Three critical factors affecting automated image species recognition performance for camera traps

Area of Science:

• Ecology
• Computer Vision
• Wildlife Biology

Background:

• Manual analysis of camera trap images for wildlife monitoring is labor-intensive and time-consuming.
• Existing deep learning studies often require massive datasets and lack focus on performance in new environments.

Purpose of the Study:

• To evaluate deep learning capabilities for species identification using moderately sized camera trap datasets.
• To compare model performance on known versus novel geographic locations.
• To establish data requirements for expected performance levels in ecological studies.

Main Methods:

• Utilized a dataset of 47,279 images from 36 locations, representing 55 species and human activity.
• Trained and compared six deep learning computer vision networks (DenseNet201, Inception-ResNet-V3, InceptionV3, NASNetMobile, MobileNetV2, Xception) using transfer learning and data augmentation.
• Evaluated model accuracy on both trained and untrained locations.

Main Results:

• DenseNet201 achieved the highest top-1 accuracy (95.6%) on trained locations, indicating deep learning's potential for smaller research projects.
• Classification recall was significantly higher and more stable with over 1,000 training images (0.971 ± 0.0137) compared to fewer than 500 images (0.750 ± 0.329).
• Models performed less accurately on untrained locations, with DenseNet201 achieving 68.7% top-1 accuracy.

Conclusions:

• Deep learning models can be effective for wildlife species identification with modest datasets, particularly in familiar environments.
• Performance degrades in new locations, highlighting the need for location-specific training or transfer learning strategies.
• An open repository is provided for ecologists to develop custom species detection models.