Data from 'High population density reduces predation risk in Antarctic fur seal breeding colonies'

This record contains the annotated images, the trained neural network and the animal detections of the study “High population density reduces predation risk in Antarctic fur seal breeding colonies”. The data were collected at Bird Island, South Georgia (54°00'24.8ʺ S, 38°03'04.1ʺ W) during the austral summer of 2020–2021 at two Antarctic fur seal breeding colonies that differ four-fold in seal density and lie approximately 200 metres apart: Freshwater Beach (FWB): an open beach intersected by a shallow stream that forms natural corridors and influences the spatial distribution of seals within the colony. Special Study Beach (SSB): a small cobblestone breeding beach separated from adjacent breeding sites by a cliff on the east side, open sea on the west and rocky ridges to the north and south. A scaffold walkway provides safe access to the animals while minimising disturbance. FWB was analysed in a previous study, which describes the training, evaluation and analysis of the FWB detection model: Berthelsen et al. (2026), Royal Society Open Science 13(5): 250931, https://doi.org/10.1098/rsos.250931. The FWB dataset is available at https://zenodo.org/records/18148442 and the associated code at https://github.com/fabrylab/AntarcticFurSealPredatorPrey. Abstract Population density can influence individual predation risk in colonial breeders through shared vigilance and predator deterrence. We investigated how predator-prey interactions are shaped by population density at two Antarctic fur seal (Arctocephalus gazella) breeding colonies at Bird Island, South Georgia, which differ four-fold in seal density. By deploying autonomous time-lapse cameras, we captured high-resolution images at one-minute intervals throughout the breeding season. Using a YOLOv8 neural network, we identified fur seal adult males, females and pups, as well as three bird species representing a predator-scavenger gradient: giant petrels (Macronectes spp.), brown skuas (Stercorarius antarcticus) and snowy sheathbills (Chionis alba). Abundance patterns corresponded to the known foraging and breeding behaviours of these species. Differences in seal density between the colonies were mainly driven by adult females and their pups, not by adult males. Ratios of predatory birds to pups were markedly lower at the high-density colony, while scavenger to pup ratios remained similar. Spatial analyses revealed that predators were largely excluded from areas of high seal density, whereas scavengers overlapped extensively with pups in both colonies. This study demonstrates the value of remote observation in resolving predator-prey interactions and illustrates how density can shape predation risk in a colonial breeder. Repository contents 1. data.zip (annotated training images)111 SSB images (about two per day across the breeding season), each as a pair of files: .jpg: full-resolution image from the time-lapse camera, rotated to upright orientation. .cdb: ClickPoints annotation database with a bounding box for every animal in the image (for opening .cdb files, see the ClickPoints documentation). Annotated classes: Antarctic fur seal adult male, adult female and pup; southern elephant seal; leopard seal; king penguin; giant petrel; brown skua; snowy sheathbill. Leopard seals were too rare and were not used for training.These images were used to train and validate the SSB detection model: they were cut into non-overlapping 512 × 512 pixel crops and randomly split into a training (80%) and a validation set (20%). 2. model.pt (trained model weights)Detection model used for all SSB images. Architecture: Ultralytics YOLOv8m (medium; 25.9 million parameters), 8 classes. Training: training crops from data.zip, with Mosaic augmentation, random translations, scaling and horizontal flips. Because birds were rare in these images, 3,000 additional training crops were generated by copy-paste augmentation with birds annotated in 162 further SSB images. The model was trained for up to 200 epochs with early stopping and reached a mean Average Precision (mAP50) of 0.81 on the validation crops. Inference: full-size images were predicted with sliced inference (SAHI; 512 × 512 pixel slices, overlap 0.2) at a confidence threshold of 0.65. The raw SSB camera images are upside down and were rotated by 180° before prediction. Detection performance (F1 scores, confusion matrix) is reported in the Supplementary Information. The annotations of the bird test set are included in the GitHub repository. 3. FWB.csv and SSB.csv (detected animals)Detections at both colonies. They contain only images taken between 09:00 and 17:00 that passed the image-quality filter. FWB detections were obtained with the hierarchical model of Berthelsen et al. (2026), SSB detections with model.pt.Columns: filename: name of the source image. timestamp: date and time the image was taken. class: detected class male: Antarctic fur seal, adult male (Arctocephalus gazella) female: Antarctic fur seal, adult female (Arctocephalus gazella) pup: Antarctic fur seal, pup (Arctocephalus gazella) giant petrel: giant petrel (Macronectes spp.) skua: brown skua (Stercorarius antarcticus) sheathbill: snowy sheathbill (Chionis alba) SSB.csv additionally contains elephant seal (southern elephant seal, Mirounga leonina) and penguin (king penguin, Aptenodytes patagonicus), which were not analysed. x_px / y_px: x and y coordinates of the top-left corner of the bounding box in the image (pixels). width_px / height_px: width and height of the bounding box (pixels). lat / lon: geographic coordinates (latitude, longitude) of the animal. x_m / y_m: position of the animal in metres relative to the camera. Associated codeThe code for the analysis, the training of the detection model and the model evaluation is available on GitHub: https://github.com/fabrylab/AntarcticFurSealDensity