Interactive comment on “ A high-resolution and harmonized model approach for reconstructing and analyzing historic land changes in Europe ”

Overall, the manuscript is of good quality, both in terms of the robustness and relevance of the work reported, and of the structure, clarity and writing of the paper itself. For the record and clarity of the discussion, I will here rediscuss some issues raised in the preliminary screening, and whenever useful the author’s responses. I don’t have major corrections to request, but I do have some remarks concerning (i) some unclear aspects of the method, (ii) some of the assumptions used for the reconstructions of past land cover changes, and (iii) remarks related to the validation. ****************** Unclear aspects of the methods: ******************

various land change types. This paper investigates if the combination of different data sources, more detailed modeling techniques and the integration of land conversion types allow us to create accurate, high resolution historic land change data for Europe suited for the needs of GHG and climate assessments. We validated our reconstruction with historic aerial photographs from 1950 and 1990 for 73 sample sites across Europe 10 and compared it with other land reconstructions like Klein Goldewijk et al. (2010Goldewijk et al. ( , 2011, Ramankutty and Foley (1999), Pongratz et al. (2008) and Hurtt et al. (2006). The results indicate that almost 700 000 km 2 (15.5 %) of land cover in Europe changes over the period 1950 to 2010, an area similar to France. In Southern Europe the relative amount was almost 3.5 % higher than average (19 %). Based on the results the specific types 15 of conversion, hot-spots of change and their relation to political decisions and socioeconomic transitions were studied. The analysis indicate that the main drivers of land change over the studied period were urbanization, the reforestation program after the timber shortage since the Second World War, the fall of the Iron Curtain, Common Agricultural Policy and accompanying afforestation actions of the EU. Compared to 20 existing land cover reconstructions, the new method takes stock of the harmonization of different datasets by achieving a high spatial resolution and regional detail with a full coverage of different land categories. These characteristic allow the data to be used to support and improve ongoing GHG inventories and climate research.

Introduction
Currently, up to 30 % of the global carbon emission is estimated to originate from human induced land use and land changes (Brovkin et al., 2004;Prentice et al., 2001). This is the case since about 1960. For earlier decades (before 1960) the contribution of land change emissions to total emissions was even higher because of lower fossil 5 fuel emissions (Brovkin et al., 2004;Houghton and Hackler, 2001;House and Prentice, 2002;Prentice et al., 2001). However, a large uncertainty in those assessments is present due to the varying anthropogenic and natural land change processes going on in parallel (Houghton et al., 2012). A main shortcoming in making an assessment of the consequences of land cover change for climate and greenhouse gas (GHG) bal- 10 ances is the lack of spatially explicit and thematic complete historic high resolution land cover change data and its conversion types that feed into these models. The historic information on land cover is needed for GHG assessments, since every current land cover type contains also the legacy of previous land cover types, such as soil carbon from residues (Houghton et al., 2012;Poeplau et al., 2011). The consideration of this appear in the same way. Since the EU-reporting is on an advanced level for GHG emissions, there is a growing demand for high-resolution, harmonized and spatially explicit land change products, to improve our understanding of the amount and extent of human induced land change processes (global and regional) (Ciais et al., 2011;Gaillard et al., 2010;Schulze et al., 2010).
At the same time, more detailed historic land use reconstructions based on actual data (such as historic maps and remote sensing) have been gathered for local case studies or small regions (e.g. Antrop, 1993;Čarni et al., 1998;Bicik et al., 2001;Petit and Lambin, 2002;Van Eetvelde andAntrop, 2004, 2009;Kuemmerle et al., 2006;Orczewska, 2009). Such studies are able to describe land conversion patterns at a fine 20 spatial, temporal and thematic detail and on the level where human-induced change processes take place. However, they are difficult to compare and combine with each other, especially cross border. On a continental level their synergistic use will remain limited, due to a lack of an accepted and commonly used reporting scheme for land use classes, including standardized definitions and harmonization levels but also as a 25 result of their limited spatial coverage and focus on regions that are often known for large historic changes.
Many land transitions in Europe have taken place affecting the land use pattern due to changes in farming or management systems (e.g. fallow land, abandoned, Introduction Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | reactivated and reforested land). These changes follow fine scale variability in environmental conditions, socio-ecological factors, such as demographic change, accessibility and cultural factors (Kuemmerle et al., 2009;Mander and Kuuba, 2004;Pinto-Correia and Vos, 2004;Prishchepov et al., 2012). Thus, they require high resolution data sets to observe and study these local heterogeneous processes. These changes 5 may have large consequences for GHG emissions and climate variables (e.g. albedo) together with European specific determinants that are crucial (e.g. management practices; Houghton et al., 2012). Based on the shortcomings of current land cover reconstructions and the needs of GHG and climate assessments, the objective of this study is to investigate if the combi-10 nation of different and new data sources, detailed region specific modelling techniques and the consideration of multiple land cover types allows us to reconstruct historic land change for Europe at a high spatial resolution for the period 1950-2010. Validation with independent data and comparison with existing land cover reconstructions is used to evaluate the research objective. 15 After presenting the methods employed to reconstruct historic land changes, this paper will analyse the regional land change hotspots over the 1950-2010 period and its major conversion types at the continental scale. The results will be compared with existing global scale historic land change databases of Klein Goldewijk et al. (2010Goldewijk et al. ( , 2011, Ramankutty and Foley (1999), Pongratz et al. (2008) and Hurtt et al. (2006), 20 henceforth referred to as Goldewijk, Ramankutty, Pongratz and Hurtt, respectively . Finally, the validation and performance assessment with independent historic highresolution data (aerial photographs from 1950 and 1990) will outline uncertainties in our allocation of land cover and its changes on a pixel level. Introduction

Overview of the method
This study uses a land change quantity and land change allocation approach. The approach simulates land conversions on the basis of land change pressures, resulting from area statistics on country level for each land category (land change quantity), and 5 allocates this information based on data that are able to indicate pixels of this land category where these changes are likely to happen (land change allocation). The preparation of the land change quantity data is explained in Sect. 2.2, the pre-processing of data for the land change allocation procedure in Sect. 2.3. The processing steps and the usage of the two data stacks are described in Sect. 2.4. To validate the perfor-10 mance of our approach, the results were compared with high-resolution aerial photos (1950 and 1990) obtained for regional case studies. This is presented in Sect. 2.5. The resulting data set of this investigation is called HIstoric Land Dynamics Assessment (HILDA).

Data sets and preparation
Focus of this work will be on EU-27 plus Switzerland, since the data for these countries are quite good, even on regional scales (spatially, thematically and temporally While remote sensing data could provide spatially explicit land cover and use information and its changes, it temporally covers only a relatively small proportion of the investigated time frame (1990s-2010 vs. 1950-2010). Some statistics instead span longer terms and some even the complete period. However, they are often just available as aggregated numbers on country scale and lack the information on spatial allocation 5 within these administrative boundaries (Verburg et al., 2011).
For recent years (from 1990 onwards) the data availability and quality (temporal, spatial and thematic) is appropriate to cover major land changes in Europe. Remote sensing data can be used for the spatial allocation of land cover classes and for crosscalibration of temporal land change trends with spatially coarse national statistics. 10 Thus, their period 1990-2010 is used to inter-calibrate the existing data sources and extrapolate the change trends using the less detailed data for the historical periods back to 1950.
The various data do not necessarily follow the same nomenclature and class definitions have to be harmonized and aggregated to make them comparable. Besides 15 the detailed analysis of existing legends (Herold and DiGregorio, 2012), the main idea was to aggregate to broad land categories in order to avoid definitional conflicts. In line with GHG accounting and climate modelling requirements five suitable land categories were defined for the modelling: -Settlements (incl. green urban areas),

20
-Cropland (incl. orchards and agro-forestry), These classes and their definitions cover 100 % of the land area in Europe and are based on the Intergovernmental Panel on Climate Change (IPCC) categories (IPCC, 2003). However, due to the lack of sufficient land information for the last 60 yr of the wetland, it was integrated in the grassland category.
The Land Cover Classification System (LCCS) (DiGrigorio and Jansen, 2000) was 5 used to harmonize all existing data sets on the five IPCC classes. An overview of the class accounting and parameter description by LCCS is given in Supplement A. The advantage of this procedure is the objective class accounting using describable and comparable class features, instead of subjective appraisals. 10 The finest scale for a cross-comparison along the data sets was the country scale, so all harmonized data were brought on that level for the analysis of land change trends. Spatially explicit data were geo-referenced on an equal area projection (Lambert Equal Area) to compare areas. Despite the harmonization process, the data sources could still differ in the overall amount of land cover area per class, e.g. due to the relatively 15 coarse spatial resolution of GlobCorine (300 m) and UMD (1 km) or due to the fixed thematic boundary of some statistical classes. It was also recognized that in the Forest Resource Assessment (FRA) reports for Mediterranean countries like Spain, shrublands were accounted in some years to forests and in other years to cropland and grassland. In these cases other data sets, for example FAOSTAT, could be used in-20 stead.

Data adjustment and analysis of land change trends
The FAO-FRA data set provides cropland and grassland back to 1946. In comparison with FAO-STAT data (back to 1961), where these two classes are separated, area relations of these two classes and their relative trends over time could be calculated for each country. This allowed the separation of the FAO-FRA cropland and grassland class before 1961.
Since settlement data were not separately reported in the statistics data (mainly included in settlement and others -FAO or other land and settlements -FRA), population 14830 Introduction For all countries and its land categories, outliers were sorted out and gaps with 5 missing data were filled. An overview of the used method per country, per class and per year is given in Supplement B. Available data, which could be used for this study, were inter-and extrapolated by the use of approximation functions that were able to describe the land change trends over the whole period. The chosen polynomial order for each class per country is also given in Supplement B.
Due to the heterogeneous data sources, the sum of all harmonized land categories may lead to varying total areas per country over time. These differences occur, if the land categories are subject to high variances in area along the used data sets at one time step. For the investigated land categories the variances were highest for grassland and lowest for settlements and forest. Reasons for these variances might be remaining 15 inhomogeneity of class definitions and inaccuracies in classification of the products itself. To correct for discrepancies between the total area per country and the sum of all land categories, the one with the highest variance, in this case grasslands, was used to match the sum of all land categories with the total area per country. 20 A simple allocation procedure was implemented to distribute the land areas within the administrative boundary to 1 km 2 pixels based on probability maps for each land category ( Fig. 1). Probability maps represent the spatially explicit likelihood of a dominating land cover. The probability maps are derived through an empirical analysis of the relations between observed land use patterns in the year 2000 and a range of 25 supposed explanatory factors conducted by Verburg et al. (2006) and Verburg and Overmars (2009)  response to the biogeophysical and socio-economic conditions. As explanatory factors Verburg and Overmars (2009) used biogeophysical factors with parameters like soil properties, precipitation, sunshine hours, altitude, slope, and socio-economic factors involving accessibility to settlements based on settlement size and population density. Logistic regressions were estimated for all land cover types and countries separately, 5 allowing different variables to explain different land cover types across the different countries. Then, the probability of finding the land cover type under the prevailing conditions was calculated for all locations on a 1 km grid. The resulting probability maps are visualized in Fig. 1. Other Land was not processed since it is treated differently in the approach than the other classes (see Sect. 2.4).

Model structure and processing
The approach processes the data in decadal time steps for each country separately. Each time step can be separated into a pre-processing phase (Fig. 2, upper box), a class-processing phase (Fig. 2, middle box) and post-processing phase (Fig. 2, lower box). 15 In the pre-processing phase it is decides which land cover map (LCM) has to be chosen. This is dependent on the time step which needs to be processed. If these time steps are 2010 or 1990 the baseline map of the year 2000 is used, otherwise the LCM of the previous time step is used.
For land allocation in the class-processing phase the model follows a process hier-20 archy. The land categories are ranked by its socio-economic value, so that settlements are calculated first, croplands second, forest third, and grasslands at last. Forest was ranked third, because its area was almost constantly increasing since 1950 according to land change quantity data (LCQ). This implies a demand for these areas. On the other hand, grassland was calculated last, since it was mainly decreasing according 25 to the LCQ data, implying a lower demand for that land. Furthermore, grassland contains pastures and natural grasslands (peatlands, highlands, etc.), so that the socioeconomic value was assumed to be lower than for the other land categories. The approach treats the other land class, which mainly consists of water, glaciers, bare soils and sandy areas, like beaches, desserts and dunes as static, and therefore it was masked from the data set. Since other land areas are small, influences from climate, tides and the meandering of rivers, were considered to be low at this spatial resolution.

5
If a class is selected for processing the next time step, the model requests information from the LCQ database on increase or decrease of the class area (Fig. 2, left vertical box). Considering a class is increasing, it masks all other classes in the LCM and selects the highest values in the relevant probability map (PM) within this mask until the right area for that class is obtained. The selected area is then converted into the according class (Fig. 2, middle box). Should the class decrease, the model masks the relevant class instead of all other classes, and picks the lowest values in the according PM equal to the LCQ area for that class. The area is then converted into unclassified area, which can be incorporated in other increasing classes later on as part of their increase mask (Fig. 2, middle box). Since the sum of all land categories is matched 15 with the total area per country (see Sect. 2.2), no unclassified pixels are left after a processed time step. All new class areas are merged (including other land) to a new time step in the post-processing phase if all classes have been processed (Fig. 2 lower box). 20 In order to check the performance, the approach was compared with other land change reconstructions available for this scale. Four relevant global models were chosen: Goldewijk, Ramankutty, Pongratz and Hurtt. Their spatial, temporal and thematic features are shown in Table 1. Since the grassland class in our approach comprises pastures and natural grasslands, the comparative assessment between these reconstructions 25 and ours was only possible for cropland. On the one hand the comparison was performed in a spatially explicit way to point out the differences of detail due to the resolution and to show similarities and discrepancies of regional hotspot patterns. It was possible to use the same class aggregation scheme for the five IPCC classes (LCCS) and for the CORINE product, since they use the same nomenclature and definitions. For this study the results were compared for 1950 and for 1990. Unfortunately, the data for 2000 were not available for all validation sites.

Land use reconstructions
The result was analysed for the period 1950-2010 ( Fig. 4) and is separately displayed for the years 2010, 1990, 1970 and 1950 The growing population of Europe within the last 60 yr (+122 Mio.) has led to the development of settlement agglomerations across the entire study area, especially in the population belt, known as the blue banana (Brunet, 1989).
Forests in Sweden increased their coverage by almost 20 % within 60 yr compared to 1950, mainly occurring between the lake Vänarn and Stockholm. In Finland the same 5 patterns occur, although more heterogeneously, for the coastal region reaching from the Saint Petersburg in Russia to the upper Gulf of Bothnia.
The Baltic States underwent a notable land transformation. The loss of cropland and the increase in forests and grassland can be determined as the main drivers for that region. 10 For the Mediterranean countries it can be concluded that the coastal areas of Italy, Spain and Southern Portugal experienced a considerable drop of cropland by simultaneous conversions into mainly grasslands and to a minor extent into forests. Especially the regions of Alentejo in Portugal and Tuscany in Italy are affected by these changes.
The forest for France increased from 109 540 km 2 (1950) to 159 540 km 2 (2010) by 15 50 000 km 2 , mainly occurring in the Provence and around Paris, which implies an increase of 45.64 % within the last 60 yr. The same conversion type occurred also in Poland, more or less spread over the whole country, reaching a forest increase of +35.14 % between 1950 and 2010. In Romania, while forests stayed almost constant, the main driver was the drop in cropland in the Transylvanian and Moldavian regions, 20 resulting in increasing grassland areas. Accumulating the land changes between every single time step, a hotspot map can be generated for the whole period (Fig. 5). The hotspot map allowed focusing just on the modelled land changes instead of the coverage, in order to analyse the spatial hotspot patterns and agglomerations of multiple land changes per pixel. This way hot 25 spots are highlighted and clustered for visualization. Moreover, it shows areas of multiple land changes that took mainly place in France, Scandinavia, the Baltic States, Czech Republic, Austria, Italy and Portugal. This could be used to calculate the overall land changes for the entire study area with varying regional amounts of land changes. Therefore, the study area was separated into four major regions: Northern Europe, Eastern Europe, Southern Europe and Western Europe (see Fig. 5 and Table 2). For the investigated period the area of affected land by land changes could be calculated as 601 154 km 2 , which is 13.79 % of the total area of all EU-27 states plus Switzerland (Table 2). If the amount of all land changes is considered (including multi-5 ple land changes) an area of 674 684 km 2 has changed, which is 15.47 % of the EU-27 plus Switzerland region. This implies that every year 0.26 % of the entire 4.36 Mio. km 2 is converted, an area similar to Northern Ireland (Fig. 5). While the amount of changes of Northern and Eastern Europe follows the total average of land changes, Western Europe was roughly 2 % below average. Contrary, Southern Europe was roughly 3.5 % 10 above average. Figure 6 separates the relative amount of all occurred land changes per region within 1950 and 2010 into their main land conversion types. The two main land conversion types for these regions were either grassland to forest or cropland to grassland, incorporating together 63 % (Eastern Europe) to almost 85 % (Southern Europe) of land 15 change areas per region. These conversion types were followed by cropland to forest, grassland to cropland and cropland to settlement.

Comparative assessment and validation
One objective of this study was to compare and evaluate our land reconstruction results with Goldewijk, Ramankutty, Pongratz and Hurtt (see Table 1). The spatial comparison 20 is displayed in Fig. 7. Since the Hurtt product is based on the Goldewijk database and rescaled to 0.5 • it was left out for the spatial pattern analysis. Due to the fact that our approach covers grasslands (incl. pastures and natural grassland) instead of pastures, the direct comparison with the global models was only possible for croplands.  the same trend, which is likely caused by the fact that global models rely on FAOSTAT data since 1960 and before on linear model based estimates.
In order to evaluate the quality of the land cover reconstruction, a comparison with independent observation data at higher resolution was made as a means of validation. In general, by comparing the two data sets, it could be recognized that the historic land reconstruction could mainly cover the main land change trends of the Gerard et al. (2010) data set (e.g. increasing areas of settlements, reforestation, cropland decrease, etc.). The sample sites of Amsterdam and Haarlem (NL) and Grenobles (FR) 10 indicate that during the backcasting to 1950, our approach was able to reduce the amount and to keep the shape of settlement areas as determined by reference data. However, in some parts, differences remain. While the historic land change approach considered the south east to be more stable, the southern region existed already in the 1950s. The urbanization of the suburbs was well captured, although the area of Haar-15 lem (middle western part) was a bit underestimated. The example of the Carpathian Mountains in Romania demonstrates that the approach was also able to cover land changes like clear-cuts in forest areas, although the patches were difficult to capture with a 1 km resolution. The fourth sample site (Vecpiebalga, LV) was in the southern section affected by afforestation. The historic land change model was capable to re-20 construct this land conversion. However, it found the land change area in the middle of the southern section, whereas it was in the left southern section according to the reference data.
Besides the visual comparison in Fig. 9, the two products were cross-validated for each of the 73 validation sites for the time steps 1990 and 1950 by comparing the It was noticed that the agreement of the forest and grassland class was negatively influenced by one outlier. This outlier was the most northern validation site in Finland, for which the reference data set derived almost a complete coverage of forest (94 %), 5 whereas the land reconstruction approach yielded grassland coverage of 94 %. Ignoring these differences in classification between the datasets would have caused an increase of R 2 of about 0.2 for forest and grassland, leading to a final R 2 of 0.90 for forest and a R 2 of 0.61 for grassland. It should, however, be noted that these high correlation levels are largely the result of persistence in land cover: the overall distribution 10 of land cover across the test sites remained the same across the two years, especially as many of the reference sites were located in relatively stable rural areas. This persistence often led to high correspondence levels in land cover model validations (Pontius et al., 2008).
In general the validation with reference data revealed that our approach could cap- 15 ture the main land change hot spots and its conversion types correctly in many cases. Both the reference data and our approach showed an increase in urban and forest areas (mainly due to cropland and grasslands losses) and a decrease in cropland and grassland areas (due to afforestation and urbanization) between 1950 and 1990. However, detailed comparison of the maps revealed larger deviations in predicting the exact 20 location of change. The area affected by change and its change rate were smaller than those of the modelled land cover for EU-27. This was because of the sampling size and a bias towards areas containing nature reserves. Therefore, it was not possible to produce statistically reliable estimates of land cover change for larger areas (Gerard et al., 2010). 25 Nevertheless, compared with the existing global land use reconstructions, the validation showed that the presented historic land reconstruction is capable to describe land changes on a higher spatial and thematic resolution leading to a realistic representation of the landscape composition and pattern, which is of high importance for reliable assessments based on such data (Verburg et al., 2012). While our approach could provide complete thematic information on land changes within validation sites, global models could only provide information on some classes with a spatial resolution that is for some of the data as coarse as a whole reference test site.

Land reconstruction
Analysing the reconstructed land conversions of the investigated period for Europe, the main conversion types were grassland to forest, cropland to grassland, cropland to forest, grassland to cropland, and cropland to settlement (Fig. 6). Together all changes led to 674 684 km 2 (15.47 %) of changed land within the last 60 yr, an area similar to 10 France (Table 2). Although we cannot determine the proximate cause and underlying driving factors of these land changes based on the analysis in this paper, some of the locations of major land changes can be related to major political decisions. Examples include the timber shortage after the Second World War, the urbanization due to the increased population, the controlled economy in countries belonging to the Russian 15 Federation until 1990, the Common Agricultural Policy (CAP) and its accompanying afforestation actions. especially world and global cities and their agglomerations. They cover the highest density of commerce, money, industries and related human capital (Fig. 10). City clusters along the blue banana were mainly affected as well as cities likes Madrid, Berlin and Paris.

afforestation actions
The total area of forest increased by 314 177 km 2 (+25.35 % of new forest land) ( Fig. 10) since 1950. This land conversion could be seen in almost every country, with the main increase in Western and Northern Europe (Fig. 6). After the two World Wars and rigorous resource exploitation due to former land use, the European forests were 10 in a critical situation. The timber shortage was induced by the economic demand for wood products and led to several national afforestation actions (FAO, 1947(FAO, , 1948. One hotspot is Southern Scandinavia. Although Sweden and Finland always exported timber for the last few centuries, they released land reforms at the beginning of the last century, which regulated the management of their forests (Meissner, 1956). Be-15 fore these land reforms, in the 19th and beginning of the 20th century, primary forests were cut by subsistence farmers using a mixed form of management between forest, cropland and grassland. Later on, large scale forest enterprises managed the land, focusing only on wood supplies (KSLA 2009). Croplands were abandoned, resulting in fallow land and afforested by the companies with seedlings, resulting decades after the 20 last land reform in new managed forest areas. The results show this transition, taking the temporal gap of cropland and forest demand into account (Fig. 5).
In the 1990's the EEC Regulation No. 2080/92 included afforestation as forestry measure in the European Law to further decrease the deficit of European timber production. Accompanying the CAP, less productive agricultural land should be converted into for- 25 est areas to steer and optimize the production of natural goods and to support the preservation of the environment (EEC, 1992(EEC, , 2005. From 2000 to 2006, afforestation actions were stipulated by the Regulation (EC) No. 1257/1999(EEC, 1999, 2005

Cropland changes after the introduction of the Common Agricultural Policy
The CAP of the European Union came into effect in 1990. By guaranteeing farmers subsidies and a standard of living, this policy forced the reorganization of agricultural land (cropland and pastures) to be more competitive for global markets (Pinto-5 Correia and Vos, 2004). Several regions (e.g. the province Alentejo in Portugal) became unattractive due to their higher management effort and lower accessibility and were converted into other land forms within just a few decades (Pinto-Correia and Vos, 2004).
In whole Europe an area of 144 733 km 2 of cropland was converted into grassland 10 and forests since the start of the CAP (1990-2010) (Fig. 11). This is an increase by 150 % in comparison to the same period before 1990   an amount of land changes, which were 4 % above the European average (Fig. 6). 85 % of the occurred land changes in this region were due to land conversions from cropland to grassland or grassland to forest, although it cannot be distinguished whether these land changes are cropland abandonment, conversion into pastures or driven by the 20 reforestation actions of the EU.

The fall of the Iron Curtain
The same conversion effects can be seen for the Baltic States (Fig. 11) mainly since 1990, but under a different political situation. Lithuania, Latvia and Estonia were part of the Soviet Union before 1990, and carried out a plan economy, resulting in large areas 25 of cropland. After the fall of the Iron Curtain, the agricultural system could not compete with the market, so that the value of wood production became more important, resulting in afforestation areas and fallow cropland (Mander and Kuuba, 2004;Prishchepov et al., 2012).
Since the beginning of this modelling period Romania has also been led by a plan economy of the Soviet Union. The main focus was on cropland due to the Mediterranean climate, but the markets in the 1990's entailed that the supply and the produc-5 tion methods were not competitive enough to survive. Large areas in the Transylvanian and Moldavian province have been turned into fallow land (Kuemmerle et al., 2009;Mueller et al., 2009).
The main land conversion types of Eastern Europe were cropland to grassland, grassland to forest and cropland to forest (Fig. 6). Together they caused 78 % of all 10 land changes in that region since 1950. Most of these changes occurred after the fall of the Iron Curtain. The effects, before and after this event, can be seen for two of these conversion types in Fig. 11.

Comparative assessment and validation
The comparison with global models revealed differences in the spatial allocation of 15 land cover. Figure 7 illustrated this for cropland. Differences could be attributed to the various distribution methods of each model, considering different assumptions for the allocation of land cover and its changes. However, the absolute differences (Fig. 8) could also originate from different baseline data sets, from processing in a non-equal area projection (all global model results are given in WGS84), a different change data 20 basis, methods for gap filling of land change data, cross country allocation procedures and wrong assumptions for areas with poor data.
The validation with the reference data revealed that our results could capture most of the overall patterns of land change, although deviations with the observed data remain. The higher inaccuracies in the results for the grassland class can also be attributed 25 to the known problems of CORINE to differentiate between cropland and grassland (Maucha and Buettner, 2005;EEA, 2006). Since our study also combines pastures and natural grassland areas it assumes the same dynamics for both land cover types, which is in reality not the case.

Methods
Due to the combination of new and more suitable data sets for Europe as well as better and more detailed modelling techniques, the results of our approach can be used to 5 considerably improve GHG and climate assessments compared to existing methods. By the use of the presented method and available data for Europe new synergies arose, like a high spatial resolution, flexibility in processing and the consideration of a full land change balance with its land conversion types.
In comparison to other land reconstructions we have only considered a relatively 10 short time period in which we could base the national land areas on available census data and other sources. Global historic models like HYDE (Ellis et al., 2012;Klein Goldewijk et al., 2010 have reconstructed land change over much longer historic periods and are therefore relying more on assumptions about management practices and class relations to process land categories over time (e.g. population/cropland ratios 15 or livestock/pasture ratios). This is because land data are rare or often not available for their covered areas and periods (centuries to millennia) for all time steps. The higher spatial-thematic detail of our study responds to the demands by the GHG community (Ciais et al., 2011;Schulze et al., 2010) providing base maps for GHG inventories and further information about the influence of land change on emissions. As a baseline year 20 we used the year 2000, where data availability, quality and overlap along the products were best. However, the approach is flexible in using different base years if new data become available. Although European level simulations of future land change were available Verburg et al., 2010) the underlying models were not directly applicable 25 to provide backcasting. Many land change models used for simulation of future scenarios account for path-dependency in the land system evolvement and are therefore not suited for reconstructing land use history in a backward mode or deal with limitations 14844 Introduction in historic data availability. The land allocation approach used in this paper is much simpler and not path-dependent and therefore more suited for the specific purpose of this paper.
The assumption of constant probability maps for the whole modelling period might lead to limitations in the allocation approach. They are econometrically fitted based on 5 the current time relations between drivers and land use. Although many factors are considered to be quite stable in time (e.g. climate-, terrain-and soil factors), this may have been different in the past for some of them (e.g., for accessibility or population density). However, the estimation of the probability maps has been done at national scale (with country specific factors) and was widely used and tested in multiple land 10 use modelling efforts in a foresight mode (Verburg and Overmars, 2009;Verburg et al., 2008Verburg et al., , 2010. The chosen class hierarchy was most suitable for adapting the real land develop-20 ments. However, it has implications on the final result that have to be considered. The hierarchy approach requires that all territorial claims of a higher ranked class are satisfied first, which is in reality not always valid. It is rather the case that each class has dominant and less dominant conversion types (e.g., increasing settlement area is incorporating 60 % of cropland, 30 % of grassland and 10 % of forest areas). On the other 25 hand, this consideration would require knowledge about gross land changes (e.g., provided by spatially explicit information or statistics which consider such a conversion matrix), instead of net land changes (e.g. provided by statistics on an administrative basis), which was not consistently available for the investigated period.

Implications for GHG and climate models
Besides the technical improvements on spatial resolution, which enables to study more fine scale variability in land changes than before, the results include new relevant land categories for GHG assessments, such as the settlement class and other land class (including inland water). Since all land categories in the presented approach cover together thematically 100 % of the land area, it enables GHG models to take a full land change balance into account. This again affects the GHG balance. The importance of historic land changes and their effect on soil organic carbon (SOC) was pointed out by Poeplau et al. (2011). The associated uncertainties of SOC estimation on the GHG balance without sufficient land change information was addressed by Ciais et al. (2011).
Furthermore, using our approach allows relating land changes with their underlying proximity causes on an improved level of detail. This is an important advancement for GHG and climate research, since it supports the study of human activity on our climate. However, this land change reconstruction processes net land change information, instead of gross change information due to the input data. Therefore, the change rate 15 will be underestimated, since the dynamic of changes within administrative boundaries is not well captured. Schulze et al. (2010) quantified the spatially inexplicit UNFCCC gross change rate per year to be 17 800 km 2 for EU-25, whereas our results have a spatially determined yearly net change rate of 11 336 km 2 for EU-27 plus Switzerland.

20
The aim of this paper was to investigate whether the combination of different data sources, more detailed modelling techniques and the integration of land conversion types allow us to create accurate, high resolution historic land change data for Europe suited for the needs of GHG and climate assessments. By the use of multiple harmonized data sources and our modelling approach, we were able to process the The categories cover 100 % of the land area, and take a full land change balance into account. This allows the consideration of land conversion types.
The results indicate that almost 700 000 km 2 (15.5 %) of land cover in Europe has changed over the period 1950 to 2010, an area similar to France. In Southern Europe the relative amount of change was almost 3.5 % higher than this average. Based on 5 the results the specific types of conversion, hot-spots of change and their relation to political decisions and socio-economic transitions were studied. The analysis indicated that the main drivers of land change over the studied period were urbanization, the reforestation program due to the timber shortage after the Second World War, the fall of the Iron Curtain, the Common Agricultural Policy and accompanying afforestation 10 actions of the EU. The validation with historic aerial photographs from 1950 and 1990 for 73 sample sites across Europe revealed that our results could capture most of the overall patterns of land change, although deviations with the observed data remain. In comparison with other land reconstructions like Klein Goldewijk et al. (2010Goldewijk et al. ( , 2011, Ramankutty and 15 Foley (1999), Pongratz et al. (2008) and Hurtt et al. (2006) it could be shown that our approach performs in line with these land reconstructions. Furthermore, the new method takes account of the harmonization of different datasets by achieving a high spatial resolution and regional detail with a full coverage of different land categories. These characteristic allow the data to be used for supporting and improving on-going BGD BGD BGD 9,2012 Reconstructing historic land change in Europe