Cape gannet movement tracking
The study took place in the Western Cape, South Africa, where we studied chick-rearing Cape gannets from Malgas Island (33.05° S, 17.93° E) during October–November from 2008 to 2015 (Fig. 1). We caught birds using a pole fitted with a loop and fitted 197 adult Cape gannets (22 in 2008, 16 in 2009, 38 in 2010, 11 in 2011, 29 in 2012, 29 in 2013, 23 in 2014, and 29 in 2015) with GPS-loggers (2008: GPS mass 65 g, i.e., 2.4 % adult body mass, Technosmart, Rom. 2009–2010: GPS mass 45 g, i.e., 1.7% adult body mass, Technosmart, Rom. From 2011: GPS mass 30 g, 1.1% of bird body mass, Catnip Technologies, Hong-Kong). Loggers were attached to the lower back with waterproof Tesa® tape and recorded position at a regular 30-s to 2-min intervals, reinterpolated over 1-min intervals. Devices were recovered after one foraging trip lasting a few hours to one week. Bird handling and tracking using these procedures do not have a measurable impact on foraging behavior19,20. We caught adult birds at-random from the colony, and previous studies showed that this resulted in a well-balanced sex-ratio preventing confounding sex effects21. All experiments were performed under permit from South African National Parks with respect to animal ethics (N° RYAP/AGR/001-2002/V1).
Cape gannet movement tactics and behavioral phases
We identified two movement trip tactics for Cape gannets: After their daytime foraging activities, some birds returned to the colony at night (rest at colony tactic) while others spent all the night at sea (rest at sea tactic). Within the GPS tracks of gannets from these two categories, we discriminated resting, foraging, and commuting phases, with a segmentation-clustering method based on smoothed speed (i.e., speed smoothed over two steps before and after the focal location) and turning angle measured at constant step length. This corresponded to the angle between the focal location, the first location entering a circle of radius equal to the median step length, and the last location inside the circle22. We fitted behavioral identification with the segclust2d package23 for the R software24. See complete details on behavioral classification for Cape gannets tracks in Appendix 1 in Courbin et al.25.
Cape fur seal movement tracking and the seascape of fear
We assessed the at-sea spatial distribution of Cape fur seals, a predator of Cape gannet fledglings7 and adults (Supplementary Data 1). We used Argos data collected from 25 lactating female seals before (2003 and 2004) and again concomitantly with gannet tracking (2012 and 2014). Seals were tracked during the same period of the year as gannets (i.e., September to November). Adult females nursing pups were selected at random and captured using a modified hoop net. Once restrained, anesthesia was induced using isoflurane gas delivered via a portable vaporizer (Stinger, Advanced Anesthesia Specialists, Gladesville, New South Wales, Australia). A satellite tag was glued to the guard hairs on the upper back. Individuals were allowed to recover from the anesthesia and resumed normal behavior within 45 min of capture. Throughout the process, the animals’ breathing was closely monitored and their flippers were repeatedly flushed with seawater to prevent hyperthermia. Seals were equipped with Argos satellite transmitters at three colonies (Fig. 1): Kleinsee (29°35’09”S, 16°59’56”E) located ~400 km to the North of the gannet colony (n = 8 seals in 2003 and 2004); Vondeling Island (33°09’11”S, 17°58’57”E), ~12 km away from the gannet colony (n = 12 seals in 2012 and 2014); and Geyser Rock (34°41’19”S, 19°24’49”E) located ~230 km to the South of the gannet colony (n = 5 seals in 2003). Seals at Vondeling Island were equipped with Argos-linked Spot-6 position transmitting tags (Wildlife Computers) following deployment procedures outlined in Kirkman et al.26. Seals at Kleinsee and Geyser Rock were equipped with ST18 and ST20 satellite-linked platform terminal transmitters (Telonics, Mesa, USA), as detailed in Skern-Mauritzen et al.27. Devices collected a well-balanced number of Argos locations during the day (n = 6080 locations) and at night (n = 6501 locations). See full details on seal tracking in Supplementary Table 6. All fieldwork was permitted by the Animal Ethics Committee of the Department of Environmental Affairs and Tourism’s Marine and Coastal Management branch, which at the time was the management authority of South Africa’s marine and coastal environment (Ref: DEAT2006-06-23).
We modeled both daytime and nighttime at-sea occurrences of seals for each colony with resource selection functions (RSF)28,29, a proxy of the fear effect for Cape gannets. RSF compared environmental features of seal’s at-sea Argos positions (i.e., further 500 m than the colony) with five times more random locations that captured the breadth of environmental conditions available to seals. We sampled random locations for each individual within the yearly area used by seals from each colony, delineated by the 95% kernel utilization distribution of the Argos locations of all seals of the colony. RSF were fitted with a generalized linear mixed model with a binomial distribution for errors. As environmental variables, we considered bathymetry (m), the slope of the bathymetry (°) and the distance to the colony (km) within the RSF. These variables were not highly correlated (|r| ≤ 0.61) and had low collinearity with a variance inflation factor VIF < 230. All continuous predictors were centered and scaled. Following statistical recommendations in Muff et al.29, we added random intercept for seal ID with a large fixed variance, a random slope for each predictor and weighted random locations by 1000. We assessed the robustness of the RSF using a leave-one-out cross validation with iteratively one of the individuals representing the testing set and the other seals representing the training set31,32. RSF with a high predictive power had a high average Spearman’s rank correlation (({bar{r}}_{s})) between the rank of the RSF scores (relative probabilities of seal occurrence) split into ten bins and the area-adjusted frequency of the Argos locations31. We ran RSF with the glmmTMB package33 for the R software24.
Finally, we spatially predicted the binned RSF scores (10 bins) for each seal colony within our study area34. We assessed the seascape of fear by overlapping the three colony-specific RSF maps and retaining the maximum value among the three maps for each pixel.
Co-occurrence of Cape gannets and fisheries
To assess the influence of fisheries on gannet movements we used vessel log-book records between 2008 and 2015. We mapped the yearly catch distribution of trawlers targeting hake (Merluccius capensis) with a 20 × 20 nautical mile resolution grid and of purse-seiners targeting anchovy (Engraulis capensis) and sardine (Sardina pilchardus) with a 10 × 10 nautical mile resolution grid. These data were made available by the branch: Fisheries Management of the Department of Forestry, Fisheries and the Environment of the Republic of South Africa.
Daytime gannet foraging habitat selection
We assessed whether Cape gannets experienced different risks of encountering seals and different levels of competition with fisheries, according to their movement tactics. For this analysis, we only kept daytime foraging locations, and estimated foraging habitat selection of gannets with a RSF for each tactic28,29. Thereby, we fitted a generalized linear mixed model with a binomial distribution for errors with random intercept for gannet ID (with a large fixed variance) and year, random slope for each predictor and weighted available locations by 1000. We determined availability using a design adapted for central place foragers in habitat selection studies, considering that individuals used areas close to the colony more frequently than elsewhere35,36. For each observed foraging trip, we simulated 10 tracks that started at the same location as the observed trip (i.e., at the colony) using a first-order vector autoregressive model37. Simulated tracks considered no habitat preference, while respecting constraints on trip structure (duration and travel speed)37. For each simulated track, we then determined available foraging locations using the segmentation-clustering method described above (see “Cape gannet movement tactics and behavioral phases” section). RSF included daytime seal encounter risk, as well as the presence of purse-seiners and trawlers, assessed through the spatial distribution of their yearly catches (log(tonnes + 1)). All continuous predictors were centered and scaled, and had low collinearity (|r| ≤ 0.62, VIF < 2)30. We assessed RSF robustness as previously described, but using a k-fold cross validation with iteratively 80% of the individuals representing the training set and the 20% remaining birds representing the testing set31,32. We ran RSF with the glmmTMB package33 for the R software24.
Cape gannet diel at-sea habitat use
We tested whether Cape gannets adjusted their at-sea behavior according to the presence of seals and fishing boats between relatively safe daytime and risky nighttime. For this purpose, we only considered gannets that did not return to the colony at night (n = 142 individuals, 72% of our sample). We estimated the probability of daytime foraging and nighttime resting, depending on both fishing activities and seal encounter risk. Thereby, we calculated daytime seal encounter risk for daytime foraging locations and nighttime seal encounter risk for nighttime resting locations. We used a generalized linear mixed model with a binomial distribution for errors and individual ID and year as random intercepts. Models also included the average monthly sea surface temperature (SST, °C) to test whether by moving away from risky areas to rest, gannets may also benefit from better thermoregulation conditions, i.e., higher SST. SST data were extracted from Aqua MODIS satellite imagery, with a 4-km-resolution grid (NASA Goddard Space Flight Center, Ocean Ecology Laboratory, Ocean Biology Processing Group; (2018): Aqua MODIS Sensor Ocean Color Data, NASA OB.DAAC.). All continuous predictors were centered and scaled. We tested candidate models with linear or nonlinear effects (natural spline with df = 4) for predictors, and selected the best model using the Akaike’s Information Criterion corrected for finite sample size. All candidate models did not include highly correlated variables (|r| < 0.4) and had low collinearity with VIF < 230. We fitted models with the lme4 package38 and model selection with the MuMIn package39 for the R software24.
Reporting summary
Further information on research design is available in the Nature Research Reporting Summary linked to this article.
Source: Ecology - nature.com
