A significant challenge for conservation science is to understand the needs of wide-ranging species during the design of nature reserves. Although numerous GIS methods exist to analyze patch composition and structure, few methods readily allow a user to incorporate species behavior and scaling. Here I describe a general approach that allows functional properties of organisms and processes to be modeled so that possible habitat fragmentation due to land use change (including housing development and roads) can be assessed. Results of this modeling approach are illustrated through landscape connectivity maps for lynx (Lynx canadensis).
A second problem is that the space between patches of habitat the matrix is typically considered to be biologically inert and isolation is computed as the Euclidean distance between patches (e.g., Doak et al. 1992). Yet, species movement or dispersal mechanisms are influenced not just by the inter-patch distance, but by the characteristics of the matrix such vegetation type, structure, and land use (Wiens et al. 1993). A third aspect is to recognize that patches, while often conceptualized as a point or "node" in reality have a spatial extent and are often irregular in shape. Much of the connectivity literature has ties to meta-population theory, which in most cases represents a patch as a single point -- usually a polygon centroid. Simplifying the geographic complexity of patches is generally done to provide adequate performance. (e.g., RAMAS GIS). When the inter-patch distance is very large relative to the variation in distances from patch centroid to patch edge, this assumption is reasonable (Moilanen and Nieminen 2002). However, edge-to-edge measures are preferred (e.g., Keitt et al. 1997), are straightforward to calculate in GIS, and are more robust to geographic complexity (Crooks 2002).
A fourth aspect is that the spatial configuration of patches is important. That is, some patches are more important than others, not because of size or habitat quality, but because of the position of a patch within the network. Critical connector patches are located at a nexus where the overall landscape connectivity is sensitive to inclusion or removal of these patches. A number of specific purpose GIS-based modeling efforts have begun to build landscape connectivity models. These have typically taken advantage of cost-distance methods in raster GIS. For example, Walker and Craighead (1997) used least-cost path modeling to identify grizzly bear corridors. Bunn et al. (2000) also used least-cost paths but used graph-theoretic metrics to better quantify how the configuration of patches related to landscape connectivity.
Building on this previous research, I describe a general modeling
approach here that incorporates an organism-centric perspective to define
landscape connectivity, especially to examine the consequences of landscape
changes such as development and roads and their possible fragmenting
effects. This approach has the following advantages:
1. Gradients in habitat quality can be recognized so that binary species- vegetation modeling can be improved to reflect species-specific use of the landscape. This also provides a way to include uncertainty in land cover data and can incorporate edge effects.
2. Patches are defined relative to a species' mobility. Generally, this is related to levels of organization, so patches refer to individual population level, where foraging behavior and home range requirements can be used to estimate the minimum patch size.
3. Matrix quality (the intervening area between patches) is explicitly recognized by modeling the "resistance to movement" through different land cover types. For example, for a lynx, it may be relatively easy to move through sagebrush, because it provides some cover, while traveling through open grasslands is more risky, and traveling through human-occupied land (urban areas) is typically avoided.
4. The potential direct and indirect (adjacent) affects of land use on habitat are explicitly incorporated. Known effects of different land use types and roads can be incorporated. This addresses important questions such as: are we likely to have increased fragmentation because of new development?
Using the general approach described above, I created a general, GIS-based tool to model landscape connectivity. The tool is an ArcView v3.2 extension that allows users to model landscape connectivity based on user-defined input parameters. There are 3 basic steps to the model: defining habitat quality, defining matrix quality, and establishing clusters of habitat.
First, a map of habitat quality needs to be prepared. This usually relies on species-affinities for certain land cover types, but other factors such as elevation gradients, distance from water, etc. can be incorporated. Then, habitat quality (values can range from 0 to 100, not just 0 or 1) is then "integrated" using the species scale of foraging (Addicott et al. 1987). Patches that are below a minimum size (e.g., related to minimum home range size) can be removed. This results in a map that represents functionally-defined patches of habitat.
Matrix quality is also derived from a land cover map and reflects the ability of a species to move through various land cover types. The matrix quality map is the "cost GRID" input to the cost-distance function (the habitat patches map supplies the "seed" locations). The matrix quality map can be best thought of as a 3-D surface, where areas of high resistance are peaks and ridges, and areas of low resistance to movement are the valleys and low areas. The path of a species moving from one patch to another would avoid peaks and ridges and instead travel around "contouring" around an area, incurring the "least-cost" path, rather than necessarily the shortest Euclidean distance.
Although using cost-distance weights or resistances for different types of land cover instead of Euclidean distance makes reasonable sense to most biologists, a significant challenge is to estimate the resistance parameters. Commonly, cost-distances for different types are simply estimated through "expert opinion", by applying knowledge of a species' life history characteristics and habitat use. An innovative example of this was Boone and Hunter (1996) examination of grizzly bear movement assuming different sensitivity to land cover types. An empirically-based method is to generate cost-distance parameters directly through examining movement data. For example, Palomares et al. (2000) and Ferreras (2001) used movement data collected through radiocollars to relate the proportion of use of different habitat types (using Jacobs' index) to the "friction" or resistance. Ricketts (2001) used maximum likelihood to estimate the resistances of matrix types for butterfly movement.
For each scenario run, a new view (named with a time-stamp) is
created and GRID output files are placed inside the view. Grids that begin
" provide output for patches of habitat that meet the
minimum requirements specified. GRIDs that begin with "Clusters
output for clusters of patches that are "connected" by dispersal ability and
matrix quality assumptions. Use the "Unique Value" legend to display, so
that same-colored patches are connected. Note that the GRIDs inside the
output view are "temporary" GRIDs, in the ArcView sense, so that they
physically located in the working directory. If you want to save these GRIDs,
make these permanent or copy them over into another directory that will not
get periodically cleaned.
Note: the purpose of this example is to illustrate the modeling process in the
Southern Rocky Mountain Ecoregion (central Colorado, south-central
Wyoming, and north-central New Mexico), not to produce a definitive map of
lynx connectivity. Four major steps in the process are described.
The first step is to create the habitat map, and for the lynx in the Southern
Rocky Mountains I converted National Land Cover data to create a habitat
map (Figure 1). The 30 m resolution was aggregated to 300 m resolution
using BLOCKSTATS (with sum) to produce a GRID with values ranging from
0 to 100, or the proportion of a cell that is habitat. Using a coarser GRID
was needed to ensure that the cost-distance functions perform reasonably.
Based on housing density values, the effective area of habitat in a cell was
reduced (e.g., urban 100% loss, suburban 75% loss, exurban 50%
loss). Cells that had a highway run through them had a 100% loss of habitat
(while smaller roads had no reduction).
Figure 1. Creation of habitat map for Lynx in the Southern Rocky Mountain Ecoregion. The coniferous and mixed forest cover types were extracted from the USGS National Land Cover Dataset (30 m) represent, in a course way, critical habitat for Lynx.
The second step is to determine habitat patches that meet the minimum
area requirement. Those patches that have an area smaller than A are
removed (Figure 2).
Figure 2. Removal of patches that are below the minimum area requirements. Different color patches represent separate habitat patches, and patches that are to be removed are shown in Red (Left) and subsequently removed (Right).
A third major step is to create the matrix quality map by reclassing the land cover map to reflect assumed resistance to movement (Figure 3). Resistance values are described in a reclass table (Table 1). Note that different assumptions about sensitivity of a species to movement through different types of land cover can be represented.
|Cover type||Low Sensitivity||High Sensitivity|
The fourth step is to identify "clusters" of patches those that are within the
prescribed dispersal distance (Figure 4). That is, a single path between
patches is not assumed, rather, patches are found to be "connected" if any
part of a patch can reach any part of another patch assuming a given
dispersal distance (that reflects matrix quality). Using this methodology
provides much improved computer performance, as compared to finding the
Figure 4. Finding "clusters of habitat patches. Individual patches are shown in unique colors (left). Groups of patches that are within the dispersal distance are considered clusters. For example, cluster "A" is composed of patches: 1, 2, and 3; "B" is composed of 4, 5, and 14; etc.
This modeling approach supports two modes of use. The first is that if
specific parameters are available for a given species, then those can be input
and the model results interpreted. But, a more robust mode of use is to
examine how connectivity changes over a range of parameter values (e.g., a
sensitivity analysis). The model input is set to handle this, as each set of
major parameters allows a list of parameter values. This enables a critical
range or threshold to be identified. For example, given a range of
sensitivities and dispersal ability, the output of a series of models can be
grouped to illustrate critical scales that might result in markedly different
connectivity regimes (Figure 5). The ability of a single species can be
compared directly to this critical scale. Also, a broad range of species can be
filtered to determine if they are likely to be sensitive to fragmentation.
Figure 5. Identifying critical scales. A critical scale of dispersal ability occurs between 10 and 30 km.
Also, by examining the output maps directly, locations on the landscape
that are critical to maintain connectivity can be identified (Figure 6). That is,
the "pinch-points" or "bottle-necks" to landscape movement can be readily
identified (Figure 7).
Figure 6. Dispersal ability ranging from 1 km to 10 km to 100 km. Different colored areas represent distinct clusters of patches, so that at short dispersal distances (Left), roads (grey lines) are seen to delimit patches and create a fragmented landscape. With greater dispersal ability, landscape connectivity increases, and important locations (I-70 west of Vail) become apparent (middle). At large distances, the landscape is fairly-well connected.
Figure 7. Pinch-points in the Southern Rocky Mountains for lynx.
Because land use and roads are explicitly incorporated in the model, and alternative scenarios can be easily generated, fragmentation can be assessed in a meaningful way by comparing natural vegetation patterns to a human-modified landscape. Often, landscape fragmentation is being caused by modifications in the intervening matrix that are occurring at a much faster rate changes directly to the habitat. Given these situations, models of landscape fragmentation must incorporate direct changes to matrix. Ideally, model- and field-based inquiry operate in tandem. This modeling approach was developed to provide improved biological realism into understanding landscape-level processes, but models are typically challenged and rightly so -- on the basis of the empirical data that provide the foundation for the estimated input parameters. Modeling is valuable when field-based data are not available. An interesting finding of a model of black-bear dispersal near Banff, Canada (Clevenger et al. 2002) suggests that a model based on expert literature performed better, as assessed by comparison to an empirical model, than an expert opinion-based model. Also, models can be used to conduct a preliminary analysis to determine what field parameters are the most important to collect.
Most models have relied on the least-cost path to derive cost-distances between patches (e.g., Bunn et al. 2000; Singleton and Lehmkuhl 2000). Least-cost paths are presumed to better approximate the actual distance a species moves, as compared to Euclidean distance, but estimating these paths is demanding on computer resources (Bunn et al. 2000). Part of the demand is related to computing distances for all pairs of patches though assumptions of uni-directional flow require only roughly half of the pairs -- O(n(n-1)/2).
A refinement to this standard methodology uses "landscape graphs"
work in progress -- that differ from traditional graphs (sensu Bunn et al.
2000; Urban and Keitt 2001; Wallis 2000) in two ways. First, the topology of
landscape graphs is computed by establishing adjacencies through Thiessen
polygons (computed either with Euclidean- or cost-allocation methods).
Practically this reduces the number of edges (connections between nodes or
patches) in a graph and hence computation complexity -- and an adjacency
list stores only connected features (Theobald 2001). Second, instead of
computing inter-patch distance using the least-cost path, I find the average
distance of paths that travel through the shared boundary of the habitat
patches allocation polygon. Though this has not been tested, this more
reasonably approximates the approach of individual-based models that allow
for a probabilistic interpretation of species moving across a landscape (e.g.,
Boone and Hunter 1996; Shippers et al. 1996). The connection between
patches is represented, not as a straight-line edge between nodes, but as a
polyline that begins at a patch edge, passes through the mid-point of the
shared boundary of the Thiessen polygons, then ends at the adjacent patch
edge. This allows a somewhat more realistic portrayal of the linkage