Caribou Detection Dataset for Weakly Supervised Learning in Arctic Aerial Surveys

Supplementary file 1 accompanies a research paper proposing HerdNet, a weakly supervised framework for wildlife detection and counting. The dataset includes five caribou herds distributed across Alaska, captured in high-resolution aerial imagery from 2017 and 2019. The work by Ghazaleh Serati was published on figshare in March 2026.

Use Cases

• Training weakly supervised object detection models based on patch-level empty vs. non-empty labels.
• Benchmarking detection algorithms for small or occluded targets in heterogeneous backgrounds.
• Developing methods to address class imbalance in aerial wildlife surveys.
• Facilitating manual counting of animals by pre-mapping regions containing caribou.

Strengths

• Model performance is reported with specific F1 scores (e.g., 93.7% on 2017 imagery).
• Includes an independent test set from a different year (2019) for validation.
• Released under a permissive CC-BY-4.0 license.
• Addresses specific, documented challenges like background heterogeneity and class imbalance.

Limitations

• The dataset is described in a 507.6 KB PDF; the actual image data files are not directly available.
• Column-level documentation and sample data are unavailable, requiring inference from the paper.
• Remaining model limitations include false positives from animal-like clutter and false negatives from occlusions in low-density areas.

Provenance

Source

figshare, authored by Ghazaleh Serati.

Collection Method

Aerial surveys providing high-resolution imagery across vast and remote Arctic regions.

Time Range

Imagery from 2017 and 2019 is mentioned.

Freshness

Last updated 2026-03-18 07:40:35.

Geography

Five caribou herds distributed across Alaska.