Current state and challenges in producing large-scale land cover maps: review based on recent land cover products

Figures & data

Table 1. Main characteristics of analysed land cover products.

	CGLS-LC100	ESA WorldCover	ELC10	S2GLC	ODSE-LULC	Dynamic World
Spatial resolution and CRS	1°/1008 (≈ 100 m)EPSG:4326	1°/12000 (≈ 10 m)EPSG:4326	10 mEPSG:3035	10 mUTM CRS	30 mEPSG:3035	10 mUTM CRS
Main input data	Images from VEGETATION instrument on PROBA-V (multitemporal)	Sentinel-1 and Sentinel-2 images (multitemporal)	Sentinel-1 and Sentinel-2 images (multitemporal)	Sentinel2 images (multitemporal)	GLAD Landsat ARD images	Sentinel-2 images
Other input data	Spectral metrics (spectral indices) Textural metrics Descriptive statistics Input data density indicator Eco-regions CopernicusDEM (90 m) auxiliary: WSF, OSM, GSW, GSHHG…	Spectral metrics (spectral indices) Descriptive statistics Eco-regions CopernicusDEM (30 m) Pixel coordinatesauxiliary: OSM, GHSL, GMW, CGLOPS permanent ice	Spectro-temporal metrics (spectral indices) Descriptive statistics Texture features Environmental covariates SRTM DEM data ERA5 (precipitation and temperature) VIIRS DNB (night light images)	Normalized differences between all spectral bands	Spectral metrics (spectral indices) VIIRS DNB (night light images) GSW TRI for Landsat green band Monthly geometric temperature Continental EU DEM (OpenStreetMap data, HRLs)	(Eco-regions, MCD12Q1, NLCD 2016, MapBiomas 2017)
Training data	Manual labelling (≈ 22 000 samples)	Manual labelling (≈ 22 000 samples)	LUCAS dataset (≈ 53 000 samples)	CLC (2012) and HRLs	LUCAS dataset and filtered CLC centroids (≈ 8.1 mil. samples)	Manual labelling (≈ 24 000 samples)
Classification algorithm	Random forest (supervised learning)	CatBoost (supervised learning)	Random forest (supervised learning)	Random forest (supervised learning)	Ensemble machine learning (random forest, gradient boosted trees, and neural networks)	Semi-supervised deep learning (FCNN)
Classification scheme, classes	FAO LCCS22 classes: 6 closed forest classes 6 open forest classes Shrubs Herbaceous vegetation Herbaceous wetlands Moss and lichen Bare / sparse vegetation Cropland, agriculture Urban/built-up Snow and ice Permanent water bodies Open sea	FAO LCCS11 classes: Tree cover Shrubland Grassland Cropland Built-up Bare / sparse vegetation Snow and ice Permanent water bodies Herbaceous wetlands Mangroves Moss and lichen	8 classes: Artificial land Bare land Cropland Grassland Shrubland Water Wetland Woodland	13 classes: Artificial surfaces Cultivated Vineyards Herbaceous Deciduous Coniferous Moors and Heathland Sclerophyllous Natural material surfaces Permanent snow Marshes Peat bogs Water bodies	CLC classification system5 CLC level-1 classes14 CLC level-2 classes43 CLC level-3 classes (excluding class ‘5.2.3 Sea and ocean’)	9 classes (similar to the IPCC Good Practice Guidance classes): Water Trees Grass Flooded vegetation Crops Shrub & Scrub Built area Bare ground Snow & Ice
Overall classification accuracy	81% (with consolidated forest classes)75% (all classes)	74%	90%	86% (all classes) 89% (after combining spectrally similar classes)	84% (level 1) 64% (level 2) 50% (level 3)	73%
Temporal coverage	Yearly, 2015–2019	2020, 2021	2018	2017	Yearly, 2000–2019	NRT (every mostly cloudless Sentinel-2 image)
Geographical coverage	Global	Global	Europe	Europe	Europe	Global

Abbreviations: Analysis Ready Data (ARD), Copernicus Global Land operations (CGLOPS), Copernicus Global Land Services Land cover map at 100 m (CGLS-LC100), CORINE (short for Coordination of information on the Environment) Land cover (CLC), coordinate reference system (CRS), digital elevation model (DEM), day/night band (DNB), Sentinel-based pan-European land cover map (ELC10), European petroleum Survey Group (EPSG), European Centre for Medium-Range Weather Forecasts (ECMWF) Reanalysis v5 (ERA5), Food and Agriculture Organization (FAO), fully Convolutional Neural Network (FCNN), Global human Settlement layer (GHSL), Global Land Analysis and Discovery laboratory at the University of Maryland (GLAD), Global Mangrove Watch (GMW), Global self-consistent, hierarchical, high-resolution geography (GSHHG), European Commission joint research center’s Global surface water (GSW; Pekel et al. 2016), high resolution layer (HRL), intergovernmental panel on climate change (IPCC), Land cover classification system (LCCS), Land use and coverage area frame Survey (LUCAS), terra and aqua combined moderate resolution imaging Spectroradiometer (MODIS) Land cover type (MCD12Q1), National Land cover database (NLCD), near Real-Time (NRT), open Data Science Europe Land use/Land cover (ODSE-LULC), OpenStreetMap (OSM), project for on-Board Autonomy – vegetation (PROBA-V), Sentinel 2 Global Land cover (S2GLC), shuttle radar topography mission (SRTM), Universal Transverse Mercator (UTM), visible infrared imaging Radiometer Suite (VIIRS), word Settlement footprint (WSF).

Figure 1. Visualizations of analysed land cover products and main data sources used for their production over the city of Trogir in Croatia. Each column represents one land cover product. First in each column is the land cover map with classes coloured according to legends shown at the bottom of each column. CGLS-LC100 land cover map is from 2019 (Buchhorn, Smets, et al. 2020), ESA WorldCover is from 2020 (Zanaga et al. 2021), ELC10 (Venter and Sydenham 2020) from 2018, S2GLC (Malinowski et al. 2020) from 2017, ODSE-LULC (only 14 level-2 classes, available at https://stac.ecodatacube.eu/) from 2019, and the Dynamic World (Brown et al. 2022) map is from 17 April 2020. All legends are complete, except the one for the CGLS-LC100, for which classes related to forests are aggregated. Between the land cover map and legend, in each column are visualizations of the main data sources that have been used for producing a particular land cover map. Copernicus Sentinel-2 image (R–G–B = B4–B3–B2) was acquired on 17 April 2020, ESA’s PROBA-V image (R–G–B = red band – NIR (near infrared) band – SWIR (shortwave infrared) band, as suggested by Veljanovski, Čotar, and Marsetič (2019)) was acquired on 16 April 2020, USGS’s Landsat 8 image (R–G–B = B4–B3–B2) was acquired on 4 April 2020, and Copernicus Sentinel-1 image (R–G–B = VH band – [(VH + VV)/2 · 1.1 + 30] – VV band (VV – vertical transmit and vertical receive, VH – vertical transmit and horizontal receive), partly adopted from Syrris et al. (2020)) was acquired on 16 April 2020. Light blue pixels on the Sentinel-1 composite correspond to urban areas (Syrris et al. 2020). Abbreviations: Copernicus Global Land Services Land Cover map at 100 m (CGLS-LC100), Sentinel-based pan-European land cover Map (ELC10), European Space Agency (ESA), Open Data Science Europe Land use/Land Cover (ODSE-LULC), Project for on-Board Autonomy – Vegetation (PROBA-V), Sentinel-2 Global Land Cover (S2GLC), U.S. Geological Survey (USGS).

Figure 2. Main steps in producing land cover products. Trapezoids represent input or output data, rectangles correspond to data that are the result of intermediate processing, ovals denote processes, and a diamond-shaped symbol represents a decision.

Pekel J-F, Cottam A, Gorelick N, Belward AS. 2016. High-resolution mapping of global surface water and its long-term changes. Nature. 540(7633):418–422. doi:10.1038/nature20584.

 PubMed Web of Science ®Google Scholar

Buchhorn M, Smets B, Bertels L, Roo BD, Lesiv M, Tsendbazar N-E, Herold M, Fritz S. 2020. Copernicus global land service: land cover 100m: collection 3: epoch 2019: globe. Zenodo. doi:10.5281/zenodo.3939050.

 Google Scholar

Zanaga D, Van De Kerchove R, De Keersmaecker W, Souverijns N, Brockmann C, Quast R, Wevers J, et al. 2021. ESA WorldCover 10 m 2020 V100. Zenodo. doi:10.5281/zenodo.5571936.

 Google Scholar

Venter ZS, Sydenham MA. 2020. ELC10: European 10 m resolution land cover map 2018. Zenodo. doi:10.5281/ZENODO.4407051.

 Google Scholar

Malinowski R, Lewiński S, Rybicki M, Gromny E, Jenerowicz M, Krupiński M, Nowakowski A, Wojtkowski C, Krupiński M, Krätzschmar E, et al. 2020. Automated production of a land cover/use map of Europe based on Sentinel-2 imagery. Remote Sensing. 12(21):3523. doi:10.3390/rs12213523.

 Web of Science ®Google Scholar

Brown CF, Brumby SP, Guzder-Williams B, Birch T, Hyde SB, Mazzariello J, Czerwinski W, Pasquarella VJ, Haertel R, Ilyushchenko S, et al. 2022. Dynamic world, near real-time global 10 m land use land cover mapping. Sci Data. 9(1):251. doi:10.1038/s41597-022-01307-4.

 Web of Science ®Google Scholar

Veljanovski T, Čotar K, Marsetič A. 2019. Large scale mosaicing and compositing of PROBA-V satellite images. Geodetski Glasnik. 53(50):5–18.

 Google Scholar

Syrris V, Corbane C, Pesaresi M, Soille P. 2020. Mosaicking Copernicus Sentinel-1 data at global scale. IEEE Trans Big Data. 6(3):547–557. doi:10/gh3ctv.

 Web of Science ®Google Scholar

Data availability statement

Data sharing is not applicable to this article as no new data were created or analysed in this study.