LCM2021 used Sentinel-2 MultiSpectral Instrument (MSI) Level-2A surface reflectance satellite data (Drusch et al., 2012; Claverie et al., 2018) acquired and pre-processed in Google Earth Engine (Gorelick et al., 2017). The images were acquired between 1 December 2020 and 31 January 2022. All 10 and 20 m spectral bands, comprising bands 2 (490 nm), 3 (560 nm), 4 (665 nm), 5 (705 nm), 6 (740 nm), 7 (783 nm), 8 (842 nm), 8a (865 nm), 11 (1610 nm) and 12 (2190 nm), were used.

Context data were used as additional inputs to the classification process to enable better classification of the required land cover classes (Rogan et al., 2003). The context data included a digital elevation model (DEM); coastline, foreshore and tidal water layers (to aid coastal classification); building and road layers (to reduce confusion between arable and urban areas) (Table 1); and freshwater and forest layers. The DEM was used to calculate slope and aspect, which were also included as context layers. National cartographic products for Great Britain (GB) were provided by the Ordnance Survey (OS), the national mapping agency of GB, whereas for Northern Ireland (NI), products were provided by a number of government organisations including the NI Statistics and Research Agency (NISRA); the Ordnance Survey Northern Ireland (OSNI); and the NI Department of Agriculture, Environment and Rural Affairs (DAERA). Slightly different context products were available for NI compared to GB (Table 1). The main difference between the OSNI and Ordnance Survey Great Britain (OSGB) context data is the lack of an NI equivalent to the GB building layer. The OS layers were converted from vector to raster data, with the “distance from” layers created for buildings, roads, rivers and water bodies. The “distance from” products were used to allow the context data products to influence a wider area rather than just the pixels they intersected with. The 10 m NEXTMap digital elevation model (DEM) was used to calculate slope and aspect, with elevation, slope and aspect included as context layers.

LCM2021 is produced through supervised classification of satellite images, an empirical process that requires training areas of known land cover type. The training areas for the classification were widely distributed to capture the range of spectral signatures typical of each class. For LCM2021, training areas were primarily harvested from existing vector data from LCM2018 (Morton et al., 2020a, b), LCM2019 (Morton et al., 2020c, d) and LCM2020 (Morton et al., 2021e, f). The method is described in Sect. 3.2.1.

The LCM spatial framework is a set of land parcel polygons summarising the landscape of the UK into real-world objects such as lakes, fields, woodlands and urban sites. It was derived from generalised digital cartography (Ordnance Survey MasterMap™ topographic layer (OSMM) for GB and Ordnance Survey of Northern Ireland (OSNI) Large-scale Vector for NI), supplemented with rural payment boundary data (Smith et al., 2007; Morton et al., 2011a). The spatial framework was first generated for LCM2007 and revised for LCM2015 onwards by fixing some minor spatial errors and making additional simplifications in terms of land parcel structure. The spatial framework is used to derive a land parcel data set from which 25 m and 1 km raster data sets are generated.

Validation data are necessary to establish the accuracy of land cover classifications (Foody, 2002). LCM2021 validation used a UK-wide data set of 35 182 points gathered from field observations, manual interpretation of aerial photography and quality-assured third-party data sets (Fig. 1). The validation data included habitat-mapping and plot data from Countryside Survey data (Wood et al., 2017), supplemented with additional points for arable land (8589 points) collected in 2020 by the Rural Payments Agency. Data from the National Forest Inventory (NFI, 2019) were used to validate the broadleaved woodland and coniferous woodland classes for GB. Further data were gathered from the 2007 LCM validation field survey (Morton et al., 2011a) and were checked against current (circa-2021) aerial photography to ensure no change had occurred; some additional manually derived points (interpreted from aerial photography) were also added, particularly for water and urban classes.

Temporal composite images (also known as temporal aggregations) are increasingly used to compress voluminous image collections and to overcome problems of data gaps caused by clouds in optical imagery (Carrasco et al., 2019; Holben, 1986). Cloud computation platforms, such as Google Earth Engine (Gorelick et al., 2017), provide users with tools to create composite images aggregated over user-defined intervals (e.g. annually, monthly, bi-monthly) and for user-defined properties (e.g. raw bands, spectral indices) and with user-defined functions (e.g. median, maximum, mean).

Seasonal composite images of Sentinel-2 Multi-Spectral Instrument (MSI) Level 2-A data (Drusch et al., 2012) were created using Google Earth Engine, with cloud masking based on the Sentinel-2 Cloud Probability layer, s2cloudless (Skakun et al., 2022), and snow masking based on the quality assurance (QA) attributes performed. Images representing median surface reflectance were aggregated over four composite periods: 1 December 2020–31 March 2021, 1 April 2021–30 June 2021, 1 July 2021–31 September 2021 and 1 October 2021–31 January 2022. Periods 1 and 4 were extended by a month into the previous and following years to reduce cloud effects. Seasonal composites were used as they capture the variability in vegetation phenology throughout the year, which aids in the separation of the different land cover classes (Carrasco et al., 2019), and for the UK, aggregation of Sentinel-2 data for four seasons provides data for all four seasons in over 99.9 % of pixels (see SI). Context layers including slope, aspect, elevation distance to coast, distance to building, distance to road and distance to freshwater; in addition, a foreshore mask, a tidal water mask and a forest mask (GB only) were integrated with the composite period satellite imagery (see Table 1 for details of the context data layers, which varied slightly between GB and NI). The addition of context layers reduces spectral confusion between different classes with similar spectral characteristics. The seasonal composites, with the added context data, were then classified.

LCM2021 is based on the 21-class nomenclature presented in Table 2. The 21 land cover classes are based on UK Broad Habitat definitions (Jackson, 2000) and are designed to cover the range of habitats found in the UK that can be reliably mapped from satellites. Detailed descriptions of the classes are given in Appendix A. Production of the classifications is split into two stages, first developing the core training areas (Sect. 3.2.1) and then developing the classification process (Sect. 3.2.2).

Classifications for all tiles were compiled into a full UK spatial coverage at 10 m pixel resolution. This produced a two-band image. Band one is the most likely land cover; band two is the probability associated with this land cover but rescaled into an integer over the interval 0 to 100. Rescaling to an integer enables classification results to be stored in 8-bit, thereby reducing data size without degrading information. The 10 m raster is the precursor for all derived products.

The ingestion into the spatial framework involved determining the majority (modal) class for each polygon. Separate GB and NI data sets were created to accommodate the different map projections. Figure 3 shows the extents of the 32 composites used to achieve complete coverage of the UK. The approximate 100×100 km tile size, based on a modified version of the Ordnance Survey 100 km tile grid, was chosen as this provides a manageable size for processing. Some tiles such as those encompassing the Western Isles, Orkney and Shetland, and Cornwall and the Scilly Isles are intentionally enlarged to avoid a sparsity of training data due to the extensive presence of sea in these tiles. Occasionally, where tile extents are modified, overlap between adjacent tiles does occur.

Once the GB and NI classification mosaics were complete, a series of minor knowledge-based corrections was applied. These included the reclassification of misclassified arable pixels to improved grassland in urban green-space areas (as denoted by the OS Open Greenspace data set) and of coastal classes misclassified inland using a coastal mask.

The LCM2021 class was extracted for each of the validation points. From these data, confusion matrices were plotted for the 21 target classes and the 10 aggregate classes used for LCM2021.

The 25 m rasterised polygon version of LCM2021 (Marston et al., 2022e, f) was validated using 35 182 points distributed across the UK (Table 3). The results are summarised in a confusion matrix, which shows how reference points for each of the classes were classified. Ideally, all the points would fall along the main diagonal, showing complete agreement between the reference data and the classification. Table 3 shows that LCM2021 has an overall accuracy of 82.6 %, with the accuracy of individual classes varying. The results of the validation are shown in a confusion matrix (Table 3), with the reference data in the columns and the classification data in the rows. The confusion matrix shows the level of agreement between the classification and the reference data, as well as the areas of disagreement or confusion. The accuracy varies with class, with the producer's accuracy ranging between high and low values of 93.9 % (saltmarsh) and 35.4 % (heather grassland) and the user's accuracy varying between 96.1 % (arable) and 42.6 % (heather grassland). For the products that use the 10 aggregate classes (see Sect. 5 for more details about the aggregate-class products), the validation suggests an overall accuracy of 86.5 % (Table 4).

The final LCM2021 product shows the expected distribution of classes across the UK (Fig. 4). At the scale shown in Fig. 4, the differences between the grassland of the west and the arable areas in the east are clear, as are the uplands in Wales and Scotland, with London, the UK's largest urban area, clearly visible.

One of the uses of LCM2021 is to produce country-level statistics (Table 5), although land cover statistics can also be produced for other types of spatial units, such as river or lake catchments or national parks and protected areas.