year of multi-sensor observation into one vector per 10 m pixel, worldwide. We test what ecological information the AlphaEarth embedding holds about shallow (< 10 m) Great Barrier Reef habitats, and where its limits lie. Every candidate output was registered with a pass gate before testing, evaluated under one fixed protocol, and reported pass or fail. Ground truth spans 104,668 photo-quadrat images at one reef over seven years, 67 further reef surveys in two eras, and independent point labels and map products. Benthic state is broadly decodable: eight channels pass at 10–100 m, led by hard coral (held-out ρ .73 at 10 m, .80 at 50 m); six replicate on nevertrained reefs. A persistence control shows this skill mostly reads stable habitat structure: an annually refreshed structure map, not a continuous ecological monitor. Year-over-year change is decodable only in the two thermal-stress event years, only at 20–50 m, and only at the reef with change ground truth (coral-loss AUC .83 in 2024); quiet-year change sits below the noise floor. Post-registration tests against two independent monitoring archives bound deployment: cover levels transfer at a quarter to a third of inter-programme agreement, but decoded trajectories fail with a directional bias (recolonised surfaces read as coral, so gains rank below chance). Ground-truth reliability caps decodability near ρ .8; naive cross-validation inflates skill by up to +.28. The catalogue and its limits define what these embeddings can and cannot yet contribute to reef monitoring, and how geospatial foundation models should be evaluated against field ecology.