GEOID-Flood: A Large-Scale Multi-Modal Benchmark Dataset for Flood Segmentation
GEOID-Flood pairs flood labels with SAR, optical, and elevation data across 219 events in 65 countries.
The benchmark includes more than 14,000 tiles from Copernicus Emergency Management Service activations over ten years. Its labels separate background, permanent water, and flooded water, with pre- and post-event Sentinel-1, pre-event Sentinel-2 composites, and DEM data co-registered. The authors use it to compare foundation models with conventional encoders across single-image, multi-temporal, and multi-modal setups. They report modest gains from foundation models, stronger flood detection from optical-SAR fusion with finetuning, and better transfer to unseen events than training on existing datasets. HF Daily Papers' note
The benchmark includes more than 14,000 tiles from Copernicus Emergency Management Service activations over ten years. Its labels separate background, permanent water, and flooded water, with pre- and post-event Sentinel-1, pre-event Sentinel-2 composites, and DEM data co-registered. The authors use it to compare foundation models with conventional encoders across single-image, multi-temporal, and multi-modal setups. They report modest gains from foundation models, stronger flood detection from optical-SAR fusion with finetuning, and better transfer to unseen events than training on existing datasets. HF Daily Papers' note
score 5