Raffaele Lafortezza, Department of Scienze delle Produzioni Vegetali, University of Bari, Via Amendola 165, Bari 70126, Italy. E-mail: firstname.lastname@example.org
Aim Few studies have attempted to assess the overall impact of fragmentation at the landscape scale. We quantify the impacts of fragmentation on plant diversity by assessing patterns of community composition in relation to a range of fragmentation measures.
Location The investigation was undertaken in two regions of New Zealand – a relatively unfragmented area of lowland rain forest in south Westland and a highly fragmented montane forest on the eastern slopes of the Southern Alps.
Methods We calculated an index of community similarity (Bray–Curtis) between forest plots we regarded as potentially affected by fragmentation and control forest plots located deep inside continuous forest areas. Using a multiple nonlinear regression technique that incorporates spatial autocorrelation effects, we analysed plant community composition in relation to measures of fragmentation at the patch and landscape levels. From the resulting regression equation, we predicted community composition for every forest pixel on land-cover maps of the study areas and used these maps to calculate a landscape-level estimate of compositional change, which we term ‘BioFrag’. BioFrag has a value of one if fragmentation has no detectable effect on communities within a landscape, and tends towards zero if fragmentation has a strong effect.
Results We detected a weak, but significant, impact of fragmentation metrics operating at both the patch and landscape levels. Observed values of BioFrag ranged from 0.68 to 0.90, suggesting that patterns of fragmentation have medium to weak impacts on forest plant communities in New Zealand. BioFrag values varied in meaningful ways among landscapes and between the ground-cover and tree and shrub communities.
Main conclusions BioFrag advances methods that describe spatial patterns of forest cover by incorporating the exact spatial patterns of observed species responses to fragmentation operating at multiple spatial scales. BioFrag can be applied to any landscape and ecological community across the globe and represents a significant step towards developing a biologically relevant, landscape-scale index of habitat fragmentation.
There are key disconnects between field ecologists and remote-sensing researchers in the ways that they assess and quantify the degree, or intensity, of fragmentation effects. Remote-sensing researchers and landscape ecologists have developed tools, such as the well-respected fragstats package (McGarigal & Marks, 1995), which enable them to calculate numerous metrics of fragmentation from maps indicating the position, size and shape of fragments within a landscape (O'Neill et al., 1988; Riitters et al., 1995). Yet these summary spatial statistics provide little information on the biological effect of the fragmentation pattern that is observed (Davidson, 1998). By contrast, field ecologists routinely measure the abundance of species or the structure of biological communities at point locations within fragmented landscapes and then relate these measures to metrics of habitat fragmentation (Lindenmayer & Fischer, 2007). Typically, field ecologists will focus on biological responses to one or a few attributes of the fragments or landscape such as area (Watling & Donnelly, 2006), edge (Laurance et al., 2002), shape (Saura & Carballal, 2004), isolation (Schmiegelow et al., 1997), landscape forest cover (Trzcinski et al., 1999) or matrix quality (Baum et al., 2004; Watling & Donnelly, 2006). Biological studies rarely attempt to explicitly quantify the landscape-level significance of the effects observed in their sampling plots. Typically, authors stop once they have explained the ecological patterns of species responses to habitat loss and fragmentation, rather than trying to scale up those point-based observations to estimate the overall effect of fragmentation on the landscape as a whole. This leaves the field in a position where it is possible to determine whether fragmentation does or does not have an impact on biodiversity in a given landscape, but unable to quantify the magnitude of that impact for that landscape. For example, the use of point data can be used to assess species responses to habitat edges and to show that edges do have an impact on species abundances (Ewers & Didham, 2008), but it is only possible to quantify the net impact on the populations of those species when species responses are combined with spatially explicit data on the distribution of habitat edges (Ewers & Didham, 2007; Ewers et al., 2009). It follows that more research is needed to develop robust methods for extrapolating biological data, collected at the plot scale, to the wider patterns of fragmentation observed with remote sensing and geographical information system (GIS) techniques routinely applied at the landscape level. Such methods would build on the current ability of researchers to relate observed biological changes to remotely sensed spatial patterns of fragmentation, and take them a step further to quantify the net impact of habitat fragmentation on landscapes. The end result would be a biologically relevant index of habitat fragmentation that directly quantifies the net impact of fragmentation on a measure of biodiversity across an entire landscape.
Biologically relevant indices of habitat fragmentation would find immediate uses in conservation biology, and are also needed to help support policy at global and other scales. The Convention on Biological Diversity has identified forest fragmentation as one of a suite of indicators for tracking progress towards its ‘2010 target’ on reducing biodiversity loss (UNEP/CBD, 2002, 2004) and policy-makers in the UK identified the development of such indicators as being of key importance (CBD Secretariat, 2001; Sutherland et al., 2006).
In this paper, we examine the impacts of fragmentation on plant communities in New Zealand forests. We develop and describe a new method for converting point-based observations into a continuous biodiversity surface representing the similarity of fragmented biological communities to those in unfragmented forest. We apply the method to two forest landscapes in New Zealand, one mountainous and naturally highly fragmented and the other lowland and fragmented to a much lesser extent. Using a multiple nonlinear regression technique that accounts for spatial autocorrelation effects, we analyse plant community composition in relation to various metrics of fragmentation. By extrapolating the best-supported regression model in a GIS, we predict a Bray–Curtis similarity value for each forest pixel within the two landscapes and use these mapped outputs to calculate a landscape-scale metric of the effects of fragmentation on biological diversity. Our method scales up data on biodiversity observed in tree plots into a landscape-level statistic reflecting the net biological impact of fragmentation.
We use this statistic to test three hypotheses. First, we hypothesized that the landscape-level impact of fragmentation on plant communities in the highly fragmented mountainous region would be greater than on the plant communities in the relatively intact lowland region. This expectation arises because the more compact shape and larger size of patches in relatively unfragmented landscapes should ensure that populations in those landscapes suffer relatively little impact from forest fragmentation (Ewers & Didham, 2007). Our second hypothesis was that ground-layer vegetation in these forests would be more influenced by fragmentation than the tree and shrub layer. This expectation arises from the difference in generation time of ground-layer plants and trees that will result in different time lags before the impacts of forest fragmentation on communities become discernible (Ewers & Didham, 2006). Deforestation and fragmentation in the study landscapes occurred less than one generation ago for many of the tree species in the region (Ewers et al., 2006), so relatively little time has elapsed in which those species and communities may have changed their spatial patterns. By contrast, ground layer plants with relatively fast generation times are expected to have more rapidly altered their patterns of community composition with respect to the changed environmental conditions in fragmented landscapes. Finally, we hypothesized that exotic plants are important components of communities growing near forest edges and would have a strong influence on the (dis)similarity between communities at forest margins and interior forest. We based this prediction on literature documenting the invasion of native forests in New Zealand by exotic species that has been facilitated by the creation of forest edges (e.g. Wiser et al. 1998; Standish, 2002).
DATA AND METHODS
Study landscapes and forest survey data
The impact of fragmentation on plant diversity was modelled by analysing changes in species composition in two contrasting regions of the South Island of New Zealand (see Fig. 1 and Table 1). The first landscape, located in south Westland, represents one of the least fragmented areas in New Zealand (Ewers et al., 2006). This landscape stretches from sea level to an altitude of 1132 m (mean 240 m) (Table 1) with forests dominated by the conifers Dacrycarpus dacrydioides (A. Rich.) Laub. growing on alluvial flats close to rivers, merging into Dacrydium cupressinum Sol. ex Lamb. on older terraces. Other species include Prumnopitys ferruginea (D. Don.) deLaub. (Podocarpaceae) and the angiosperm Weinmannia racemosa L. f. (Cunoniaceae) in the lowlands, which merge into Nothofagus forests on the slopes (Wardle, 1991; Duncan, 1993). Variation in the structure and composition of these forests is mainly determined by natural disturbances, probably including catastrophic earthquake events such as those observed further north on the west coast (Wells et al., 2001), rather than human-induced impacts. However, humans have altered the distribution of forest cover in the lowlands through agricultural expansion leading to the creation of non-natural forest fragments and edges embedded within a pasture matrix.
Table 1. Summary characteristics of the two study landscapes at plot and landscape level.
Lowland rain forest
Species present > 1% of plots.
106 species in common with tree and shrub species.
Mean similarity (BC) of ground-layer data (95% CI)
Mean (95% CI) similarity (BC) of tree and shrub layer data
Number of exotic species per plot (95% CI)
By contrast, the second landscape comprises highly fragmented forests on the drier side of the Southern Alps (Ewers et al., 2006). The dominant tree species in this region is Nothofagus solandri, consisting of subspecies solandri (Hook. f.) Oerst. (black beech) below 600 m altitude and subspecies cliffortioides (Hook. f.) Poole. (mountain beech) at higher altitude (Wardle, 1984). The lower boundaries of the forest extent are mostly human-made, usually as a result of pre-European (800–200 years bp) and post-European (0–200 years bp) burning. In some areas the hill-slope forests have also been cleared by fire and now provide rough grazing for sheep (Wardle, 1984). However, these forests are also heavily fragmented by natural processes because the steep mountain slopes are divided by frequent scree slopes and give way to alpine grasslands at an altitude of around 1300 m.
In the two regions, forest survey data were drawn from the New Zealand National Vegetation Survey Databank. In particular, we used data collected in the period 1979–85, using a standard reconnaissance method (Allen & McLennan, 1983; Allen, 1992) in which a visual estimate was made of the percentage of foliage cover of all vascular plant species within each of six height tiers. The tiers represented ground-cover species and any seedlings of shrubs or trees (T6, < 0.30 m), shrub (T5, 0.3–2 m), lower subcanopy (T4, 2–5 m), upper subcanopy (T3, 5–12 m), canopy (T2, 12–25 m) and emergent above-canopy vegetation (T1, > 25 m) (Miller, 2004). Cover was estimated by cover classes of: < 1, 1–5, 6–25, 26–50, 51–75 and 76–100% (Allen, 1992). All forest survey data were collected in 20 × 20 m plots. A total of 7799 forest plots were used in this study: 4637 plots in south Westland and 3162 plots in the montane Nothofagus-dominated forest.
Calculation of fragmentation metrics
To characterize the fragmentation patterns within these two landscapes, we used indigenous forest cover data stored in the New Zealand Topographic Database (NZTopo, http://www.linz.govt.nz/topography/topographic-data); this database provides information on a wide range of land features, including forest vegetation boundaries in vector format, at a scale of 1:50,000. The data were derived from aerial photographs taken mostly in the 1970s and 1980s (black and white, scale 1:25,000), which is roughly consistent with the period during which plots were surveyed in the two regions. Vector polygons were converted into 25 × 25 m grid cells. We defined the landscape for analysis as being the land surface located within a 5000-m buffer around the survey plots.
Many metrics of landscape fragmentation are closely correlated with one another (Riitters et al., 1995; McGarigal & Cushman, 2002; Neel et al., 2004) because they are based on related properties of fragmented landscapes such as fragment area, edges, shape, isolation and the characteristics of the surrounding land cover (Hargis et al., 1998; Griffith et al., 2000). We selected nine fragmentation measures that previous studies have reported to be relatively uncorrelated with each other (Table 2) (Cain et al., 1997; Trani & Giles, 1999; Honnay et al., 2003; De Clercq et al., 2006), although we found that in our study areas some of these measures were still significantly inter-correlated (see Tables S1 & S2 in Supporting Information). The nine measures were grouped into component-based metrics, which are based on patch-level calculations, and landscape-based metrics, which examine the attributes of neighbouring cells around a focal cell (see Table 2 for more details). Metrics were calculated from the rasterized forest-cover data (cell size = 25 m) using fragstats v.3.3 (McGarigal & Marks, 1995). Landscape-level metrics (termed ‘class-level metrics’ in fragstats) were calculated using neighbourhood radii of 500 m and 1000 m.
Table 2. Fragmentation measures used in this study.
Type of measure
Landscape-based metrics were computed using a moving window of radius of 500 m and 1000 m.
The same description applies to each measure computed within a radius of 1000 m (e.g. PD.1000).
The shape index: describes the deviation of each forest patch from circular: a circle has a value of 1 whereas forest fragments with irregular shapes will have higher values.
Metrics have been grouped into component-based metrics, which are based on patch-level calculations, and landscape-based metrics, which examine the attributes of neighbouring cells around a focal cell. A detailed description of each metric can be found in McGarigal & Marks (1995).
Average distance to the nearest neighbouring forest patches
ENN.500 ≥ 0
Calculation of similarity indices
Our analyses are based on the premise that community composition within control plots, located deep inside continuous indigenous forest, is representative of the community that would have occurred in the absence of forest fragmentation, and that community composition will be more similar between control sites than between a control site and a location that is affected by forest fragmentation (Ewers et al., 2009). To assess the similarity of communities at different point locations, we used the Bray–Curtis index (BC), one of the most common measures of community similarity in the ecological literature (Magurran, 2004):
where Yij is the abundance of species i in site j, Yik is the abundance of species i in site k, and the summation is over all species found at the two sites. BC values range from 0 (no species in common) to 1 (identical abundance of all species). Using vegetation cover scores as a proxy for abundance (Duncan et al., 1997; Wiser & Buxton, 2008), we calculated the average BC between each plot that we regarded as being potentially affected by fragmentation (henceforth PFA plots) and the nearest control plot that differed in elevation by less than 100 m from the PFA. To avoid rare species greatly skewing BC estimates, we used only species that were present in more than 1% of plots (Tables 1 & S3). In the case of the lowland rain forest (south Westland), plots less than 1500 m from the nearest forest edge were initially selected as PFA plots on the basis of previous studies investigating the spatial scale of edge effects (Ewers & Didham, 2008). Using this approach, 4564 out of 4637 plots were designated as PFA plots and the remaining 73 plots were denoted as control plots (0.95 quantile = 1170 m). To validate our arbitrary choice of 1500 m, we plotted PFA–control plot similarity against the distance to the forest edge and examined the shape of the curve (Fig. 2). For the lowland rain forest, we found that the value of edge distance at which BC reached 0.99 of the asymptote value was 1482 m. Consequently, we considered our choice of control plots as appropriate for the scope of the study. For the Nothofagus-dominated landscape, BC reached 0.99 of its asymptotic value at an edge distance of 1027 m, so we therefore decided that 1000 m was an appropriate threshold distance from the edge for distinguishing control plots (n= 465) from PFA plots (n= 2697).
In both landscapes, we calculated similarity values for two layers of plants: the ground layer (T6), consisting of angiosperm herbs, fern species and seedlings of woody species, and an amalgamation of all other canopy layers (T1–T5) consisting mostly of woody angiosperm and conifer species but also tree ferns and other ferns such as Blechnum discolor. If a woody species appeared in more than one of the T1–T5 layers then we used the maximum cover score observed. The classification of control and PFA plots was based on species composition data from the ground-cover layer (T6). In addition we ran a set of analyses in which exotic species were excluded, to test the hypothesis that exotic species invading from forest edges were contributing substantially to the patterns of plot similarities observed (Wiser et al. 1998).
Multiple non-linear regression on similarity indices
Because BC is a proportional value, we arcsin-square-root transformed it before using it in regression analyses as recommended by Sokal & Rohlf (1995). When back-transformed, all BC values that we predicted from the final regression relationships were bounded between 0 and 1. Preliminary inspections of plotted against the selected fragmentation metrics revealed nonlinear relationships that could be effectively modelled using a three-parameter asymptotic exponential function (e.g. Fig. 2):
where x is a fragmentation metric, and a, b and c are parameters estimated by least-squares regression. Parameter a is the asymptotic value that tends toward when comparing two plots deep in the interior of the forest (i.e. two control plots), while a−b is the similarity when x= 0 (for example, on the edge of a patch when x is distance from the forest edge). Parameter c describes the steepness of the response curve.
We extended this approach to nonlinear multiple regression, allowing us to model the effects of several fragmentation metrics by expanding on the single-variable function:
Here x1 and x2 are two fragmentation metrics (more could be included) and a, b1, c1, b2, c2, b3 and c3 are the parameters to be estimated by regression. The final term in the model, DIST, is the distance between the PFA and control plots and is included to explicitly account for the effects of spatial autocorrelation, based on the recognition that communities located far apart in space are likely to be less similar in composition than communities located close together, and that this effect could potentially confound our estimates of the biological impacts of forest fragmentation.
Nonlinear multiple regression models were fitted using the nls function in R v.2.6.2 (R Development Core Team, 2004). We started by entering each fragmentation metric into the model as a single term (alongside an autocorrelation term) and noted which of the alternative models had the lowest Akaike information criterion (AIC) (Burnham & Anderson, 2002). We then constructed a series of models containing two fragmentation measures, one of which was the term from the best-supported single-term model, the other being another fragmentation metric. We took the best-supported two-term model and attempted to construct more complex models, containing three or more fragmentation indices. We followed this approach because models with many terms often failed to converge, so building complex models from simpler ones in a forward stepwise manner was the most pragmatic approach. For each combination of landscape × tier × full or native-only community subset, we selected the best set of three candidate models, as defined by AIC model selection criteria, and calculated Akaike weights (wi) to determine the probability that any given model was the ‘best’ of the candidate set (Burnham & Anderson, 2002). The degree of support for the final models was established by comparing their AIC values with those of ‘null’ models containing only an intercept term and a spatial autocorrelation term.
Generating similarity maps and calculating landscape-level fragmentation impacts
We used the best-supported regression model to generate continuous GIS surfaces for each landscape, representing the predicted similarity of each grid cell to control sites deep within the forest. For example, if the modelling identified that distance from edge (DIST.EDGE) was the only significant explanatory variable (over and above the effects of spatial autocorrelation), then for a grid cell with coordinates [x,y] the predicted value would be
where DIST.EDGExy is the distance from the forest edge to grid cell [x,y]. These predictions were used to generate a map in which all grid cells designated as control sites (based on their distance from forest edges) have the same predicted BC value, equal to , representing the average similarity between interior plots that are DIST metres apart. When making our predictions, we assigned DIST the value of the mean inter-plot distance for all PFA–control sites in each dataset. In the regression analysis (equation 3), this variable was computed as the distance between any given pair of PFA-to-control plots, and so has a specific value only for pixels that can be paired with a control pixel. When predicting BC values for new pixels, the location of an appropriate ‘control’ pixel for those new pixels is not clearly defined, so an exact value of DIST cannot be estimated directly from a GIS. Consequently, we used the mean DIST value taken from all PFA-to-control plot pairs in each dataset: 30.3 km (lowland rain forest landscape; range 0.1–143.9 km) and 39.7 km (Nothofagus forest landscape; range 0.1–217.5 km). This approach effectively treats each pixel in the landscape as being the same distance from a control plot, thereby standardizing for the potentially confounding effect of spatial autocorrelation on our predictions of BC values. Sensitivity analysis showed that our final estimates of fragmentation impacts were insensitive to our choice of DIST, which influenced our estimates of the impact by less than 1% and had no discernible influence when DIST was greater than approximately 10 km (Fig. S1). The resulting maps of predicted BC values represent the extent to which the community in each grid cell is expected to differ from that of the control sites.
Finally, we computed a landscape-scale indicator to quantify the overall effects of fragmentation on species communities:
where the mean value of pred(BCxy) is computed over all forested grid cells. If there is no biological effect of fragmentation, then the best supported regression model would predict that all grid cells in the PFA sites and control sites would have BC values equal to BCmax and BioFrag would have a value of 1. By contrast, BioFrag approaches 0 when fragmentation generates very strong effects that have resulted in a complete turnover in the composition of biological communities. In reality, BioFrag can never actually be 0, because there must be at least one control site within the landscape from which to estimate the similarity of PFA, and the presence of these control sites ensures the value of the numerator will always be greater than 0.
Landscape-level variation in α-diversity
In addition to calculating the BioFrag index, which analyses changes in community composition among forest locations (i.e. β-diversity), we assessed the effects of fragmentation on species richness within communities (i.e. α-diversity). We avoid the inclusion of rare species by considering only those species that were present in more than 1% of plots. In both regions, we computed the number of species surveyed in each plot. We fitted a linear model using generalized least squares (gls) using the nlme package in R v.2.6.2 and determined the extent to which α-diversity was explained by fitting a full model containing a linear combination of all nine fragmentation measures, then using model simplification procedures (AIC) to produce a minimal model containing only significant terms. The gls function allowed us to model the effects of spatial autocorrelation on α-diversity using a spherical correlation structure (corSpher class) (R Development Core Team, 2004).
Regression analysis of similarity indices
A preliminary screening of similarity indices for all analyses in response to the nine fragmentation measures indicated a positive correlation with distance from edge (DIST.EDGE) and patch size (P.SIZE), meaning the similarity between plant communities in PFA and control plots increased with increasing values of the distance to the nearest forest edge (Fig. 2) and/or with fragment size. By contrast, similarity indices were negatively correlated with other fragmentation measures such as the percentage of non-forest (PNF), edge density (ED) and patch density (PD).
In the lowland rain forests of the Westland district, the most plausible model for ground-cover species (T6) included distance from edge (DIST.EDGE), percentage of non-forest (PNF.1000) and edge density (ED.1000) as explanatory variables (Tables 3 & 4). For trees and shrub species (T1–T5), the best-supported model included distance from edge (DIST.EDGE) and percentage of non-forest (PNF.500) (Tables 3 & 4). In the Nothofagus-dominated landscape, the ground-cover species were best modelled with patch size (P.SIZE) and edge density (ED.500), whereas for species in T1–T5 the best-supported model included distance from edge (DIST.EDGE) and nearest neighbour distance (ENN.500; Tables 3 & 4).
Table 3. Model selection statistics for candidate models predicting changes in community composition in two New Zealand landscapes that vary in terms of fragmentation history and intensity.
Models were fitted by nonlinear least square regression and selected using the Akaike information criterion (AIC). Models are ranked from worst to best fitting, with the best-fitting model highlighted in bold; wi are the Akaike weights computed considering the difference in AIC between each model and the best-fitting model in the candidate set. Abbreviations of variable names are described in Table 2 and full equations for the best-fitting models are presented in Table 4.
(A) Lowland rain forest
Native + exotic
DIST.EDGE + PNF.1000
DIST.EDGE + ED.1000
DIST.EDGE + PNF.1000 + ED.1000
Tree + shrub species (tiers 1–5)
DIST.EDGE + MSI.1000
DIST.EDGE + PNF.1000
DIST.EDGE + PNF.500
DIST.EDGE + PNF.1000
DIST.EDGE + ED.1000
DIST.EDGE + PNF.1000 + ED.1000
Tree + shrub species (tiers 1–5)
DIST.EDGE + MSI.1000
DIST.EDGE + PNF.1000
DIST.EDGE + PNF.500
(B) Nothofagus forest
Native + exotic
P.SIZE + PD.500
P.SIZE + ED.1000
P.SIZE + ED.500
Tree + shrub species (tiers 1–5)
DIST.EDGE + ENN.1000
DIST.EDGE + ENN.1000
DIST.EDGE + ENN.500
P.SIZE + PD.500
P.SIZE + ED.1000
P.SIZE + ED.500
Tree + shrub species (tiers 1–5)
DIST.EDGE + ENN.1000
DIST.EDGE + ENN.1000
DIST.EDGE + ENN.500
Table 4. Regression models explaining variation in plant community composition for two fragmented landscapes in New Zealand.
ΔAIC = difference between the Akaike information criterion (AIC) value for the current model and value for a null model incorporating spatial autocorrelation effects only: y=a−b2 exp(−c2 DIST).
Analyses were run separately for ground cover and tree + shrub communities, and for the full community and for the community with exotic species excluded. The response variable for all models is arcsin-square-root transformed Bray–Curtis values (equation 3). The last column reports the value of the BioFrag index for each of the landscape × community combinations that were analysed. Codes for all variables in the regression models are explained in Table 2.
All models showed significant effects of forest fragmentation on plant communities, with models explaining up to 35% of the variation in PFA–control community similarity (Table 4). In the Westland district, models explained 15–16% of all variation in community similarity index between PFA and control plots, whereas in the Nothofagus-dominated landscape, the best-supported models explained up to 35% of the variation in community similarity. This particular model, explaining patterns of ground cover (T6), found that community similarity was positively correlated with patch size and negatively correlated with edge density (Table 4).
Effects of excluding exotic species
Exotic species represented approximately 10% of total species richness in the two landscapes (Table 1), and we found that excluding them from the total list of species did not significantly alter the model selection statistics. Best models were the same in each tier group regardless of whether exotic species were included or excluded; in no cases did removing the exotic species change the sign of a parameter estimate, although model parameters were slightly different in most cases. Overall, and contrary to our expectations, exotic plants did not drive the patterns of (dis)similarity observed between PFA and control sites (Tables 3 & 4).
Analysis of α-diversity
On the whole, fragmentation metrics explained little of the spatial variation in species richness (i.e. α-diversity). In the lowland rain forest landscape, percentage of non-forest (PNF.1000) (slope =−0.32; P= 0.004) was the only variable with a significant effect on ground-cover (T6) species richness. No fragmentation metrics explained the α-diversity of tree and shrub species. In the Nothofagus-dominated landscapes, edge density (ED.500) (slope = 0.07; P= 0.041) was the only significant factor explaining variation in α-diversity for ground-cover species, and none of the fragmentation metrics had significant effects on species richness in tiers T1–T5. Similar results were obtained for both study areas when exotic species were excluded from the analysis.
Landscape-level assessment of forest fragmentation impacts
In the two landscapes, we used the best supported of our multiple regression models (Table 4) to generate continuous grid maps representing the predicted similarity of each pixel to control sites located deep within the forest (Fig. 3). These similarity maps provide a visual indication of the likely spatial variation of community composition across the landscapes as a result of forest fragmentation.
Similarity maps were used to derive the value of the BioFrag index for each landscape (equation 5, Table 4): BioFrag is a landscape-level estimate of the impact of forest fragmentation on plant communities that explicitly incorporates the spatial pattern of forest cover. BioFrag values ranged between 0.86 (T6) and 0.90 (T1–T5) in the Westland district, suggesting that plant communities in the lowland rain forest landscape are relatively little affected by fragmentation. In the Nothofagus-dominated landscape, BioFrag ranged between 0.68 (T1 and T5) and 0.86 (T6) indicating a higher fragmentation effect on tree and shrub communities than on ground-cover composition.
A biologically meaningful index of habitat fragmentation
We devised a method for exploiting spatial patterns of community similarity to generate an index of fragmentation impacts, BioFrag, that accounts for the exact spatial patterns of habitat cover in a landscape. BioFrag represents two important advances over alternative measures of fragmentation effects. First, BioFrag is derived from observed patterns of biological responses to fragmentation metrics, rather than relying solely on descriptors of spatial patterns of habitat cover. Simple pattern descriptions assign a ‘value’ to a landscape, but ignore the fact that different taxa will respond in different manners to that pattern. For example, multi-taxa studies in human-modified and fragmented landscapes have empirically demonstrated that the biological value of a landscape is taxon dependent (Laurance et al., 2002; Barlow et al., 2007). By relying on observed patterns of community changes in response to habitat fragmentation, BioFrag values are implicitly taxon-specific and will, for example, be different for bird and tree communities in the same landscape.
The second advantage of BioFrag is the spatially explicit nature by which it is calculated. This enforces the principle that not all forest is equal. It has long been established that small areas of isolated forest do not support the same biological value as equal areas of forest that are part of a large, continuous forest tract (Pardini et al., 2005; Giraudo et al., 2008). BioFrag encompasses this principle by weighting the index value of a forested pixel according to its position in a landscape using variables such as the size of the forest patch it is embedded within, the amount of forest surrounding the pixel, and the distance of that pixel to the forest edge. Consequently, the BioFrag index accounts for the exact pattern of fragmentation within a landscape, with the result that landscapes of equal area and with the same proportion of habitat cover will have different index values that depend on the spatial pattern of habitat cover.
We calculated BioFrag index values for two contrasting landscapes in New Zealand. Overall, BioFrag values were relatively high (typically >0.8), indicating that the influences of fragmentation on these forest plant communities were only moderate. This suggests that in these regions a large proportion of the between-plot variation in community composition is due to underlying variation in factors such as climatic conditions, soil composition and fertility that do not spatially covary with the fragmentation metrics. The high BioFrag values also reflect the fact that the two landscapes we studied contained extensive areas of continuous forest. The impacts of fragmentation are likely to be much greater in landscapes that are more heavily fragmented. Despite the relatively low impact of fragmentation, we observed important differences within and between the two landscapes. Index values were approximately equal for ground-cover communities in both landscapes, despite the massive difference in the degree of fragmentation among landscapes. This result shows that BioFrag generates values that have a very intuitive property: biological communities that are minimally impacted by fragmentation, as indicated by the limited ability of fragmentation metrics to explain patterns of community composition, return similar BioFrag values when assessed at the landscape scale, regardless of whether they are assessed in a landscape with low or high degrees of habitat fragmentation.
By contrast, BioFrag values in the heavily fragmented montane forests were considerably lower (0.68) for the tree and shrub community than for the ground-cover community (0.86). This pattern occurred despite the ground-cover community having much stronger regression models than the tree and shrub community, giving a superficial impression that it is the ground-cover community that is more strongly impacted by habitat fragmentation. The reason for this apparent incongruity in the results is that while the ground-cover community responds strongly to features of the landscape, the effect size of that response is small relative to the effect size of the response observed in the tree and shrub community. Similar statistical issues have been raised elsewhere (Northcott, 2008), showing how variables that appear to have a greater statistical significance can have lower impacts on a response than other seemingly less significant variables. By extrapolating regression equations defining community responses to fragmentation metrics across the full landscape, BioFrag effectively reports the net magnitude of the biological impact of fragmentation. By contrast, measures of goodness-of-fit for the regression equation that feeds into BioFrag provide an estimate of the strength of the statistical relationship between communities and fragmentation metrics. Consequently, BioFrag provides a measure of fragmentation impacts that is based more on effect sizes than on statistical significance, providing additional support for our statement above that BioFrag values provide a biologically relevant measure of the impact of habitat fragmentation.
Beyond landscapes: scaling BioFrag estimates to nations and biomes
Our spatially explicit landscape-scale index of community change, BioFrag, shows how the fine-scale configuration of habitat loss sums across a landscape to determine changes in biodiversity at a larger spatial scale. Within a specific habitat, BioFrag estimates can be calculated at almost any spatial scale that is smaller than the spatial extent of that particular habitat. For example, using our parameter estimates we could generate a BioFrag estimate for every 100 × 100 km, 10 × 10 km or even 1 × 1 km grid square within the landscapes we studied. However, scaling BioFrag estimates to quantify fragmentation impacts across larger spatial scales such as nations and/or biomes represents an important challenge. Developing rigorous methods for scaling BioFrag from individual landscapes to much larger spatial extents is of crucial importance for assessing progress towards the Convention on Biological Diversity (CBD) 2010 target to reduce rates of biodiversity loss (UNEP/CBD, 2002, 2004). The 2010 target is a global target, although the actions, monitoring and reporting of progress will be done by individual nations which vary greatly in terms of their land area and the diversity of habitats they contain.
Our analyses showed that the floral communities inhabiting two different habitat types within a single island of New Zealand responded differently to different components of the fragmented landscape. Therefore, a direct extrapolation of parameter estimates from one type of habitat would not generate a reliable estimate of fragmentation impacts in a second type of habitat, and it is not yet possible to generate a single BioFrag estimate for a region encompassing multiple habitat types. We propose two potential routes for generating BioFrag estimates at very large spatial scales. First and most obviously, one could generate an individual BioFrag estimate for each distinct habitat type within a nation or biome, and combine those estimates using a weighted average to reflect the relative extent of each habitat within the larger region. Second, one could estimate community responses to the individual fragmentation parameters in a number of landscapes across the nation or biome. Parameter values for communities inhabiting landscapes that were not sampled could then be predicted from the sampled landscapes using spatial interpolation techniques. This approach would allow a researcher to predict BioFrag values in unsampled areas and, consequently, across much larger spatial scales than can be directly sampled, paving the way to generating national-scale estimates of the impacts of fragmentation and thereby contribute to assessing the rate of loss of biodiversity that is required by the CBD (UNEP/CBD, 2004).
Community responses to habitat fragmentation
The process of calculating the BioFrag index requires the user to link remotely sensed patterns of habitat cover with field-based biological observations, generating useful information about the spatial determinants of community changes in fragmented landscapes. For example, we found that landscape edge density was an important determinant of ground-cover plant communities in both landscapes, as well as explaining variation in the α-diversity in the montane ground-cover community. This finding should be interpreted cautiously, as edge density was significantly correlated with several other landscape metrics that we used in this study, yet it is consistent with other literature suggesting that many herb species have a general preference for edge-disturbed locations (Ross et al., 2002; Hobbs & Yates, 2003). In the lowland rain forest, we also detected an effect of distance from edge, and landscape forest cover on ground-cover communities, whereas in the montane forests, our results suggested that the size of forest fragments was more relevant. By contrast to the ground-cover community, spatial variation in the tree and shrub plant communities of both landscapes was partially explained by distance from edge, with landscape forest cover and nearest neighbour distance playing important roles in the lowland and montane landscapes, respectively. Across the eight landscape × community combinations that we analysed, the direction, or polarity, of the relationship between the biological data and landscape metrics was remarkably consistent. For example, plots in landscapes with high edge density that were close to forest edges or were in small forest patches tended to have reduced similarity between the PFA and control plots than other plots. Similarly, community similarity was reduced in landscapes with low relative to high landscape forest cover. We found that communities differed in the strength of their relationship with landscape metrics, such that patterns detected in one landscape were weaker or absent in another. However, in no instance did we detect a reversal of polarity, in which the relationship between community similarity and a given landscape metric had the opposite sign for different communities or landscapes. This suggests that the mechanisms underlying the responses of the communities we analysed are remarkably consistent across landscapes.
Taken together, our analyses of the spatial determinants of plant community patterns highlight two important points. First, they show that different communities respond to different fragmentation metrics. We expect that this difference would be even more noticeable when comparing fragmentation patterns of floral with faunal communities, which differ greatly in terms of their habitat requirements, modes of reproduction and dispersal mechanisms, and all of which influence the way in which a given species will respond to the spatial structure of a landscape. Second, we found that in different landscapes the same communities were influenced by different characteristics of the fragmented landscapes. Although edge effects appear to be a significant driver of the impacts of fragmentation for both plant communities in both landscapes, the relative importance of other variables such as fragment size and isolation varied among landscapes. This result may reflect the fact that the communities in the two landscapes are different and have spatial patterns that are governed by different ecological requirements of those species or by the different histories of the two landscapes. Moreover, it is possible that the correlations between community composition and variables such as fragment size and isolation may vary among landscapes with differing amounts of habitat, a process that could also result in apparently variable effects of fragment size and isolation on tree communities.
We used BioFrag index values to test three specific hypotheses regarding the responses of plant communities to forest fragmentation. First, we had predicted that plant communities in the heavily fragmented landscape would be more negatively affected than those in the less fragmented landscape. Our results provided equivocal support for this hypothesis, which was supported for the tree and shrub community. By contrast, and contrary to our second hypothesis, the impacts of fragmentation on the herb layer community were similar in both landscapes, and considerably smaller than the influence on the tree and shrub community in the heavily fragmented landscape. However, in line with our hypothesis, we found that fragmentation metrics had significant impacts on the α-diversity of the ground-cover community in both landscapes, but did not detect a similar impact on the tree and shrub community. Our final hypothesis, that the impacts of fragmentation would be driven by exotic species, was not supported. Excluding exotic species from the analyses did not result in significant changes to the observed relationships between community composition and fragmentation metrics, or to the final estimates of the impacts of fragmentation as indicated by BioFrag values. This does not suggest that exotic species do not respond to patterns of fragmentation, as we know that forest edges in New Zealand act as a focal point for plant invasions (Wiser et al. 1998). What this result does indicate is that the responses of native species to patterns of fragmentation are at least as strong as, or perhaps even override, the responses of exotic species to those same spatial patterns of forest cover.
LIMITATIONS AND FUTURE DIRECTIONS
Fragmentation affects the structure of landscapes, causing shifts in the diversity and distribution of plant species that could lead to extinctions of populations in forest fragments (Arroyo-Rodríguez et al., 2007; Barbaro et al., 2007). Quantifying the biological effects of fragmentation at multiple spatial scales is a prerequisite to establishing strategies and policies for counteracting loss of biodiversity at the global, regional and national levels (Millennium Ecosystem Assessment, 2005) and for quantifying progress towards global targets to reduce the rate of loss of biodiversity (CBD Secretariat, 2001).
Recent advances in remote sensing and geospatial technologies allow the routine assessment of fragment properties at the landscape scale, such as the distribution of patch sizes and shapes, and the spatial configuration of forest patches (Riitters et al., 1995; McGarigal & Cushman, 2002). Nevertheless, new methods are still needed for these structural properties to be interpreted in terms of biological and functional impacts on ecosystems and landscapes. We developed the BioFrag index as an important step towards this goal, but it is not the final step. BioFrag estimates the impact of fragmentation on communities within a single habitat type, but ignores the fact that many species and many important ecological processes and services are not bounded by the edges of that habitat (Fischer & Lindenmayer, 2006). Rather, forest patches are just one habitat embedded within landscape mosaics that incorporate multiple ecosystems. Moreover, the individual habitats comprising landscape mosaics change over the large spatial scales that are relevant to assessments of the impacts of fragmentation at national and global scales. Biologically meaningful measures of regional-scale landscape mosaics still need to be developed before we can translate complex spatial patterns of land cover and land use in terms of biological or ecological impacts. This next generation of landscape metrics will need to integrate patterns of biodiversity across the full range of habitats and land-cover types that occur within a given landscape. These metrics are required to quantify current trends in the impacts of land-use change on biodiversity and ecosystems. Ultimately, indices that extend beyond BioFrag will provide essential support for quantifying progress towards globally agreed policy targets on biodiversity and habitat conservation.
The authors of this paper acknowledge the use of data drawn from the National Vegetation Survey (NVS) Database, and thank the many researchers involved in collecting and storing this information. We are grateful to Dr Robert Allen (Landcare Research, New Zealand) for permission to include his Harper/Avoca dataset in our study, and to two referees for their constructive comments on the manuscript. This study was supported by a grant from the Isaac Newton Trust (Cambridge, UK), Conservation Biology Initiative 2007–08 and by resources from the Global Environment Facility's 2010 Biodiversity Indicators Partnership project (via UNEP-WCMC). It forms a contribution to the work of the 2010 Biodiversity Indicator Partnership.
Raffaele Lafortezza is a Research Associate at the University of Cambridge, UK and Lecturer at the University of Bari, Italy. He is a landscape ecologist and GIS analyst with an interest in geospatial analysis applied to forest conservation and biodiversity assessment at multiple scales.