High-resolution land-use land-cover change data for regional climate modelling applications over Europe-Part 2: Historical and future changes

Anthropogenic land-use and land cover change (LULCC) is a major driver of environmental changes. The biophysical impacts of these changes on the regional climate in Europe are currently extensively investigated within the WCRP CORDEX Flagship Pilot Study (FPS) LUCAS "Land Use and Climate Across Scales" using an ensemble of different Regional Climate Models (RCMs) coupled with diverse Land Surface Models (LSMs). In order to investigate the impact of realistic LULCC on past and future climates, high-resolution datasets with observed LULCC and projected future LULCC 5 scenarios are required as input for the RCM-LSM simulations. To account for these needs, we generated the LUCAS LUC Version 1.0 at 0.1° resolution for Europe (Hoffmann et al., 2021c, b). The plant functional type distribution for the year 2015 (i.e. LANDMATE PFT dataset) is derived from the European Space Agency Climate Change Initiative Land Cover (ESA-CCI LC) dataset. Details about the conversion method based on a cross-walking procedure and the evaluation of the LANDMATE PFT dataset are given in the companion paper by Reinhart et al. (submitted). Subsequently, we applied the land-use change 10 information from the Land-Use Harmonization 2 (LUH2) dataset, provided at 0.25◦ resolution as input for CMIP6 experiments, to derive realistic LULC distribution at high spatial resolution and at annual timesteps from 1950 to 2100. In order to convert land use and land management change information from LUH2 into changes in the PFT distribution, we developed a Land Use Translator (LUT) specific to the needs of RCMs. The annual PFT maps for Europe for the period 1950 to 2015 are derived from the historical LUH2 dataset by applying the LUT backward from 2015 to 1950. Historical changes in the forest 15 type changes are considered using an additional European forest species dataset. The historical changes in the PFT distribution of LUCAS LUC follow closely the land use changes given by LUH2 but differ in some regions compared to remotely-sensed 1 https://doi.org/10.5194/essd-2021-252 O pe n A cc es s Earth System Science Data D icu ssio n s Preprint. Discussion started: 10 August 2021 c © Author(s) 2021. CC BY 4.0 License.

species are zero in the dataset. Consequently, no additional information for the time evolution is available for the deciduous coniferous PFT. The allocation of the tree species to the remaining three tree PFTs provided in table 2.   Table 2. Allocation of the tree species data (McGrath et al., 2015;Naudts et al., 2016) to the tree PFTs. Please note that Naudts et al. (2016) distinguished the needleleaf species in temperate and boreal, even if they are the same species.

Translating land use changes to PFT changes
In order to convert the land use change information given by LUH2 into PFT changes, an algorithm with a fixed set of transition 165 rules is developed (Tables 4 and 5). In a first step LUH2 classes are grouped into main land use classes, denoted here as LUT classes (Table 3)  The transition rules are defined to ensure that the changes in cropland are as close to the LUH2 changes as possible. In contrast to other LUTs, urban transitions are included in the LUT. Following the recommendations by Ma et al. (2020) and Hurtt et al. (2020), in the LUT natural vegetation is only cleared and converted into grassland for land-use class transitions to pasture, while it remains for transitions from non-forested vegetation to rangeland. An exception to this general rule is the transition from forest to rangeland when the land will be used for livestock grazing. Since the vegetation fractions differ between the LANDMATE PFT map, used as the basemap for LUCAS LUC, and LUH2 (e.g. the spatial distribution of forest fraction), the rules are designed to be flexible. In order to ensure that crop and urban 185 changes are as close as possible to the changes provided by LUH2, transition to crops are not as strict regarding the treatment of the PFTs that occupied the grid cell previously. For example, during the transition of forest to cropland (FOR2CRO in table   4) the LUT checks if enough forest is available. If this is not the case shrub PFTs are reduced and in a subsequent step also the grass PFTs, given that the sum of forest and shrub PFTs is still smaller than the transitions. The reduction of a PFT group is done until its fraction is zero.

190
For each transition in tables 4 and 5, the relative PFT-fractions remain constant within each PFT group (Table 1) are presented and discussed as these regions are representative to illustrate the strengths and weaknesses of the high-resolution LULCC reconstruction.

Historical PFT time series based on ESA-CCI LC PFTs
Together with the high-resolution land cover maps, the ESA-CCI provides a dedicated user tool to re-project, re-sample the LULC maps and to translate the LULC classes into model specific PFTs. During the re-sampling from the native ∼300 m 235 horizontal resolution, the LULC class fractions are automatically preserved as fractions per re-sampled grid cell. The user tool provides a generic translation table but also gives the possibility to include user-defined translations. Further, the involvement of climate data within the translation process is possible to a limited extent. In order to prepare the ESA-CCI-based PFT maps for the present comparison, the generic table provided by ESA is used under consideration of the modifications introduced by Poulter et al. (2015). In addition to adjustments of the LULC class translation, an urban-PFT is added (Table A3). The

Historical PFT time series based on MODIS
The Collection 6 Terra and Aqua combined Moderate Resolution Imaging Spectroradiometer (MODIS) land cover datasets (C6 MCD12Q1) provide ready to use PFT maps as one of their 13 science data sets. For C6 MCD12Q1, several processing 245 steps were refined to eliminate known issues from the MODIS Collection 5 data sets, such as the excessively high interannual variability (Abercrombie and Friedl, 2015). Additional information on the MODIS data processing can be found at https://lpdaac.usgs.gov/products/mcd12q1v006/. The annual maps are available globally from 2001-2018 in ∼500 m horizontal resolution. The 12 MODIS PFTs follow the PFT definition that was developed for the National Center for Atmospheric Research land surface model (NCAR LSM) (Bonan et al., 2002). Table A2 shows the 12 PFTs including their description.

250
The 13 science datasets are provided having six different LULC classifications, including the PFT classification. All LULC classification are generated through employment of a supervised classification algorithm (Sulla-Menashe and Friedl, 2018).
For the comparison with the LUCAS LUC dataset, the MODIS PFT maps are aggregated to 0.1 • horizontal resolution.

Cropland
Cropland changes in LUCAS LUC correspond well with the changes in LUH2, with some exceptions (Fig. 3a

Grassland
While LUH2 shows a decrease in grassland (i.e. land use class pasture) over the Iberian Peninsula and Poland (Fig. 3h), LUCAS

275
LUC shows an increase in grassland (Fig. 3g). The reduction in grassland in LUH2 is mainly driven by conversion of nonforested vegetation to pasture, which compensates for the increase in pasture converted from cropland. However, in LUCAS LUC the non-forested vegetation that can be converted into grassland is shrubland (Tables 4 and 5), which is sparsely present in Poland and the Iberian Peninsula (not shown). This limits to what extent the LUT can increase grassland by decreasing shrubs.
Consequently, other transitions (e.g., from cropland to pasture) dominate the land cover change signal for grassland. While The changes in grassland are quite small in ESA POULTER except for Southern Russia and Northwestern Kazakhstan, where a dual pattern of decrease and increase can be seen (Fig. 3i). The decrease in grassland in this region results from a conversion into cropland (3c), which LUH2 does not capture.
The time series of grassland changes for the three PRUDENCE regions show substantial differences between the datasets ( Fig. 4d-f). Since LUH2 does not provide grassland cover, pasture area changes are plotted instead. This is likely the reason 290 why LUCAS LUC changes divert more strongly from LUH2 when changes in grassland are compared.

Forest
Forest changes in LUCAS LUC match the LUH2 changes closely (Fig. 3d, cover increased due to forest growth instead of active afforestation (Potapov et al., 2015), which seems to be captured by the satellite-based ESA POULTER dataset but not by LUH2.
Forest cover increases in both LUH2 and LUCAS LUC in the PRUDENCE regions EA, ME, and MD ( Fig. 4a-c). In EA, ESA POULTER also shows a increase in forest cover but with a smaller magnitude while MODIS forest cover shows both 310 increase and decreases. The difference between LUH2 and LUCAS LUC compared to the two satellite-based datasets are more substantial in ME. Here, ESA POULTER shows a strong decrease while MODIS shows both strong increases and decreases for some years. Such year-to-year variations seem to be unrealistic. Hence, a detailed investigation of the MODIS dataset for this region is needed. In the IP region, LUCAS LUC, LUH2, and MODIS show an increase in forest cover while ESA POULTER shows a decrease.

Cropland
Cropland decrease is even more widespread for most parts of Europe when starting from 1950 (Fig. 6b) instead of 1992 (Fig.   3b). Especially, in Mid-Europe a steep decline in cropland cover is visible from the 1950s to the 1970s (Fig. 12c). On the other hand, increase in Northern Africa, Iran, and along the Nile Delta are larger for the longer time period. In addition, the areas with increasing cropland cover in Southern Russia and Eastern Ukraine expanded. Following the results of Thebo et al. (2014), irrigated cropland also appears increasingly around cities (e.g., around Paris in France, Casablanca in Morocco, Tripoli in Libya).

Grassland
The grassland cover increases in many of the east European countries from 1950 to 2015 with the exception of Estonia,

350
Slovenia, and the Czech Republic (Fig. 6d), which show a decrease in grassland. In addition, Portugal and Turkey show increase in grassland. Decreasing grassland cover is found in Central and Southern Europe as well as in Scotland and parts of Russia. In contrast to the shorter period from 1992 to 2015, the grassland increases in Northern Germany, Northern France, England, and Ireland between 1950 and 2015.

355
The expansion in forest cover is more pronounced for the period 1950 to 2015 (Fig. 6e) than for the period 1992 to 2015 (Fig. 3). Especially in the Alps, Balkan, Caucasus, Scotland, Estonia, and Lithuania forest cover increases substantially. In addition, for the longer time period forest cover increase is found in Sweden and Finland. An increase in forest cover in

Urban
The urban fraction increase for the period 1950 to 2015 (Fig. 6e) is larger and more wide-spread compared to the signal from 1992 to 2015 (Fig. 3j). A strong urbanization signal is visible for Madrid, Paris, and Moscow. In addition, a larger scale increase 365 is found in England, Northern Italy, Benelux region, Western Germany, Poland, and Slovenia.

Future LULC
Future changes in land cover fractions between 2015 and 2100 for the eight different scenarios (Table A1)

Cropland
While cropland is found to decrease in the historical period from 1950 to 2015 over most of Europe (Fig. 6a), only for the SSP3/RCP7.0 scenario continent-wide decrease is projected until 2100 (Fig. 7h) The temporal evolution of aggregated cropland area shows that the changes are not steady for all scenarios (Fig. 12). For instance, the cropland area for Europe and in particular for the ME region shows rapid increase from 2050 onwards in the 380 SSP5/RCP3.4OS while staying rather constant before (Fig. 12a,c). The two SSP1 scenarios diverge in their evolution around 2025 for Europe and the IP and ME regions but rather converge at the end of the century (Fig. 12a,b,c). While showing a slight decrease in cropland cover aggregated for Europe, the SSP5/RCP8.5 cropland area stays almost constant in the three PRUDENCE regions.
Whereas the SSP1-based scenarios show small changes, the SSP3/RCP7.0 scenario projects large changes predominantly in the Middle East (Urmia Basin) and Turkey where a strong increase in irrigated cropland is expected.
The strongest changes for Europe are projected by the SSP5/RCP3.4OS scenario which shows a continent-wide increase in 390 irrigated cropland (Fig. 8d) and of cropland in general (Fig. 7d). Exceptions like regions along the Po river in Italy and along the Euphrates and Tigris in Iraq show a decrease of irrigated cropland for this scenario. In contrast to the continent-wide increase in irrigated cropland in the scenario SSP5/RCP3.4OS, the SSP4-based scenarios with the pathways RCP3.4 (Fig. 8c) and RCP6.0 (Fig. 8f)

Grassland
The SSP1-based scenarios as well as SSP4/RCP3.4 and SSP5/RCP3.4OS (Fig. 9a-d) show a strong decrease in grassland cover 400 for most of Europe. For SSP1/RCP1.9 and SSP1/RCP2.6 this is due to the conversion of grassland to forest 10a,b) while for SSP4/RCP3.4 and SSP5/RCP3.4OS grassland is mainly converted to cropland (Fig. 9c,d). Similarly, the increase in grassland in Southern Russia in the SSP4/RCP3.4 scenario is due to the conversion from cropland to grassland. The decreases are not as strong in the SSP2/RCP4.5 and SSP4/RCP6.0 scenarios where also large regions with increase are visible (Fig. 9e,f). The SSP3/RCP7.0 scenario shows a number of regions with a large grassland cover increase such as Spain, Germany, Norway, 405 the Alps, and Carpathian Mountains (Fig. 9g). The increase in Norway and in the Alps is compensated by deforestation (Fig.   10g). Almost no changes in grassland cover are visible for the SSP5/RCP8.5 scenario over Europe (Fig. 9h). However, a large increase is found in Northern Africa and the Middle East, which is compensated by a decrease in cropland (Fig. 7h). The block-like structures that are found for the cropland changes are also visible for the grassland changes.
Except for SSP5/RCP3.4OS, which shows a strong abrupt decrease from 2050 to 2055, increases and decreases in grassland 410 cover over Europe are steady from 2015 onwards (Fig. 13a). The abrupt change in SSP5/RCP3.4OS is even more pronounced in the ME and EA regions (Fig. 13c,d). In the latter region, also the SSP4/RCP3.4 scenario shows a steep decline in grassland cover after 2050. While grassland cover strongly increases in one scenario (Fig. 13b) in the IP region grassland cover either stays almost constant over time or decreases in the other two regions.

415
A strong forest cover increase is found for SSP1/RCP1.9, SSP1/RCP2.6 ( Fig. 10a,b) and to a lesser extent for SSP2/RCP45 ( Fig. 10a,b). These are the scenarios for which the added tree cover fraction files are used because the original LUH2 dataset underestimates the strong afforestation signal (Sect. 2.3.2). In the SSP1-based scenarios, the largest increase is found in Ireland, England, Northern France, and in Russia near the border to Ukraine. Also, a widespread increase is visible for most countries in Central and Eastern Europe, Southern Sweden and Finland, and Northern Spain. The forest increase is mainly compensated 420 by a decrease in grassland ( Fig. 9 and shrubland (not shown) and to a lesser extent by declining cropland (Fig. 7). The only region with substantial forest reduction is Western Russia. This decrease is also visible in the SSP4/RCP3.4, SSP5/RCP3.4OS, and SSP2/RCP4.5 scenarios (Fig. 10c-e). As for cropland and grassland, SSP5/RCP8.5 shows only small forest cover changes The aggregated forest cover changes for Europe show the steep increase from 2016 onwards (Fig. 14a). In the ME region, the forest cover increase levels off around 2050 while the IP and EA regions show steady increases 14b-d). Especially in the ME and EA regions, the magnitude of the increase is orders of magnitude larger than for the changes in the historical period. In contrast, the afforestation in the SSP2/RCP4.5 scenario starts in 2050 and continues until 2100 with a magnitude comparable to historical changes. A substantial deforestation signal in the SSP2/RCP45 is visible in the IP and EA region from 2050 onwards.

430
In contrast to the historical period, all scenarios show both decreases and increases in urban fraction between 2015 and 2100 ( Fig. 10). Urban changes are largely driven by the SSP scenarios (i.e. population dynamics) resulting in almost identical changes in the LUCAS LUC dataset in scenarios based on the same SSP scenario. A widespread urbanization signal can be found in the SSP5-based scenarios for Europe except for Eastern Europe, which shows a decrease in urban fraction (Fig.   10d,h). The increase in urban fractions is particularly strong in Great Britain. The SSP1-based scenarios show a increase with 435 a smaller magnitude in West-and South Europe, respectively, as well as Scandinavia and a decrease in the eastern European countries including some parts of Germany (Fig. 11a,b). Based on the SSP4/RCP6.0 scenario, only a few urban areas in Spain, France, Italy, England, and the Czech Republic exhibit an increase in urban fraction. In East-and Central Europe urban fraction decreases, with Germany experiencing the largest decrease. The SSP3/RCP7.0 scenario is the only scenario which does not show a decrease in urban fraction in Russia. Instead, it shows decreases in Western and Central European countries.

440
The time series analysis of the aggregated changes for Europe shows that all scenarios project an increase in urbanization until 2050 (Fig. 15a). Thereafter, only in the SSP5-based scenarios the total urban area increases further until 2100. For the other scenarios, urban fraction remains constant or declines. There are however regional differences. In the EA region, all scenarios show a peak in urbanization until 2050 and a decreases until 2100 with the exception of SSP3/RCP7.0, where the peak is already reached by 2040 (Fig. 15d). The IP and ME regions show a similar temporal evolution of urban areas (Fig.   445 15b,c). However, the decrease is stronger in ME. Here, the total urban area in 2100 is even smaller than the total area in 1950 in the SSP3/RCP7.0 scenario.

Discussion
The newly generated LULCC dataset LUCAS LUC was tailored towards the requirements of future downscaling experiments within the FPS LUCAS and EURO-CORDEX. The need for high-resolution land cover input is met by employing the ESA-450 CCI LC dataset, which has a ∼300 m grid, as a basemap for the year 2015. Since most of the state-of-the-art LSM employ a PFT land cover classification the ESA-CCI LC was converted into PFTs. This step also helps dealing with mixed ESA-CCI LC land cover classes, which can be conveniently converted into classes with similar properties. The uncertainty introduced by the cross-walking procedure was investigated by Georgievski and Hagemann (2019). Their results show that this uncertainty can be of high relevance, comparable to the uncertainties in historical changes for the global forest cover. For the LUCAS LUC dataset a more detailed cross-walking procedure was applied taking into account the dependency of the composition of ESA-CCI LC land cover classes on the climatic conditions (Reinhart et al., submitted). This likely introduces more uncertainty, e.g., due to the climate dataset used to compute the HLZs. For instance, in mountainous regions, it is important to have a high-resolution climate dataset in order to capture the highly varying climate zones. In addition, the boundaries between different HLZs can be shifted if the datasets show differences in temperature and rainfall amounts. Such as management driven changes e.g. due to wood harvesting. Ceccherini et al. (2020) showed that averaged over Europe such 485 changes are small compared to forest harvesting but can be larger in regions with frequent forest fires (e.g., Portugal). Also the different definition of forest in LUH2, which is based on a biomass density threshold, and ESA-CCI LC, based which is based on tree cover, could have a substantial impact on the forest transitions. A detailed analysis of the ESA POULTER time series is needed to investigate if these processes caused the discrepancies in forest changes LUH2/LUCAS LUC and ESA-CCI LC.
Another possibility is the already mentioned uncertainty originating from the CWP. However, the impact of the CWP on the 490 computed land cover changes has not been analyzed so far. Notable differences between LUCAS LUC and ESA POULTER are also found in the urban land cover changes. Here, ESA-CCI LC seems to largely overestimate the urbanization in Europe the climate (Valmassoi et al., 2019). Hence, there is a need for high-resolution European-wide dataset with information on the 510 distribution and the development of irrigation methods.
Future land cover changes are even larger than the historical changes for some of the available scenarios. Hence, substantial policy changes would be necessary to reach the amount of land cover conversions in the densely populated Europe, where land ownership is both public and private. In addition, there are large regional differences. The two SSP1-based scenarios, which are the low-end scenarios, show a strong afforestation signal compensated by a decrease in grass-and shrubland. Also, 515 noticeable changes in cropland (both increase and decrease) are projected for these scenarios. Hence, the LULCC induced climate change signal might be comparable to the greenhouse-gas induced signal in regions with large LULCC for some seasons (Hirsch et al., 2018). For instance, Davin et al. (2019) showed that for a extreme afforestation scenario temperature shows the smallest land cover changes except for urbanization, where it has the largest signal together with the other SSP5based scenario (i.e. SSP5/RCP3.4OS). Therefore, it might be harder to detect LULCC induced regional climate changes giving the strong greenhouse gas forcing. In contrast, the SSP3/RCP7.0 scenario, which has also a large greenhouse gas forcing, shows large-scale cropland decreases and regions with deforestation (e.g., Alps and Scandinavia) as well as afforestation (e.g.,

525
The large block-like features appearing in the cropland and grassland change signals in all scenarios might be attributed to the harmonization process within the LUH2 workflow. Annual changes in cropland, grazing land, and urban areas are computed and aggregated to a 2 • grid and subsequently disaggregated to the final 0.25 • grid (Hurtt et al., 2020). It is therefore likely that the disaggregation step did not fully dissolve the grid structure of the coarse 2 • grid. For GCMs or ESMs with a typical resolution of around 1 • this might not have caused any issues. However, for RCMs with a typical resolution of about 530 0.1 • , for which the LUCAS LUC dataset has been created, the impact of such structures in the LULCC needs to be carefully investigated.
Within the LANDMATE project, a short documentation summarizes the technical information on the LANDMATE PFT and 540 LUCAS LUC dataset.

Conclusions
The need of the RCM community for a high-resolution LULCC dataset is met using high-resolution PFT maps based on the ESA-CCI LC dataset and land use change information from the LUH2 dataset that were translated into PFT changes using a newly developed LUT. The resulting LUCAS LUC dataset is tailored towards RCM requirements. Urbanization, which is 545 mostly discarded by LUTs, is included as well as changes in irrigated cropland. For the historical period, also changes in the broad-/needleleaf forest ratio are considered employing an additional forest type dataset by McGrath et al. (2015). impact of LULCC on the regional climate change signals can be investigated. For most of Europe, past and future trends in 550 cropland, forest, and urban areas in LUCAS LUC are consistent with the LUH2 dataset albeit with a slight underestimation of the magnitude. A comparison with two satellite-based PFT datasets revealed substantial differences in the trend of some land cover classes. However, the differences between the ESA-CCI LC based dataset and the MODIS-based dataset are also quite large showing the uncertainty related to the approaches employed to estimate LULCC.
The future LULCC for the eight SSP/RCP scenarios show substantial changes that can exceed the observed historical 555 LULCC in Europe. Hence, the regional climate change signals, simulated by RCMs, are likely to be affected by these changes and should, therefore, be considered in upcoming downscaling experiments. Especially when downscaling projections for the low-end scenarios (i.e. SSP1/RCP1.9 and SSP1/RCP2.6), which show a strong afforestation signal, the biogeophysical effect of LULCC is expected to be of the order of the greenhouse-gas induced effects in some regions (Hirsch et al., 2018). In contrast, for the high-end scenario SSP5/RCP8.5 LULCC in Europe are small compared to the other scenarios except for urbanization.

560
While the current dataset is provided on a 0.1 • grid for Europe in order to be suited for the EURO-CORDEX EUR-11 grid, the method could be applied to generate data at even higher resolution, e.g., needed for convective permitting RCM experiments . However, a downscaling of the coarse land-use changes provided by LUH2 would be necessary, e.g. by using a spatial disaggregation model (Chen et al., 2020).
The LUCAS LUC dataset can also be prepared for other CORDEX regions because most of the input data is provided 565 globally . However, the quality of conversion of ESA-CCI LC classes into PFTs depends to some extent on the availability of high-resolution climate data, needed for the Holdridge-based cross-walking procedure. In addition, data on forest type conversion for the historical period might not be available for other regions.  Table A1. Specfication of the land-use change scenarios provided by LUH2 (Hurtt et al., 2020) and the assumptions for land-use and land cover developments in the different scenarios (Hurtt et al., 2020;Popp et al., 2017;Riahi et al., 2017;van Vuuren et al., 2011)   (http://www.uerra.eu) and the Copernicus Climate Change Service, and the data providers in the ECA&D project (https://www.ecad.eu).
We thank the European Space Agency (ESA) for making the Land cover products publicly available. Special thanks go to the FPS LUCAS partners for providing useful comments in order to improve the dataset.