The study was carried out in Gran Paradiso National Park (45°31′ N, 7°16′ E, north-western Italian Alps; Fig. 1) in June–July 2001 and 2002. The park area extends for about 70 000 ha. The bottom of the valleys is covered by mixed-broadleaved woods (Castanea sativa, Quercus spp., Betula alba, Populus spp., Fraxinus spp., Acer spp.), whereas coniferous woods of European larch Larix decidua mixed with Norway spruce Picea abies, arolla pine Pinus cembra and, occasionally, silver fir Abies alba dominate from around 1000 m a.s.l. to the treeline (1600–2200 m a.s.l). Progressing up the slopes, trees are replaced by a shrub belt (Juniperus, Rhododendron, Vaccinium, Arctostaphylos, Salix spp.), wide alpine pastures and natural grassland dominated by Festuca and Nardus spp. Above 3000 m a.s.l., rocks, screes and snowbeds dominate the alpine landscape, up to the highest peaks and glaciers of Gran Paradiso massif (4061 m a.s.l). Non-woody vegetation below the timberline consists mainly of semi-natural grasslands dominated by grasses (Gramineae) and green alder Alnus viridis shrub growing on disturbed sites (wet and steep slopes, avalanche tracks).
Figure 1. Location of the study area and plots (solid circles); water courses are also shown. The distribution of plots matches that of pastures and abandoned pastures (alpeggi) in the study area; the central zone was under-sampled because it includes bare rocks and glaciers above 3000 m a.s.l.
Download figure to PowerPoint
In the study area the decline of stocking levels and summer transhumance activities has precipitated a marked reduction in the extent of grassland below the treeline (Fleury et al. 2001). In abandoned pastures, shrubs such as Berberis, Rhododendron, Juniperus, Alnus, Salix spp., bracken Pteridium aquilinum and European larch are encroaching rapidly from field borders and colonizing previously open habitats (Cavallero et al. 1997). In pastures that are still exploited, grazing pressure is normally weak, as it occurs for very short periods: 1–2 (sometimes 3) months in July–September
bird survey and quantification of local habitat structure, grazing pressure and landscape parameters
The grasslands used in this study covered a wide range of altitudes (1000–2800 m a.s.l) and differed greatly in topography, size (from less than 1 to 1200 ha) and degree of shrub and tree cover. In such heterogeneous landscapes, traditional survey methods such as fixed radius point counts or transects are inappropriate, as differences in bird detectability would bias intersite comparisons. We therefore used a standardized area count method (Bibby et al. 2000), surveying birds in circular plots of radius 50 m. Counts lasted 15 min. During the first 5 min of the recording period the observer stood still and quiet at the centre of the plot, as in a standard point count, while in the latter part of the count the observer moved around and stopped at suitable vantage points to look and listen, recording all birds seen or heard within the plot. This method is particularly useful when comparing community structure of habitats that differ in vegetation density. Each census plot was visited twice (in June and July of the same year) and the largest values from the two censuses were used as a measure of bird species abundance per plot; overall, 350 plots were visited on two occasions (Fig. 1). Each day a 5–18-km long transect was walked, in an ascent of 800–1500 m. Plots along daily transects were located 200–6000 m away from each other; we were very careful not to record the same individuals on neighbouring plots by tracking birds with binoculars if they were flushed off.
For each plot we calculated the following five variables: vegetation (grass and shrub) height (H, mean value of 40 measurements per plot taken with a wooden dowel subdivided into 1-cm units; in grazed pastures these were taken prior to grazing); heterogeneity of vegetation height [CV = H/SD (H) × 100]; percentage shrub cover; percentage tree cover; and percentage boulder/stone cover (percentages estimated by eye). Grazing pressure was estimated, after interviewing local people and consulting the archives of the park as: 0, abandoned pasture where grazing no longer occurred; 1, low to moderate grazing intensity (grasslands were either rapidly crossed by sheep/cows or grazed by free-ranging livestock); 2, high grazing intensity (high stocking levels, with cattle kept in fenced pastures until the whole area was grazed); 3, very high utilization (haymaking associated with grazing, so that the herb layer was kept uniformly short). In the interest of clarity, the terms ‘high’ and ‘very high’ have a relative value, as the pastoral activities in the study area were largely extensive. Two landscape variables were calculated from land cover data in the Gran Paradiso GIS database (digitized from 1 : 10 000 aerial photographs) using ArcView (version 3.1, ESRI, CA): the amount of contiguous grasslands (total grassland area) around each census plot and the distance between each plot and the nearest woodland (nearest-neighbour distance from woodland).
Plots on each transect that were close together were more likely to be similar in bird community composition than those far apart. Accordingly, we tested for plot autocorrelation within transects with Moran's I coefficient (Moran 1950), a weighted correlation coefficient used to detect departures from spatial randomness. This index varies between −1 and 1, and a high value indicates positive autocorrelation between plots along a transect. The autocorrelation between the study plots was low (mean I=−0·02, range −0·40–0·20, n= 70 daily transects). Monte Carlo tests with 100 permutations of the data set were used to test for the significance of spatial autocorrelation coefficients via randomization, allowing us to determine if the observed coefficients were significantly different from a random pattern. In all cases, probability levels were > 0·05, suggesting that plots were spatially independent. Moran's I and Monte Carlo randomization were performed with RookCase Software (Sawada 1999).
We classified bird species into four major ecological groups: open habitat–grassland species (hereafter grassland species, species that require open fields both for breeding and foraging); ecotone species (species that use grassland and woodland alternatively); shrub species (species that dwell in shrubby areas); and woodland species (species typical of forest and open forest habitats) (see the Appendix).
Bird community structure at each individual plot was expressed in terms of diversity (Shannon index: H′ = –∑ pi × ln pi, where pi is the relative frequency of species i) and abundance of the four ecological groups. Results obtained by considering species richness per plot were similar to those achieved using Shannon diversity; hence only the latter index was considered here, as a measure of avian α-diversity.
To reveal patterns in the local habitat structure and compensate for multicollinearity we summarized habitat attributes with a principal components analysis (PCA) on standardized data (zero mean and unit SD). PCA condensed the original information on three derived axes (PC1, PC2 and PC3) having the benefit of being orthogonal and uncorrelated (Jongman, ter Braak & Van Tongeren 1995). This gave three main factors with eigenvalue > 1 that explained 73·5% of the variability of the data set. Percentage shrub cover and vegetation height showed the highest correlation with PC1 scores (factor loadings R= 0·87 and 0·80, respectively); PC2 was positively correlated with percentage stone cover (R = 0·78) and negatively correlated with tree cover (R = −0·66), whereas heterogeneity of vegetation height provided the major loading on PC3 (R = 0·97). Hereafter PC1, PC2 and PC3 are termed SHRUB-HEIGHT, STONE-TREE and VEGHETERO, respectively.
Mean diversity and abundance of ecological groups were used as dependent variables in stepwise multiple regression analyses using (i) local habitat variables (three PCA scores), (ii) landscape variables, (iii) grazing pressure levels and (iv) elevation as quantitative predictors. To attain a normal distribution, these variables were log-transformed (y = log(x + 1)). Because this transformation proved to be ineffective with the variables bird diversity, abundance of shrub species, abundance of ecotone species and abundance of woodland species, in these four cases normality was achieved through Box–Cox transformation [y = (xλ − 1)/λ, with λ= 0·92 for bird diversity and abundance of woodland species, λ= 0·98 for the abundance of shrub species and λ= 0·50 for the abundance of ecotone species; Box & Cox 1964].
Before carrying out these regressions, we further tested for correlations between the variables. Grazing pressure proved to be negatively correlated with both elevation (R = −0·31) and SHRUB-HEIGHT (R =−0·30; with 9% of the variability in SHRUB-HEIGHT being explained by grazing). To control for the effects of vegetation structure and topography on grazing, we used the residuals of the regression between grazing levels vs. elevation and SHRUB-HEIGHT rather than raw values in all the regression models.
In keeping with Herrando & Brotons (2002), we used a parsimonious approach assuming that bird diversity and abundance were affected primarily by elevation, local habitat and grazing pressure, and then by the landscape. A stepwise multiple regression analysis was carried out in two steps. First, bird diversity and abundance of the four ecological groups were used as dependent variables and (i) elevation, (ii) SHRUB-HEIGHT, (iii) STONE-TREE, (iv) VEGHETERO and (v) residuals of grazing pressure (partialling out the effects of elevation and SHRUB-HEIGHT) were used as predictors. Secondly, we used the residuals of the former regression as the dependent variable and (i) total grassland area and (ii) nearest-neighbour distance from woodland as the predictors.
Differences in environmental variables (local habitat and landscape) and bird community among the four grazing levels were also tested by means of one-way anovas on log- or Box–Cox-transformed data (in the case of environmental variables, bird diversity, abundance of grassland, edge, shrub and woodland species) and χ2 tests of independence (occurrence of the most common species); in this analysis the altitudinal ranges of 1000–1900 (montane belt) and 1900–2800 (subalpine and alpine belt) m a.s.l. were considered separately.
In order to verify the accuracy of our approach, one-sample Kolmogorov–Smirnov tests were always carried out to test whether the residuals of previous models fit a Gaussian distribution: in all the cases Kolmogorov–Smirnov d was < 0·06, and P > 0·10, suggesting that all residuals were still normally distributed.