S12-Coasts

S12-Coasts is a global, multimodal, patch-based geospatial dataset designed for coastline and water-body segmentation tasks. The dataset consists of 6,637 georeferenced image patches, each with a spatial size of 512 × 512 pixels, paired with automatically generated binary water masks. Data modalities Sentinel-2 (optical): Multi-band optical imagery providing spectral information suitable for discriminating water, land, and coastal features under cloud-free or low-cloud conditions. Sentinel-1 (SAR): Synthetic Aperture Radar imagery enabling robust water detection under all-weather and day/night conditions and complementing optical observations in cloudy or turbid environments. Each patch contains co-registered Sentinel-1 and Sentinel-2 observations covering the same spatial extent, enabling multimodal learning and data fusion approaches. Reference labels Reference water masks were generated automatically using the Ensemble Object-Based Clustering (EOBC) framework. EOBC integrates multiple independent shoreline and water-body vector datasets, including: NOAA shoreline datasets (CUSP), OpenStreetMap (OSM) water and coastline features, HydroLAKES lake polygons. These vectors are fused with satellite imagery through object-based clustering and consensus rules to derive binary water / non-water labels. This fully automated approach removes the need for manual annotation while ensuring spatial consistency across regions.