Dataset & Model: Fine-scale spatiotemporal predator-prey interactions in an Antarctic fur seal colony

Description

This repository contains the annotated image dataset and the pre-trained neural network weights developed for the research paper: "Fine-scale spatiotemporal predator-prey interactions in an Antarctic fur seal colony".

The data was collected at Freshwater Beach, Bird Island, South Georgia, and serves to analyze predator-prey dynamics using computer vision.

Manuscript abstract

Density critically shapes population dynamics, with high densities exacerbating intraspecific competition and disease transmission, while low densities increase predation risk. To investigate spatiotemporal density patterns and predator-prey interactions in an Antarctic fur seal (Arctocephalus gazella) colony, we deployed an autonomous camera that captured minute-by-minute high-resolution images throughout a breeding season. Using a YOLO-based neural network, we identified adult males, females and pups, and avian predator-scavenger species: giant petrels (Macronectes spp.), brown skuas (Stercorarius antarcticus) and snowy sheathbills (Chionis alba). Analysing a dataset of 4.1 million automated detections from over 10,000 high-quality images, we found spatiotemporal abundance patterns corresponding with the known foraging and breeding behaviours of these species. Strong temporal associations also emerged between the abundance of pups and two of the avian species. Fine-scale spatial analyses further revealed that pups typically remained near other pups and adult females but avoided avian predators and territorial males. Notably, the proximity of adult fur seals of both sexes reduced pup predation risk, defined as the distance between the pup and the nearest bird, whereas proximity to other pups did not. This study provides a framework for studying densitydependent interactions in wild populations and highlights the value of remote observation in ecological research.

Repository Contents

1. data.zip (Dataset)

This archive contains the training and validation data used in the study. The dataset is organized as paired files:

• .JPG: High-resolution raw images from the autonomous camera system.

• .cdb: Corresponding annotation files created using ClickPoints. These databases contain the ground-truth labels for animals.

For information on opening .cdb files, please refer to the ClickPoints Documentation.

2. model.h5 (Model Weights)

This file contains the final trained weights for the object detection network.

• Architecture: Custom YOLO-based object detection head.

• Backbone: EfficientNet B7 (pre-trained on ImageNet).

• Framework: Implemented in Python using TensorFlow and Keras.

• Purpose: To classify and locate species within the Antarctic fur seal colony at Bird Island.

3. Detection Data (.csv)

These files contain the spatiotemporal data extracted from the images. automated_detections.csv contains the output from the YOLO-based neural network, while manual_detections.csv contains human-verified ground truth annotations.

Columns:

• image_id / filename: Identifiers linking the detection to specific source images in the dataset.

• timestamp: Date and time of image

• position_x_px / position_y_px: The X and Y coordinates of the top-left corner of the bounding box in the image (pixel space).

• width_px / height_px: Dimensions of the bounding box in pixels.

• class: The annotated class

• latitude / longitude: Real-world geographic coordinates of the detected animal.

• position_x_m / position_y_m: The position of the animal in meters relative to the camera (where the camera is the origin at 0.0, 0.0). 

• laplace: The Laplace score (Variance of the Laplacian), a metric used to quantify image sharpness/quality.

Associated Code

The source code used to train this network and perform the ecological analysis described in the paper is available on GitHub: https://github.com/fabrylab/AntarcticFurSealPredatorPrey