On the Transferability of Earth Observation Foundation Models to Coastal Environments

Earth Observation (EO) Foundation Models (FMs) are primarily trained and evaluated on terrestrial imagery, leaving their transferability to coastal and marine environments largely unexplored. We benchmark EO FMs and supervised baselines on three aquatic segmentation tasks, spanning marine surface features and submerged aquatic vegetation across temperate and tropical waters. We further introduce a tropical coastal dataset comprising 188,857 unlabeled Sentinel-2 RGB patches and use it for domain-adaptive self-supervised pretraining of DINOv3. Results show that standard EO FMs transfer unevenly across benchmarks, providing evidence of a coastal domain gap. Coastal adaptation consistently improves performance, with the largest gains under a frozen-encoder protocol observed on the tropical benchmark most closely aligned with the adaptation corpus. Our findings highlight the impact of domain coverage in EO FM pretraining and demonstrate that domain adaptation can improve their transferability to underrepresented coastal settings. The code, the pretraining dataset, and the weights will be available.