Winter selection of landscapes by woodland caribou: behavioural response to geographical gradients in habitat attributes

Authors

  • Daniel Fortin,

    Corresponding author
    1. NSERC-Université Laval industrial research chair in silviculture and wildlife, Département de Biologie, Université Laval, Québec, Québec, Canada GIV 0A6;
      *Correspondence author. E-mail: daniel.fortin@bio.ulaval.ca
    Search for more papers by this author
  • Réhaume Courtois,

    1. Ministère des Ressources naturelles et de la Faune, Direction du développement de la faune, 880 chemin Sainte-Foy, 2eétage, Québec, Québec, Canada G1S 4X4;
    Search for more papers by this author
  • Pierre Etcheverry,

    1. Activa Environnement Inc., 84 St Germain Est, bureau 2080, Rimouski, Québec, Canada G5L 3H8;
    Search for more papers by this author
  • Claude Dussault,

    1. Ministère des Ressources naturelles et de la faune, Direction de l’aménagement de la faune, 3950 boulevard Harvey, 4e étage, Saguenay, Québec, Canada G7X 8L6; and
    Search for more papers by this author
  • André Gingras

    1. Ministère des Ressources naturelles et de la faune, Direction de l’aménagement de la faune, 818, boulevard Laure, Sept-Îles, Québec, Canada G4R 1Y8
    Search for more papers by this author

*Correspondence author. E-mail: daniel.fortin@bio.ulaval.ca

Summary

  • 1Understanding animal–habitat relationships is central to the development of strategies for wildlife management and conservation. The availability of habitat attributes often changes along latitudinal and longitudinal axes, and animals may respond to those changes by adjusting their selection. We evaluated whether landscape selection by forest-dwelling woodland caribou Rangifer tarandus caribou varied along geographical gradients in habitat attributes.
  • 2Centroids (n = 422) of track networks made by caribou in winter were recorded during aerial surveys conducted over 161 920 km2 of boreal forest in Québec, Canada. Autologistic models were estimated by comparing the characteristics of landscapes (201 km2) centred on each centroid to an equal number of randomly located landscapes, with an autocovariate controlling for the non-independence among caribou locations.
  • 3The availability of habitat attributes varied along longitudinal and latitudinal gradients, and caribou altered their landscape selection with respect to those gradients.
  • 4Information Theory provided substantial support for only one model. The model revealed that the probability of occurrence of caribou increased with the abundance of conifer forests over most of the study region, but this positive response gradually became negative towards the southern portion of the region. The association between caribou and lichens changed from being negative west of the study region to being positive in the eastern part. Availability of landscapes dominated by lichen decreased from west to east. Finally, caribou generally displayed an aversion to areas with high road density, a negative association that became positive in the southern part of the study region.
  • 5Synthesis and applications. Under current legislation in Canada, the critical habitat of woodland caribou must be defined, and then protected. Our autoregressive models can help to identify landscapes to prioritize conservation efforts. The probability of occurrence of caribou was related to different landscape characteristics across their range, which implies that the typical habitat of woodland caribou differs spatially. Such behavioural plasticity could be problematic for defining critical habitat, but we showed that spatial variation in landscape selection was organized along geographical gradients. Our study illustrates how geographical trends in habitat selection can guide management and conservation decisions.

Introduction

Habitat selection is a fundamental process that structures animal distribution. Behavioural responses to spatial heterogeneity may differ among members of a population (e.g. Fortin et al. 2003; Osko et al. 2004; Forester et al. 2007), or among populations (Rettie & Messier 2000; Ferguson & Elkie 2005). For example, certain populations of woodland caribou Rangifer tarandus caribou (Gmelin) prefer clear-cuts over burned areas, whereas other populations show the opposite preference (Rettie & Messier 2000). Such flexibility in selection may influence our evaluation of animal–habitat relationships, which in turn can have management and conservation consequences. Plasticity in the response of animals to landscape attributes has led Osko et al. (2004) to cast doubt upon the usefulness of habitat suitability indices because such models assume fixed habitat selection. Understanding the causes of variations in habitat selection, and determining how trends in selection can be organized in space and time constitute an important ecological challenge.

Variation in resource availability may offer a potential explanation for the observed plasticity in resource selection. A change in resource availability can lead to a shift in selection when consumers only require a limited amount of a resource and when the use of this resource places the animal in a trade-off situation (Mysterud & Ims 1998; Gillies et al. 2006). Mysterud & Ims (1998) referred to the change of habitat selection with availability as a functional response in resource selection. The detection of a functional response often requires the availability of limiting resources to vary above and below a threshold, with the consequence that functional responses appear especially likely to be detected over large spatial extents because resource availability would be more likely to vary broadly.

Animals respond to broad-scale gradients in abiotic and biotic characteristics of their habitat. Habitat characteristics commonly show latitudinal and longitudinal gradients, and systematic trends in life-history traits of animals or in their behavioural response to landscape features are commonly observed along those gradients. For example, intraspecific estimates of body mass often increases with latitude among mammals (Bjørnstad, Falck & Stenseth 1995; Jones et al. 2005; Herfindal et al. 2006). The strength of density dependence also varies among mammal populations along latitudinal and longitudinal axes (Hanski & Korpimaki 1995; Post 2005). Similarly, habitat selection may be organized spatially along major latitudinal and longitudinal gradients. The presence of such geographical trends in animal–habitat relationships needs to be quantified and integrated into conservation plans for wildlife species.

The forest-dwelling ecotype of woodland caribou is an excellent candidate to examine variation in habitat selection along latitudinal and longitudinal gradients. Populations of forest-dwelling caribou are declining all across North America (Bergerud 1988; Courtois et al. 2003b; Schaefer 2003), and the ecotype has been classified as threatened by the Committee on the Status of Endangered Wildlife in Canada (Thomas & Gray 2002). Woodland caribou are a high-priority management and conservation species across North America (Johnson et al. 2002). Habitat use by caribou can vary regionally (Courtois et al. 2004; Ferguson & Elkie 2005), indicative perhaps of functional responses in habitat selection. The general characteristics of woodland caribou habitat include mature coniferous forests, peatlands and open areas with terrestrial and arboreal lichens (Courtois et al. 2004; Ferguson & Elkie 2004; McLoughlin, Dunford & Boutin 2005). Areas brought to an early successional stage by logging or forest fire tend to be avoided (Schaefer & Pruitt 1991; Schaefer 2003; Courtois et al. 2004). Many of these habitat attributes follow broad spatial latitudinal or longitudinal trends. For example, forest harvesting and road densities tend to decrease in a northerly direction (Schaefer 2003). The eastern part of the boreal forest in Québec, Canada, may receive as much as twice the amount of precipitation as the western part of this forest (Boucher, De Grandpré & Gauthier 2003). The fire cycle, thus, is longer in the eastern than in the western boreal forest of Québec (Boucher et al. 2003). The response of woodland caribou to broad-scale spatial gradients in landscape attributes remains unclear.

Our objective was to identify the relationship between winter distribution of forest-dwelling caribou and characteristics of landscapes encountered over a 161 920-km2 section of the boreal forest, and to determine whether landscape selection varies along latitudinal and longitudinal axes. Our study focused on the selection of landscapes because woodland caribou have vast requirements for space (Schaefer et al. 2001), and it seems especially useful to investigate habitat–caribou relationship at a scale that is relevant to managers. Ecosystem-based management has an increasingly strong influence in current forest management practices. One related strategy consists of emulating patterns of natural forest disturbance (Bergeron et al. 1999) which, for caribou, implies management of large areas. For example, Courtois et al. (2004) offer forest management guidelines for woodland caribou that involve the protection of areas covering 100–250 km2 of, preferentially, wintering caribou habitat. The response of individuals to landscape characteristics should be evaluated over broad spatial extents for the information to be useful for the conservation of woodland caribou. Finally, we focused on winter because it is a critical season for temperate ungulates (Clutton-Brock, Major & Guinness 1985; Saether et al. 1996). For example, the rate of increase of caribou populations has been positively related to the quality of winter habitat (Wittmer, Sinclair & McLellan 2005).

Methods

study region

The 161 920 km2 study region was bounded by 48°10′N–52°15′N and 61°12′W–78°52′W (Fig. 1). This section of the boreal forest includes the southern half of the range of forest-dwelling caribou in Québec. The region is composed of conifer forests (62%), mixed wood forests (15%), early-seral-stage forests (8%; generally resulting from fire or forest harvesting), lakes (5%), and areas covered by terrestrial lichens (5%). Conifer stands are dominated by black spruce Picea mariana (Mill.), balsam fir Abies balsamea (L.) Mill., and jackpine Pinus banksiana (Lamb.). The study area is subject to forest harvesting, especially south of 51·5°N. Mean annual temperatures range between –2·5 and 0 °C, and mean annual precipitation varies from 600 to 1400 mm. Woodland caribou occur throughout the region at low densities (approximately 1·5 caribou 100 km−2), with individuals organized as small groups (Courtois et al. 2003b). Deer are absent from the region. Moose Alces alces (L.) are found at a density of 4·3–10·0 individuals 100 km−2. Little information exists on wolf Canis lupus (L.) and black bear Ursus americanus (Pallas), but they are expected to occur at low densities (Courtois et al. 2007).

Figure 1.

Distribution of land cover types and roads within the study area in the boreal forest, Québec, Canada.

surveys of caribou track networks

Areas used intensively by caribou were identified from aerial surveys conducted between 1999 and 2005, generally from late February and late March (see Supplementary Material Fig. S1). Caribou track networks were located by an experienced team (pilot, navigator-observer, two observers) travelling by fixed wing aircraft (Navajo 350) at 200 km h−1 at an altitude of 200 m, along equidistant north–south transects 2·1 km apart (see Courtois et al. 2003a for details). The day following the detection of caribou tracks, the movement paths of caribou were followed through their snow networks by helicopter (Astar 350A or Bell 206) flying at low speed (100 km h−1) and altitude (100 m), and the areas used intensively were delimited. Those areas were recorded on 1:50 000 topographic maps, and then georeferenced. Track network centroids were calculated using the ReturnCenter function of ArcView 3·2a (Environmental Systems Research Institute Inc. 2000). The resulting 422 centroid locations constitute the basis of our analyses. An evaluation of the aerial survey technique estimated that 90% of all caribou track networks should have been detected in the study region (Courtois et al. 2003a).

habitat characteristics

Habitat attributes were determined from classified 1 km resolution satellite images (SPOT-VEGETATION) taken between 1999 and 2001 (Ministère des Ressources Naturelles du Québec 2002). The original classification recognized 16 land cover types, which we collapsed into seven classes: conifer forest (dominated by open or closed conifer stands with mosses), mixed forest (dominated by both conifer and deciduous stands), early-seral-stage forest (dominated by recent fires or cutblocks, covered with regeneration < 2 m high), lake, lichen (dominated by lichens with 10–40% of conifers or by shrubs and lichens), peatland, and other (dominated by deciduous forest, bare ground, moss and shrubs, agricultural lands, or unclassified areas). The proportional area of class ‘other’ was < 3% and was not considered in the analyses. Information on the road network was obtained from digital topographic maps having a spatial resolution of 1:20 000.

statistical analyses

To investigate landscape selection, we first defined landscapes as aggregations of spatial units (1 km2 pixels) with similar characteristics. Variograms were built for each of six land cover classes (i.e. all but the class ‘other’) to characterize their spatial patterns of variation based on 10 000 pixels located across the study region. The lag distance where the sill was reached for all cover classes (i.e. the overall range) indicated the spatial extent beyond which spatial dependency vanished. Circular landscapes were created based on a radius corresponding to that lag distance.

Landscape selection was evaluated by fitting Resource Selection Functions (RSFs) based on a use/availability design (Manly et al. 2002; Johnson et al. 2006). Most RSFs included an autocorrelation term, Autocov (sensu Augustin, Mugglestone & Buckland 1996). Autocov was the weighted average of the number of cells, among the 168 adjacent 1 km2 cells (i.e. a grid of 13 × 13 cells centred on the focal cell), in which a centroid of track networks was present. The weight given to a particular 1 km2 cell was the inverse of its Euclidean distance (1/h) to the focal cell (details follow Augustin et al. 1996). Because the survey technique led to the total coverage of the study region (Courtois et al. 2003a), Autocov was directly calculated from equation 1 of Augustin et al. (1996). We investigated the influence of the number of cells (8, 24, 48, 80, 120, 168, and 224) on AICc, and found that the decrease in AICc levelled off at 168 cells. Condition indices used to assess multi-colinearity increased slightly beyond 168 cells. Thus, we calculated Autocov based on 168 adjacent cells.

Characteristics of landscapes centred on the centroids of track networks (n = 422) were compared by logistic regression to the characteristics of an equal number of landscapes (n = 422) randomly located throughout the study region. The areas (km2) encompassed by each of the six cover types and road density (km km−2) characterized used and random landscapes. We tested for the presence of latitudinal and longitudinal gradients for each of the six land cover types and road density using data from the randomly generated landscapes defining resource availability. We then built a Pearson's correlation matrix that related the distance along the longitudinal (X, km) or the latitudinal axis (Y, km) to each of the seven landscape covariates. Candidate RSFs were built by considering the interaction between X or Y and habitat attributes that had a correlation coefficient r ≤ –0·25 or r ≥ + 0·25. Changes in habitat selection with availability might occur without being linked to geographical gradients, in which case non-linear responses to changes in the availability of habitat attributes could be detected by adding squared terms for certain covariates in RSFs (Boyce et al. 2003). We built candidate RSFs, some of which accounted for geographical patterns in habitat selection, while others considered quadratic responses to landscape attributes.

All models were screened for colinearity. Such problems were considered negligible when condition indices were less than 10 (Belsley, Kuh & Welsch 1980). Otherwise, the correlated variables were standardized by subtracting their means (Neter, Wasserman & Kutner 1985), and colinearity was re-assessed. The inclusion of peatlands in RSFs led to colinearity even after standardization; thus, this covariate was excluded from RSFs. For the same reason, the interaction between X or Y and mixed forest, as well as the quadratic term for mixed forest were excluded from candidate models. The remaining covariates, with interaction and quadratic terms were maintained for inclusion in candidate RSFs. The most parsimonious of the candidate model was selected based on the differences in Akaike's Information Criterion (AIC) corrected for small sample size (ΔAICc) and AICc weights (w). Robustness of the best RSF was evaluated by k-fold cross-validation (Boyce et al. 2002). The procedure was done five times. Each time, an RSF was built using a random sample of 80% of used landscapes. RSF scores were then calculated for the remaining 20%, with RSF scores = exp[β0 + β1x1 + β2x2 ... ], where β0 is the intercept of the RSF model and β1 and β2 are the coefficients of independent variables x1 and x2, respectively. Model goodness of fit was assessed by calculating Spearman-rank correlation (rs) to measure how successful the model built on 80% of used landscapes was at describing the other 20% of the data (see Boyce et al. 2002). High inline image (n = 5) indicate a strongly predictive model. RSF scores should reflect the relative probability of occurrence of caribou in landscapes (Johnson et al. 2006), and we used the best-validated model (recalculated after cross-validation using 100% of the data) to calculate relative probabilities of occurrence over the study region.

Results

Spatial autocorrelation was observed for most land cover classes, in particular, conifer forests (Fig. 2). The range of spatial autocorrelation was ≤ 8 km for all six cover classes. Landscapes were thus defined by 8 km radii, corresponding to 201 km2 areas. Geographical gradients in landscape attributes were detected across the study region (Table 1). The availability of conifer forest increased from west to east, whereas the availability of mixed forest and areas dominated by lichens and by peatlands decreased. Conifer forests and areas dominated by lichens increased with latitude, while availability of mixed forests and road density decreased.

Figure 2.

Variograms estimated for six landscape cover types (1 km resolution) in a 161 920 km2 section of the boreal forest in Québec, Canada, based on classified spot images taken in 1999–2001.

Table 1.  Mean availability of six land cover classes (km2) and road density (km km−2) for 422 landscapes (201 km2) randomly located across a 161 920 km2 portion of the boreal forest, together with the Pearson correlation coefficient relating landscape attributes to the distance along latitudinal (Y) and longitudinal (X) axes. The mean percentage cover is also given for all land cover classes. A positive correlation corresponds to an increase in the abundance of a given landscape attribute in northerly or easterly directions
Landscape attributeMean ± SDRangePer cent coverCorrelation with X (P)Correlation with Y (P)
Lake9·6 ± 16·20–1254·80·07 (0·13)0·15 (0·002)
Conifer forest125·2 ± 55·50–20162·30·55 (< 0·0001)0·40 (< 0·0001)
Early-seral-stage forest15·4 ± 30·20–2017·7–0·05 (0·28)–0·14 (0·003)
Mixed forest31·0 ± 49·30–18415·4–0·20 (< 0·0001)–0·65 (< 0·0001)
Lichen9·3 ± 18·90–1314·6–0·27 (< 0·0001)0·34 (< 0·0001)
Peatland4·6 ± 20·20–1702·3–0·46 (< 0·0001)0·13 (0·005)
Road Density0·4 ± 0·60–3·2 0·05 (0·26)–0·66 (< 0·0001)

Caribou responded to the spatial gradients by altering their landscape selection along the longitudinal and latitudinal axes (Table 2). Among the candidate models, the top-ranking RSF (Model 1) was an autologistic models that accounted for geographical changes in the selection of conifer forests, areas dominated by lichens, and road density. The worst model (Model 15) was the one excluding the autocovariate. According to AICc weights, Model 1 was at least 155 times (i.e. 0·779/0·005) more likely to adequately represent landscape selection by woodland caribou than any of the models including quadratic terms (Models 4–7). This result points out that spatial variation in landscape selection was not simply a response to differential habitat availability, but was at least partially linked to geographical gradients. However, the relatively low weight of alternate Models 2–3 (Table 2) indicated that landscape selection was not necessarily associated with the strongest latitudinal and longitudinal gradients (Table 1). Finally, a comparison of AICc weights between Model 1 and Models 8–14 showed the relative importance of each habitat attribute and interaction term on landscape selection.

Table 2.  Candidate models of landscape selection for woodland caribou in winter in the boreal forest of Québec, Canada. A number of estimated parameters including the intercept (K), Akaike's Information Criterion (AICc), the difference in AIC (ΔAICc), and AICc weight (w) are provided. Comp corresponds to the variables: lake, conifer forest, lichen, early-seral-stage forest, mixed forest, transformed road density (road density0·5), transformed distance along the longitudinal (X0·5) and latitudinal (Y0·5) axes, and Autocov. Models were estimated with individual observations of lichen, mixed forest, road density0·5, X0·5, and Y0·5 standardized by subtracting their mean value to avoid multi-colinearity
No.ModelKAICcΔAICcw
  • Includes (road density), instead of (road density)0·5, to end up with a model testing for quadratic effects by considering (road density) + (road density)2.

1Comp; conifer forest × Y; lichen × X; road density × Y13833·20·00·779
11(Comp, without road density); conifer forest × Y; lichen × X;11837·13·90·112
14Comp; conifer forest × Y; lichen × X;12838·65·40·053
9(Comp, without lichen); conifer forest × Y; road density × Y11840·87·60·017
13Comp; conifer forest × Y; road density × Y12841·38·10·014
8(Comp, without conifer forest); lichen × X; road density × Y11842·69·30·007
3Comp; conifer forest × Y; lichen × Y; road density × Y13843·19·90·006
5Comp; (lichen)2; conifer forest × Y; road density × Y13843·310·10·005
2Comp; conifer forest × X; lichen × X; road density × Y13843·610·40·004
12Comp; lichen × X; road density × Y12844·511·30·003
4Comp; (conifer forest)2; lichen × X; road density × Y13846·513·30·001
6Comp; (road density)2; conifer forest × Y; lichen × X;13857·123·90·000
10(Comp, without mixed forest); conifer forest × Y; lichen × X; road density × Y12867·934·70·000
7(Comp, without X and Y); (conifer forest)2; (lichen)2; (road density)211892·559·20·000
15(Comp, without Autocov); conifer forest × Y; lichen × X; road density × Y12948·9115·70·000

The top ranking model (Model 1) was a robust autologistic RSF: k-fold cross-validation showed a strong association between predicted and observed classes of RSF scores (inline image = 0·95, P < 0·0001). This model indicated that caribou selected landscapes with high coverage of lakes and avoided those with abundant mixed and early-seral-stage forests (Table 3). The probability of occurrence of caribou increased with the abundance of conifer forests over most of the study area, but this positive response gradually became negative in a southerly direction (Fig. 3). Caribou generally displayed an aversion for landscapes with high road density (Table 3). This negative association became positive, however, in the very southern part of the study area (Fig. 3). Finally, caribou were positively associated with lichens east of the study area, but this association gradually became slightly negative in westerly direction (Fig. 3).

Table 3.  Most parsimonious model of landscape selection by woodland caribou in winter. The autologistic model corresponds to Model 1 in Table 2, and was estimated based on centroids of track networks located from aerial surveys conducted between 1999 and 2005 over a 161 920 km2 section of the boreal forest, Québec, Canada
CovariateβSEWald statisticsP
Intercept–1·1790·19835·45< 0·0001
Lake0·0130·0064·370·04
Conifer forest0·0030·0040·370·54
Lichen0·0070·0080·930·33
Early-seral-stage forest–0·0180·0069·500·002
Mixed forest–0·0380·00730·44< 0·0001
Road density–0·2040·3930·270·60
X–0·0540·0303·190·07
Y–0·1760·0983·260·07
Conifer forest × Y0·0060·00213·210·0003
Lichen × X0·0060·0029·090·003
Road density × Y–0·480·1787·360·007
Autocov72·489·15862·78< 0·0001
Figure 3.

The relative probability of occurrence for woodland caribou in winter as a function of road density and the area of 201 km2 landscapes that was comprised of conifer forests and lichens. Relative probabilities were estimated based on Model 1 (Table 3) for each of three habitat attributes, while holding the other variables constant at their mean. The relative probabilities are provided for three locations on the longitudinal (East = 68°30′W, West = 77°45′W, Centre = 73°45′W) or latitudinal (South = 49°35′N, Centre = 50°35′N, North = 51°35′N) axes to illustrate the gradual changed in landscape selection by caribou along latitudinal and longitudinal gradients.

Overall, landscape selection led to important spatial variations in the relative probability of occurrence of caribou across Québec's boreal forest (Fig. 4). Landscapes of low probability of occurrence were strongly associated with mixed forests and roads. The negative response became particularly apparent for roads because the negative effect of these linear features was mapped as linear segments of low occurrence probabilities (Fig. 4).

Figure 4.

The relative probability of occurrence of woodland caribou throughout a 161 920 km2 section of the boreal forest in Québec, Canada, based on a resource selection function (Table 3) built from caribou track networks observed during aerial surveys conducted in 1999–2005.

Discussion

The influence of habitat attributes on landscape selection and on the winter distribution of forest-dwelling woodland caribou changed systematically along longitudinal and latitudinal axes within the boreal forest of Québec, Canada. Our study thus supports the hypothesis that landscape selection can vary over space, in response to change in the availability of specific habitat elements, or possibly, exogenous geographical processes such as climate. Changes in habitat selection with variations in availability can be detected by including quadratic terms in RSFs (Boyce et al. 2003). We found, however, that models with such terms did not describe with as much parsimony the observed distribution of woodland caribou compared to models with geographical interaction terms. Spatial patterns in landscape selection thus reflected more than non-linear responses to local variations in landscape composition by being linked to broad-scale spatial gradients. In other words, we not only found that landscape selection varied with resource availability (e.g. increased in the strength of selection for lichens, Fig. 2, as their availability decreased, Table 1), but we also observed that the response varied systematically over broad gradients. It is worth noting that, although selection patterns were tied to geographical gradients, caribou did not necessarily respond to the strongest gradients. For example, the spatial gradient in conifer forests was the strongest along the longitudinal axis (Table 1), but spatial variation in landscape selection was best explained by accounting for the latitudinal than the longitudinal gradient (Table 2). Overall, our study demonstrates that landscape selection is somewhat plastic for this ecotype, with part of its spatial variability being structured along major geographical gradients in vegetation cover and disturbance history.

Our evaluation of landscape selection also confirms many caribou–habitat associations previously reported. We found that landscape selection was positively associated with the availability of water bodies. Caribou select frozen wetlands and areas with more lakes (Ferguson & Elkie 2004, 2005), presumably because wind prevents the accumulation of snow on these large ice-covered open areas and decreases predation risk by improving the ability to detect or escape from predators (Mysterud & Ostbye 1999). The spatial distribution of woodland caribou largely reflects a behavioural response to risk. For example, caribou avoided early-seral-stage forests. Mortality risk has been positively linked to the amount of early-seral stands within ranges of caribou populations (Wittmer et al. 2007). Early-seral-stage forests included both recently burned and harvested forests. Fires negatively affect caribou habitat (Joly et al. 2003) by reducing lichen availability (Schaefer & Pruitt 1991; Terry, McLellan & Watts 2000; Dunford et al. 2006). Forest harvesting could be detrimental to caribou populations by disturbing individuals, reducing food biomass and increasing predation risk (Bergerud 1988; Cumming & Beange 1993; Vors et al. 2007). Many studies have also reported an avoidance of roads by caribou (e.g. Dyer et al. 2002; Apps & McLellan 2006), as we observed over most of the study region. Roads can facilitate travel by predators (James & Stuart-Smith 2000) and hunters (Bergerud 1974). Caribou also avoided mixed forests. The use of mixed forests can be detrimental to caribou in part because they are suitable for moose, and high moose abundance often translates into high risk of predation by wolves (Seip 1992; McLoughlin et al. 2005; Wittmer et al. 2007).

Mature conifer forests represent key cover types for woodland caribou (Terry et al. 2000; Joly et al. 2003; McLoughlin et al. 2005). Caribou were generally positively associated with landscapes largely comprised of conifer forests. These forests are suitable winter habitat because they offer reduced predation risk for caribou (Seip 1992; O’Brien et al. 2006; Wittmer et al. 2007) and other ungulates (Fortin et al. 2005). Moreover, mature conifer forests support arboreal lichen (e.g. Schaefer & Pruitt 1991; Schaefer 1996; O’Brien et al. 2006), an important food source in winter, especially in areas of deep snow (Johnson, Parker & Heard 2001). Terrestrial lichens also are central to the winter diet of caribou (Johnson et al. 2001; Courtois et al. 2004) and, in the eastern part of the study region, we detected a strong positive selection for landscapes with abundant areas dominated by terrestrial lichens.

management implications

In Canada, forest-dwelling caribou are protected under ‘Schedule 1’ of the Canadian Species at Risk Act (SARA; Species at Risk Public Registry 2007). Conservation efforts rely on the identification and preservation of critical habitat. Habitat protection regulations recognize the importance of protecting wintering areas, but do not generally specify the characteristics of those areas (Courtois et al. 2004). Our autoregressive RSF can help to identify landscapes that should be prioritized for conservation. For example, Courtois et al. (2004) proposed a forest management strategy based on the concentration of forest harvesting in large management areas, on the protection of large forested blocks, and on the maintenance of connectivity among protected blocks. Our model can be used to delineate both forested blocks and corridors that should be protected. We found that it would become particularly important to prioritize the protection of landscapes largely comprised of conifers when north of 50°N. Also, the preservation of landscapes with important coverage of terrestrial lichens would increase in value in an easterly direction, as the availability of large areas of lichens decreases. We caution, however, that the information provided by our landscape selection model should be considered at the proper scale (i.e. 201 km2 landscapes). For example, we found that the negative relationship between the probability of occurrence of caribou and road density became positive in the southern portion of the study areas. J. L. Frair et al. (unpublished data) studied an isolated population of woodland caribou located south of the study area, where road density is relatively high. They found that road density had no effect on the probability of occurrence of woodland caribou in late winter, although it had a positive effect in the spring. But they also found that, even when caribou did not avoid areas of high road density, individuals selected areas far from roads once in those landscapes (i.e. at a finer scale). Therefore, the positive association with road density that we detected in the south should not be interpreted as an attraction of caribou for roads. Instead, the observation simply indicates that, in the southern portion of the ecotype's range, the presence of caribou is not inconsistent with the presence of roads in the landscape. Future studies investigating population trends in southern populations are necessary to determine the quality of these habitats.

Our landscape selection model provides information about paths most likely to be used by caribou and, consequently, locations for potential corridors connecting discrete landscape units in managed forests. For example, O’Brien et al. (2006) defined connectivity within a network of habitat patches by evaluating the most likely paths for woodland caribou based on RSFs. Our study points out that spatial differences in corridor attributes could be expected in distinct parts of the boreal forest.

Although RSFs have been used as a projection tool (e.g. Carroll et al. 2003), their application should generally be restricted to the area and the period for which they have been developed (Boyce et al. 2002). The large size of our study region offered many advantages. First, the final RSF directly reflects landscape selection for approximately half the range of forest-dwelling caribou in Québec. Secondly, because of the large spatial extent considered here, our analyses characterized the response of caribou over a broad range of habitat attributes. In fact, certain spatial trends would not have been detectable by considering a smaller spatial extent. Our best RSF thus provides a global picture of caribou–habitat association over a large section of the ecotype's range.

Finally, our finding of a systematic response of forest-dwelling woodland caribou to geographical gradients can help in the management of broad areas. Osko et al. (2004) correctly pointed out that wildlife–habitat relationship models may lack repeatability and therefore transferability, including those developed for woodland caribou (cf. Rettie & Messier 2000). High behavioural plasticity can be problematic for the elaboration of habitat management guidelines because it could require specific models for all sites within a management region. Our study shows that there might be an alternative. Part of the variation among sites or populations can be tied to geographical gradients in landscape attributes. On this basis, a restricted number of strategically selected sites could be surveyed along geographical gradients, and serve as basis for interpolation within the management region.

Acknowledgements

This work was supported by the Natural Sciences and Engineering Research Council of Canada and the Ministère des resources naturelles et de la faune du Québec, and it is a contribution of the ‘Chaire de recherche industrielle CRSNG–Université Laval en sylviculture et faune.’ We thank Gaétan Daigle for his statistical advice, Nicolas Courbin for GIS assistance, and Cheryl Johnson, Steve Cumming, Mark Hebblewhite and two anonymous referees for their comments on the paper. We are grateful to Sophie Brugerolle for help in assembling the database. Special thanks to Danielle St-Pierre in Chibougamau for providing data, and the many people who conducted the aerial surveys.

Ancillary