Researchers have introduced GEOID-Flood, a large-scale multi-modal benchmark designed to evaluate geospatial foundation models for flood segmentation tasks. Derived from Copernicus Emergency Management Service activations, the dataset encompasses 219 flood events spanning 65 countries over a ten-year period. It comprises over 14,000 tiles featuring co-registered pre- and post-event Sentinel-1 imagery (available in both GRD and RTC formats), pre-event Sentinel-2 composites, and Digital Elevation Models (DEM). Crucially, the dataset includes manually validated labels that distinguish between background areas, permanent water bodies, and transient flooded water. The study evaluates foundation models against conventional encoders using three distinct protocols: single-image, multi-temporal, and multi-modal. Findings indicate that while foundation models provide a consistent but modest performance advantage, the most effective method for resolving transient flooding is the fusion of optical and SAR data through fine-tuning. Furthermore, the results demonstrate that models trained specifically on the GEOID-Flood benchmark exhibit superior transferability to previously unseen flooding events compared to models trained on existing datasets, highlighting the benchmark's utility for developing robust geospatial AI for disaster management.