OSMlanduse: A 10 m Resolution Land Use Map for the European Union

Introduction

Humans have reshaped more than three quarters of the world’s land surface in the last century. Detailed information about how land is used—whether for agriculture, forest, urban development or natural cover—is essential to understanding environmental change, guiding policy and managing resources. Until now, continent-scale land use maps have been limited by coarse spatial resolution, infrequent updates or restricted licensing. Traditional products like CORINE Land Cover offer only 100 m or coarser resolution with update cycles spanning several years, and often under restrictive terms.

The democratization of satellite remote sensing and citizen-science mapping has created a new opportunity to generate high-resolution, openly licensed land use datasets. In this article, we introduce OSMlanduse—a fully open, 10 m resolution land use map of the European Union. By fusing volunteer-contributed OpenStreetMap (OSM) labels with Copernicus Sentinel-2 imagery and modern deep learning methods, OSMlanduse delivers detailed and timely land use information across 28 countries, with 89% overall accuracy.

Why High-Resolution Land Use Matters

Land use and land cover data underpin a broad range of applications:

• Environmental monitoring: tracking deforestation, wetland loss and urban sprawl.
• Agriculture management: optimizing crop planning, irrigation and yield forecasting.
• Urban planning: guiding sustainable infrastructure development and green space allocation.
• Disaster response: mapping vulnerable zones for floods and wildfires.

For example, wetlands provide critical ecosystem services such as water purification and flood buffering, yet they often occupy narrow strips or fragmented patches that coarse products miss entirely. Similarly, distinguishing urban parks from surrounding built-up areas requires fine spatial detail to guide heat island mitigation and green infrastructure design.

The Open Data Revolution: OSM and Sentinel-2

OpenStreetMap (OSM) is the world’s largest volunteer-driven mapping project. Contributors worldwide add roads, buildings, land use tags and points of interest, creating a freely accessible spatial database. While incredibly rich in some regions, OSM coverage can vary due to local mapping activity, leaving gaps in others.

Copernicus Sentinel-2 satellites, operated by the European Space Agency (ESA), capture multispectral imagery at 10 m resolution. Their optical sensors record visible (blue, green, red) and near-infrared bands, enabling us to distinguish vegetation, soil and water. Sentinel-2 data are free and updated every five days at the equator, offering a consistent, continent-wide view.

By combining OSM labels with Sentinel-2 imagery, OSMlanduse overcomes each source’s limitations: volunteer-verified labels add fine detail where available, and deep learning fills gaps using satellite patterns where OSM data are sparse.

Methodology Overview

1. OSM Label Extraction

• Downloaded the March 2020 OSM planet dump and extracted relevant land use tags for each EU country.
• Translated common tags (e.g., landuse=forest, natural=scrub, amenity=grave_yard) into 13 classes following the CORINE Land Cover 2-level legend.
• Direct OSM labels covered 61.8% of the EU area, represented as vector polygons with sub-meter precision.

2. Sentinel-2 Composite Generation

• Retrieved three years (2017–2020) of Sentinel-2 multispectral imagery, focusing on RGB and near-infrared bands at 10 m resolution.
• Applied cloud and shadow masking using the Fmask algorithm, then constructed a medoid “best pixel” composite to minimize atmospheric noise and seasonal gaps.
• Projected the composite to the European grid (EPSG:3035) for consistent processing.

3. Deep Learning Classification

To bridge unlabelled areas, we harnessed deep learning:

• Trained country-specific convolutional neural networks (ResNet-style) on balanced samples of OSM-labeled pixels, learning spectral and textural patterns for each land use class.
• Separate models per country captured local spectral variations—field crops in southern climates differ from northern grassland, for example.
• Classification of missing regions was followed by spatial smoothing and merging with original OSM labels, preserving volunteer contributions.

4. Post-processing and Harmonization

• A sieve filter removed isolated misclassified pixels smaller than a 64-pixel cluster, smoothing the final map.
• Remaining outlier pixels were reclassified to align with the CORINE legend.
• Predicted areas and direct OSM labels were merged, giving priority to sub-meter OSM vector data.

Dealing with OSM Tag and Data Quality

Volunteer-driven data sources like OSM inherently vary in completeness and thematic depth. Contributors may use different tags for similar features based on local conventions. To address this:

• We curated a standardized mapping of common OSM tags to CORINE classes, documented in our GitHub repository.
• Spatial clustering analysis identified and removed spurious or outdated labels, improving training data quality.
• Countries with sparse OSM coverage relied more heavily on Sentinel-2 classification, while those with dense OSM data preserved vector detail at sub-meter resolution.

Technical Validation

Rigorous accuracy assessment followed international remote sensing standards:

• Stratified random sampling allocated reference points proportionally across classes.
• A crowd-sourced campaign with over 60 volunteers yielded 4 616 interpreted points using high-resolution Bing and Google imagery. Each point was independently labeled by at least three interpreters.
• Confusion matrices revealed the greatest class confusion between “Artificial non-agricultural vegetated areas” and “Shrub and/or herbaceous vegetation associations,” reflecting their similar spectral signatures.
• Despite these challenges, the final map achieved an overall accuracy of 89%, with user’s and producer’s accuracies exceeding 90% for most classes.

Data Records and Access

OSMlanduse is openly licensed under the Open Database License (ODbL) and distributed as GeoTIFFs:

• 28 GeoTIFFs (EU countries) + 1 GeoTIFF (United Kingdom), each at 10 m spatial resolution.
• Pixel values correspond to the 13 CORINE Land Cover Level 2 classes; metadata files include a legend and processing history.
• Download via DOI: https://doi.org/10.11588/data/IUTCDN
• Interactive map viewer: https://osmlanduse.org

GeoTIFFs can be seamlessly integrated into GIS software. Users can query, visualize, and analyze land use patterns, calculate zonal statistics or combine with other environmental datasets.

Applications and Use Cases

Agriculture and Resource Management

Detailed field-scale land use maps enable precision agriculture: farmers can detect early signs of crop stress, plan irrigation routes and identify underutilized land.

Biodiversity and Conservation

Fine-scale mapping of wetlands, forests and semi-natural areas helps conservationists monitor habitat fragmentation, design ecological corridors and prioritize restoration projects.

Urban Planning and Infrastructure

City planners can analyze green space distribution, assess impervious surface cover, and model urban heat islands to inform sustainable development and climate adaptation strategies.

Disaster Risk and Response

Accurate delineation of floodplains, landslide-prone zones and wildfire susceptibility regions supports proactive disaster risk reduction and emergency response planning.

Education and Citizen Science

The open nature of OSMlanduse fosters community engagement: mappers can validate and improve local labels, educators can incorporate real-world data into GIS and remote sensing coursework, and NGOs can conduct targeted environmental assessments.

Community, Licensing and Future Directions

OSMlanduse embodies the spirit of open science:

• All processing scripts are available under GNU GPL3 on GitHub: https://github.com/schultzheidelberg/OpenStreetMap-land-use-for-Europe-2020
• ODbL licensing allows both commercial and non-commercial reuse with attribution, lowering barriers for integration into applications and research pipelines.
• Future plans include integrating radar (Sentinel-1) and LiDAR data for improved wetland and canopy mapping, as well as automating periodic updates through cloud computing platforms.
• We encourage feedback, contributions and regional extensions via pull requests and issues on the project repository.

Conclusion

OSMlanduse delivers the first openly licensed, continent-scale land use map of the European Union at 10 m resolution. By harnessing volunteer-driven OpenStreetMap labels, Copernicus Sentinel-2 imagery and modern deep learning, it achieves an 89% overall accuracy and up to 99% class accuracies. The dataset is fully reproducible, global-ready and free to download, offering unprecedented opportunities for environmental monitoring, sustainable planning, disaster management and community engagement. We invite researchers, planners and citizen scientists to explore, adapt and improve OSMlanduse as part of a growing ecosystem of open geospatial data.

Acknowledgments

This work was funded by the EU Horizon 2020 LandSense project (Grant No. 689812). We thank all volunteer contributors in mapathon and crowdsourcing campaigns, and the technical teams at FAO SEPAL and Heidelberg University’s high-performance computing center for data preprocessing and storage services.