Landsat-derived European land-cover maps on a 10 m grid for 2010 and 2020
Authors/Creators
- 1. Department of Earth and Environmental Sciences, Lund University
- 2. School of Geosciences, University of Edinburgh
- 3. Royal Botanic Garden Edinburgh
- 4. Tencent
- 5. School of Artificial Intelligence and Robotics, Hunan University
- 6. Smart Group, School of Engineering, Institute for Imaging, Data and Communications, The University of Edinburgh
- 7. Faculty of Environmental and Life Sciences, University of Southampton
- 8. School of Informatics, University of Edinburgh
- 9. LIRA, Observatoire de Paris
- 10. Institute for Logic, Language and Computation, University of Amsterdam
Description
This dataset contains Landsat-derived land-cover maps for the European study region for periods centred on 2010 and 2020. The maps are provided on a 10 m grid and cover approximately 4.90 million square kilometres.
The 2010 map was produced from Landsat-5 TM surface-reflectance composites for 2009–2011, and the 2020 map from Landsat-8 OLI surface-reflectance composites for 2019–2021. Both maps therefore represent multi-year compositing periods rather than observations from a single calendar year. Growing- and non-growing-season observations were classified separately, and their class probabilities were combined to produce the final maps.
Mapping used a globally trained model that jointly reconstructs multispectral information and classifies land cover. Paired Landsat and Sentinel-2 observations and expert land-cover labels were used during training; map production required Landsat imagery alone. The classification comprises nine classes: water, trees, grass, flooded vegetation, crops, shrub and scrub, built area, bare ground, and snow and ice.
Independent assessment against LUCAS field observations yielded 78.8% agreement for the 2020 map. The 2010 map achieved 69.2% agreement with LUCAS 2009 and 69.5% with LUCAS 2012; agreement at the respective stable reference subsets was 73.6% and 74.8%. These values describe agreement at the selected survey locations, rather than area-weighted accuracy across the entire mapped region.
The 10 m grid represents land-cover inference from 30 m Landsat observations and should not be interpreted as native 10 m imagery. The maps support land-cover characterization for the two periods. Direct comparison of their class labels can generate apparent changes through classification errors in either period; land-cover change estimates therefore require additional validation.