Reliable general one-shot and ten-shot performance for AlphaEarth

Establish adequate and reliable general one-shot and ten-shot performance for the AlphaEarth Foundations Model in extreme low-data regimes.

Background

The AlphaEarth Foundations Model is presented as a geospatial embedding model designed to support tasks including change detection and post-disaster building-damage analysis. However, its benefits become less consistent when only extremely small amounts of labeled data are available.

The authors report substantial variability across 1-shot evaluations and identify the lack of dependable general performance in 1-shot and 10-shot settings as an unresolved research frontier. This is particularly relevant immediately after disasters, when labeled examples are scarce.

References

For 1000-fold 1-shot trials, AEF error reductions were only $>$ 1.0$\times$ in the ~90\% confidence interval in 5/15 evaluations with average variation ±0.49$\times$[...] we consider the extreme degree of variability indicative that adequate general 1-shot or 10-shot performance remains an unsolved research frontier.

— Revisiting the Task of Building Damage Assessment by Satellite Imaging  (2610.13857 - Cioppa et al., 9 Oct 2026) in Section 6, subsection “The case of AlphaEarth”