Open-source AI model for UK camera trap species classification
The DeepFaune UK Model is an open-source species classification model for UK camera trap images. It has been developed to help conservation organisations, ecologists, universities, citizen science groups, and statutory agencies process large volumes of camera trap data without needing commercial platforms or machine learning expertise.
The model classifies UK-relevant species, including common mammals, selected bird species, domestic animals, people, and operational utility classes such as calibration poles. It is designed to run as part of a two-stage detect-then-classify pipeline, making it a natural companion to our Trap Tracker UK Mammals detection models.
DeepFaune UK is released as part of our wider aim to make conservation AI more accessible, transparent, and useful to the organisations that need it most.
Model Overview
DeepFaune UK is a Vision Transformer (ViT-L/DINOv2) classifier fine-tuned on a curated UK camera trap dataset. Unlike a model trained from scratch or adapted from general-purpose image weights, DeepFaune UK is fine-tuned from the DeepFaune European camera trap backbone – a foundation already adapted to the specific conditions of camera trap imagery (infrared, motion blur, awkward poses, cluttered backgrounds). This camera-trap-aware starting point is then specialised to UK species.
This matters because camera trap AI performance depends heavily on the species, landscapes, camera types, lighting conditions, and deployment contexts represented in the training data – and because a backbone that already understands camera trap imagery can transfer more robustly to new deployments than weights trained on everyday photographs.
The model is intended to provide a strong, practical classification stage for UK camera trap analysis, particularly for users who need local, offline, inspectable, and non-commercial tools.
How it works: a two-stage pipeline
DeepFaune UK is a classifier, not a standalone detector. It names an animal once that animal has been located and cropped from the full image. In practice it is run in two stages:
- Detection: a camera trap animal detector (for example the DeepFaune detector, or a TrapTracker UK detection model) locates animals, people, or vehicles in the full frame.
- Classification: each detected animal is cropped and passed to DeepFaune UK, which assigns a UK species label and a confidence score.
People and vehicles are handled at the detection stage and are not passed to the species classifier, which supports responsible, privacy-aware processing.
Supported Classes
DeepFaune UK classifies 30 UK-relevant classes, spanning mammals (including hedgehog, red and grey squirrel, roe / red / fallow deer, muntjac, fox, badger, pine marten, rabbit, wild boar, and domestic cattle, sheep, goat, horse, cat and dog), selected birds (including pheasant, wood pigeon, house sparrow, carrion crow, goshawk, buzzard, capercaillie, and curlew), people, and an operational calibration-pole class.
Training Data
The model was fine-tuned on labelled animal crops drawn from UK camera trap images collected through conservation deployments over nearly a decade.
The dataset includes a wide range of practical camera trap conditions, including RGB images, infrared night-time images, motion blur, partial occlusion, multi-animal frames, juveniles and adults, manufacturer overlays and timestamps, and varied UK habitats and deployment settings.
The raw training images are not released, because many were provided under partner-specific data agreements that pre-date these open model releases.
Training Setup
The model was fine-tuned using:
- Architecture: Vision Transformer ViT-L (DINOv2), fine-tuned from the DeepFaune European camera-trap backbone.
- Approach: transfer learning, with a linear-probe stage followed by full-tuning of a fresh 30-class UK classification head.
- Input resolution: 518 x 518.
- Training hardware: Blade with 8 x NVIDIA RTX A6000 GPUs
- Framework: PyTorch / timm
- Expert format: ONNX
The classification head is sized to 30 UK classes, and the trained model is fully self-contained – at inference it does not depend on the original DeepFaune weights.
Performance
On a held-out UK test set, DeepFaune UK achieves a macro-averaged F1 score of 0.992 across all 30 classes, with strong and even performance across species. The lowest-scoring classes are closely related deer (red, fallow, roe), reflecting genuine visual similarity rather than systematic error.
On the out-of-distribution evaluation (camera deployments not represented in training), the two stage pipeline produced high-confidence, ecologically plausible predictions across a 262-image set, with the great majority of detections classified at high confidence. Formal accuracy on out-of-distribution data should be validated against ground truth for any specific deployment; as with all camera trap AI, performance in a new setting depends on how closely that setting matches the training conditions.
Intended Use
This model is intended for non-commercial conservation and ecological monitoring, including:
- UK camera trap species classification biodiversity monitoring ecological research conservation project workflows university and student research citizen science projects local offline image processing pre-filtering large image collections before manual review
The model can be used through TrapTracker tools or integrated into other non-commercial workflows using the ONNX weights, paired with a suitable camera trap detector.
Not Intended For
The model is not intended for:
- commercial resale or paid inference services use as a general global wildlife model legally sensitive identification tasks surveillance or monitoring of people replacing expert ecological validation making conservation decisions without human review classifying species not included in the model classes
Where the results are used in scientific, regulatory, or conservation decision-making, users should include appropriate manual checking and validation.
Why We Are Releasing It
Camera trap AI has become increasingly important for biodiversity monitoring, but access to high-performing species recognition tools is often limited by commercial platforms, proprietary weights, pay-per-image models, or cloud-based inference.
This release is intended as a practical counterweight to that trend. The aim is simple: a conservation organisation with limited funding, no machine-learning team, and no GPU should still be able to process camera trap images locally and obtain useful species-level classifications.
Acknowledgements and Licence
DeepFaune UK is a derivative of the DeepFaune classification model and is built on the DeepFaune consortium’s camera-trap-adapted backbone. We gratefully acknowledge their work. Users of DeepFaune UK must also acknowledge the DeepFaune initiative.
Please cite DeepFaune alongside this model:
- Rigoudy, N., Dussert, G., the DeepFaune consortium, Spataro, B., Miele, V. & Chamaillé-Jammes, S. (2023) The DeepFaune initiative: a collaborative effort towards the automatic identification of the European fauna in camera-trap images. European Journal of Wildlife Research.
IMPORTANT — licence: Because DeepFaune UK is fine-tuned from DeepFaune’s model parameters, it inherits DeepFaune’s model-parameter licence, which is a Creative Commons ShareAlike licence. This means DeepFaune UK must be released under the same ShareAlike terms and with attribution to the DeepFaune consortium.
How To Cite
If you use the model in research, reports, publications, or conservation outputs, please cite both:
- Democratising Camera Trap AI: An Open-Source Model for Detecting UK Mammals.
- and the DeepFaune initiative (Rigoudy et al., 2023, above).
Download
By downloading the model, you agree to use it in accordance with its ShareAlike licence terms, to acknowledge both this model and the DeepFaune initiative, and to acknowledge the model appropriately in any resulting publications, reports, or public outputs.
The model is hosted on the Zenodo platform.
Updates
DeepFaune UK will continue to be developed as new data become available and as the underlying model architecture improves. Future releases will be made available periodically on the Zenodo platform, with clear versioning so that users can track changes, improvements, and compatibility across model updates.