A Multi-Label Sentinel-2 Dataset of Surface Water, Wetlands, and Riparian Landscapes in Germany

Lukas Pasold, Felix Waigner, Luca Eisentraut, Ricardo Buettner

This dataset contains 10,000 Sentinel-2 true-color image chips for multi-label classification of surface-water, wetland, and riparian landscape features across Germany. Each image is annotated independently with four binary labels: river or canal, standing water, wetland, and riparian woodland. These labels result in 16 possible combinations, including images containing none of the four categories. The dataset is fully balanced and comprises 625 image chips for each label combination. Each image is provided as a 224 × 224 pixel RGB PNG file and represents a 512 m × 512 m geographic footprint. Label annotations were derived from OpenStreetMap geometries and recomputed for the complete footprint of every image. The river-or-canal category includes geometries tagged waterway=river, waterway=canal, waterway=riverbank, and natural=water combined with water=river or water=canal. A chip received this label when its footprint contained at least 180 m of a retained linear river or canal feature or at least 3,500 m² of a retained river or canal polygon. The standing-water category includes geometries tagged natural=water in combination with water=lake, water=reservoir, or water=pond. The wetland category is based on geometries tagged natural=wetland. For both categories, at least 5,000 m² of the respective mapped area had to fall within the complete image footprint. The riparian-woodland category includes woodland geometries tagged natural=wood or landuse=forest. A chip received this label when at least 5,000 m² of retained woodland occurred within its footprint and a mapped river, canal, lake, reservoir, or pond was located within the footprint or its surrounding 100 m search buffer. The riparian-woodland label therefore represents woodland associated with mapped surface water and should not be interpreted as a general forest-presence label. The four labels are not mutually exclusive. A single image may contain any combination of rivers or canals, standing water, wetlands, and riparian woodland and receives all applicable labels. Source locations were drawn from retained OpenStreetMap line and polygon geometries and spatially distributed across Germany using a round-robin selection procedure. Positive samples were selected separately for each of the 15 non-empty label combinations. Samples without any of the four labels were generated from neighboring locations and accepted only when none of the category-specific visibility thresholds was met within the complete image footprint. Streams, ditches, drains, swimming pools, basins, technical storage tanks, and covered or underground waterways were excluded from the dataset definition. Consequently, the absence of a label denotes the absence of the corresponding retained OpenStreetMap category at the specified visibility threshold; it does not necessarily imply the complete absence of water, vegetation, or other related landscape features. Image chips were generated from Sentinel-2 Level-2A imagery using bands B04, B03, and B02 as red, green, and blue channels. Imagery was retrieved through the Copernicus Data Space Ecosystem Sentinel Hub Process API for the period from July 11, 2025, to July 10, 2026. The collection was generated using a maximum cloud-cover threshold of 10 percent and least-cloud-cover mosaicking. The repository includes the complete image collection and a compact CSV table containing image filenames and the four binary labels. The dataset can be used for supervised multi-label image classification, benchmarking of computer vision models, representation learning, surface-water and wetland mapping, riparian landscape analysis, and environmental remote-sensing research.