- Top of page
Conservation planning for populations of rare or threatened species is an inherently spatial process (Abbitt, Scott & Wilcove 2000). Delineating the spatiotemporal extent of a threatened species’ range is the first step in most conservation strategies. At finer scales, research, monitoring and planning efforts focus on identifying and protecting important habitat resources and reducing or mitigating limiting factors such as excessive sources of human-caused mortality (Loehle & Li 1996; Ceballos & Ehrlich 2002). The spatial representation of that information can serve as a powerful tool in the design and placement of ecological reserves, mitigating developments, directing remediation efforts and planning further research (Mladenoff, Sickley & Wydeven 1995; Flather, Knowles & Kendall 1998).
Various techniques are available to quantify and spatially represent species–environment relationships. Techniques range in sophistication from geographical information system (GIS) queries formulated from expert opinion to empirically derived statistical models (Guisan & Zimmerman 2000). The choice of technique is often guided by the availability of data, but consideration must also be given to project objectives and the ecology of the focal species (Austin 2002). Even when those criteria are considered, a number of issues, including inaccurate or unrepresentative data, poorly documented, complex or subjective methods, unclear interpretation of results and failures to evaluate the internal and external validity of predictions, can limit the usefulness of results and, in some cases, lead to incorrect conclusions when attempting to meet conservation objectives (Khagendra & Bossler 1992; Conroy & Noon 1996; Flather, Knowles & Kendall 1998).
Across central British Columbia, Canada, mountain caribou Rangifer tarandus caribou (Gmelin) are found at low densities (Heard & Vagt 1998). These animals inhabit mountain ranges during winter and primarily forage on arboreal lichens (Bryoria spp. and Alectoria sarmentosa), which are most abundant on old trees (Terry, McLellan & Watts 2000). During the past century, the distribution and abundance of mountain caribou has decreased considerably, leading to their being listed as endangered by provincial and federal conservation agencies. Proposed reasons for the decline and current threats include historical patterns of excessive hunting, loss of important habitats, reduction in connectivity of populations, increases in the distribution and abundance of predators, and displacement due to disturbance from industrial and recreation activities (Spalding 2000; Mountain Caribou Technical Advisory Team 2002).
Conservation of mountain caribou is facilitated through government-directed planning initiatives and legislation that recognizes and protects important habitats. Across the northern and central portion of mountain caribou range, these efforts are guided by small-scale maps that delineate areas as affording high, medium or low habitat values for caribou. The identification of boundaries and rankings was based on a number of sources including expert opinion, radio-telemetry and survey data, and continuous input from foresters and the public. However, data sources, map creation and evolution were poorly documented and habitat rankings are largely subjective.
We present a technique, resource selection functions (RSF), that combines GIS data and animal location information to generate spatially explicit predictive resource selection models. We used RSF to model and predict mountain caribou occurrence at two spatial scales: resources important within the vegetative patch; and topographic factors limiting the distribution of mountain caribou across multiple watersheds we term landscapes. We assumed patch occupancy was conditional on topography at the larger landscape scale and generated a single map predicting caribou occurrence across central British Columbia. Our principal objective was to develop and implement methods necessary to refine existing maps used by provincial land management agencies to identify and rank mountain caribou habitats. Secondarily, we evaluated and discussed the strengths and limitations of predictive RSF as a tool for conservation planning of rare and threatened species.
- Top of page
In general, the presented RSF coefficients agree with previous studies of habitat selection by the Yellowhead and other populations of mountain caribou found across British Columbia. During early winter, caribou of the Yellowhead population selected for mid-elevation forests on slopes of 16–30% that were dominated by subalpine fir (Terry, McLellan & Watts 2000). Finer-scale site investigations within those stands revealed that caribou selected foraging paths with more accessible biomass of the arboreal lichens Alectoria sarmentosa and Bryoria spp. During late winter, caribou selected higher elevation subalpine parkland habitats. Those authors used the same data as this study, but different analyses, scales of selection and habitat delineations. Pooling winter seasons, we found that Yellowhead caribou strongly selected for alpine forest, alpine and stands of subalpine fir. At the scale of multiple landscapes, mountain caribou demonstrated the highest relative probability of occurrence at elevations of approximately 1100 m and slopes of 35°. Those are steeper and lower areas than reported by Terry, McLellan & Watts (2000). The discrepancy results from our multivariate model and through sampling a wider variety of topography.
This is not the first application of RSF to issues of wildlife conservation. Among numerous examples, Apps et al. (2001) used logistic regression to model habitat selection of mountain caribou found across south-eastern British Columbia. Results were used to illuminate habitat selection patterns and generate maps of predicted caribou distribution. Both products allowed forest planners and wildlife biologists to designate important caribou habitats within regional land-use plans. Mladenoff, Sickley & Wydeven (1995) used logistic regression to estimate the amount and spatial distribution of potential habitat available for wolf Canis lupus L. recolonization across northern Minnesota, Wisconsin and upper Michigan, USA. Further observations confirmed that those models performed well at predicting the distribution of recolonizing wolf packs (Mladenoff, Sickley & Wydeven 1999). Mace et al. (1999) used RSF to model the cumulative effects of human activities on grizzly bear Ursus arctos L. habitat. RSF have a variety of other applications, including risk analysis for land-use change, habitat-based population viability analyses and population estimates (Boyce, Meyer & Irwin 1994; Boyce & McDonald 1999; McDonald & McDonald 2002).
Although the objectives were similar, the modelling approach employed here differed from past efforts. Conventional logistic regression has served as the statistical framework for the majority of previously published RSF studies. For the patch-scale analyses we used conditional fixed-effects logistic regression. The technique is widely employed in other fields of study, but only now is beginning to appear in the ecological literature (McCracken, Manly & Vander-Heyden 1998; Compton, Rhymer & McCollough 2002). CFE regression allowed us to control for variation in temporal and spatial factors between clusters defined by animal locations. The CFE approach also permits a more precise estimate of availability, eliminating issues of selection, parameterization and validity of home-range analyses and relaxes the assumption of equal accessibility of all habitats across large home range areas (Arthur et al. 1996; Alldredge, Thomas & McDonald 1998; Arthur & Schwartz 1999).
We also presented an approach that relates habitat selection and animal distribution across spatial scales. Typically, researchers define scale according to the spatial extent of available habitats (Bradshaw et al. 1995; Poole, Heard & Mowat 2000). Although multiscale studies can reveal changes in patterns of selection it is difficult to determine the most appropriate scale for management or the most ecologically relevant scale to the species of interest (Apps et al. 2001). We defined two hierarchically related scales of selection that we assumed influenced the distribution of mountain caribou across the study area. Using logic founded in probability theory and simple image arithmetic we then calculated the joint relative probability of a caribou selecting patch A and topographic feature B [P(A and B) = P(A) × P(B)]. However, the RSF models do not represent independent events and we calculated scaled relative probabilities. It is more appropriate to consider the final map as the relative probability of occurrence of caribou in vegetative patches weighted by the relative probability of occurrence across the larger study area as opposed to a true joint probability.
Variation in spatial and temporal scales can be measurement or phenomenon based. Therefore, poor definitions of scale and associated concepts can lead to confusion when interpreting the objectives and results of multiscale studies (Dungan et al. 2002). For mountain caribou, resource selection probably occurs over a range of scales, from the feeding site to the annual range of each population to the wider distribution of the ecotype. RSF coefficients describing selection may change in a non-linear fashion according to the scale of measurement or the behavioural phenomenon of interest (Johnson et al. 2002). Considering the objectives of this study and the added complexity of interpreting results from modelling efforts at several arbitrary spatial scales, we chose to constrain observations and generalize inferences to two spatial scales. When interpreting the results, it is sufficient to consider selection functions as representative of the topographic conditions limiting the distribution of caribou across multiple watersheds and the behaviours caribou would demonstrate when moving within and between patches over a time span of approximately 7–33 days. We assumed that covariates in candidate models for the patch- and landscape-scale RSF represented resources or ecological factors important to caribou at those scales.
The primary objective of this project was the prediction of caribou distribution across the larger study area. Applying individual RSF models to the known range of a sample population is ecologically and statistically appropriate. Using those models to predict the distribution of caribou beyond the range of the populations from which samples were drawn can be problematic. It was imperative to develop models using a procedure that maintained the generalities of caribou–habitat relationships and that did not overfit sample data to the final predictive model (Olden & Jackson 2000). Automated statistical algorithms such as forward or backward stepwise procedures are commonly used to test a large number of combinations of independent variables iteratively and identify the model with the best statistical fit. These techniques lead to incorrect statistical inference and exploit random variations in sample data that result in models that fit well but generalize poorly to the larger population (Hurvich & Tsai 1990; Derksen & Keselman 1992; Menard 1995). We a priori defined a small set of ecologically plausible models and used AIC to select the most appropriate model from that set. When properly used, AIC guards against overfitting and provides a measure of best inference, given the data and the set of proposed models, which is not reliant on arbitrary levels of significance (Anderson, Burnham & Thomas 2000).
To maintain the external validity of the patch-scale model we applied vegetation covariates that had uniform definitions across the study area. Such a strategy probably sacrificed some capacity to model resource selection within the range of the Yellowhead population, but allowed us to apply those models to the wider distribution of mountain caribou. Extrapolation, however, was hindered by forest inventory types that were classified too coarsely to represent the variation in selection responses we might expect from mountain caribou across the larger study area. Most notably, the ecological characteristic of the alpine class varied considerably but was represented by a single code within the forest inventory (Johnson et al. 2003). This led to poor prediction of caribou occupancy across the steep and high portions of the study area, which was rectified using covariates for topography at the broader landscape scale. We had insufficient data to assess the predictive power of the patch-scale model beyond the range of the Yellowhead population. Therefore, caution should be exercised when interpreting and applying results.
Although the RSF models had good predictive power within the range of data used for construction, we did not parameterize all factors dictating mountain caribou distribution. Inclusion of only habitat covariates resulted in simplistic models and maps that represent the potential not the current distribution of mountain caribou. Predators, disturbance from industrial or recreation activities, size and connectivity of habitat patches, and historical declines in distribution and abundance, probably resulted in exaggerated estimates of occurrence. From a conservation perspective, however, we were interested in identifying potential habitats that may support caribou following the identification, understanding and remediation of limiting factors.
By definition RSF are proportional to the probability of use of a resource unit and allow prediction of relative probabilities of occurrence (Manly et al. 2002). The metric of interest for conservation and management is the relative probability of occurrence that does not necessarily correlate with habitat quality (VanHorne 1983; Hobbs & Hanley 1990). Where source–sink dynamics are present, RSF models may predict a high probability of occurrence, but those locations may negatively affect population productivity (Mattson & Merrill 2002). Thoughtless application and poor interpretation of RSF could result in the incorrect designation of habitat importance and ineffective or perhaps harmful conservation initiatives. We assumed that mountain caribou used resources out of proportion to their availability in direct accordance with survival and reproductive benefits. As with most long-lived low productivity species, the data necessary to evaluate natality and survival are lacking, making it difficult to test explicitly that assumption or build models that relate habitat to population processes (Johnson et al. 2005). In the case of mountain caribou, land-use managers and habitat biologists should consider the spatial adjacency of limiting factors such as human access and the distribution of predators when using relative probabilities of occurrence to assess habitat quality.
These results must also be considered in a temporal context. Vegetation communities are dynamic, leading to changes in habitat availability and the strength of selection for particular resources by caribou. Unlike other populations of mountain caribou, we saw no effect of stand age on selection (Apps et al. 2001). We attribute that result to the relatively homogeneous distribution of old subalpine fir stands across the range of the Yellowhead population. Increased natural or human disturbance would lead to younger stands with fewer arboreal lichens and presumably infrequent use of those stands by caribou during winter. Stand age may become an important predictor of occurrence following widespread disturbance.
We do not have a measure of truth by which to compare the absolute accuracy of the RSF or expert opinion maps, but it is clear that both approaches have distinct advantages and limitations. When solicited informally, the opinions of experts can be collected inexpensively and the GIS analyses necessary to represent that information spatially are relatively simple. In some cases, however, expert knowledge may be unreliable, variable or unavailable. Furthermore, it can be difficult to document and present the dialogue necessary to generate consensus on criteria used to designate habitat values. Rigorous methods are available for soliciting expert opinion, but they come with costs in time and financial resources (Alder & Ziglio 1996; Dixon 1997).
The greatest advantage of an RSF approach for conservation mapping is that methods, data and results are easily documented and relatively transparent. Geographic and temporal range of animal data can be evaluated for bias, precision of coefficient estimates are presented, and numerous methods are available to assess model fit and predictive capacity (Pearce & Ferrier 2000; Manel, Williams & Ormerod 2001; Boyce et al. 2002). Although caribou locations were available for these retrospective analyses, other geographical areas or species may require the initiation of expensive mark–relocation studies, which are typically conducted over periods of 2–4 years. RSF analyses also have limiting assumptions that may be restrictive for certain study designs, species or data sets (Alldredge, Thomas & McDonald 1998). Quantitative habitat use vs. availability approaches also have been criticized as being time, place and definition specific, with few links to mechanistic processes (Hobbs & Hanley 1990; Garshelis 2000). Most problematic to applications of conservation and management is the underlying assumption that probability of occurrence is related to habitat quality.
Ultimately, researchers should strive to incorporate expert opinion within RSF. Selection of appropriate scales of analysis and definition of ecologically plausible RSF models is best guided by current understanding of the study species. Past research and knowledge is also crucial for interpreting RSF models, which are often developed using GIS data that serve as proxies for the mechanistic responses of animals to resources or disturbance factors that occur at a range of spatial scales.
application of rsf models to mountain caribou conservation
The results of this work have a number of direct and indirect applications to the management of mountain caribou populations found across central British Columbia, but are not without limitations. RSF is a flexible and powerful tool for modelling habitat use, yet resulting coefficients describe only behaviours represented by the sample of radio-telemetry locations. Biases associated with collaring caribou or collecting location data will be reflected in coefficients describing the strength of selection for or avoidance of particular habitat types. Furthermore, reliability of predictive maps is dictated by the quality of the forest inventory data used to build and extrapolate RSF models across the study area.
Given those limitations, predictive maps, when applied appropriately, may serve to identify contiguous areas across central British Columbia with a high potential of being good caribou habitat. Identification of such areas will assist with large-scale land-use planning, forestry management and recovery efforts for threatened populations. However, the results of this work are likely to be inappropriate for stand-level habitat management. The resolution of the spatial data and the suspected response of caribou to finer-scale habitat attributes require that site inspections or more refined modelling efforts guide forest harvest prescriptions. RSF coefficients and predictive maps also can serve to generate hypothesis for future research and direct population inventories across areas where relatively little is known of the distribution of caribou.
We recommend that these models and maps, and results from similar applications, serve as a starting point within an adaptive framework for conservation. As demonstrated by Mladenoff, Sickley & Wydeven (1999), observations from further research and inventory efforts can be used in an iterative fashion to assess and update RSF and resulting maps. Such approaches to interpretation, application and ultimately revision are essential when generating out-of-sample predictions across large diverse geographical areas.