Explain natural-image foundation-model transfer to satellite critical-infrastructure classification

Investigate how and why representations learned by the natural-image foundation model DINOv3 ViT-L/16 transfer to 10 m Sentinel-2 critical-infrastructure classification, including which feature types transfer, at which spatial resolutions, and for which downstream tasks, and determine how this transfer should inform Earth-observation foundation-model selection and pretraining-objective design.

Background

DINOv3 ViT-L/16, which was pretrained on natural RGB web images rather than Earth-observation imagery, achieved the strongest or near-strongest performance in the Infra-Bench CLS evaluation, outperforming the evaluated Earth-observation foundation models on several critical-infrastructure classes. The authors identify this result as an unresolved question in the literature because the mechanisms enabling transfer from natural imagery to satellite-based infrastructure classification are not established.

The unresolved problem concerns both explanation and characterization: determining which learned features are useful, how transfer depends on spatial resolution and task type, and whether the findings should influence the design of Earth-observation-specific pretraining objectives.

References

A natural-image model outperforming EO-specialized foundation models on CI classification is a non-obvious finding, and one that is an open question in the literature. Understanding which feature types transfer, at which resolutions, and for which tasks, would guide FM selection for downstream applications and clarify where EO-specific pretraining investment is warranted.

— Infra-Bench CLS: A Global, Open-Source Benchmark for Critical Infrastructure Classification with Earth Observation Foundation Models  (2609.09482 - Guthrie et al., 8 Sep 2026) in Section Discussion, second key finding; Section Future work, fourth extension