Habitat fragmentation caused by woody plant encroachment inhibits the spread of an invasive grass

Authors


*Correspondence author. E-mail: kmalofs@mail.utexas.edu

Summary

1. Although habitat fragmentation and species invasions are widely recognized threats to biodiversity, few empirical studies have examined both threats together. Here we test the hypothesis that habitat fragmentation may limit the spread of invasive species across landscapes.

2. In central Texas, patches of herbaceous vegetation become increasingly fragmented (e.g. smaller and more isolated) as landscapes undergo woody plant encroachment. However, during this process, fragmentation and loss of herbaceous habitat are not completely correlated, which allowed us to separate the effects of fragmentation from the effects of habitat loss.

3. At three sites in central Texas, we recorded the occurrence of the invasive bunchgrass Bothriochloa ischaemum in randomly-located plots in herbaceous patches. We measured fragmentation of the herbaceous habitat around these plots at five scales, using aerial photographs.

4.Bothriochloa ischaemum occurrence was significantly negatively related to most aspects of fragmentation. These negative relationships were stronger after the effects of habitat loss were removed from the effects of fragmentation per se.

5. The strength of the negative relationship between B. ischaemum occurrence and fragmentation varied among sites, probably reflecting the influence of grazing and other management practices.

6.Synthesis and applications. Our results indicate that fragmentation, separated from habitat loss, can have positive as well as negative conservation effects. Fragmentation, if wisely used, could be a useful tool in the management of invasive plant species like B. ischaemum. Increasing the degree of habitat fragmentation may slow the spread of invasive species across landscapes.

Introduction

Although landscape connectivity is generally regarded as desirable and habitat fragmentation as undesirable for conservation purposes, the potential role of connectivity in facilitating the spread of introduced species has been recognized for more than 20 years (Forman & Godron 1986; Simberloff 1988). Concern has been expressed repeatedly (e.g. Simberloff et al. 1992; Whittaker 1998; Proches et al. 2005; Lindenmayer et al. 2008), but the hypothesis that connectivity increases, and fragmentation decreases, the rate of spread of invasive species, remains largely untested. Here we report the results of a test of this hypothesis. Based on these results, we discuss how fragmentation may in some instances be useful in conservation.

To test the hypothesis that fragmentation, including reduced connectivity, is slowing an invasion, we quantified habitat loss, habitat fragmentation and the distribution of an invasive grass in the same landscapes. Habitat loss, habitat fragmentation, and species invasions are three important threats to biodiversity world-wide (e.g. Wilcove et al. 1998; Sala et al. 2000; Ewers & Didham 2006), but few previous studies have investigated the potential interactions between the effects of these factors (Didham et al. 2007). It has been suggested that important insights may be gained by integrating landscape and invasion ecology (With 2002; Chabrerie et al. 2007), but few studies of invasive species have considered the spatial configuration of the landscape being invaded (With 2002; Andersen et al. 2004). This omission is surprising given that natural resource managers may have greater control over habitat characteristics than they do over arriving propagules. However, understanding the relationships between landscape spatial configuration and the spread of invasive species is challenging because they may be nonlinear, include thresholds, and be sensitive to the spatial resolution at which indices are measured (Tischendorf 2001).

Recent models have suggested that habitat fragmentation may slow the spread of species with limited long-distance dispersal but may promote the spread of species with better dispersal (With 2002, 2004; Pearson & Dawson 2005). There are only a few previous empirical studies of the role of habitat configuration on the spread of invasive species through landscapes, with inconsistent results. Thiele, Schuckert & Otte (2008) found that habitat connectivity and other indicators of low levels of fragmentation are positively correlated with the presence of an invasive plant species. In contrast, Bartuszevige, Gorchov & Raab (2006) failed to find a relationship between the presence of an invasive plant species and connectivity, but found a strong positive relationship with the amount of habitat edge (which increases with fragmentation). In experimental landscapes with equal amounts of habitat, Damschen et al. (2006) did not find a significant effect of corridors on the abundance or proportion of introduced species in the species pool.

We tested the role of habitat fragmentation per se in the spread of the invasive grass Bothriochloa ischaemum (L.) Keng, which has relatively limited dispersal. We hypothesized that habitat fragmentation has slowed the spread of B. ischaemum and therefore its presence in our study plots would be negatively related to the degree of fragmentation of the surrounding landscape. The effects of habitat fragmentation per se are difficult to separate from those of habitat loss because these processes often co-occur and because many metrics of fragmentation are strongly related to habitat amount (Gustafson & Parker 1992; Andrén 1994; Fahrig 2002, 2003). Habitat fragmentation per se has been defined as the breaking apart of habitat independent of habitat loss, generally leading to increased number of patches, decreased mean patch size, and increased mean patch isolation (Fahrig 2003). We took advantage of the widespread existence of landscapes in central Texas in all stages of woody plant encroachment, in which habitat loss (of herbaceous habitat suitable for B. ischaemum) and fragmentation are sufficiently uncorrelated to allow us to separate the effects of habitat loss from the effects of habitat fragmentation. We separated the two effects using principal components analysis and by calculating residual metrics from models of various individual measures of fragmentation as functions of habitat amount (McGarigal & McComb 1995; Trzcinski, Fahrig & Merriam 1999; Villard, Trzcinski & Merriam 1999; Sleeman et al. 2005; Yamaura, Katoh & Takahashi 2006).

Materials and methods

Over recent decades, the relative abundance of woody plants has dramatically increased in savannas around the world (Van Auken 2000). In central Texas, woody plant encroachment decreases the amount of open herbaceous habitat while the remaining habitat becomes more fragmented. The vegetation of much of the eastern Edwards Plateau falls on a continuum from savannas (grasslands with patches of woody plants) to woodlands with remnant glades of herbaceous vegetation. There is variation in the configuration and degree of fragmentation of herbaceous habitat in landscapes with the same amount of open herbaceous (i.e. non-woody) vegetation. Dominant woody species are Ashe juniper Juniperus ashei J. Buchholz and plateau live oak Quercus fusiformis Small. Dominant grass species include B. ischaemum, Bothriochloa saccharoides (Sw.) Rydb., Nasella leucotricha (Trin. & Rupr.) Pohl, Schizachyrium scoparium (Michx.) Nash, and several species of Bouteloua and Aristida.

This study was conducted at three sites on the eastern Edwards Plateau: Balcones Canyonlands National Wildlife Refuge (BC), Pedernales Falls State Park (PF) and Freeman Ranch (managed by Texas State University; FR). These sites differed in grazing history and in the frequency and intensity of woody plant removal (Table 1).

Table 1.   Management histories of three study sites and frequency of B. ischaemum observed in plots at each site
SiteYears since last grazedPrescribed fire and mechanical clearing of woody plantsN plotsN plots B. ischaemum present
Balcones Canyonlands National Wildlife Refuge (BC)∼15Often, extensive8458 (69%)
Freeman Ranch (FR)Grazed during studySome, limited8057 (71%)
Pedernales Falls State Park (PF)30+Little8530 (35%)

The Eurasian perennial bunchgrass King Ranch bluestem Bothrichloa ischaemum was introduced to Texas as early as 1930 (Gabbard & Fowler 2007) and has been widely planted for erosion control, revegetation, fodder, and rangeland improvement. It is now common throughout central and southern Texas (Turner et al. 2003). Bothriochloa ischaemum is the only common non-native invasive plant in the savannas of the eastern Edwards Plateau, where it often forms dense, nearly pure stands with low native plant species diversity (Gabbard & Fowler 2007). On the eastern Edwards Plateau the distribution of B. ischaemum is unrelated to slope, fire, or grazing history; however, it is never found under the canopies of woody plants (Gabbard & Fowler 2007). Thus for the purposes of this study, B. ischaemum‘habitat’ is any upland open herbaceous area not under the canopy of woody plants, regardless of extent. Bothriochloa ischaemum seems to have quite limited dispersal ability (N. L. Fowler, pers. obs.). Although B. ischaemum dispersal has not been explicitly measured, it forms virtual monocultures which spread in advancing fronts. The structure of B. ischaemum spikelets does not appear to support wind or animal dispersal.

Data collection

Between 5 May and 13 June 2007, we performed a census of 249·1-m radius plots in the three sites. Plots were centred on points randomly selected within herbaceous habitat using ArcGIS 9.2. Each point was located with GPS and aerial photographs. All plant species present at each plot were recorded.

We selected the landscape metrics listed in Table 2 using criteria of interpretability, practicality, descriptiveness, and bias (Jaeger 2000; Neel, McGarigal & Cushman 2004; Sleeman et al. 2005). These eight fragmentation metrics will be referred to as ‘initial metrics’ to distinguish them from other measures of fragmentation derived from them (see below). Attempts to identify a universal set of independent metrics which describe the major components of landscape configuration suggest it is necessary to consider the behaviour of each metric with regard to the specific processes and landscapes being investigated (Tischendorf 2001; Cushman, McGarigal & Neel 2008). We chose metrics that describe landscapes rather than patches because fragmentation is a landscape-level process (McGarigal & Cushman 2002). We first selected landscape metrics often related to colonization (e.g. nearest neighbour distance) or to extinction (e.g. mean patch area), to which we added metrics particularly suited to describing our study landscape, in which patches are rarely compact or obvious (Fig. 1) (e.g. landscape shape index and area-weighted mean fractal dimension). Measures of connectivity which rely upon estimations of dispersal distance were not practical for this study because dispersal of B. ischaemum has not been quantified.

Table 2.   Description of habitat amount and 8 fragmentation metrics used to describe habitat configuration and their relationship to fragmentation (for further description and equations see McGarigal et al. 2002). Metrics that are positively related to the degree of fragmentation are expected to be negatively related to B. ischaemum presence, and vice versa
MetricRelationship to fragmentationDescription
Habitat amountNegativePercentage of the landscape comprised of the herbaceous habitat
Mean patch areaNegativeMean area (m2) of herbaceous habitat patches converted to hectares
Clumpiness indexNegativeProportional deviation of joins between herbaceous habitat cells from that expected from a spatially random distribution; ranges from −1 (maximal disaggregation) to 1 (maximal aggregation)
Patch cohesionNegativePhysical connectedness of herbaceous habitat given as a percentage; calculated using all patches as (1 - ∑p/∑ (pa))(1 - 1 /√N)−1 (100); where p is patch perimeter (m), a is patch area (m2), and N is the total landscape area (m2)
Mean nearest-neighbour distancePositive or negative (see discussion)Mean for all herbaceous patches of edge-to edge
Euclidean distance (m) to the nearest neighbouring herbaceous habitat patch from cell centre to cell centre
Number of patchesPositiveNumber of patches of herbaceous habitat, adjacent cells (by the 8 neighbour rule) belong to the same patch
Total edgePositiveTotal length (m) of herbaceous habitat patch perimeter
Landscape shape indexPositiveTotal perimeter of herbaceous habitat patches over minimum perimeter with maximal aggregation
Area-weighted mean fractal dimensionPositiveSum for all herbaceous patches of (2 ln (.25 p)/ln a)(a/∑a); where p is patch perimeter (m) and a is patch area (m2); approaches 2 with increased patch shape complexity
Figure 1.

 Typical configuration of habitat in a 56-m radius landscape shown as (a) an aerial photograph (orthophotoquad) and (b) an herbaceous (black cells) and woody (white cells) vegetation binary layer created from the photograph.

Ideally, fragmentation metrics should assess changes in landscape configuration that are independent of habitat amount. However, most metrics of fragmentation are related to habitat amount (Fahrig 2003; Neel et al. 2004). We included two metrics that were developed to describe fragmentation per se: the clumpiness index (McGarigal et al. 2002; Neel et al. 2004) and patch cohesion (Schumaker 1996; but see Neel et al. 2004 and our results below). We also removed the effects of habitat amount by calculating ‘residual metrics’ (see below).

One-metre resolution Digital Orthophotoquads (DOQs) of all study sites taken in 2004 were obtained from the Texas Natural Resources Information System (TNRIS). Using supervised maximum likelihood classification in ArcGIS 9.2, these photographs were converted to binary rasters where each cell represented either woody or herbaceous habitat (Fig. 1). With these binary rasters, we calculated habitat amount and the eight initial fragmentation metrics (Table 2) for landscapes surrounding each of the 249 random points (see above). In landscape studies it is important to choose a scale of analysis relevant to the organism or process being investigated (McGarigal & Cushman 2002). We anticipated that the relationship between landscape configuration and B. ischaemum occurrence may not be detectable for landscapes which are too small or too large. Therefore, we measured fragmentation at five scales: circular landscapes centred on the given point with a 10, 25, 56, 75 or 100-m radius. A 56-m radius landscape is approximately 1 ha. We used the FragStatsBatch script for ArcGIS 9 which calls fragstats 3.3 (McGarigal et al. 2002) to calculate fragmentation metrics. The distance from each point to the nearest road or trail was also measured using ArcGIS 9.2.

Statistical analyses

Examining site-to-site and scale-to-scale variation, distance to roads or trails and metric correlations

A χ2 goodness-of-fit test was used to examine differences in the rate of B. ischaemum occurrence among sites. Logistic regressions between B. ischaemum presence/absence and the distance to the nearest road or trail were performed to test the influence of roads and trails on the occurrence of B. ischaemum. We examined the range and distribution of each of the eight initial fragmentation metrics and of habitat amount at each of the five scales to identify important scales and to compare sites. A 56-m (1 ha) radius landscape scale was used for most subsequent analyses (see Results regarding this choice). One-way anova was used to test differences in habitat amount and fragmentation metrics among sites. Pearson product–moment correlations were calculated to examine relationships among initial fragmentation metrics and between initial fragmentation metrics and habitat amount. All statistical analyses, with the exception of principal components analysis (see below), were performed using sas 9.1 (SAS 2003).

Description of fragmentation per se

We used three statistical techniques to describe the variation in habitat fragmentation per se (i.e. differences in fragmentation above and beyond those due to differences in habitat amount). These analyses were performed at a single scale, using 56-m radius landscapes. In landscapes where all herbaceous habitats occurred in a single patch the distance between patches (nearest-neighbour distance) could not be measured. The 13 plots in such landscapes were dropped from all analyses of fragmentation per se.

We calculated residual fragmentation metrics by determining which type of model best predicted the value of each initial fragmentation metric as a function of habitat amount. For most of the fragmentation metrics the best model was a linear or quadratic polynomial (fitted with the sas glm procedure, with appropriate transformations to improve the normality of residuals); the model with the largest R2 was considered to be the best model (see Table S1 and Fig. S1 in Appendix S1, Supporting information). The fragmentation metric patch cohesion was an exception: a hyperbolic model was considered, by visual inspection, to provide the best fit for this metric (fitted with sas nlmixed procedure). A second exception was the fragmentation metric clumpiness, for which no model was fitted, because the relationship between it and habitat amount was so weak (see Results). From the best model for each fragmentation metric, except clumpiness, we obtained the residual value (i.e. inline image) for each landscape i. These residual values are our ‘residual fragmentation metrics’ and are free of most of the confounding effects of habitat amount (McGarigal & McComb 1995; Villard et al. 1999), although the effects of fragmentation relative to habitat amount can be underestimated (Sleeman et al. 2005; Koper, Schmiegelow & Merrill 2007). We calculated Pearson product–moment correlations between each pair of initial fragmentation metrics and habitat amount, and between each pair of residual fragmentation metrics and habitat amount.

We also used principal components analysis (PCA), performed in PC-ORD (McCune & Mefford 1999), to summarize the eight initial fragmentation metrics into principal components representing gradients of habitat fragmentation (e.g. Trzcinski et al. 1999; Sleeman et al. 2005). This technique does not factor out the effect of habitat amount a priori, thus making no assumption that habitat amount is more important than fragmentation in explaining differences in configuration, but it may obscure non-linear relationships between metrics (Cushman et al. 2008). Therefore we also performed a PCA incorporating the seven residual metrics and the clumpiness index for which we did not calculate a residual metric (e.g. McGarigal & McComb 1995; Yamaura et al. 2006). This PCA accounts for non-linear relationships between metrics and particularly between metrics and habitat amount. We determined the usefulness of each principal component with the broken-stick criterion (Jackson 1993). Principal components with eigenvalues greater than expected under the broken-stick (random) distribution were judged meaningful. To determine the relationships between principal component axes and habitat amount and between each other, Pearson product–moment correlations were calculated.

Presence/absence of B. ischaemum

We examined the relationship between local occurrence data (from plots) and habitat configuration in surrounding landscapes of various sizes. Logistic regressions of B. ischaemum occurrence (presence/absence) against each of the eight initial fragmentation metrics and habitat amount were performed at each of five scales. At the 56-m scale, we performed logistic regressions of B. ischaemum occurrence against each residual metric and all significant principal components. The presence/absence of B. ischaemum was the dependent variable in each analysis. The independent variables in each regression were one of the metrics (or principal components), site, and a metric*site interaction term. We determined whether the relationship between the occurrence of B. ischaemum and each metric varied substantially among sites by comparing models with and without the metric*site interaction term. If including the metric*site interaction term decreased the Akaike information criterion (AIC) of a model by two or more (Burnham & Anderson 2002), we fitted three separate models, one per site. Otherwise, the pooled data set was retained and the interaction term was dropped from the model.

Results

Relationship between the occurrence of b. ischaemum and fragmentation

Regardless of scale, nearly all significant relationships between B. ischaemum occurrence and initial fragmentation metrics demonstrate a negative relationship between B. ischaemum and fragmentation (e.g. B. ischaemum was more likely to be present in plots located in landscapes with less edge and higher patch cohesion) (Fig. 2, Table 4, also see Table S2, Appendix S2). The strength of the relationship between B. ischaemum occurrence and the initial fragmentation metrics often varied among sites and was significant more often at PF than other sites (Table 4). Using residual metrics improved our ability to detect the negative relationship between B. ischaemum occurrence and fragmentation across sites: there was less evidence of variation among sites and more significant relationships for all sites pooled (Table 4). Further supporting the negative relationship between occurrence and fragmentation, B. ischaemum was more likely to be present in plots further to the right on the first axis of each PCA (Fig. 3, Table 4). Occurrence of B. ischaemum was significantly negatively related to habitat amount (measured in landscapes with 56-m radius and larger) only at PF. Consistent with this, B. ischaemum was more likely (> χ2 = 0·0027) to be present in PF plots further to the left on the second axis of the PCA of initial fragmentation metrics, an axis that was positively correlated with habitat amount (Fig. 3, Table 4).

Figure 2.

 Logistic regression models of the probability of B. ischaemum presence against initial and residual patch cohesion at FR and initial and residual total edge at PF. Fragmentation metrics measured in 56-m radius landscapes.

Table 4.   Results of logistic regressions of B. ischaemum occurrence v. habitat area, fragmentation metric, or principal component in 56-m radius (1 ha) landscapes. For some metrics, sites were pooled (see Methods). Far right column indicates model support for a negative relationship between occurrence and fragmentation
Fragmentation metricEstimated metric value Supports
Sites pooledBCFRPFHypothesis
  1. *P <0·05, **P <0·01, ***< 0·005.

  2. aNo residual calculated for clumpiness index, insignificant relationship with habitat amount.

  3. bThis axis primarily represents habitat amount, not fragmentation.

Negatively related to fragmentation
 Habitat amountInitial0·0060·009−0·03**n/a
 Mean patch areaInitial0·88   
Residual2·40   
 Clumpiness indexaInitial0·221·6010·90*yes
 Patch cohesionInitial0·020·31*−0·14yes
Residual0·070·64**−0·01yes
Positive or negative relationship with fragmentation
 Mean nearest-Initial−0·040·06 0·39***n/a
 Neighbour distanceResidual0·21*  n/a
Positively related to fragmentation
 Number of patchesInitial−0·01   
Residual−0·02*  yes
 Total edgeInitial−0·0004−0·0002−0·002***yes
Residual−0·0004−0·0005−0·002***yes
 Landscape shape indexInitial−0·15**  yes
Residual−0·27***  yes
 Area-weighted mean fractal dimensionInitial−3·37*  yes
Residual−3·96   
 1st principal component of initial metrics 0·23***  yes
 2nd principal component of initial metrics 0·030·58−0·62**n/ab
 1st principal component of residual metrics 0·26***  yes
Figure 3.

 Plot of first principal component vs. second principal component at PF. Open circles represent plots with B. ischaemum absent and solid circles represent plots with B. ischaemum present.

General description of b. ischaemum distribution

Bothriochloa ischaemum occurred significantly less often in plots at PF than the other sites (Table 1, χ22 = 28·00, < 0·0001). Pooling all sites, the mean distance from a plot to the nearest road or trail was 119-m. There was no significant site*distance interaction (> χ= 0·41) and after pooling sites there was no significant relationship between the occurrence of B. ischaemum in a plot and the distance between that plot and the nearest road or trail (> χ=0·18). Therefore distance to roads was not included in further analyses.

Scale-to scale and site-to-site variation in fragmentation

Some metrics were relatively invariant to scale (mean nearest-neighbour distance, clumpiness index, patch cohesion; Fig. S2, Appendix S2). Others increased predictably in mean and range with increasing scale, such as number of patches, total edge and landscape shape index (Shen et al. 2004). In general, varying spatial scale did not change the sign of the relationship between the local presence of B. ischaemum and the degree of fragmentation in the surrounding landscape. Comparing scales, the intermediate scale, 56-m radius, was significant for the most metrics (Table S2, Appendix S2). Over the five scales analysed, we found no change in the significance of the relationship between B. ischaemum occurrence and landscape shape index or mean patch area. The relationship with habitat amount and patch cohesion, however, was insignificant at smaller scales. For the majority of the remaining metrics the relationship was insignificant, or significant only at PF, at larger scales. We therefore selected the 56-m radius scale for analyses separating the effects of fragmentation per se from habitat amount.

Site-to-site differences in most aspects of fragmentation were relatively scale invariant although the magnitude of the effect may have changed with scale (Fig. S2, Appendix S2). At the 56-m scale, landscapes in BC were the least fragmented: they had on average larger mean patch area (F246, 2 = 13·74, < 0·0001) and smaller total edge (F246, 2 =31·97, < 0·0001), landscape shape index (F246, 2 = 21·26, < 0·0001) and area-weighted mean fractal dimension (hereafter fractal dimension) (F246, 2 = 44·51, < 0·0001). Landscapes in FR were on average the most fragmented: they had on average a larger number of patches (F246, 2 = 36·92, < 0·0001) and total edge (F246, 2 =31·97, < 0·0001). The mean amount of habitat in PF landscapes was significantly smaller than that at other sites (F246, 2 = 14·03, < 0·0001).

Correlations between fragmentation metrics and habitat amount

Using the initial (uncorrected) measures of fragmentation, habitat amount was negatively correlated with degree of fragmentation (e.g. strong negative correlation with number of patches, strong positive correlation with mean patch area, Table 3). However, the relationship between clumpiness index and habitat amount was so weak (= 0·17) that a residual clumpiness index was not calculated. Residual patch cohesion was, however, calculated because initial patch cohesion was strongly related to habitat amount (= 0·71), despite the fact that this metric was developed to measure fragmentation apart from habitat amount (Schumaker 1996; but see Neel et al. 2004). Residual metrics had little correlation with habitat amount (|r| ≤ 0·20, Table 3), as expected. Although almost all of the initial fragmentation metrics were strongly related to habitat amount, separating out habitat amount changed the relatively strong relationships between them very little (|r| > 0·40; Table S3, Appendix S3).

Table 3.   Pearson product–moment correlations between initial fragmentation metrics, habitat amount and PCA axes and between residual fragmentation metrics, habitat amount and PCA axis (in parentheses)
Fragmentation MetricHabitatamount1st principal component2nd principal component
  1. aNo residual calculated for clumpiness index, insignificant relationship with habitat amount.

Negatively related to fragmentation
 Habitat amount0·39 (0·06)0·76
 Mean patch area0·69 (0·19)0·52 (0·21)0·54
 Clumpiness index0·17a0·79 (0·93)−0·01
 patch cohesion0·71 (0·20)0·22 (0·43)0·86
Positive or negative relationship with fragmentation
 Mean nearest-neighbour distance−0·40 (−0·05)0·46 (0·74)−0·64
Positively related to fragmentation
 Number of patches−0·44 (-0-)−0·69 (−0·63)−0·38
 Total edge−0·05 (-0-)−0·88 (−0·95)0·37
 Landscape shape index−0·55 (-0-)−0·96 (−0·97)−0·14
 Area-weighted mean fractal dimension−0·27 (-0-)−0·81 (−0·73)0·29

Principal components analysis of fragmentation metrics

The strong correlations among fragmentation metrics, even after the effect of habitat amount had been removed, suggested some redundancy among fragmentation metrics; a principal components analysis addresses this. By the broken-stick method, the first 2 axes of the PCA of the initial fragmentation metrics were significant and together these axes explained 72·39% of the sampled variance. Most of the information about fragmentation was captured by the first axis, which explained 49·60% of the sampled variance. The first axis was negatively correlated with metrics which increase with habitat fragmentation (e.g. number of patches) and positively correlated with metrics which decreased with fragmentation (e.g. mean patch area, Table 3). Additionally, the first axis was strongly correlated with those fragmentation metrics which were least correlated to habitat amount (including total edge, fractal dimension and clumpiness index). The second axis of this PCA primarily represented habitat amount; it was strongly correlated with metrics which were strongly correlated to habitat amount (e.g. mean patch area and patch cohesion). Only the first axis of the PCA of residual fragmentation metrics was significant; it explained 55·09 % of the sampled variance and was otherwise very similar to the first axis of the PCA of initial fragmentation metrics.

Discussion

Fragmentation limits invasion

Bothriochloa ischaemum presence was negatively related to the degree of fragmentation of the surrounding landscape, and to the overall gradients of fragmentation represented by principal components. (Recall that we are referring to the fragmentation of the herbaceous habitat, not the encroaching woody matrix. Also, note that some fragmentation metrics increase as fragmentation decreases; we refer here and below to fragmentation itself, not to the signs of particular fragmentation metrics.) These relationships support our hypothesis that fragmentation slows the spread of this invasive species, and by extension the spread of other species, particularly those with limited dispersal. The negative relationships between B. ischaemum presence and fragmentation were more evident when habitat loss was removed from the measures of fragmentation, strengthening the conclusion that fragmentation per se had affected B. ischaemum distribution. Studies of B. ischaemum dispersal and establishment may reveal the mechanisms underlying these relationships.

Almost all of the separate measures of fragmentation revealed the expected negative relationship between fragmentation and B. ischaemum presence (Table 4). We suspect that the effect of fragmentation on B. ischaemum is primarily due to its effects on colonization rates. High patch cohesion, high clumpiness, low total edge, and low landscape shape index (all indicators of low fragmentation) are indicators of a landscape in which B. ischaemum seed has fewer barriers to dispersal. Only two measures of fragmentation, nearest-neighbour distance and patch area, did not show a clear negative relationship between B. ischaemum presence and fragmentation of the surrounding landscape. There was a positive relationship between B. ischaemum presence and the distance between patches (mean nearest-neighbour distance). Fragmentation generally leads to increased patch isolation (Fahrig 2003), but the distance between patches may decrease as other aspects of fragmentation increase with habitat loss (Tischendorf 2001; Fahrig 2003), as it did in our system. As woody plant encroachment proceeds and fragmentation increases, the break-up of large patches creates groups of new patches that are close together. Mean patch area, the only fragmentation metric never significantly related to B. ischaemum presence, may be more strongly related to extinction rates than colonization rates. In our system, there is a highly irregular and interlaced arrangement of herbaceous and woody habitat; for example, a single ‘patch’, as defined by contiguous pixels of herbaceous habitat, can look like beads on a string (Fig. 1). For this reason, the other fragmentation metrics better describe this landscape.

Site-to-site variation

Bothriochloa ischaemum presence was most closely related to fragmentation at site PF, probably because this species was so common at the other two sites that the relationship was obscured. Together with the conclusion of Gabbard & Fowler (2007) that the potential distribution of B. ischaemum in this region is limited only by the availability of open (i.e. non-woody) upland area, this suggests that the invasion of B. ischaemum at sites BC and FR was much further advanced than it was at PF. As a working ranch, FR had the most off-road traffic and was the only site being grazed during this study; both vehicles and cattle may have dispersed B. ischaemum seed and thus weakened the effects of fragmentation there. The extensive management of woody plant encroachment by prescribed fire and mechanical removal at BC may have obscured the relationship between B. ischaemum and fragmentation by changing habitat configuration and reducing the number of highly fragmented landscapes at BC in comparison to other sites.

Relationship between invasion and habitat loss

Because habitat fragmentation occurs with, and covaries with, habitat loss, we might expect that B. ischaemum would be rarer in landscapes with less herbaceous habitat (more woody cover). However, at site PF B. ischaemum occurred more often in landscapes with less herbaceous habitat. A possible explanation is that habitat loss may have facilitated the invasion of the remaining habitat by altering the composition of the native herbaceous community (Shea & Chesson 2002). In fact, plots in landscapes with less herbaceous habitat did have fewer native species (K. M. Alofs, unpublished data). Species can be lost as the minimum amounts of habitat required for populations to persist in landscapes are reached and populations in more fragmented landscapes may require more habitat to persist (Fahrig 2002). The negative relationship between the amount of suitable habitat in a landscape and B. ischaemum presence may have been detectable only at site PF because the other two sites had enough herbaceous habitat overall to maintain native species populations.

Habitat fragmentation per se vs. habitat loss

Few other empirical studies have measured the effects of fragmentation per se or compared them to the effects of habitat loss (Simberloff 2000; McGarigal & Cushman 2002; Fahrig 2003). These studies have found that the relationship between fragmentation and native biodiversity can be positive or negative (Fahrig 2002, 2003), the effects of habitat fragmentation may be important only when habitat amount is low (Andrén 1994; Betts et al. 2006), and the effects of habitat loss tend to be stronger than those of habitat fragmentation (Fahrig 2002, 2003). In contrast, B. ischaemum occurrence was more closely related to habitat fragmentation than it was to habitat amount (the former relationship was detected at all three sites, the latter only at one site). This may reflect the fact that the published studies are mostly of species or population persistence following habitat fragmentation, while the present study is of the spread of an invasive species through fragmented habitat.

Limitations of this study

We detected a negative relationship between B. ischaemum presence and fragmentation despite the limited sample sizes of this observational study and despite several factors that would be expected to obscure or weaken the underlying relationship. We know very little about the history of B. ischaemum at these sites; the dates and locations of its introductions might help to explain some of the unexplained site-to-site and plot-to-plot variation in B. ischaemum distributions. As mentioned above, a better understanding of B. ischaemum dispersal mechanisms and the size and shape of its dispersal kernel might help explain some of the observed variation. Finally, as discussed above, fragmentation may have reduced native species richness, thus making communities more susceptible to invasion, and to some degree counter-acting the direct negative effects of fragmentation on B. ischaemum.

Management implications

The negative relationship between fragmentation and the presence of this invasive grass suggests that habitat fragmentation can limit the spread of invasive species, particularly those with limited dispersal. Land managers should consider fragmentation as a tool. For example, where woody plant encroachment is occurring, as it was in our sites, woody plant removal that creates small isolated open patches rather than corridors between patches or a single large open area may be effective in slowing the spread of B. ischaemum and similar invasive species. Additionally, managers should consider maintaining strips of woody plants or other unsuitable habitat as potential barriers to the spread of invasive species (With 2002), particularly between infested and pristine sites. In our system, a strip of the width of one adult Q. fusiformis or J. ashei may be sufficient to slow the spread of B. ischaemum between patches. Areas protected or encircled by such barriers may be good sites in which to invest resources for invasive plant eradication and the restoration of native communities. Although fragmentation is unlikely to permanently protect an area from invasion, it may slow down invasion sufficiently to make these management efforts worthwhile.

Although our results show that increasing landscape connectivity or creating corridors may facilitate the spread of an invasive species with limited long-distance dispersal, this should not be interpreted as implying that corridors are not useful in many conservation scenarios. In each case, it is important to examine whether the net effect of fragmentation is positive or negative (Levey et al. 2005).

Acknowledgements

This work was funded by a NSF Doctoral Dissertation Improvement Grant (DEB-0710348), the NSF Graduate Research Fellowship Program and an Ariel Appleton Fellowship from the Research Ranch Foundation. We thank the staff at Balcones Canyonlands National Recreation Area, Pedernales Falls State Park and Freeman Ranch (Texas State University) for access to research sites. We are grateful to S. Meadows, G. Casares, L. Aguilar, H. Gillespie, H. López-Fernández and A. González for their assistance with data collection. Two reviewers and the editor provided helpful comments.

Ancillary