Skip to content

Synthetic Images Helped Train Muskox Detectors When Real Aerial Photos Were Scarce

A herd of muskox in Bering Land Bridge National Preserve
Research · Other research

A study published on 13 February 2026 tested whether synthetic images could compensate for the shortage of labelled aerial photographs available for training automated muskox (Ovibos moschatus) detectors. The researchers used DALL-E 2, a diffusion-based image generator, to create aerial-style training images and then tested the resulting object-detection models on real muskox photographs from northern Quebec, Canada.

The study compared zero-shot models trained without any real muskox images with few-shot models that combined a small real training set with synthetic imagery. The generated images required substantial quality control: 84% were discarded because of unrealistic animals, anatomy or viewing angles, leaving 160 synthetic images for model development. A separate set of 996 real aerial photographs was reserved for testing.

Synthetic-only training already detected most muskoxen

Models trained only on synthetic images detected more than 80% of muskoxen in the real test photographs. Increasing the amount of synthetic training data improved precision, recall and F1 score, a combined measure of precision and recall, although the gains levelled off as more synthetic images were added and the models still performed below the baseline trained on real images.

Combining real and synthetic images improved recall

When the small real training set was combined with synthetic images, recall — the share of actual muskoxen that the detector found — increased from about 85% to as much as 93%, while overall performance became broadly comparable with the real-image baseline. The gain came with a trade-off: precision fell by about 4%, so the models missed fewer muskoxen but produced somewhat more false detections that would need human review. The increase in recall was not statistically significant, and the balance between precision and recall did not improve significantly overall. The authors therefore propose synthetic imagery as a way to start or strengthen wildlife detectors before large real datasets exist and then refine the models as field imagery accumulates.

Sources and image credit

  1. Durand, S., Foucher, S., Delplanque, A., Taillon, J., & Théau, J. (2026). Lacking data? No worries! How synthetic images can alleviate image scarcity in wildlife surveys: A case study with muskox (Ovibos moschatus). Remote Sensing in Ecology and Conservation, 12(4), 593–610. https://doi.org/10.1002/rse2.70056Primary peer-reviewed research article, first published 13 February 2026. The study tested synthetic diffusion-generated images for zero-shot and few-shot muskox detection in real aerial imagery.

Image: Bering Land Bridge National Preserve — muskox herd photographed by park biologist Marci on 2 November 2009 while observing the population and recording herd composition; Wikimedia Commons / Flickr; representative monitoring image, not an image from the 2026 study · CC BY 2.0 — https://creativecommons.org/licenses/by/2.0/ · Source

About this content: This story was produced with AI assistance within an editorial workflow developed by Wildlife Vagabond. Editorial responsibility remains with Wildlife Vagabond.How AI is used

Other research

SPECIES IN THIS STORY

Species in this story

Muskox Ovibos moschatus Explore species

Independent research and conservation news archive

Did you find this information useful?

Wildlife Vagabond is independently built and maintained. Voluntary support helps cover source verification, hosting and continued work on research and conservation news.

Support Wildlife Vagabond

The news archive will remain freely available.