GL-FCR: A Global–Local Feature Context and Reconstruction Decoder for Transferable Foundation Model-Based Flood Mapping

Accurate and transferable flood segmentation from Earth observation (EO) imagery remains challenging because models trained on specific geographic regions and environmental conditions often exhibit limited generalization. This study proposes Global–Local Feature Context and Reconstruction (GL-FCR), a task-specific downstream decoder designed to adapt compact EO foundation-model representations for spatially detailed flood mapping. Integrated with Prithvi-EO-2.0, GL-FCR enhances the reconstruction of dense flood maps while maintaining contextual information from the pretrained representation. On the SEN1Floods11 standard test set, Prithvi + GL-FCR achieved a flood F1 score of 91.08%, compared with 90.02% on the geographically held-out Bolivia subset in the original experiment. Across three random seeds, GL-FCR achieved mean flood IoU scores of 83.07 ± 0.50% and 81.53 ± 0.27% on the standard test and Bolivia subsets, respectively, demonstrating reproducible performance across training runs. Additional SAR-only and optical–SAR experiments showed that TerraMind-B + GL-FCR achieved the strongest SAR-only performance, with F1 scores of 81.53% and 82.28% on the standard test and Bolivia subsets, respectively, whereas adding SAR to strong clear-scene optical representations provided limited additional benefit. Furthermore, zero-shot cross-dataset evaluation across all 509 WorldFloods v2 scenes yielded a flood F1 score of 84.56% and a flood IoU of 73.24% for Prithvi + GL-FCR, compared with 83.35% and 72.15%, respectively, for the Prithvi native decoder, without additional training or adaptation. Overall, the results demonstrate that task-specific global–local reconstruction can effectively leverage EO foundation-model representations for accurate and transferable flood segmentation across geographic regions and datasets.