An Ecologist-Friendly R Workflow for Classifying Camera Trap Images
Author Information
Author(s): Luca Petroni, L. Natucci, A. Massolo
Primary Institution: University of Pisa
Hypothesis
Can a user-friendly R workflow streamline the classification of camera trap images using deep learning?
Conclusion
The developed R workflow achieved over 90% accuracy in classifying camera trap images of various species.
Supporting Evidence
- The workflow achieved an overall precision of 0.962 and a recall of 0.945.
- Model accuracy reached 91.8% for unclassified images and 97.1% for those classified with >90% confidence.
- Training was conducted on a dataset of 21,861 images, including 21,500 for training and 5,080 for validation.
Takeaway
This study created a simple way for scientists to quickly sort pictures of animals taken by cameras, making it easier to study wildlife.
Methodology
The workflow integrates deep learning models and image augmentation techniques to classify camera trap images in R.
Potential Biases
Potential biases in model performance due to imbalanced datasets and reliance on previously tagged images.
Limitations
The workflow is currently limited to images rather than videos and may require customization for different operating systems.
Participant Demographics
Images collected from various camera trap sites in Central Italy targeting medium and large mammals.
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website