Skip to main content

Site fidelity and behavioral plasticity regulate an ungulate’s response to extreme disturbance

Cite this dataset

Kreling, Samantha et al. (2022). Site fidelity and behavioral plasticity regulate an ungulate’s response to extreme disturbance [Dataset]. Dryad.


1. With rapid global change, the frequency and severity of extreme disturbance events are increasing worldwide. The ability of animal populations to survive these stochastic events depends on how individual animals respond to their altered environments, yet our understanding of the immediate and short-term responses of animals to acute disturbances remains poor.

2. We focused on animal responses to the environmental disturbance created by megafire. Specifically, we explored the effects of the 2018 Mendocino Complex Fire in northern California, USA on the behaviour and body condition of black-tailed deer (Odocoileus hemionus columbianus). We predicted that deer would be displaced by the disturbance or experience high mortality post-fire if they stayed in the burn area.

3. We used data from GPS collars on 18 individual deer to quantify patterns of home range use, movement, and habitat selection before and after the fire. We assessed changes in body condition using images from a camera trap grid. The fire burned through half of the study area, facilitating a comparison between deer in burned and unburned areas.

4. Despite a dramatic reduction in vegetation in burned areas, deer showed high site fidelity to pre-fire home ranges, returning within hours of the fire. However, mean home range size doubled after the fire and corresponded with increased daily activity in a severely resource-depleted environment. Within their home ranges, deer also selected strongly for patches of surviving vegetation and woodland habitat, as these areas provided forage and cover in an otherwise desolate landscape. Deer body condition significantly decreased after the fire, likely as a result of a reduction in forage within their home ranges, but all collared deer survived for the duration of the study.

5. Understanding the ways in which large mammals respond to disturbance like wildfire is increasingly important as the extent and severity of such events increases across the world. While many animals are adapted to disturbance regimes, species that exhibit high site fidelity or otherwise fixed behavioural strategies may struggle to cope with increased climate instability and associated extreme disturbance events.


Study area

We conducted our fieldwork at the University of California Hopland Research and Extension Center (HREC), located in southern Mendocino County (39°00′ N, 123°04’ W; Fig. 1). The 21.4 km2 study area is situated at the wildland-urban interface, a key zone of interest for fire science, and is bordered to the south by development (a town and major highway) and to the north by protected wildlands. The study area is comprised of a heterogeneous mixture of habitat types, including chaparral/shrubland, oak woodland, and grassland, and deer in the study area access resources in all of these habitats. The region has a Mediterranean climate, with mild seasons and winter rains. Topography of the study area is characterized by rugged inclines and several ravines through which water drains in the wet season. The property has a number of agricultural pastures with low fences that deer can easily cross, and the deer population is free-ranging and wild.

In 2018, California had its worst fire season in recorded history at the time of the study (surpassed in 2020), in terms of area burned, with 8,527 fires burning nearly 7,700 km2 (NIFC 2018). On July 27, 2018, the Mendocino Complex Fire broke out north of the study area (Costafreda-Aumedes et al. 2018). Between the date of ignition and September 18, 2018, the fire burned a total of 1,858 km2, becoming the second largest fire in California history. On July 27, the Mendocino Complex Fire entered HREC, burning the study area until July 28. Up to eight weeks after the initial event, trees continued to smolder, and small fires emerged. The fire burned roughly 65% of HREC (13.8 km2), with burns concentrated in the northern half of the study area across a range of different habitat types including oak (Quercus spp.) woodlands, madrone (Arbutus menziesii) forests, manzanita (Arctostaphylos spp.) shrubland, and grasslands (Fig. 1). We treated the Mendocino Complex Fire as a natural experiment, and our study design was therefore constrained by the data collection methods already in place at the time of the fire.

Monitoring deer movement

To monitor movement of black-tailed deer in the study area, we deployed GPS collars on 18 adult deer between July 2-19, 2018, including 16 female deer and 2 male deer (Supplementary Table 2). Deer were captured using Clover traps and were manually restrained without the use of chemical immobilizers. Of the 18 collared deer, 13 had home ranges within the burn perimeter, facilitating the comparison of deer movement in burned and unburned areas.

We used Vectronic VERTEX Plus Collars and Lotek IridiumTrack M collars for female deer, and ATS Iridium Lite G2110L expandable collars for male deer. Vectronic collars recorded GPS locations every hour, and ATS collars every two hours. We chose to use a longer fix rate for the ATS collars on males to maximize collar lifespan, given the difficulty of capturing male deer (they less readily enter Clover traps, which accounts for the smaller sample size of males in our study). We remotely monitored deer for multiple days after capture to ensure there were no lasting negative effects from handling without interrupting deer behavior.

To compare deer movement before and after the fire, we subset our data such that there were the same number of pre- and post-fire GPS data points for each individual in the study. This resulted in pre- and post-fire periods ranging from 15 - 25 days, depending on how many days before the fire a given individual had been collared. For all analyses of GPS collar data, we removed the first 24 hours of post-capture data to ensure paths were representative of typical behavior. We believe that this cut-off is conservative, as visual inspection revealed that deer resumed normal activity within hours of release. We also removed three erroneous GPS locations from the dataset, given that they were far from the study area with no nearby consecutive points within 2 km.

Home ranges and displacement

We used the Local Convex Hull method (LoCoH) to determine home range size (Getz et al. 2007). We calculated 95% isopleths for each individual for the pre- and post-fire periods, as well as on a monthly basis from July to December, using the T-LoCoH and adehabitatHR packages in R (Calenge 2006, Lyons and Getz 2018, Lyons 2018). We used a k-nearest-neighbors approach with k = 15 neighbors, which we determined to be an acceptable k value for all individuals based on isopleth area curves and isopleth area-edge ratio plots (Dougherty et al. 2018). When calculating isopleth area, we did not consider a temporal effect (s = 0). We used paired Welch’s unequal variance t-tests to compare deer home range size before and after the fire for female deer only, given that male deer had significantly larger home ranges than female deer, and a low sample size prevented an independent analysis of males (we instead report summary metrics for the male deer).

To determine the displacement distance of deer as a result of the fire, we identified the point during the fire and the 3-day post-fire period that was farthest from the pre-fire LoCoH home range centroid and calculated the Euclidean distance between points (Calcagno 2013). We calculated the distance between the centroids of pre- and post-fire isopleths for each deer to examine if, and how far, deer shifted their home ranges after the fire. We calculated displacement for all 18 deer collared pre-fire, including those with home ranges inside and outside of the fire perimeter.

 Movement metrics

We calculated pre-fire and post-fire movement metrics for each individual deer, including average step length, mean turn angle correlation (TAC), mean time to return (hours an animal spends before returning to a given radius), and mean residence time (number of hours spent inside a given radius), using the amt package in R (Signer 2018). The radius was set equal to mean step length, following Abrahms et al. (2017). To understand if deer movement became more directed after fire, we calculated the straightness index, a measure of path tortuosity that ranges from 0 to 1, where 1 represents perfect linearity between distance and trajectory length (Benhamou 2004). After initial exploration suggested differences in TAC, mean time to return, and mean residence time between males and females, we excluded males from the analysis (a low sample size prevented an independent analysis of male deer, and the fix rate of the collars was different for males and females).

Resource selection functions

We used resource selection functions (RSF) to examine patterns of deer habitat selection before and after the fire, for the deer with home ranges in the burn perimeter (n=13). We generated 4 random points for each GPS location for each deer within the 95% Minimum Convex Polygon corresponding to the combination of their pre- and post-fire home ranges. We used the lme4 package to run logistic regressions (GLMMs; Bates et al. 2015). Given small sample sizes of male deer, we combined male and female deer in RSF models, and explored the effect of sex as a fixed effect in model selection. We modelled pre- and post-fire time periods separately.

We used a hypothesis-driven approach to select covariates that we believed to influence deer movement, based on our understanding of the study system and on previous studies of black-tailed deer in the region (Bose et al. 2018; Supplementary Table 1). The covariates we considered in the RSFs were sex, vegetation type, elevation, slope, aspect (northness and eastness), ruggedness, distance to streambed, and surviving vegetation (post-fire model only). We confirmed that Variance Inflation Factor (VIF) < 3 for all covariates, a common cut-off for multicollinearity (O’Brien 2007). We then used an information theoretic approach to model selection, using a backwards stepwise approach from the full model and selecting the best model based on AIC (Burnham & Anderson 2002).

To create the vegetation type layer, we hand-digitized vegetation classes from high resolution (<1m) National Agriculture Imagery Program aerial imagery (2014-2015) to create a vegetation classification layer of the study area. In 2015, we ground-truthed the vegetation classification for the entire HREC study area, visiting 50 random points and validating their digital classification (accuracy was 98%). For our analyses, we simplified land cover classes into three categories: shrubland (chaparral), woodland, and grassland.

We obtained elevation and slope data from the ASTER Global Digital Elevation Model (NASA and METI 2011). We derived aspect from this DEM data, and calculated northness (cosine of the aspect layer) and eastness (sine of the aspect layer). We also calculated ruggedness, which considers variability in both slope and aspect within a neighborhood of 2,500 musing the Vector Ruggedness Measure tool for ArcGIS, which was adapted from Hobson (1972). We created a raster layer of distance from streambeds (seasonal streams, which were mostly dry during the study period) on the study site. We obtained stream vector data from the National Hydrography Dataset and calculated the distance from any given cell in the raster to the nearest stream. Finally, we created a layer of post-fire surviving vegetation using the near infrared and shortwave infrared bands from 3m-resolution satellite imagery acquired on August 8, 2018, five days post-fire by calculating the Normalized Burn Ratio (NBR; Imagery courtesy of Planet Labs, Inc.). Positive NBR values were classified as vegetated (value 1), and negative NBR values were classified as burnt (value 0; Escuin et al. 2007).

To assess the predictive ability of the models, we validated the top models using area-adjusted cross-validation, following Boyce et al. (2002). We ran 1,300 bootstrapped iterations with replacement in which we randomly subset the data, training the model on 80% of the data and withholding 20% for testing. We ran 100 iterations for each of the 13 deer, separating the deer for the area-adjusted cross-validation given differences in available habitat in each deer’s home range. For each iteration, we divided the study area (the given deer’s MCP home range) into 10 bins based on deciles of predicted risk for the test data and calculated the Spearman rank coefficient between bin rank and the mean area-adjusted frequency of deer locations from the test data, for all iterations across deer combined.

Assessing deer body condition

We used images from camera traps to assess the effects of wildfire on deer body condition. Beginning in 2016, we deployed a grid of 36 motion-activated Reconyx Hyperfire PC900 and HC600 infrared cameras (Supplementary Figure 1). We placed each camera trap at the centroid of a hexagonal grid cell, spaced 750 m apart from cameras in the six neighboring grid cells (the area of each grid cell was 0.37 km2). To facilitate comparison across camera sites, we placed cameras at the most suitable location within 50 m of the predetermined grid cell center to maximize detection probability, by facing game trails for example. Cameras were unbaited and mounted 1 meter high in steel cases on trees, or on steel posts when there were no trees nearby.

Of the 36 cameras, 25 cameras were in burned areas within the Mendocino Complex Fire perimeter, and 11 cameras were in unburned areas (Supplementary Table 3). Memory cards in six cameras in the burn area were not salvageable due to fire damage and excluded from the analysis (n=6). We additionally excluded cameras that were operational for <40 days during either the pre- or post-fire time period. Five of the cameras were non-functioning after the fire, but we recovered data from the memory cards and replaced the cameras between August 1 and August 8 with Bushnell Trophy Cams. An additional 3 cameras inside the fire perimeter did not capture any deer photographs suitable for estimating body condition index (BCI) from. One camera outside of the burned area was also excluded due to vegetation blocking the camera pre-fire. Otherwise, all cameras in the burned and unburned areas were operating continuously before, during, and after the fire. This resulted in a total of 10 cameras outside the burn area and 15 cameras inside the burn area.

Following Smiley (2017), we categorized records of adult male and female deer from each camera into a BCI. BCI ranges from 0-5 based on the visibility of five bone regions (scapula, spinal ridge, ribs, tuber ischium, and tuber ilium), and is correlated with subcutaneous fat storage (see full details in Smiley 2017). Each photo was reviewed by one person. We defined independent camera records as those that occurred at least 15 minutes after the previous record. We removed photographs from analysis if more than 60% of the deer’s body was not visible due to lighting, picture quality, or deer position. We were unable to identify individual deer at the camera traps, but we know that the study area hosts a high density of deer. We designed the camera grid such that each grid cell was larger than an individual deer’s home range, and we believe that the camera traps were set far enough to limit amount of resampling the same individuals based on home range size (typical home range size: 0.1 – 0.3 km2; camera grid cell: 0.37 km2).

We then compared body conditions of deer pre- and post-fire, inside and outside (control group) of the burn perimeter. We defined the 60-day pre-fire period as June 1 - July 27, 2019, and the 60-day post-fire period as July 28 - September 30, 2019 (120 total days). We chose these time periods to be long enough to capture a representative sample of animal activity, but not so long that seasonal influences could greatly affect our results. We then used linear regression models to evaluate BCI as a function of whether the observation was inside or outside of the burn perimeter, time period (before or after fire), days since fire (set to 0 for all pre-fire observations), and interaction terms (between burn and time period, and between burn and days since fire). We did not include time period and days since fire in the same models due to collinearity. We included camera location as a random effect to account for the possibility of resampling individuals. We compared models using AIC, and evaluated model fit using the MuMIn package in R to calculate conditional pseudo-r-squared (Barton 2018). To evaluate potential spatial autocorrelation in BCI across camera sites, we calculated Moran’s I for mean BCI for male and female deer across cameras in the pre- and post-fire periods, using the ape package in R (Paradis & Schliep 2018).


California Department of Fish and Wildlife, Award: P1680002