Introduction

Semi-natural areas are an inseparable part of European cultural landscapes, and were shaped through millennia of livestock grazing and other low intensity farming practices (Benthien et al. 2018). Many European semi-natural areas are among the most species rich ecosystems in the world, and host some of Europe’s most threatened species and habitats (Wilson et al. 2012). Such high biodiversity levels have mostly evolved as a combination of traditional low intensity pastoral and grassland cutting activities and distinct local soil, terrain and climate conditions that limited otherwise more intensive uses (Johansen et al. 2019; Schermer et al. 2016). Besides supporting high plant species richness, semi-natural areas are vital pollinator habitats due to their high flower diversity, and ample space for nesting (Holzschuh et al. 2006; Öckinger and Smith 2007). In addition, they are often landscapes characterized by large variation in plant functional types with different shares of trees, shrubs and grasslands such as agroforestry areas (Plieninger et al. 2015).

Semi-natural areas in Europe are the result of long-term extensive agricultural activities, mostly grazing and/or direct conversion of natural areas such as forests into more open areas often found in areas with harsh environmental conditions. This has led to increased niche diversity in European landscapes, and semi-natural systems still harbour far higher species diversity than intensive modern farming systems. Various semi-natural landscapes in Europe have been subject to overgrazing in recent decades (Sartorello et al. 2020), leading to heavily altered vegetation states. At the same time, there is a nuanced balance as grazing in or close to nature areas can lead to impoverishment of their biodiversity and ecosystem functions, potentially leading to irreversible (or difficult to restore) damage. As semi-natural areas generally have low carrying capacities due to limiting climatic, soil and terrain conditions, even low livestock densities can lead to degradation of soil resources and biodiversity (see for example Britton et al. 2005; Kosmas et al. 2016; Pakeman et al. 2003; Pulido et al. 2018; Quaranta et al. 2020). Overgrazing is therefore often disputed as the open landscapes resulting from this are seen as part of the cultural landscape (Monbiot 2022).

Spread across all environmental zones in the European Union (EU), semi-natural areas demonstrate exceptional diversity of different agricultural practices throughout history. One of the most well-known examples are the Dehesas and Montados in Spain and Portugal, archetypical silvo-pastoral landscapes of oak woodland, some arable cropping and pastures of different intensities (Joffre et al. 1988; Pinto-Correia et al. 2011). Semi-natural areas of high importance for cultural heritage have become part of regional identities in other EU countries as well. The Burren in Western Ireland is a pastoral landscape defined by a mosaic of limestone patches, grassland, heathland and wetlands (Proctor 2010). Mellanbygd landscapes of southern Sweden, are characterized by semi-open pastures scattered with trees and shrubs (Ihse and Skånes 2008; Torralba et al. 2018). Steljniki, open birch woodlands with an understory of ferns and clubmoss, are a result of low intensity grazing and bedding harvesting in southeastern Slovenia (Paušič and Čarni 2012). Furthermore, the European Alps (and other European mountain ranges such as the Apennines, Carpathians and Pyrenees) have been shaped by transhumance pastoralism, resulting in, for example, the Almen, Alpeggi and Alpages, seasonal mountain pastures used during summer (Paris et al. 2019; Schermer et al. 2016; Schirpke et al. 2017).

Semi-natural habitats have experienced a strong decline during the last 50 years, and are threatened by abandonment and intensification of traditional land use on these areas all across Europe (Hodgson et al. 2005; Poschlod et al. 2005; Prévosto et al. 2011; Schils et al. 2022; WallisDeVries et al. 2002; Weiss et al. 2014), making them among most threatened ecosystems on the continent (EEA 2020a; IPBES 2018; Stoate et al. 2009). The distribution and type of livestock grazing is essential to their maintenance as abandonment of grazing could lead them to disappear. The abandonment of semi-natural habitats leads to changes in species composition (Hilpold et al. 2018; Marini et al. 2009; Valkó et al. 2018) and landscape homogenization (Kaligarič and Ivajnšič, 2014; Rey Benayas 2007). Extensive grazing of shrubland and forest areas has historically prevented the build-up of dry biomass in Mediterranean areas and the abandonment of these practices, therefore, contributes to the increased fire risk in southern Europe (Moreira and Russo 2007; Viedma et al. 2015). Converting semi-natural areas to more intensive uses leads to biodiversity loss (Humbert et al. 2021; Mangels et al. 2017), but also soil and vegetation disturbance, which could be irreversible (Ceulemans et al. 2013; Gilbert et al. 2009). Preserving and restoring semi-natural areas has, therefore, become an important objective in EU’s nature conservation agenda, which has set ambitious targets for the maintenance and restoration of semi-natural habitats (Benthien et al. 2018; Halada et al. 2011; Holland et al. 2017). Despite the importance for biodiversity and landscape conservation, on one hand, and demonstrated potential negative impacts of over-grazing in such fragile ecosystems on the other, the role of grazing and the spatial distribution of grazing activities is uncertain. Understanding the spatial context of grazing on semi-natural areas is necessary to preserve these areas and revert the negative impacts of the intensification of farming on the environment, particularly biodiversity loss, in the EU. Without such information, identifying priority areas for conservation and restoration, as well as addressing the drivers and enablers leading to abandonment and continuation of the use of semi-natural areas will not be possible.

The EU is among the regions with most detailed and regularly updated spatial data on land use and land cover, derived from satellite imagery (see for example (EEA 2020b, 2020c, 2020d). Nevertheless, data on the spatial pattern and distribution of grazing in the agricultural landscape are not available or are limited to individual member states (where they are mostly derived from censuses with limited spatial resolution). Livestock grazing is also very difficult to map using satellite imagery and automated procedures, with one of the main challenges being able to distinguish between different uses and intensities of grasslands (Estel et al. 2018; Kuemmerle et al. 2013). For example, grasslands could be mown, or grazed (pastures) and it is not easy to separate these two from satellite imagery (Estel et al. 2018). New methods, able to map livestock grazing using high resolution satellite imagery and based on NDVI or even directly identifying the presence of livestock using high-resolution imagery have been developed, however they are data intensive and are applicable only for small areas (Castro et al. 2020; Kolecka et al. 2018; Masters 2021; Plaza et al. 2022). Low grazing intensities on semi-natural grassland are very difficult to detect as they create only a very limited spectral signal. Furthermore, grazing occurs on a wide range of vegetation types where automated methods generally have not been tested yet. Novel approaches, combining existing data on land use and land cover, with contextual territorial conditions are therefore necessary to map the spatial distribution of grazing across larger scales.

In this study, we map the spatial distribution of grazing in semi-natural areas in the EU Member States and the United Kingdom (hereafter referred to as EU27 +). We present an innovative approach, combining data and methods from different disciplines (agronomy, conservation science, remote sensing, spatial analysis and machine learning).

Methods

Methods overview

We present a multi-method approach (Fig. 1), combining expert knowledge with advanced spatial distribution modeling. First, we identified CORINE land cover (CLC) classes that contain primarily semi-natural vegetation (Paracchini et al. 2008). This enabled us to create a land cover mask for the estimated distribution of semi-natural areas in Europe. Secondly, we obtained expert-based estimates on grazing shares across different land cover types in different environmental zones for the EU27 + by performing an extensive survey among livestock grazing and conservation experts from each Member State. Thirdly, we identified how the socio-economic and biophysical contexts can explain the occurrence of grazing and mapped the spatial distribution of grazing within the semi-natural area mask. To do that, we applied maximum entropy modelling—a machine learning approach, using a set of pan-European spatially explicit data on land use and socio-economic and biophysical geographic characteristics as explanatory variables. For grazing occurrences, we used in-situ data from the pan-European ‘Land Use and Coverage Area frame Survey’ (LUCAS). The result of the maximum entropy spatial distribution modeling were probability maps, which we then used together with the expert-derived grazing shares, to allocate grazing across semi-natural areas in the EU27 + . Finally, we compared our results with two independent datasets on observed grazing in the analysed region: a grazing map for Austria, and data on livestock presence obtained from social-media photo uploads for all of the EU27 + . All our analyses were performed using publicly available data and open-source software. All data were processed using QGIS (QGIS Development Team 2020) or R (R Core Team 2021).

Fig. 1
figure 1

Flowchart of methodology for mapping grazing in semi-natural areas of the EU27 + 

Full size image

Grazing share estimates

Despite the abundance of statistical data on various agricultural characteristics for the EU27 + there is no publicly accessible data on the surface area that is grazed on a national level. We therefore first performed an extensive qualitative survey among 70 livestock experts from across the EU27 + (Supplementary Material). In each EU Member State and the United Kingdom, we contacted grazing experts, including ecologists, agricultural experts and agricultural officials. The survey was conducted between January and December 2020, and the exchange with the experts was done in several phases.

The process started by sending the experts a survey on grazing in semi-natural areas (S1). Each of the grazing experts filled the survey, where they defined which agricultural and semi-natural land cover classes (using the CLC classification, explained below) are grazed. Then, we asked them to provide estimated shares of grazing per class for those classes that are being grazed, differentiating between environmental zones within Member States. All experts provided input for different environmental zones in their respective countries and could provide suggestions on different levels. For example, in some countries specific regions had considerably different grazing shares.

This was followed by interviews with all experts to clarify and estimate uncertainties in the actual land use of CLC classes. These bilateral exchanges were done with the help of on-screen shared land cover maps and street view photographs of areas covered with particular land cover types to confirm or modify their first grazing estimates. After the interviews, we prepared a summary and shared it with the expert for further feedback and potential adjustment of their estimates. We prepared tables with the calculated potential grazing areas using the land cover classes they identified as being grazed. Based on experts’ responses and collected statistical information, estimations on the grazed share and the livestock densities for the considered semi-natural (and agricultural) CLC classes in the respective environmental zones for the EU27 + were compiled. The estimations were also matched across environmental zones and national livestock husbandry practices were taken into account, by discussing with the surveyed experts on the shares of livestock kept indoors (e.g. most of goats in the Netherlands are kept inside), transhumance systems (such as in the Alps and in the Mediterranean), and potential grazing livestock we did not consider (such as pigs grazing in the Dehesa/Montado landscapes of Spain and Portugal). While the experts provided some estimations on trends of grazing, these were mostly qualitative (e.g. grazing has decreased considerably in the past decades), or related to the fact, that the extent of open semi-natural land cover types has decreased, meaning we could not use them to prepare a time series of estimates. Our estimates therefore present the current distribution of extensive livestock grazing on semi-natural land cover.

Final verification of the estimates for grazing per CLC class, grazing shares and livestock densities was done during four webinars with experts grouped by environmental zones. All experts who provided information at one point during the process were invited to join. With the exceptions of Luxembourg and Malta, at least one expert provided feedback for each country. This iterative approach was finished in December 2020, resulting in data on semi-natural land cover classes that are grazed and the share of grazing per CLC class. We provide the grazing shares in the Supplementary Material (S1), together with details on the grazing experts per country and the survey questions.

Land cover data and environmental stratification

We used the CORINE Land Cover—CLC (EEA 2020b) data for the year 2018 to identify the location and spatial distribution of semi-natural land cover classes in the EU27 + that are grazed. CLC is a publicly available regularly updated land cover map, available at a 100-m spatial resolution and a minimum mapping unit of 5 ha. More details on this land cover classification are available in Bossard et al., (2000) and Feranec et al., (2016).

The experts were asked to provide feedback on a range of agricultural and semi-natural classes (see S1). For this study we utilized their feedback on the occurrence and share of livestock grazing for the following CLC classes: Natural grasslands, Moors and heathland, Sclerophyllous vegetation, Transitional woodland/shrub, Sparsely vegetated areas, Inland marshes, Peat bogs and Salt marshes. This means we did not analyse potential grazing outside this land cover mask for semi-natural areas, which is presented in the Supplementary material S2. We performed individual analyses of the spatial distribution of grazing on the level of European environmental zones. The climatic stratification of Europe (Metzger et al. 2005) was used to delineate environmental zones in the EU27 + (delineated in S2). Stratifying the EU27 + into more homogeneous zones enabled us to better account for biophysical drivers and limitations for grazing, as opposed to performing the analysis on the whole EU27 + level or individual countries.

Grazing observations

The Land Use/Cover Area frame Statistical Survey—LUCAS (EUROSTAT 2020a) provides (some) information on signs of grazing, and served as our sample of grazing observations. LUCAS is a regularly and systematically performed field survey performed across the whole territory of the EU27 + on a regular 2 km grid. All LUCAS observations are geographically referenced points (all observations have geographic coordinates) with information on land cover/land use, agro-environmental and soil data. LUCAS data are regularly used in European environmental assessments, with different land use and environmental characteristics being used to model the spatial distribution of a specific environmental process or characteristic across EU Member States. Among others, LUCAS data have been used to model soil erodibility (Panagos et al. 2012), landscape elements such as hedgerows, grass margins and other linear features (van der Zanden et al. 2013), and to estimate land cover area across the EU (Gallego and Bamps 2008). We used data from all existing LUCAS surveys since 2006, available in the form of a harmonized database (d’Andrimont et al. 2020). In total, land use information is available for 651,780 unique locations across the European Union. Together with other soil, terrain and land use characteristics, LUCAS surveyors also reported signs of grazing, such as presence of pastures, livestock, manure, infrastructure (e.g. fences) or livestock trampling. Such signs of grazing were observed on 31,899 locations on all CLC classes across the EU27 + , and on 6,766 locations within our analysed semi-natural CLC classes (Fig. 2). These observations served as our sample for ‘presence of grazing’, in our spatial distribution of grazing models. To ensure sufficiently large samples in our spatial distribution modelling, we considered a minimum of 50 points per environmental zone. Most locations with observed signs of grazing were located in all Mediterranean zones, and the least in the northern environmental zones. Nevertheless, such large differences between the environmental zones did not impact the analysis, as it was performed for each environmental zone separately. Despite having a considerable number of records with observed grazing in the EU27 + , our sample was still limited. The LUCAS data does not systematically sample within all areas that are grazed (and those that are not grazed), as the observation depends on the time when the surveyors are present. In addition, points that fall in remote and high-altitude areas are not surveyed. Until the 2015 survey the altitude limit was 1000 m, and after that 1500 m. Nevertheless, on locations above these altitude thresholds, surveyors performed photo interpretation to identify land use characteristics of the location (including observed signs of grazing). Given the potential mismatch between grazing and observations, our sample presents a collection of pseudo-absence data (incomplete information on absence of grazing).

Fig. 2
figure 2

Distribution of LUCAS points with observed signs of grazing on semi-natural land cover (n = 6766) across different environmental zones in the EU27 + . Gray lines present borders of environmental zones and countries (shown in more detail in the Supplementary material). LUCAS points are color-coded according to the environmental zone they are located in

Full size image

Determinants of grazing occurrence

We used a set of 24 potentially explanatory location factors to empirically determine how grazing (as observed in LUCAS locations) can be explained by territorial characteristics. We used the most recent data on terrain, soil, climate and socio-economic characteristics (Table 1).

Table 1 Explanatory location factors behind grazing in semi-natural areas
Full size table

To study the influence of terrain, we focused on altitude and slope that can pose constraints to high-intensity grazing (Havlík et al. 2011). Furthermore, we chose bio-physical variables that can limit (or facilitate) grazing. We considered five soil variables that describe different physical and chemical soil characteristics: proportions of clay and sand, available water capacity, bulk density and the share of coarse fragments. Soil variables define the suitability or level of intensity for crop and fodder production, or livestock density by characterizing soil structure, fertility, and ability to store water (Gardner et al. 1999). Mean annual temperature and precipitation, temperature range and precipitation seasonality were selected as climatic variables as they determine potential vegetation growth and represent suitability for growing crops and fodder—extensive, semi-natural areas could therefore occur in areas less suitable in climatic terms (Licker et al. 2010; Panagos et al. 2015).

Easier access to markets was assumed to increase the likelihood for grazing. We used different market accessibility indices (Verburg et al. 2011) to capture the impacts of access to financial services, agricultural inputs, lower transport costs and overall better accessibility. Proximity to roads and major cities was assumed to be important, as it can describe remoteness or easiness or difficulty to reach administrative institutions and facilities (Hinojosa et al. 2018). We considered it by using data on road density and travel time to nearest city (Nelson et al. 2019). In addition, we used two variables representing differences between countries within an environmental zone. We focused on farmer’s age, which has been identified among the most important socio-economic variables behind persistence of semi-natural and multifunctional areas, and abandonment and reforestation (Fayet et al. 2022). Here, we used data on farmers’ age, expressed as ratio of young (< 35 years) famers per older (> 50 years) farmers (Debonne et al. 2022)- We also considered the grazing share per country and CLC, which is expected to capture differences in culture and agricultural traditions of countries within the same environmental zone. Furthermore, the presence of nature protection through NATURA2000 designated areas was included to represent nature protection policies.

Spatially explicit data on location factors (except the digital elevation model), were available on a spatial resolution of 1 × 1 km2, which was the operating resolution of our grazing distribution modelling. The data were projected to Lambert Azimuthal Equal Area Europe ETRS89, which is the projection in which pan-European data are commonly provided in.

Maximum entropy modeling

Having an incomplete sample of grazing observation, characterized by a lack of absence points of grazing meant that we could not use traditional statistical methods where presence and absence data are combined (e.g. regression, such as in Glavan et al. 2017). We therefore used maximum entropy, that can account for incomplete pseudo-absence data, using the open source platform MaxEnt (Phillips et al. 2017). Maximum entropy is suitable to study the factors influencing the spatial distribution when there are only limited presence-only data available (Elith et al. 2011; Fithian and Hastie 2013). It has good predictive power particularly with small, limited datasets (Wisz and Guisan 2009). It is a machine learning technique to model spatial distributions (ranges), using a set of environmental data in grid form (e.g. soil, terrain, climate), and georeferenced observations of the studied characteristic (environmental process, state of environment, or species). Maximum entropy expresses the probability (likelihood) of each location (grid cell) to “host” the studied environmental characteristic. It does so, by providing a predicted probability of the assumed conditions (environmental grids) for the environmental characteristic, in our case, grazing. Using maximum entropy modelling, we calculated the most optimal grazing distribution based on the spatial distribution of LUCAS observations and our set of spatial explanatory variables (Phillips et al. 2006). The results of maximum entropy in this study can therefore be interpreted as predicted probability of presence of grazing. Maximum entropy is mostly used in ecological studies to study species distribution, but has been applied in numerous assessments of agricultural management and use. For instance, it has been used to study the crop distribution, changes in their spatial extent due to climate change, and carbon sequestration potential of land use (Duan and Zhou 2013; Liu et al. 2015; Luedeling and Neufeldt 2012; Machovina and Feeley 2013). Additionally, it has recently been used in Europe to study the spatial distribution of organic farmers (Malek et al. 2019b), and adaptation to irrigated agriculture in the Mediterranean region (Harmanny and Malek 2019). We built maximum entropy models in R (R Core Team 2021), using the maxent function from the dismo package (Hijmans et al. 2020).

Maximum entropy models were built separately for each environmental zone (Supplementary Material). LUCAS grazing records served as our sample of presence records, with all other semi-natural areas serving as the background grazing mask. Maximum entropy uses the background as locations where presence/absence cannot be confirmed as it is unmeasured, and where the likelihood for presence and absence data is equal (Merow et al. 2013). In this analysis, we randomly selected 100 times the number of observed grazing points per environmental zone from the background grazing mask. The observation of ‘signs of grazing’ in the LUCAS survey does not always mean that in the area grazing was actually observed, but signs that indicate that the area was recently grazed. Because of that, and other uncertainties in the surveying mentioned above, we used a default prevalence of 0.7. The default prevalence describes the certainty of presence records and is defined between 0 (low likelihood) and 1 (completely sure about the presence of the observed phenomenon).

The data distribution within grazing points was evaluated, to ensure that location factors carry sufficient difference between grazing and background points and are hence relevant predictors. Correlation between location factors was determined to reduce the likelihood of inflated variable importance due to inclusion of correlated factors (Lesschen et al. 2005). We excluded factors with a high level of collinearity (where the correlation coefficient > 0.8), as suggested by Menard (2002). If correlated factors were from the same group (i.e. terrain, soil, socio-economic or climate), only one variable was included based on better model fit. If location factors were correlated, but from different groups, none was excluded, as it was assumed that they still present different relationships and dynamics. For example, in the Alpine South environmental zone, sand and silt content exhibit high negative correlation, and we removed silt content from our model. At the same time, elevation and annual mean temperature also exhibit high negative correlation, however we kept both variables in this case, as both limit grazing in a different way. Elevation impacts access and poses physical limits to where livestock can graze, and annual mean temperature impacts plant growth and potential feed accessibility.

Two outputs of MaxEnt include the identified statistical relationships between the explanatory factors and grazing, as well as the spatial distribution (a spatially explicit map) of grazing probabilities. The result of maximum entropy modelling are response curves to identify the direction of the influence of each considered variable. We used the ‘Permutation importance’ values (%), standard output of maximum entropy modelling, to describe the contribution of each variable to the final result. Permutation importance values are estimates of how different variables contribute to the spatial distribution of grazing in terms of effect size, by breaking the link between the variable and the model (Phillips et al. 2017). To assess the predictive performance of our maximum entropy models, we calculate the Area Under Curve (AUC), the most common approach to evaluate MaxEnt model fit. AUC presents the probability that a random presence location is ranked higher than a random background point (Merow et al. 2013), with values over 0.5 indicating better performance than a random model. Nevertheless, the AUC here is calculated using presence (grazing observations) against background data (our spatial mask of semi-natural area where grazing could potentially occur), considering that presence data is fully certain (without considering default prevalence).

Grazing allocation

The spatial distribution modeling identifies areas that are more or less likely to be grazed, as it provides the probability of grazing for each location in the analysed semi-natural area mask. We used these probability maps to allocate the estimated share of grazed areas for each country, environmental zone and land cover type (based on expert inputs). We used the grazing share for the different CLC, country and environmental zone grazing estimates combinations within the background mask, to allocate grazing across the MaxEnt derived probability maps.

We first excluded areas regarded unlikely to be grazed due to physical characteristics, using different pan-European high-resolution data (Table 1). These characteristics included imperviousness (areas with > 30% imperviousness), wetness and water (permanent water), tree cover density (areas with > 30% tree cover on natural grasslands and moors and heathland, and areas with > 50% tree cover on transitional woodland/shrub). Then, we allocated grazing separately for each CLC class, country and environmental zone combination. We selected the pixels of the probability maps that have the highest probabilities for grazing, based on the share of grazing as defined by the surveyed experts. For example, if a particular land cover class was defined as being 30% grazed, we selected the same share of cells with highest probabilities for grazing. We refer to these locations estimated to be grazed as our final grazing map.

Comparison with other available data on livestock grazing

We compared our probability map, and the final allocated map of grazing in semi-natural areas with two independent spatial datasets on observed grazing. Given that all maps represent slightly different aspects of grazing and that all include certain levels of uncertainty we refer to the process as ‘comparison’ rather than validation. First, we compared our results with the grazing map for Austria from INVEKOS (Open Data Österreich 2021). INVEKOS data potentially overestimate the grazed areas, as the whole unit, for example parcel, is marked as grazed, even if just a small portion of it is actually grazed. Furthermore, some grazing may occur outside the land area included in INVEKOS. However, they still present one of the rare publicly accessible spatially explicit data on grazing in the EU (and globally), and are linked to payments (subsidies) for different agricultural activities, among them grazing, meaning they are regularly updated. The reported locations of grazing and non-grazing of the two maps were compared in a matrix. We then calculated the agreement of grazing and non-grazing, and the overall agreement. The grazing and non-grazing agreement was defined by comparing the share of locations that were reported as being grazed (or not grazed) in both maps. Additionally, we compared the probability map at locations which INVEKOS defines as being grazed with the values of our grazing map. This way we could compare our actual spatial distribution of grazing map with the INVEKOS data, without potential uncertainties related to expert derived estimated grazing shares. Here, we prepared kernel density plots, to understand the differences in the distribution of probability values. MaxEnt probability maps will only have a perfect normal distribution and a mean of 0.5 if all location factors are normally distributed. Therefore, we included the distribution of all probability values in the entire extent of semi-natural areas in Austria, to visualize potential shifts away from a mean of 0.5.

Secondly, we compared our results to data on observed grazing obtained from photographs uploaded on social media (Tieskens et al. 2017; Zanten et al. 2016). This dataset consists of user uploaded photographs of non-urban landscapes uploaded to Flickr, Panoramio and Instagram. The content of photos was identified using artificial intelligence (Google Image Search), resulting in keywords describing them. For this study, we used photos with keywords associated with grazing: livestock, shepherd, goat, cow, pasture, herd, sheep, grazing, cattle (and similar variations and combinations of these keywords). As this dataset only consists of presence of grazing and no absence data, we only determined the grazing agreement rate. Additionally, we analyzed the differences in data distribution of the probability values at the social media locations and compared them to the probabilities of the allocated grazing map and overall probabilities (same principle as Austria). In total, 1269 observations from the Tieskens et al. (2017) dataset were within our semi-natural areas mask, which we used in our comparison. For more details on how the photographs were derived from social media please see the studies of Tieskens et al. (2017) and Zanten et al. (2016).

Results

Spatial distribution of grazing

The expert survey first enabled us to estimate the amount of grazing on a range of CLC classes in the EU27 + (presented in more detail in the Supplementary Material S1). The analysis presented here focused on CLC semi-natural classes and shows that across the EU27 + about 21% of semi-natural land cover is grazed. This ranges from nearly none in Finland and Sweden (without reindeer) and Latvia to around 60% in Cyprus, Ireland and the United Kingdom (S1). Whereas less than 1% of semi-natural land cover types in the combined Alpine North, Boreal and Nemoral zone are grazed, over a third are still being grazed in many parts of Europe: Atlantic Central and North (55.7%), Alpine South (39.4%), Mediterranean Mountains (37.9%), and Mediterranean North and South (37.3%). The shares of grazing across land cover classes vary from 6.8% for Transitional woodland/shrub to 82.4% for Natural grasslands (S1). This demonstrates that large portions of the CLC class 321’Natural grassland’, which are defined as “under no or moderate human influence” (EEA 2020b), are actually grazed. This is expected in the light of human land use history in Europe and extensive grazing would appear compatible with the term ‘under moderate human influence’. In geo-botanical terms a better name for this class would probably be ‘(Semi-)natural grassland’.

We identify a large diversity of enabling and limiting conditions of grazing in European semi-natural areas, demonstrated by the main variables influencing the spatial distribution of grazing across different environmental zones (Table 2, Fig. 3, Supplementary material S3 and S4). In addition, this enabled us to map the spatial distribution of grazing probability (Fig. 4) and to allocate grazing within a pre-defined semi-natural area landscape mask (Figs. 5 and 6). The distribution of grazing determined by the different variables representing a range of landscape characteristics are visualized in more detail in the Supplementary material, and here we summarize the main observations. In general, we see that grazing in semi-natural areas is influenced more by terrain, climate and socio-economic characteristics, and less by soil conditions. Nevertheless, the distribution of grazing observations and resulting maximum entropy model fits suggest, that in most environmental zones, areas with more favourable socio-economic and environmental characteristics (within the range of marginal lands as predefined by the semi-natural land cover mask) are more frequently grazed.

Table 2 LUCAS points available for MaxEnt modelling and brief overview of the MaxEnt results for each environmental zone
Full size table
Fig. 3
figure 3

Variable contributions to the spatial distribution of grazing for 4 selected environmental zones: Alpine South, Atlantic Central, Mediterranean Mountains and Pannonian zone. Figures for other zones are provided in the supplementary material (S4). Note that the scale of x axis varies between the environmental zones

Full size image
Fig. 4
figure 4

Probability map for grazing in analysed semi-natural land cover mask for the EU27 + estimated with MaxEnt models

Full size image
Fig. 5
figure 5

Map of grazed semi-natural areas in the EU27 + (green areas). Detailed maps for some areas are presented in Fig. 6

Full size image
Fig. 6
figure 6

Grazed semi-natural areas in a) Ireland and United Kingdom, b) Portugal and Spain, c) Bulgaria and Greece, and d) Alps (France, Italy, Austria and Slovenia). All maps are on the same scale

Full size image

Terrain, represented by elevation or slope, was identified among the most important variables in all environmental zones, except the Mediterranean South. Grazing on semi-natural vegetation is mostly observed in areas with lower elevation compared to the elevation distribution in the environmental zones (S3), except in the Continental, Lusitanian and the Mediterranean zones, where it is more likely at higher elevation. In these zones transhumance remains common, and animals graze in higher altitudes during the summer period. Higher elevation areas in these environmental zones receive more rainfall, crucial in the summer period when lowlands have less rainfall. Similar is observed for slope, except that grazing occurs on areas with gentler slopes in the Mediterranean zones.

Climatic variables are mostly represented by temperature seasonality, which is among the main variables in the Alpine South, all three Mediterranean regions and the Continental zone. In the Alpine and Continental zone, grazing is more likely in areas with lower seasonality (a lower standard deviation of monthly mean temperatures), and in the Mediterranean regions in areas with slightly higher temperature seasonality. While in the Alpine zones this means that areas with less extreme temperatures (both in the winter and in the summer) are grazed, in the Mediterranean grazing occurs in areas, that have a milder climate (in terms of summer heat) and hence better growing conditions for vegetation in the summer.

Precipitation seasonality has the next highest influence and is among the main variables in the Continental and Lusitanian zone, and the Mediterranean South. In the Continental and Lusitanian zones, grazing is considerably more likely in areas with a lower precipitation seasonality (i.e. less severe drought conditions in the summer). In the Mediterranean South however, it is more common in areas with a higher precipitation seasonality. This could however be explained by the fact that grazing occurs in areas that have more precipitation in wetter periods (e.g. winter), that can support more vegetation throughout the year.

Annual mean climate characteristics (temperature and precipitation) are most important in the Pannonian and Lusitanian zone, and the Mediterranean South. In the Pannonian and Lusitanian zone, grazing is observed in areas with a lower annual mean temperature, but (slightly) higher annual precipitation. In the Mediterranean south, grazing is observed in areas with (slightly) higher temperatures and precipitation.

Other territorial characteristics described by the grazing share per CLC and tree cover density were identified among the most important variables in the Alpine South, all three Mediterranean zones, Pannonian zone and the Atlantic central zone. In general, grazing is found in areas with medium tree cover densities, except in the Pannonian zone, where it is more commonly observed in areas with a lower tree cover density. This confirms the diversity of semi-natural areas, that cannot always be explained by a dominant land cover type (e.g. grassland or woodland), but are in fact mosaic landscapes consisting of grasslands, trees, shrubs, and other land cover types.

Socio-economic variables most often identified as important are farmer’s age (in the Alpine South, the three Mediterranean zones, and Atlantic North and Central), and market access and travel time to cities (important in the Continental, Pannonian, Atlantic North and Central and the Lusitanian zone). In all zones where farmer’s age was identified as important, grazing occurrence is found more in areas with a more equal spread of the young to old farmer ratio compared to all semi-natural areas. In the Alpine South it occurs in slightly more areas with a higher% of young farmers. In both Atlantic zones, grazing occurs more likely in areas with a lower share of young farmers. In all three Mediterranean zones, grazing occurs in areas with a lower% of young farmers, similar to all semi-natural areas. However, it is more common in areas with a medium (up to 30%) share of younger farmers and is unlikely in areas with considerably high% of young farmers—despite the fact, that several regions where we can find semi-natural areas have considerably high shares of young farmers in the Mediterranean. Semi-natural areas are more likely to be grazed in areas with better market access in the two Atlantic regions (and to an extent also in the Pannonian zone), but in areas with lower market access in the Continental and Lusitanian zones. In the Atlantic zones, grazing is also slightly more likely in areas closer to major cities, whereas in the Continental and the Lusitanian zone this is not the case.

Soil characteristics were rarely identified as important, although silt content is important in the Continental and Lusitanian zone. This suggests that soil characteristics are not important for limiting or driving grazing in semi-natural areas. However, it is worth noting, that among all the variables we considered soil conditions are most commonly characterized by a normal distribution (S3).

Our spatial distribution models demonstrate a good fit (Area Under Curve, AUC > 0.7) in all environmental zones (Table 2), with models in some regions exhibiting excellent fits (e.g. Pannonian, Lusitanian and Continental zones). The AUC demonstrates how well we are able to explain the grazing distribution with the selected geographic variables. On the other hand, it could also be an indication of whether grazing in such areas is at all influenced by geographic characteristics, or general socio-economic trends (e.g. overall development or demography) that perhaps are not captured by available spatial data.

Comparison to other observations of grazing

We first compared our maps with the grazing map of Austria, looking at both the grazing agreement (the share of locations, where grazing was mapped in both maps, in relation to the total grazing area mapped in this study), and non-grazing agreement (agreement in non-grazing in relation to total non-grazing within the grazing mask of our study). The two maps agreed in 54% of the locations in terms of grazing, and in 71% in terms of non-grazing, with overall agreement being 62% (Fig. 7). A full comparison between the two datasets was however not entirely possible. Our analysis did not include the land cover class pastures (as it mostly represents intensively managed grasslands), whereas the INVEKOS grazing data includes such parcels. The INVEKOS data maps 21% more grazed areas within the semi-natural areas mask compared to our expert-derived estimates of grazed semi-natural land cover area. A large portion of the disagreement therefore lies within the discrepancy between the amount of grazing of the two grazing maps, and different extents of semi-natural areas (based on CLC) rather than the different spatial distributions.

Fig. 7
figure 7

Comparison of our grazing map with the INVEKOS grazing map for Austria

Full size image

Compared to the distribution of grazing probability values of the entire area (S5), grazing mapped by INVEKOS was more often found in locations with grazing probabilities above 0.5. Below this threshold, a random model would perform better. The probability values of grazing areas from the INVEKOS map were relatively evenly distributed over the entire range of probabilities. Probability values at locations, where grazing was allocated as result of this study, were predominantly above 0.5 and generally higher than the values at the INVEKOS grazing locations (S5). Social media observations of grazing across the EU27 + were furthermore used to validate our results. From the total of 1269 observations that were within our background mask, 901 agreed with our grazing map, presenting an agreement of 71% (Fig. 8). We observe particularly high agreement in the United Kingdom, the French Alps and the Pyrenees. Comparing the probability values at social media grazing observation locations with the probability values at allocated grazing in our map and with overall probability values, we found that probability values at the locations of social media observations are higher than the background mask and the grazing map, indicating rather high agreement (S5).

Fig. 8
figure 8

Locations of social media observation in comparison with our grazing map: agreements (green) and disagreements (red) with mapped grazing

Full size image

Discussion

Understanding the spatial pattern of low-intensity grazing in the European Union is important to the conservation of habitat types dependent on extensive grazing systems (Halada et al. 2011). In addition, this could provide numerous other co-benefits: maintenance of cultural landscapes, lower environmental impacts of agriculture (compared to high intensity grazing), and improved animal welfare. At the same time it is important to identify areas potentially degraded due to negative impacts of grazing. Another important environmental dimension associated with grazing of semi-natural vegetation is the reduction of wildlife risk due to the removal of biomass via low-intensity grazing (Álvarez-Martínez et al. 2016; Colantoni et al. 2020; Rouet-Leduc et al. 2021). Given current socio-economic trends it will be important to provide targeted support for such systems in many areas of Europe to avoid further abandonment of grazing. Conversely, a reduction of grazing intensity and change of current practices will be necessary in other places to achieve biodiversity targets. We provide a continental scale systematic analysis of how different contextual territorial characteristics impact the spatial distribution of grazing, which resulted in a map of grazed semi-natural vegetation for the wider European Union region. Our approach to map livestock grazing uses regularly updated data on locations of observed grazing and territorial characteristics. This makes our approach easy to update in the future, with new grazing observations or estimated shares of grazing per land cover type being obtained.

Spatial determinants of grazing

The influence of individual drivers or limitations behind grazing is comparable with other research on land use, grazing and land abandonment in the EU. Most studies however focus on the drivers of either intensification or abandonment, and do not analyse persistence of grazing (van Vliet et al. 2015). Similar to other studies, we demonstrate that the use of marginal lands for grazing can be described by an interaction of multiple socio-economic and environmental factors (Hersperger and Bürgi 2009; Lasanta et al. 2017; Rey Benayas 2007). The spatial mask within which we analysed occurrence of grazing focuses on semi-natural land cover, which has generally remained in Europe only on land not suitable for modern crop farming and/or intensive livestock systems. This means, that comparisons with regions with more favourable socio-economic and biophysical conditions are not entirely possible, as extensive grazing generally occurs on marginal lands. Nevertheless, our results agree with other Pan-European studies in identifying that within the marginal semi-natural areas, extensive grazing is more likely to occur in areas with more favorable biophysical and socio-economic conditions, where grazing could be more profitable (Lasanta et al. 2017; Levers et al. 2018; Terres et al. 2015).

Existing studies on land use and abandonment in Europe mostly focus on economic and institutional drivers and farmers’ characteristics, with contextual biophysical conditions mostly being represented by accessibility or remoteness of agricultural areas (MacDonald et al. 2000; van Vliet et al. 2015). In our approach we also consider a wider spectrum of potential territorial characteristics (e.g. climate and soil). Areas where grazing was abandoned in the past decades are concentrated in mostly hilly and mountain areas (García-Ruiz and Lana-Renault 2011; Lasanta et al. 2017; MacDonald et al. 2000), corroborative to our findings of grazing mostly occurring in areas with lower elevation and shallow slopes. We identified that grazing depends on the share of young (or old) farmers. Demography has been identified among most important drivers of land use change in different parts of the world (see for example Asselen et al. 2013; Geist and Lambin 2002), but there is a lack of evidence on how co-occurring demographic changes manifest themselves in land use change in Europe (van Vliet et al. 2015).

Despite the fact that we considered numerous spatial factors of grazing, evidence from other studies suggests that low income in extensive farming systems, lifestyle changes, presence of holiday homes and alternative employment opportunities (Debonne et al. 2022; Hinojosa et al. 2016; Malek et al. 2019a), together with agricultural labour, farm structure and farming conditions (mechanization, yields and field size), are also important determinants of land abandonment (Hinojosa et al. 2016; Levers et al. 2018). Due to the lack of empirical data (on a comparable spatial resolution), we could not consider such characteristics.

Mapping the spatial distribution of grazing

Mapping the presence and intensity of grazing or other types of grassland use remains extremely challenging (Erb et al. 2013; Kuemmerle et al. 2013). This is partly due to the inability to map grazing from satellite imagery, but also due to the large diversity of grassland and pasture systems and the definitions used to describe them in different countries (Smit et al. 2008). Other studies on mapping the spatial distribution of grazing on such scales are rare, particularly when it comes to including other semi-natural vegetation types beyond extensive grassland. Neumann et al. (2009) mapped the spatial distribution of all livestock in Europe (herbivores and monogastrics), using livestock statistics on subnational level, which were allocated across different CORINE land cover types using a combination of expert opinion and regression models. The approach, resulting in a raster map of 1-km cells, however did not explore the role of semi-natural vegetation as a fodder source and did not consider potentially different grazing in different environmental zones.

Estel et al. (2018) mapped grassland management intensity in Europe using MODIS NDVI. They mapped grassland mowing frequencies across Europe and considered the spatial distribution of livestock of Neuman et al. (2009) as proxy for grazing. Such an approach however potentially did not capture grazing in marginal areas, such as semi-natural vegetation, as only 1 of 5 clusters of grassland management was coinciding with semi-natural grasslands; it also did not include other vegetation types used for grazing (low mowing and grazing intensities). Our study has used agronomic knowledge on the fodder source of different livestock systems (Paracchini et al. 2008), the description of land cover classes in the CLC technical compendium (Bossard et al. 2000; Feranec et al. 2016), and expert-derived estimates of the share of grazing in different land cover types which represent semi-natural vegetation across the EU27 + . Additional geo-spatial analysis building on a maximum entropy approach enabled further refinement of grazing distribution estimates. Our spatial distribution of grazing probabilities suggests a similar pattern compared to simulated transitions of abandonment of semi-natural areas (Ceaușu et al. 2015; Verburg and Overmars 2009). This way, our results could also be used to track abandonment of grazing in areas where we identified low probability of grazing, by for example looking at tree cover change and shrub encroachment.

Limitations

The first limitation of our approach is the inability to separate between different types of extensive livestock grazing systems, and different types of grazing animals. Both distinctions might be important for different reasons. First, extensive grazing systems vary considerably across the European Union and exhibit high diversity. While we tried to address this by operating on individual environmental zones, within them different grazing systems have developed through history. Even in Atlantic Europe, which is among the most homogeneous European environmental zones in terms of biophysical characteristics, different grazing systems have developed on different ecosystems, among others grasslands (Aldezabal et al. 2015; Duckworth et al. 2000), heathlands (Rosa García et al. 2013), and shrublands (Ross et al. 2016). Other regions, particularly in Mediterranean Europe, are home to even larger diversity of extensive grazing systems. Separating between livestock is also important, as different types of animals have different impacts on the environment (Rosa García et al. 2013). Different types of animals (not only species, also breeds) also graze different ecosystems (see for example Garmendia et al. 2022; Jones et al. 2016; Šumrada et al. 2021; Török et al. 2014), meaning that any future policies to prevent abandonment or negative impacts of grazing need to consider that different grazing animals graze on areas with different biophysical and socio-economic characteristics—something, we were not able to consider due to lack of spatial data on grazing by different animals.

Other major limitations of our approach are related to available data. First, the allocated grazing fully depends on the land cover classes identified as grazed and the estimates on their grazed share. The grazing estimates could be biased towards past levels of grazing. To ensure that the estimates are as recent as possible, we performed our survey in an iterative matter throughout the course of one year. While we surveyed numerous local experts on grazing across the European Union (from each Member State and the UK) and ran a consolidation and cross-checking exercise, these are still estimates. Having adequate geo-spatial survey data (on subnational or even land cover level) would considerably reduce this limitation, and in the future, efforts should be taken to collect such information at country at/or EU level. For example, our comparison with Austrian grazing data indicates somewhat large uncertainties related to the share of grazed areas, as the Austrian dataset maps 21% more. Available statistical information, such as the Farm Structure Survey (EUROSTAT 2020b) could not be used, as it only provides information on the type of farm, or total livestock numbers (including livestock that does not necessarily graze). In addition, including existing spatially explicit data on livestock grazing and presence, such as a global distribution of livestock (Robinson et al. 2014), was not possible due to several reasons, such as different land cover data, and exclusion of protected areas from grazing, many of them actually being identified as grazed in our study. Nevertheless, our approach could easily be updated and improved using different grazing input data, once (if) they become available.

Secondly, the locations of observed grazing are also subject to uncertainty. The LUCAS survey also involves an accuracy assessment, which estimated high accuracy for land cover types (> 90%). Nevertheless, an accuracy assessment for observed signs of grazing is not available. Despite the fact that LUCAS surveyors are trained and instructed on how to spot signs of grazing, their observations were done only once. This means that grazing could have also occurred at a different time without major signs of grazing in those locations, or that there were still signs of grazing on locations, that were actually abandoned. In addition, not all locations across the EU that are potentially grazed were sampled (such as remote or high-altitude areas). We reduced this uncertainty by using a method that can deal with pseudo-absence (presence) data.

In addition, we performed our analysis using CORINE Land Cover data, which has its own limitations even though its production and validation follows clear technical guidelines. One issue inherent in the design of any land cover map produced from satellite imagery is that it records land cover primarily, with land use strongly associated with certain land cover classes, e.g. arable land. The thematic resolution of such land cover maps has a strong impact on analysing different ecological characteristics (Buyantuyev and Wu 2007), and excluding other classes could have lead to underestimation of semi-natural areas. We tried to overcome that by consulting experts who can match personal field observations and grazing system knowledge with different (semi-natural) land cover types. Secondly, the assignation of a specific land cover code to a certain CLC polygon follows the majority rule, which means that the dominant land cover type is assigned to the entire polygon. Thus, other land cover types can co-occur in the same polygon which means that the spatial distribution of grazing connected to a certain land cover type is only representative at larger unit level, e.g. at 1 km2 grid or larger. However, as our analysis aimed to detect grazing patterns at Member State or EU-level, this is not an important limitation for this study.

Nevertheless, having more detailed information on actual land use from Member States, for example using national Land Parcel Information Systems (LPIS), could improve our approach considerably, particularly by reducing potential biases of estimated grazing. Unfortunately, LPIS data are mostly not publicly accessible, not easy to process, or they have a coarse thematic resolution (sometimes consisting only of a few aggregated land cover classes). Furthermore, they do not cover all area grazed by livestock due to Common Agricultural Policy (CAP) rules imposing limits on what kind of vegetation structure and tree density is acceptable as eligible for CAP payments. This is particularly relevant in countries with significant areas of common land and/or a tradition of transhumance. Another way to reduce the limitations of using categorical land cover maps would be to consider grassland productivity derived from satellite imagery, that could serve as proxy for grazing intensity (Hankerson et al. 2019). Nevertheless, the productivity of grassy vegetation in semi-natural areas is mostly low across all environmental zones, making it difficult to identify any clear grazing patterns (Smit et al. 2008).

Recent advances in the spectral and temporal resolution of satellite images (e.g. latest Sentinel missions), could lead to improvements in mapping livestock grazing. Using data from the latest satellite images (together with in-situ data) has been demonstrated as a potential tool to map different intensities of grazing on a local scale (Saadeldin et al. 2022). However, approaches beyond local case studies are rare and demand extensive data collection (Kolecka et al. 2018; Plaza et al. 2022). This means that despite good technical progress in the last decade, it has not yet been possible to upscale remote sensing data for detection of livestock grazing (Masters 2021). Combining dense time series of satellite images with cloud computing platforms, could be a way forward in identifying also subtler changes to grassland use intensity (Pazúr et al. 2021). Developing improved spatial grazing data would benefit from substantial investment in field-based grazing observations combined with an expanded range of territorial variables as we demonstrate in our landscape context approach. Further useful spatial variables could be generated using high-resolution satellite imagery, for example, to capture crop and vegetation phenology or the diversity and density of landscape features (such as hedgerows and stone walls).

Conclusions

To develop policies and other measures at areas or vegetation types with highest risk of grazing abandonment, and to reduce damage from overgrazing or other negative management trends, we need methods and data that help us understand where semi-natural areas are still being grazed. We provide both in this study by combining expert knowledge and land cover data. Our study demonstrates how it is possible to develop a Europe-wide data set on semi-natural vegetation and the likely share of grazing per land cover class and country. We also show how geo-spatial data on grazing observations and a wide range of variables linked to vegetation productivity can be combined to estimate the likely distribution of grazing within this semi-natural land cover mask using approaches from other environmental domains such as species distribution modeling. From this analysis we conclude that there is no single dominant driver or limitation behind the spatial distribution of livestock grazing in semi-natural areas. Any policy will therefore need to consider the multitude of local characteristics: soil, climatic and terrain, as well as socio-economic. Moreover, grazing occurs in different locations depending on local and regional context, which vary widely across the European Union—even within single countries. Therefore, we expect that any measures to support biodiversity-friendly grazing practices need to be very targeted. What could work for semi-natural grazing in mountainous areas might not be appropriate in lowland and coastal regions. Considerable further investments in suitable input data, both via ground surveys and satellite programs, are, however, needed to consolidate and improve our work and to develop more tailored policy recommendations. This would help to maintain and develop extensive grazing systems that are suitable for managing priority semi-natural habitats and potential other co-benefits to the European society such as the reduction of wildfire risk.