Dataset for "Leveraging machine learning and accelerometry to classify animal behaviours with uncertainty"

AWD Biologging is a dataset comprising 6,311 instances of timestamped tri-axial raw acceleration signals (16 Hz) and associated behavior annotations from five free-ranging African wild dogs (Lycaon pictus) in the Okavango Delta, Botswana (19°31′S, 23°37′E). The dataset spans 72.21 hours of matched data collected between 3 September 2021 and 16 August 2022. This dataset was introduced in the paper "Leveraging machine learning and accelerometry to classify animal behaviours with uncertainty." [1]

Behavioral annotations were assigned to 72.21 hours of acceleration data across five classes: feeding, resting, moving, running, and vigilant. Labels were derived by aligning acceleration data with timestamped video observations collected from vehicles. Audio-assisted labeling was additionally applied to expand feeding, moving, running, and vigilance annotations, as these behaviors produced distinct and reliably identifiable audio profiles. A summary of behavior durations (hours), disaggregated by behavior class and labeling source, is provided below.

This repository consists of two files:

1. matched_acceleration_data_out.csv: This file includes the timestamped tri-axial acceleration data, behavior labels, and source of labels (Audio or Video).
   Columns: behavior, behavior_start, behavior_end, duration, acc_x, acc_y, acc_z, Source
2. matched_acceleration_metadata_out.csv: This file contains metadata features about the acceleration data, including individual and time information from when the data was collected.
   Columns: individual ID, year, am/pm, half day [yyyy-mm-dd_am/pm] 

A PyTorch interface for using the AWD Biologging dataset for behavior prediction with uncertainty quantification is available on GitHub. 

Note: All capture and collaring were approved by the University of Washington Animal Care and Use Committee (Protocol #4514-01) and the Botswana Department of Wildlife and National Parks.

[1] Leveraging machine learning and accelerometry to classify animal behaviours with uncertainty. Medha Agarwal, Kasim Rafiq, Ronak Mehta, Briana Abrahms, and Zaid Harchaoui. To appear in Methods of Ecology and Evolution (MEE), 2025. BioRxiv Report.

@article{agarwal2025leveraging,
    title={Leveraging machine learning and accelerometry to classify animal behaviours with uncertainty},
    author = {Medha Agarwal and Kasim Rafiq and Ronak Mehta and Briana Abrahms and Zaid Harchaoui},
    year={2025},
    journal={Methods of Ecology and Evolution}
}

[2] Increasing ambient temperatures trigger shifts in activity patterns and temporal partitioning in a large carnivore guild. Kasim Rafiq, Neil R. Jordan, Krystyna Golabek, John W McNutt, Alan Wilson, and Briana Abrahms. In Proceedings of the Royal Society B, 2010. Journal Paper. 

@article{rafiq2023increasing,
  title={Increasing ambient temperatures trigger shifts in activity patterns and temporal partitioning in a large carnivore guild},
  author={Kasim Kafiq and Neil R. Jordan and Krystyna Golabek and John W. McNutt and Alan Wilson and Briana Abrahms},
  journal={Proceedings of the Royal Society B},
  volume={290},
  number={2010},
  pages={20231938},
  year={2023},
  publisher={The Royal Society}
}