Can CNN‐based species classification generalise across variation in habitat within a camera trap survey?
Problem
Manually annotating vast numbers of camera trap images is extremely time-consuming, and automated classifiers may fail to generalise across habitats with different disturbance levels.
Why it matters
Species data from camera traps underpins ecological monitoring and conservation decisions that annotation bottlenecks can delay.
What the AI did
Trained convolutional neural networks to classify species in camera trap images, comparing three architectures, training dataset configurations, and bounding-box cropping across tropical forest disturbance levels.
Results
Overall accuracy 72.8%, improved by excluding rarest species and adding extra training images
Impact
Shows automated classification is feasible at moderate accuracy but generalisability across habitats remains limited.
Record
| Venue | Methods in Ecology and Evolution |
|---|---|
| Published | 2022 |
| Work type | article |
| Authors | Danielle L. Norman, Philipp H. Bischoff, Oliver R. Wearn, Robert M. Ewers, J. Marcus Rowcliffe, Benjamin C. Evans, Sarab S. Sethi, Philip M. Chapman et al. |
| Citations | 31 (OpenAlex, at import time) |
| References | 32 |
| Model or method | CNNs; three architectures not named in source |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1111/2041-210x.14031 |
| Open access | hybrid |
| Keywords | camera traps, CNN, species classification, habitat disturbance, generalisability |
Importance rating
real but narrower advance
Camera trap annotation bottlenecks delay species data underpinning ecological monitoring and conservation; real but narrower benefit, with only moderate accuracy and poor generalisability shown.