Landscape Ecology

, Volume 26, Issue 5, pp 617–628

Matrix is important for mammals in landscapes with small amounts of native forest habitat

Authors

    • School of Geography, Planning & Environmental ManagementThe University of Queensland
  • Clive A. McAlpine
    • School of Geography, Planning & Environmental ManagementThe University of Queensland
  • Hugh P. Possingham
    • The Ecology Centre, School of Biological SciencesThe University of Queensland
  • Craig J. Miller
    • CSIRO Sustainable Ecosystems
  • Greg S. Baxter
    • School of Geography, Planning & Environmental ManagementThe University of Queensland
Research Article

DOI: 10.1007/s10980-011-9602-6

Cite this article as:
Brady, M.J., McAlpine, C.A., Possingham, H.P. et al. Landscape Ecol (2011) 26: 617. doi:10.1007/s10980-011-9602-6

Abstract

Acknowledgment that the matrix matters in conserving wildlife in human-modified landscapes is increasing. However, the complex interactions of habitat loss, habitat fragmentation, habitat condition and land use have confounded attempts to disentangle the relative importance of properties of the landscape mosaic, including the matrix. To this end, we controlled for the amount of remnant forest habitat and the level of fragmentation to examine mammal species richness in human-modified landscapes of varying levels of matrix development intensity and patch attributes. We postulated seven alternative models of various patch habitat, landscape and matrix influences on mammal species richness and then tested these models using generalized linear mixed-effects models within an information theoretic framework. Matrix attributes were the most important determinants of terrestrial mammal species richness; matrix development intensity had a strong negative effect and vegetation structural complexity of the matrix had a strong positive effect. Distance to the nearest remnant forest habitat was relatively unimportant. Matrix habitat attributes are potentially a more important indicator of isolation of remnant forest patches than measures of distance to the nearest patch. We conclude that a structurally complex matrix within a human-modified landscape can provide supplementary habitat resources and increase the probability of movement across the landscape, thereby increasing mammal species richness in modified landscapes.

Keywords

Terrestrial mammalsMixed effects modelsLandscape mosaicAustralia

Introduction

Persistence of wildlife in human-modified landscapes depends on the heterogeneity of habitat resources within the entire landscape mosaic (Bennett et al. 2006). However, the relative importance of local-level habitat factors, landscape-level habitat factors and the human-modified matrix are still poorly understood. Local habitat factors such as habitat structural complexity can have an important influence on predation risk (Bertolino 2007) and local population dynamics (Spencer et al. 2005), while landscape configuration can have an important influence on dispersal success and therefore population structure, species persistence and richness (Henein et al. 1998; Westerberg et al. 2005). A landscape mosaic is considered here to encompass original habitat, including modified habitat such as ‘edge’ and matrix of varying land use. The landscape matrix, defined here as the human modified area of the landscape that surrounds remnant (or native) habitat, has historically been ignored and considered non-habitat (e.g., island biogeography theory); however, its importance is now increasingly being acknowledged (Pita et al. 2007; Franklin and Lindenmayer 2009). The matrix can be important for both its influence on animal movement through the landscape (Ricketts 2001; Hein et al. 2003; Revilla et al. 2004; Bender and Fahrig 2005) as well as for the potential resources it can provide (Tubelis et al. 2007; Harper et al. 2008), especially in landscapes where fragmentation or the amount of remnant habitat are major limiting factors (Tubelis et al. 2004).

Previous attempts to disentangle the importance of multiple elements of the landscape mosaic have met with various difficulties. As landscapes become modified and more intensely developed for human use, multiple factors that influence wildlife often change simultaneously. For example, habitat loss and fragmentation often exhibit collinearity with both anthropogenic activities and habitat condition (Moffatt et al. 2004). The amount of natural habitat available both in an individual patch and across a landscape is unquestionably critical to the persistence of a rich native faunal assemblage (Radford and Bennett 2007). However, their collinearity with matrix and patch attributes has obscured the relative importance of landscape mosaic properties (Ewers and Didham 2006). Other studies have not measured all relevant attributes across the entire landscape mosaic. Pita et al. (2007) found that matrix attributes were more important than patch attributes for patch occupancy and spatial population structure of a threatened mammal habitat specialist (Microtus cabrerae), although landscape-scale habitat attributes could be correlated with matrix attributes, obscuring their relative importance. Vieira et al. (2009) found that landscape configuration was more important than matrix attributes for small mammal richness but did not consider potential habitat attributes of the matrix that could be correlated with native habitat configuration and could be contributing to the results.

In spite of the increasing recognition of the importance of attributes across the entire landscape mosaic, overlooking potential habitat attributes of the matrix is still common in both research and conservation practice. Conservation strategies often have a patch-centric focus maintaining habitat quality in patches, or creating stepping stones or corridors of habitat to aid wildlife dispersal between patches. The landscape context of the remnant patch habitat and the land use intensity or habitat attributes of the matrix that could potentially have positive and negative effects on wildlife are often ignored. Information gained from unravelling the relative importance of attributes of the entire landscape mosaic, including the matrix, will help guide on-ground conservation efforts in directing resources to where they will have the greatest benefit, while also advancing ecological theory and landscape conceptualisation regarding species persistence in human-modified landscapes.

We addressed the question: how important is the matrix to mammal species richness, relative to other landscape attributes (e.g., distance to nearest patch) and patch attributes (e.g., habitat structure), while holding the amount of remnant forest habitat (as both patch size and total amount in landscape) relatively constant across all landscapes? To address this question, we developed an a priori multi-scale conceptual model of mammal species–environment relationships within human-modified landscapes (Fig. 1) and from this tested a small set of models. Within the conceptual model are four different testable models that also represent alternative management actions for mammal conservation. Embedded in the model are influences at the local patch level (‘patch model’), the landscape level (‘landscape configuration model’), as well as landscape matrix attributes (‘matrix model’), which each have an individual effect on mammal species richness, while combined they have an effect as the whole landscape mosaic (‘mosaic model’).
https://static-content.springer.com/image/art%3A10.1007%2Fs10980-011-9602-6/MediaObjects/10980_2011_9602_Fig1_HTML.gif
Fig. 1

Conceptual model of the factors influencing mammal species in modified landscapes of eastern Australia, including final variables used to represent these factors in statistical modelling in parentheses. See Table 1 for variable descriptions

Patch model

This model contains attributes at the patch level important to mammal species richness in modified landscapes including vegetation structure (Knight and Fox 2000; Garden et al. 2007), resource availability (Mortelliti and Boitani 2008), nature of the edge (Lidicker 1999) and disturbance (Fox and Fox 2000; Goosem 2000). Patch size and floristic attributes were controlled (see Brady et al. 2009 for details).

Landscape configuration model

This model contains two attributes of spatial configuration of remnant forest important to mammal species richness in modified landscapes. The amount of remnant forest in landscapes was controlled at low levels (~15%). Proximity of habitat patches affects dispersal of individuals and can influence the potential for the ‘rescue-effect’ of populations (Brown and Kodric-Brown 1977), and therefore the long-term viability of species in modified landscapes. This can be especially important for species that need to use multiple patches within their home range to survive and/or if most individual patches are too small to maintain populations of some species in the face of environmental stochasticity (Hanski and Ovaskainen 2000).

Matrix model

The matrix model contains variables relating to human development intensity of the matrix, presence of feral predators, as well as habitat attributes of the matrix such as vegetation structure and potential resources that are important to mammals in modified landscapes. These variables are hypothesised to affect the likelihood of species’ safely using the matrix, whether for supplementing resources or crossing to other areas of resources including finding mates. Species are often better able to persist in a modified landscape if there are also resources to use in the matrix (Harper et al. 2008).

Mosaic model

The mosaic model is the ‘global’ highly parameterised model, containing all variables from other models. All variables combined may best explain mammal species richness because the whole landscape mosaic acts as a unit, important to species movement around, and persistence in, modified landscapes. Variables from each model may be important, such that combined they account for more variability in species richness.

Building a small set of models, each with sound ecological support to clearly represent multiple working hypotheses, allows confirmatory inference about the importance of each model (Burnham and Anderson 2002). We applied a generalised linear mixed effects modelling and model averaging approach to quantify the effect size and the relative importance of matrix, patch and landscape level attributes for native terrestrial mammal species richness. Species richness was chosen because: (1) it allows reflection on a large body of previous literature examining landscape patterns, including theoretical work; and (2) it has been suggested that information for groups of species can be more useful to conservation planners and land managers than for single species (Huggett 2005).

Methods

Study area

The study area was Toowoomba Regional Council in subtropical southeast Queensland, Australia (see Brady et al. 2009 for map of study area). Throughout the area, historical grazing land is in various stages of conversion to low density rural residential and high density residential land uses. Native forest mainly persists as small isolated remnants, apart from larger remnant patches on steep escarpment land. On average 14.7% (±3.3%) remnant vegetation remains within the study landscapes (Brady et al. 2009).

Study design

A study landscape was defined as the area within a 500 m radius (78.5 ha area) around a focal patch (remnant tall forest) edge. Nineteen study landscapes were selected based on rigorous criteria related to matrix land use type and intensity, focal patch characteristics such as size, shape, remnant vegetation type and location in the landscape, landscape composition, and land use history (see Brady et al. 2009). Criteria were assessed by analyses in ArcView 9.2 (Environmental Systems Research Institute, Redlands, California), extensive ground-truthing, plus consultation with landholders. Landscapes were only sampled if matrix sites were able to be located in areas that were representative of the entire matrix area (by way of matrix land use type, intensity and history) around the focal patch. Within each landscape, three site ‘types’ were surveyed: patch ‘core’, patch ‘edge’ and ‘matrix’, giving a total of 57 sites. Patch edges were defined by the limit of the continuous forest canopy (Harper et al. 2005). Edge sites were located along this edge, and were equidistant (50 m) between the core and matrix sites; survey design diagrams are available in Brady et al. (2009). This landscape design allowed sampling of whole landscapes while ensuring different site types within landscapes shared similar influences. Landscapes spanned the widest and most evenly spread range of “matrix development intensity”, quantified by a weighted road length (WRL) metric that considers multiple road attributes, giving importance to the ecological impact of roads within study landscapes. This measure is derived from the entire matrix area in the landscape. Matrix development intensity was also highly correlated with landscape housing density.

Mammal surveys and response variable

Small and medium sized terrestrial mammals were surveyed by a combination of Elliott traps, wire traps, hair funnels, scats and direct sightings; further details in Brady et al. (2010) (supplementary material). Sampling effort was equal across all sites and landscapes. All 57 sites were sampled twice, once in spring 2007 and once in summer 2007–2008 when study animals are expected to be most active, with at least 4 months between first and second sampling, resulting in 11,970 detection nights. Data from the first and second sampling were combined. The response variable, species richness, was calculated as the number of mammal species detected at each site (see Fig. 2).
https://static-content.springer.com/image/art%3A10.1007%2Fs10980-011-9602-6/MediaObjects/10980_2011_9602_Fig2_HTML.gif
Fig. 2

Abundance (minimum number known alive; MNKA) of species in core, edge and matrix landscape elements

Explanatory variable selection and exploratory data analysis

Thirty-six environmental variables were measured across multiple spatial scales and a full floristic survey was conducted for each core, edge and matrix site, during summer of 2007–2008. These environmental variables were used to characterise the matrix intensity gradient by a detailed analysis of habitat attributes across the study landscapes (Brady et al. 2009). From these variables, we a priori selected 12 explanatory variables that best represented the attributes of interest in the conceptual model. Spearman’s Rank correlations were performed between all variables. When explanatory variables were highly correlated (r > 0.5, Booth et al. 1994), the variable that showed higher importance to species richness in a Principal Components Analysis (run in CANOCO 4.5) was kept, resulting in ten final explanatory variables; four matrix variables, four patch variables and two landscape configuration variables (Table 1).
Table 1

Descriptions and summary statistics for explanatory variables used in predictive models

Variable (code)

Description

Range

Mean

SD

Matrix development intensity (MI)

Length of road in landscape (500 m radius) weighted by ecological impact and correlated with housing density. Described in detail in Brady et al. (2009)

0–29,620

10,585

7540

Feral predators (MPETS)

No. domestic pets (dogs and cats) in matrix determined by survey of all properties in matrix site

0–14

4.6

3.5

Hides and roosts in matrix (MHIDROO)

Sum of number (count) of dead trees, tree hollows, old stumps, logs and other potential hides and roosts such as ground cavities, rocks and human garbage across entire matrix site

0–7

1.9

1.7

Litter cover (CLICOV)

Mean % ground covered by litter estimated in four randomly placed 1 × 1 m quadrats

52–100

81.2

16.5

Tree layer <10 m high (CTREL10, MTREL10)

Number (count) of trees <10 m high in four randomly placed 10 × 10 m quadrats. Core and matrix values treated as separate covariates in analysis

0–35 (c)

0–15 (m)

10.3 (c)

3.6 (m)

10.3 (c)

4.8 (m)

Edge disturbance (EMICDEN20)

Mean number (count) of stems at 20 cm above ground in 50 cm radius from 4 randomly placed points

1–62

25.7

16.4

Edge contrast (EDGHARD)

Categorical ranking of 0 = Indistinct, 1 = Low, 2 = Moderate, 3 = Severe for distinctness of patch edge and matrix habitat structure

1, 2, 3

Nearest neighbour (NEARNEIGH)

Distance (m) to closest remnant vegetation patch of any size from any edge of focal patch, measured in ArcGIS and ground-truthed

90–780

279.5

165.8

Nearest large patch (NEARLGE)

Distance (m) to closest remnant vegetation patch >20 ha from any edge of focal patch, measured in ArcGIS and ground-truthed

350–2520

971.1

573.4

Abbreviations in parentheses are analysis codes. First letter of Analysis Code denotes the site type that variable was measured in (c patch core, e patch edge, m matrix)

Responses to landscape modification are often highly species specific. Because we were interested in terrestrial species richness, we selected explanatory variables that were more likely to be important for a rich assemblage than those that might evoke strongly different responses from species. We also tried to choose variables important for species movement throughout modified landscapes, regardless of the reasons for movement or the sensitivity of a species to fragmentation or matrix effects. For example, the response of mammals to the structure and density of shrub layers is species specific (Hockings 1981), often dependent on body size. For this reason, we chose an understorey tree layer; cover of trees less than 10 m high, as a measure of structural complexity that may be more important to mammals as a group. This is also the layer often missing in disturbed habitats, patch or matrix, especially those historically disturbed by grazing (such as in our study region), due to a lack of tree regeneration and might also represent overall ecosystem health or resilience. The density of stems at 20 cm above ground along the edge of the patch was related to the level of disturbance at the patch edge including degree of vegetation thickening and weed invasion. Edge conditions related to edge permeability and patch–matrix similarity can be important for functional landscape connectivity and species use of the matrix (Haynes and Cronin 2006).

Statistical modelling

All statistical modelling was performed in R 2.7.0 (R Development Core Team 2008). We modelled native mammal species richness using generalised linear mixed-effects models using the ‘glmmML’ package version 0.81-3 (Brostrom 2008). The glmmML function estimates model parameters by maximum likelihood and calculates Akaike Information Criteria (AIC) values. We converted AIC to AICc (Sugiura 1978; Hurvich and Tsai 1989; Anderson and Burnham 2002). The general form of the multivariate Poisson models was:
$$ \ln \left( {E\left( {Y_{ij} } \right)} \right) = \beta^{\prime } X_{ij} + b_{i} , $$
(1)
where Yij is the species richness for site i in landscape j; β is a vector of model coefficients; Xij is a vector of explanatory variables for site i in landscape j; bi is a normally distributed random-effect for site i, with a mean of zero and variance of σ2.

By incorporating a common ‘random-effect’ mixed-effects models are a good way to account for spatial dependencies in hierarchically structured data (Rhodes et al. 2009). In our dataset, a core, edge and matrix site is clustered within each landscape. To ensure our modelling accounted for any spatial dependencies, we compared spline correlograms of the raw count data with spline correlograms of the residuals of all models tested. Correlograms were produced using the ‘ncf’ package version 1.1-1 (Bjornstad 2008). All explanatory variables were standardized to have a mean of zero and standard deviation of one to allow comparison of model parameter estimates.

Seven a priori models that represented all different combinations of the patch, matrix and landscape configuration variables from the conceptual model were fitted to the species richness data, with site ‘type’ included in every model, plus site ‘type’ was modeled on its own. An intercept-only model was also fitted. AICc values were used to assess the relative support for each model and determine a 95% confidence set of models. The relative probability of each model being the best model of the set (as opposed to the ‘truth’; Anderson and Burnham 2002) was also assessed by calculating the Akaike weights (wm) for each model, with a weight >0.9 indicating strong evidence for the model (Burnham and Anderson 2002). Goodness of fit of the landscape mosaic model (Anderson and Burnham 2002) and the most parsimonious model were assessed by the model’s residual deviance, and by visually assessing half-normal plots of the model residuals (Faraway 2006). The function ‘res.glmmML’ was used to calculate model residuals (Rhodes et al. 2009).

The relative importance of patch, matrix and landscape subsets of variables was derived by summing the Akaike weights across all the models where the variables occurred. The larger the sum of the weight, the more important the subset is, relative to the other subsets. The direction and magnitude of the effect size of each individual explanatory variable was determined based on model-averaged parameter estimates. We calculated the model averaged parameter estimates and their unconditional standard errors using Akaike weights over all models in the 95% confidence set of models (Σwm ≈ 0.95) in which the parameter occurred (Burnham and Anderson 2002).

Various a posteriori analyses were also conducted to help explore potential mechanisms contributing to the modelling results. Combinations of variables that potentially influence dispersal success such as matrix vegetation structure, resources available in the matrix and distances (across matrix) to further remnant habitat were selected. Interaction effects were also tested between a posteriori variables and site type. Models were tested by the same procedures as described above.

Results

Mammal species

Across the 57 sites sampled, 12 native species were detected, including nine small-medium sized mammal species and three medium-large macropod species. Highest species richness recorded at a site was five. Small mammal species included the bush rat (Rattus fuscipes), the pale field rat (Rattus tunneyi), the yellow-footed antechinus (Antechinus flavipes), the brown antechinus (Antechinus stuartii), the northern brown bandicoot (Isoodon macrourus), the long-nosed bandicoot (Perameles nasuta), the rufous bettong (Aepyprymnus rufescens) and the brushtail possum (Trichosurus vulpecula). The medium-large macropods included the swamp wallaby (Wallabia bicolour), the red-necked wallaby (Macropus rufogriseus) and the eastern grey kangaroo (Macropus giganteus). These species were pooled as a single macropod entity as identification to a species-level was not always possible. The brushtail possum (Trichosurus vulpecula) was the most frequently caught species but was not included in analyses because it is an arboreal species that responds positively to landscape modification. All eleven species included in species richness analyses were detected in patch core sites, five of the seven small-medium sized species and all medium-large macropod species were detected in patch edge sites and three of the seven small-medium sized mammal species and all medium-large macropod species were detected in matrix sites (Fig. 2).

Explanatory variables

Overall there was a low level of correlation among the final ten explanatory variables. However, the number of pets (i.e. feral predators; see Table 1) in the matrix (MPETS) was positively correlated with matrix development intensity (MI) (r = 0.558, P < 0.01) and edge hardness (EDGHARD) (r = 0.516, P < 0.01). Matrix development intensity was also correlated with two variables related to human disturbance that were removed from the analysis; human disturbance of patch cores (no. people encountered in patch core during sampling) (r = 0.593, P < 0.01) and level of human use of matrix directly adjacent to patch edge (no. people encountered during sampling) (r = 0.783, P < 0.001). Matrix development intensity was retained as a factor in models.
Table 2

Akaike Information Criteria values (AICc) for all a priori candidate models, model ranking according to their Akaike weights (wm) and the relative importance of variable subsets (matrix, patch or landscape) by sum of Akaike weights across models that contained the subset (denoted by ✓; four models each)

Models

Matrix variables

Patch variables

Landscape variables

AICc

wm

Matrix

  

50.13

0.674

Patch

 

 

53.3

0.138

Matrix + patch

 

54.0

0.097

Matrix + landscape

 

55.38

0.049

Site type only

   

56.80

0.024

Patch + landscape

 

59.37

0.007

Mosaic (all variables)

59.63

0.006

Landscape

  

60.04

0.005

Intercept-only

   

64.54

0.0005

Relative importance

0.826

0.248

0.067

  

Models above the dashed line are in the 95% confidence set of models

Ranking of a priori models and explanatory variable subsets

There was strong support for the matrix model as the best candidate model (Table 2). Evidence ratios (Burnham and Anderson 2002) indicated that the matrix model was 4.9 times more likely to be the best model than the patch model, which was the second ranked model and 134.8 times more likely to be the best model than the landscape model, which was the least likely of all parameterized models tested. The 95% confidence set of models (Σwm = 0.96), however, contained four models, revealing model selection uncertainty. Therefore there was insufficient evidence, based on the Akaike weights, that any single model was the best performing model for explaining mammal species richness. The residual deviance divided by the degrees of freedom was <1 for both the matrix model (0.681) and the landscape mosaic model (0.624) indicating that the data fit the models well and that there were no problems with over-dispersion of the data (Burnham and Anderson 2002). All points on the half-normal plot fell close to the 1:1 line suggesting that there are no major departures from model assumptions. Spatial autocorrelation was adequately accounted for in the residuals of all models. The matrix variables as a group were relatively more important than the patch and landscape configuration variables. The sum of the Akaike weights for the four models containing matrix variables was 0.826, while for the four models containing patch and landscape configuration variables the sum of Akaike weights was 0.248, and 0.067 respectively (Table 2).

Model averaged parameter estimates

The structural complexity of the matrix, represented by cover of trees <10 m high, had a strong positive effect on species richness (Fig. 3). Matrix development intensity followed by number of feral predators in the matrix had a strong negative effect on species richness. Nearest large forest neighbour, edge contrast and cover of trees <10 m high in the patch core had high unconditional standard errors, approximately two times larger than the parameter estimate, indicating high parameter uncertainty. Parameter uncertainty of edge contrast and core structural complexity reflects the high level of variability in their effect across model combinations.
https://static-content.springer.com/image/art%3A10.1007%2Fs10980-011-9602-6/MediaObjects/10980_2011_9602_Fig3_HTML.gif
Fig. 3

Model-averaged parameter estimates and their unconditional standard errors showing the direction and magnitude of effect of individual explanatory variables on mammal species richness, derived from models within the 95% confidence set of models (Table 3) that contained the parameter. Dashed lines indicate negative effects and line thickness is weighted by the model-averaged parameter estimate

A posteriori modelling

Based on the revealed importance of the matrix, models containing combinations of matrix development intensity (MI), cover of trees <10 m high in the matrix (MTREL10) and nearest neighbour distance (NEARNEIGH) were tested to explore potential mechanisms behind the a priori results. These attributes were considered to be influential on dispersal success across the matrix. Interaction effects were also tested between site type and: (1) the number of hides and roosts in the matrix (MHIDROO), (2) cover of trees <10 m high in the matrix (MTREL10), and (3) stem density at 20 cm on the patch edge (EMICDEN20) to further explore effects of these variables on species’ use of the landscape.

Of all models tested, the best was composed of the explanatory variables matrix development intensity and cover of trees <10 m high in the matrix (Table 3). This model had a lower AICc and a better fit to the data (deviance ratio = 0.71) than any other model tested, including models that had landscape configuration variables such as nearest neighbour distance. Species richness in the matrix increased with cover of trees <10 m high in the matrix (MTREL10) and this did not change with matrix development intensity. Species richness also increased in the matrix with the number of hides and roosts in the matrix, however, the effect was not as strong as for MTREL10. Species richness in the matrix decreased with edge density of stems at 20 cm above ground.
Table 3

A posteriori modeling exploring variables important to dispersal success across matrix and species’ use of landscape; Akaike Information Criteria values (AICc) and coefficients (coef) and their standard errors [SE (coef)] for interaction effects with site type

Variables

AICc

Combination models

 

 Site type + MI + MTREL10

48.38

 Site type + MI + MTREL10 + NEARNEIGH

49.54

 Site type + MI + NEARNEIGH

54.91

 Site type + MTREL10 + NEARNEIGH

54.93

Variables

Coef

SE (coef)

AICc

Interaction models

   

 Site type: EMICDEN20

  

62.82

 Core: emicden20

−0.4094

0.2099

 

 Edge: emicden20

−0.1618

0.2062

 

 Matrix: emicden20

−0.3835

0.2100

 

 Site type: MHIDROO

  

64.24

 Core: mhidroo

0.3039

0.1532

 

 Edge: mhidroo

0.1801

0.1703

 

 Matrix: mhidroo

0.2462

0.1610

 

 Site type: MTREL10

  

67.12

 Core: mtrel10

0.11565

0.1898

 

 Edge: mtrel10

0.08994

0.1922

 

 Matrix: mtrel10

0.28448

0.1740

 

Discussion

The primary aim of this study was to determine the relative importance of landscape mosaic properties for native terrestrial mammal species richness in highly modified landscapes of southeast Queensland, Australia. By controlling factors such as remnant forest habitat amount and patch size, that usually change along an urbanisation gradient and have often been confounded with matrix effects in previous studies, we have shown that the matrix can have an important independent effect on mammals. Our approach demonstrated that the character of the matrix between remnant forest patches can be considerably more important than the distance between those patches or local habitat attributes of the patches themselves. Furthermore, model averaged parameter estimates revealed that of all variables, matrix development intensity had the strongest negative effect, while structural attributes of (potential) habitat in the matrix had a strong positive effect on species richness. The modelling does not suggest that landscape configuration or remnant patch habitat attributes are unimportant per se, rather that in our study landscapes where only small amounts (~15%) of native habitat remain, the complexity of matrix habitat is the most important influence.

The matrix may be impacting mammal species through at least three mechanisms. Firstly, the matrix may be acting as a selective filter to dispersal (Anderson et al. 2007). Mammal persistence is often reliant on dispersal between patches in highly modified landscapes, with species in this study known to operate as metapopulations (Marchesan and Carthew 2008). The best model we tested contained two variables that can influence movement of a species across the matrix; matrix development intensity and cover of trees <10 m high in the matrix. Higher matrix development intensity can increase the resistance of the matrix to dispersal by reducing the likelihood of safe dispersal through greater predation by increased domestic or exotic predators (Riley et al. 2005; Campos et al. 2007) and a higher density of roads increasing chances of death by road-kill (Trombulak and Frissell 2000). McAlpine et al. (2006) found road density was more important than mean nearest patch neighbour distance to koala occurrence in fragmented rural–urban landscapes in southeast Queensland, Australia. In our study, increased cover of trees <10 m high in the matrix adds structural complexity providing cover and reducing perceived predator risk, without impeding foraging or locomotion in the way that structural layers such as dense shrub or grass layers can for some species (Vernes 2003). Presence of a tree layer may also create small stepping stones for species across the matrix. Regrowth or secondary forest in the matrix of a previously cleared landscape can allow highly fragmentation-sensitive species to re-colonise fragments after local extinction (Antongiovanni and Metzger 2005).

Secondly, the matrix could impact species through disturbance of remnant patch habitat by animals or humans emanating from the surrounding matrix, and this may be more important for habitat specialists that avoid the matrix (Pita et al. 2007). Pita et al. (2007) suggest that the mechanism through which the Mediterranean farmland matrix impacts on Cabrera voles (Microtus cabrerae) may be cattle from the surrounding matrix disturbing vegetation structure of habitat patches. In our study, matrix development intensity was highly correlated with human disturbance of patch cores (human disturbance was removed from further analysis) (Brady et al. 2009). This can result in wildlife avoiding otherwise suitable habitat (George and Crooks 2006). Disturbance of forest patches can have a greater impact on small habitat patches (Fox and Fox 2000) and can even result in more detrimental effects than patch size (Michalski and Peres 2005).

Finally, habitat attributes of the matrix can influence resources available to a species across the landscape and determine a species’ use of the matrix, not just for dispersal, but also for supplementing resources (Tubelis et al. 2007). An ability to use the matrix has been shown to increase an animal’s survival ability in modified landscapes (Laurance 1991; Viveiros de Castro and Fernandez 2004). Our a posteriori modeling revealed that species richness in the matrix increased with both the number of hides and roosts in the matrix and with the cover of trees <10 m high occurring in the matrix, and these effects did not change with matrix development intensity. Use of agricultural matrix by colobus monkeys (Colobus angolensis palliatus) in Kenya increased with food tree cover and matrix vegetative cover >6 m high (Anderson et al. 2007). Use by birds and small mammals of an exotic pine plantation matrix in Australia increased with age of the plantation (Lindenmayer and Peakall 2000; Tubelis et al. 2004), which could partly be attributed to increased resources and structural similarity to remnant habitat with matrix age (Tubelis et al. 2004). Collectively these results suggest that, regardless of the development intensity of the matrix, modified landscapes with a structurally complex matrix that can provide supplementary resources would conserve a richer faunal assemblage than a landscape with a structurally simple matrix.

Our results contrast, however, with results of Vieira et al. (2009) who found patch isolation (measured by shortest distance to nearest patch, i.e. the same metric as our study) was relatively more important than matrix attributes for small mammal richness in Atlantic Forest remnants of Brazil. Vieira et al. (2009) measured the type of economic activity (peri-urban, agriculture, cattle) and pattern of property ownership (peri-urban, low income rural, high income rural) to define the nature of matrix land use. The type of economic activity had a small effect on species richness in patches. While these measures of matrix land use could have distinct spatial and temporal patterns of vegetation composition and structure that could influence mammal use of the matrix and dispersal between patches, actual vegetation attributes of the matrix were not described. Small differences in matrix management activities can make important differences to matrix vegetation structure that can allow animal movement through the matrix (Prevedello and Vieira 2010). In Australian agricultural landscapes, for example, scattered paddock trees are important habitat elements (Fischer and Lindenmayer 2002; Manning et al. 2006), increasing the structural complexity and heterogeneity of the landscape mosaic (Haslem and Bennett 2008), but are under threat by changes in farming practices (Maron and Fitzsimons 2007). Bakker (2006) found that presence of structural microhabitat features in the matrix that were a component of forest habitat for red squirrels (Tamiasciurus hudsonicus) allowed greater movement of squirrels through the matrix. However, those same structural habitat attributes can be added or removed at the whim of individual landholders (Brady et al. 2009). We therefore advocate that in order to unravel matrix effects, potential habitat attributes of the matrix must be analysed to gain an understanding of processes operating at the scale of the species under study.

Implications for conservation

Our findings emphasize the necessity of explicitly acknowledging the importance of the matrix in any conservation strategy or land use planning process. For example, results of this study confirm the importance of any regrowth in the matrix with the number of trees <10 m high having the greatest positive effect on mammal species richness across landscapes. Therefore policies to protect regrowth of native vegetation can have positive benefits for biodiversity. To only focus on distribution and maintenance of habitat in designated parks and reserves may be insufficient for conservation of species on local to regional scales. However, our results do not say that matrix habitat alone can maintain species richness; therefore matrix management should not be done at the expense of patch habitat. Finally, from a landscape management perspective the matrix can be more important than patch attributes or landscape configuration for a very practical reason. In already developed landscapes there is often no opportunity to change patch or landscape habitat attributes, whereas with education, legislation and appropriate incentives individual landholders can alter matrix attributes enough to alter wildlife movement or habitat value of an entire landscape.

Acknowledgments

We gratefully acknowledge funding from the University of Queensland and CSIRO, B. Triggs for analysing traces, the 60 landholders who allowed access to their properties for surveys and the manuscript reviewers and editor for helpful suggestions.

Copyright information

© Her Majesty the Queen in Rights of Australia 2011