Copernicus Global Land Cover Layers—Collection 2

: In May 2019, Collection 2 of the Copernicus Global Land Cover layers was released. Next to a global discrete land cover map at 100 m resolution, a set of cover fraction layers is provided depicting the percentual cover of the main land cover types in a pixel. This additional continuous classiﬁcation scheme represents areas of heterogeneous land cover better than the standard discrete classiﬁcation scheme. Overall, 20 layers are provided which allow customization of land cover maps to speciﬁc user needs or applications (e.g., forest monitoring, crop monitoring, biodiversity and conservation, climate modeling, etc.). However, Collection 2 was not just a global up-scaling, but also includes major improvements in the map quality, reaching around 80% or more overall accuracy. The processing system went into operational status allowing annual updates on a global scale with an additional implemented training and validation data collection system. In this paper, we provide an overview of the major changes in the production of the land cover maps, that have led to this increased accuracy, including aligning with the Sentinel 2 satellite system in the grid and coordinate system, improving the metric extraction, adding better auxiliary data, improving the biome delineations, as well as enhancing the expert rules. An independent validation exercise conﬁrmed the improved classiﬁcation results. In addition to the methodological improvements, this paper also provides an overview of where the di ﬀ erent resources can be found, including access channels to the product layer as well as the detailed peer-review product documentation.


Introduction
Land is an important asset for human beings. The globalization of the world's economy and the increase of its population, however, have large environmental consequences and put unprecedented pressure on land management [1][2][3]. To understand these consequences and to act upon them, accurate characterization of land cover and land-use change is essential. This also means records of vegetation characteristics and land cover (LC) need to be available. While long-term consistent in situ observations for large areas remain scarce, satellite remote sensing has become a major source of information for the monitoring of vegetation dynamics since the 70s [4]. Satellite systems like Landsat, National Oceanic and Atmospheric Administration's (NOAA) AVHRR, and MODIS, among others, provide long-term records of reflectance data on a global scale [5][6][7]. These datasets have been used to generate state-of-the-art global discrete LC maps at low to medium resolution [8,9] like the GlobCover [10], LC-CCI [11], and MODIS [12] products. Gong et al. [13] produced the first global high-resolution LC map at 30 m resolution, attaining an overall classification accuracy of 65%. Nevertheless, all these LC maps have a major drawback-they have fixed legends and therefore rigid characterization of LC types. Land cover maps are used by various user groups to better understand the current landscapes and the impact we have as global citizens. But at the same time, these LC maps are often hard to use in applications that require up-to-date information, such as forest monitoring, sustainable development goals monitoring, and biodiversity monitoring. The variety in the needs between all these different users makes it hard to create a one-size-fits-all LC map. More elastic LC mapping approaches are thus needed to guarantee high-quality products for these users [14].
As part of the European Copernicus service which aims to provide systematic global monitoring of the Earth's land surface, the Copernicus Global Land Service has recently released yearly global LC maps at 100 m resolution from 2015 onwards [15]. Copernicus' Global Land Cover product aims to provide dynamic LC layers that allow creating custom LC maps that fit the needs of different map users. Therefore, next to a discrete LC map, a set of cover fraction layers is delivered, depicting the percentual cover of main LC types in a pixel. This allows the user to tailor the LC product to his application and needs, and to harmonize it to other available LC maps.
The first Copernicus Global Land Service Land Cover Map at 100 m (CGLS-LC100)-Collection 1-was provided for the 2015 reference year over the African continent in July 2017. It was derived from 100 m time-series of the vegetation instrument on board of the PROBA satellite (PROBA-V) [16,17], a database of high-quality LC reference sites and several ancillary datasets. We integrated spectral, temporal and spatial features derived from the PROBA-V time series to improve classification accuracy as recommended by Zhai et al. [18] and Eberenz et al. [19]. Based on the success of this demonstration product for Africa, showing high quality with an overall accuracy of 74.3% and alignment to other continental LC maps [20][21][22][23], Collection 2 of the CGLS-LC100 product was released in May 2019, extending the map to a global coverage for the reference year 2015 [15]. However, Collection 2 was not just a global up-scaling, but also includes major improvements in the map quality, reaching now around 80% or more of overall accuracy for each continent. Further improvements include the switch from the system World Geodetic System 1984 (WGS84) to the Universal Transverse Mercator (UTM) coordinate system and the alignment to the Sentinel-2 tiling grid in order to improve mapping quality in the high latitudes [24]. Moreover, this gridding system facilitates the continuation of the yearly CGLS-LC100 product when switching from PROBA-V to Sentinel-2 as an input data source.
The objective of this communication is to introduce potential users to the product, to give a short overview to the methodology, accuracy assessment, and product comparisons. Moreover, the different data access channels to the datasets as well as more detailed product documentation are provided.

General Overview
In this section, a broad overview of the CGLS-LC100 production workflow is given. For a more detailed explanation of the whole workflow, including detailed technical descriptions of algorithms, used ancillary data, and produced intermediate products, we refer to the Algorithm Theoretical Basis Document (ATBD) [25].
It is important to note that the overall workflow used in the Copernicus Global Land Service Land Cover processing workflow is not sensor-specific, and can be applied to any satellite data, irrespective of the sensor or resolution (Figure 1). Where sensors like Sentinel-1, Sentinel-2, and Landsat were up to now mainly used for regional/continental LC mapping applications [26,27], the CGLS-LC100 Collection 2 product uses PROBA-V sensor data, despite the workflow is already tested with Sentinel-2 and Landsat. The CGLS-LC100 product is generated by combining high-quality external data and several proven individual methodologies for processing Earth Observation (EO) data ( Figure 1). Each of the workflow blocks in Figure 1, EO data pre-processing, classification/regression pre-processing, classification/regression and product generation, and validation and comparison are described in the sections below.

EO Data Pre-Processing
One main improvement in Collection 2 of the GCLS-LC100 product was the reprocessing of the PROBA-V archive to improve the quality of the EO input data. The complete PROBA-V L1C archive (unprojected, Top-of-Atmosphere data) was translated into the PROBA-V UTM Analysis Ready Data (ARD) archive following the recommendations for ARD datasets by Dwyer et al. [28]. The geometric processing included the projection into the UTM coordinate system, fully aligned with the Sentinel-2 tiling grid in naming as well as tile dimensions [29]. This reduced distortions in the High North, and will allow continuity of the service beyond the lifetime of the PROBA-V sensor. Through an improved atmospheric correction, the reflectance values were converted to surface reflectance (Top-of-Canopy).
From the PROBA-V UTM ARD, the 5-daily PROBA-V UTM multi-spectral image data with a Ground Sampling Distance (GSD) of~0.001 degrees (~100 m), is used as primary EO data, while PROBA-V UTM daily multi-spectral image data with a GSD of~0.003 degrees (~300 m) as a secondary source. In addition to the Status Mask (SM) cleaning for each image, a temporal cloud and outlier filter was applied to further clean the data, built on a Fourier transformation [30][31][32]. As a final step, median compositing was carried out in order to archive regular 5-daily time steps in the 100 m and 300 m PROBA-V EO time series, upgrading the PROBA-V UTM ARD archive into the PROBA-V UTM ARD+ archive. From this cleaned and outlier screened data, a Data Density Indicator (DDI) was calculated, which was used as an input quality indicator in the supervised learning process.
As a final step in the input EO data preparation, to improve the data density in the 5-daily 100 m time series, the 100 m and 300 m EO datasets were fused using a Kalman filtering approach, based on Sedano et al. [33] and tested by Kempeneers et al. [34], but further adapted to heterogonous surfaces. This upgraded the PROBA-V UTM ARD+ archive into the PROBA-V UTM ARD++ archive, containing temporally cleaned, consistent and dense 5-daily image stacks for all global land masses at 100 m resolution.

Classification/Regression Pre-Processing
A next step is the calculation of the metrics, which are used as general descriptors to enable distinguishing the different LC classes. The PROBA-V UTM ARD++ archive is used as input for this metrics generation step. Instead of using the original reflectance values, more descriptive Vegetation Indices (VI's) are used ( Figure 1). Overall ten VI's are generated using the four reflectance bands of PROBA-V, in detail: Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Structure Intensive Pigment Index (SIPI), Normalized Difference Moister Index (NDMI), Near-Infrared reflectance of vegetation (NIRv), Angle at NIR, HUE, and VALUE of the Hue Saturation Value (HSV) color system transformation, Area Under the Curve (AUC), and Normalized Area Under the Curve (NAUC) (for detailed VI's definition see ATBD [25]). Since the whole time series stack of the VI's for an epoch (three years period giving reference year +/-1 year) would be too many proxies to be used in supervised learning, the time dimension in the data stack was condensed. Temporal descriptors were derived through a harmonic model, fitted through the time steps of each of the VI's based on a Fourier transformation [30,32]. In addition to the seven parameters of the harmonic model using three frequencies above the zero-frequency, used as metrics for the overall level and seasonality of the time series, 11 descriptive statistics (mean, standard deviation, minimum, maximum, sum, median, 10th percentile, 90th percentile, 10th -90th percentile range, time step of the first minimum appearance, and time step of the first maximum appearance) and one textural metrics (median variation of center pixel to median of the neighbors) were generated for each VI. Additionally, four external metrics including the height, slope, aspect, and purity derived at 100 m were derived from a Digital Elevation Model (DEM). Overall, 270 metrics were extracted from the PROBA-V UTM ARD++ archive for the 2015 epoch which includes spectral, temporal, and spatial features as suggested by Zhai et al. [18] and Eberenz et al. [19].
In addition to the input EO data, other data sources were needed as well. Prior to using these external ancillary data, they were checked for consistency and if needed, re-warped to the UTM coordinate system, resampled to 100 m, retiled into the Sentinel-2 tiling grid, and post-processed to usable ancillary data products ( Figure 1). Since water surfaces and built-up surfaces are not classified in the CGLS-LC100 algorithm but fused in from these ancillary data sets, the most important ancillary dataset in the CGLS-LC100 processing line is JRC's Global Surface Water dataset [35] and DLR's World Settlement Footprint (WSF) [36].
Training data has been collected through the Geo-Wiki engagement platform by a group of 20 experts trained by IIASA staff [37]. Note that the training data collection is not through the use of volunteer efforts but instead by a group of highly trained experts involved in the data collection. Therefore, a specific branch of Geo-Wiki, which is only accessible to the experts, has been consolidated for collecting training data at the required resolution and grid. The training data was collected via a 10x10 grid at 10 m spatial resolution through manual classification using Google Maps and Microsoft Bing images. Consequently, the training data not only includes the land cover type, but also the cover fractions of the main LC classes per 100 m resolution cell. Such fractions can be easily translated into different legends. Here, we used LC definitions as suggested by the United Nations (UN) Land Cover Classification System (LCCS) [38]. Overall, 141,000 unique 100x100 m training locations were used in the CGLS-LC100 Collection 2 workflow (see Figure 2 for the geographical distribution of the training data). In order to adapt the classification/regressor algorithm to sub-continental and continental patterns, the classification/regression of the data was carried out per biome, with the biome clusters defined by the combination of several global ecological layers. In the classification and regression preparation, the metrics of the training points were analyzed for inter-specific outliers in the pure endmembers, as well as screened via an all-relevant feature selection approach for the best metrics combinations (best band selection) for each biome cluster in order to reduce redundant information. Therefore, in a first step, the metrics are sorted by their separability, and in the second step, reduced via iteratively removing metrics which are proved by a statistical test to be less relevant than random probes (see [25] for more details).

Classification/Regression and Product Generation
The optimized training data set, together with the quality indicator of the input data (DDI dataset, see Section 2.2) were inputs for the supervised classification algorithms. As the core classifier, Random Forest (RF) techniques were used. All the generated models were optimized via 5-folded cross-validation in order to estimate the optimal classifier/regressor parameters.
RF classification was used to produce a hard classification, showing the discrete class for each pixel, as well as the predicted class probability. This class probability is provided as a quality indicator layer in the final product. In a second step, the vegetation coverage per pixel was generated via RF regression, based on the cover fractions collected for all training points, for each class separately. Moreover, the standard deviation of the applied RF model for each pixel was outputted and added in the product as additional quality layers. Note that in the final product, the cover fraction layers for built-up (urban), permanent inland water bodies and seasonal inland water bodies were derived from the processed ancillary data products.
As a last step, the different classification outputs (discrete map, probability maps as well as the regression maps), were combined with ancillary data sets to produce the final CGLS-LC100 product through a set of expert rules. For example, the classes for permanent snow, built-up, and permanent water were added via rules on the corresponding cover fraction layers. The produced LC map uses a hierarchical legend based on the UN LCCS. Compatibility with existing global LC products was hereby taken into account.

Validation and Comparison
The quality of the CGLS-LC100 product was assessed quantitatively and qualitatively in an independent validation approach (Figure 1). A detailed description of the methods can be found in the validation report [39]. In short: For the quantitative evaluation, the CGLS-LC100 discrete map was assessed using an independent validation dataset which includes 21,752 sampling locations across the globe (2843 to 3616 samples per continent, Table 2 and see Figure 3 for geographical distribution). The sampling protocol followed the probability sampling scheme [20]. Similar to the training data collection, validation data was collected on the Geo-Wiki platform, but using an independent branch from the training data collection [37]. The validation data collection platform was only accessible to around 30 regional experts with knowledge on satellite-based LC analysis and image interpretation, who contributed to the validation data collection [39]. More detailed information on the validation data collection and datasets used for collection is provided in [39,40]. Based on the collected reference information, fractions of LC types were calculated and used to validate the CGLS-LC100 cover fraction layers. Specifically, the mean absolute error (MAE) and root mean square error (RMSE) were calculated. The LC fractions were further translated to a CGLS-LC100 discrete map legend based on the legend descriptions and used to validate the discrete map. More specifically, a confusion matrix was calculated based on the mapped and reference LC types. The error matrix was corrected by sample inclusion probabilities [20]. A detailed quantitative comparison of the CGLS-LC100 and existing global LC datasets can be found in the validation report [39]. In summary, for quantitative comparison, Tsendbazar et al. [40] combined six available reference datasets consisting of around 25,000 points and checked the correspondence of the resulting dataset with the Globcover 2009, LC-CCI 2010, and MODIS2010 global LC maps and as well as the CGLS-LC100 product. Similarly, the CGLS-LC100 validation dataset was also used to assess the Globeland30 2010 for comparison with the CGLS-LC100 2015 Collection 2 maps.
All 20 layers forming the CGLS-LC100 Collection 2 product are distributed via several channels for viewing, on-demand analyses and direct download. See Chapter 4 for more details on the data access channels.

The CGLS-LC100 Product and Accuracy Assessment
The full product specifications (file naming, file format, product content, and characteristics) for each of the 20 layers in the CGLS-LC100 Collection 2 product [15] can be found in the Product User Manual (PUM) [41]. This also includes a full description of the Metadata, LC classes, and definitions, color coding at different classification levels, and found limitations of the dataset.

Global Discrete Map and Cover Fraction Layers
The main product, the CGLS-LC100 discrete map layer (Figure 4), provides a primary LC scheme at three classification levels, 12 classes at level 1 and up to 23 classes at level 3, with classes according to the LCCS scheme [38]. Accomplished is the discrete map layer by the CGLS-LC100 forest-type-layer showing the distribution of evergreen needle-leafed and broad-leafed as well as deciduous needle-leafed and broad-leafed forests. A novelty of the CGLS-LC100 product is the generation of vegetation continuous fields that provide proportional estimates for vegetation cover for all base classes, i.e., trees, shrubland, herbaceous vegetation, cropland, moss and lichen, bare ground/sparse vegetation, permanent snow and ice, built-up and permanent water cover. These nine layers are complemented by a seasonal water cover fraction layer indicating the area of seasonal water within a pixel. Note: this extra layer is not showing the water occurrence, but if a certain area of a pixel had water coverage up to 11 months in the year. An overview of these 10 cover fraction layers is shown in Figure 4.

Quality Indicators
The CGLS-LC100 product delivers, next to the discrete map and cover fraction layers, eight additional quality layers, which can be used to evaluate pixel-wise the quality of the input EO data as well as the RF classification and regression quality: • DDI: The Data Density Indicator indicates the availability of input data from the PROBA-V UTM ARD+ MC5 archive for 100 m and 300 m resolutions. It is a score between 0 = no input data available and 100 = best data availability. Overall, 19 single time-series statistics are calculated and combined via a scoring approach into the DDI. For more details see the ATBD [25]. • Discrete class probability: The probability of the discrete classification indicates the quality of the discrete classification and is provided as a number between 0 and 100, in steps of 1%. The class probability is calculated pixel-wise using the RF model and scoring the classification trees within the RF forest. The higher the probability, the more confident we are that the given class is correct.

•
Cover fraction standard deviation: The standard deviation of the percentage cover regression indicates the quality of the associated cover fraction layer. It is provided as a number between 0 and 100, in steps of 1%. This indicator is given only for the six classes calculated by the CGLS-LC100 workflow (tree, shrubland, herbaceous vegetation, cropland, moss and lichen, and bare/spare vegetation). The standard deviation is calculated pixel-wise over the single regression results of the trees within the RF forest. The lower the standard deviation for a class in a pixel, the higher the confidence that the result is correct.

Accuracy Assessment
A full validation report for the CGLS-LC100 Collection 2 product is available for download [39]. It includes a detailed description of the validation sampling design, sampling data collection, and methods regarding the accuracy assessment including detailed confusion matrixes on global and continental level for classification levels 1 and 2. We report here only a short summary of the results. Table 1 shows the confusion matrix for the discrete layer at classification level 1 (base classes) of the CGLS-LC100 Collection 2 product at a global scale. The overall accuracy is 80.2% +/-0.7% (confidence intervals at 95% confidence level).
In terms of class-specific accuracies, forest, bare/sparse vegetation, snow/ice, and permanent water are mapped with very high accuracies (>85%). The class accuracies of herbaceous vegetation, croplands and built-up are moderate (70%-85%). Built-up class has a higher confusion error with croplands, forest, and herbaceous vegetation. While herbaceous wetlands, lichen/moss, and shrubs have lower-class accuracies (<65%). Specifically, herbaceous wetlands have high confusion errors with herbaceous vegetation and shrubs have high confusion errors with forest and herbaceous vegetation classes which are expected considering the spectral similarity of these classes. Lichen/moss class has higher confusion errors with herbaceous vegetation. All the class accuracies are estimated with high confidence, having confidence intervals less than 8%.
The overall accuracies for the discrete layer at classification level 1 of the CGLS-LC100 Collection 2 product were also calculated for each continent ( Table 2). All continents are mapped with overall accuracies around 80% with the lowest at 77.1% for North America and the highest at 83.3 % for Asia. See the full validation report [39] for in-depth confusion matrices per continent at classification level 1 as well as for detailed validation on classification level 2 (separating open and closed forests). Table 3 lists the mean absolute error (MAE) and root mean square error (RMSE) for the fractional cover layers of the CGLS-LC100 Collection 2 product at a global scale. Rare classes, such as snow/ ice, water, and built-up area, show the lowest errors with MAE of maximum 0.8% and RMSE lower than 6%. Followed by lichen/moss, crops and bare/sparse vegetation cover fraction layers with an MAE 5% and RMSE~15%. Higher MAE and RMSE (MAE~9%, RMSE~17%) are viewable in the tree and shrubland cover fraction layers. Herbaceous vegetation fraction product has the highest error with MAE of 17.3% and RMSE of 28.1%. This can be due to difficulty in separating herbaceous vegetation from other LC types. Table 1. Confusion matrix for the discrete layer at classification level 1 of the CGLS-LC100 Collection 2 product at a global scale. Adapted from [39].   Table 2. Overall classification accuracy of the discrete layer at classification level 1 of the CGLS-LC100 collection 2 product at continental scale. Adapted from [39]. South America 3017 79.6 1.5 Table 3. Mean absolute error (MAE) and root mean square error (RMSE) of the fractional cover layers of the CGLS-LC100 collection 2 product at global scale. Adapted from [39].

Quantitative Comparison of CGLS-LC100 and Existing Global LC Datasets
When compared against the combined reference dataset, the CGLS-LC100 2015 Collection 2 map shows at global scale a higher agreement (by 3%) to the combined reference dataset than the other tested four global LC maps (Table 4) [39]. Similarly, at a continental scale, the CGLS-LC100 2015 Collection 2 map has consistently higher agreements than the other maps (Table 4) [39]. Only in South America, MODIS 2010 map has slightly higher accuracy than the CGLS-LC100 2015 Collection 2 map. Please note, that the objective of this comparison is only to indicate the generic quality of the CGLS-LC100 2015 Collection 2 map as compared to commonly used previous global LC maps. One must be cautious to compare directly the overall agreement of these maps due to the differences in the reference years and legend definitions and other differences between the combined reference dataset and the global LC maps [40]. Anyhow, this result indicates that the CGLS-LC100 2015 Collection 2 map has a comparable, if not better, quality as previous LC maps. Additional comparison studies verified this indication [20][21][22][23]. Table 4. Quantitative comparison of the discrete layer at classification level 1 of the CGLS-LC100 Collection 2 product with other recent global land cover products. Adapted from [39].

Data Access Channels
To improve the accessibility, the CGLS-LC100 Collection 2 product as well as the product documentation is distributed via several access channels:

•
The LC layers are available for viewing through the Global Land Cover viewer, available at https://land.copernicus.eu/global/lcviewer. It displays the various LC layers (discrete map, cover fractions, false-color combinations of cover fractions) on a map, allows us to download the data in 20x20 degree tiles in the EPSG:4326 projection and reports on LC statistics per administrative area.

•
All products are uploaded on the Zenodo platform for long-term archiving and assigned a concept DOI as well as a version DOI. The concept DOI will resolve all the time into the newest collection of the dataset. Moreover, the single layers of the CGLS-LC100 product can be downloaded as single global files in the ESPG:4326 projection (Table 5).

•
All products were ingested into the Google Earth Engine Data Catalog -so now the access through the Google Earth Engine (GEE) offers on-demand analyses, visualization and data download in Pseudo-Mercator projection for customized boundary boxes. Search for the term "Copernicus Global Land Cover Layers: CGLS-LC100 Collection 2" in the GEE Data Catalog.

•
The product is also available through the Geo-Wiki engagement platform available at https: //www.geo-wiki.org/. The user can compare the CGLS-LC100 product with other global and continental LC products as well as provide feedback or help to collect additional training points for the next collection of the product.

•
More information and documentation about the CGLS-LC100 product is available from the Copernicus Global Land Service web site at https://land.copernicus.eu/global/products/lc as well as accessible through the "Copernicus Global Land Service: product documentation" community on the Zenodo platform at https://zenodo.org/communities/copernicus-land-product-documents/

Conclusions and Outlook
Copernicus currently delivers a dynamic global LC product at 100 m spatial resolution for the reference year 2015 (CGLS-LC100 Collection 2). Next to a discrete layer, the product also includes continuous field layers or "fraction maps" for all basic LC classes that provide proportional estimates for vegetation/ground cover for the LC types. This continuous classification scheme may depict areas of heterogeneous LC better than the standard classification scheme and, as such, can be tailored for application use (e.g., forest monitoring, rangeland management, crop monitoring, biodiversity and conservation, monitoring environment and security in Africa, climate modeling, etc.). Moreover, the cover fractions layers allow generating LC maps in different legend schemes (i.e., Common International Classification of Ecosystem Services (CICES) or LCCS). The Minimum Mapping Unit of 100 m and consistent mapping approach allows the usage of the global product on several scales, i.e., from regional to global planning.
The highly adaptable CGLS-LC100 mapping approach was already used by other research groups, e.g., the NatureMap group, a project of the United Nation World Conservation Monitoring Centre, created a global forest management layer at 100 m with our support [42]; or the Food and Agriculture Organization of the United Nations (FAO) funded WaPOR (Water Productivity Open-access Portal) project which generated yearly agriculture maps for Africa starting at 2010 extending the crop class into irrigated and rainfed agriculture [26].
Our assessments show that the overall accuracy of the discrete layer at classification level 1 of the CGLS-LC100 Collection 2 product reaches 80.2+/-0.7% at a global and continental scale. The targeted overall accuracy of 80% has therefore been met. Among the fractional LC layers, snow and ice, built-up, water and lichen/moss fraction maps show the lowest errors, followed by crops and bare and sparse vegetation fraction types. On the other hand, fractional herbaceous vegetation cover layer has the highest error. Our independent quantitative comparisons show that the CGLS-LC100 2015 Collection 2 map has higher accuracy than other available global LC maps.
The third edition of the CGLS-LC100 layers (Collection 3) is currently under preparation covering the 2015 reference year and annual LC changes from 2016 to 2019 over the entire globe. Firstly, the map for 2015 will be improved with additional training data (~40,000 points) focusing on regions with lower accuracies. Secondly, for users looking to map land change processes, such as desertification, de-or re-forestation, urbanization, the impact of major infrastructure developments and so on, it is important to compare LC maps across different years. The new and additional yearly LC layers facilitate these processes. Continuity of the service from 2020 onwards is feasible through the use of Sentinel-1 and Sentinel-2 EO data in the processing line, or going back prior to 2015 through the use of Landsat EO data.