Reusable Change Signal: An Outcome-Blind Audit of Annual Earth Embeddings

Precomputed annual Earth embeddings offer a reusable measure of landscape change without task-specific training. We evaluate whether AlphaEarth embedding displacement tracks mapped impacts and adds value beyond specialized optical indices. In an outcome-blind cohort of ten Monitoring Trends in Burn Severity (MTBS) fires, displacement recovered substantial within-fire severity ordering (median Spearman rho=0.712), but showed no consistent advantage over a prespecified Sentinel-2 baseline based on burn-ratio and vegetation-index change. TESSERA produced the same qualitative result. Across eight development cases involving lava emplacement, forest loss, and drought, AlphaEarth displacement was positively associated with every mapped reference, although effect sizes varied. These findings support annual embeddings as portable change representations, but not as replacements for task-specific measures or operational hazard detectors.