A Multi-Label Sentinel-2 Dataset for Road and Rail Classification

This dataset contains 10,000 Sentinel-2 true-color image chips for multi-label road and railroad presence classification across France. Each image is annotated independently for the presence of road infrastructure and railroad infrastructure, resulting in four possible label combinations: road and rail present, rail present only, road present only, and neither road nor rail present. The collection comprises 2,500 chips for each of these four combinations. Each image is provided as a 224 × 224 pixel RGB PNG file and represents a 512 m × 512 m geographic footprint. Road sample locations were derived from OpenStreetMap geometries tagged highway=motorway, highway=trunk, highway=primary, highway=secondary, highway=tertiary, highway=motorway_link, highway=trunk_link, highway=primary_link, highway=secondary_link, highway=tertiary_link. Railroad sample locations were derived from geometries tagged railway=rail. Locations were spatially balanced across the retained road and railroad networks and selected using line-sample centers spaced at 2.5 km intervals. Candidates were accepted only when their complete image footprints did not overlap previously accepted samples and satisfied the respective road and rail intersection rules. Road and rail segments marked as tunnels were excluded from positive sample selection, and candidates whose complete image footprints intersected retained road or rail tunnel segments were rejected. Images labeled with both road and rail were required to intersect retained road and retained rail geometries within their complete 512 m × 512 m footprints. Images labeled rail only were required to intersect retained rail geometries but not retained road geometries. Conversely, images labeled road only were required to intersect retained road geometries but not retained rail geometries. Thus, road and rail labels are not mutually exclusive: a single image may contain both infrastructure types and receives both labels accordingly. Road and rail geometries not included in the dataset definition, such as residential roads, service roads, tracks, paths, cycleways, rail yards, sidings, spurs, crossovers, and disused, abandoned, proposed, razed, or under-construction rail segments, were excluded from the respective label definitions. Locations with neither road nor rail label were generated from neighboring candidate locations in the local surroundings of retained road and rail networks. They were accepted only when their complete 512 m × 512 m footprints did not intersect retained road or retained rail geometries and did not overlap accepted image footprints. The absence of both labels therefore denotes the absence of the road and rail categories included in the dataset definition; it does not imply the absence of all transport-related 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 1, 2025, to June 30, 2026. The collection was generated using a maximum cloud-cover threshold of 10 percent and least-cloud-cover mosaicking. The repository includes the image collection and CSV metadata tables containing the image filename and the corresponding labels. The dataset can be used for supervised multi-label image classification, benchmarking of computer vision models, representation learning, road and rail infrastructure mapping, and research on transport infrastructure.