Linking landscape characteristics to local grizzly bear abundance using multiple detection methods in a hierarchical model

1. What they did

Most habitat management for grizzly bears has been based on where radio-collared bears choose to go (habitat selection). Few studies have linked habitat to how many bears actually live in an area, and those did not account for bears being present but not detected. The authors wanted to find out which landscape features best explain how many grizzly bears live in different parts of a large area.

In the summer of 2000, they collected bear hair in and around Glacier National Park in northwestern Montana, using two methods:

• Hair traps. A single strand of barbed wire around several trees with scent lure poured on forest debris in the middle. They aimed for one trap in each 8 x 8 km grid cell in each of five 14-day sessions, moving traps between sessions (633 trap sites in all).
• Rub trees. Trees that bears naturally rub on, fitted with short pieces of barbed wire to catch hair. They surveyed 907 rub trees along 1185 km of trails, mostly in the park, visiting each about five times.

Genetic analysis of the hair gave the species, sex and identity of each bear. The authors then built a model that estimated the number of bears in each grid cell while separately estimating how likely each method was to detect a bear there. They tested 11 landscape features, such as road density, building density, moist (mesic) habitat, avalanche chutes, greenness, precipitation, terrain ruggedness, how strongly bears were protected on each type of land, and past human-caused bear deaths. They also tested whether bear numbers in neighbouring cells were linked (spatial autocorrelation). Males and females were analysed separately.

2. What they found

• The hair identified 222 grizzly bears: 108 males and 114 females. 71% of grizzly bear samples (809) could be matched to an individual.
• Females: road density and per cent mesic habitat best explained how many females lived in a cell. There were more females where road density was lower and where there was more mesic habitat. Spatial autocorrelation was not supported (included in only 34.6% of models). Detection of females rose with rub tree sampling effort.
• Males: road density best explained how many males lived in a cell, with more males where road density was lower. Spatial autocorrelation was supported (63% of models). Detection of males rose with both rub tree and hair trap effort, and fell over the course of the summer. When the authors used narrower priors, precipitation became important and spatial autocorrelation dropped out, so they kept the results from the less informative priors.
• Predicted numbers of bears per cell matched the raw counts of bears detected fairly well (r = 0.72 for females and 0.78 for males).
• Male and female predicted numbers were similar even though there are about 40 males for every 60 females, because males have larger home ranges and are counted in more cells. The average seasonal home range of adult females was 231 km² (a personal communication), larger than one grid cell.
• West-central Glacier National Park held a cluster of males beyond what the landscape features predicted.
• The number of hair samples in a hair trap was a poor guide to the number of bears caught there (r = 0.459).
• The data were sparse (64% of rub tree cell sessions had no rub tree sampling), so the results show only relative numbers of bears, not absolute counts.

3. What they concluded

• Road density was the most important landscape feature for both sexes, consistent with earlier studies that found grizzly bears avoid areas with many roads.
• The study supports the assumption that closing and removing roads may increase the number of bears where mesic habitat and areas of low road density are nearby.
• Landscape features worked the same way for abundance as in earlier habitat selection studies, suggesting habitat selection analyses give a good index of grizzly bear habitat quality in this ecosystem.
• Their approach offers a new way to combine several detection methods in one abundance model and to test whether spatial autocorrelation belongs in the model. It is most useful where detection varies with sampling effort, time or habitat, and where different methods detect different parts of the population.
• They recommend sampling as evenly as possible across sessions and space.

4. How tracking was or was not used

No tracks or trailing were used. The study relied on one kind of natural sign: hair left on trees that grizzly bears rub on. Field crews surveyed rub trees along trails and collected hair from them, alongside hair from baited hair traps. All identifications of species, sex and individual came from genetic analysis, not from reading the signs. The paper does not describe how rub trees were found or recognised; it refers to Kendall et al. (2008).

5. Tracker or observer certification

Not reported. The paper thanks more than 450 biologists, field and laboratory technicians, data specialists and many volunteers, but does not say who found and surveyed the rub trees or what their experience was. Because every identification was made genetically, the findings do not depend on observers’ sign identification skills.

6. Other technologies, skills or innovations

• Non-invasive genetic sampling: species from the G10J microsatellite, sex from the SRY and amelogenin markers, and individual identity from six microsatellite loci, with up to 10 more loci used to confirm close matches and careful error checking.
• An extended N-mixture model that handles two detection methods with their own detection sub-models, in one hierarchical Bayesian model.
• Bayesian variable selection with indicator weights, and a conditional autoregressive term to test spatial autocorrelation.
• Landscape layers from remote sensing, including the LANDFIRE vegetation cover map (for mesic habitat), DAYMET precipitation, and a new avalanche chute index from remote imagery.

7. Implications for conservation (from the paper)

• Identifying the landscape features that best explain local bear numbers can help prioritise conservation actions for threatened populations, and those features can be monitored as an early warning of decline. Habitat monitoring is one requirement for removing this population from the threatened list.
• Closing and removing roads, as some land managers already do, may increase bear numbers where mesic habitat and low road density are nearby.
• Managers cannot easily add mesic habitat, but could manage for bear foods. Because mesic habitat may change with climate change and fire, monitoring it may give early insight into changes in bear numbers.
• Better records of mortality, measures of people’s attitudes toward bears and better detection information could explain more of the variation, especially where fewer bears were found than expected.