Introduction

Landscape heterogeneity is considered a key factor for biodiversity conservation and ecosystem-service provisioning on farmland (e.g. Benton et al. 2003; Bassa et al. 2012; Schippers et al. 2015). During the past decades, however, farmland landscapes have become ever more homogeneous, which is often associated with the intensification of agriculture, and it is regarded as a major cause for the loss of biodiversity and of number of services such as habitat provisioning, pollination and water quality (e.g. Stoate et al. 2001; Tscharntke et al. 2005; Kleijn et al. 2009; Stoate et al. 2009). As a consequence, agri-environment policies and management strategies have been devised to maintain or enhance farmland heterogeneity, involving for instance support to farmers for increasing the amount of natural habitats, the diversity of crop types or land uses (compositional heterogeneity), and the spatial complexity of patch shapes and of their distribution in the landscape (configurational heterogeneity) (Fahrig et al. 2011). In general, however, limited consideration has been given to the influence of farm management strategies on landscape characteristics, and vice versa, though these interactions can impact on farmer’s income and thus limit the practical application of agri-environmental management (Fahrig et al. 2011; Bamière et al. 2011).

Farm management strategies are expected to affect both compositional and configurational landscape heterogeneity (Dunning et al. 1992; McGarigal and Marks 1995), as they shape for instance the type and diversity of crops, the size of fields to enable the use of heavy machinery, the amount and spatial distribution of natural habitats, and the prevalence of non-crop elements such as ponds, hedges, and scattered trees, among others (e.g. Deffontaines et al. 1995; Calvo-Iglesias et al. 2009; Carmona et al. 2010). On the other hand, however, farm management options are constrained by a number of biophysical and structural features such as soil quality, the presence of rocky outcrops, and the amount of forest habitats, which are also reflected in landscape pattern (e.g. Persson et al. 2010; Ribeiro et al. 2014). In this context, farm managers may have one of two contrasting approaches, either modifying the landscape to suit their needs (i.e., landscape makers), or adapting farm management to existing landscape constraints (i.e., landscape takers). The first approach is often associated to agriculture intensification, involving for instance levelling of fields, removal of hedgerows and rocks, and drainage of wetlands, to allow for the intensive production of certain crop types (e.g. Tscharntke et al. 2005; Concepción et al. 2007; Ruiz and Domon 2009; Ferreira and Beja 2013; Lomba et al. 2014). The second approach is more associated with low-intensity systems, and implies farming strategies that can be implemented without major changes to landscape characteristics (e.g. Bignal and McCracken 1996; Lomba et al. 2015). Although this dichotomy probably represents the extremes of a continuum of farmer attitudes towards the landscape, it may provide a useful framework to understand how farm management shapes, or is shaped by, the landscape, and how this interferes with the development of agri-environment strategies for safeguarding farmland heterogeneity.

In common with other studies examining the relationships between agriculture and biodiversity and ecosystem services (e.g. Oñate et al. 2007; Calvo-Iglesias et al. 2009; Pointereau et al. 2010; Carmona et al. 2010; Bamière et al. 2011), the farming-system approach may provide a convenient starting point to describe the interactions between farm management and landscape characteristics, as it helps simplifying the very large range of activities carried out by farmers. The concept of farming system stems from agricultural economics, and it is based on a comprehensive analysis of all farm management activities, including land use, animal husbandry, farming practices, resources involved, and the interdependencies between them, within a given political and socio-economic environment (Reboul 1976). Under this concept, farms may be aggregated based on the similarity of resource bases, production patterns and management strategies, which are likely to impact on the landscape in similar ways, and to show similar responses to biophysical conditions, as well as policy and market changes (Dixon et al. 2001; Ferraton and Touzard 2009). Therefore, it may be hypothesised that each farming system should be associated with a particular set of landscape characteristics, which may be more or less variable within each system depending on agriculture intensity level. Specifically, it may be expected that (i) more intensive farming systems should behave as landscape makers, transforming the landscape to meet their requirements, and thus showing low variability in landscape characteristics across areas farmed using the same system (i.e., low landscape beta diversity); and (ii) less intensive systems should behave as landscape takers, adapting to the constraints imposed by landscape features, and thus showing higher variability across areas under the same system (i.e., high landscape beta diversity).

In this paper we address these issues by assessing how farming systems are related to both landscape compositional and configurational heterogeneity, based on a case study developed in an agricultural landscape of southern Portugal. Using a typology of farming systems developed previously (Ribeiro et al. 2014), we mapped the spatial distribution of farming systems (i.e., a map of landscape units composed of blocks of contiguous farms sharing the same farming system) and characterised landscape patterns at the farming-system level (i.e., a spatial scale intermediate between the farm and the landscape levels), seeking to answer the following questions: (1) are landscape patterns different across farming systems?; (2) if yes, are these differences more associated to landscape composition or configuration?; (3) are there significant differences in landscape variability (beta diversity) across areas within each farming system?; and (4) is this landscape variability associated with the intensity of the farming system? Implications of the results for designing agro-environment policies to retain farmland heterogeneity are then discussed.

Methods

Study area

The study was carried out on a high nature value farmland area with ca. 210,000 ha in southern Portugal (approx. lat.: 38°N; long.: 8°W) (Fig. 1). The landscape is dominated by lowland agricultural systems, with undulating relief and altitudes ranging between 100–200 m above sea level. Climate is Mediterranean, with hot dry summers and mild rainy winters. The area encompasses the Special Protection Area (SPA) of Castro Verde, designated under EU Directive 79/409/CEE (Birds Directive) to protect populations of globally threated species such as lesser kestrel (Falco naumanni), great bustard (Otis tarda) and little bustard (Tetrax tetrax) (BirdLife International 2004). The typical landscape is a mosaic of dry cereal crops in rotation with long term fallows, and low density livestock grazing (Delgado and Moreira 2000). Since 1995, part of the area has benefited from an agri-environment scheme under the European Common Agricultural Policy (CAP) aiming to protect traditional farming systems (Santana et al. 2014). This scheme encourages agricultural practices considered favourable to conservation, including support to traditional rotation of cereals and fallows, the maintenance of low livestock densities, the growth of crops benefiting steppe birds, and the creation and maintenance of wildlife watering spots (Santana et al. 2014). In recent years the traditional farming system has been declining, with many farmers converting to specialized livestock systems, or to more intensive crop systems where soil quality and water availability makes it possible (Ribeiro et al. 2014).

Fig. 1
figure 1

Location of the study area in southern Portugal, showing the Special Protection Area (SPA) of Castro Verde, the area included in the Agri Environment Scheme (AES) of Castro Verde, and the main urban areas

Full size image

Farming systems and landscape patterns

In a previous study using spatially explicit agricultural data at the farm level, Ribeiro et al. (2014) identified four main farming systems occurring in the region in 2000–2002: (1) the Traditional system, typically comprising farms with a mosaic of dry cereals and fallows in rotation, and pastures grazed by sheep at moderate densities; (2) the Crops system, including farms with dry cereals and other annual (mostly irrigated) crops, and with no pastures, no livestock and little fallow land; (3) the Cattle system, comprising farms dominated by pastures and small areas of dry cereal, and with high densities of grazing cattle; and (4) the Sheep system, including farms with a crop composition similar to the Cattle system, but with lower livestock densities and composed strictly by sheep.

The intensity of each farming system was estimated following Brookfield (1972, 1993), from estimations of farm income per hectare, calculated by multiplying the unitary gross margin of the different crops and activities (Rosário 2005) by their relative proportions in each farming system. Although agricultural intensity is often measured from the input side (e.g., fertilizers or pesticides consumption per unit of land; Turner and Doolittle 1978; Lambin et al. 2000; Dietrich et al. 2010), here we measured it from the output side (e.g. tons of cereal or beef per unit of land) because data on input consumptions was unavailable, and because it does not involve any presumptions about the effect of inputs on productivity. Since we were comparing farming systems with distinct productions, we used a monetary surrogate to measure and compare outputs across farming systems (Turner and Doolittle 1978; Dorsey 1999).

A digital map of farming systems was produced in a geographical information system, by merging all neighbouring farms classified in the same system (Fig. 2). In this way we obtained geographic landscape blocks, each corresponding to groups of farms under a single farming system. A random sample of 1 km2 circular land plots (ca. 564 m radius) was then extracted from the farming systems map, subject to three criteria: (i) each plot should be completely enclosed within the same farming system; (ii) plots should not overlap; and (iii) a minimum of 30 plots should be extracted per farming system, with no upper limit. A sample of 241 circular land plots was thus obtained, of which 85 were in the Traditional system, 33 in the Crops system, 91 in the Cattle system, and 32 in the Sheep system (Fig. 2). Plots were then overlaid on a land use/land cover map (LUC) of the study area, representing the agricultural landscape in the spring of 2001.

Fig. 2
figure 2

Distribution of the randomly selected circular land plots (black dots) used to extract landscape variables within the Traditional, Crops, Cattle and Sheep farming systems in the study area (grey areas) in 2000–2002

Full size image

Mapping considered 13 LUC categories (Table 1), based primarily on farm parcel level data on agricultural land uses, extracted from a spatially explicit database maintained by the Portuguese Ministry of Agriculture (Ribeiro et al. 2014), and complemented with the following data: (i) shrubland and (ii) “montado” (open cork oak Quercus suber and holm oak Q. rotundifolia woodlands) cover, obtained from the 1990 digital land cover map (IGP 2012); and (iii) bare soil cover (including ploughed fields), mapped using Landsat ETM + (Landsat 7) imagery (dated April 1, 2001). The shrubland and “montado” layers were updated using aerial photographs from 2001, thereby correcting for eventual changes that occurred since 1990. The proportion of the different LUC categories summed more than one, because we considered the land uses (e.g., crops) occurring under the tree canopy of “montado” (e.g., Bugalho et al. 2011).

Table 1 Summary statistics of variables used to characterize landscape composition and configuration in the 241 circular plots of 1 km2
Full size table

To describe landscape pattern we used 15 variables quantifying the proportions and diversity of dominant LUC types (landscape composition), and four variables describing configurational heterogeneity (landscape configuration) (Table 1). Variables were selected from a wide range of metrics commonly used in landscape ecology studies (McGarigal and Marks 1995), following the recommendation to avoid redundant or highly correlated variables (e.g. Cushman et al. 2008; Schindler et al. 2008; Bassa et al. 2012). Landscape variables were computed for each of the 241 circular land plots (Table 1), using the Patch Analyst extension to ArcGIS 10 (Elkie et al. 1999).

Data analysis

Landscape variables expressing proportions (Table 1) were arcsin(sqrt(x)) transformed and the remaining were log10(x + 1) transformed to improve normality and stabilize variances (McDonald 2009). Overall local landscape patterns were explored using a principal components analysis (PCA) performed on a correlation matrix of the 19 transformed variables. Principal components (PC) with an eigenvalue larger than 1 were retained (Kaiser 1960), and a varimax rotation was applied to simplify and improve the interpretability of the solution. A one-way analysis of variance (ANOVA) was used to test for differences among the mean scores on each PC of plots classified in each farming system.

Linear discriminant analysis (LDA) was then performed on PC to assess the extent to which landscape patterns differed across farming systems, and whether the latter could be predicted from the former. A stepwise forward variable selection procedure was conducted using the Wilks’ lambda test, starting with an initial model that includes the variable which separates groups the most, and then adding further variables contingent on the Wilk’s lambda criterion (Roever et al. 2014). The procedure was stopped when no additional variable was significant at p < 0.05. Prediction accuracy was assessed based on a confusion matrix, implemented with leave-one-out cross-validation. Because sample sizes were uneven among farming systems, we used Cohen’s kappa to correct for agreements occurring by chance between observed and predicted categories (Titus et al. 1984).

Inertia ellipses were used to help visualizing the distribution of observations within each farming system along the axes in PCA and LDA plots, considering the default probability of ca. 66 % corresponding to a one standard deviation length. To characterise variability in landscape pattern among plots classified in the same farming system, we computed an adimensional beta diversity index (Anderson et al. 2006). Specifically, we computed the average Euclidian distance to group centroids within each farming system using the first two PC coordinates, and then standardized the index by dividing their values by the maximum distance obtained for the four farming systems.

Statistical analyses were implemented in R version 3.1.1 (R Development Core Team 2014) using the “psych” package (Revelle 2014) for the PCA, the “klaR” package (Weihs et al. 2005; Roever et al. 2014) for the Wilks’ lambda test and the “MASS” (Venables and Ripley 2002) for the LDA. The “candisc” package (Friendly and Fox 2014) was used to extract the standardized canonical coefficients of the discriminant functions.

Results

Overall patterns

There was a spatial trend for the Crops system to be predominant in the northern section of the study area, while the Sheep system was more common in the South. Both the Traditional and the Cattle systems were more evenly spread throughout the study area (Fig. 2). The sampled landscape plots were dominated by pastures, cereal crops, and fallow fields, which together accounted for more than 80 % of the area (Table 1). The average number of different LUC was 3.9 per plot, and the average patch number was six. “Montados” were common in the region, covering ca. 20 % of the area of the plots (Table 1).

The PCA returned seven PC with eigenvalues larger than one, together accounting for 69 % of the overall variance (Table 2). The first PC (hereafter named “Heterogeneity”) represented a gradient of local landscape heterogeneity, showing a joint increase in the number of patches, edge density, patch size variation, shape complexity (AWMSI), land cover richness and diversity. The second PC (“Specialization”) represented a gradient from landscapes dominated by annual crops (cereals and other annual crops) to landscapes dominated by permanent pastures, thereby separating landscapes associated with crop production from the ones specialized in livestock production. PC3 (“Permanent crops”) was mainly related to increasing proportion of permanent crops (mainly olive orchards), and also of wetlands. PC4 (“Built-up areas”) was associated to the presence of built-up areas, PC5 (“Leguminous crops”) reflected the proportion of leguminous crops, PC6 (“Forage”) represented the proportion of forage crops, and PC7 (“Fallows”) was associated to both fallows and bare soil areas, with opposite coefficient signs. Plot coordinates in PC were significantly different across farming systems (ANOVA, p < 0.05), except for PC4 (p = 0.340) and PC5 (p = 0.422) (Fig. S1 in supplementary material).

Table 2 Standardized loadings of variables on the principal components (PC) retained (eigenvalues > 1) from a principal components analysis of landscape metrics, and rotated using varimax
Full size table

The highest average local landscape heterogeneity, revealed by the position of the group centroids along the heterogeneity axis (PC1), occurred in the Traditional system, followed by the Cattle and Crop systems, whereas the lowest heterogeneity was found in the Sheep system (Fig. 3a). However, the inertia ellipses suggested that the Sheep system was the most variable spatially, as its landscape plots were found highly scattered along the entire heterogeneity axis (PC1). In contrast, the Crop system presented the lowest dispersion of plots along this axis (Fig. 3a).

Fig. 3
figure 3

Scatterplot of the 241 landscape plots in the first two axis extracted from a principal components analysis with varimax rotation (a) and in the first two axis of a linear discriminant analysis (b). The centroids and inertia ellipses are provided for each farming system

Full size image

Discrimination of farming systems based on landscape patterns

Variable selection using Wilks’ lambda returned five PCs that significantly contributed to the separation of farming systems (Table 3), which were subsequently used in the LDA. Built-up areas (PC4) and Leguminous crops (PC5) were discarded because they failed to improve the group separation power of the model.

Table 3 Summary of the Wilk’s lambda stepwise forward variable selection test performed on the 7 PC
Full size table

The first discriminant function (LD1) captured most of the between-group variance (78 %), mainly separating the Crops from the other systems (Fig. 3b). Specialization (PC2) was the most important variable contributing to LD1 (Table 4), with landscapes associated to the Crops system showing higher proportions of cereals and other annual crops, whereas landscapes associated to the Sheep system had a higher proportions of pastures. The Traditional and Cattle systems were both located close to the origin (Fig. 3b), indicating that they were little differentiated by the specialization gradient described by PC2. However, there was a tendency for landscapes in the Traditional system to be closer to the Crops system, due to their higher proportion of cereals. Likewise, the Cattle system tended to be closer to the Sheep system due to their higher proportion of pastures.

Table 4 Standardized canonical coefficients of the discriminant functions
Full size table

The between-group variance captured by the second function (LD2; 16 %) was much lower than that of LD1, thereby showing a much lower discrimination ability of farming systems. Nevertheless, LD2 contributed to a weak separation of Traditional and Cattle systems from the Sheep system, with the former being associated with higher local landscape heterogeneity (PC1) and higher proportion of forages (PC6) (Fig. S1 in supplementary material). The third discriminant function (LD3) captured a marginal 6 % of the between group variance, and it was mainly associated with PC7 (Fallows).

The overall percent agreement rate achieved by LDA predictions with leave-one-out cross-validation was 55.7 % (Table S1 in supplementary material), corresponding to an overall Cohen’s kappa of 0.35 after correcting for chance agreements. In pairwise classification tables for individual farming systems, the proportions of correct classifications were much higher for the Crops (92.5 %; Cohen’s kappa = 0.68) and Sheep (87.1 %; Cohen’s kappa = 0.44) systems, than for the Traditional (65.6 %; Cohen’s kappa = 0.25) and Cattle (64.3 %; Cohen’s kappa = 0.24) systems.

Landscape beta diversity and agricultural intensity

Landscape beta diversity was highest for the Sheep system, intermediate for the Cattle and Traditional systems, and lowest for the Crops system (Fig. 4). There was a significant inverse relationship between beta diversity and agricultural intensity (R = −0.96, p = 0.030), though care should be taken in the interpretation of statistical testing due to small sample sizes.

Fig. 4
figure 4

Relationship between the landscape beta diversity index and the index of agricultural intensification, for four farming systems identified in the study area

Full size image

Discussion

Our study pointed out some differences in landscape patterns across farming systems, and that these were far more associated to differences in landscape composition than in configuration. Also, we found differences in landscape beta diversity across farming systems, and that there was an inverse relationship between beta diversity and agricultural intensity. Overall, therefore, our results agree with the idea that farms associated with the more intensive systems may operate under a narrower range of landscape patterns, thereby shaping the landscape to suit their needs. In contrast, less-intensive systems may be compatible with a wider range of landscape patterns, thereby adapting to existing landscape features. These results have implications for agri-environment policies, as they suggest that maintaining landscape patterns to achieve biodiversity and ecosystem service goals require due consideration of the farming systems operating in those landscapes.

Results of our study are in line with previous research showing that farming systems have an influence on landscape pattern and dynamics (e.g. Deffontaines et al. 1995; Calvo-Iglesias et al. 2009; Carmona et al. 2010). In our case, this relationship was mostly a consequence of differences in land uses and cover between systems, mainly associated with the higher amount of cereals and other annual crops in the Crop system, and with the higher amount of pastures in the Sheep system. The Traditional and the Cattle systems had an intermediate position in this gradient, though with stronger affinities between the Traditional and the Crop systems, and between the Cattle and the Sheep systems. This suggests, therefore, that one of the main mechanisms whereby farmers affect landscape pattern is through a decision to produce either crops or livestock, thus leading to a compositional dominance of either cereals and other crops or pastures, or a mix between the two in the Traditional system.

The relatively weak relation between farming systems and local landscape configuration (sensu Fahrig et al. 2011) observed in our study was unexpected, because there was in the study region a dominant gradient from simple to heterogeneous landscapes (with higher number of patches, more variable patch sizes, more complex shaped patches, and higher edge densities). However, farming systems were little differentiated along this landscape gradient, though there was a tendency for the highest local heterogeneity to be found in the Traditional system, followed by the Cattle and Crop systems, and then by the Sheep system. The highest heterogeneity found for the Traditional system was probably a consequence of the mixture of different crop types and pastures characterising the system, which may promote a more complex patchwork of land uses. This is also in line with the major importance for biodiversity conservation of this system, because the diversity of land uses and cover is beneficial for a wide range of species with contrasting requirements, and for species requiring a blend of complementary habitats (e.g. Reino et al. 2010; Santana et al. 2014). It should be noted, however, that variability within each farming system was high, and so high to medium levels of local landscape heterogeneity could be found in areas dominated by any of the farming system.

In line with expectations, there was marked variation in landscape beta diversity among farming systems, with a tendency for a gradient from higher to lower beta diversity to be inversely related to agriculture intensification. The Sheep system is a case in point, corresponding to the less intensive system and the one showing the wider range of landscape patterns. Farmers may thus choose to specialise on sheep production largely irrespective of landscape patterns, either because this production system is not strongly constrained by the landscape or, in alternative, because the farmer’s income is not sufficiently high to allow investment for changing the landscape to meet their needs. In contrast, the Crop system was the most intensive and was associated with the narrowest range of landscape beta diversity. This was probably because specialised crop production demands landscapes with specific characteristics, due for instance to the need of large fields where pivot irrigation systems and heavy machinery can operate. The Traditional and the Cattle systems were intermediate in the beta diversity and intensification gradients, probably because they can accommodate a relatively wide range of landscape patterns, though they may still be able to change the landscape where this is necessary to suit their needs.

Overall, our results are consistent with the landscape taker versus landscape maker dichotomy, with the Sheep system falling in the former category, the Crop system falling in the latter, and the Cattle and Traditional systems probably having an intermediate position. It should be borne in mind, however, that our study is based on a snapshot of farming systems and landscape patterns, and so it is unknown whether the most intensive systems actually changed the landscape to suit their needs (i.e., acted as true landscape makers), or whether they were adopted in landscapes that were already suitable at the outset. Probably both mechanisms were influential, as there is evidence that temporal dynamics in farming systems are shaped by biophysical and structural characteristics of the territory such as soil fertility or the presence of forests limiting agricultural development (Ribeiro et al. 2014). There is also some evidence, however, that the most intensive systems actively change the landscape, by for instance increasing field size, removing hedgerows, ponds, scattered trees and non-crop elements, and reducing crop diversity (e.g. Tscharntke et al. 2005; Concepción et al. 2007; Ruiz and Domon 2009; Ferreira and Beja 2013; Lomba et al. 2014).

Our results have implications for policies and management strategies aiming to retain landscape heterogeneity on farmland, and thus to the conservation of biodiversity and ecosystem services (e.g. Fahrig et al. 2011). We found that more intensive farming systems were associated with reduced landscape beta diversity, thereby suggesting that agricultural intensification may contribute to landscape homogenization at the regional scale. The negative effects of homogenization have been demonstrated in several studies, both in our area (e.g. Reino et al. 2010; Santana et al. 2014) and elsewhere (e.g. Benton et al. 2003; Bassa et al. 2011, 2012), thereby pointing out the need to avoid regional dominance by a single, intensive farming system. Also, we found that less-intensive systems may occur under a relatively wide range of landscape patterns, suggesting that they may be better able than intensive systems to adapt to landscape constraints. As such, these systems may be more compatible with agri-environment schemes aimed at preserving heterogeneity or to retain important landscape features (e.g., ponds and hedgerows), because these should interfere little with farm management operations. As a consequence of the former two findings, it is likely that in a landscape with a mixture of farming systems with different intensity levels, agri-environment schemes should primarily target at the less intensive ones, with the double objective of promoting landscape features important for biodiversity and ecosystem conservation, and avoiding transitions between less- and more-intensive farming systems (e.g. Ribeiro et al. 2014). The latter may be particularly important, because intensification often results in landscape changes that are damaging to biodiversity and ecosystem services (e.g. Stoate et al. 2001; Tscharntke et al. 2005; Kleijn et al. 2009; Stoate et al. 2009), and these may be irreversible through voluntary policies like the agri-environment schemes due to the high costs required to compensate the investments and eventual loss of income by farmers. Furthermore, intensive systems may have limited flexibility to accommodate landscape patterns compatible with biodiversity conservation (e.g. hardly reversible loss of non-crop elements).

Overall, our results underline the importance of considering farming systems as a tool for designing agri-environment policies and to inform landscape management strategies. The farming system approach may be particularly useful because it explicitly recognises the existence of tracts of farmland that are managed similarly and, for this reason, react to the same market and policy changes, which eventually lead to changes on landscape composition and configuration, both at the local and at the regional level (Deffontaines et al. 1995; Calvo-Iglesias et al. 2009; Carmona et al. 2010; this study). As different farming systems often coexist in the same region, cost-effective design of agri-environment schemes should thus be based on a correct identification of dominant farming systems, on their compatibility with environmental management objectives, and on the priority and type of support that should be given to each one for promoting biodiversity conservation and the delivery of ecosystem services. This might provide an opportunity to tailor agri-environment schemes to socio-ecological specificities, without the high costs that would be needed for developing and implementing farm-level management strategies.