The importance of key marginal habitat features for birds in farmland: an assessment of habitat preferences of Red-backed Shrikes Lanius collurio in the Italian Alps

Abstract

Capsule Key marginal habitat features maintain their importance even when they occur at very low density.

Aims To assess the importance for breeding birds of key habitat elements, such as isolated shrubs, hedgerows and untilled vegetation patches.

Methods We investigated the habitat preferences of Red-backed Shrikes Lanius collurio in an agricultural environment where key habitat elements have been dramatically reduced. Based upon data from territorial maps, a fine-scale model of habitat preference was developed and the results of three different analytical approaches (binary logistic regression, multimodel inference and hierarchical partitioning) were compared.

Results The occurrence of Red-backed Shrikes was positively influenced by the extent of non-grazed/mown grassland, isolated bushes and hedgerows, and negatively influenced by woodland cover. The model highlighted the importance of bushes/hedgerows despite their very scarce occurrence in the study areas (overall, average cover only 7.94%). Breeding densities were rather low, but mean densities were slightly higher in pasture zones in which bushes/hedgerows availability was higher than elsewhere.

Conclusion These results confirm the crucial importance of these key marginal elements even in depauperated farmland landscapes. Maintaining and increasing their availability is a key management option for the encouragement of breeding Red-backed Shrikes, and potentially for other species which are also declining in these farmland habitats.

In the second half of the 20th century the European populations of many bird species which utilize farmland have undergone rapid decline (Tucker & Evans 1997, Pain & Pienkowski 1997, Gregory et al. 2008). The most important causes of this decline have been recognized in the intensification of agriculture, encouraged by national and European policies (Common Agricultural Policy) to increase agricultural yields (Pain & Dixon 1997, Donald et al. 2001; Stoate et al. 2009), with population declines most marked in countries where agricultural intensification has been longer and more pervasive (Donald et al. 2001, Geiger et al. 2010). The increase of agriculture production has been reached through mechanization, use of chemical products and removal of many ‘unproductive’ marginal elements of the landscape such as hedgerows, bushes, trees and uncultivated or semi-natural lands. These marginal elements represent key habitats for many species and are probably more important for farmland bird populations than landscape heterogeneity per se (Heikkinen et al. 2004, Batáry et al. 2011, Kleijn et al. 2011). The availability of good foraging habitats close to safe nesting sites is a critical factor for birds breeding on farmland (Benton et al. 2002), and many farmland bird species are associated with marginal features (see e.g. Brambilla et al. 2008, 2009).

The other side of the deep changes in agriculture practices is represented by the abandonment of many areas where agricultural intensification was not possible or too expensive. Land abandonment is leading to a dramatic loss of habitat for many species tied to open and semi-open landscapes (Pain & Pienkowski 1997, Brambilla et al. 2010b).

The aim of this article is to investigate the importance of marginal habitat features for threatened species, by studying habitat preferences of Red-backed Shrikes Lanius collurio in Trentino, Italian Alps. The study region still hosts an appreciable population of Red-backed Shrikes; however, many habitat traits usually selected by this species are poorly represented, having been removed by intensification and/or abandonment in many areas. Therefore, our study region qualifies as an ideal site to assess the importance of residual, marginal features in depauperated landscapes.

Red-backed Shrikes are particularly sensitive to both agricultural intensification and land abandonment (Brambilla et al. 2007), and they appear to be tied to the marginal features accompanying cultivated or grazed areas (Brambilla et al. 2009, Brambilla & Ficetola 2012). The populations of this species underwent a strong decline between the 1970s and the 1990s, disappearing from large areas (Lefranc 1997), while a short-term slightly positive trend is reported from the Pan-European Common Bird Monitoring Scheme for the period 1990–2003 (PECBMS 2008). By investigating habitat preferences in a population inhabiting a landscape which has undergone impoverishment of those habitat resources that appear to be essential for Red-backed Shrikes, we hoped to assess the importance of marginal features. If such marginal features are key resources even if scarce, habitat selection in the study species should be driven by the availability of those features; otherwise, if marginal features lose their importance when severely depleted, habitat suitability would be affected by other elements.

METHODS

Study area

The Trento Province (N. Italy, 6206 km², elevation 67–3769 m asl) includes part of the Central and Eastern Italian Alps. The terrain is mainly mountainous (20 % is over 2000 m asl) with extensive woodland (56 % of the total surface, Provincia Autonoma di Trento 2007). Urbanized and cultivated areas are concentrated in the valleys. In the last 50 years the land surface area exploited for farmland activities has declined at middle–high elevation; at the same time, agricultural intensification at low elevation has led to an uniform landscape of high-intensity vineyards and orchards in large parts of the valleys. The area of meadows and pastures decreased from 148 769 ha in 1970 to 120 161 ha in 2000, while arable land declined from 7662 ha to 3678 ha (Provincia Autonoma di Trento, in Pedrini et al. 2005). Intensification in valley floors and abandonment at middle elevations are probably the main causes of the decline of birds on farmland in Trentino.

Some hundreds of Red-backed Shrike pairs (Pedrini et al. 2005) still breed in the remaining grassland areas, within a broad altitudinal range, with important sites in valley floors, middle elevations and uplands.

Fourteen study areas were chosen within the Trento province (6 in 2009; 8 in 2010). These were situated in locations where species occurrence was already known from large-scale breeding bird distribution data (Pedrini et al. 2005). Study areas included open and semi-open habitats, such as grasslands, pastures, herbaceous cultivations, but also portions of orchards, vineyards, small woodlots, sparse buildings. Study areas were delimited by discontinuities in the landscape, such as edges of large woodlands, large roads, urban habitats and rivers. Each area (range: 18.8–116.9 ha) was large enough to host potentially from a few to several breeding pairs of Red-backed Shrikes (Sutherland 2006). Study areas include all the important habitats for the species in Trentino (meadows, pastures, cultivated lands) in a broad altitudinal range (200–1800 m asl).

Fieldwork

Territory mapping

Each study area was surveyed at slow pace two to five times. The number of visits varied according to area, terrain, number of contacts with the target species. All contacts were plotted on detailed aerial photographs, recording all territorial and breeding behaviours indicating site occupancy; these included nests, male and female seen together, singing male, aggressive encounters, calling of juveniles (Sutherland 2006). Simultaneous observations of different pairs or males are particularly useful for the definition of the position and shape of the breeding territories of the species (Karlsson 2004; Brambilla et al. 2007, 2009).

Surveys were carried out in the morning (06:00–12:00 hours), between 1 June and 15 July in both years, to avoid possible recording of migrant (passage) birds which are still common in May in northern Italy (Brambilla et al. 2009). A lower number of individuals was observed during the first ten days of June of both years, probably because of later arrivals and/or settlements, which were also recently reported for other areas (Brambilla et al. 2009), and in accordance with the ringing data available for Italy (Spina & Volponi 2008).

Recording habitat variables

Thirty-two detailed habitat variables (Table 1) not available as geographic information system (GIS) layers were directly measured in the field. After territory mapping, the habitat variables were measured within 1 ha (100 m × 100 m square) territory and control squares; 1 ha approximates the average size of breeding territory of the Red-backed Shrikes in the study region (Brambilla et al. 2007, 2009). Squares were placed so as to cover most of the area defended and used by each pair. Within each study area, an equal number of control squares was randomly scattered over areas known to be unoccupied by the species and with tree cover lower than 50 % of the area; sites with a higher tree cover would be clearly unsuitable for the species, and were therefore unsuitable as control squares for modelling fine-scaled habitat preferences (Brambilla et al. 2009).

Table 1. Habitat variables measured in the field within the 100 m × 100 m territory and control squares.

Habitat variables	Characteristics
Mown grassland, Pastures, Grassland neither grazed nor mown,	Area covered (%)
Dry grassland, Maize, Cultivations, Woodland, Isolated bushes,
Isolated trees, Bramble bushes, Heath, Uncultivated land,
Lucerne, Vineyards, Orchards, Urbanized areas, Rocky areas, Bare soil,
Waterbodies, Wetlands, Trees plantations
Hedgerows, Tree rows, Paved roads, Unpaved roads, Fences	Area covered (%), length (m)

Habitat variables were measured during the last territory mapping sessions and during dedicated surveys (mid-July–August). Topographic variables (elevation, aspect and slope) were calculated in a GIS environment as mean values for each square by using the Digital Terrain Model of the Trento Province.

Statistical analysis

The data collected during 2009 were used to develop a first fine-scale model. The data set was simplified by reducing the number of factors. Habitat variables which occurred in less than 8 sample squares considering both territory and control squares were removed. These were: coverage of dry grassland, maize, cultivations, bramble bushes, heath, uncultivated land, lucerne, vineyards, orchards, urbanized areas, rocky areas, bare soil, waterbodies, wetlands, tree plantations and coverage and length of tree rows, paved roads, fences.

Habitat cover variables were square-root-arcsine transformed (Fowler & Cohen 2002). Aspect (a circular variable) was reclassified for obtaining a linear factor using a score reclassification: 330–360° and 0–30°: north; 30–60°: northeast; 60–90°: east-northeast; 90–120°: east-southeast; 120–150°: southeast; 150–210°: south; 210–240°: south-southwest; 240–270°: west-southwest; 270–300°: west-northwest; 300–330°: northwest (Brambilla et al. 2009).

A t-test was carried out for a comparison between mean values of habitat variables in territory and control squares.

To assess the relative importance of habitat variables on the probability of species occurrence within a sample square (modelled by comparing occupied versus unoccupied squares), three alternative statistical procedures were carried out and compared: stepwise binary logistic regression, multimodel inference, and hierarchical partitioning. We avoided including strongly intercorrelated factors (r > 0.6) in model selection. The factors which were indicated to be the most important according to the three analyses were used to build two models through logistic regression analysis (procedure enter in SPSS).

Logistic regression was carried out in SPSS for Windows 17.0 (Norušis 2009) using a stepwise backward procedure, which sequentially excludes non-significant factors (P > 0.1) on the basis of the likelihood ratio test.

Multimodel inference, based on an information-theoretic approach (Anderson et al. 2000, Anderson 2008), was performed on a set of 12 models. The set was built including 2–4 habitat variables in each model in such a way that every variable was considered at least once (Merli & Meriggi 2006). Akaike's Information Criterion corrected for small samples (AIC_C), differences between AIC_C values (Δ _i ) and Akaike weights w were calculated for each model in order to identify the ‘best’ models in the set. The area under the curve (AUC) of the Receiver Operating Characteristic (ROC) plots was used to assess the discrimination ability of the models (Pearce & Ferrier 2000). This value varies between 0.5 (discrimination ability no better than random) and 1.0 (perfect discrimination ability; Pearce & Ferrier 2000). Model averaging was carried out on the seven best models to assess the relative importance of the habitat variables. For each factor we calculated: the sum of the Akaike weights of the models which included the variable, the regression coefficient β weighted and averaged on the models and the unconditional sampling variance with 95 % confidence interval (Anderson et al. 2000).

Hierarchical partitioning analysis was used to check the results of the previous analysis (Brambilla et al. 2010a). This method was carried out in r 2.12 (R Development Core Team 2010) by means of hier.part package (Walsh & Mac Nally 2009) and allowed the estimation of both independent (I) and joint (J) explanatory power of each variable, considering all the 2^K possible models (Mac Nally 2000).

Checking the validity of models

The data collected in 2010 in eight new areas were classified according to the models obtained from the 2009 data. The discrimination ability of the models over the new data set was tested by calculating the AUC of the ROC plots.

Breeding densities

Breeding densities were calculated as number of breeding pairs per 10 ha per study area (Table 2). To assess the potential effect of habitat variables on breeding density, we first checked for possible correlations between breeding densities and each predictor (calculated as mean percentage value over both control and territory squares of each study area), and second, checked for an effect of the cumulative percentages of hedgerows, bushes and isolated bushes (summarized to obtain one factor representing all nesting and perching elements) on the breeding density. Finally, as we noted that areas with pastures hosted on average a higher number of marginal elements than areas without pastures (11.0 % versus 5.2 %; t = –2.74, P = 0.018), we checked for a difference in breeding densities also between areas with and without pastures.

Table 2. Breeding densities of Red-backed Shrikes in the 14 study areas in Trento Province (Italian Alps) in 2009 and 2010.

Study area	Surface (ha)	No. of pairs	Breeding density (pairs per 10 ha)
M.ga Brigolina	31.5	7	2.22
Dossi di Omalga	19.1	4	2.09
Vigolo Vattaro	24.6	4	1.63
M.ga Trenca	60.1	9	1.50
Foci dell'Avisio	28.5	4	1.40
Bordala	61.1	8	1.31
Roncegno	27.6	3	1.09
Malghèt	18.8	2	1.06
Celado	88.2	9	1.02
Barco	40.9	4	0.98
Santa Giuliana	41.2	4	0.97
Prada di M.te Terlago	116.9	10	0.86
Viote del Bondone	100.7	6	0.60
Torcegno	21.9	1	0.46

RESULTS

Thirty-three breeding pairs were recorded in 2009 in 6 study areas, and 42 in 2010 in the other 8 study areas. Comparing the mean values of habitat variables in territory and control squares, there was a significant difference with respect to non-grazed/mown grassland (t = –2.59, P = 0.014) and woodland (t = 1.55, P = 0.042) in the 2009 data set, and for non-grazed/mown grassland (t = –2.53, P = 0.014), isolated bushes (t = –4.03, P < 0.001), unpaved roads (t = 2.10, P = 0.041) and length of unpaved roads (t = 2.27, P = 0.027) in the 2010 data set. In territory squares the mean values of non-grazed/mown grassland and isolated bushes were higher than in control squares, whereas the opposite was true for the mean values of woodland and unpaved roads, which were lower in territory squares.

The output of stepwise binary logistic regression was the following equation (8th step of the backward stepwise procedure):

In the set used for multimodel inference, seven models had some reasonable level of support (Δ _i  < 10; Anderson et al. 2001; Table 3). The other five models had Δ _i >13 and w values near to 0, so they could not provide useful information. The difference between the two model groups was confirmed by AUC values (Table 3). The seven best models included the same variables selected by logistic regression, besides mown grassland and aspect.

Table 3. The set of 12 models used for multimodel inference ranked according to their AIC_c values. ngmg = non-grazed/mowed grassland, b = bushes, ib = isolated bushes, w = wood, mg = mowed grassland, h = hedgerows, a = aspect, e = elevation, ur = unpaved roads, s = slope, gg = grazed grassland, it = isolated trees.

Habitat variables	K	–2log-likelihood	AICc	Δ i	w	AUC
ngmg, b, ib w	5	69.68	80.68	0.00	0.58	0.82
ngmg, b, w	4	75.33	83.99	3.31	0.11	0.77
ngmg, w	3	78.15	84.54	3.86	0.08	0.73
ngmg, b, mg, ib	5	73.74	84.74	4.06	0.08	0.79
ngmg, b, h, w	5	73.83	84.83	4.15	0.07	0.78
ngmg, b, a, w	5	74.97	85.97	5.29	0.04	0.78
w, h, ib, b	5	75.31	86.31	5.63	0.04	0.78
w, h, ib, a	5	82.97	93.97	13.29	0.00	0.70
mg, b, w, e	5	84.05	95.05	14.37	0.00	0.70
mg, b, a, ur	5	87.33	98.33	17.65	0.00	0.64
mg, ib, it, e	5	87.95	98.95	18.27	0.00	0.59
gg, s, ur, b	5	88.72	99.72	19.04	0.00	0.61

Model averaging allowed us to determine the information of all models considered (i.e. the seven best models of the previous set in this specific case). Results showed a clear effect of four variables on the occurrence of the species: non-grazed/mown grassland, bushes and isolated bushes (positive effect), and woodland (negative effect; Table 4). The influence of hedgerows is less clear and less strong, because of the low averaged β and the confidence interval which includes 0. Values of the other two variables (mown grassland and aspect) are much lower; therefore, the variables with appreciable influence are the same as those included in logistic regression equation.

Table 4. Effect of seven habitat variables on the occurrence of Red-backed Shrikes as estimated from multimodel inference. The effect of aspect and mown grassland is negligible (confidence intervals of β encompass 0). The u.s. var. is the unconditional sampling variance.

Habitat variables	Σw	no. of models	β aver.	u.s. var.	–0.95	+0.95
Non grazed/mown grassland	0.99	6	6.90	1.48	4.01	9.80
Wood	0.92	6	–3.10	1.10	–5.24	–0.95
Bushes	0.91	6	3.15	1.21	0.79	5.52
Isolated bushes	0.69	3	5.29	1.90	1.57	9.01
Hedgerows	0.11	2	0.43	0.41	–0.37	1.23
Mown grassland	0.08	1	0.04	0.08	–0.12	0.20
Aspect	0.04	1	0.01	0.03	–0.05	0.06

Hierarchical partitioning analysis substantially confirmed the results of the previous analyses (Table 5). The five variables with higher independent contribution are the same as those included in the logistic regression equation. It was necessary to exclude two variables because with 11 factors the analysis produced a rounding error (a potential bias already mentioned in hier.part package guide, Walsh & Mac Nally 2009). Isolated trees and slope were excluded because of their low importance as pointed out by the previous analyses.

Table 5. Independent contribution (I) of single habitat variables on the occurrence of Red-backed Shrikes as estimated from hierarchical partitioning

Habitat variables	Independent contribution (I) (%)
Non-grazed/mown grassland	33.44
Bushes	18.67
Isolated bushes	13.01
Wood	12.48
Hedgerows	11.17
Grazed grassland	4.83
Mowed grassland	2.94
Orientation	2.89
Unpaved roads	0.57

On the basis of the outcome of the three different procedures, two models were developed: the first, including all the five variables which were indicated to be important, is the same as that found from the logistic regression analysis (see above), and had an AUC equal to 0.84 (good discrimination power). A second and more synthetic model was built excluding hedgerows, because of the lower importance resulted in multimodel inference:

with AUC = 0.82. Standardized residuals of both models approached a normal distribution.

Model validation

Applying the 2009 models to the 2010 data set, their discrimination capacity was found to be lower with respect to the original data (2009 data set). Therefore, the coefficients of the variables were re-calculated by applying the model to the entire data set. Two new equations were thus obtained:

and

The first model classified the 2009 data set with AUC = 0.82, and the 2010 data set with AUC = 0.82. The second classified the 2009 data set with AUC = 0.78, and the 2010 data set with AUC = 0.82. In both cases, standardized residuals approached a normal distribution. The five-variable model confirmed its better discrimination ability on the entire data set with a minimal difference between the 2009 and 2010 data sets.

Breeding densities and residual marginal elements

The five variables included in the best model occupied a very low area of the squares (Table 6), with the exception of non-grazed/mown grassland which had high percentages in some partially abandoned areas. The other elements with positive effect on the occurrence of the species (isolated bushes, bushes and hedgerows) which are used as breeding sites and hunting perches, had a very low cover in both territory and in control squares.

Table 6. Mean values ± standard error of the habitat variables included in the best model calculated for territory squares, control squares and the entire data set.

Habitat variables	Territory squares	Control squares	Entire data set
Non-grazed/mown grassland	16.79 ± 2.97	6.45 ± 1.90	11.62 ± 1.81
Wood	3.44 ± 0.77	7.25 ± 1.26	5.35 ± 0.75
Hedgerows	1.37 ± 0.48	0.67 ± 0.38	1.02 ± 0.31
Isolated bushes	3.00 ± 0.37	1.13 ± 0.24	2.07 ± 0.23
Bushes	5.49 ± 1.05	4.21 ± 1.04	4.85 ± 0.74

No significant correlation was found between bird breeding density and the single habitat variables (all P > 0.12), while the correlation between breeding density and the factor representing the sum of nesting and perching elements was marginally significant (r = 0.49, P = 0.079).

Consistent with the higher cover of marginal elements in pastures, breeding densities were marginally higher in study areas with pastures (mean = 1.52 pairs/10 ha) than without (mean = 1.01 pairs/10 ha; t = –2.06, P = 0.061) (Fig. 1).

Figure 1. Left: mean values (±1 se) of breeding density (pairs per 10 ha) of Red-backed Shrikes in study areas with (1) or without (0) pasture. Right: mean % values (±1 se) of hedgerows and bushes (sum) in study areas with (1) or without (0) pasture.

DISCUSSION

The fine-scale models obtained using three different approaches showed a good discrimination ability over the two completely independent data sets. They highlighted the importance of five habitat variables, four of which positively affected the probability of Shrike occurrence. Only one of the variables which are important for Shrikes, the non-grazed/mown grassland, was represented well within the whole area of the study. This habitat represents a source of invertebrate prey for Red-backed Shrikes (Tryjanowski et al. 2003, Casale & Brambilla 2009). Hedgerows, bushes and isolated bushes were included in the model of habitat preferences despite their scarcity in the whole area of the study (overall average cover = 7.94%), and all have a strong influence on the probability of the occurrence of Red-backed Shrikes according to the regression coefficients of the best ‘final’ model. Results showed that even a very limited availability of these marginal features is enough to influence the occurrence of breeding pairs of Red-backed Shrikes. Considering the surveyed areas, portions with bushes and hedgerows represent islands of suitable sites scattered within wider grassland areas (which are also islands at a different scale, given that in the region of the study below 2000 m asl the land is mainly covered by woodlands and intensively managed orchards and vineyards). The scarcity of bushes and hedgerows is the likely reason for the rather low breeding density. Breeding densities were higher in pasture areas, where the availability of the key habitat factors was significantly higher; grazed habitats often qualify as optimal habitats for Red-backed Shrikes, because of high food and nesting sites availability (Lefranc 1997; Brambilla et al. 2007). Similar studies carried out in a neighbouring region (Lombardy) reported much higher cover values of shrubs and hedgerows in both territory and control plots for Red-backed Shrikes (Brambilla et al. 2009, 2010b) in otherwise comparable habitats (same elevation and land use), with associated much higher breeding densities of Red-backed Shrikes (up to 10 pairs/10 ha; Casale & Brambilla 2009). This indirectly suggests how scarcity of these habitat features may actually limit breeding densities in our study areas.

These findings have potentially important conservation implications, as they suggest that even limited (and not expensive) efforts to ensure a minimal cover of these key factors in farmland areas (by maintaining or restoring bushes and hedgerows) could have positive impact on the density of Red-backed Shrikes. These findings suggest that even highly simplified grasslands could become attractive to Red-backed Shrikes with suitable management of these marginal habitat features. The importance of grassland areas with untilled herbaceous vegetation was also confirmed. Red-backed Shrikes positively respond to the introduction of management promoting untilled herbaceous vegetation (Casale & Brambilla 2009), and this would also have positive effects on the occurrence of many other farmland species (Casale & Brambilla 2009).

In conclusion, our work further confirms the importance of marginal habitat elements in farmed landscapes, suggesting that they can play a crucial function even though they may be scarce. We suggest that the maintenance or increase of such habitat elements should be encouraged and promoted also in depauperated landscapes.

ACKNOWLEDGEMENTS

We are grateful to C. Tattoni, F. Rizzolli and M. Segata for kind help with the work and to V. Fin and L. Sottovia for support. The present work has been partly financed by the Autonomous Province of Trento in the context of the Rural Development Programme and of the Natura 2000 network.